Energy Disaggregation Competition Survey (Responses)
 Share
The version of the browser you are using is no longer supported. Please upgrade to a supported browser.Dismiss

 
View only
 
 
Still loading...
ABCDEFGHIJKLMNOPQRSTUVWXYZAAABACADAEAFAGAHAIAJAK
1
TimestampWhat is your affiliation?Are you happy to use your real name (or your company's real name) in the competition?Would you like the competition to run continuously or yearly?
Assuming we meet all your requirements, would you be happy to participate in a NILM competition?
Performance metricsCan your algorithm disaggregate American appliances?Which "disaggregation tracks" would you be interested in competing?
If we could provide data from any geographical region, which would be most relevant for your algorithm(s)?
Kaggle or custom-built platform?What if we don't provide any training data? Would you still be able to compete?Do you plan to train *only* on data provided by the competition?Are you happy to declare that you have trained on additional data?Can you train your algorithm twice?Which features do you *absolutely require* in the input data?Which input features, additional to those you listed above, can your algorithm optionally exploit?What metadata about each house do you absolutely require?What metadata about each house can you optionally exploit?Which features does your algorithm output?What's the minimum duration of training data that you require?How much "warm up" data does your algorithm require for disaggregation?Is this sufficient data: 1 year of data across 5 homes for training & 1 year of data across 5 different homes for testing; all data from the US
For which categories does your algorithm estimate energy consumption for?
Please list which appliances your algorithm can detect
Which non-appliance categories does your algorithm detect?
For the input data, what is the fastest sample rate that your algorithm can handle?
For the input data, what is the slowest sample rate that your algorithm can handle?
We intend to provide data at five temporal resolutions: 1 Hz; every 10 seconds; minutely; half-hourly and hourly. Are you happy with this set of temporal resolutions?
If the competition provided only a *single* sample rate then which would you prefer?
What is the temporal resolution of your algorithm's output?Should the competition try to measure the time it takes each algorithm to perform disaggregation?Any other suggestions?
After "warming up", what's the minimum duration of data that your algorithm can disaggregate per house?
How many instances of each appliance do you require in the training dataset?Which NILM use-cases are most relevant to your work?Would you be happy to submit an executable program instead of data?
2
26/09/2016 17:10:16CompanyI will remain anonymous, even if my algorithm does well.
Continuously. Teams can submit disaggregated estimates once per day and the published metrics will update immediately. The dataset will not change. There is a risk that teams will overfit to the competition data or manually disaggregate the data.
Yes
I'm not sure. My algorithm should be able to disaggregate American appliances but all of my research and development has been done on non-American appliances.
I'd actually *prefer* the competition to not provide any training data.
Technically, I *could* train my algorithm on just the data provided by the competition. But the requirement to train twice would be so much hassle that it would prevent me from participating.
Real (active) powerYesOne monthIndividual appliances, End-use categories (e.g. "cooling")secret ;)0.1 Hz0.1 HzSame as the aggregate input
3
28/09/2016 14:55:35Non-profit organisationI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.YesMy algorithm can definitely disaggregate American appliancesDomestic electricity consumption, Domestic water consumption, Domestic natural gas consumption
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Real (active) power, CurrentReal (active) power, Reactive power, Apparent power, Voltage, CurrentRough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), Building type (e.g. detached, mid-terrace, bungalow etc.)
Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), The fuel used for space heating (e.g. electric, natural gas, oil, CHP or district heating)
CurrentNo minimum.One weekThat's fineEnd-use categories (e.g. "cooling")MinutelyMinutelyYes. That set of temporal resolution is fine.Provide only a low sample rate (e.g. minutely) to level the playing fieldSame as the aggregate inputNo
4
29/09/2016 11:37:30AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.YesI think four of them are enough.
I'm not sure. My algorithm should be able to disaggregate American appliances but all of my research and development has been done on non-American appliances.
Yes, I will only train on data provided by the competition.Apparent power, Phase angle, Mains frequencyInternal temperature, External temperature, Appliance state reported by smart appliancesA list of appliancesApparent power, Phase angleOne quarterOne monthThat's fineIndividual appliancesFridge / freezer, Dish washer, Cooking oven, Air conditioning unit, Computer, Kettle, Toaster, Microwave, Televisioncooling, heating> 1 Hz and ≤ 1 kHz1 Hz
Provide two or three sample rates. That way, we still get to explore the difference in disggregation performance between different sample rates but we also get to use more test data per sample rate.
1 HzSame as the aggregate inputNo
I think it should be defined how to calculate the score for ranking.
5
30/09/2016 16:39:56
6
09/10/2016 06:16:00AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesMy algorithm can disaggregate American appliances
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Americas
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) powerReactive power, Apparent power, Voltage, CurrentA list of appliancesRough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)Real (active) powerOne weekNo "warm up" data is required.That's fineIndividual appliances1 HzEvery 10 seconds (0.1 Hz)Yes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.No minimum1-3Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances.
7
10/10/2016 12:29:40AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Europe
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Real (active) powerApparent powerA list of appliancesReal (active) powerOne weekOne weekThat's fine
Individual appliances, User behaviours (e.g. "cooking dinner on Tuesday night used X kWh")
Fridge / freezer, Clothes dryer, Clothes washer, Dish washer, Kettle, Toaster, Microwave> 1 Hz and ≤ 1 kHzEvery 10 seconds (0.1 Hz)Yes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputYes, formally measure execution timeOne day1-3Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., List most or all of the appliances in each home
8
10/10/2016 15:04:47CompanyIf no cheating is possible, reveal name would be OKYessorry, don't have the time right to go through this question right now :(In the cas of American appliances, we would need to perform specific trainingNatural gas disaggregation, Disaggregation of data from commercial buildings (e.g. office blocks), Separate tracks for different geographical regions
Europe, Oceania (Australia & Melanesia & Micronesia & Polynesia)
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I'd actually *prefer* the competition to not provide any training data, so you can double the size of the testing dataset.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) power, External temperatureReactive power, Apparent power, Power factor, Internal temperatureA list of appliances, Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), Number of occupantsConstruction date (e.g. to infer the thermal performance of the building)Real (active) powerOne monthNo "warm up" data is required.for testing, considering the variety of combinations, we feel that the appropriate size would be more like 100 homes...
Individual appliances, End-use categories (e.g. "cooling"), User behaviours (e.g. "cooking dinner on Tuesday night used X kWh")
Fridge / freezer, Lighting, Clothes dryer, Clothes washer, Dish washer, Cooking oven, Air conditioning unit, Electric space heater, Electric domestic water heater
cooling, heating, base load
1 HzMinutelyYes. That set of temporal resolutions is fine.MinutelyWeekly
since execution time is related to the harware, that doesn't seem like a fair/accurate measurement (but imposing a maximum duration of a few minutes could be a way to limit cheating)
One week
I will be training on additional data so it doesn't matter how much data is provided by the competition
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances.
9
10/10/2016 16:48:51
10
10/10/2016 19:09:22CompanyI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!
Continuously. Teams can submit disaggregated estimates once per day and the published metrics will update immediately. The dataset will not change. There is a risk that teams will overfit to the competition data or manually disaggregate the data.
YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Europe
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Real (active) power, Reactive power, Voltage, Current, Total harmonic distortion, Mains frequencyReal (active) power, Reactive powerOne weekOne monthThat's fineIndividual appliances, End-use categories (e.g. "cooling")> 1 kHz and ≤ 1 MHz1 Hz
Provide two or three sample rates. That way, we still get to explore the difference in disggregation performance between different sample rates but we also get to use more test data per sample rate.
1 HzSame as the aggregate inputYes, formally measure execution timeNo minimum>9
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
11
11/10/2016 08:45:17Company
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
EuropeRun a supervised NILM competition on Kaggle but also build a web application to run more detailed analysis of the results.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
No, I will train on additional data (either a private dataset or additional public datasets)Yes, I'm happy to declare that I have trained on additional dataReal (active) power, Reactive powerA list of appliancesReal (active) power> 1 Hz and ≤ 1 kHz1 Hz1 HzYes, formally measure execution timeOne day>9
12
11/10/2016 15:52:36AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.YesMy algorithm can disaggregate American appliancesWater disaggregation, Natural gas disaggregation, Reactive power disaggregationAmericas, Europe
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I could not compete.No, I will train on additional data (either a private dataset or additional public datasets)Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) powerReal (active) power, Reactive power, CurrentA list of appliancesReal (active) power, Reactive powerOne yearOne yearThat's fineIndividual appliancesFridge / freezer, Clothes dryer, Clothes washer, Dish washer, Cooking oven, Kettle, Microwave1 HzMinutelyYes. That set of temporal resolutions is fine.MinutelySame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.No minimum1-3
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
13
11/10/2016 15:54:38CompanyI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Once per year. Maybe high-level results could be revealed before the formal announcement.YesAbsolute and relative error sound good.My algorithm can disaggregate American appliances
Water disaggregation, Natural gas disaggregation, Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regions
Americas, Europe
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I'd actually *prefer* the competition to not provide any training data, so you can double the size of the testing dataset.
No, I will train on additional data (either a private dataset or additional public datasets)Yes, I'm happy to declare that I have trained on additional dataOur algorithm does not require any training data provided by the competition and we would want to ensure that we are evaluated as such.Real (active) power, External temperatureRough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)
A list of appliances, To know whether each home has electric space heating / water heating / air conditioning., Number of occupants, Construction date (e.g. to infer the thermal performance of the building), Building type (e.g. detached, mid-terrace, bungalow etc.)
Real (active) powerWe would prefer no training data.
Individual appliances, End-use categories (e.g. "cooling"), User behaviours (e.g. "cooking dinner on Tuesday night used X kWh")
Fridge / freezer, Lighting, Clothes dryer, Dish washer, Cooking oven, Swimming pool (or spa) pump, Air conditioning unit, Electric space heater, Electric domestic water heater, Electric vehicle charging, Televisionheating/cooling1 HzHourlyWe would prefer replacing 30min with 15min15 minutelyDepends on the sampling rateNo, don't make any attempt to measure execution time.Good luck!
I will be training on additional data so it doesn't matter how much data is provided by the competition
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., List most or all of the appliances in each home
14
11/10/2016 16:37:30Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.I'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Water disaggregation, Natural gas disaggregation, Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regions
Americas
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Real (active) power, Reactive power, Apparent power, Voltage, Current, Power factor, Phase angle, Total harmonic distortion, Mains frequencyMovement detectors (e.g. PIR, ultrasonic, door sensors)A list of appliances, To know whether each home has electric space heating / water heating / air conditioning.Number of occupants, Building type (e.g. detached, mid-terrace, bungalow etc.)Real (active) power, Reactive power, Apparent powerOne monthOne monthThat's fineIndividual appliancesFridge / freezer, Clothes dryer, Clothes washer, Dish washer, Cooking oven, Kettle, Toaster, Microwave, Television1 HzEvery 10 seconds (0.1 Hz)Yes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputYes, formally measure execution time
probably, depends on my schedule.
One year7-9
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
15
11/10/2016 16:38:30Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesI agree ( MSE/RMSE also) My algorithm can disaggregate American appliances
Water disaggregation, Natural gas disaggregation, Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regions
Americas, Asia, Europe
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
No, I will train on additional data (either a private dataset or additional public datasets)Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) power, Reactive power, Apparent power, Current
Water consumption, Natural gas / heating oil consumption, Internal temperature, External temperature, Sunshine, Appliance state reported by smart appliances, Movement detectors (e.g. PIR, ultrasonic, door sensors), Proximity of users' mobile phones to the house / internal navigation
A list of appliances, To know whether each home has electric space heating / water heating / air conditioning.Number of occupants, Building type (e.g. detached, mid-terrace, bungalow etc.)
Real (active) power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset), Human behaviours (e.g. "cooking dinner"), Water consumption of target appliances, Natural gas / heating oil consumption of target appliances
One dayOne monthThat's fine
Individual appliances, End-use categories (e.g. "cooling"), User behaviours (e.g. "cooking dinner on Tuesday night used X kWh"), Energy consumed per occupant in a multi-occupant home (e.g. "John used X kWh; Eve used Y kWh")
Fridge / freezer, Lighting, Clothes dryer, Air conditioning unit, Electric space heater, Electric domestic water heater, Toaster, Microwave, Television1 HzMinutely, 15 minutely, Half hourly, HourlyYes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.One month
I will be training on additional data so it doesn't matter how much data is provided by the competition
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
16
11/10/2016 16:58:23Government R&D
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesMy algorithm can disaggregate American appliancesWater disaggregation, Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Supervised disaggregationAmericas
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
We could only declare if we've used additional data if our team name remains anonymous for the duration of the competition
I will only submit once; training on whatever data I can get my hands on. Technically, I *could* train my algorithm on just the data provided by the competition. But it would be too much hassle to train twice.
Real (active) power, Reactive power, Water consumption, Appliance state reported by smart appliancesTotal harmonic distortion, Movement detectors (e.g. PIR, ultrasonic, door sensors)A list of appliances, To know whether each home has electric space heating / water heating / air conditioning.
Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), Number of occupants, Construction date (e.g. to infer the thermal performance of the building)
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset), Water consumption of target appliances
One weekNo "warm up" data is required.That's fineIndividual appliances, End-use categories (e.g. "cooling")
Fridge / freezer, Lighting, Clothes dryer, Clothes washer, Dish washer, Cooking oven, Swimming pool (or spa) pump, Air conditioning unit, Electric space heater, Electric domestic water heater, Electric vehicle charging, Computer, Kettle, Toaster, Microwave, Television
> 1 Hz and ≤ 1 kHz1 HzYes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.
Thank you for the initiative!
No minimum>9
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
17
11/10/2016 17:02:06CompanyI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!It might be good to do a Kaggle competition once as a trial of what it could be. I'm in favor of a yearly winner while having a continuous leader board. There is something to be said to have a deadline and a point were leaders can be acknowledged.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Water disaggregationAmericas, Europe
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
only if training data is provided would you get participation from the wider data science communityNo, I will train on additional data (either a private dataset or additional public datasets)
It must be assumed that those versed in the field will use additional data. "declaring" or not it must be assumed that people will use additional data.
I will only submit once; training on whatever data I can get my hands on. Technically, I *could* train my algorithm on just the data provided by the competition. But it would be too much hassle to train twice.
Real (active) power, Voltage, Current, Water consumption, Natural gas / heating oil consumption, Appliance state reported by smart appliances
Reactive power, Apparent power, Phase angle, Mains frequency, Internal temperature, External temperature, Sunshine, Movement detectors (e.g. PIR, ultrasonic, door sensors), Proximity of users' mobile phones to the house / internal navigation
A list of appliances, To know whether each home has electric space heating / water heating / air conditioning., Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), Number of occupants, Construction date (e.g. to infer the thermal performance of the building), Building type (e.g. detached, mid-terrace, bungalow etc.)
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset), Human behaviours (e.g. "cooking dinner"), Water consumption of target appliances, Natural gas / heating oil consumption of target appliances
One monthNo "warm up" data is required.always more is betterUser behaviours (e.g. "cooking dinner on Tuesday night used X kWh")Fridge / freezer, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Swimming pool (or spa) pump, Air conditioning unit, Electric space heater, Electric domestic water heater, Toaster, Microwave
water use: shower, toilet, sink
1 Hz15 minutely
Provide two or three sample rates. That way, we still get to explore the difference in disggregation performance between different sample rates but we also get to use more test data per sample rate.
1 HzSame as the aggregate inputYes, formally measure execution timeOne month>9Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
18
11/10/2016 18:25:37CompanyI will remain anonymous throughout the competition, even if my algorithm does well.Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesMy algorithm can disaggregate American appliancesDisaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")Americas, Europe
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I'd actually *prefer* the competition to not provide any training data, so you can double the size of the testing dataset.
I'm not sure (please select this option if there is any chance that you might train on additional data)
We could only declare if we've used additional data if our team name remains anonymous for the duration of the competition
I will only submit once; training on whatever data I can get my hands on. Technically, I *could* train my algorithm on just the data provided by the competition. But it would be too much hassle to train twice.
Real (active) power, Reactive power, Voltage, Current, Power factor, Phase angle, Total harmonic distortion, Mains frequencyA list of appliancesA list of appliances
Real (active) power, Reactive power, Voltage, Current, Power factor, Phase angle, Total harmonic distortion, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
Half a yearOne monthI'd prefer 6 months x 10 homes for both training and test setsIndividual appliances> 1 kHz and ≤ 1 MHz> 1 kHz and ≤ 1 MHz1 HzSame as the aggregate inputYes, formally measure execution timeOne month
I will be training on additional data so it doesn't matter how much data is provided by the competition
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances.
19
11/10/2016 18:50:21AcademicI will remain anonymous throughout the competition, even if my algorithm does well.My algorithm can disaggregate American appliances
Water disaggregation, Natural gas disaggregation, Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regions
Europe
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.
Real (active) power, Reactive power, Apparent power, Voltage, Current, Power factor, Phase angle, Natural gas / heating oil consumption, Internal temperature, Appliance state reported by smart appliances
Reactive powerRough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)Construction date (e.g. to infer the thermal performance of the building)Power factor> 1 kHz and ≤ 1 MHz>9
20
11/10/2016 19:25:36Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Disaggregation of data from commercial buildings (e.g. office blocks), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I could not compete.Yes, I will only train on data provided by the competition.Real (active) powerA list of appliances
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One weekOne weekThat's fineIndividual appliances> 1 Hz and ≤ 1 kHzYes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.
A dataset with the aim of predicting private information would be interesting. However, such a dataset will be hard to get (fortunately, from the privacy view)
No minimum1-3
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
21
11/10/2016 19:30:16AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regions
Europe
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I'd actually *prefer* the competition to not provide any training data, so you can double the size of the testing dataset.
No, I will train on additional data (either a private dataset or additional public datasets)Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) power, Reactive power, Apparent power, Voltage, CurrentPower factor, Phase angle, Total harmonic distortionA list of appliances
To know whether each home has electric space heating / water heating / air conditioning., Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
I'd prefer 6 months x 10 homes for both training and test setsIndividual appliancesFridge / freezer, Lighting, Clothes dryer, Dish washer, Cooking oven, Air conditioning unit, Computer, Kettle, Toaster, Microwave, Television
Vampire Loads and continuously variable
> 1 Hz and ≤ 1 kHzNo. Keep life simple (at least for the first competition). Just provide a single sample rate.1 HzDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.
I will be training on additional data so it doesn't matter how much data is provided by the competition
Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
22
11/10/2016 20:36:59AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesMy algorithm can disaggregate American appliancesDisaggregation of data from commercial buildings (e.g. office blocks)AmericasRun a supervised NILM competition on Kaggle but also build a web application to run more detailed analysis of the results.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Real (active) power
Reactive power, Apparent power, Voltage, Current, Power factor, Phase angle, Total harmonic distortion, Mains frequency, Internal temperature, External temperature, Sunshine
A list of appliances
Real (active) power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One monthNo "warm up" data is required.That's fineIndividual appliancesFridge / freezer, Lighting, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Microwave, Any appliance > ~30W as long as there's training data available> 1 kHz and ≤ 1 MHz15 minutelyYes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.No minimum1-3
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
23
11/10/2016 21:29:16CompanyNot yet sure if we want to participateUndecided. Depends on the design.NoIdentification of abnormally high end use load categories in a specific home or building, for use in improving efficiency of electric, heating fuel and water use."NILM" refers to loads, not devices. We analyze loads.Water disaggregation, Natural gas disaggregation, Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Custom-defined load categoriesAmericasI think you're putting the cart before the horse. Define the competition first before building the platform.Lack of ground truth data is the key challenge.No, I will train on additional data (either a private dataset or additional public datasets)"Training" is not involved in our process, so maybe we don't fit your definition of NILM.
We've tested on analysis on thousands of homes, so would not modify them based on a small set of test data; especially if it was atypical (e.g. experimental data).
External temperature, hourly energy (kWh), not powerWater consumption, Natural gas / heating oil consumption, survey responses to targeted questions from occupantsRough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)Number of occupants, Construction date (e.g. to infer the thermal performance of the building), Building type (e.g. detached, mid-terrace, bungalow etc.), Directed Q&A with occupantsMonthly consumption by load categoryOne yearOne yearDepends what you are testingEnd-use categories (e.g. "cooling")
We can't identify any of these devices with 100% certainty, so we don't claim to. Every home is different, many devices have similar load profiles, and false positives have a terrible effect. This is the "bane" of the NILM industry and your competition -- if it highlights such poor results -- will not help with this issue.
Heating, Cooling, Always On, Recurring and Variable/Behavioral. All 5 for electricity, 3 for heating fuel, and 3 for water.
15 minutely
Hourly, This is for electricity. Different answers for water and heating fuel.
you can't compare results across different resolutionsHourlyMonthlyThe key issue (for scale) is whether there is any human interventionI'll send some emails...One month
I will be training on additional data so it doesn't matter how much data is provided by the competition
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances.
24
11/10/2016 22:12:03AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!
Continuously. Teams can submit disaggregated estimates once per day and the published metrics will update immediately. The dataset will not change. There is a risk that teams will overfit to the competition data or manually disaggregate the data.
YesThat's fineI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Water disaggregation, Natural gas disaggregation, Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Americas, Oceania (Australia & Melanesia & Micronesia & Polynesia)
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Real (active) power
Internal temperature, External temperature, Sunshine, Proximity of users' mobile phones to the house / internal navigation
A list of appliances
To know whether each home has electric space heating / water heating / air conditioning., Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), Number of occupants, Building type (e.g. detached, mid-terrace, bungalow etc.)
Real (active) power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One yearOne weekThat's fineIndividual appliances
Fridge / freezer, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Swimming pool (or spa) pump, Pond pump, Swimming pool (or spa) heater, Air conditioning unit, Electric space heater, Electric domestic water heater, Electric vehicle charging, Kettle, Toaster, Microwave
> 1 Hz and ≤ 1 kHzHalf hourlyYes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputYes, formally measure execution timeNo minimum4-6
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
25
11/10/2016 22:19:32AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesMy algorithm can disaggregate American appliances
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regions
EuropeUse a ready-made web application for the competition, which is free and open source. Probably have a look if mooshak can do this.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional data
I will only submit once; training on whatever data I can get my hands on. Technically, I *could* train my algorithm on just the data provided by the competition. But it would be too much hassle to train twice.
Real (active) powerReactive powerA list of appliancesA list of appliances
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One weekOne weekThat's fineIndividual appliances, End-use categories (e.g. "cooling")1 HzEvery 10 seconds (0.1 Hz)Yes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputYes, formally measure execution timeNo minimum1-3
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
26
12/10/2016 01:32:17Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.Water disaggregation, Natural gas disaggregation, Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.)
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
We could only declare if we've used additional data if our team name remains anonymous for the duration of the competition
I will only submit once; training on whatever data I can get my hands on. Technically, I *could* train my algorithm on just the data provided by the competition. But it would be too much hassle to train twice.
4-6
27
12/10/2016 06:02:42Company
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Water disaggregation, Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Americas, Europe
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
No, I will train on additional data (either a private dataset or additional public datasets)Yes, I'm happy to declare that I have trained on additional data
I will only submit once; training on whatever data I can get my hands on. Technically, I *could* train my algorithm on just the data provided by the competition. But it would be too much hassle to train twice.
Real (active) powerReactive power, Apparent powerA list of appliancesTo know whether each home has electric space heating / water heating / air conditioning.
Real (active) power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One monthOne weekI'd prefer 6 months x 10 homes for both training and test setsIndividual appliancesFridge / freezer, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Air conditioning unit, Electric space heater, Electric domestic water heater, Kettle, Toaster, Microwave, Television1 Hz15 minutelyYes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputYes, formally measure execution timeOne day7-9Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
28
12/10/2016 09:51:51AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!
Continuously. Teams can submit disaggregated estimates once per day and the published metrics will update immediately. The dataset will not change. There is a risk that teams will overfit to the competition data or manually disaggregate the data.
YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regions
Americas, Europe
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I could not compete.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) power, Reactive power, Apparent power, Voltage, Current, Power factor, Total harmonic distortion, Mains frequencyA list of appliancesTo know whether each home has electric space heating / water heating / air conditioning.One weekOne weekThat's fineIndividual appliancesFridge / freezer, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Electric space heater, Computer, Kettle, Toaster, Microwave, Television> 1 kHz and ≤ 1 MHz> 1 Hz and ≤ 1 kHz
Provide two or three sample rates. That way, we still get to explore the difference in disggregation performance between different sample rates but we also get to use more test data per sample rate.
1 HzHourlyDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.One day1-3Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
29
12/10/2016 14:53:15AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.)Asia
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I could not compete.Yes, I will only train on data provided by the competition.Real (active) power, Reactive power, Power factor, Phase angle, Total harmonic distortion, Appliance state reported by smart appliancesReal (active) power, Reactive power, Voltage, Current, Power factor, Phase angle, Total harmonic distortionA list of appliances, To know whether each home has electric space heating / water heating / air conditioning.A list of appliancesReal (active) powerNo minimum.No "warm up" data is required.That's fineIndividual appliancesLighting, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Electric domestic water heater, Electric vehicle charging, Computer, Kettle, Toaster, Microwave, Television> 1 MHz> 1 Hz and ≤ 1 kHzYes. That set of temporal resolutions is fine.1 HzHourlyDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.No minimum>9Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
30
12/10/2016 19:41:32AcademicI will remain anonymous throughout the competition, even if my algorithm does well.Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.Yes
I think that the per appliance on-threshold should be learned for each appliance independently (and not a fixed threshold). Just learn what the baseline load is for each appliance and add some epsilon to it.
I'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Disaggregation of data from residential buildings.
Americas, EuropeCould you create multiple Kaggle competitions, at least for multiple datasets (and also build a web application to run more detailed analysis of the results)
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
It seems like there should be two tracks: those that use additional data and those that do not
I will only submit once; training on whatever data I can get my hands on. Technically, I *could* train my algorithm on just the data provided by the competition. But it would be too much hassle to train twice.
Real (active) powerA list of appliances, To know whether each home has electric space heating / water heating / air conditioning., Number of occupantsReal (active) powerOne monthNo "warm up" data is required.That's fineIndividual appliances, End-use categories (e.g. "cooling")Fridge / freezer, Clothes dryer, Clothes washer, Dish washer, Cooking oven, Air conditioning unit, Electric space heater, Kettle, Microwave1 HzMinutelyYes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputNo, don't make any attempt to measure execution time.No minimum>9Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
31
12/10/2016 22:47:12Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
how about monthly or a quarter year?Yes
I would maybe add the mean absolute percentage error (MAPE) in the first case. I found at least two different approaches for accuracy - the metrics should therefore be formalized somewhere.
I'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Water disaggregation, Natural gas disaggregation, Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regions, How about some Grid Characteristics?
Africa, Asia, Oceania (Australia & Melanesia & Micronesia & Polynesia)
Run a supervised NILM competition on Kaggle but also build a web application to run more detailed analysis of the results.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) power, Reactive power, Voltage, Current, Power factor, Total harmonic distortion, HarmonicsMains frequencyA list of appliances
To know whether each home has electric space heating / water heating / air conditioning., Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), Number of occupants, 1,2 or 3 phase household and maybe maximum fuse capacity.
Real (active) power, Reactive power, Voltage, Current, Power factor, Total harmonic distortion, Harmonics
One quarterNo "warm up" data is required.Its fine, but cross validation is impossible to prevent cheating?Individual appliances, start-up events
Fridge / freezer, Lighting, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Swimming pool (or spa) pump, Pond pump, Air conditioning unit, Electric space heater, Electric domestic water heater, Computer, Kettle, Toaster, Microwave, Television, if training samples are provided
resistive, inductive, capacitive
> 1 kHz and ≤ 1 MHz1 kHz
Provide two or three sample rates. That way, we still get to explore the difference in disggregation performance between different sample rates but we also get to use more test data per sample rate.
1 Hzminutely or even lowerYes, formally measure execution time
Try to get a high frequency dataset :)
No minimum7-9Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
32
13/10/2016 11:26:19CompanyI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.NoMy algorithm can disaggregate American appliancesDisaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Separate tracks for different geographical regionsAmericas, EuropeRun a supervised NILM competition on Kaggle but also build a web application to run more detailed analysis of the results.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Real (active) power, Reactive powerA list of appliancesReal (active) powerOne quarterNo "warm up" data is required.That's fineIndividual appliances
Fridge / freezer, Lighting, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Swimming pool (or spa) pump, Pond pump, Swimming pool (or spa) heater, Air conditioning unit, Electric space heater, Electric domestic water heater, Electric vehicle charging, Computer, Kettle, Toaster, Microwave, Television
> 1 Hz and ≤ 1 kHzMinutelyNo. Keep life simple (at least for the first competition). Just provide a single sample rate.1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.No minimum1-3
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
Yes, I'd be happy to upload a trained executable program to the competiton.
33
13/10/2016 20:03:36Company
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
YesMy algorithm can disaggregate American appliances
Water disaggregation, Natural gas disaggregation, Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Americas, Asia, Europe
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) power, Reactive power, Apparent power, Current, Power factor, Phase angle, Total harmonic distortion, Mains frequency, ~10 kHz sampling rate
Real (active) power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One weekOne weekThat's fineIndividual appliances> 1 kHz and ≤ 1 MHz> 1 kHz and ≤ 1 MHzYes. That set of temporal resolutions is fine.~10khzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.One day
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
Yes, I'd be happy to upload a trained executable program to the competiton.
34
14/10/2016 03:31:17Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesMy algorithm can disaggregate American appliancesDisaggregation of data from commercial buildings (e.g. office blocks)Americas
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
We could only declare if we've used additional data if our team name remains anonymous for the duration of the competition
I'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) power, Reactive power, Voltage, CurrentApparent power, Power factor, Mains frequencyA list of appliancesRough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
No minimum.No "warm up" data is required.That's fineIndividual appliancesFridge / freezer, Lighting, Clothes dryer, Air conditioning unit, Computer, Microwave> 1 kHz and ≤ 1 MHz> 1 Hz and ≤ 1 kHzYes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.No minimum>9
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
Yes, I'd be happy to upload a trained executable program to the competiton.
35
14/10/2016 09:22:06Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesSounds reasonable.I'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Separate tracks for different geographical regions
Americas, Europe, Oceania (Australia & Melanesia & Micronesia & Polynesia)
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
No, I will train on additional data (either a private dataset or additional public datasets)Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) powerVoltage, Current, Appliance state reported by smart appliancesA list of appliances, Number of occupants
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One monthOne monthThat's fineIndividual appliancesFridge / freezer, Clothes washer, Dish washer, Kettle, Toaster, Microwave> 1 Hz and ≤ 1 kHzMinutely
Provide two or three sample rates. That way, we still get to explore the difference in disggregation performance between different sample rates but we also get to use more test data per sample rate.
1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.One day
I will be training on additional data so it doesn't matter how much data is provided by the competition
Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.No, I definitely cannot upload an executable program. This would be a deal breaker.
36
15/10/2016 23:19:09CompanyI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.Yes
I think the first case of weekly summary of estimated total energy consumed by each appliance it not a bad option because they may have high precision but the NILM algorithm is not good. It may happen for example one appliance worked 2 houres total yesterday but the system detected it worked just 1h. But today, that appliance works just 1h in total but system detectes it works 2h. Then, the weekly summary of working houre is correct but the NILM did not work well.

I propose a metric call the precision in detecting "appliance type" and "their habit". So that, it can cover your both three use-cases: in their type, their status ON/OFF at each time steps, and their total power usage per week.
I'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.)Europe
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Voltage, Current, Mains frequencyReal (active) power, Reactive power, Internal temperature, External temperatureA list of appliances
To know whether each home has electric space heating / water heating / air conditioning., Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)
Real (active) power, Reactive power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One monthOne weekI'd prefer 6 months x 10 homes for both training and test setsEnd-use categories (e.g. "cooling")> 1 MHz> 1 MHzIt is better to provide original data current and voltage at high resolutionAs high sampling as possibleminuteNo, don't make any attempt to measure execution time.One month
I will be training on additional data so it doesn't matter how much data is provided by the competition
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances.No, I definitely cannot upload an executable program. This would be a deal breaker.
37
17/10/2016 13:13:18AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!
Continuously. Teams can submit disaggregated estimates once per day and the published metrics will update immediately. The dataset will not change. There is a risk that teams will overfit to the competition data or manually disaggregate the data.
YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Disaggregation of data from commercial buildings (e.g. office blocks), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")Americas, Europe
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I'd actually *prefer* the competition to not provide any training data, so you can double the size of the testing dataset.
Yes, I will only train on data provided by the competition.Voltage, Current
Real (active) power, Reactive power, Apparent power, Power factor, Phase angle, Total harmonic distortion, Mains frequency
A list of appliances
To know whether each home has electric space heating / water heating / air conditioning., Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), Number of occupants
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset), Human behaviours (e.g. "cooking dinner"), Energy consumed by each appliance
One weekOne weekThat's fineIndividual appliances, End-use categories (e.g. "cooling")Fridge / freezer, Lighting, Clothes dryer, Dish washer, Air conditioning unit, Electric space heater, Electric domestic water heater, Computer, Kettle, Toaster, Microwave, Television, iron, coffee machine, juicer, mixer> 1 kHz and ≤ 1 MHz1 HzYes. That set of temporal resolutions is fine.2 kHzSame as the aggregate inputYes, formally measure execution timeOne day1-3
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
Yes, I'd be happy to upload a trained executable program to the competiton.
38
17/10/2016 13:20:14Company
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Separate tracks for different geographical regionsAmericas, Europe
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional data
Our algorithm already has many parameters coming from data that we collected. It would be a big hassle to make the algorithm "forget everything" and use only parameters coming from the new training data (building new models, etc).
Real (active) power, Reactive power, VoltageA list of appliances, To know whether each home has electric space heating / water heating / air conditioning.Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), Number of occupantsReal (active) powerOne monthOne week
One of the main challenges we found in the NILM problem was "generalisation" of the problem. It is relatively easy to make a NILM software working in 1 house, or 2, or even 5, because you can fine tune the algorithm for those particular cases. The actual difficulty is having an algorithm capable to work on, say, 50 homes or more so that the only chance is to make a generalistic algorithm. I'd prefer 1 month 60 homes! (I know it is much more difficult, it is just my view on it).
Individual appliances, End-use categories (e.g. "cooling")Fridge / freezer, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Swimming pool (or spa) pump, Pond pump, Air conditioning unit, Electric space heater, Electric domestic water heater, Kettle, Toaster, MicrowaveBase load1 HzEvery 10 seconds (0.1 Hz)Yes. That set of temporal resolutions is fine.1 HzDailyNo, don't make any attempt to measure execution time.
This competition looks great. It is noticeable that you guys have been long thinking about it! We keep talking!
One day4-6Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances.No, I definitely cannot upload an executable program. This would be a deal breaker.
39
18/10/2016 00:57:51CompanyI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Run the competition continuously but have a changing dataset to avoid teams overfitting to competition data and manual disaggregation. YesMy algorithm can disaggregate American appliancesDisaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Separate tracks for different geographical regions
Oceania (Australia & Melanesia & Micronesia & Polynesia)
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onVoltage, CurrentTotal harmonic distortionA list of appliances
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One weekNo "warm up" data is required.That's fineIndividual appliancesFridge / freezer, Lighting, Cooking hob, Cooking oven, Swimming pool (or spa) heater, Electric space heater, Computer, Kettle, Toaster, Television, Untested on other appliance types but may still work> 1 kHz and ≤ 1 MHz> 1 kHz and ≤ 1 MHzHappy with the current resolutions, but a sample rate of in the KHz range would be better for us. 10 - 20KHz would be ideal.Same as the aggregate inputIt would be interesting to have a formal measurement, without it being used in competition judging. Ie. an FYI especially if we are providing executables.No minimum1-3Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
Yes, I'd be happy but wonder if the executable needs to be self contained? ie. could the executable access other resources such as a database of signatures in 'the cloud'
40
20/10/2016 20:47:00AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!
Continuously. Teams can submit disaggregated estimates once per day and the published metrics will update immediately. The dataset will not change. There is a risk that teams will overfit to the competition data or manually disaggregate the data.
YesSee my ENEF paper http://makonin.com/doc/ENEF_2014.pdfMy algorithm can disaggregate American appliancesWater disaggregation, Natural gas disaggregation, Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regions
Africa, Americas, Asia, Europe, Oceania (Australia & Melanesia & Micronesia & Polynesia)
What about HeroX?
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) powerCurrent, Water consumption, Natural gas / heating oil consumptionA list of appliancesTo know whether each home has electric space heating / water heating / air conditioning.Real (active) powerOne monthNo "warm up" data is required.That's fineIndividual appliances
Fridge / freezer, Lighting, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Swimming pool (or spa) pump, Pond pump, Swimming pool (or spa) heater, Air conditioning unit, Electric space heater, Electric domestic water heater, Electric vehicle charging, Computer, Kettle, Toaster, Microwave, Television
1 HzHourlyYes. That set of temporal resolutions is fine.Every 10 seconds (0.1 Hz)Same as the aggregate inputYes, formally measure execution timeNo minimum
I will be training on additional data so it doesn't matter how much data is provided by the competition
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
Yes, I'd be happy to upload a trained executable program to the competiton.
41
20/10/2016 21:11:57AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesMy algorithm can disaggregate American appliancesAmericas
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
No, I will train on additional data (either a private dataset or additional public datasets)Yes, I'm happy to declare that I have trained on additional data
I will only submit once; training on whatever data I can get my hands on. Technically, I *could* train my algorithm on just the data provided by the competition. But it would be too much hassle to train twice.
Real (active) powerReal (active) power, Reactive power, Apparent powerA list of appliances
Real (active) power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One weekNo "warm up" data is required.That's fineIndividual appliancesFridge / freezer, Clothes dryer, Clothes washer, Dish washer, MicrowaveEvery 10 seconds (0.1 Hz)Yes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.No minimum1-3
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
Yes, I'd be happy to upload a trained executable program to the competiton.
42
20/10/2016 21:27:59Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.YesF1 score based on on/off of the applianceMy algorithm can disaggregate American appliances
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Americas, Europe
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Voltage, CurrentReal (active) powerOne quarterOne monthI'd prefer 6 months x 10 homes for both training and test setsIndividual appliancesThose appliances for which ground truth was provided> 1 kHz and ≤ 1 MHz> 1 kHz and ≤ 1 MHzhigh frequency data1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.No minimum>9Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.No, I definitely cannot upload an executable program. This would be a deal breaker.
43
21/10/2016 02:43:44AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regions
Americas, Asia, Europe, Oceania (Australia & Melanesia & Micronesia & Polynesia)
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I'd actually *prefer* the competition to not provide any training data, so you can double the size of the testing dataset.
Yes, I will only train on data provided by the competition.Real (active) power, Reactive powerInternal temperatureno needA list of appliances
Real (active) power, Reactive power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One weekOne weekThat's fine
Individual appliances, End-use categories (e.g. "cooling"), User behaviours (e.g. "cooking dinner on Tuesday night used X kWh")
Fridge / freezer, Lighting, Clothes dryer, Clothes washer, Dish washer, Cooking oven, Swimming pool (or spa) pump, Pond pump, Swimming pool (or spa) heater, Air conditioning unit, Electric space heater, Electric domestic water heater, Computer, Kettle, Toaster, Microwave, Television
> 1 kHz and ≤ 1 MHzMinutely, 15 minutely
Provide two or three sample rates. That way, we still get to explore the difference in disggregation performance between different sample rates but we also get to use more test data per sample rate.
1 HzSame as the aggregate inputDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.One day1-3
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
Yes, I'd be happy to upload a trained executable program to the competiton.
44
21/10/2016 09:18:07AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Disaggregation of data from commercial buildings (e.g. office blocks), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")Europe
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I could not compete.Yes, I will only train on data provided by the competition.Voltage, Current
Internal temperature, External temperature, Sunshine, Appliance state reported by smart appliances, Movement detectors (e.g. PIR, ultrasonic, door sensors), Proximity of users' mobile phones to the house / internal navigation
A list of appliancesNumber of occupants, Construction date (e.g. to infer the thermal performance of the building), Building type (e.g. detached, mid-terrace, bungalow etc.)
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
No minimum.That's fine> 1 MHz1 HzNo. Keep life simple (at least for the first competition). Just provide a single sample rate.1 HzSame as the aggregate inputYes, formally measure execution timeNo minimum>9Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.Yes, I'd be happy to upload a trained executable program to the competiton.
45
21/10/2016 12:34:21Company
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesMy algorithm can disaggregate American appliancesResidentialAmericas
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I'd actually *prefer* the competition to not provide any training data, so you can double the size of the testing dataset.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) power, Voltage, Current, Phase angle, Total harmonic distortionInternal temperature, External temperature, SunshineRough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)
A list of appliances, To know whether each home has electric space heating / water heating / air conditioning., Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), Number of occupants, Construction date (e.g. to infer the thermal performance of the building), Building type (e.g. detached, mid-terrace, bungalow etc.)
One monthThat's fineFridge / freezer, Clothes dryer, Clothes washer, Dish washer, Cooking oven, Swimming pool (or spa) pump, Swimming pool (or spa) heater, Air conditioning unit, Electric space heater, Computer, Microwave, Television> 1 kHz and ≤ 1 MHzHourly, MonthlyYes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputYes, formally measure execution time
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
Would need more details on uploading an EXE. Probably not.
46
22/10/2016 01:31:33HobbyistI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Once per year. The competition results would be revealed once per year (probably at a NILM workshop). The main disadvantage is that people will have to wait a long time before figuring out if their algorithm performs well.NoMean Squared ErrorMy algorithm can disaggregate American appliancesWater disaggregationAmericas
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I'd actually *prefer* the competition to not provide any training data, so you can double the size of the testing dataset.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onCurrent, Water consumptionA list of appliancesWater consumption of target appliancesHalf a yearNo "warm up" data is required.That's fineIndividual appliancesClothes washer, Dish washerMinutelyMinutelyYes. That set of temporal resolutions is fine.MinutelySame as the aggregate inputYes, formally measure execution time
Glad to see you're carrying through with this!
One day
I will be training on additional data so it doesn't matter how much data is provided by the competition
Total WATER consumed by appliances over some time periodYes, I'd be happy to upload a trained executable program to the competiton.
47
23/10/2016 00:58:18Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.NoI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Americas, Asia, Oceania (Australia & Melanesia & Micronesia & Polynesia)
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Real (active) power, Mains frequencyReactive power, Internal temperatureA list of appliancesNumber of occupants
Real (active) power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One weekOne weekThat's fineIndividual appliancesFridge / freezer, Clothes dryer, Clothes washer, Cooking oven, Air conditioning unit, Computer, Microwave1 Hz1 Hz, Every 10 seconds (0.1 Hz)Yes. That set of temporal resolutions is fine.1 HzSame as the aggregate inputYes, formally measure execution timeNo minimum1-3
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.
No, I definitely cannot upload an executable program. This would be a deal breaker.
48
25/10/2016 14:20:05Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesEnergy Usage Accuracy (Relative Difference) per appliance as supposed by Krish (EPRI) would be niceI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Americas, Europe
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I'd actually *prefer* the competition to not provide any training data, so you can double the size of the testing dataset.
Yes, I will only train on data provided by the competition.Real (active) power, Reactive powerApparent power, Voltage, Current, Power factor, Phase angle, Total harmonic distortion, Mains frequencyA list of appliances, Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)Real (active) powerNo minimum.No "warm up" data is required.That's fineIndividual appliances> 1 kHz and ≤ 1 MHzEvery 10 seconds (0.1 Hz)
Provide two or three sample rates. That way, we still get to explore the difference in disggregation performance between different sample rates but we also get to use more test data per sample rate.
1 HzSame as the aggregate inputNo, don't make any attempt to measure execution time.No minimum
I will be training on additional data so it doesn't matter how much data is provided by the competition
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
Depends on the terms.
49
26/10/2016 14:40:36CompanyI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesMy algorithm can disaggregate American appliances
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Asia
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I could not compete.No, I will train on additional data (either a private dataset or additional public datasets)Yes, I'm happy to declare that I have trained on additional dataI can't train my algorithm on the provided data.Real (active) power, Appliance state reported by smart appliances
Internal temperature, External temperature, Sunshine, Movement detectors (e.g. PIR, ultrasonic, door sensors), Proximity of users' mobile phones to the house / internal navigation
A list of appliancesRough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)
Real (active) power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
That's fineIndividual appliancesFridge / freezer, Lighting, Clothes dryer, Clothes washer, Dish washer, Air conditioning unit, Electric space heater, Electric domestic water heater, Microwave1 HzHourly
Provide two or three sample rates. That way, we still get to explore the difference in disggregation performance between different sample rates but we also get to use more test data per sample rate.
Every 10 seconds (0.1 Hz)Same as the aggregate inputNo, don't make any attempt to measure execution time.Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances.No, I definitely cannot upload an executable program. This would be a deal breaker.
50
27/10/2016 12:03:24AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year but with a separate test dataset for each case. Both datasets should be update annually. The results to the first is published immediately while the results to the second is published once a year.YesMy algorithm can disaggregate American appliances
Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Africa, Americas, Asia, Europe
Custom-build a web application to run the competition (Jack has funding until December to write this code). Open-source. The community would have total control over the competition. We can have multiple tracks (e.g. different temporal resolutions) and multiple metrics. One disadvantage is that we would not automatically get noticed by the majority of the ML community. But that could be solved by additional publicity activities.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Real (active) powerReactive power, Current, Internal temperature, External temperature, Appliance state reported by smart appliancesA list of appliancesTo know whether each home has electric space heating / water heating / air conditioning.
Real (active) power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One monthNo "warm up" data is required.That's fineIndividual appliances, End-use categories (e.g. "cooling")Fridge / freezer, Lighting, Clothes dryer, Clothes washer, Dish washer, Cooking hob, Cooking oven, Air conditioning unit, Kettle, Microwave> 1 kHz and ≤ 1 MHzMinutely
Provide two or three sample rates. That way, we still get to explore the difference in disggregation performance between different sample rates but we also get to use more test data per sample rate.
1 HzSame as the aggregate inputNo, don't make any attempt to measure execution time.4-6
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
Yes, I'd be happy to upload a trained executable program to the competiton.
51
04/11/2016 08:06:10CompanyI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!
Continuously. Teams can submit disaggregated estimates once per day and the published metrics will update immediately. The dataset will not change. There is a risk that teams will overfit to the competition data or manually disaggregate the data.
My algorithm can disaggregate American appliances
Water disaggregation, Natural gas disaggregation, Disaggregation of data from commercial buildings (e.g. office blocks), Disaggregation of data from industrial buildings (e.g. factories, power stations, etc.), Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")
Americas, Europe
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
Yes, I will only train on data provided by the competition.Real (active) power, Reactive power, Apparent power, Voltage, Current, Power factor, Phase angle, Total harmonic distortion, Appliance state reported by smart appliancesAll
A list of appliances, To know whether each home has electric space heating / water heating / air conditioning., Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation), Number of occupants, Construction date (e.g. to infer the thermal performance of the building), Building type (e.g. detached, mid-terrace, bungalow etc.)
AllReal (active) power, Apparent powerThat's fine
Individual appliances, End-use categories (e.g. "cooling"), User behaviours (e.g. "cooking dinner on Tuesday night used X kWh"), Energy consumed per occupant in a multi-occupant home (e.g. "John used X kWh; Eve used Y kWh")
> 1 MHzMinutelyNo. Keep life simple (at least for the first competition). Just provide a single sample rate.1 HzDailyNo, don't make any attempt to measure execution time.7-9
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
Yes, I'd be happy to upload a trained executable program to the competiton.
52
07/11/2016 08:48:18AcademicI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Europe
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onVoltage, CurrentMains frequencyA list of appliances, Rough geographical location (you might need this, for example, so you can get historical weather data to assist in disaggregation)To know whether each home has electric space heating / water heating / air conditioning.
Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset), Custom features
Individual appliances> 1 kHz and ≤ 1 MHz> 1 Hz and ≤ 1 kHz> 100 Hz and < 1MHz> 1kHzDon't formally measure execution time. Instead just ask teams to informally declare roughly how long their algorithms take to run.Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity.Can't decide now
53
08/11/2016 15:09:00Academic
I will register anonymously. I may reveal my real name later, e.g. if my algorithm does well, or if I beat my arch rival :)
Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.YesI'm not sure how well my algorithm will perform on American appliances but I'm happy to give it a shot!Purely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean"), Separate tracks for different geographical regionsEurope
Run two competitions: 1) an unsupervised NILM competition on Kaggle. One aim would be to attract new researchers to NILM. It would use a single dataset and a single metric and would be a very cleanly defined machine learning challenge. 2) A supervised NILM competition, run on a custom platform. The aim would be to assess the performance of existing NILM algorithms. Test across a few metrics, and would separate the teams into a few different tracks (e.g. teams who trained on external data versus those who did not; and high freq disag versus low freq; etc.)
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
I'm not sure (please select this option if there is any chance that you might train on additional data)
Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) power, Reactive power, Apparent power, Voltage, CurrentReal (active) power, Reactive power, Apparent power, Voltage, CurrentA list of appliancesA list of appliancesReal (active) powerOne yearThat's fineIndividual appliances1 Hz1 HzHourlyYes, formally measure execution time>9
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
No, I definitely cannot upload an executable program. This would be a deal breaker.
54
11/11/2016 17:14:56CompanyI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!
Continuously. Teams can submit disaggregated estimates once per day and the published metrics will update immediately. The dataset will not change. There is a risk that teams will overfit to the competition data or manually disaggregate the data.
YesMy algorithm can disaggregate American appliancesPurely unsupervised disaggregation (basically: trying to identify repeating patterns, without knowing what those patterns "mean")Americas, Asia
Pay Kaggle to run a supervised NILM competition. Assume winners would *not* have to hand over their intellectual property or code. ADVANTAGES: Kaggle has an enormous community of machine learning experts. Running a competition on Kaggle could help to encourage more machine learning researchers to work on NILM. Kaggle also has mature features such as a discussion forum. DISADVANTAGES: Kaggle competitions only allow for a single metric, a single dataset, and a single track.
I can compete if the competition doesn't provide training data but I'd prefer you to provide training data.
No, I will train on additional data (either a private dataset or additional public datasets)Yes, I'm happy to declare that I have trained on additional dataI'm happy to train my algorithm twice: once on *just* the data provided; and once on any data I can get my hands onReal (active) power, Reactive powerTotal harmonic distortion, Appliance state reported by smart appliancesA list of appliances, To know whether each home has electric space heating / water heating / air conditioning.
Real (active) power, Reactive power, Appliance event labels (e.g. labelled state transitions, like in the BLUED dataset)
One weekOne weekI'd prefer 6 months x 10 homes for both training and test setsIndividual appliancesFridge / freezer, Clothes dryer, Clothes washer, Dish washer, Swimming pool (or spa) pump, Swimming pool (or spa) heater, Air conditioning unit, Electric domestic water heater, Kettle, Toaster, Microwave1 Hz1 Hz, Every 10 seconds (0.1 Hz)Yes. That set of temporal resolutions is fine.1 Hz1024 sec (~17 min)Yes, formally measure execution time
Great job on forming NILM community!! :)
One day1-3Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances.No, I definitely cannot upload an executable program. This would be a deal breaker.
55
27/03/2017 15:08:16Non-profit organisationI'm happy to use my real name throughout the competition, even if my algorithm doesn't do very well!Both continuously and once per year. We could run the competition continuously and release "official" results at each NILM workshop.Yesno code jetDisaggregation of data from commercial buildings (e.g. office blocks), domestic
Africa, Americas, Asia, Europe, Oceania (Australia & Melanesia & Micronesia & Polynesia)
don't know jet
I'd actually *prefer* the competition to not provide any training data, so you can double the size of the testing dataset.
No, I will train on additional data (either a private dataset or additional public datasets)I cannot declare that I have trained on additional data, even if my team is anonymous.no code jetReal (active) powerReal (active) power, Total harmonic distortion, SunshineReal (active) powerOne dayOne weekThat's fine
I'm not sure if I can join the competition. I'm trying to get internal company funding the support disaggregation of household smart meter data. I would like to create/support an opensource cominity with code/funding in time. My current dream: "Have a opensource disaggration allogaritm in the domoticz.com software package before the summer. ".
One day1-3
Total energy consumed by appliances / behaviours / categories over some time period. e.g. a weekly summary of the energy consumed by individual appliances., Telling users which appliances are on at any given moment. e.g. telling users when they leave the house that their clothes iron is switched; or generating a log of appliance activity., List most or all of the appliances in each home
Yes, I'd be happy to upload a trained executable program to the competiton.
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
Loading...
 
 
 
Form responses 1