1 of 44

Data Driven Model for Medium Voltage Cable Fault in Rome

Yiming Jiang

Stanley Li

CUSP NewYork 07/12/2023

2 of 44

Top 10 Power Failures

2

Click here to add the title text content

3 of 44

The Major Power Outages in the U.S

延迟符

1542

83%

78%

3

Click here to add the title text content

4 of 44

❓WHY and HOW Should We Do

Maintain the reliability of the electrical grid in densely populated urban area.

Design and fit models on datasets

Urban

Challenge

DSO’S

DEMAND

Data

Modeling

延迟符

Data Analysis

Distribution System Operators need to increase the resilience of the grid and prevent outages

Aggregate data from different sources and conduct the data analysis

4

Click here to add the title text content

5 of 44

AGENDA

05

03

01

02

04

Introdution

Database Contruction

Data

Analysis

Data

Modeling

Policy Recommendation

& Strategy

延迟符

行业PPT模板http://www.1ppt.com/hangye/

6 of 44

7 of 44

areti S.p.a.

2.8 million

Supply a basin of about 2.8 million resident inhabitants

31000km

8.4kV & 22kV

添加

文本

延迟符

One of the leading companies in the energy sector, especifically in electricity distribution, in Italy.

Provides electrical services to the metropolitan city of Rome with over31,000 km of the grid

The medium level typically includes 8,4kV and 22kV.

7

Click here to add the title text content

8 of 44

Operation Mechanism

Medium Voltage Distribution System

Complex Network Infrastructure

Primary substations (which transform high voltage electricity into medium voltage electricity).

The secondary substations (from medium voltage to low voltage)

延迟符

8

Click here to add the title text content

9 of 44

Operation Mechanism

Medium Voltage Distribution System

Medium Voltage distribution line

Areti’s MV grid is constituted by about 1600 lines, and the lines

are constituted by 73 primary substations and 13000 serially connected secondary substations.

Each MV line is connected by 2 primary substations.

延迟符

9

Click here to add the title text content

10 of 44

Fault Protection Mechanism

Areti's MV grid, equipped with two-way current flow during fault situtation(Only single direction during normal operation)
Each substation houses two circuit breakers, a key setup for efficient fault handling.
These breakers isolate the affected branch, effectively containing the issue.
SCADA's technical sophistication enhances this robust system.

延迟符

10

Click here to add the title text content

11 of 44

Database Contruction

PART 2

Data Source

延迟符

Data Merging

12 of 44

Data Source

1

2

4

Objective

Geographic Information System (GIS): Provides constitutive and georeferenced information

SCADA System: Provides topological and electric load data.

Weather Data Providers: Provide measurements of weather parameters

延迟符

3

Fault Database: A historical repository of past grid faults.

Create A New Dataset

SOURCES

12

Click here to add the title text content

13 of 44

Data Merging

Weather Data

Line Loads Data

Branch Netwrok

Aata

Fault Data

13

Click here to add the title text content

14 of 44

Missing Data Analysis

Findings

"BRANCH CODE" and "MINIMUM SECTION", "% OVERHEAD", and "INSTALLATION TYPE" are closely linked, indicating similar patterns of missing data.

A similar degree of missing data found between "MATERIAL" and "% COPPER".

"AVERAGE TEMPERATURE" and "RAINFALL" form a cluster, suggesting a common cause for their missing values

14

Click here to add the title text content

15 of 44

Data Source

Objective: create a new database that includes both fault and normal data for each day by
merging the fault dataset with the power grid datasets.

Data Sources:

Geographic Information System (GIS): Provides constitutive and georeferenced information about our grid assets.

SCADA System: Provides real-time topological and electric load data.

Weather Data Providers: Provide measurements of weather parameters like temperature and precipitation.

Fault Database: A historical repository of past grid faults.

15

Click here to add the title text content

16 of 44

Data Source

Objective: create a new database that includes both fault and normal data for each day by
merging the fault dataset with the power grid datasets.

Data Sources:

Geographic Information System (GIS): Provides constitutive and georeferenced information about our grid assets.

SCADA System: Provides real-time topological and electric load data.

Weather Data Providers: Provide measurements of weather parameters like temperature and precipitation.

Fault Database: A historical repository of past grid faults.

16

Click here to add the title text content

17 of 44

Data Merging

17

Click here to add the title text content

18 of 44

18

Click here to add the title text content

19 of 44

Statistical Analysis——Length

添加标题 3

Skewness=3.36

Right Skewness

Majority of the grid lengths are relatively short

延迟符

19

Click here to add the title text content

20 of 44

Statistical Analysis——Material

Material distribution

44.6% Aluminum

Copper makes up 29.5%, 26% consist of a mixture of both materials.

延迟符

Length by Material

Mixed-materials branches are longer than aluminum and copper

it might be more prone to failure.

20

Click here to add the title text content

21 of 44

Failure Analysis

Seasonal

Line faults occur in the summer months, peaking in July.

it might have relation with rainfal and temerprture.

Amount

In July a cumulative total of 175 faults.

In 2022, a cumulative total of 574 faults

21

Click here to add the title text content

22 of 44

Failure Analysis

22

Click here to add the title text content

23 of 44

Failure Analysis

23

Click here to add the title text content

24 of 44

Data Modelling

PART 4

Regression

延迟符

Classification

25 of 44

Regression VS Classification

Predicting Faults Count with Regression Models:

Estimate how factors like line length, load, and materials might influence the likelihood of faults.
For instance, if a line is longer or carries more load, is it more likely to have a fault?

Categorizing Faults with Classification Models:

Classify branches as either likely to have a fault or not likely, based on the same factors.
This is helpful when we want a straightforward way to categorize the lines, like creating a priority list for inspections or maintenance.

25

Click here to add the title text content

26 of 44

Regression VS Classification

Predicting Faults Count with Regression Models

Categorizing Faults with Classification Models

For instance, if a line is longer or carries more load, is it more likely to have a fault ?

This is helpful when we want a straightforward way to categorize the lines, like creating a priority list for inspections or maintenance.

Monthly Fault Count

Fault or No Fault

VS

延迟符

26

Click here to add the title text content

27 of 44

Poisson Regression

Why Poisson Regression?
Traditional linear models fall short because they may predict negative count
Advantages

Predicts fault frequency based on various influencing factors.
Coefficients provide direct insight into the change in expected counts.
Handles both continuous and categorical variables effectively.

Limitations

Assumes equidispersion and independence of events, which might not always hold.

27

Click here to add the title text content

28 of 44

Poisson Regression

Backward Selection: 8 variables selected

Accuracy: R-squared value is 0.85

What Increases Faults:

The higher the Maximum Current Load, the more faults we see.
When Average Temperature rises, we tend to see more faults too.

What Decreases Faults:

More faults occur when the Minimum Current Load
The more branches a line has or the higher the voltage level, the fewer faults we see.

28

Click here to add the title text content

29 of 44

Negative Binomial Regression

Why Negative Binomial Regression?:

Better at handling data with high variability, giving more reliable predictions when Poisson Regression may not fit the data well, especially when data variability is high.

Strengths:

Works well with different types of data (numbers, categories, etc.).
Gives a more accurate picture of what causes faults in our power grid.

Things to Note:

It's a bit more complex than our standard model and might be harder to interpret.
Might not work as well when data variability is low.

29

Click here to add the title text content

30 of 44

Negative Binomial Regression

Model Performance: The model explains about 52.40% of the changes in FAULT COUNT.

Backward Selection: 3 variables selected

Impact of Factors:

AVERAGE TEMPERATURE: Has a significant influence on the FAULT COUNT.
MAX of Current Load: Shows a small effect on the FAULT COUNT.
MIN of Current Load: Does not noticeably affect the FAULT COUNT.

Goodness-of-Fit Measures: Deviance (2.6790) and Pearson chi2 (1.56) help us understand how well the model fits our data.

30

Click here to add the title text content

31 of 44

Regression Model Evaluation

The Poisson regression model, with a larger number of variables, gives a more accurate result, while the Negative Binomial model produces a surprisingly decent result with only three variables.
The trend of Current loads looks similar to the monthly fault count, indicating a strong correlation between them

31

Click here to add the title text content

32 of 44

Branch Classisfication

Imbalanced Data

Problem:Biased towards the majority class and perform poorly on the minority class.

Solution:Oversampling the minority class using the Synthetic Minority Over-sampling Technique (SMOTE).

32

Click here to add the title text content

33 of 44

Branch Classisfication

Logistic Regression

XGBoost

KNN

Decision Tree

Random Forest

LightGBM

MODELS

延迟符

33

Click here to add the title text content

34 of 44

Branch Classisfication

Fitting Models

Use grid search to optimize the hyperparameters of machine learning models

34

Click here to add the title text content

35 of 44

Branch Classisfication

Feature Importance

Objective: To determine the importance of features in our model
Methodology:

'Split' feature importance evaluates the frequency a feature is used to split the data in the model.
Higher values suggest a greater reliance

"PROGRESSIVE BRANCH" and "LENGTH": Emerged as important factors for classification models.

"% OVERHEAD" and "INSTALLATION TYPE": Showed lower importance, indicating their less frequent use in decision splits.

35

Click here to add the title text content

36 of 44

Branch Classisfication

Feature Importance

Objective: To determine the importance of features in our model
Methodology:

'Split' feature importance evaluates the frequency a feature is used to split the data in the model.
Higher values suggest a greater reliance

"PROGRESSIVE BRANCH" and LENGTH": Emerged as important factors.
"% OVERHEAD" and "INSTALLATION TYPE": Showed lower importance, indicating their less frequent use in decision splits.

36

Click here to add the title text content

37 of 44

Classisfication Model Evaluation

Why Confusion Matrix?

Helps us see when the model is correct and when it's making mistakes.

Key Elements:

True Positives (TP): The model correctly predicted the positive cases.
True Negatives (TN): The model correctly predicted the negative cases.
False Positives (FP): The model incorrectly predicted positive cases (these are actually negative).
False Negatives (FN): The model incorrectly predicted negative cases (these are actually positive).

37

Click here to add the title text content

38 of 44

Policy Recommendation

& Strategy

PART 05

延迟符

39 of 44

Policy Recommendation

Most pimiary substations have fewer than 13 faults in a year. Predict maintenance cost accordingly.

Predictive cost of maintenance

Effective management of factors like temperature and voltage levels can minimize faults.

Enhanced Weather Management

Abnormal current loads may indicate potential faults; Maintain line load within optimal limits

Current Load Monitoring

Fault incidents have been found to occur more frequently in the summer mont

Account for Seasonality

延迟符

39

Click here to add the title text content

40 of 44

Areti.s.p.a

1

Focus on Vulnerable Lines

2

Continue Corportation

Involve lobbying for financial support programs for electricity distribution companies, much like those that exist in the United Kingdom.

These programs can help offset the costs of necessary improvements to the network and contribute to the overall stability of the power system.

延迟符

Power lines with small (below 100mm) minimum diameters and carrying a voltage of 20kV are more prone to faults.

Implementing a more focused approach on these lines could include more frequent inspections and preemptive repairs.

40

Click here to add the title text content

41 of 44

Sustainability Plan

Make

Sustainabilty

Plan

延迟符

41

Click here to add the title text content

42 of 44

McKinsey’s Words

The average temperature on Earth has increased by 1.1°C since the end of the 19th century. Climate change is also causing a rise in extreme climate events: the probability of extreme temperatures in summer was as much as 15% higher in 2011–15 than in 1961–80

42

Click here to add the title text content

43 of 44

THANK YOU

44 of 44

Branch Classisfication

Fitting Models

We trid Logistic Regression', 'KNN', 'Decision Tree', 'Random Forest', 'XGBoost', 'LightGBM'
Use grid search to optimize the hyperparameters of machine learning models

44

Click here to add the title text content