XI INTERNATIONAL CONFERENCE

“INFORMATION TECHNOLOGY AND IMPLEMENTATION” (IT&I-2024)

The Machine Learning Model Development Lifecycle for Prediction of Electrical Energy Market Volumes

Anatoliy Doroshenko�Dmitry Zhora�Oleksii Zhyrenkov

Institute of Software Systems, Kyiv, Ukraine

Introduction

Modern Python-based libraries for machine learning and data analysis:

• scikit-learn
• TensorFlow
• Keras
• PyTorch

Modern MLOps platforms for model deployment and version tracking:

• BentoML
• MLflow
• Kubeflow
• SageMaker
• TFX

The usage of machine learning techniques de-facto became a standard for modern systems that need to provide a prediction. The regression algorithms available in scikit-learn library are used in current research to build the forecasting model for electrical energy markets.

Input Dataset

On July 1st, 2019, Ukraine adopted the European model that assumes the following four markets for electrical energy: bilateral, day-ahead, intraday, and balancing. The bilateral market can be also referenced as a future or forward market. In Ukraine, the total amount of deals is recorded every hour. The markers are organized in a way to balance the energy price and volatility. The dataset used for this research matches the time range from July 1st, 2020, to December 31st, 2021.

Outside Temperature

Dataset Augmentation

It is common for real-world processes that the dynamics of monitored parameters is affected by other factors that are not available in the original dataset. In particular, the electricity production is influenced by outside temperature and by consumer activity cycles.

One of the challenges is to provide the representation of time in a way that close moments in time would be interpreted as close by machine learning algorithm. The solution that is convenient from computational perspective is to calculate the sine and cosine functions when the corresponding argument represents the phase of the cycle. These periodic data series were calculated with the help of an algorithm written in Python.

Dataset Resampling

As original data record contains the figures representing the current hour it makes sense to add two types of columns in the dataset: the parameters that represent the past history and parameters that represent the future to be forecasted. The sliding time window of 24 hours was considered in both directions. Overall, 188 parameters were added for every record.

The obtained dataset had 13'129 records. For hyperparameter tuning it was split into training and testing parts with proportion of 80% to 20%. The random split functionality is provided by method from scikit-learn library. The obtained datasets were saved into .csv files, so that different machine learning algorithms considered later are evaluated with equal conditions.

Model Metrics

Three metrics shown in table below were used to compare the input parameter sets and different regression algorithms. Each metric measures the discrepancy between test set and forecasted data for selected output column representing one of electrical energy trading volumes. The R2 score was used to make a decision. The nearest neighbors regression algorithm was used to evaluate the performance of input parameters, it provides quite competitive results and has limited number of hyperparameters to tune.

Feature Selection

The scikit-learn library provides automation facilities to select most informative input parameters. The following classes are worth mentioning: GridSearchCV, LassoCV and SelectFromModel. The latter option was used in current research. The alternative would be to manually evaluate up to 2ⁿ combinations, here n represents the total number of possible inputs. The periodic parameters were not added to the history as such parameters precisely indicate the moment in time. The final set of input features obtained with SelectFromModel class contained 60 entries out of 106.

selected_columns = \

    ['Bilateral', 'DayAhead', 'Intraday', 'Balancing', 'BilateralM1',

    'DayAheadM1', 'IntradayM1', 'BalancingM1', 'DayAheadM2',

    'IntradayM2', 'BalancingM2', 'IntradayM6', 'BilateralM7',

    'DayAheadM7', 'IntradayM7', 'BalancingM7', 'BilateralM8',

    'DayAheadM8', 'IntradayM8', 'BilateralM9', 'DayAheadM9',

    'IntradayM9', 'BilateralM10', 'DayAheadM10', 'IntradayM10',

    'BilateralM13', 'BilateralM14', 'IntradayM14', 'BalancingM14',

    'BilateralM15', 'DayAheadM15', 'IntradayM15', 'BalancingM15',

    'BilateralM16', 'DayAheadM16', 'IntradayM16', 'BilateralM17',

    'DayAheadM17', 'IntradayM17', 'BilateralM18', 'DayAheadM18',

    'IntradayM18', 'IntradayM19', 'BilateralM21', 'BilateralM22',

    'DayAheadM22', 'IntradayM22', 'BilateralM23', 'DayAheadM23',

    'IntradayM23', 'BalancingM23', 'TempUkr', 'TempKiev', 'CosDay',

    'SinWeek', 'CosWeek', 'SinMonth', 'CosMonth', 'SinYear', 'CosYear']

Prediction Error (MWh)

Error Histogram

R2 Score Comparison

MAPE Comparison

MAE Comparison (MWh)

MLOps Constituents

The paradigm of MLOps includes aspects like best practices, sets of concepts, as well as a development culture when it comes to end-to-end conceptualization, implementation, monitoring, deployment, and scalability of machine learning products.

MLOps is aimed at productionizing machine learning systems by bridging the gap between development (Dev) and operations (Ops).

Principles: CI/CD automation, workflow orchestration, reproducibility; versioning of data, model and code; collaboration; continuous ML training and evaluation; ML metadata tracking and logging; continuous monitoring; and feedback loops.

Existing Instruments

BentoML and FastAPI

1. Register model in model registry
2. Create Service and workers
3. Add additional logic layer
4. Package model as Bento
5. Deploy as docker image

Model Deployment

Conclusion

The production-ready forecasting solution requires multiple stages:

• data preprocessing and augmentation
• selection of input parameters
• selection of machine learning algorithm
• hyperparameter optimization
• model training and serialization
• creation of REST API layer
• creation of Docker image
• networking and autoscaling configuration
• deployment into Kubernetes cluster

The ONNX standard for model serialization allows for interoperability between different languages and platforms. The usage of popular machine learning libraries like scikit-learn greatly simplifies the research phase. The deployment platforms like BentoML automatically provide features for service resilience and scalability.