Electricity Consumption Forecasting using Support Vector Regression with the Mixture Maximum Correntropy Criterion.

Area of Science:

• Electrical Engineering
• Data Science
• Machine Learning

Background:

• Electricity consumption forecasting (ECF) is vital for electricity market operations.
• Traditional Support Vector Regression (SVR) struggles with nonlinear, non-Gaussian electricity consumption (EC) data and outliers due to its Mean Square Error (MSE) cost function.
• Existing methods lack robustness in handling statistical properties of errors in non-Gaussian scenarios.

Purpose of the Study:

• To develop a novel, robust forecasting method for electricity consumption.
• To address the limitations of traditional SVR in handling outliers and non-Gaussian data.
• To improve the accuracy and reliability of ECF.

Main Methods:

• Developed a robust Support Vector Regression (SVR) method by replacing the Mean Square Error (MSE) cost function with the mixture maximum correntropy criterion (MMCC).
• Investigated key factors influencing electricity consumption (EC) through data statistical analysis, identifying historical temperature and EC as primary inputs.
• Utilized MMCC-SVR for ECF, incorporating historical temperature and EC data.

Main Results:

• The proposed MMCC-SVR method demonstrated superior performance in electricity consumption forecasting compared to traditional SVR.
• The MMCC criterion effectively handled non-Gaussian data and outliers, leading to more accurate predictions.
• Empirical testing on real-world EC data from Guangzhou, China, validated the method's effectiveness.

Conclusions:

• The MMCC-SVR method offers a significant improvement in ECF accuracy, particularly for complex, real-world datasets.
• This robust approach enhances the reliability of electricity market operations through better forecasting.
• The study highlights the importance of appropriate cost functions for handling non-Gaussian and outlier-prone data in forecasting.