Adaptive electricity consumption forecasting approach for universal environments.

Area of Science:

• Power Systems Engineering
• Data Science
• Machine Learning

Background:

• Accurate electricity consumption forecasting is vital for power system stability and intelligent management.
• Traditional forecasting models often fail to account for user heterogeneity and concept drift.
• Concept drift, caused by changing data distributions, significantly impacts forecast accuracy over time.

Purpose of the Study:

• To develop an adaptive electricity consumption probability forecasting method for universal environments.
• To address limitations of traditional methods in handling user heterogeneity and concept drift.
• To improve the accuracy and adaptability of electricity consumption forecasts.

Main Methods:

• Implemented a nonmonotonic correlation elimination-based recursive feature selection for adaptive feature determination.
• Utilized a joint loss function combining point and probability forecasting for precise online error quantification.
• Incorporated a buffer for storing pattern-changing batch data and dynamically adjusted weights to mitigate concept drift.

Main Results:

• The proposed Adaptive Electricity Consumption Forecast for Universal Environments (AECF-UC) method demonstrated superior performance against mainstream approaches.
• AECF-UC achieved average RMSE, pinball loss, and CRPS of 0.3041, 0.0567, and 0.1683, respectively.
• The joint loss function improved prediction accuracy by approximately 6% compared to single-loss methods.

Conclusions:

• The AECF-UC method exhibits significant advantages in universality and adaptability for electricity consumption forecasting.
• The proposed approach effectively handles user heterogeneity and concept drift in dynamic environments.
• This adaptive forecasting model contributes to more stable operation and intelligent management of power systems.