Practical Applications of Machine Learning in Manufacturing: Virtual Power Plants and Coordinated Computing and Power Management

Abstract

This paper uses publicly available UCI electricity consumption data to examine how machine learning affects day-ahead load forecasting and energy storage dispatch. Hourly loads from 20 large electricity users are aggregated to simulate a virtual power plant. The experiments compare naive baselines with direct regression models and use the weekly load pattern to develop residual models that predict deviations from the load at the same hour of the previous week. The gradient boosting residual model achieves a mean absolute percentage error of 3.49%, a 13.2% reduction from the baseline of 4.02%, and reduces root mean squared error by 28.0%. The error in the period of peak falls from 1.69% to 1.22%, and the mean absolute error in daily peak load falls by 25.4%. By using a 15 MW/30 MWh storage system, assumed the rates are based on the time it uses, and a fixed peak-shaving target, peak-shaving not only effectively increases from 78.2% to 79.0% over the roughly 38-day test period but also the saves more electricity cost rise from CNY 511,000 to CNY 515,000. Weekly residual learning improves the accuracy forecasting with a larger error reductions during peak periods than the full test period. Better forecasts bring additional peak-shaving and cost benefits although the storage capacity and dispatch rules limit gains. This paper also shows that how the approach can potentially support coordinated computing and power management through shifts in task execution times; However, the performance in computing task scheduling might requires more validation.