Data-driven residential electricity load forecasting method for virtual power plants

Abstract

Virtual power plants (VPP) integrate distributed energy resources and residential loads to enhance the flexibility and reliability of power systems. Accurate residential electricity load forecasting is therefore essential for VPP scheduling and demand-side management. However, residential load series exhibit strong non-stationarity, periodic fluctuations, and complex temporal dependencies, which limit the forecasting performance of conventional models. To address these challenges, this study proposes a hybrid forecasting framework that integrates seasonal-trend decomposition using loess (STL), a bidirectional gated recurrent unit (Bi-GRU), and an attention mechanism. The original load sequence is decomposed into trend, periodic, and residual components to separate long-term evolution, regular fluctuations, and stochastic disturbances. These components are then processed by the Bi-GRU to capture bidirectional temporal dependencies, while the attention mechanism dynamically emphasizes informative time steps associated with load peaks and abrupt variations. The proposed method achieved MAE of 7.2, RMSE of 11.85, and R² of 0.88. Compared with the best-performing baseline, Bi-GRU, the proposed framework reduced MAE and RMSE by 16.28% and 15.96%, respectively, while improving R² by 6.02%. The resulting trajectory can serve as a load-side input for day-ahead VPP planning; its economic and operational effects require further validation in closed-loop scheduling.