Adaptive Residual-Transformer Ensemble for Day-Ahead Net-Load Forecasting in PV-Integrated Microgrids
Accurate short-term net-load forecasting is critical for the reliable operation of power systems integrating renewable energy sources such as photovoltaic (PV) systems. The inherent variability of PV generation and human-driven demand patterns complicates energy scheduling and storage management. To address these challenges, this study proposes a hybrid residual-transformer ensemble framework that integrates adaptive blending and behavioral baseline learning to improve the accuracy and robustness of forecasts. For gross load forecasting, the framework separates training into nighttime and daytime phases: a single nighttime model is trained on all days, while daytime models are weekday-specific and operate on residuals from monthly–weekday baselines representing habitual load behavior. Instead of directly predicting total load, the model predicts deviations from a holiday-aware baseline that can incorporate recent anomalies computed over the preceding week. These residuals are learned using PatchTST transformer networks and merged with the baseline through an adaptive weighting rule that increases the weight on the baseline during anomalous periods. PV generation is modeled independently using dual Sequence-to-Sequence (Seq2Seq) Transformer networks trained under sunny and non-sunny regimes, incorporating meteorological inputs comprising solar irradiance and temperature, regime blending, and daylight masking to maintain physical consistency. The final net-load forecast is obtained by subtracting the PV predictions from the blended gross load estimates. Evaluations using one year of 15-minute data from three real-world sites spanning distinct climates (a University of Hawaii building, an Australian residential prosumer, and a Madeira Island prosumer) demonstrate improved accuracy and robustness relative to the baseline, supporting the practical feasibility of the framework for renewable-rich distribution systems.