A hybrid deep learning–ensemble model for accurate and explainable electricity demand forecasting in renewable energy communities

Accurate, interpretable, and scalable short-term electricity demand forecasting is a crucial requirement for transitioning towards decentralized, citizen-driven energy systems, particularly within renewable energy communities (RECs). The proposed framework utilizes long short-term memory (LSTM) networks and extreme gradient boosting (XGBoost) to capture both the sequential dynamics of consumption and the nonlinear interactions between household and community-level characteristics. A stratified household selection strategy based on contribution-level binning is employed to ensure coverage across distinct demand regimes while maintaining computational efficiency. Based on 15-min smart meter data from a renewable energy community in Loureiro, Portugal, the hybrid model achieved the lowest prediction errors at both scales, with a best-case household-level mean absolute percentage error (MAPE) of 1.25% and a community-level average MAPE of 3.14%, corresponding to explained variance ( R 2) values of 0.9420 and 0.9558, respectively. Although XGBoost performed slightly better for specific low-variability households, the hybrid approach consistently outperformed the standalone models in terms of accuracy and robustness. The model’s interpretation is enhanced by SHAP, which reveals distinct demand drivers for consumers, including memory-based characteristics and climatic variables like temperature and solar radiation. The Interpretive, multi-level prediction framework provides a reliable basis for practical REC applications, such as battery planning, peer-to-peer trading, and real-time demand response.