An interpretable and uncertainty-aware hybrid framework for short-term load forecasting in modern energy systems
ENERGY REPORTS, cilt.16, ss.109462-109478, 2026 (Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 16
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.egyr.2026.109462
- Dergi Adı: ENERGY REPORTS
- Derginin Tarandığı İndeksler: Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.109462-109478
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
Accurate short-term load forecasting (STLF) is critical for the reliable operation of modern power systems, the integration of renewable energy resources, and informed participation in electricity markets. However, the increasing variability of residential, prosumer, and electric vehicle (EV) demand profiles introduces significant challenges related to non-stationarity, predictive uncertainty, and the limited interpretability of existing forecasting models. To address these challenges, this study proposes an interpretable and uncertainty-aware hybrid forecasting framework that integrates signal decomposition, deep temporal learning, and ensemble modeling within a unified decision-support pipeline. Load signals are decomposed using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to extract multi-scale temporal components, while a hybrid long short-term memory (LSTM)–Transformer architecture captures both short-term dynamics and long-range temporal dependencies. In parallel, a stacking ensemble combining gradient boosting, multilayer perceptron (MLP), and Ridge regression is employed to enhance robustness and generalization across heterogeneous load profiles. Predictive uncertainty is explicitly quantified through Monte Carlo (MC) Dropout, and model transparency is ensured via SHAP-based feature attribution. The proposed framework is evaluated on four real-world datasets encompassing residential consumption, prosumer demand with photovoltaic generation, EV charging, and heterogeneous household load profiles. Experimental results demonstrate that the framework achieves high forecasting accuracy while simultaneously providing well-calibrated uncertainty estimates and interpretable insights. Notably, the stacking ensemble attains root mean squared error (RMSE) and mean absolute error (MAE) values as low as 0.0359 and 0.0209, respectively, confirming improved forecasting performance under diverse and non-stationary conditions. The probabilistic forecasts are well-calibrated, with prediction interval coverage probability (PICP) values exceeding 95%, and the model maintains consistent performance across multiple load profiles.