This study aims to develop a robust and interpretable forecasting framework for short-term electricity demand that supports sustainable energy demand management, particularly under conditions of demand volatility and unexpected events.
This study proposes a hybrid Prophet-XGBoost forecasting framework that combines a decomposable time-series model with machine learning–based residual learning. The model was validated using 425 days of operational data from the Taiwan Power Company, utilizing a 30-day out-of-sample test set to simulate real-world forecasting.
The hybrid model consistently improves upon the standalone Prophet baseline and remains highly competitive with standalone XGBoost. Anomaly detection identifies demand irregularities linked to exogenous shocks, such as national holidays, improving robustness and interpretability.
The framework offers energy utilities and system operators a practical decision-support tool for improving short-term load forecasting, anticipating demand shocks, and strengthening grid reliability.
This study contributes a decision-oriented hybrid forecasting framework that combines statistical interpretability with machine learning accuracy and explicit anomaly detection for energy sector management.
