Research on explainable BiGRU deep learning framework for short term load forecasting in smart power systems

This study proposes a novel forecasting framework based on a parallel multi-input Bi-GRU architecture combined with a sliding window-based Multi-Input Multi-Output (MIMO) prediction strategy. The key innovation lies in the feature-specific Bi-GRU encoding, where historical load, weather, and calendar data are independently processed before fusion, allowing the model to better capture temporal dependencies across diverse input types. Unlike prior models, our approach uniquely combines a structured feature-specific Bi-GRU encoding with interpretability using SHAP analysis, offering both enhanced predictive accuracy and transparent model decision-making for real-world deployment in power systems. The proposed Bi-GRU MIMO model serves as the core architecture. To benchmark its effectiveness, an attention-enhanced LSTM model is included strictly for comparative analysis. Experiments on real-world data from Singapore’s National Electricity Market System (NEMS) show that the Bi-GRU MIMO model improves prediction accuracy by 1.64% and reduces MAE by 2.19% compared to conventional models, highlighting its potential for enhancing smart grid decision-making.