Toward robust household load forecasting in smart grids using uncertainty-aware and cross-regional hybrid deep learning

Accurate and transferable household (HH) load forecasting is essential for the stable and real-time operation of renewable-energy-integrated smart grids; however, forecasting models trained on region-specific data often exhibit limited generalization across heterogeneous climatic and behavioral domains. To address this challenge, this study proposes a Cross-Regional Uncertainty-Aware Hybrid (CRUA-Hybrid) deep-learning (DL) framework that unifies dual dynamic feature gating, hybrid temporal representation learning, Transformer-based attention fusion, and adversarial domain adaptation for robust and transferable household load prediction. Specifically, a Dynamic Feature Gating (DFG) mechanism performs dual temporal and channel-wise gating to selectively emphasize salient information from input load sequences. A hybrid encoder combining Temporal Convolutional Network (TCN) and Gated Recurrent Unit (GRU) branches captures both short-term and long-term temporal dependencies, while a Transformer-based Cross-Model Attention Fusion (CMAF) module enhances the global contextualization of multi-branch representations. In addition, a Gradient Reversal Layer (GRL) enables adversarial domain adaptation between the Germany source domain and the Morocco target domain to improve cross-regional transferability. Experimental results show that, compared with the strongest baseline methods, the proposed model achieves performance improvements of 69.8% in RMSE, 15.1% in sMAPE, 26.6% in PMAE, 4.5% in Peak-MAE, 62.5% in CRPS, and 1.4% in PICP, indicating improved forecasting accuracy, peak prediction performance, and uncertainty calibration and strong adaptation to unseen HH and regions in renewable-energy-enabled smart grid environments.