Short‐Term Residential Load Forecasting With <scp>FG</scp> ‐ <scp>LSTM</scp> ‐ <scp>TA</scp> : A Hybrid Deep Learning Approach Integrating Feature Gating and Temporal Attention for Enhanced Prediction Accuracy

ABSTRACT This paper introduces FG‐LSTM‐TA, a hybrid deep learning model that integrates feature‐gated LSTM networks with a temporal attention mechanism for high‐accuracy short‐term residential load forecasting. The architecture combines selective feature gating, sequential temporal modeling, and attention‐based context weighting to enhance learning efficiency and adaptivity. The model was evaluated using real‐world 30‐min interval electricity consumption data from Ahvaz, Iran, across four scenarios involving both individual and aggregated households. Experimental results demonstrate that FG‐LSTM‐TA consistently outperforms baseline models (LSTM and CNN) and advanced benchmarks (GRU, Transformer, and CNN‐LSTM), achieving R 2 scores up to 98%, lower training losses, and shorter computation times. Compared to other models, it maintains robust performance even in high‐noise or low‐amplitude environments. A compact sensitivity analysis further confirms the model's responsiveness to dominant temporal and weather‐related features. These results establish FG‐LSTM‐TA as a scalable, efficient, and interpretable forecasting solution for smart grid applications such as demand‐side control and energy optimization.