Dual Attention‐Empowered Bidirectional Long Short‐Term Memory Network for Short‐Term Electric Load Forecasting

Accurate load forecasting ensures sufficient energy generation, effective distribution, and grid stability. Short‐term load forecasting (STLF) helps address challenges such as unit commitment, spot pricing, load switching, and dispatching. With advancements in machine learning and data analytics, STLF systems can better adapt to changing electricity consumption patterns and provide more reliable predictions for grid operators and energy traders. In recent years, deep learning networks have significantly improved the multistep reliability of STLF due to their hidden layer features, which enable effective pattern recognition. However, conventional deep learning models often struggle to identify the most relevant data points when processing long sequences. To address this, a hybrid STLF method is proposed in this study, incorporating a dual‐attention‐based bidirectional long short‐term memory (DA‐BiLSTM) network for both commercial and residential datasets. The proposed method is compared with conventional LSTM, convolutional neural networks combined with LSTM (CNN‐LSTM), and quantile networks hybridized with LSTM (Q‐LSTM) to demonstrate its robustness. It improves the mean absolute percentage error (MAPE) on the commercial dataset by 30.1, 16.8, and 8.2% compared to LSTM, CNN‐LSTM, and Q‐LSTM, respectively. For the residential dataset, the improvements are 41.2, 52.3, and 13.8% over the same models.