Electricity Load Forecasting with Bi-LSTM Model Based on Self-Attention Mechanism

A novel forecasting model of the bidirectional LSTM with self-attention (Bi-LSTM-SA) is introduced to address the need to upgrade the accurate projection of electrical load forecasting. The Bi-LSTM-SA model provides a reliable approach to capturing complex temporal correlations. The behavior of the model is analyzed using the peak interval-weighted root mean square error metric, which specifically assesses prediction accuracy during peak load periods. The trial results reveal that the Bi-LSTM-SA exceeds the others in terms of capacity. Bi-LSTM-SA model outperforms other models, including the unmodified LSTM, LSTM-SA, and the basic Bi-LSTM, as well as RNNs, TCNs, and GRUs, on various evaluation criteria. As a result, the superior predictability of the Bi-LSTM-SA model underscores its opportunity to significantly improve the reliability of electric load forecasting, thereby establishing its pivotal role in the field.