Enhancing Short-Term Load Forecasting with a Hybrid CNN-BiLSTM Model and Attention Mechanism

Electric load forecasting is a critical technology for optimizing energy distribution and grid management. Current works tend to suffer from the issue of insufficient capture of longterm dependencies when handling long time-series data, thus significant errors and overfitting in multi-step forecasting tasks still remains. This study proposes a hybrid model that integrates Convolutional Neural Networks (CNN), Bidirectional Long ShortTerm Memory (BiLSTM), and an attention mechanism to enhance the accuracy and robustness of multi-step load forecasting. Experimental results demonstrate that the proposed CNNBiLSTM+Attention model achieves root mean square errors (RMSE) of 0.1573, 0.1356, and 0.1269 for 1-day, 3-day, and 7-day prediction tasks, respectively, with correlation coefficients (CC) generally around 0.98, which significantly outperforms the baseline models. Additionally, the model exhibits faster convergence and stronger generalization capabilities during training, effectively mitigating overfitting.