Short-Term Power System Load Forecasting Based on TCN-BiGRU with CBAM

Short-term load forecasting plays a pivotal role in precisely evaluating the variations of regional power loads and facilitating scientific operational decision-making in modern power systems. To fully exploit the intrinsic temporal features embedded in load data and thereby improve the forecasting precision, a hybrid prediction model is proposed, which integrates a convolutional block attention module (CBAM), a temporal convolutional neural network (TCN), and a bidirectional gated recurrent unit (BiGRU). The K-medoids clustering method is used to cluster the dataset with date, weather, and other factors as inputs. The CBAM-TCN module captures the key features through the hybrid channel-space attention mechanism, and the BiGRU module explores the temporal dependencies of the load data through the bidirectional gating mechanism. Taking the publicly accessible ISO-NE power load dataset as a case study, the proposed model in this study is systematically analyzed and rigorously validated. Furthermore, the results demonstrate that the proposed load forecasting method exhibits remarkable advantages in forecasting accuracy compared with single neural network models and hybrid neural network models. Accordingly, this outcome provides a robust theoretical and practical foundation for power system planning and contributes to safeguarding the stable operation of power grids.