Short-Term Load Forecasting Method Based on Improved Temporal Fusion Transformer

To address the challenges of strong multiscale spatiotemporal feature coupling and complex long-term dependencies in short-term power load data, this paper proposes a hybrid short-term load forecasting (STLF) method based on an improved Temporal Fusion Transformer (TFT) model. First, a parallel dilated convolutional network (DCNN) is constructed as a feature extraction module, leveraging convolutional kernels with different dilation rates to capture local periodic patterns and multiscale spatiotemporal correlations in load sequences. Second, a Bidirectional Gated Recurrent Unit (BiGRU) is used to replace the Long Short-Term Memory (LSTM) structure in the traditional TFT, reducing model complexity by simplifying the gating mechanism while enhancing the ability to capture dynamic temporal features. Finally, an improved TFT architecture employs a hierarchical attention mechanism for explicit modeling of multi-source covariates. Experimental results demonstrate that the proposed model achieves higher accuracy than benchmark models on a public dataset, both on weekdays and non-working days.