Multi-scale feature fusion enhanced short term power load forecasting deep learning network

Customer load is affected by multi-scale factors. Aiming at the problem that insufficient multi-scale feature fusion will lead to low prediction accuracy, this paper proposes a deep network prediction method based on multi-scale feature fusion enhancement. Firstly, variational mode decomposition (VMD) algorithm is used to decompose load data into multiple sub-sequences with different features. Secondly, multiple time convolutional networks (TCN) are used to extract and learn features of these different sub-sequences and meteorological factors, and multiple features extracted by TCN network and date factors encoded by One-Hot are fused into a feature vector containing multi-scale information. The new feature vector of the Multi-Head Attention mechanism is used for feature enhancement, and the final predicted value is obtained by using the fully connected layer. Finally, the experimental results show that the root-mean-square error of the proposed method is 100.28 MW and the average absolute percentage error is 1.2%. Compared with the existing model, the proposed method has a higher prediction accuracy.