Abstract: Short-term electricity load has the characteristics of high volatility and randomness disturbed by a variety of factors, which may affect the accuracy of load forecasting. In order to fully extract the features in the load data and improve the accuracy of short-term load forecasting, a long- and short-term temporal networks with attention (LSTNet-Attn) short-term load forecasting model based on modal decomposition is proposed. Firstly, by using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), the model processes the original load time series containing a large number of high-frequency components with complex frequency components, obtaining several intrinsic mode functions (IMF) containing different frequency components after frequency separation. Secondly, the date features are constructed on the basis of the collected features, and the redundancy problem of the input data dimensions is solved by the Boruta algorithm. Then, the LSTNet-Attn prediction model is constructed on the above basis, which includes a neural network module, an AR module and an attention mechanism module. The neural network module extracts the highly non-linear long- and short-term features and linear traits in the input load data, the AR module solves the insensitivity of the neural network for the linear feature recognition, and the attention mechanism enables more weights to be assigned to important features in order to capture the global and the local associations, optimizing the model to improve the prediction accuracy. Finally, the model is validated using the Umass Smart Dataset MIT dataset. Compared with the commonly used forecasting models and through the model ablation studies, the results demonstrate that the model effectively improves the accuracy of the load forecasting.