A short-term load forecasting method based on restructured second modal decomposition and LSTNet-Atten was proposed in order to address the issues of high randomness and volatility in power load data, as well as the relatively low prediction accuracy. Complete ensemble empirical mode decomposition with adaptive white noise was employed for the preliminary decomposition of the load sequence during the data preprocessing phase. Then the randomness and volatility of the original signal was reduced. Similar subsequences were reorganized and aggregated based on their sample entropy values. Variational modal decomposition was applied to further decompose the components with higher complexity obtained from the reorganization in the feature engineering phase. The correlation between input influencing factors and load data was evaluated by using Pearson, Spearman, and maximum information coefficient methods, while evidence theory was utilized to optimize the feature dimensions of the input data. The LSTNet-Atten forecasting model was reconstructed in the model construction phase. Convolutional modules were used to mine local dependencies within sequences, while recurrent and skip recurrent modules extract both long-term and short-term features from the data to enhance its predictability. The autoregressive module was used to enhance the ability of the neural network to recognize linear features and improve the predictive performance of the model. More weight was given to important features by increasing temporal attention in order to achieve the capture of global and local connections. Experimental results on a Valencia regional load dataset indicate that the proposed method reduces sequence prediction error by up to 66.69% compared with other classical deep learning models with a 5.04% increase in fitting coefficient, demonstrating higher prediction accuracy and robustness.