Abstract: In order to improve the accuracy of power load forecasting, a novel short-term load forecasting method based on multiple seasonal-trend decomposition using Loess and two-stage attention LSTM is proposed. The multiple seasonal-trend decomposition using Loess algorithm was used to decompose the original load series into trend component, daily periodic component, weekly periodic component and residual component. The feature attention mechanism and temporal attention mechanism were introduced to establish a two-stage attention LSTM load forecasting model. The potential information was extracted from the two dimensions of feature and time. Each component was predicted by this model and the prediction results were superimposed to obtain the final load forecasting value. The experimental results show that the proposed forecasting model has higher prediction accuracy and stability than other forecasting models.