Abstract: Aiming at the problem of insufficient prediction accuracy caused by factors such as strong volatility of power load data, multi-feature coupling and noise interference, a short-term power load forecasting method based on secondary decomposition reconstruction and improved NGO optimized TCN-BiLSTM-Attention combination is proposed. First, the Spearman coefficient analysis method is used to select influencing features with strong load correlation. Then, in order to reduce the randomness and volatility of the load sequence, the load sequence is decomposed into a set of modal components using fully adaptive noise ensemble empirical mode decomposition; secondly, based on sample entropy calculation, each mode is aggregated into three frequency bands of high, medium and low, and then the high-frequency component is secondary decomposed using variational mode decomposition to further reduce the complexity of the load sequence. Finally, the reconstructed modal components and the screened strongly correlated features are input into the proposed combined model for prediction, and the final prediction value is obtained by weighted fusion of the prediction results of each component. An example analysis is carried out using power load data from a certain place in Australia. The experimental results show that compared with other comparative models, the RMSE, MAE and MAPE of the proposed method are reduced by 57.9%, 54.6% and 55.6% respectively, which proves that the model can effectively improve the accuracy of power load forecasting.