The power system must precisely predict future demand for power supply scheduling based on the forecast. Electricity forecasting has served as the foundation for the power sector’s adjustment of tariffs, substation regulation, total load forecasting, and peak avoidance management. In this research, we propose a multi-head dual-stream attention mechanism CNN-LSTM model for power forecasting based on external auxiliary information. While LSTM can discover correlations between time series, CNN is capable of extracting features from multivariate variables. Higher-order features of external auxiliary data and sequence data are acquired by CNN and fed into the multi-head dual-stream attention layer, along with the temporal aspects that are captured. Provide feed to the LSTM layer for capturing the temporal relationships of the sequence data, set weights among various features to highlight the elements that are significant for prediction, and ultimately output the future period of power. The CNN-LSTM model of external auxiliary features for multi-head dual stream attention mechanism is constructed after parameter adjustment. According to experimental results, the model performs better than deep learning prediction algorithms like CNN and LSTM in terms of average absolute error and root mean square error metrics. It also outperforms the state-of-the-art model in terms of error reduction under MAE evaluation metrics, with a 7% improvement in accuracy and stability. Overall the method extracts spatio-temporal features and builds an end-to-end model with significantly better prediction results than traditional methods.