Abstract: In recent years, the imbalance between the power supply and demand during the peak load periods has become increasingly prominent, increasing the costs of the power grid operation. A daily peak load forecasting method based on Bi-LSTM (bi-directional long short-term memory) and the feature correlation analysis is proposed to improve the prediction performance. Some descriptive and curve indicators are used to analyze the daily peak load characteristics of the users in different industries and the Copula function is applied to analyze the connection between the multi-dimensional time series data quantitatively. The bi-directional long and short-term memory network model is constructed to predict the magnitude of the daily peak load in the future. Due to the common existence of the multi-peak feature in electricity regularity, the continuous historical occurrence moments are converted into several discrete time periods and then the load occurrence forecasting is carried out based on the result of magnitude prediction and the XGBoost algorithm. Experimental results on the real dataset show that the proposed model has high prediction accuracy for the high-value load and contributes to improve the capabilities of the "peak shaving and valley filling" and other power supply services for the power grids.