Abstract: In short-term load forecasting, deep learning models with recurrent units are widely used. However, the weight sharing structure used during training is time-invariant, ignoring different influence of input features（weather, date, historical load value,etc.） on load changes at different moments, ie., the weight sharing structure cannot track the fluctuations in the importance values of input features. To solve this problem, this paper proposes a forecasting method based on mutual information（MI） and bidirectional long short-term memory（BILSTM） network considering fluctuations in the importance values of features. The MI method is used to extract the importance value of input features at different moments, and constitute the fluctuant matrix of the importance values of input features, which is used as coefficients to correct the original input feature. Then, the corrected features are substituted into the BILSTM network to complete the training and forecasting, which makes up for the defect that the weight sharing structure cannot track the fluctuations in the importance values of input features and further improves the forecasting accuracy. Finally, taking the load data of the actual grid in a certain region as an example, the effectiveness of the proposed shortterm load forecasting method is verified.