Influenced by meteorology, production, holiday factors and so on, user loads fluctuate randomly. General linear regression methods are difficult to describe the nonlinear trends of load changes. Machine learning algorithms can be used to fully explore the correlation between data before and after the sequence. In order to improve the predictive performance of the model, a user load prediction method combining micro meteorological monitoring and deep learning algorithms is proposed. Firstly, the mRMR method is used to select variables from the input meteorological and load historical data, reducing the input dimension, thereby reducing the complexity of the model and improving the efficiency of the prediction algorithm. Then, the dimensionality reduced variable historical data is input into LSTM to establish an mRMR-LSTM user load prediction model. The load forecasting experiment was conducted using historical data of users from six industries in the city. The results showed that, while ensuring the efficient operation of the model, the model has a good predictive effect on user loads in most industries.