With the continuous growth of electricity demand, the problem of hidden dangers in customers’ electricity use has become increasingly prominent. Traditional electricity safety monitoring methods often rely on a single data source and empirical rules, resulting in low prediction accuracy and slow response speed. This paper uses data cleaning and preprocessing to standardize the original data, eliminate noise data, and fill in missing values. Then, feature engineering is used to extract key feature indicators. Based on these features, this paper uses random forest and support vector machine (SVM) algorithms for classification modeling to determine whether there are abnormal or hidden danger risks in customers’ electricity use. In addition, a long short-term memory (LSTM) neural network is used to predict time series data and identify potential long-term risk trends. In terms of multi-source data fusion, this paper introduces an ensemble learning method to fuse the prediction results of multiple models through a weighted average strategy to improve accuracy and robustness. Finally, a threshold algorithm and decision tree are used for risk assessment and warning generation to provide early warnings for potential safety hazards and provide users with corresponding safety measures. The experimental results show that the random forest model has the best overall performance, with an accuracy of 91.2%, a recall rate of 87.6%, an F1 score of 89.3%, and an AUC value of up to 0.943, which is superior to other models in terms of overall accuracy and stability.