As a common safety hazard in low-voltage electricity use, fault arcs are difficult to perceive effectively on the eve of failure due to their concealment and randomness. Existing protection methods usually take measures after the occurrence of faults, which can easily cause electrical fires. To address these issues, a warning method for low-voltage electrical safety hazard is proposed on basis of multi-dimensional features and random forests. Next, a random forest model is built and hyper-parameters are optimized, with the goal of minimizing node information entropy to complete model training, so that enhances the overall performance and learning efficiency of the model. Finally, experimental verification shows that the proposed method achieves a prediction accuracy of over 99.4% with different loads, and its prediction accuracy is higher than that of four traditional classification prediction models.