Research on the Identification Method of Series Arc Fault Based on the Feature Sensitivity Analysis and the TVA Coefficient Optimized Random Forest

This paper aims to tackle the difficulties in identifying series arc faults with feature aliasing in low-voltage electricity scenarios. Supported by feature sensitivity analysis, a method for identifying low-voltage series arc faults using a time-variance-accuracy coefficient optimized random forest (TVA-RF) model is proposed. Firstly, in the quantitative analysis of real sample feature sensitivity, it is discovered that the main reason for the above issue is that the sensitivity of the sample features decreases in certain scenarios, thereby interfering with the judgment of traditional arc detection methods. Secondly, the random forest model, with its anti-interference capabilities derived from its ensemble operation mechanism, is utilized as the core model of proposed method. Thirdly, this paper constructs a TVA coefficient to achieve multi-objective optimization of random forest hyper-parameters and institutes a program module that uses low-voltage load state samples for training the random forest model. Additionally, a program for denoising and extracting time-frequency features is embedded into the input of the trained random forest model. Finally, the TVA-RF method hit 99.978% accuracy in laboratory tests, outperforming traditional methods in accuracy, training efficiency, and calculation speed. In real low-voltage systems, it achieved 98.096% accuracy. Overcoming feature overlap interference, the TVA-RF method accurately identified low-voltage series arc faults.