Construction of Power Equipment Fault Prediction and Diagnosis Model Based on LightGBM and Particle Swarm Optimization

Abstract

 With the increasing intelligence of power systems and the complexity of equipment, traditional fault prediction based on experience or statistical methods is difficult to cope with multi-dimensional nonlinear data. Existing research has limited accuracy in high-dimensional features, complex working conditions and cross-device applications. This paper combined LightGBM with particle swarm optimization (PSO), used LightGBM (Light Gradient Boosting Machine) to model nonlinear features and handle sample imbalance, and used PSO to achieve automatic hyperparameter tuning to build a high-precision and highly adaptable power equipment fault prediction and diagnosis model. In the design of the intelligent fault identification system, this paper first processed the raw data such as voltage, current, temperature, vibration, etc., and extracted time series statistics, frequency domain features and change rate features to construct input variables. Then, the LightGBM model is used to establish a nonlinear mapping relationship between the equipment operation status and the fault mark, and the PSO algorithm is introduced to guide the particle iterative search for the optimal hyperparameter combination of LightGBM by minimizing the validation set loss function as the fitness function. Finally, this paper used the SMOTE (Synthetic Minority Over-sampling Technique) method to enhance fault class samples to alleviate the sample imbalance problem, and set category weights to improve recognition accuracy. Experiments show that the proposed method has a maximum fault identification accuracy of 0.84 and a recall rate of 0.95 (partial discharge). The measured detection delay on the edge computing platform is as low as 9.1 seconds (insulation degradation), and the fault classification F1 score is 0.94 (partial discharge), which are significantly better than traditional models. The area under the ROC curve is 0.87, achieving high-precision and low-latency power fault warning.