Fault prediction for distribution network based on CNN and LightGBM algorithm

Abstract

Fault prediction plays a significant role in enhancing the safety, reliability, and stability of distribution network. However, the problem of enormous time-series data and discrete data makes the prediction great challenge. The imbalance between normal and fault samples will reduce the accuracy of the model. In addition, the influence of time-series variables on distribution network is direct and continuous, so the time-series feature extraction is the key technique for fault prediction. In this work, we propose a fault prediction method for distribution network based on CNN and LightGBM algorithm. This method deeply learns feature of time-series data by utilizing CNN, and improves the adaptability for imbalanced dataset by training LightGBM submodels. Experimental results based on fault dataset of a district in Chongqing from 2017 to 2018 show that fault prediction performance can be ameliorated by utilizing this method.