Abstract: Feature selection is important to improve the model accuracy and reduce overfitting. The 10 kV power distribution network is complex and there are too many features for a data mining model to work. Before modeling power fault data, the dimensionality reduction and model selection is necessary. In order to solve this problem, a Bayesian information criterions (BIC) model selection algorithm along with backward selection algorithm was proposed. BIC aims to reduce the complexity of model and the backward selection can reach fast convergence. Experiments show that the algorithm works well. It is proven that the algorithm proposed here is of advantage to improve model accuracy and data training efficiency.

Key words: distribution network fault data, dimensionality reduction, Bayesian information criterions, backward selection algorithm