Abstract: This paper proposes a method for fault feature selection of chillers. Firstly, the Fisher Score is used to eliminate a few features that are extremely insensitive to fault categories, and then the improved Lightning Search Algorithm is used to determine the weight of features and the number of features that should be selected. Thus, the final chiller feature subset is obtained. Experiments are carried out on ASHRAE Research Project 1 043 data, and a subset of chiller fault features containing 13 parameters is obtained, most of which are temperature parameters. Furthermore, four methods including k-Nearest neighbors (KNN), random forest (RF), BP neural network and gated recurrent unit (GRU) are used to obtain the diagnostic accuracy of each fault. Partial fault diagnosis accuracy is also improved compared with the original data, verifying the effectiveness of the selected feature subset.