Considering the misidentification of non-intrusive load identification method based on electrical features, when used for household loads which have similar electrical load features, this paper proposes features combination of household loads utilization which presented by load switching time, working period and number of switching, and original electrical features as new load identification features. On this basis, this paper presents extreme gradient boosting algorithm based on improved chaotic particle swarm. First, build extreme gradient boosting model with regression trees as base classifier. Furthermore, ease over fitting by adding regularization terms to object function and reduction coefficient. Meanwhile, chaos thought is applied to particle swarm algorithm to promote global parameter optimization, and get extreme gradient boosting algorithm with parameters improved by chaotic particle swarm. At last, we use AMPds public data set to test the algorithm this paper proposed, and the comparative analysis of testing results show the features and load identification algorithm have nice effect.