Partial discharge (PD) type recognition for power transformer involves small sample problem because of lack of effective training instances. In this paper, a novel ensemble probabilistic neural network (PNN) based on feature subsets (a.k.a. FS-EPNN) was proposed for classifying PD types. Firstly, forty-four statistical parameters were extracted from two-dimension diagram histograms based on the phase resolved partial discharge pattern. Secondly, in order to reduce the feature dimensionality of samples, we split the entire feature set and grouped features into a certain number of feature subsets instead of traditional approach such as PCA, which may cause information loss. Then, an equal number of PNN classifiers was built according to training instances under each feature subset. Finally, the type of PD was determined by simple majority voting. Experimental results show that, the proposed FS-EPNN outperforms BPNN, PNN based on PCA and single PNN in terms of recognition accuracy for PD.