Precise identification and classification for power quality disturbances is significantly important to analyze and comprehensively cope with power quality problems. Based on wavelet and improved neural tree techniques, a new classification methodology for power quality disturbances is proposed. In the method, the disturbance signal is decomposed into different frequency bands, whilst energy values and wavelet coefficient entropies of the base, harmonic and high frequency bands are calculated as eigenvalues respectively. The root mean produced in the disturbance process of the base wave band is calculated as a supplement, which is then combined with the energy values and wavelet coefficient entropies as eigenvectors for judging the disturbances. Thereafter the eigenvectors are normalized and input into the improved neural tree classifier, composed of neural network, decision trees and classification rules, for training and classifying. Simulation results demonstrate the method has a small amount of calculation to extract eigenvalues and the obtained eigenvectors can adequately reflect the difference information for different disturbance signals. The improved neural tree classifier combines respective superiorities of the neural network and decision tree in pattern classification, thus the classifier presents good convergence, global optimality and generalization, and can effectively identify seven common power quality disturbances with a simple structure and high accuracy.