In order to improve the capability of entropy method in time-frequency feature presentation of fault signals of transmission lines, a new method is proposed for power system fault diagnosis based on Tsallis wavelet singular entropy (TWSE) with hierarchical variable step size. Firstly, the collected voltage signals are transformed by wavelet decomposition and single branch reconstruction, which is used to construct time-frequency matrix. Secondly, the singular value decomposition theory combines with the Tsallis entropy theory, and the time-frequency matrix processed by TWSE with 1 sliding step size is used to determine the fault occurrence time. Then, the method calculates TWSE with 1/4 period sliding step size to obtain the feature vector for fault diagnosis from the one period after the fault happened of the three-phase voltage reconstruction coefficient. Finally, the TWSE feature vector is input to the classifier based on the extreme learning machine (ELM) to realize fault diagnosis of transmission line. Simulation results show that the new method has better feature representation ability for fault transient signal, and classification result is not affected by fault time, transition resistance and fault location. Compared with the SWSE fault diagnosis method based on wavelet singular entropy, the method of TWSE has better diagnosis effect. This work is supported by National Natural Science Foundation of China (No. 51307020).