Abstract: This paper proposed an adaptive S-transform (AST) which directly controlled the standard deviation of window function to manage resolution. The Gaussian window was directly matched with the main value interval of the signal spectrum to optimize the time-frequency resolution. The process was independent of the S transform (ST) feedback result, improving the optimization speed. Based on the time-frequency matrix of AST, four features were extracted to characterize 20 typical types of power quality disturbances (PQDs). The feature vector had low dimension and strong recognition ability. The feature vectors extracted by AST and ST were identified by Extreme learning machine (ELM) and Probabilistic neural network (PNN) with different noises. The results show that AST can extract more accurate features than ST with better time-frequency resolution and noise resistance.