Short-term wind power prediction is a popular issue in the research field. This paper proposes t location-scale distribution to describe the errors distribution of wind power prediction. Based on the measured data in wind power plants, autoregressive integrated moving average model and back propagation neural network are adopted to analyze the errors of two forecast models respectively, proving that the t location-scale distribution can describe effectively the frequency distribution of forecast errors, and the specific research data show that the goodness of fit of t location-scale distribution is better than that of normal distribution. Parameters of t location-scale distribution, as the indicators for judging the accuracy degree of the prediction algorithm, make it available to analyze its performance directly.