Mixture Density Network Based on Truncated Distribution and Genetic Algorithm For Wind Power Forecasting

Abstract

Due to the excellent wind power probabilistic prediction performance, Mixture Density Network (MDN) is used in short-term wind power forecasting, but the density leakage problem the Not a Number (NaN) loss problem and the choice of hyperparameters in the MDN seriously affect the model performance. GA-TDMDN is proposed in this paper for wind power probabilistic forecasting. GA-TDMDN uses truncated distribution as kernel function to solve density leakage. For the NaN loss problem that occurs during model training, different output layer activation methods and improved loss function are used for different mixture component parameters, so that the shape of the truncated normal distribution can be better controlled. Genetic Algorithms (GA) is used to optimize key hyperparameters in the MDN structure. The experimental results show that it is feasible to use truncated distribution to solve the density leakage problem, and using the GA algorithm to optimize the model structure can improve the model performance