Wind Power Prediction based on Potts Model Feature Optimization

With the development of new energy, the wind power prediction problem has been a popular research direction. Most of the forecasting models only focus on the relationship between variables and the forecasting series, ignoring the complex interactions between wind power variables and the autocorrelation in time. We propose a new time-series predictor-variable optimization method, PMFO (Potts Model Feature Optimization), in conjunction with the Potts model in classical aerodynamics, and combine it with the BERT model in the neural network approach to obtain more accurate predictions. First, the aerodynamic field of the Potts model is analyzed as an excitation model for the feature planning of wind power generation variables. In this model, we treat the time series of wind generation variables as a kind of path sum of infinitesimal trajectory increments, where each infinitesimal increment is controlled by the Potts gas model. This is followed by a modified input layer into a BERT network for power prediction to obtain the results. Finally, we validate the effectiveness of the method with a real power grid dataset, and also embed the method into other neural network models, all of which show varying degrees of performance improvement, demonstrating the excellence of the method in this paper.