Deep Belief Network-Based Hammerstein Nonlinear System for Wind Power Prediction

The wind power systems have the features of complex physical relationship, nonlinearity, and randomness, which pose great challenge to establish wind power system model and make a reasonable power prediction. In this article, a deep belief network (DBN)-based Hammerstein system for wind power prediction is developed by applying separable signals, in which the Hammerstein system is made up of static nonlinear block and dynamic linear block in series. With the goal of examining the nonlinear and linear information encompassed within temporal series data, DBN and autoregressive exogenous (ARX) model is used to elucidate the potential distribution properties inherent in wind power systems. To achieve a prediction model with a high degree of precision, separable signals are used to decouple the static nonlinear and dynamic linear characteristics. Furthermore, to decrease burden and increase the accuracy of prediction model, quartile data cleaning technique including horizontal and vertical dimensions is used for eliminating the abnormal data of wind power systems. The presented methodology is validated on wind power plant, and the simulation results verify that the developed DBN-based Hammerstein system has significant advantage over other prediction models involved in this article for prediction accuracy and generalization capability.