Hybrid Neural Network Based on GRU with Uncertain Factors for Forecasting Ultra-short-term Wind Power

Abstract

The non-stationarity and stochastic nature of wind power bring difficult challenges to large-scale grid-connected of wind power. The ultra-short-term forecasting of wind power is used for balancing load and the optimal optimization of spinning reserve, which has high requirements for prediction accuracy. The neural network can solve the problem of feature selection, but in the task of wind power prediction, it is of great concern to find the optimal input features and model structure by mining physical correlation among features. Inspired by the physical formula for wind power, an uncertain factor is calculated, which caused by both environmental disturbance and wind turbine state changes. This paper proposes a method to predict ultra-short-term wind power, which using the features associated with wind power and the uncertain factors. Time series features are predicted through the Gated Recurrent Unit (GRU) Neural Network, and finally all the features were fused to form a hybrid neural network. The effectiveness of the proposed method has been confirmed on the real datasets derived from a wind field. Compared with the conventional time series dependent methods, our proposed method shows more reasonable results in terms of accuracy and availability.