上海科技大学知识管理系统

With increasing penetration of renewables, wind power forecasting (WPF) is of great importance for the stable operation of modern power grid. The rapid development of artificial intelligence (AI) enables higher accuracy for WPF. This paper proposes a data-driven ultra-short term WPF method that incorporates physics information within the WPF problem. Firstly, the physics model of WPF problem is built, and the mapping from the wind speed to the wind power is approximated through physics. With the predicted wind speed from numerical weather prediction (NWP) and the physics model, the preliminary wind power output can be approximated. Secondly, the residual between the preliminary predicted value and the actual wind power is described by a data-driven network. Specifically, the structures of temporal convolutional network (TCN) and graph convolutional network (GCN) are adopted to extract the time-space features within the input data, and to determine the mapping between the input data and the residual. Finally, the preliminary predicted value and the predicted residual value are added together to get the final predicted wind power. Experimental results using practical wind farm data show that the proposed ultra-short term WPF method presents higher forecasting accuracy compared to methods that are only based on physics or data-driven method.