An adaptive spatiotemporal fusion graph neural network for short-term power forecasting of multiple wind farms

Reliable short-term wind-power forecasting is crucial for the safe operation of power systems that accommodate large amount of wind power. With the development of large-scale wind power bases, wind power spatiotemporal correlations of multiple wind farms (WFs) can be used to improve short-term wind power prediction accuracy and efficiency. However, the time-varying and coupling characteristics of wind power spatiotemporal correlations present significant challenges for the accurate prediction of multi-WF power. In this paper, we propose an adaptive spatiotemporal fusion graph neural network for short-term power forecasting of multiple wind farms. Based on regional weather forecasts and historical wind power data, an adaptive directed graph is generated to model the dynamic spatial correlations between the WFs and the impact of regional weather factors on the wind power of different WFs. Graph convolution modules and temporal recursive modules are then alternately arranged to fuse the spatial and temporal features of wind power and to output the wind power prediction results for multiple WFs and time points. The case study results show that the proposed model outperforms the investigated baseline models in terms of prediction accuracy and that each model component has a unique contribution.