A method for ultra-short-term wind power forecasting of large-scale wind farms based on adaptive spatiotemporal graph convolution

To reduce the adverse effects of wind power uncertainties on power systems, a precise ultra-short term power prediction method for wind farm clusters was proposed. The power prediction results of multiple wind farms with multiple time steps were output synchronously. Based on the statistical results of long-term historical data, the wind farm cluster is abstracted into a graph topology structure, and the graph structure is adaptively adjusted to the dynamic change of spatial correlation during network training. Then, a spatiotemporal graph neural network is constructed to extract spatiotemporal features by integrating multi-source features of multi-wind farms with improved multi-head attention mechanism. The multi-task learning mechanism is introduced to optimize the output layer of the network and synchronously output the multi-step power forecasting results of all wind farms. Taking a wind power cluster composed of 18 wind farms in Jilin, China as the research object, the normalized root mean square error of the 4h ahead forecasting result is 0.1161, which is 0.0173 lower than that of the persistence method.