Ultra-short-term wind power interval prediction based on multi-task learning and generative critic networks

Wind power forecast has played a significant role in modern power systems operation. Meanwhile, interval forecast, as a practical way to represent wind power uncertainty, has attracted considerable attention. In this paper, we propose a novel wind power interval forecast method for multiple wind farms in near regions based on machine learning techniques. First, existing interval forecast methods mainly utilize meta-heuristic algorithms to train the networks, which however, suffer from heavy computation burden and local convergence problem. To remediate this problem, a interval forecast method called Generative Critic Networks (GCN) is proposed, which applies gradient descent algorithm in the parameters optimization and further improve the forecasting performance by a function approximation. Second, considering the spatial correlation of neighboring wind farms, the prediction of these outputs can be regarded as related tasks, thus Multi-Task Learning (MTL) is used as a base to achieve a joint interval forecast of multiple wind farms. Therefore, a unified deep learning model, Multi-Task GCN (MTGCN), is formed to achieve high-quality PIs of multiple wind farms. Finally, experimental results on different datasets show that the proposed algorithm can obtain high-quality prediction interval than other methods, leading to a reduction of at least 9.5% in the interval width. © 2023 Elsevier Ltd