A Spatiotemporal Directed Graph Convolution Network for Ultra-Short-Term Wind Power Prediction

Abstract

The expansion of wind generation and the advance in deep learning have provided feasibility for multisite wind power prediction motivated by spatiotemporal dependencies. This paper introduces a novel spatiotemporal directed graph convolution neural network to sufficiently represent spatiotemporal prior knowledge and simultaneously generate ultra-short-term multisite wind power prediction. At first, a spatial dependency-based directed graph is established to learn the intrinsic topology structure of wind farms taking sites as graph nodes and Granger causality-defined spatial relation as directed edges. Subsequently, a unified spatiotemporal directed graph learning model is presented by embedding the multi-scale temporal convolution network as a sub-layer into the improved graph convolution operator, where the temporal features of each node are extracted by the above sub-layer to capture time patterns with different lengths, and the improved graph convolution layer is introduced by redefining <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">K</i> -order adjacent nodes to further share and integrate the deep spatiotemporal knowledge on the graph containing temporal features. Finally, under a comprehensive training loss function, this method is capable of improving the accuracy of each site for 4h-ahead prediction along with decent robustness and generalization. Experiment results verify the superiority of the proposed model in spatiotemporal correlation representation compared with classic and advanced benchmarks.