Short-term wind speed forecasting based on spatial-temporal graph transformer networks

• Published in: Energy (Oxford) 2022-08, Vol.253, p.124095, Article 124095
• Main Authors: Pan, Xiaoxin, Wang, Long, Wang, Zhongju, Huang, Chao
• Format: Article
• Language: English
• Subjects: Alternative energy sources; Benchmarks; Energy sources; Errors; External attention mechanism; Feature extraction; Forecasting; Graph convolution; Mathematical models; Modules; Multilayer perceptrons; Nodes; Performance enhancement; Renewable energy sources; Short-term wind speed forecasting; Spatial and temporal correlations; Spatial dependencies; Temporal variations; Topology; Transformers; Turbines; Wind power; Wind speed; Wind turbines
• ISSN: 0360-5442; 1873-6785
• DOI: 10.1016/j.energy.2022.124095

Wind energy is a widely concerned renewable energy source. Accurate short-term wind speed forecasting is helpful for the stable operation of wind power systems, which is crucial to the wind power industry. In this paper, a Spatial-Temporal Graph Transformer Network (STGTN) is proposed to improve the performance of short-term wind speed forecasting. The proposed model consists of a temporal feature extraction module and a spatial feature extraction module and thus it can capture the temporal and spatial correlations between wind turbine nodes. A transformer based on the external attention mechanism and the graph convolutional layer is proposed to extract spatial features while a multilayer perceptron is employed to derive temporal features. Since the graph convolutional layer relies on the Euclidean spatial topology input, the location distribution of wind turbine nodes is not considered in the proposed model. To verify the performance of the STGTN model, five wind speed forecasting methods (with and without spatial dependencies) are employed as benchmarks. Experimental results show that the proposed model performs the best in terms of the mean absolute error, root mean square error and mean absolute percentage error for each forecasting horizon. •A wind speed forecasting method based on spatiotemporal information is proposed.•Euclidean spatial information is used to improve forecasting robustness.•Transformer with graph convolution is developed to capture wind speed features.•The effectiveness of the proposed method is verified on real datasets.