Spatiotemporal forecasting using multi-graph neural network assisted dual domain transformer for wind power

Abstract

Accurate prediction of wind power generation is crucial for operational and maintenance decision in wind farms. With the increasing scale and capacity of turbines, incorporating both temporal and spatial characteristics has become essential to improve prediction accuracy. In this paper, a novel spatiotemporal multi-step wind power forecasting method using multi-graph neural network assisted dual domain Transformer is proposed. Specifically, to adequately represent the heterogeneous dependencies among wind turbines, multi-relational graphs are constructed and integrated into a unified graph via attention mechanisms. Subsequently, the spatiotemporal fusion module (STFM) is developed using graph convolutional network and one-dimensional convolutional neural network to capture temporal and spatial features simultaneously. Moreover, the time–frequency dual domain Transformer (DDformer) is devised to fully utilize the information extracted by the STFM. Sequence learning in DDformer is performed through three perspectives, including multi-head self-attention mechanism, intrinsic mode function attention mechanism, and residual connection. Finally, the comprehensive evaluation metrics are formulated to assess the overall performance of wind power forecasting at both individual turbine and entire farm levels. Extensive simulations on a real-world dataset are conducted for multi-step forecasting, covering time horizons ranging from 10 min to 6 h ahead. In the case study, the proposed method consistently outperformed advanced benchmarks and ablation models, achieving average comprehensive normalized mean absolute error and normalized root mean square error of 5.8469% and 8.9461%, respectively, with improvements of 38.35% and 33.72%. Overall, the effectiveness of multi-step forecasting makes this study provide valuable insights into a new framework for wind power forecasting.