Collaborative data reconstruction and power prediction of wind turbine clusters

Wind power forecasting presents significant challenges due to missing data issues. Existing approaches frequently neglected inter-turbine correlations within wind farms and uncertainties in the data, which resulted in suboptimal performance and decreased precision. To resolve these challenges, this study proposes a collaborative multi-turbine wind power forecasting framework, which integrates an enhanced diffusion denoising probabilistic model (DDPM) with a temporal graph neural network (TimeGNN). The enhanced DDPM approach was utilized to directly learn the conditional distribution of observed turbine data, addressing uncertainties and correlations in data reconstruction. It employs a two-dimensional attention method to capture temporal dependencies and feature correlations in turbine power data, facilitating accurate data reconstruction and supplying high-quality input for subsequent forecasting tasks. The TimeGNN model was utilized to predict the power output of multiple turbines, effectively harnessing the spatiotemporal features of the turbine data for collaborative forecasting, thereby improving prediction performance. Experimental results demonstrate that the enhanced DDPM-TimeGNN method outperforms in both data reconstruction and downstream forecasting tasks. In comparison to the traditional mean reconstruction method, the proposed forecasting approach achieves a 5.27 % reduction in root mean square error within the case study.