Enhanced wind power prediction via adaptive fusion of multi-dimensional spatial graph and global features

Accurate power prediction is essential for optimizing power grid scheduling and improving wind farm operational efficiency. However, many existing approaches inadequately model the complex global interdependencies between meteorological variables and power generation dynamics, which constrains further improvement of prediction model performance. Therefore, we propose a hybrid prediction framework that combines the adaptive fusion of multidimensional spatial graphs and global features. Firstly, we construct heterogeneous spatial graphs by quantifying the spatio-temporal correlations among multivariate working condition parameters. Secondly, sparse graph theory and node embedding algorithms are employed to preserve salient topological structures while capturing latent global correlations. Subsequently, by integrating the Graph Attention Network (GAT) and the Long Short-Term Memory (LSTM) network as the backbone and incorporating an adaptive gate mechanism, we design and propose a prediction model, namely GLG (GAT-LSTM with Adaptive Gate Mechanism). The GAT component captures spatial dependencies through attention-weighted aggregation, while the LSTM module extracts global dynamics to generate graph-level representations. Moreover, a feature fusion module is developed to extract global spatial features by constructing gating units and employing the attention mechanism for graph-level feature fusion. The experimental results indicate that the proposed method outperforms the common and classical power prediction methods in mean squared error (MSE), mean absolute error (MAE), and R-squared (R2).