Non-stationary GNNCrossformer: Transformer with graph information for non-stationary multivariate Spatio-Temporal wind power data forecasting

The spatiotemporal prediction of wind power is of great significance for the grid connected operation of multiple wind farms in the wind power system. However, due to the complex temporal and spatial dependencies among multiple wind farms, developing advanced models to make accurate wind power predictions under their mutual influence is equally challenging. Furthermore, most of existing models are not ideal for long-term prediction of multivariate and non-stationary wind farm power datasets. To solve these problems, this paper proposes a novel Transformer-based model named Non-stationary GNNCrossformer for non-stationary multivariate Spatio-Temporal forecasting, utilizing Nonstationary-Two-Stage-Attention for both non-stationary cross-time dependency and cross-dimension dependency, as well as using the new graph convolutional neural network with Chebyshev interpolation for extracting temporally conditioned topological information from multiple wind farms efficiently. To tackle the dilemma between series predictability and model capability, we also propose Series Stationarization to complement Nonstationary-Two-Stage-Attention. While series stationarization makes sequence representation more generalized, the Nonstationary-Two-Stage-Attention can be devised to recover the intrinsic non-stationary information into temporal dependencies by approximating distinguishable attentions learned from raw series. Besides, the new graph convolutional neural network with Chebyshev interpolation can converge faster, be more robust, and have stronger generalization ability than the traditional one with Chebyshev approximation. In our experiment, two real-world wind power datasets were used to validate the proposed model. Numerical experiments have demonstrated the effectiveness and robustness of the proposed method compared to state-of-the-art spatiotemporal models.