A Scale-Aware Graph Neural Network Framework via Multivariate Variational Mode Decomposition for Multi-Day Wind Speed Forecasting

Recent deep learning advances improve predictive performance but often increase computational and memory costs. This limits use in resource-constrained settings. Meanwhile, meteorological data exhibit strong multiscale characteristics. Training such signals with single-scale models can cause scale mixing and spectral bias, which degrade performance in extreme events and long-term forecasting.

Motivated by these challenges, this study explores an alternative strategy that enhances forecasting performance through scale-aware data preprocessing rather than increased model complexity. Multivariate Variational Mode Decomposition (MVMD) is integrated with graph neural networks (GNNs) to separate multi-scale temporal variability before spatial learning. Surface wind forecasting over Taiwan is characterized by complex atmospheric dynamics associated with typhoons, Meiyu fronts, and monsoon systems. It provides a challenging case for 72-hour wind speed forecasting.

ERA5 reanalysis data at a 0.25° spatial resolution and 12-hourly intervals over East Asia (5–40°N, 105–140°E) are used to construct a scale-aware spatio-temporal forecasting framework. The training dataset spans 2000 to 2016, the validation dataset spans 2016 to 2020, and the testing dataset spans 2020 to 2024. Raw surface wind fields are decomposed into five intrinsic mode functions (IMFs) using MVMD, with the number of modes selected based on a balance between root mean square error (RMSE), signal-to-noise ratio (SNR), and orthogonality index (OI). These scale-separated wind components with selected background meteorological variables (temperature, mean sea-level pressure, sea surface temperature, and 500-hPa variables) are incorporated into a three-layer Graph Attention Network (GATv2) model. The model is trained for one-step-ahead prediction, and multi-day forecasts are generated through an autoregressive rollout strategy that does not rely on additional temporal sequence encoders.

The MVMD–GATv2 model was compared with a baseline GATv2 trained directly on raw surface wind fields. Model performance is evaluated using mean absolute error (MAE), RMSE, and anomaly correlation coefficient (ACC). Preliminary results show that RMSE at the 12-hour forecast point decreased from 1.7 to 0.8. In addition to improved accuracy, ongoing analyses within this comparison framework focus on examining the evolution of errors across lead times and quantifying training costs. Further analyses assess the interpretability of scale-separated representations and explore boundary-related effects. In summary, these findings highlight the potential of MVMD as a scale-aware data preprocessing strategy that improves the accuracy, stability, and interpretability of graph-based regional wind predictions.