Integrating large-scale wind generation into electrical grids requires accurate short-term power forecasts at the regional level, accounting for dependencies across multiple geographically distributed farms. Researchers have introduced ST-DualFormer, a transformer architecture designed to handle the complex spatio-temporal patterns inherent in regional wind power forecasting (RWPF).
The key innovation separates temporal and spatial encoding into two parallel processing streams. The temporal encoder captures how wind power evolves over time at each farm, while the spatial encoder identifies correlations and dependencies between farms across a region. Rather than relying on predefined spatial connection graphs, the model uses attention mechanisms to dynamically learn relationships among all farms simultaneously, offering greater flexibility in capturing real-world wind patterns.
When tested on data from 28 wind farms, ST-DualFormer achieved a normalized mean absolute error (nMAE) of 5.25% and normalized root mean squared error (nRMSE) of 7.53% for three-day-ahead forecasts. These performance metrics substantially outperformed conventional graph-based and other transformer-based baseline methods currently in use.
The dual-stream design offers practical advantages for grid operators. By independently modeling how weather patterns drive temporal variations while simultaneously capturing the spatial coordination effects across a region, the architecture can integrate multiple data sources—including meteorological observations and historical generation records—more effectively than single-pathway models.
Testing on subsets from the Weather2K-R dataset provided initial evidence that the architecture generalizes beyond its original training region, suggesting potential applicability across different wind-rich geographies. This transferability is commercially significant, as utilities could potentially apply trained models to new regional deployments with minimal retraining.
For grid operators managing increasing wind penetration, improved regional forecasting directly enhances unit commitment decisions, reduces reserve margins needed for variability, and minimizes curtailment. As renewable energy deployment accelerates globally, such advances in machine learning-based forecasting become increasingly valuable for reliable, efficient grid operation.
