Harnessing Synoptic-Scale Information in Wind and Photovoltaic Energy Forecasting Using Machine Learning

A supervised machine-learning regression framework is presented for forecasting wind and photovoltaic (PV) power generation by integrating local and synoptic-scale meteorological information. The approach is evaluated across multiple sites, including 39 wind and 18 PV stations in Mexico, and 3 wind and 8 PV stations in China. For each station, an XGBoost regression model is trained to predict hourly energy production using local meteorological variables, derived from ERA5 reanalysis data for Mexico and on-site measurements for the Chinese stations.

To assess the added value of large-scale atmospheric information, dimensionally reduced synoptic-scale predictors extracted from ERA5 using self-organizing maps and principal component analysis are incorporated. These predictors are designed to represent dominant atmospheric circulation patterns potentially influencing local renewable energy production. Model performance is assessed through station-specific cross-validation, comparing configurations with and without synoptic-scale features across multiple predictor combinations.

Results indicate that the inclusion of synoptic-scale atmospheric patterns can improve short-term power forecasts at several locations, although the overall gains are generally modest. The analysis suggests that improvements in local meteorological inputs are likely to yield larger increases in forecast skill than further refinement of synoptic-scale representations. Nevertheless, the proposed framework demonstrates clear operational relevance: when customized for individual stations, synoptic-scale information can contribute to improved forecasting performance while maintaining the computational efficiency of machine-learning-based methods.