Climate change impacts on solar energy generation in the continental United States, forecasts from deep learning

Large-scale solar promises a low-carbon energy alternative. However, solar production in North America given anticipated climate change has been studied only seasonally in terms of solar irradiance. This work integrates more of the predictive potential of climate-change models by exploring other environmental variables, such as humidity and temperature. Here, a Continental US (CONUS) model is produced by deep learning using 2,593 NREL simulated solar power stations. Daily forecasts using 17 Global Climate Models (GCM’s) through 2099 are summarized monthly. Results suggest power production factors change between +4% and -19% over 93 years. These results suggest more, but still modest, potential declines than previous solar irradiance-based studies. The modest impact is encouraging. For some areas, climate model variability unfortunately yielded statistically insignificant trends and practical application is less clear. For future evaluations, this work suggests the potential importance of additional variables, monthly interval summary, and accounting for model variability. • Deep learning was effective at modeling solar power output with limited environmental input information. • Results show both geographic and seasonal changes in average solar energy output across CONUS. • Results emphasize the value of analyzing projections in monthly timeframes. • The incorporation of climate inputs in addition to solar irradiance enhanced predictive performance. • Varied results presented by different GCM’s benefitted from trend analysis.