The increasing penetration of solar power in the grid has motivated a strong interest in developing real-time performance models that estimate solar output based on a deployment’s unique location, physical characteristics, and weather conditions. Solar models are useful for a variety of solar energy analytics, including indirect monitoring, forecasting, disaggregation, anonymous localization, and fault detection. Significant recent work focuses on learning “black box” models, primarily for forecasting, using machine learning (ML) techniques, which leverage only historical energy and weather data for training. Interestingly, these ML techniques are often “off the shelf” and do not incorporate well-known physical models of solar generation based on fundamental properties. Instead, prior work on physical modeling generally takes a “white box” approach that assumes detailed knowledge of a deployment. In this paper, we survey existing work on solar modeling, and then compare black-box solar modeling using ML versus physical approaches. We then i) present a configurable hybrid approach that combines the benefits of both by enabling users to select the parameters they physically model versus learn via ML, and ii) show that it significantly improves model accuracy across 6 deployments.
