Staring at the sun

Abstract

Developing accurate solar performance models, which estimate solar output based on a deployment's unique physical characteristics and weather, is increasingly important as the aggregate energy generated from solar rises. Since manually developing white box physical models based on site-specific information requires expert knowledge and thus does not scale, recent research focuses on black box approaches that use training data to automatically learn a custom machine learning (ML) model. Unfortunately, this approach requires months-to-years of training data, and often does not incorporate well-known physical models of solar generation, which reduces its accuracy. To address the problem, we develop a physical black-box modeling approach that leverages many of the same fundamental properties as existing white-box models.Rather than manually determining values for physical model parameters, our approach automatically calibrates them by finding values that best fit the data. This calibration requires much less data (as few as 2 datapoints) than training a ML model, as the physical model already embeds the complex relationship between the input parameters and solar output. In developing our approach, we isolate the effects of 10 different weather metrics on solar output from nearly 343 million hourly weather and solar readings, or 78,435 aggregate years, gathered from 11,205 solar sites. We show that our physical model accurately describes weather's effect on solar output at all sites, obviating the need for training custom ML models using weather metrics. Instead, we augment our physical model by applying ML to learn only the relationships that are unique to each site, specifically non-weather-based shading. We evaluate our approach on solar and weather data from 100 sites, and show it yields higher accuracy than current state-of-the-art ML approaches.