We began work using infrared satellite channels in an effort to improve pre-dawn forecasts of early morning solar generation. We then added infrared channels to our daytime algorithm to help discriminate snow cover from clouds and better account for the impacts of aerosols, thin cirrus, cloud phase and other subtle features on observed surface irradiance. We have progressed to a hybrid system in which a boosted regression tree machine learning algorithm is used to train two models, one for use in forecasts issued at night and at times of very low solar zenith angles, but trained during the day, using infrared satellite channels only (and excluding the near-IR channels, which are impacted by reflected near-IR solar radiation during the day), and one for use in forecasts issued in day-time, trained against all channels.
The output of this forecast process is an estimate of the two-dimensional field of the fraction of clear-sky radiance over a large area (e.g. the western United States). To forecast the short-term evolution of this field, we use a feature-tracking algorithm to extrapolate cloud motion and deformation for the next several hours at 5 minute time-steps. For locations where observations of solar generation are available in real time, these forecasts are, together with a timeseries of recent observed generation, passed to a second machine learning model, which is trained on several months of archived generation timeseries and satellite forecasts.
Here we present diagnostics of the relative weights of the various IR and visible channels and discuss the physical interpretation of the functional form of their contributions to the final forecast via SHAP plots (Lundberg and Lee, 2017). We will also explore the lead-time dependence of the relative importance of cloud tracking and timeseries predictors to forecast skill over a range from five minutes to three hours.