Solar Irradiance Forecasting with Visible Spectrum Sky View Images and Random Forest Regression

Integrating solar photovoltaic (PV) energy into the electrical grid is challenging due to its inherent production variability stemming from clouds. Solar forecasts and grid inertia from our current electrical infrastructure help maintain grid stability and availability in the wake of PV production fluctuations. However, to transition to future high-renewable-energy-penetration grids, grid operators will require improved forecasting to proactively deploy grid-balancing assets and maintain grid reliability. In this work, we use 3-channel, 8-bit-depth visible spectrum sky-view images from a year-long dataset captured at 15-second intervals, and smart persistence features to predict broadband irradiance in North Cape, Canada over several forecasting horizons. Multiple features are extracted from a single sky-view image to train and test random forest regression models and evaluate feature importance over forecast horizons from 1 minute to 1 hour. The practicality and usefulness of the models are evaluated using testing and training computational times, and the prediction accuracies relative to the smart persistence model, respectively. The models achieve a positive skill score in the range of 0 to 0.1, outperforming the smart persistence model for all forecast horizons.