Solar Forecasting: The Value of Using Satellite Derived Irradiance Data in Machine Learning Based Forecasts

Solar forecasts lower the cost of solar power and reduce the barriers to firm power generation. Solar forecasting conventionally has relied on advecting near real time observations and numerical weather predictions (NWPs). Observation based methods have limitations in cost and operational feasibility. NWPs generally have coarse spatial and temporal resolution, resulting in forecasts that may be overly general for a solar plant's location. Machine learning (ML) based forecasts have the potential to extract and blend observations and NWP data in an optimal blend, adding forecast skill. A primary drawback is ML based forecasts usually require training data. This paper will quantify the ML based forecast skill of using satellite derived irradiance data in lieu of ground, and the relationship between length of input training data and trained forecast skill gained. Climate and regional effects will be investigated by testing sites across the globe. Importantly these ML based forecasts will be compared to persistence forecasts and current NWP forecasts as a baseline.