Use of Physics to Improve Solar Forecast: Part II, Machine Learning and Model Interpretability
Machine learning (ML) models have been applied to forecast solar energy; however, they often lack clarity of interpretability and underlying physics. This work addresses such challenges by developing a hierarchy of ML models that gradually introduce predictors to improve the forecast accuracy based on a physics-based framework. Three ML models (ARIMA, LSTM, and XGBoost) are examined and compared with four physics-informed persistence models reported in Part I and the simple persistence model to assess the improvement of different models. The 7-year measurements at the U.S. Department of Energy's Atmospheric Radiation Measurement's Southern Great Plains Central Facility site are used for forecasts and evaluations. The results reveal that the step-by-step introduction of predictors leads to different improvements for models at different hierarchical levels. Comparison of the ML models with persistence models shows that LSTM and XGBoost outperform all the persistence models, with LSTM having the overall best performance; however, ARIMA underperforms the four physics-informed persistence models. This study demonstrates the importance and utility of incorporating physics into ML models in improving forecast accuracy by introducing a hierarchy of physics-based predictors, distinguishing predictor contributions, and enhancing the ML interpretability. The combined use of Global Horizontal Irradiance (GHI) and Direct Normal Irradiance (DNI) significantly improves the forecast accuracy compared to using individual irradiances alone because the pair contains more information on cloud-radiation interactions.