UC San Diego

Solar forecasts are important for low-cost integration of solar energy into the smart grid. Accurate intra-hour predictions of irradiance quantify the variability of solar power at ground level, reduce the uncertainty in power output from solar farm, and are important for real- time grid balancing and management. A multilayered-hybrid- algorithm method is developed to generate real-time intra- hour prediction intervals (PIs) for both global and direct solar irradiance. This forecasting method integrates stochastic learning methods for the prediction of solar irradiation and local sensing techniques for the introduction of exogenous inputs. The research of the proposed forecasting method consists of four objectives : (1) Development of a smart forecasting engine based on advanced stochastic learning methods. (2) Development of an image-based cloud detection system using a cost- competitive fish-eye camera. (3) Integration of the smart forecasting engine with the cloud detection system to create a high-fidelity forecasting model. (4) Development of a hybrid algorithm to provide prediction intervals for the integrated forecasting model. The forecasting method introduced here is deployed in real-time and achieves forecast skills up to 20% over the reference persistence model. Real-time PIs generated from this method achieve coverage probabilities which are consistently higher than the nominal confidence level (90%) regardless of weather condition