UC San Diego

To effectively manage the growing density of solar power installations, grid operators are increasingly relying on solar forecasts to predict irradiance and power from minutes to days ahead. For forecast horizons of five hours or greater, large-scale computational fluid dynamics models known collectively as numerical weather prediction (NWP) is the most accurate method of predicting solar irradiance. Though more accurate than statistical regression or imagery based techniques, NWP is consistently erroneous and generally under-represents cloud cover; approximately half of all cloudy days are incorrectly forecast as clear. Furthermore, these errors are exacerbated for regions with dynamic cloud cover such as the California coast. On average, this results in irradiance over-predictions of 20-25%. Overall, these weaknesses limit NWP's utility for solar power forecasting. However, by coupling meteorological theory with statistical techniques and model improvements, a specialized forecasting system can be developed that significantly improves forecast accuracy. In this study, several advanced NWP techniques are implemented to improve forecast accuracy. First, the accuracy of the existing operational NWP is evaluated. By relating forecast accuracy to meteorological regime, model-output-statistics (MOS) is used to bias-correct forecasts and improve mean absolute error (MAE) by approximately 5%. Additionally, MOS was used to characterize forecast certainty; meteorological regimes with historically erroneous forecasts were prescribed with wider confidence intervals. Though valuable for operational forecasting, statistical models such as these fail to address the model weaknesses that prevent the accurate simulation of clouds. To directly improve cloud simulations, a high-resolution NWP specialized for solar irradiance forecasting in coastal areas was developed and implemented operationally (WRF- SRAF). Using Weather Research and Forecasting Model (WRF) architecture, WRF-SRAF was used to predict solar irradiance for days in which marine layer stratocumulus (MLS) clouds were observed in San Diego, CA. In general, it was found that simulated MLS clouds dissipated 1.9 hours earlier than observed and that model error was primarily attributed to inaccurate initial conditions and incorrect treatment of the surface energy balance. To improve model initializations, a method of direct-cloud assimilation was implemented. In this method, satellite observations were co-located with the model domain. For areas with observed cloud cover, modeled water vapor was raised to super-saturation, populating clouds in the initial conditions. Lastly, a cloud dissipation model was derived in order predict the influence of the surface heat fluxes on cloud evaporation. It was found that, due to extraordinarily dry boundary conditions, simulated sensible heat fluxes were 2-3 times larger than observed. Subsequently, a background uncertainty in the surface boundary conditions was established and used to motivate a systematic perturbation to future simulations. Using this surface modification in an ensemble setting, solar forecast accuracy improved by over 7% and the dissipation of MLS clouds was more accurately characterized