Implementation and Validation of a Surface Parameterization Ensemble for Solar Power Prediction

Numerical weather prediction (NWP) has several well documented sources of error which limit its practical utility. In general, NWP models such as the Global Forecast System (GFS) and North American Mesoscale (NAM) models predict too little cloud cover and relatively poorly resolve variability in cloud, particularly marine stratus. Direct cloud assimilation into a mesoscale NWP model can significantly reduce error in forecast cloud cover due to improved initialization. However, the positive impact of this initialization diminishes for long forecast horizons, which specifically serve day-ahead transactions and reserve planning for utilities. Furthermore, cloud assimilation fails to address the source of the error in the model physics and potentially unrealistic interactions between the land surface model (LSM) and the lower atmosphere. An ensemble-based cloud forecast system may be implemented in which model physics parameterization is diversified. In this way, a diverse set of cloud fields can be resolved per prediction, ultimately allowing for reduced forecast error and better representation of forecast uncertainty.

Sensible and latent heat exchange, between the surface and the atmosphere, encapsulated by the Bowen Ratio, is particularly important for cloud formation. One effect of increasing the latent heat flux from the surface is to lower the lifting condensation level (LCL) and to decreases atmospheric stability – factors which promote cloud cover. In this study, the Weather Research and Forecasting (WRF) model, developed at the National Center for Atmospheric Research (NCAR) is used to simulate cloud cover using a variety of Bowen ratios over a 4-month period over the southern United States. For each day, five Bowen ratios were used. For each member, the resulting irradiance forecasts were applied to photovoltaic power plant models for conversion to a solar power forecast. The relative accuracy of each member was then compared to the direct-cloud assimilation and persistence based statistical techniques. The ideal combination of all techniques was then used to determine the relative weightings of each forecast. Overall, it is shown that for very short forecast horizons, the persistence based technique is weighted most. For several hours to a day in advance, the direct-cloud assimilation technique receives greatest weight. For forecast horizons beyond the day-ahead, the Bowen ensemble proves to be most skillful. We speculate on the causes for relative contributions, assess the relative impacts of technique and integrated observation data sources on forecast performance and application to its end user, and describe ways to effectively combine the techniques for an improved operational power forecast.