It is well established that most operational numerical weather prediction (NWP) models consistently over-predict irradiance. While more accurate than imagery-based or statistical techniques, their applicability for day-ahead solar forecasting is limited. Overall, error is related to the expected meteorological conditions. For regions with dynamic cloud systems, forecast accuracy is low. Specifically, the North American Model (NAM) predicts insufficient cloud cover along the California coast, especially during summer months. Since this region represents significant potential for distributed photovoltaic generation, accurate solar forecasts are critical. To improve forecast accuracy, a high-resolution, direct-cloud-assimilating NWP based on the Weather and Research Forecasting model (WRF-CLDDA) was developed and implemented at the University of California, San Diego (UCSD). Using satellite imagery, clouds were directly assimilated in the initial conditions. Furthermore, model resolution and physics parameterizations were chosen specifically to facilitate the formation and persistence of the low-altitude clouds common to coastal California. Compared to the UCSD pyranometer network, intra-day WRF-CLDDA forecasts were 17.4% less biased than the NAM and relative mean absolute error (rMAE) was 4.1% lower. For day-ahead forecasts, WRF-CLDDA accuracy did not diminish; relative mean bias error was only 1.6% and rMAE 18.2% (5.6% smaller than the NAM). Spatially, the largest improvements occurred for the morning hours along coastal regions when cloud cover is expected. Additionally, the ability of WRF-CLDDA to resolve intra-hour variability was assessed. Though the horizontal (1.3 km) and temporal (5 min) resolutions were fine, ramp rates for time scales of less than 30 min were not accurately characterized. Thus, it was concluded that the cloud sizes resolved by WRF-CLDDA were approximately five times as large as its horizontal discretization. (C) 2013 Elsevier Ltd. All rights reserved.