Improved Solar Power Forecasting Using Cloud Assimilation into WRF

Transcription

1 1 Improved Solar Power Forecasting Using Cloud Assimilation into WRF P. Mathiesen 1,2, C. Collier 1, J. Parkes 1, L. Landberg 1, and J. Kleissl 2 1 GL Garrad Hassan 2 University of California, San Diego Abstract Though currently the best available method for predicting irradiance at forecast horizons of longer than 5 hours, the operational numerical weather prediction (NWP) models are consistently erroneous. Generally, NWP overpredict irradiance, implying an under-prediction of cloud cover. In frequently cloudy regions this error is especially prominent. Overall, model error can be attributed to several sources: Domain resolution, model physics parameterizations, and inaccurate initial conditions. To address these error sources, a high-resolution, cloud-assimilating NWP for solar irradiance forecasting was developed and implemented at the University of California, San Diego, and GL-Garrad Hassan, America, Inc (WRF- CLDDA). Here, model resolution and physics parameterizations were chosen to accurately simulate local cloud conditions. Furthermore, a cloud-assimilation system was implemented to directly initialize clouds into the model. Simulations were performed for 5/1/2011 to 6/30/2011 for coastal California, a region with complex meteorology and frequent summertime cloud cover. Overall, WRF-CLDDA MAE was 5.2% (10.2%) smaller than the operational NWP for intra-day (day-ahead) irradiance forecasts. Index Terms Solar Forecasting, Numerical Weather Prediction, Direct-Cloud Assimilation I. INTRODUCTION The accurate characterization of cloud fields, their evolution, and their optical properties is critical for solar irradiance forecasting. For short-term forecasting, imagery based cloud-advection techniques [1]-[2] provide excellent characterizations of cloud fields and cloud motion. However, clouds are highly dynamic and cloud properties can change drastically over just a few hours. As such, the accuracy of frozen-cloud advection techniques diminishes significantly over the first six hours. For longer forecast horizons, physicsbased weather models (numerical weather prediction (NWP)) are generally regarded as the most accurate method for predicting solar irradiance [1]. Though more accurate than cloud-motion techniques, previous studies have conclusively demonstrated that NWP irradiance forecasts are consistently and systematically erroneous [3]-[9]. Regardless of model, irradiance NWP forecasts are generally biased high. This consistent underprediction of cloud cover demonstrates the limitations of the current operational NWP for solar irradiance forecasting. Coarse model resolutions and simple physics parameterizations contribute to NWP cloud cover error. The operational NWP models generally have spatial resolutions on the order of 10 km or larger. In this configuration, it is impossible to resolve fine-scale cloud features and little irradiance variability is predicted. Past studies [10]-[11] have found that as model resolution is increased, so did the ability to accurately characterize cloud-cover. Furthermore, the parameterization of physical processes, specifically the simulation of cloud microphysics and planetary boundary layer (PBL) mixing, has a large impact on cloud and irradiance forecast accuracy. [12] thoroughly catalogued the effect that different physics parameterizations have on NWP simulated cloud fields. Additionally, accurate model initialization is critical for NWP forecast accuracy. To minimize initialization error, observation data can be assimilated into the initial conditions. Notably, the Rapid Update Cycle (RUC, [13]) uses a cloudanalysis system to assimilate cloud observations into the model initial conditions. In this system, satellite imagery, radar data, and local cloud cover reports are used to construct an observed three-dimensional cloud field matrix. Clouds are built into the initial conditions by directly modifying the model hydrometeors (cloud and water mixing ratios) and the state variables which support them [14]-[17]. Similar systems [18] are in use in the Center for Analysis and Prediction of Storms (CAPS) Advanced Regional Prediction System (ARPS, [19]). In this study, a new, high-resolution, cloud-assimilating NWP model is developed and tested at the University of California, San Diego (UCSD) for solar irradiance forecasting (WRF-CLDDA). Using fine spatial resolution, physics parameterizations that promote cloud-cover formation, and a cloud-assimilation system, this model is specifically designed to minimize the errors typically associated with NWP irradiance forecasts. Using WRF-CLDDA, irradiance forecasts are produced for 5/1/11 6/30/11 and validated against the dense UCSD pyranometer network. Forecast accuracy is compared with the North American Mesoscale Model (NAM) for all varying sky conditions. Overall, it is shown that WRF-CLDDA is significantly more accurate than 703

2 2 the NAM for both intra-day and day-ahead irradiance forecasts. increases, the length of time for which the columnar assumption is valid diminishes and the time of day dependency on accuracy is magnified. II. NUMERICAL WEATHER PREDICTION (NWP) MODELS A. The North American Mesoscale Model (NAM) One of several operational NWP for the continental United States, the North American Mesoscale Model (NAM) is based on the Weather and Research Forecasting Non-Hydrostatic Mesoscale Model (WRF-NMM) and is published by the National Centers for Environmental Prediction (NCEP). The NAM uses 12 km horizontal resolution and 60 vertical hybrid sigma-level terrain-following coordinates. Temporally, the time step size is 30s and output is available hourly to a maximum forecast horizon of 36 hours. The NAM model is initialized by the NAM Data Assimilation System. Data is assimilated using the three-dimension variational (3DVAR) grid-point statistical interpolation (GSI) method [20]. Assimilated data include temperature, pressure, relative humidity, and wind magnitudes/directions as determined from radiosonde, satellite, and other observations [21]. Initialized shortly before the North American sunrise, irradiance forecasts from the 12 UTC initialization are slightly more accurate. As such, this was the only NAM data used in this study. Within the NAM, cloud evolution is primarily dependent on the 2-class Ferrier microphysics package [22] and radiative transfer is calculated using the Geophysical Fluid Dynamics Laboratory (GFDL-SW, [23]). Full details of the NAM are available in [24]-[25]. B. The Weather and Research Forecasting Model (WRF) The Weather and Research Forecasting (WRF, [26]) model is community developed and highly customizable numerical weather prediction model that is maintained by the National Center for Atmospheric Research (NCAR). In this study, WRF V3.3 was configured with three-nests of horizontal resolutions of 12 km, 4 km, and 1.3 km (Figure 1, centered at the University of California, San Diego). Boundary conditions for the outer domain were derived from the NAM. To facilitate low-altitude marine layer fog and stratus formation, the domain was vertically divided into 50 levels with 15 below 1000 m. The model time-step size was defined dynamically, with a maximum allowable CFL criterion of 0.8. To establish the irradiance forecast, downward short-wave irradiance was output every 5 min. to a maximum forecast horizon of 15 hours. Additionally, cloud microphysics are parameterized using the 6-class hybrid single/double-moment Thompson microphysics package [27]. Planetary boundary layer (PBL) mixing is parameterized using the Mellor-Yamada-NakanishiNiino scheme (MYNN, [28]). Finally, irradiance is calculated using the rapid radiative transfer model (RRTM, [29]-[30]). The RRTM is a spectrally-dependent model and accounts for differences in atmospheric absorption, reflection, and transmittance at various wavelengths. Like the GFDL-SW model, the RRTM is columnar. As model resolution Fig. 1: WRF-CLDDA nest configuration (Δx = 12 km, 4 km, 1.33 km) centered at the University of California, San Diego. Inset shows location of the UCSD pyranometer network. III. DATA SOURCES A. GOES Imagery Observed cloud field information is determined using Geostationary Operational Environmental Satellite (GOES) imagery. GOES Surface and Insolation Products (GSIP) level-2 data provides derived measurements of cloud-toptemperature (CTT) with approximate 4 km spatial resolution [31]. To ensure high-quality data, CTT data is spatially convolved to 12 km horizontal resolution to remove noise. Furthermore, data that is missing or outside of historical limits is ignored. Cloud vertical position is derived from the intersection of WRF-simulated temperature profiles and observed CTT. B. DEMROES Irradiance Network The Decision Making using Real-time Observations for Environmental Sustainability (DEMROES) network is a highdensity system of weather stations located at the University of California, San Diego (Fig. 1, inset). Meteorological data such as wind speed/direction, temperature, and precipitation is recorded at frequency of 1 Hz. Additionally, global horizontal solar irradiance at the surface is measured to within ±5% [32] using Licor LI-200SZ photodiode pyranometers. To match WRF-CLDDA/NAM GHI output, 1s irradiance data is temporally averaged to 5 min. or 1 hr. Furthermore, to ensure that similar spatial scales are being compared between NWP and observed data, DEMROES GHI measurements are averaged across WRF-simulated grid cells prior to comparison. In this study, only the WRF-CLDDA grid cell containing 5 DEMROES stations was analyzed. IV. METHODOLOGY A. WRF Cloud Assimilation System (WRF-CLDDA) 704

3 3 To improve the accuracy of irradiance forecasts, the initial cloud field was specified from observation. Since the benefit of data assimilation drastically diminishes over the first few hours [33], relying on assimilated state variables to produce accurate initial cloud fields becomes impractical as the assimilated data is nearly always outdated. To circumvent the cloud microphysics package and immediately populate the initial conditions, direct-cloud assimilation has been implemented in this study. In direct-cloud assimilation, cloud hydrometeor and water vapor mixing ratios are directly injected into the model initialization, eliminating the need for cloud spin-up and ensuring that cloud initializations are not obsolete. Observations Cloud Analysis System WRF Model Initialization - Pre- processing - Ingest NAM Spin- Up Time (2 hrs) GOES Imagery Initial Cloud Fields Direct Comparison Cloud Contingency Table Cloud Building/Clearing Updated Model Fields - qcloud - qvapor To initialize a cloud in the model, the water vapor mixing ratio (qvapor) is raised to 110% of saturation (1) in observed cloudy cells. qvapor = 1.1ws (1) Here, ws is the saturation water vapor mixing ratio and is a function of pressure, P, and the saturation water vapor pressure, es (2). ws = ε es P es (2) Here, ε is the ratio of gas constants of moist and dry air. es (3) is calculated from model-simulated temperature profiles using the Clausius-Clapeyron Equation. L 1 1 es = 6.11exp v Rv T (3) Excess qvapor is subsequently converted to cloud water and cloud ice via the microphysics scheme based on temperature. In this way, clouds are effectively and nearly instantaneous initialized in the initial conditions. Similarly, for clear observed grid cells, qvapor is reduced to a maximum of 75% of saturation and qcloud set to zero. To ensure numerical stability, cloud assimilation is not performed within five points of the boundary and large introduced qvapor and qcloud gradients smoothed to ensure numerical stability. Full details on the data-assimilation method are available in [34]. Figure 3 shows an example of the cloud assimilation result for 6/12/2011. Intra- day forecast Fig. 2: WRF-CLDDA cloud assimilation system overview. Figure 2 depicts a general outline of the WRF-CLDDA forecast system. This system of direct-cloud assimilation is based off of previous cloud analysis systems such [14]-[18]. First, the WRF model is initialized with 12 UTC NAM data. To assimilate cloud cover, GOES satellite imagery is colocated onto the WRF grid. Cloud-top location is calculated independently for each column via the intersection of the WRF-simulated vertical temperature profile and observed cloud-top-temperature. For sea-bound and coastal columns, cloud-top location is fixed to the top of the temperature inversion. Using a constant cloud thickness assumption, a second 2-D map of cloud base is derived for the entire WRF domain. Finally, the 2-D cloud-top and cloud-base location maps are combined to produce a three dimensional binary matrix of cloud location that is co-located with the WRF domain. Fig. 3: Comparison of GOES cloud mask (Fig. 3a) and columnar qvapor of assimilated initial conditions for 6/12/2011. V. RESULTS Qualitatively, WRF-CLDDA accurately predicts the evolution of cloud cover, even 12 hours after initialization. Figure 4 compares the forecast clear sky index (kt*) against GOES satellite imagery. kt* (ktm) is the ratio of forecast (observed) irradiance to the expected irradiance given clear conditions (4). kt * = GHI Forecast GHI CSK (4) 705

4 4 A clear sky index of 1 indicates completely clear conditions or 100% of GHICSK. Fig. 4: Comparison of WRF-CLDDA kt* forecast (a) with GOES satellite imagery (b) for 5/17/2011 off of the California coast. The spatial positioning of WRF-CLDDA-predicted clouds matches observed well. On 5/17/2011 at 07 PST, an expansive cloud system was observed off of the California coast (Fig. 4b, 7 PST). On this day, a low pressure center was located over northern California, creating strong on-shore flow. This pushed the coastal clouds to the east where they dissipated through mixing with warm, continental air. In the wake of this cloud system, few clouds were produced and clear conditions were observed at sea. Overall, this pattern was captured well by WRF-CLDDA. The cloud system was accurately assimilated into the initial conditions (Fig. 4a, 7 PST) and its movement and dissipation were accurately characterized throughout the day. Quantitatively, irradiance error is summarized in Table 1. Here, intra-day (day-ahead) persistence forecasts are calculated by multiplying the hourly-average ktm from 24 (48) hours prior with clear sky irradiance. As expected, WRFCLDDA is more accurate than the persistence models; dayahead rmae is 11.2% smaller. However, WRF-CLDDA only improves over the 24 hr persistence model by 1.9%. Since the presence of marine layer stratus (the dominant cloud feature in coastal California) is dependent on slowly changing, large scale meteorological conditions (e.g. the general circulation pattern and marine temperature inversion characteristics), coastal cloud cover patterns are often similar on consecutive days. Under these conditions, the 24 hr persistence model generally captures the irradiance signal well. Table 1: Summary of hourly-average error metrics for persistence, NAM, and WRF-CLDDA irradiance forecasts at UC San Diego. Error (% GHICSK) Overall Intra-Day Pers. NAM Day-Ahead WRF Pers. NAM WRF rmbe rmae rstderr Consistent with previous studies [5], NAM intra-day forecasts are positively biased (rmbe = 18.1%) indicating an under-prediction of coastal cloud cover. Comparatively, WRF-CLDDA irradiance forecasts are only slightly positively biased (rmbe = 2%), representing a large improvement in accuracy. In terms of rstderr, however, WRF-CLDDA improves by only 1.1%. Since relative standard error is the bias-corrected rrmse it represents the random component of forecast error and can be interpreted as the spread of the forecast error distribution. Thus, though WRF-CLDDA is much less biased, intra-day forecast errors are similarly as random as the NAM. In general, NWP rmae and rrmse increase by 1%-3% between intra-day and day-ahead forecasts [6]. Here, NAM rmae similarly increases to 29.5% (a jump of 2.7%) for dayahead forecasts. NAM rmbe, however, decreases to 10.3%. Consequently, since less of the forecast error can be contributed to a systematic bias, NAM rstderr increases by 5.5% to 39% indicating that the day-ahead NAM error is more random. 706

5 5 Significantly, WRF-CLDDA day-ahead and intra-day accuracies are approximately equal. WRF-CLDDA day-ahead rstderr is 31% (an improvement of 8% over the NAM). In general, WRF-CLDDA accurately predicts the overnight reformation of marine layer cloud cover. Furthermore, clouds assimilated within the large outer WRF-CLDDA domain likely increase day-ahead accuracy. Since the outer domain of WRF-CLDDA is very large (1500 km x 1500 km), clouds (especially frontal systems) can be assimilated and advected for 24 hours or more before impacting the region of interest and improving the day-ahead irradiance forecast. Previously, [8] found WRF rmbe of 62% - 75% (Table 2) for observed overcast conditions (kt m < 0.4). Overall, it was concluded that WRF predicted clear skies too frequently and that WRF accuracy decreases as observed cloud fraction increases. Here, considering only observed overcast conditions WRF-CLDDA intra-day rmbe was 16.8%. While still over-predicting irradiance, this demonstrates that the combination of model configuration and data assimilation increases the likelihood that cloudy conditions will be forecast correctly. Clear conditions, however, were slightly negatively impacted. [8] found rmbe between 2% and 4%, while WRF- CLDDA rmbe was -5.5%. Thus, for some clear observations, clouds were formed or failed to evaporate in WRF-CLDDA. Table 2: Summary of hourly-error metrics by sky condition. rmbe (%) rrmse (%) WRF- WRF- Description [8] NAM [8] NAM CLDDA CLDDA Overcast Cloudy Mostly Clear VI. SUMMARY AND CONCLUSIONS Previous studies have shown that the operational NWP models (e.g. the NAM) are inaccurate for solar irradiance forecasting, typically over-predicting GHI and underpredicting cloud cover. This error is seasonal and location dependent. Specifically, these errors are exacerbated for summer-time coastal California, when marine layer fog and stratus conditions are common. The source of this error can be traced to three primary sources: Domain discretization size, insufficient physics parameterizations, and poor model initializations. To address these sources of error, a high-resolution, cloudassimilating NWP based on the Weather and Research Forecasting (WRF) model was implemented at the University of California, San Diego for solar irradiance forecasting. Fine horizontal (Δx = 1.33 km) and vertical resolutions were used to promote low-altitude cloud formation. Cloud microphysics, solved by a 2-class scheme in the NAM, was parameterized by the more complex 6-class Thompson model. Additionally, a direct-cloud assimilation system was employed to populate the model initial conditions with cloud cover. Using this system, intra-day irradiance forecasts were created for 5/1/2011 and 6/30/2011 and validated against the UCSD pyranometer network. Overall, WRF-CLDDA intra-day forecasts had rmbe of 2% compared to 18.1% for the NAM. Furthermore, rmae was 21.6%, 5.2% smaller than the NAM. rstderr, a measure of error randomness, was only 1.1% smaller than the NAM as a large portion of NAM error was attributed to systematic bias. For day-ahead forecasts ( > 24 hours), WRF- CLDDA accuracy did not degrade; rmae was 19.3% and rstderr 31%. NAM accuracy, however, deteriorated significantly as rstderr increased by 5.5% to 39%. In general, NWP rmae and rrmse increase by 1%-3% between intra-day and day-ahead forecasts [6]. Here, however, it was found that NAM forecast accuracy deteriorated at a faster rate than average while WRF-CLDDA accuracy was maintained. Additionally, WRF-CLDDA accuracy was established as a function of sky condition. Overall, clear conditions were most accurate with rrmse of 30.3%. However, the largest improvements over the NAM and previous studies were for overcast conditions, indicating that WRF-CLDDA is well suited for solar irradiance forecasting especially for regions with frequently cloudy conditions. VII. ACKNOWLEDGMENTS GOES satellite products were provided by the Cooperative Institute of Research for the Atmosphere (CIRA) at Colorado State University under the direction of Matt Rogers and Steve Miller. Helpful conversations with Stan Benjamin, John Brown, Curtis Alexander and others at the Earth Systems Research Laboratory (NOAA) in Boulder, CO. Patrick Shaw and Daran Rife of GL-Garrad Hassan provided assistance and guidance for WRF modelling. VII. REFERENCES [1] Perez, R., Kivalov, S., Schlemmer, J., Hemker, K., Renné, D., and Hoff, T.E., Validation of short and medium term operational solar radiation forecasts in the US. Solar Energy. 84(12), [2] Chow, C.W., Urquhart, B., Lave, M., Dominguez, A., Kleissl, J., Shields, J., and Washom, B., Intra-hour forecasting with a total sky imager at the UC San Diego solar energy testbed. Solar Energy. 85, [3] Remund, J., Perez, R., and Lorenz, E., Comparison of solar radiation forecasts for the USA. In: 2008 European PV Conference, Valencia, Spain. [4] Lorenz, E., Hurka, J., Heinemann, D., and Beyer, H., Irradiance forecasting for the power prediction of grid-connected photovoltaic systems. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 2(1), [5] Mathiesen, P., and Kleissl, J., Evaluation of numerical weather prediction for intra-day solar forecasting in the continental United States. Solar Energy. 85(5), [6] Perez., R., Beauharnois, M., Hemker, K., Kivalov, S., Lorenz, E., Pelland, S., Schlemmer, J., and Van Knowe, G., Evaluation of numerical weather prediction solar irradiance forecasts in the US. Proceedings of Solar 2011, American Solar Energy Society. Raleigh, NC. [7] Pelland, S., Galanis, G., and Kallos, G., Solar and photovoltaic forecasting through post-processing of the Global Environmental Multiscale Numerical Weather Prediction Model. Progress in Photovoltaics: Research and Applications. [8] Lara-Fanego, V., Ruiz-Aria, J.A., Pozo-Vázquez, D., Santos-Alamillos, 707

6 6 F.J., and Tovar-Pescador, J., Evaluation of the WRF model solar irradiance forecasts in Andalusia (southern Spain). doi: /j.solener [9] Mathiesen, P., Brown, J., and Kleissl, J., Geostrophic wind dependent probabilistic irradiance forecasts for coastal California. IEEE Transactions on Sustainable Energy. 99, 1-9. [10] Tselioudis, G., and Jakob, C., Evaluation of midlatitude cloud properties in a weather and climate model: Dependence on dynamic regime and spatial resolution. Journal of Geophysical Research. 107(D24), 4781, doi: /2002jd [11] Lin, W., Zhang, M., and Wu, J., Simulation of low clouds from the CAM and the regional WRF with multiple nested resolutions. Geophysical Research Letters. 36, L08813, doi: /2008/gl [12] Otkin, J., and Greenwald, T., Comparison of WRF modelsimulation and MODIS-derived cloud data. Monthly Weather Review. 136, [13] Benjamin, S., Dévényi, D., Weygandt, S., Brunadge, K., Brown, J., Grell, G., Kim, D., Schwartz, B., Smirnova, T., and Smith, T., 2004a. An hourly assimilation-forecast cycle: The RUC. Monthly Weather Review. 132, [14] Benjamin, S., Kim, D., and Brown, J., Cloud/hydrometeor initialization in the 20-km RUC using GOES and radar data. Proceedings of the 10 th Conference on Aviation, Range, and Aerospace Meteorology. American Meteorological Society, Portland, OR, USA. [15] Benjamin, S., Weygandt, S., Brown, J., Smith, T., Smirnova, T., Moniger, W., and Schwartz, B., 2004b. Assimilation of METAR cloud and visibility observations in the RUC. Proceedings of the 11 th Conference on Aviation, Range, and Aerospace and the 22 nd Conference on Severe Local Storms. American Meteorological Society, Hyannis, MA, USA. [16] Weygant, S., Benjamin, S., Dévényi, D., Brown, J., and Minnis, P., Cloud and hydrometeor analysis using METAR, radar, and satellite data within the RUC/Rapid-Refresh model. Proceedings of the 12 th Conference on Aviation, Range, and Aerospace. American Metrological Society, Atlanta, GA, USA. [17] Hu, M., Weygandt, S., Xue, M., and Benjamin, S., Development and testing of a new cloud analysis package using radar, satellite, and surface cloud observation within GSI for initializing rapid refresh. Proceedings of the 18 th Conference on Numerical Weather Prediction. American Meteorological Society, Park City, UT, USA [18] Albers, S., McGinley, J., Birkenheuer, D., and Smart, J., The local analysis and prediction system (LAPS): Analyses of clouds, precipitation and pressure. Weather and Forecasting. 11(3), [19] Xue, M., Wang, D., Gao, J., Brewster, K., and Droegenmeier, K., The advanced regional prediction system (ARPS), storm-scale numerical weather prediction and data assimilation. Meteorology and Atmospheric Physics. 82, [20] Rogers, E., DiMego, G., Black, T., Ek, M., Ferrier, B., Gayno, G., Janjic, Z., Lin, Y., Pyle, M., Wong, V., and Wu, W.S., The NCEP North American mesoscale modeling system: Recent changes and future plans. Preprints, 23 rd Conference on Weather Analysis and Forecasting and the 19 th Conference on Numerical Weather Prediction. Omaha, NE. American Meteorological Society, 2A.4. [21] Wu, W.S., Purser, R.J., and Parrish, D., Three dimensional variational analysis with spatially inhomogeneous covariances. Monthly Weather Review. 130, [22] Rogers, E., Black, T., Ferrier, B., Lin, Y., Parrish, D., and DiMega, G., Changes to the NCEP meso eta analysis and forecat system: Increase in resolution, new cloud microphysics, modified precipitation assimilation, modified 3DVAR analysis. Accessed 9/6/12. [23] Lacis, A., and Hansen, J., A parameterization for the absorption of solar radiation in the Earth s atmosphere. Journal of Atmospheric Sciences. 31(1), [24] Janjic, Z., Gall, R., and Pyle, M.E., Scientific documentation for the NMM solver. NCAR Technical Note NCAR/TN 477+STR. [25] Janjic, Z., Black, T., Pyle, M., Ferrier, B., Chuang, H.Y., Jovic, D., McKee, N., Rozulmalski, R., Michalakes, J., Gill, D., Dudhia, J., Duda, M., Demirtas, M., Nance, L., Slovacek, T., Wolff, J., Bernadet, L., McCaslin, P., Stoelinga, M., User s guide for the NMM core of the Weather Research and Forecast (WRF) modeling system version pgs. [26] Skamarock, W., Klemp, J., Dudhia, J., Gill, D., Barker, D., Duda, M., Huang, X., Wang, W., and Powers, J., A description of the advanced research WRF version 3. NCAR Technical Note NCAR/TN 475+STR. [27] Thompson, G., Rasmussen, R., and Manning, K., Explicit forecasts of winter precipitation using an improved bulk microphysics scheme. Part I: Description and sensitivity. Monthly Weather Review. 132, [28] Nakanishi, M., and Niino, H., An improved Mellor-Yamada level- 3 model: Its numerical stability and application to a regional prediction of advection fog. Boundary Layer Meteorology. 119, [29] Mlawer, E., and Clough, S., Shortwave and longwave enhancements in the rapid radiative transfer model. Proceedings of the 7 th Atmospheric Radiation Measurement (ARM) Science Team Meeting. US Department of Energy, CONF [30] Mlawer, E., Taubman, S., Brown, P., Iacono, M., and Clough, S., Radiative transfer for inhomogenous atmospheres: RRTM, a validated correlated-k model for the longwave. Journal of Geophysical Research. 102(14), [31] Sengupta, M., Heidinger, A., and Miller, S., Validating an operational physical method to compute surface radiation from geostationary satellites. Proceedings of the SPIE Conference. San Diego, CA. [32] Campbell Scientific, LI200S Pyranometer instruction manual. Campbell Scientific Technical Specifications, Revision 2/96. [33] Novakovskaia, E., Sloop, C., and Guo, Z., The impact of real-time observations on the WRF model surface forecasts. 11 th EMS Annual Meeting, 8, EMS [34] Mathiesen, P., Collier, C., and Kleissl, J., A high-resolution, cloud-assimilating numerical weather prediction model for solar irradiance forecasting. Solar Energy. Submitted 8/