1. Introduction
There are several types of aerosols in Earth’s atmosphere, which come from both natural and anthropogenic sources, including dust, sea salt, black carbon, organic carbon, sulfates, nitrates, and volcanic ash. Each of these types of aerosols exist in time-varying concentrations and size distributions in the atmosphere, and interact both with moisture and radiation. Accurately forecasting the three-dimensional movement and concentration of these aerosols is a challenging problem.
The amount of scattering and absorption of incoming solar radiation caused by aerosols can be summarized by the total aerosol optical depth (AOD). In general, higher AOD values correspond to reductions in solar irradiance at the surface. This is particularly true for direct normal irradiance (DNI), because any scattering or absorption by definition reduces DNI at the surface; this is the aerosol direct effect. By contrast, global horizontal irradiance (GHI), which is calculated from a combination of direct and diffuse irradiance, is less sensitive to changes in AOD. In clear-sky (i.e., cloud free) conditions, aerosols have an impact on DNI that is 3–4 times as large as that on GHI (Gueymard 2012; Ruiz-Arias et al. 2013), partly due to offsetting errors in the estimation of direct and diffuse irradiance. Accurate forecasts of AOD will reduce forecast error in both DNI and GHI during clear-sky conditions, which will in turn reduce forecast errors in power generation at concentrated solar power (CSP) plants, which convert DNI into power, and to a lesser degree for photovoltaic (PV) solar plants, which convert GHI into power.
Typically, atmospheric chemistry transport models, such as the WRF Model coupled with chemistry (WRF-Chem) (Grell et al. 2005; Fast et al. 2006), NOAA’s Community Multiscale Air Quality (CMAQ) model (Byun and Schere 2006; P. Lee et al. 2017), or various global models that assimilate satellite-based observations of aerosols, such as NASA’s GEOS-5 model (Rienecker et al. 2008; Molod et al. 2015; Buchard et al. 2015) or the Copernicus Atmosphere Monitoring Service (CAMS) model (Copernicus 2020; Engelen 2019; Rémy et al. 2019), are used to model the emission, transport, and deposition of aerosols with the greatest accuracy. Some of these models have been extensively validated for AOD recently, such as CAMS alongside NASA’s MERRA-2 (Gelaro et al. 2017; Randles et al. 2017; Buchard et al. 2017) reanalysis (Gueymard and Yang 2020; Fu et al. 2022). Unfortunately, adding numerous 3D variables to track several aerosol species causes these models to be often too computationally expensive, slow, or coarse for the needs of other applications, such as operational forecasting for solar energy generation. Thus, various approximations and assumptions about the aerosols must be made to reduce computational cost and speed up model run time.
In WRF-Solar (Jiménez et al. 2016a,b), an NWP model specifically tailored for solar forecasting and resource assessment applications (e.g., Haupt et al. 2018, 2020; J. A. Lee et al. 2017; Verbois et al. 2018; Dasari et al. 2019) and that is currently under active development (e.g., Yang et al. 2021; McCandless and Jiménez 2020; Juliano et al. 2022), a simple parameterization of aerosol optical properties is employed in conjunction with the RRTMG shortwave radiation scheme (Iacono et al. 2008). In this parameterization, the two-dimensional AOD at 550 nm (AOD550) field is specified with time stamps throughout the model simulation, with assumptions made about the Ångström exponent (AE), aerosol single-scattering albedo (SSA), and asymmetry parameter (ASY), which differ according to the dominant type of aerosol known or assumed to be present (Ruiz-Arias et al. 2014). The simplest approach to provide this AOD550 field is to use a prescribed aerosol climatology of some sort, where the AOD is held fixed throughout a single simulation and only varies on a monthly basis. As such, when using an aerosol climatology, there is no flow dependency or ability to capture the effects of transient features such as dust storms or wildfire smoke plumes on forecasted DNI and GHI. Alternatively, AOD550 can be prescribed from a forecast model such as GEOS-5 or CAMS, which would provide flow-dependent total aerosol loading at (typically) 1-hourly intervals, and thus hopefully better capture the actual AOD than is possible with static climatologies.
Prior research has also demonstrated that for the WRF-Solar aerosol-radiation parameterization analyzed here, the main source of irradiance error in clear-sky conditions is not the radiation parameterization but rather the AOD (Ruiz-Arias et al. 2014; Jiménez et al. 2016a). In particular, Ruiz-Arias et al. (2014) found that the errors in irradiance predictions are within the observational error when the aerosol distribution is known. Thus, quantifying the impact of AOD representation in WRF-Solar with readily available options is the focus here.
The goal of this study is to assess different representations of aerosols in WRF-Solar, from two climatologies and two global forecast models, for their impact on predictions of DNI and GHI during clear-sky conditions at high-quality observing sites across CONUS. Additionally, these WRF-Solar configurations are compared with clear-sky DNI and GHI forecasts from the CAMS model, to provide an independent benchmark. This validation, performed over more than 8 months of daily forecasts, can inform solar forecasting practitioners of the relative value of different aerosol representations. The originality of this research relies in providing a statistically robust characterization of the advantages of using aerosol forecasts from two prominent global chemistry models with respect to the more standard use of aerosol climatologies in regional NWP models.
The paper is organized as follows. Section 2 describes the observations, model configurations, and validation metrics used in this study. Section 3 is a brief validation of the AOD550 fields used in the WRF-Solar simulations. Section 4 presents the results of the instantaneous clear-sky irradiance analysis, comparing the four WRF-Solar configurations. Section 5 presents the results of the hourly average clear-sky irradiance analysis, comparing the four WRF-Solar configurations and CAMS. Section 6 summarizes and concludes the paper.
2. Methods
a. Observations
To validate the model forecasts described below, we use high-quality irradiance instruments at the SURFRAD (Augustine et al. 2000, 2005) and the NOAA Solar Radiation (SOLRAD) (Hicks et al. 1996) networks. SURFRAD and SOLRAD have seven sites each, for a total of 14 sites across CONUS. To exclude the influence of clouds on the analysis, because clouds cause much larger reductions in GHI and DNI than aerosols do, we only use GHI and DNI observations when the clear-sky flag is turned on in the Radiative Flux (“RadFlux”) Analysis (Long and Shi 2008) processed data. SURFRAD and SOLRAD both report instantaneous observations every 1 min, and to compare with instantaneous WRF-Solar forecasts as outlined below, we use instantaneous observations every 15 min. Any observations with a solar zenith angle >80° were excluded from analysis.
Radiative transfer models treat the sun as a point source, while pyrheliometers measure the DNI with a half-angle aperture of about 2.5° (Blanc et al. 2014). The pyrheliometer therefore measures a part of incoming radiation in the vicinity of the sun that is considered as diffuse radiation by radiation transfer schemes in NWP models. This inconsistency leads to a mismatch between the model output and the measurement. The order of magnitude of this difference differs according to the study. For instance, one study evaluated this effect to account for an error of about 10 W m−2 in most clear-sky conditions (Oumbe et al. 2012), while another study reported that it could create an inconsistency of as much as 300 W m−2 under the presence of thin cirrus clouds (Qin et al. 2021). The error is therefore much smaller in clear skies, which are the focus of this paper. In the meantime, numerous validation papers evaluated the accuracy of model output without considering this aspect (e.g., Ruiz-Arias and Gueymard 2018; Marchand et al. 2020). This is the standard approach and is the one that we use in this paper.
b. WRF-Solar configurations
WRF-Solar (Jiménez et al. 2016a,b; Haupt et al. 2018) is a set of enhancements to the WRF Model (Powers et al. 2017; Skamarock et al. 2019) that is specifically designed to improve solar irradiance and solar power forecasting. These enhancements include feedbacks between clouds, aerosols, and radiation. The community version of WRF-Solar is in the public domain and is included with the WRF distribution on GitHub beginning with WRF, version 4.2. For this study we use WRF-Solar as implemented in WRF v4.2.1.
All WRF-Solar simulations here use an identical 600 × 354 domain that covers CONUS with Δx = 9 km and 45 vertical levels with a 50-hPa model top. All simulations are initialized from GFS 0.25° forecast data at 0900 UTC (0500 eastern daylight time/0400 eastern standard time) daily from 19 November 2019 to 30 July 2020, and extend to a lead time of 45 h to span two complete daylight periods across the entire CONUS. The model time step is 45 s, with gridded output every 15 min. All simulations use the RRTMG shortwave and longwave radiation schemes (Iacono et al. 2008), and they are called every model time step. All simulations also use the Grell–Freitas cumulus scheme (Grell and Freitas 2014), MYNN level-2.5 atmospheric boundary layer scheme (Nakanishi and Niino 2006), Jiménez (revised MM5) surface layer scheme (Jiménez et al. 2012), and the Noah land surface scheme (Ek 2003). The Thompson microphysics scheme (Thompson et al. 2008) is also used, except for the fourth experiment as noted below, which uses the Thompson–Eidhammer aerosol-aware microphysics scheme (Thompson and Eidhammer 2014).
The four WRF-Solar experiments in this study differ only in their source of aerosols. Table 1 concisely summarizes the four experiments, the aerosol dataset used by each experiment, and the aerosol inputs that are used by and/or imposed onto WRF-Solar. Each aerosol source serves as an input to the RRTMG radiation scheme, where aerosol–radiation interactions are modeled. For comparison with the SURFRAD and SOLRAD GHI and DNI during clear skies only, we use the clear-sky GHI (SWDOWNC) and clear-sky DNI (SWDDNIC) variables from WRF-Solar. These variables account for the effect of aerosols in the atmospheric column scattering and absorbing radiation, and are identical to what the GHI (SWDOWN) and DNI (SWDDNI) are in the absence of any cloud cover in the model.
Summary of the four WRF-Solar experiments in this study, their aerosol datasets, and what is used by and/or imposed onto WRF-Solar in each dataset.
The first experiment, “WRF-Solar+Climo Aero,” uses the default Tegen climatology (aer_opt=1) (Tegen et al. 1997). This Tegen climatology has six aerosol types (organic carbon, black carbon, sulfate, sea salt, dust, and stratospheric aerosol) on a relatively coarse 5° longitude by 4° latitude global grid with monthly variations. WRF-Solar calculates AOD from this 3D climatology of aerosol properties. This experiment is considered the baseline, as it is the simplest treatment of aerosols while still including aerosols in WRF-Solar.
The second experiment, “WRF-Solar+GEOS-5 Aero,” imposes the AOD550 field every 1 h from NASA GEOS-5 model forecasts issued at 0000 UTC daily. GEOS-5 is a global Earth system model that is used for both weather and climate applications, and has several components, including for atmospheric chemistry and aerosols (Rienecker et al. 2008; Molod et al. 2015). GEOS-5 model output is on a 0.3125° longitude × 0.25° latitude grid at hourly temporal resolution out to 10 days (the 1200 UTC cycle extends to 5 days, and the 0600 and 1800 UTC cycles extend only to 30-h lead time). GEOS-5 assimilates AOD observations from the MODIS instruments on the Aqua and Terra satellites, and also uses ground-based measurements from AERONET (Holben et al. 1998) sites for bias correction and calibration (Buchard et al. 2015).
The third experiment, “WRF-Solar+CAMS Aero,” imposes the AOD550 field every 1 h from the CAMS model forecasts (Copernicus 2020). CAMS is a global 0.4° × 0.4° analysis and forecast system for aerosols and atmospheric chemistry (Engelen 2019; Rémy et al. 2019), and it is driven by the ECMWF Integrated Forecast System (IFS) atmospheric model (ECMWF 2019). CAMS forecasts initialized at 0000 UTC are used here. CAMS output is available in real-time at 1-h frequency out to 5 days for both the 0000 and 1200 UTC cycles.
The fourth experiment, “WRF-Solar+TE Microphys,” uses the monthly water-friendly and ice-friendly aerosol climatology (aer_opt=3) from the Thompson–Eidhammer aerosol-aware microphysics scheme; this climatology is ultimately derived from seven years of GEOS-4 analyses (Thompson and Eidhammer 2014). As with the Tegen climatology, the Thompson–Eidhammer aerosol climatology varies monthly, but it has a much finer grid spacing of 0.5° longitude by 1.25° latitude. We also note that in WRF-Solar+TE Microphys, while the aerosol distribution at model initialization is provided by the monthly climatology, the aerosols are advected by the model, and so AOD550 is calculated internally to WRF at each radiation time step. In the other three WRF-Solar experiments above, aerosols are not advected by the model and remain static either through the entire simulation (as in WRF-Solar+Climo Aero) or between AOD550 imposition times (as in WRF-Solar+GOES-5 and WRF-Solar+CAMS).
c. CAMS description
For hourly average irradiance assessment (section 5), we also bring in the CAMS model. CAMS, which is operated by the ECMWF on behalf of the European Commission, is dedicated to the monitoring and forecasting of aerosols. This mission is ensured by the continuous development and operation of an atmospheric composition component model called IFS-AER (Rémy et al. 2019).
The IFS is an NWP system dedicated to operational meteorological forecasts. It was extended in the Monitoring Atmospheric Composition and Climate (MACC) and CAMS projects to forecast and assimilate aerosols (Morcrette et al. 2009; Benedetti et al. 2009), greenhouse gases (Engelen et al. 2009; Agustí-Panareda et al. 2014), and reactive trace gases (Flemming et al. 2015; Huijnen et al. 2019). As a result, IFS-AER denotes the IFS extended with the bin-bulk aerosol scheme used to provide aerosol products in the CAMS project (Rémy et al. 2019). The model includes different parameterizations for aerosol sources, sinks, and its chemical production. Satellite products as well as numerous further measurements are assimilated to the model to forecast aerosols, greenhouse gases, and reactive trace gases.
CAMS is delivering operational near-real-time (NRT) forecasts as well as reanalyses of global atmospheric composition [e.g., CAMS Interim Reanalysis (CAMSiRA) (Flemming et al. 2017), CAMS Reanalysis (CAMSRA) (Inness et al. 2019)]. In this study the 0000 UTC operational forecasts for the surface shortwave radiation clear-sky (ssrdc), the clear-sky direct solar radiation at the surface (cdir), and the aerosol optical depth at 550 nm (aod550) are used.
d. Validation metrics
To validate the clear-sky irradiance forecasts from the four WRF-Solar experiments against observed clear-sky irradiance observations from SURFRAD and SOLRAD stations, we use standard metrics, namely, the mean absolute error (MAE) and mean bias error (MBE; forecast minus observation). When validating the AOD550 datasets against AERONET observations, we also use root-mean-square error (RMSE).
3. Validation of AOD550 datasets in WRF-Solar
Before validating the clear-sky irradiance, it is useful to evaluate briefly the AOD550 forecasts and climatologies that are used in the four WRF-Solar experiments. For a bulk overall validation, Table 2 presents the AOD550 mean, median, and maximum values for the observations and all four WRF-Solar experiments, and Fig. 1 depicts the distributions as boxplots. From these overall statistics, WRF-Solar+Climo Aero has the smallest MBE and RMSE, and is tied with WRF-Solar+GEOS-5 Aero for the smallest MAE. However, the maximum AOD550 in WRF-Solar+Climo Aero is only 0.17, so it entirely misses all the high-AOD outliers. The overall distribution of AOD550 in WRF-Solar+GEOS-5 Aero is most similar to the observations, while WRF-Solar+CAMS Aero has the highest, mean, median, RMSE, MAE, and MBE for AOD550, as well as the largest interquartile range. These findings, with GEOS-5 forecast AOD550 generally having lower MAE and smaller MBE than CAMS forecast AOD550, are consistent with Lee et al. (2021). In that study, they validated two years of AOD550 forecasts from three operational models against observations from 20 AERONET sites in the Middle East, finding the best performance was from GEOS-5, with CAMS generally having larger errors.
Boxplots of AOD550 at AERONET stations, for observations and WRF-Solar experiments at all 15-min valid times for 45-h simulations initialized at 0900 UTC daily from 19 Nov 2019 to 30 Jul 2020. Only matched pairs for which both the observation and all WRF-Solar forecasts are valid are included.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
Mean, median, and maximum AOD550, RMSE, MAE, and MBE values at AERONET stations, for observations and WRF-Solar experiments at all 15-min valid times for 45-h simulations initialized at 0900 UTC daily from 19 Nov 2019 to 30 Jul 2020. Only matched pairs for which both the observations and all WRF-Solar forecasts are valid are included in these statistics.
Additionally, when looking at the geographic distribution of AOD550 MAE (Fig. 2) and MBE (Fig. 3) at AERONET stations, the patterns compare well overall to the clear-sky DNI MAE and MBE that will be discussed in section 4a below (see Figs. 4 and 5). While most AERONET stations are not collocated with SURFRAD or SOLRAD stations, it is apparent that regions with higher AOD550 MAE tend to have higher clear-sky DNI MAE, and regions with overestimations of AOD550 (positive MBE) have underestimations of clear-sky DNI (negative MBE). This pattern is most pronounced for WRF-Solar+CAMS Aero, which saw positive MBE for nearly every AERONET station in the domain, with the largest magnitudes in the western United States, which correlated very well with negative clear-sky DNI MBE across CONUS, with the largest magnitudes also in the western United States. From this analysis, we are confident that the different AOD550 fields are responsible for most of the difference in clear-sky DNI in WRF-Solar across the four experiments, as expected.
AOD550 MAE at AERONET sites (filled circles) for all cycles (0900 UTC daily from 19 Nov 2019 to 30 Jul 2020) and lead times (15 min–45 h) for four WRF-Solar experiments: (a) WRF-Solar+Climo Aero, (b) WRF-Solar+GEOS-5 Aero, (c) WRF-Solar+CAMS Aero, and (d) WRF-Solar+TE Microphys.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
As in Fig. 2, but for AOD550 MBE.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
Clear-sky DNI MAE at SURFRAD (circles) and SOLRAD (squares) stations, for all lead times (0–45 h), and start dates 19 Nov 2019–30 Jul 2020 at 0900 UTC daily, for four WRF-Solar experiments: (a) WRF-Solar+Climo Aero, (b) WRF-Solar+GEOS-5 Aero, (c) WRF-Solar+CAMS Aero, and (d) WRF-Solar+TE Microphys.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
As in Fig. 4, but for MBE.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
4. Results for instantaneous clear-sky irradiance
a. Results by station
First, we examine the clear-sky DNI MAE (Fig. 4) and MBE (Fig. 5) as a function of station, combining all initialization and lead times. The SURFRAD and SOLRAD stations are indicated by filled circles and squares, respectively, colored according to the data value. As with Figs. 2 and 3 above for the AOD550 validation, Fig. 4a is for WRF-Solar+Climo Aero, Fig. 4b is for WRF-Solar+GEOS-5 Aero, Fig. 4c is for WRF-Solar+CAMS Aero, and Fig. 4d is for WRF-Solar+TE Microphys. For most of these sites, the clear-sky DNI MAE is lowest in WRF-Solar+GEOS-5 Aero nationwide (Fig. 4b), with the largest reductions in MAE relative to the baseline WRF-Solar+Climo Aero (Fig. 4a) occurring in the eastern United States. For most sites, clear-sky DNI MAE is highest for WRF Solar+CAMS Aero (Fig. 4c), with a particularly high MAE for the Hanford, California (HNX), SOLRAD station. WRF-Solar+TE Microphys (Fig. 4d) differs little from WRF-Solar+CAMS Aero. These geographic patterns correlate well with the AOD550 MAE in Fig. 2.
The clear-sky DNI MBE (Fig. 5) is smallest in magnitude for most sites in WRF-Solar+GEOS-5 Aero (Fig. 5b), though AOD550 appears to be still too high in the northern and eastern United States (with negative DNI biases) and too low in the southwestern United States (with positive clear-sky DNI biases). For sites in the western United States, the clear-sky DNI MBE is most negative for WRF-Solar+CAMS Aero (Fig. 5c), with a particularly large negative MBE for HNX. The combination of high MAE and a large-magnitude negative MBE for clear-sky DNI at HNX indicates that the likely cause is from CAMS AOD550 being far too high relative to observed AOD550. In the eastern United States, however, WRF-Solar+CAMS Aero has the smallest-magnitude DNI MBE values of all four experiments, indicating that the CAMS AOD550 is more accurate in the eastern United States than the western United States. This finding highlights an important point, that while one model may perform better than another model on average over a large continental (or global) region, there are smaller regions where that relationship may not be true—model skill can depend on the region of interest. As with the clear-sky DNI MAE, the clear-sky DNI MBE geographic patterns correlate well with AOD550 MBE in Fig. 3. This result is consistent with previous findings that indicate that the performance of the WRF-Solar clear-sky DNI is mainly controlled by the quality of the AOD at 550 nm (Ruiz-Arias et al. 2014; Jiménez et al. 2016a).
Also, from both Figs. 4 and 5, we can see that the Salt Lake City, Utah (SLC), SOLRAD station has consistently high clear-sky DNI MAE and MBE in comparison with other stations in the western United States, which stems from an ∼10% lower observed clear-sky DNI on average at SLC relative to other stations (not shown). The potential causes for this include either a systematic instrument error or a systematic missing aerosol source in all aerosol models and climatologies, but delving further into investigating this matter is beyond the scope of this study. We note, however, that for the AERONET site near Salt Lake City, the AOD550 errors do not stand out from surrounding stations (Figs. 2 and 3), pointing to instrument measurement and/or calibration error as a more likely explanation for the large clear-sky DNI errors at SLC during this time period.
For clear-sky GHI MAE (Fig. 6), there is little difference between the four experiments, and all stations except SLC have a small clear-sky GHI MAE. This finding confirms what was previously known, that aerosol loading has much smaller impact on GHI than it does on DNI. There is a little more variability across the experiments for clear-sky GHI MBE (Fig. 7), but again, MBE values are generally small in magnitude at most stations, except for the SLC, HNX, and Seattle, Washington (SEA), SOLRAD sites, which show noticeably larger clear-sky GHI MBE values than any other sites nationwide. Additionally, WRF-Solar+CAMS Aero appears to have the smallest clear-sky GHI MBE values in the southwestern United States, which has relevance for PV energy forecasting, as that region has the best solar resource in the nation. Attribution of the cause of these clear-sky GHI MBE spatial patterns is more complex and less directly attributable to the aerosol source than for clear-sky DNI; as GHI includes contributions both from the direct and diffuse components of irradiance, there are compensating effects from aerosols in GHI. In any case, for most of the United States, the choice of aerosol source has little impact on clear-sky GHI skill in WRF-Solar on average. There are specific cases with large AOD, such as heavy smoke from wildfires, when aerosol forecasts with time stamps should provide superior performance over climatologies. However, these events are relatively rare, and so would have only a small impact in clear-sky GHI forecast skill over long validation periods.
Clear-sky GHI MAE at SURFRAD (circles) and SOLRAD (squares) stations, for all lead times (0–45 h), and start dates 19 Nov 2019–30 Jul 2020 at 0900 UTC daily, for four experiments: (a) WRF-Solar+Climo Aero, (b) WRF-Solar+GEOS-5 Aero, (c) WRF-Solar+CAMS Aero, and (d) WRF-Solar+TE Microphys.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
As in Fig. 6, but for MBE.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
b. Results by lead time
Second, we examine results as a function of lead time, combining all initialization times and all stations. Time series of clear-sky DNI MAE and MBE are displayed in Figs. 8a and 8b, respectively. For most lead times, consistent with the maps presented above, WRF-Solar+GEOS-5 Aero (cyan squares) has the smallest clear-sky DNI MAE, with WRF-Solar+TE Microphys (yellow diamonds) typically in second place, especially outside of the early morning in the eastern United States. Because the Thompson–Eidhammer microphysics aerosol climatology is ultimately derived from GEOS-5 analyses, this should not be too surprising. WRF-Solar+CAMS Aero (green triangles) has the highest (worst) clear-sky DNI MAE at most lead times. For clear-sky DNI MBE, the same ordering generally holds true, with WRF-Solar+CAMS Aero being clearly the worst performer, with strongly negative clear-sky DNI MBE throughout both day 1 and day 2. This finding is consistent with Fig. 5c, which showed WRF-Solar+CAMS Aero clear-sky DNI MBE being strongly negative at most stations.
Clear-sky DNI (a) MAE and (b) MBE as a function of lead time for WRF-Solar simulations initialized from 19 Nov 2019 to 30 Jul 2020. The four experiments are WRF-Solar+Climo Aero (blue dots), WRF-Solar+GEOS-5 Aero (cyan squares), WRF-Solar+CAMS Aero (green triangles), and WRF-Solar+TE Microphys (yellow diamonds).
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
Conversely, for clear-sky GHI MAE (Fig. 9a), there is not much difference between the four experiments through all lead times, though WRF-Solar+GEOS-5 Aero seems to perform somewhat worse on average by this metric. After the first few hours of each day, all four experiments display a positive clear-sky GHI MBE (Fig. 9b), with WRF-Solar+GEOS-5 Aero generally having the largest (worst) bias, and WRF-Solar+CAMS Aero generally having the smallest (best) bias. Again, these findings are consistent with the maps of errors by site shown in Figs. 6 and 7.
As in Fig. 8, but for GHI.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
We expected to find that imposing prognostic AOD550 onto WRF-Solar, which in principle should better characterize day-to-day variability in aerosol loading, would lead to improved DNI forecasts in particular, relative to using climatological AOD550. That expectation only appears to be the case for GEOS-5 AOD550, and not for CAMS AOD550, which often had the highest clear-sky DNI errors through most lead times on day 1 and day 2. Additionally, the irradiance errors are largely similar through both day-1 and day-2 lead times, indicating that any additional advantage of prognostic AOD550 over climatological AOD550 on day 2 relative to day 1 is small, at best.
c. Results by month
Third, we examine results as a function of simulation month, to evaluate whether there are seasonal trends that emerge in the data. In these plots, a single value for each month is calculated by aggregating all initialization times, lead times, and stations.
Consistent with results discussed above, WRF-Solar+GEOS-5 Aero has the lowest (or nearly the lowest) clear-sky DNI MAE across all months (Fig. 10a), with WRF-Solar+TE Microphys performing very similarly except during March–May 2020. A similar pattern is observed for the clear-sky DNI MBE (Fig. 10b), with WRF-Solar+GEOS-5 Aero generally having the smallest-magnitude bias, and WRF-Solar+TE Microphys performing similarly except for a large negative MBE in spring, likely indicating that AOD550 was too high. Interestingly, this raises the question of whether the shutdowns and reductions in emissions across the United States in spring 2020 caused by the COVID-19 pandemic can explain this performance difference, as the GEOS-5 aerosols (which are constrained by data assimilation) lead to more skillful clear-sky DNI in WRF-Solar than an aerosol climatology based off analyses from prior years (2001–07) of the same model. Another potential explanation is that the Thompson–Eidhammer aerosol climatology simply has too much aerosol in the spring over the United States. To identify whether either of these potential explanations is correct, simulations through additional spring seasons would be required. For both clear-sky DNI MAE and MBE, WRF-Solar+CAMS Aero was generally the worst-performing experiment for most months, with WRF-Solar+Climo Aero usually the third-best model through most months.
Clear-sky DNI (a) MAE and (b) MBE for all forecast cycles initialized in every month, aggregated over all sites. The four experiments are WRF-Solar+Climo Aero (blue dots), WRF-Solar+GEOS-5 Aero (cyan squares), WRF-Solar+CAMS Aero (green triangles), and WRF-Solar+TE Microphys (yellow diamonds).
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
For the clear-sky GHI MAE (Fig. 11a), it is seen that WRF-Solar+GEOS-5 Aero and WRF-Solar+TE Microphys have the smallest errors through the winter, but then have the largest errors from late spring into the summer months. Even so, the difference between all four experiments is not large. The differences between the models are slightly larger for clear-sky GHI MBE (Fig. 11b), with WRF-Solar+GEOS-5 Aero and WRF-Solar+TE Microphys having the smallest-magnitude biases until spring and summer, when they have the highest (worst) bias. Because WRF-Solar+GEOS-5 Aero generally had the highest clear-sky GHI MAE and MBE when looking at all model cycles together (Fig. 9), rather than month by month, we can conclude that the higher errors are dominated by errors in the summer months. We speculate that this could perhaps be caused by the higher frequency of clear-sky conditions across much of the United States from late spring into summer.
As in Fig. 10, but for GHI.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
d. Overall results
Fourth, we examine overall results, where we rank the four experiments by aggregating all initialization times, all lead times, and all stations. For clear-sky DNI (Figs. 12a,b), a clear ranking of the four experiments emerges, with WRF-Solar+GEOS-5 Aero performing best with the lowest MAE and lowest-magnitude MBE, WRF-Solar+TE Microphys performing second best in both scores, WRF-Solar+Climo Aero performing third best, and WRF-Solar+CAMS Aero having the worst overall MAE and MBE for clear-sky DNI. These overall rankings are consistent with the clear-sky DNI results previously discussed, and also with the AOD550 validation in Table 1, but with a stronger penalty in clear-sky DNI errors for WRF-Solar+Climo Aero due to the complete lack of high AOD values in that climatology. Thus, if DNI is the variable of importance (say, for CSP energy forecasting), then the choice of aerosol model or climatology with WRF-Solar makes a substantial difference during clear-sky conditions.
Clear-sky DNI (a) MAE and (b) MBE over all sites, all lead times, and all simulations. The four experiments are WRF-Solar+Climo Aero (blue), WRF-Solar+GEOS-5 Aero (cyan), WRF-Solar+CAMS Aero (green), and WRF-Solar+TE Microphys (yellow).
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
For clear-sky GHI (Figs. 13a,b), all four experiments have a similar MAE, with WRF-Solar+Climo Aero being best. For MBE, WRF-Solar+CAMS Aero has the lowest (best) error. For both clear-sky GHI MAE and MBE, WRF-Solar+GEOS-5 Aero has the largest (worst) error, with WRF-Solar+TE Microphys performing only slightly better. Again, consistent with the results previously discussed, it is clear that the choice of aerosol model or climatology with WRF-Solar only has a modest difference in GHI forecast skill.
As in Fig. 12, but for GHI.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
5. Results for hourly averaged clear-sky irradiance
To compare the performance of WRF-Solar against an operational global model, we obtained 1-h average (time ending) clear-sky GHI and DNI from CAMS, from the CAMS operational forecast model initialized daily at 0000 UTC. Archived instantaneous irradiance was unavailable, but hourly accumulated irradiance from every model time step was available.
We used clear-sky GHI and DHI and converted hourly average clear-sky DHI into clear-sky DNI by dividing by the cosine of the hourly average observed solar zenith angle (from every 1-min observation in the preceding hour), when the hourly average solar zenith angle is < 80°.
The 15-min gridded output from WRF-Solar for clear-sky GHI and DNI are converted to time-ending 1-h average values at the SURFRAD and SOLRAD sites. Note that constructing hourly averages from every 15-min datum introduces a sampling error in the early morning and late afternoon, when both GHI and DNI are changing quickly in clear skies, as compared with the middle of the day when they change more slowly, so the middle of the day provides the fairest comparison between WRF-Solar and CAMS.
For the SURFRAD and SOLRAD observations, hourly averages are calculated separately for validation of the CAMS and WRF-Solar model data, to account for the different sampling within the hour from the two models. First, for the CAMS comparison, hourly average observed GHI and DHI are calculated from every 1 min of data in the preceding hour, subject to the conditions that every minute in the hour be flagged as having clear sky, and that the hourly average solar zenith angle is < 80°. Hourly average observed irradiance values that do not meet these two conditions are excluded from analysis. Requiring every minute be clear is a stringent condition, and some stations are left with small sample sizes as a result (or some times of day with no clear hours over these 8+ months), but it ensures that the validating dataset is not contaminated by any cloud cover. Second, for the WRF-Solar comparison, the observations valid at the WRF-Solar gridded output times (every 15 min) are then averaged to hourly values, under the same two conditions as above. Additionally, hours that were not flagged as clear by the more-stringent every-1-min requirement were excluded from the dataset for the comparison with WRF-Solar. In this way, both CAMS and WRF-Solar are compared with the same set of clear-sky hours at each site, even as the hourly averages are determined in a way that is most fair to how each model’s hourly average irradiance is calculated. (We note here that if WRF-Solar irradiance were saved at every model time step at each station, then the observations could be averaged from every-1-min data just as for the CAMS comparison, rather than every-15-min data.) If the CAMS model data are validated with hourly average observations from every-15-min data, or if the WRF-Solar model data are validated with hourly average observations from every-1-min data, there are substantial increases in error, particularly in early morning and late afternoon (not shown). These artificially increased errors are entirely due to sampling error, from mismatching how the model and observation hourly averages are constructed.
In the analysis that follows in this section, the overall conclusions drawn from the instantaneous (every 15 min) analysis in section 3 for the four WRF-Solar experiments changes little for 1-h average irradiance, so our focus here is primarily on how CAMS compares with the WRF-Solar experiments.
a. Results by station
Maps of the overall MAE and MBE of clear-sky DNI from the four WRF-Solar experiments and CAMS are shown in Figs. 14 and 15, respectively. CAMS clear-sky DNI MAE is generally comparable to or somewhat larger than the WRF-Solar experiments at most sites. SEA stands out as a station with much higher clear-sky DNI MAE for CAMS than in any of the WRF-Solar experiments. Overall, WRF-Solar+GEOS-5 Aero is still the best-performing experiment.
One-hour-average clear-sky DNI MAE at SURFRAD (circles) and SOLRAD (squares) stations, for day-1 and day-2 lead times, and start dates 19 Nov 2019–30 Jul 2020 at 0900 UTC daily, for five experiments: (a) WRF-Solar+Climo Aero, (b) WRF-Solar+GEOS-5 Aero, (c) WRF-Solar+CAMS Aero, (d) WRF-Solar+TE Microphys, and (e) CAMS.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
For clear-sky DNI MBE, the values for CAMS are higher than any of the WRF-Solar experiments at most sites, with generally mildly positive biases in the eastern United States and strongly positive MBE values at SLC and SEA. Positive clear-sky DNI MBE values would ordinarily lead to a conclusion that the model had insufficient AOD. However, as can be seen for WRF-Solar+CAMS Aero (Fig. 15c), DNI MBE values are negative at all sites, and more strongly negative at sites in the western United States than in the other WRF-Solar experiments; this leads to the conclusion of too-high AOD in the western United States, and smaller overestimates of AOD in the eastern United States. This conclusion is in fact backed up by the validation of AOD550 in section 3 and Fig. 3, which shows that CAMS had a positive MBE across the entire domain, with strongly positive MBE in the western United States. Thus, we conclude that WRF-Solar behaves as expected according to the AOD evaluation in section 3. As we stated above in section 4a, the performance of the WRF-Solar clear-sky DNI is mainly controlled by the quality of the AOD at 550 nm, which is consistent with previous studies (Ruiz-Arias et al. 2014; Jiménez et al. 2016a).
As in Fig. 14, but for MBE.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
The analysis performed in this section with CAMS as a baseline for comparison opens important additional questions, however. The discrepancy between WRF-Solar+CAMS Aero and the CAMS model itself potentially arises at least in part from the CAMS AOD550 field but not the AE being imposed on WRF-Solar; thus, the assumptions made in WRF-Solar about the multispectral distribution of the aerosols, which are also impacted by the relative humidity in the model, could have led to too much attenuation of DNI, especially in comparison with the CAMS model itself. Quantifying the impact of ingesting both the AOD550 and AE in WRF-Solar on clear-sky irradiance requires further investigation with several additional months of model simulations for analysis. Other possible contributing factors include the respective radiation schemes, and how they interact with aerosols; WRF-Solar uses RRTMG, while CAMS (since IFS cycle 43r3) uses ecRad (Hogan and Bozzo 2018). For instance, it would be instructive to use collocated AOD and DNI observations to analyze the CAMS DNI bias as a function of the AOD bias. After characterizing this relationship, then a detailed investigation of ecRad itself would need to be conducted to understand the cause of the observed behavior, of a large bias in CAMS AOD550 coupled with a smaller than expected bias in CAMS clear-sky DNI. Exploring the specific assumptions or mechanisms in the respective radiation schemes (RRTMG and ecRad) that contribute to the difference in DNI statistics is certainly an important issue and could lead to improved parameterizations, but it is beyond the scope of the current study—identifying the impact on clear-sky irradiance forecasts from WRF-Solar using four aerosol representations—and is thus left for future research.
For the GHI MAE (Fig. 16), CAMS and all four WRF-Solar experiments have similarly low values at all stations. For the GHI MBE (Fig. 17), values are generally small at most stations, though CAMS has the most stations with a negative MBE (10 of 14) of all five models, which is consistent with other statistics presented later in this section. Also of note is that SLC does not have a large GHI MAE or MBE in this hourly average analysis, unlike in the instantaneous analysis (Figs. 6 and 7). Interestingly, this larger GHI MAE and MBE at SLC was apparent in all four WRF-Solar experiments and in CAMS, when compared with hourly average observations from every-15-min data alone (not shown). By further restricting the hourly average observed irradiances only to hours in which every 1-min observation was flagged as clear sky, these anomalously high GHI MAE and MBE values disappeared. Investigating the cause of this potential data issue at the SLC SOLRAD station is beyond the scope of this study, but we speculate that there may be an issue with some incorrect clear-sky flags in the RadFlux data at this site.
One-hour-average clear-sky GHI MAE at SURFRAD (circles) and SOLRAD (squares) stations, for day-1 and day-2 lead times, and start dates 19 Nov 2019–30 Jul 2020 at 0900 UTC daily, for five experiments: (a) WRF-Solar+Climo Aero, (b) WRF-Solar+GEOS-5 Aero, (c) WRF-Solar+CAMS Aero, (d) WRF-Solar+TE Microphys, and (e) CAMS.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
As in Fig. 16, but for MBE.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
b. Results by lead time
The clear-sky DNI MAE and MBE as a function of lead time are presented in Figs. 18a and 18b. For the four WRF-Solar experiments, the magnitudes are generally smaller than for the instantaneous analysis (Fig. 8), but the relative performance of the four experiments remains unchanged, with WRF-Solar+GEOS-5 Aero retaining the smallest DNI MAE and MBE in general, and WRF-Solar+TE Microphys performing next best. All the WRF-Solar experiments, except with GEOS-5 AOD550 imposed, exhibit negative biases in DNI ranging from approximately −(15–20) W m−2 in the middle of the day to −(50–100) W m−2 in early morning and late afternoon, indicating substantially higher scattering by aerosols than was observed. The CAMS DNI MAE is generally competitive with, if with somewhat higher MAE, than WRF-Solar+GEOS-5 Aero. For DNI MBE, CAMS has near-zero bias at most lead times, as does WRF-Solar+GEOS-5 Aero, indicating that the multispectral aerosol profile in the CAMS model itself is roughly appropriate and leads to only small errors in DNI.
One-hour-average clear-sky DNI (a) MAE and (b) MBE as a function of lead time for WRF-Solar simulations and CAMS initialized from 19 Nov 2019 to 30 Jul 2020. The five experiments are WRF-Solar+Climo Aero (blue dots), WRF-Solar+GEOS-5 Aero (cyan squares), WRF-Solar+CAMS Aero (green upward triangles), WRF-Solar+TE Microphys (yellow diamonds), and CAMS (red downward triangles).
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
Figures 19a and 19b show the clear-sky GHI MAE and MBE as a function of lead time. Here, all the WRF-Solar experiments have similarly small clear-sky GHI MAE values of 4–8 W m−2, which is generally within instrument error. The CAMS clear-sky GHI MAE is only slightly higher, generally ranging from 6 to 11 W m−2. CAMS shows a larger difference from the WRF-Solar experiments in clear-sky GHI MBE, where CAMS maintains a consistently negative bias over the range from −3 to −10 W m−2.
As in Fig. 18, but for GHI.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
As mentioned above in section 4a, the discrepancy between the CAMS (red) and WRF-Solar+CAMS Aero (green) performance in all these statistics is possibly due in part to information being lost about the multispectral distribution of aerosols by only ingesting the CAMS AOD550 field in WRF-Solar, and not the AE as well, and potentially to other key differences between the RRTMG and ecRad radiation schemes.
c. Results by month
Stratifying results by month for DNI MAE and MBE (Fig. 20) and GHI MAE and MBE (Fig. 21) yields results that are consistent with the instantaneous analysis (Figs. 10 and 11) for the WRF-Solar experiments and the performance of CAMS discussed above. For DNI MAE and MBE, CAMS performs competitively with WRF-Solar+GEOS-5 Aero. For GHI MAE, CAMS performs similarly well as the four WRF-Solar experiments, and with a small but consistent negative GHI MBE, unlike the WRF-Solar experiments. No discernible seasonal trends are apparent, except for the large increase in DNI MAE and MBE magnitudes in spring 2020 for the WRF-Solar+TE Microphys experiment. Determining the cause of this requires further investigation, including whether this overestimation of aerosol loading is consistent across multiple springs over CONUS, or whether this was a feature more limited to spring 2020 alone.
One-hour-average clear-sky DNI (a) MAE and (b) MBE for all forecast cycles initialized in every month, aggregated over all sites. The five experiments are WRF-Solar+Climo Aero (blue dots), WRF-Solar+GEOS-5 Aero (cyan squares), WRF-Solar+CAMS Aero (green upward triangles), WRF-Solar+TE Microphys (yellow diamonds), and CAMS (red downward triangles).
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
As in Fig. 20, but for GHI.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
d. Overall results
The overall summary results for DNI MAE and MBE are displayed in the bar charts in Fig. 22, with the GHI MAE and MBE shown in Fig. 23. Overall, WRF-Solar+GEOS-5 Aero has the lowest DNI MAE (21 W m−2), with CAMS being second best (25 W m−2). For DNI MBE, CAMS has the smallest-magnitude bias (+3 W m−2), whereas WRF-Solar+GEOS-5 Aero is second best (−5 W m−2), indicating that their errors generally cancel out on average. By contrast, the other three WRF-Solar experiments have much larger negative DNI MBE values, ranging from −22 to −30 W m−2. The substantial reduction in error in CAMS and WRF-Solar+GEOS-5 Aero further reinforces the importance of accurate aerosol data for good DNI predictions.
One-hour-average clear-sky DNI (a) MAE and (b) MBE over all sites, all day-1 and day-2 lead times, and all simulations. The five experiments are WRF-Solar+Climo Aero (blue), WRF-Solar+GEOS-5 Aero (cyan), WRF-Solar+CAMS Aero (green), WRF-Solar+TE Microphys (yellow), and CAMS (red).
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
As in Fig. 22, but for GHI.
Citation: Journal of Applied Meteorology and Climatology 62, 2; 10.1175/JAMC-D-22-0059.1
For GHI MAE, all four WRF-Solar experiments range from about 6 to 8 W m−2, while CAMS is only marginally higher just above 8 W m−2. These small differences are not of any practical significance, which is to be expected due to GHI being far less sensitive to aerosol loading than DNI. For GHI MBE, the two largest-magnitude biases are for CAMS (−5 W m−2) and WRF-Solar+GEOS-5 Aero (+3 W m−2), which are still very small.
6. Summary and conclusions
This study examined the impacts of aerosol input dataset (forecast or climatology) on WRF-Solar predictions of DNI and GHI in clear-sky conditions at high-quality observing sites. The four aerosol inputs used were the relatively coarse Tegen climatology (WRF-Solar+Climo Aero) that is the default aerosol input in WRF, GEOS-5 forecast AOD550 imposed onto WRF-Solar (WRF-Solar+GEOS-5), CAMS forecast AOD550 imposed onto WRF-Solar (WRF-Solar+CAMS Aero), and the Thompson–Eidhammer microphysics water-friendly and ice-friendly aerosol climatology (WRF-Solar+TE Microphys). WRF-Solar simulations were generated once daily at 0900 UTC from 19 November 2019 to 30 July 2020 and run for 45 h to encompass a full day-2 forecast for irradiance. For the validating observations, 1-min frequency DNI and GHI data with clear-sky flags were used from the 7 SURFRAD and 7 SOLRAD sites across CONUS.
First, analysis was conducted for instantaneous model and observation values every 15 min, both for the AOD550 field used by each experiment, and also for clear-sky irradiances. By analyzing MAE and MBE of AOD550, clear-sky DNI, and clear-sky GHI, the following conclusions emerged:
-
The source of aerosols in WRF-Solar has a large impact on predicted clear-sky DNI, with the largest errors in AOD550 producing the largest errors in clear-sky DNI using WRF-Solar parameterizations. This finding is consistent with previous studies that showed the importance of having a high-quality AOD dataset under clear skies with WRF-Solar. Imposing GEOS-5 forecast AOD550 onto WRF-Solar resulted in markedly lower MAE (29 W m−2 overall) and smaller MBE (−1 W m−2 overall) than using the Thompson–Eidhammer aerosol climatology, the Tegen aerosol climatology, or CAMS forecast AOD550, which had the highest clear-sky DNI MAE (48 W m−2 overall) and worst MBE (−33 W m−2 overall). Hence, WRF-Solar is behaving as expected, producing the best results in clear-sky DNI when using the best AOD550 dataset (GEOS-5 forecasts).
-
There are some regional patterns in the clear-sky DNI errors. For instance, while WRF-Solar+CAMS Aero had negative DNI MBE everywhere (except SLC, which may have a data quality issue), the biases were more strongly negative in the western United States and more weakly negative in the eastern United States, suggesting that either CAMS AOD550 was more strongly overpredicted in the western United States than in the eastern United States (which it was, as seen in the AOD550 validation), and/or that assumptions about the multispectral properties of the aerosol loading in WRF-Solar were more incorrect in the western United States due to different types of aerosols dominating. Also, imposing the GEOS-5 AOD550 led to negative clear-sky DNI MBE at stations in the northern and eastern United States, but weakly positive DNI MBE at stations in the southwestern United States. Characterizing the errors in aerosol datasets for regions of interest is crucial for accurate DNI forecasts in clear-sky conditions, as is properly characterizing the multispectral properties of the aerosols that impact scattering.
-
The source of aerosols in WRF-Solar has little impact on predicted clear-sky GHI. Neither the MAE nor MBE change much across the four experiments with different aerosol datasets. In cases where aerosol loading departs significantly from aerosol climatologies (e.g., wildfire smoke or severe dust storms), however, it is expected that imposing aerosol forecasts would bring some benefit to clear-sky GHI forecasts (Juliano et al. 2022).
Second, analysis was conducted for hourly averaged model and observation clear-sky DNI and GHI values at the same stations, both for the four WRF-Solar experiments and also for the CAMS model. The following conclusions emerged:
-
WRF-Solar results were largely consistent between the instantaneous and hourly average analysis, despite the much smaller sample sizes in the hourly average dataset, reinforcing the conclusions made in that analysis.
-
CAMS clear-sky GHI and DNI was competitive overall with WRF-Solar+GEOS-5 Aero, the best of the WRF-Solar configurations. CAMS clear-sky DNI MAE was somewhat higher at every station, however, and DNI MBE had the opposite sign (and typically a smaller magnitude) at most stations than WRF-Solar+GEOS-5 Aero. CAMS clear-sky GHI MBE was also consistently negative at all lead times, unlike all four WRF-Solar experiments.
-
The differences in clear-sky DNI between CAMS and WRF-Solar with CAMS AOD550 imposed were quite large. This finding was unexpected, and points to the need for additional research to explain why. In any case, using CAMS AOD clearly hurts WRF-Solar performance for DNI relative to other aerosol representations; this performance is consistent with the AOD validation, in which CAMS AOD550 has the largest errors when validated against AERONET stations.
-
Sampling error is crucial to account for when analyzing hourly averages of solar irradiance, particularly in early morning and late afternoon when the solar zenith angle (and thus irradiance) changes quickly. Care must be taken to maintain consistency in sampling for fair comparisons between models and observations. Also, infrequent sampling (every 15 min vs every 1 min) can artificially introduce errors into analysis. Thus, we also recommend that WRF-Solar users turn on both the solar diagnostics package and activate time series output (through a tslist file) to obtain irradiance and other solar diagnostic values at every model time step at selected locations.
For the difference between WRF-Solar+CAMS Aero and CAMS irradiance errors, it is possible that errors in WRF-Solar clear-sky DNI are introduced by simplified assumptions about the multispectral aerosol properties when only ingesting AOD550, as WRF-Solar does. Alternatively, or additionally, there may be other key differences between the two radiation schemes (RRTMG in WRF-Solar, ecRad in CAMS) that account for the reduced sensitivity to AOD550 in CAMS versus WRF-Solar. Additional research is required to explore these issues thoroughly and develop improved parameterizations.
The above conclusions have important ramifications for solar power forecasting in clear-sky (i.e., cloud-free) conditions, regardless of whether instantaneous or hourly average irradiance is desired. If forecasting for PV plants, where GHI is the relevant variable, aerosol climatologies are likely sufficient in models like WRF-Solar in most situations. The exception to this would be during events when the actual aerosol loading is substantially different from the aerosol climatology, such as with dense smoke plumes from wildfires (Juliano et al. 2022). This is of particular relevance in the western United States, where the bulk of installed solar capacity is located, and because of the increasing severity and extent of drought-induced wildfires that is expected in coming years due to climate change (Abatzoglou and Williams 2016; Yoon et al. 2015; Crockett and Westerling 2018; Juliano et al. 2022). Of course, NWP models like GEOS-5, CAMS, and HRRR can only start predicting such plumes and their increased AOD after the fires or smoke plumes have been observed by satellites and subsequently inserted or assimilated into those models.
If forecasting for CSP plants, where DNI is the relevant variable, accurate representation of aerosols in the NWP model is absolutely crucial to getting accurate DNI forecasts in clear-sky conditions. Configured properly with aerosols, WRF-Solar can outperform leading global models like CAMS, with the additional advantage of producing forecasts on higher frequencies than hourly averages. If using a model like WRF-Solar that does not have explicit aerosol modeling capability of its own at this time, it is also important to consider the multispectral properties of the aerosols, and explore ingesting not only AOD550, but also the AE. The composition and vertical distribution of the aerosols, as well as the water vapor and ozone distribution, can also play an important role in the attenuation of incoming solar radiation. Another area that requires additional research is the integration of an explicit circumsolar radiation correction to DNI in WRF-Solar, as has recently been implemented in other radiative transfer models (Qin et al. 2021; Xie et al. 2020).
Even so, errors in predicted solar irradiance from aerosols, while they may be the “low-hanging fruit” that can be most easily addressed for improved forecasts, are still dwarfed by errors introduced by inaccurate representation of clouds in NWP models, both in location and depth and in microphysical and optical properties. Additional research is required and is ongoing to continue improving irradiance forecasts in all sky conditions.
Acknowledgments.
This material is based upon work supported by the National Center for Atmospheric Research (NCAR), which is a major facility sponsored by the National Science Foundation under Cooperative Agreement 1852977. The authors gratefully acknowledge direct funding for this work from NASA under Grant 80NSSC18K0330. We acknowledge high-performance computing support from Cheyenne (Computational and Information Systems Laboratory 2019) provided by NCAR’s Computational and Information Systems Laboratory, sponsored by the National Science Foundation. The authors also thank Branko Kosović of NCAR and three anonymous reviewers for helpful comments that improved the paper.
Data availability statement.
The WRF-Solar (AOD550 and clear-sky irradiance) and CAMS (clear-sky irradiance) forecast data at station locations for each model cycle, as well as the SURFRAD, SOLRAD, and AERONET observation data used in this study, are available publicly in the NCAR GDEX Repository (Lee et al. 2022). SURFRAD RadFlux data are publicly available at https://gml.noaa.gov/aftp/data/radiation/surfrad/RadFlux/, SOLRAD RadFlux data are publicly available at https://gml.noaa.gov/aftp/data/radiation/solrad/RadFlux/, and AERONET data are publicly available at https://aeronet.gsfc.nasa.gov/.
REFERENCES
Abatzoglou, J. T., and A. P. Williams, 2016: Impact of anthropogenic climate change on wildfire across western US forests. Proc. Natl. Acad. Sci. USA, 113, 11 770–11 775, https://doi.org/10.1073/pnas.1607171113.
Agustí-Panareda, A., and Coauthors, 2014: Forecasting global atmospheric CO2. Atmos. Chem. Phys., 14, 11 959–11 983, https://doi.org/10.5194/acp-14-11959-2014.
Augustine, J. A., J. J. DeLuisi, and C. N. Long, 2000: SURFRAD—A national surface radiation budget network for atmospheric research. Bull. Amer. Meteor. Soc., 81, 2341–2358, https://doi.org/10.1175/1520-0477(2000)081<2341:SANSRB>2.3.CO;2.
Augustine, J. A., G. B. Hodges, C. R. Cornwall, J. J. Michalsky, and C. I. Medina, 2005: An update on SURFRAD—The GCOS surface radiation budget network for the continental United States. J. Atmos. Oceanic Technol., 22, 1460–1472, https://doi.org/10.1175/JTECH1806.1.
Benedetti, A., and Coauthors, 2009: Aerosol analysis and forecast in the European Centre for medium-range weather forecasts integrated forecast system: 2. Data assimilation. J. Geophys. Res., 114, D13205, https://doi.org/10.1029/2008JD011115.
Blanc, P., and Coauthors, 2014: Direct normal irradiance related definitions and applications: The circumsolar issue. Sol. Energy, 110, 561–577, https://doi.org/10.1016/j.solener.2014.10.001.
Buchard, V., and Coauthors, 2015: Using the OMI aerosol index and absorption aerosol optical depth to evaluate the NASA MERRA aerosol reanalysis. Atmos. Chem. Phys., 15, 5743–5760, https://doi.org/10.5194/acp-15-5743-2015.
Buchard, V., and Coauthors, 2017: The MERRA-2 aerosol reanalysis, 1980 onward. Part II: Evaluation and case studies. J. Climate, 30, 6851–6872, https://doi.org/10.1175/JCLI-D-16-0613.1.
Byun, D., and K. L. Schere, 2006: Review of the governing equations, computational algorithms, and other components of the models-3 community multiscale air quality (CMAQ) modeling system. Appl. Mech. Rev., 59, 51–77, https://doi.org/10.1115/1.2128636.
Computational and Information Systems Laboratory, 2019: Cheyenne: HPE/SGI ICE XA System. NCAR Advanced Research Computing, https://doi.org/10.5065/D6RX99HX.
Copernicus, 2020: CAMS near-real-time global analysis and forecast daily data. ECMWF, accessed 3 August 2020, https://www.ecmwf.int/en/forecasts/dataset/cams-near-real-time-global-analysis-and-forecast-daily-data.
Crockett, J. L., and A. L. Westerling, 2018: Greater temperature and precipitation extremes intensify western U.S. droughts, wildfire severity, and Sierra Nevada tree mortality. J. Climate, 31, 341–354, https://doi.org/10.1175/JCLI-D-17-0254.1.
Dasari, H. P., S. Desamsetti, S. Langodan, R. Attada, R. K. Kunchala, Y. Viswanadhapalli, O. Knio, and I. Hoteit, 2019: High-resolution assessment of solar energy resources over the Arabian Peninsula. Appl. Energy, 248, 354–371, https://doi.org/10.1016/j.apenergy.2019.04.105.
Eck, T. F., B. N. Holben, J. S. Reid, O. Dubovik, A. Smirnov, N. T. O’Neill, I. Slutsker, and S. Kinne, 1999: Wavelength dependence of the optical depth of biomass burning, urban, and desert dust aerosols. J. Geophys. Res., 104, 31 333–31 349, https://doi.org/10.1029/1999JD900923.
ECMWF, 2019: Operational configurations of the ECMWF Integrated Forecasting System (IFS). ECMWF, accessed 3 August 2020, https://www.ecmwf.int/en/forecasts/documentation-and-support.
Ek, M. B., 2003: Implementation of Noah land surface model advances in the national centers for environmental prediction operational mesoscale Eta model. J. Geophys. Res., 108, 8851, https://doi.org/10.1029/2002JD003296.
Engelen, R., 2019: Implementation of IFS cycle 46r1—Copernicus Services. ECMWF Confluence Wiki, accessed 3 August 2020, https://confluence.ecmwf.int/display/COPSRV/Implementation+of+IFS+cycle+46r1.
Engelen, R., J. Flemming, A. Inness, and M. Suttie, 2009: Assimilation of IASI observations in the context of the GEMS and MACC projects. ECMWF/EUMETSAT NWP-SAF Workshop on the assimilation of IASI in NWP, Shinfield Park, Reading, ECMWF, 123–130, https://www.ecmwf.int/node/9294.
Fast, J. D., W. I. Gustafson, R. C. Easter, R. A. Zaveri, J. C. Barnard, E. G. Chapman, G. A. Grell, and S. E. Peckham, 2006: Evolution of ozone, particulates, and aerosol direct radiative forcing in the vicinity of Houston using a fully coupled meteorology-chemistry-aerosol model. J. Geophys. Res., 111, D21305, https://doi.org/10.1029/2005JD006721.
Flemming, J., and Coauthors, 2015: Tropospheric chemistry in the integrated forecasting system of ECMWF. Geosci. Model Dev., 8, 975–1003, https://doi.org/10.5194/gmd-8-975-2015.
Flemming, J., V.-H. Peuch, and L. Jones, 2017: Ten years of forecasting atmospheric composition at ECMWF. ECMWF Newsletter, No. 152, ECMWF, Reading, United Kingdom, 5–6, https://www.ecmwf.int/en/newsletter/152/news/ten-years-forecasting-atmospheric-composition-ecmwf#top.
Fu, D., M. Liu, D. Yang, H. Che, and X. Xia, 2022: Influences of atmospheric reanalysis on the accuracy of clear-sky irradiance estimates: Comparing MERRA-2 and CAMS. Atmos. Environ., 277, 119080, https://doi.org/10.1016/j.atmosenv.2022.119080.
Gelaro, R., and Coauthors, 2017: The modern-era retrospective analysis for research and applications, version 2 (MERRA-2). J. Climate, 30, 5419–5454, https://doi.org/10.1175/JCLI-D-16-0758.1.
Grell, G. A., and S. R. Freitas, 2014: A scale and aerosol aware stochastic convective parameterization for weather and air quality modeling. Atmos. Chem. Phys., 14, 5233–5250, https://doi.org/10.5194/acp-14-5233-2014.
Grell, G. A., S. E. Peckham, R. Schmitz, S. A. McKeen, G. Frost, W. C. Skamarock, and B. Eder, 2005: Fully coupled “online” chemistry within the WRF model. Atmos. Environ., 39, 6957–6975, https://doi.org/10.1016/j.atmosenv.2005.04.027.
Gueymard, C. A., 2012: Temporal variability in direct and global irradiance at various time scales as affected by aerosols. Sol. Energy, 86, 3544–3553, https://doi.org/10.1016/j.solener.2012.01.013.
Gueymard, C. A., and D. Yang, 2020: Worldwide validation of CAMS and MERRA-2 reanalysis aerosol optical depth products using 15 years of AERONET observations. Atmos. Environ., 225, 117216, https://doi.org/10.1016/j.atmosenv.2019.117216.
Haupt, S. E., and Coauthors, 2018: Building the Sun4Cast system: Improvements in solar power forecasting. Bull. Amer. Meteor. Soc., 99, 121–136, https://doi.org/10.1175/BAMS-D-16-0221.1.
Haupt, S. E., and Coauthors, 2020: Combining artificial intelligence with physics-based methods for probabilistic renewable energy forecasting. Energies, 13, 1979, https://doi.org/10.3390/en13081979.
Hicks, B. B., J. J. DeLuisi, and D. R. Matt, 1996: The NOAA Integrated Surface Irradiance Study (ISIS)—A new surface radiation monitoring program. Bull. Amer. Meteor. Soc., 77, 2857–2864, https://doi.org/10.1175/1520-0477(1996)077<2857:TNISIS>2.0.CO;2.
Hogan, R. J., and A. Bozzo, 2018: A flexible and efficient radiation scheme for the ECMWF model. J. Adv. Model. Earth Syst., 10, 1990–2008, https://doi.org/10.1029/2018MS001364.
Holben, B. N., and Coauthors, 1998: AERONET—A federated instrument network and data archive for aerosol characterization. Remote Sens. Environ., 66 (1), 1–16, https://doi.org/10.1016/S0034-4257(98)00031-5.
Huijnen, V., and Coauthors, 2019: Quantifying uncertainties due to chemistry modelling—Evaluation of tropospheric composition simulations in the CAMS model (cycle 43R1). Geosci. Model Dev., 12, 1725–1752, https://doi.org/10.5194/gmd-12-1725-2019.
Iacono, M. J., J. S. Delamere, E. J. Mlawer, M. W. Shephard, S. A. Clough, and W. D. Collins, 2008: Radiative forcing by long-lived greenhouse gases: Calculations with the AER radiative transfer models. J. Geophys. Res., 113, D13103, https://doi.org/10.1029/2008JD009944.
Inness, A., and Coauthors, 2019: The CAMS reanalysis of atmospheric composition. Atmos. Chem. Phys., 19, 3515–3556, https://doi.org/10.5194/acp-19-3515-2019.
Jiménez, P. A., J. Dudhia, J. F. González-Rouco, J. Navarro, J. P. Montávez, and E. García-Bustamante, 2012: A revised scheme for the WRF surface layer formulation. Mon. Wea. Rev., 140, 898–918, https://doi.org/10.1175/MWR-D-11-00056.1.
Jiménez, P. A., and Coauthors, 2016a: WRF-Solar: Description and clear-sky assessment of an augmented NWP model for solar power prediction. Bull. Amer. Meteor. Soc., 97, 1249–1264, https://doi.org/10.1175/BAMS-D-14-00279.1.
Jiménez, P. A., S. Alessandrini, S. E. Haupt, A. Deng, B. Kosovic, J. A. Lee, and L. Delle Monache, 2016b: The role of unresolved clouds on short-range global horizontal irradiance predictability. Mon. Wea. Rev., 144, 3099–3107, https://doi.org/10.1175/MWR-D-16-0104.1.
Juliano, T. W., and Coauthors, 2022: Smoke from 2020 United States wildfires responsible for substantial solar energy forecast errors. Environ. Res. Lett., 17, 034010, https://doi.org/10.1088/1748-9326/ac5143.
Lee, J. A., S. E. Haupt, P. A. Jiménez, M. A. Rogers, S. D. Miller, and T. C. McCandless, 2017: Solar irradiance nowcasting case studies near Sacramento. J. Appl. Meteor. Climatol., 56, 85–108, https://doi.org/10.1175/JAMC-D-16-0183.1.
Lee, J. A., C. A. Gueymard, and P. A. Jiménez, 2021: Performance of aerosol optical depth forecasts over the Middle East: Multi-model analysis and validation. ESS Open Archive, https://doi.org/10.1002/essoar.10506487.1.
Lee, J. A., P. A. Jiménez, J. Dudhia, and Y.-M. Saint-Drenan, 2022: WRF-Solar AOD550 & clear-sky irradiance forecasts & verifying observations over CONUS. NCAR, accessed 2 August 2022, https://doi.org/10.5065/v94m-y402.
Lee, P., and Coauthors, 2017: NAQFC developmental forecast guidance for fine particulate matter (PM2.5). Wea. Forecasting, 32, 343–360, https://doi.org/10.1175/WAF-D-15-0163.1.
Long, C. N., and Y. Shi, 2008: An automated quality assessment and control algorithm for surface radiation measurements. Open Atmos. Sci. J., 2, 23–37, https://doi.org/10.2174/1874282300802010023.
Marchand, M., Y.-M. Saint-Drenan, L. Saboret, E. Wey, and L. Wald, 2020: Performance of CAMS radiation service and HelioClim-3 databases of solar radiation at surface: Evaluating the spatial variation in Germany. Adv. Sci. Res., 17, 143–152, https://doi.org/10.5194/asr-17-143-2020.
McCandless, T., and P. A. Jiménez, 2020: Examining the potential of a random forest derived cloud mask from GOES-R satellites to improve solar irradiance forecasting. Energies, 13, 1671, https://doi.org/10.3390/en13071671.
Molod, A., L. Takacs, M. Suarez, and J. Bacmeister, 2015: Development of the GEOS-5 atmospheric general circulation model: Evolution from MERRA to MERRA2. Geosci. Model Dev., 8, 1339–1356, https://doi.org/10.5194/gmd-8-1339-2015.
Morcrette, J.-J., and Coauthors, 2009: Aerosol analysis and forecast in the European Centre for Medium-Range Weather Forecasts Integrated Forecast System: Forward modeling. J. Geophys. Res., 114, D06206, https://doi.org/10.1029/2008JD011235.
Nakanishi, M., and H. Niino, 2006: An improved Mellor–Yamada level-3 model: Its numerical stability and application to a regional prediction of advection fog. Bound.-Layer Meteor., 119, 397–407, https://doi.org/10.1007/s10546-005-9030-8.
Oumbe, A., Z. Qu, P. Blanc, H. Bru, M. Lefèvre, and L. Wald, 2012: Modeling circumsolar irradiance to adjust beam irradiances from radiative transfer models to measurements. EMS Annual Meeting 2012, Lodz, Poland, PSL, https://hal-mines-paristech.archives-ouvertes.fr/hal-00741624.
Powers, J. G., and Coauthors, 2017: The weather research and forecasting model: Overview, system efforts, and future directions. Bull. Amer. Meteor. Soc., 98, 1717–1737, https://doi.org/10.1175/BAMS-D-15-00308.1.
Qin, Y., T. R. McVicar, J. Huang, S. West, and A. D. L. Steven, 2021: Application of small angle approximation in circumsolar irradiance modelling. Sol. Energy Adv., 1, 100001, https://doi.org/10.1016/j.seja.2021.100001.
Randles, C. A., and Coauthors, 2017: The MERRA-2 aerosol reanalysis, 1980 onward. Part I: System description and data assimilation evaluation. J. Climate, 30, 6823–6850, https://doi.org/10.1175/JCLI-D-16-0609.1.
Rémy, S., and Coauthors, 2019: Description and evaluation of the tropospheric aerosol scheme in the European Centre for medium-range weather forecasts (ECMWF) integrated forecasting system (IFS-AER, cycle 45R1). Geosci. Model Dev., 12, 4627–4659, https://doi.org/10.5194/gmd-12-4627-2019.
Rienecker, M. M., and Coauthors, 2008: The GEOS-5 Data Assimilation System-Documentation of versions 5.0.1, 5.1.0, and 5.2.0. NASA Tech. Memo. NASA/TM-2008-104606, Vol. 27, 97 pp., http://gmao.gsfc.nasa.gov/pubs/docs/Rienecker369.pdf.
Ruiz-Arias, J. A., and C. A. Gueymard, 2018: Worldwide inter-comparison of clear-sky solar radiation models: Consensus-based review of direct and global irradiance components simulated at the earth surface. Sol. Energy, 168, 10–29, https://doi.org/10.1016/j.solener.2018.02.008.
Ruiz-Arias, J. A., J. Dudhia, C. A. Gueymard, and D. Pozo-Vázquez, 2013: Assessment of the level-3 MODIS daily aerosol optical depth in the context of surface solar radiation and numerical weather modeling. Atmos. Chem. Phys., 13, 675–692, https://doi.org/10.5194/acp-13-675-2013.
Ruiz-Arias, J. A., J. Dudhia, and C. A. Gueymard, 2014: A simple parameterization of the short-wave aerosol optical properties for surface direct and diffuse irradiances assessment in a numerical weather model. Geosci. Model Dev., 7, 1159–1174, https://doi.org/10.5194/gmd-7-1159-2014.
Schuster, G. L., O. Dubovik, and B. N. Holben, 2006: Angstrom exponent and bimodal aerosol size distributions. J. Geophys. Res., 111, D07207, https://doi.org/10.1029/2005JD006328.
Skamarock, W. C., and Coauthors, 2019: A description of the Advanced Research WRF Model version 4.1. NCAR Tech. Note NCAR/TN-556+STR, 145 pp., https://doi.org/10.5065/1dfh-6p97.
Tegen, I., P. Hollrig, M. Chin, I. Fung, D. Jacob, and J. Penner, 1997: Contribution of different aerosol species to the global aerosol extinction optical thickness: Estimates from model results. J. Geophys. Res., 102, 23 895–23 915, https://doi.org/10.1029/97JD01864.
Thompson, G., and T. Eidhammer, 2014: A study of aerosol impacts on clouds and precipitation development in a large winter cyclone. J. Atmos. Sci., 71, 3636–3658, https://doi.org/10.1175/JAS-D-13-0305.1.
Thompson, G., P. R. Field, R. M. Rasmussen, and W. D. Hall, 2008: Explicit forecasts of winter precipitation using an improved bulk microphysics scheme. Part II: Implementation of a new snow parameterization. Mon. Wea. Rev., 136, 5095–5115, https://doi.org/10.1175/2008MWR2387.1.
Verbois, H., R. Huva, A. Rusydi, and W. Walsh, 2018: Solar irradiance forecasting in the tropics using numerical weather prediction and statistical learning. Sol. Energy, 162, 265–277, https://doi.org/10.1016/j.solener.2018.01.007.
Xie, Y., M. Sengupta, Y. Liu, H. Long, Q. Min, W. Liu, and A. Habte, 2020: A physics-based DNI model assessing all-sky circumsolar radiation. iScience, 23, 100893, https://doi.org/10.1016/j.isci.2020.100893.
Yang, J., J.-H. Kim, P. A. Jiménez, M. Sengupta, J. Dudhia, Y. Xie, A. Golnas, and R. Giering, 2021: An efficient method to identify uncertainties of WRF-Solar variables in forecasting solar irradiance using a tangent linear sensitivity analysis. Sol. Energy, 220, 509–522, https://doi.org/10.1016/j.solener.2021.03.044.
Yoon, J.-H., B. Kravitz, P. J. Rasch, S.-Y. S. Wang, R. R. Gillies, and L. Hipps, 2015: Extreme fire season in California: A glimpse into the future? [in “Explaining Extreme Events of 2014 from a Climate Perspective”]. Bull. Amer. Meteor. Soc., 96 (12), S5–S9, https://doi.org/10.1175/BAMS-D-15-00114.1.
