You are on page 1of 15

Solar Energy 74 (2003) 381395

Direct solar transmittance and irradiance predictions with broadband models. Part II: validation with high-quality measurements
Christian A. Gueymard 1
Solar Consulting Services, 174 Bluebird Lane, Bailey, CO 80421, USA Received 9 December 2002; accepted 13 May 2003

Abstract A thorough investigation of the performance of broadband direct irradiance predictions using 21 solar radiation models, along with carefully measured radiation data and ancillary meteorological data, is detailed here. A sensitivity study and a detailed error analysis show that precipitable water, and even more so, turbidity, are the two most critical inputs, whose accuracy conditions the resulting uncertainty in irradiance predictions. Large prediction uncertainties result from the use of time / space interpolated or extrapolated data of precipitable water and turbidity. So that the results of performance assessment studies like this one can be of any signicance, it is necessary to rely on highly accurate precipitable water and turbidity data from collocated instruments with an appropriate sampling rate. An experimental assessment of the performance of all models has been conducted, using nearly 5000 data points from ve different sites covering a large range of geographical and climatic conditions. Direct irradiance measured with rst-class instruments at these sites are compared to model predictions where precipitable water and turbidity are determined from collocated sunphotometric measurements. This experimental assessment is found to be less stringent than the theoretical assessment (in Part 1 of this investigation), while conrming its main results. The same four high-performance models as in Part 1 are nally recommended: CPCR2, MLWT2, REST and Yang (in alphabetical order). Remarkably, they can predict direct irradiance under a variety of atmospheric conditions within the uncertainty of modern and well-maintained pyrheliometers, provided that good quality inputs of precipitable water and turbidity are used. The MLWT2 model produces the best results, with the lowest bias and variance for any irradiance value. 2003 Elsevier Ltd. All rights reserved.

1. Introduction The accurate measurement of direct normal irradiance (DNI) is essential in many disciplines, including solar resource assessment, solar energy applications, atmospheric sciences, climate change, and air quality. These radiation data can be used either directly, for instance to evaluate the solar resource at a specic site, or indirectly, for instance to monitor trends in atmospheric turbidity and pollution over long periods of time.

E-mail address: Gueymard). 1 ISES member.

chris@solarconsultingservices.com

(C.A.

Comparing predictions from solar radiation models to measured data has always been a logical way of assessing their performance (see e.g. Battles et al., 2000; Davies et al., 1988; Gueymard, 1993a; Olmo et al., 2001). When dealing with clear-sky conditions only, a main nding in these studies was that the accuracy to which clear-sky direct irradiance could be predicted with the best broadband radiation models was close to the measurement accuracy. Since then, progress in pyrheliometry has been made so that eld measurements are generally more consistent and accurate than before. New instrumentation has appeared (e.g. active cavity radiometers) with improved traceability to the World Radiometric Reference (WRR). Furthermore, there has been a strong and rapidly increasing need for very accurate reference data aimed at specic scientic

0038-092X / 03 / $ see front matter 2003 Elsevier Ltd. All rights reserved. doi:10.1016 / S0038-092X(03)00196-8

382

C. A. Gueymard / Solar Energy 74 (2003) 381395

tasks (e.g. the verication of radiative transfer models, general circulation models, and climate models, or the ground truth for remote-sensed data). As a consequence, the quality of measured datasets has improved over the last decadeeven though the number of such rst-class instrumented sites is still insufcient. There has been also major progress in the availability and accuracy of ancillary measurements, such as those of precipitable water and aerosol turbidity, which are the two most important atmospheric variables affecting direct irradiance under clear skies. In the rst part of this contribution (Gueymard, 2003a; hereafter Part 1), the theoretical accuracy of 21 direct radiation models has been assessed using appropriate theoretical benchmarks. The reader is also referred to Part 1 for a description of all these models. This second part consists in a complete re-evaluation of the experimental part of an earlier general study mentioned above (Gueymard, 1993a), as well as of similar investigations based on measured data only (Battles et al., 2000; Olmo et al., 2001). Even though these studies were timely and provided a wealth of information about the real world performance of radiation models, they generally did not address a fundamental question, which impacts any performance assessment based on comparisons between modelled and measured data: What part of the apparent error in the predicted irradiance is attributable to errors in the input data? The originality of the present study is twofold: (i) it addresses the above question directly through theoretical developments involving a general uncertainty analysis; and (ii) it also addresses this question indirectly by carefully selecting the needed reference benchmark data from experimental sources of the highest quality and largest climatic diversity.

differences between modelled and measured data are caused by shortcomings in the model, inaccurate input data, or simply inaccurate DNI measurements used to test the model. DNI is now considered to have a relatively low uncertainty if measured with modern instrumentation and careful calibration techniques: about 3% for the thermopile pyrheliometers used in networks, about 0.5% for glazed secondary-standard active-cavity pyrheliometers, and about 0.1% for the unglazed primary-standard instruments that collectively maintain WRR. The ancillary atmospheric measurements under scrutiny here are those that are usually needed as inputs to a model, i.e. those listed in Table 1 of Part 1. From this list, only pressure, p, can be measured with great accuracy. Because DNI is not very sensitive to pressure, even a rough estimate of p using an empirical function of elevation and, possibly, latitude (e.g. Gueymard, 1993b) is usually sufcient. Sensitivity calculations can be easily performed using a simple equation to predict the combined Rayleigh and mixed gas transmittance. That from Choudhurys model (Choudhury, 1982) is used here because of its good accuracy (see Table 4 in Part 1): T R T g 5 [1 2 0.606m 9 /(6.43 1 m 9 )][1 2 0.0075m 9
0.875

] (1)

2. Input data uncertainty and error analysis

2.1. Individual inputs and their uncertainties


As suggested by the results of previous investigations (e.g. Gueymard, 1993a) and conrmed by the theoretical results discussed in Part 1 of this study, a number of existing models are potentially able to predict DNI with an accuracy close or comparable to that of direct eld measurements, inasmuch as the input atmospheric data these models rely on are available and accurate. Therefore, in order to use high-performance models and obtain accurate predictions, high-quality and very accurate experimental data need to be obtained for the ancillary atmospheric measurements required by these models. The sensitivity of modelled DNI to errors in input data needs to be investigated and quantied, because input errors can propagate and considerably affect DNI predictions. It is desirable to know to what extent observed

where m 9 is the absolute, or pressure-corrected, air mass, i.e. m 9 5 m ( p / 1013.25). It follows that, when everything else is known, an unlikely large error of 5% in pressure (and thus in m 9 ) would translate into a slight error of 0.4% in DNI for an overhead sun (m 95 1) or 1.4% for a very low sun (m 95 10). Ozone has only a limited effect on DNI, as revealed in Table 5 and Fig. 4 of Part 1. The driving variable, the ozone columnar amount, u o , can be obtained from different sources: ground observations from hundreds of sites of the World Ozone and Ultraviolet Data Center (WOUDC) network, satellite observations covering the whole world, or climatological averages based on a latitude / longitude grid. The uncertainty in u o for instantaneous calculations might vary between about 5 and 30%, depending on source of data and recourse to space / time interpolation. A simple sensitivity study can be performed using Choudhurys model again, and shows that, for a large 30% error in u o , the resulting error in DNI is only 0.3% for Z 5 08 and 1% for Z 5 858. These errors, like those due to pressure, are well within the instrumental error in routine DNI measurements. Water vapor, nitrogen dioxide, and aerosols, however, can be major sources of extinction. When this is compounded with their generally high uncertainties due to instrumental limitations and lack of space / time resolution, relatively high uncertainties in DNI might result. The sensitivity of DNI to uncertainties in precipitable water, w, can be estimated easily using the rather accurate

C. A. Gueymard / Solar Energy 74 (2003) 381395

383

transmittance equation in Choudhurys model (see Table 6 in Part 1): T w 5 1.0121 2 0.11 (0.8 mw 1 0.000631)
0.3

T a 5 exp h 2 mb [0.6777 1 0.1464 mb 2 0.00626 (m b )2 ] 2 1.3 j (3) is used here because its accuracy has been demonstrated in Table 7 and Fig. 9 of Part 1. Sample calculations show that an error of 10% in b translates into an error of 0.3% in DNI for mb 5 0.02, 3% for mb 5 0.2, 10% for mb 5 1, and 21% for mb 5 5. Like in the case for precipitable water, turbidity varies rapidly over time and space, so that instantaneous errors of 50% or more in b can occur if no collocated measurement is available, which can lead to considerable uncertainty in DNI. The important relative effects on DNI of b, w and m is also demonstrated in a different way in Fig. 1, using version 2.9.2 of the SMARTS code (Gueymard, 1995, 2001, 2003b) and considering ideal aerosols characterized by a wavelength exponent following Angstroms recommendation, a 5 1.3. The sensitivity of DNI, Ebn , to a variation or an error in b, DEbn /Db, noticeably increases when b decreases or m increases. For instance, DEbn /Db 657 W/ m 2 for m 5 1, b 5 0.5, and mb 5 0.5, whereas DEbn /Db 2054 W/ m 2 for mb 5 0.5 again, but with rather m 5 5 and b 5 0.1. For the latter conditions, a 60.05 absolute error in b yields a 6103 W/ m 2 error in Ebn .

(2)

For dry conditions and a high sun (mw 5 0.2 cm), a 10% error in w translates into a 0.2% error in DNI. At the other extreme (mw 5 50 cm), the error in DNI becomes 1.5%. Even though these resulting errors may seem low or negligible, they must be regarded as generally optimistic. In an ideal case, w would be measured on the very site where DNI needs to be evaluated, with a good time resolution (e.g. at least once an hour), and with a performing instrument (microwave radiometer, sunphotometer, or GPS receiver). Only in this very favourable case an uncertainty , 10% can be considered for w. But because precipitable water is highly variable over time and space, any recourse to interpolation will dramatically increase the uncertainty in w. This is for example the case when using twice-daily radiosonde data from a distant site, or gridded data from satellite observations. When w must be empirically estimated from ground data of temperature and humidity, separate tests have revealed that the instantaneous error in w may well be . 100%, depending on atmospheric conditions and on the universality of the empirical model used. In such unfavorable cases, the propagation of error to DNI may reach 215% or more. When DNI needs to be predicted under polluted conditions, nitrogen dioxide (NO 2 ) becomes a non-negligible absorber, as discussed in Section 3.1.5 of Part 1. Measurements of the vertical column of NO 2 , u n , are exceptional (Kambezidis et al., 2001; Schroeder and Davies, 1987), so that the main source of usable data would be from networks of air samplers that measure the NO 2 concentration at ground level in and around urban or industrial areas. It is possible to relate these ground concentrations to the total column amount if the height of the well-mixed pollution layer is known. It generally varies between 0.5 and 2 km, but uctuates over time and is rarely measured. Therefore, an instantaneous error of 100% in the tropospheric NO 2 amount is likely. Sensitivity calculations are performed here with the MLWT2 model, which was presented in Part 1. For Z 5 608 (air mass 2), this results in 0.5% error in DNI for background pollution (0.5 matm cm of NO 2 ), 1.2% error for moderate pollution (2 matm cm of NO 2 ), and 9% error for heavy pollution (20 matm cm of NO 2 ). Previous investigations have demonstrated the important role of precise aerosol information and correct aerosol transmittance modelling in the prediction of solar irradiance (Gueymard, 1993a; Olmo et al., 2001). Under most cloudless conditions, the main source of error in predicted DNI results from inaccuracies in turbidity. These are essentially a direct function of the product mb, where b turbidity coefcient. The aerosol transis the Angstrom mittance of the model of Yang et al. (2001)

Fig. 1. Predicted DNI for reference atmospheric conditions, as a function of b, w and m.

384

C. A. Gueymard / Solar Energy 74 (2003) 381395

2.2. Overall error analysis


All the sources of uncertainty reviewed above must be combined to obtain the overall uncertainty in modelled DNI. Five important assumptions are retained here: (i) DNI can be obtained as the extraterrestrial irradiance, E0n , multiplied by a product of six individual broadband transmittances: Ebn 5 E0n T Rg T o T ns T nt T w T a (4)

(ii) the transmittance for each extinction process i is expressed as exp (2m i ti )as in the MLWT and REST models described in Part 1where ti is the corresponding optical depth and m i the specic optical mass for extinction i ; (iii) Z is not too large so that a common optical mass can be used for all processes (m i m); (iv) the solar position is known exactly at any moment, so that m is perfectly determined; and nally (v) ideally perfect transmittance functions are used. The overall uncertainty in Ebn resulting from imperfect input data can thus be obtained as DEbn ]] Ebn 5

DNI is a strong function of air mass, and is only comparable to the uncertainty in measured data when m is less than about 25 with the optimistic scenario, depending on conditions. The uncertainty in individual predictions of DNI becomes very large with the pessimistic scenario, reaching 22% for m 5 2 and 75% for m 5 5. The most part of this total uncertainty can be traced back to the effect of turbidity. The large uncertainty difference between the optimistic and pessimistic scenarios, and the magnitude of the possible errors associated with the latter, are two aspects of DNI prediction under clear skies that seem to have been understated in the literature so far. The major conclusion that can be drawn from this sensitivity study and error analysis is that, in order to experimentally evaluate the performance of DNI models with sufcient condence, it is absolutely necessary to rely only on the optimistic scenario. More specically, the availability of reliable, accurate and collocated turbidity data are absolutely essential in this kind of exercise. Because such a circumstance is not frequent, it is more difcult to perform a detailed and signicant performance assessment study when based on measured data than on reference modelled data (as in Part 1).

tn 1 ] Du n u n

HS D FS D S S D S D S
DE0n ]] E0n
2

1 m2
2

tRg ]] Dp p

to 1 ] Du o u o ta ] D b

tw ] Dw w

D bD GJ
2 2

3. Instrumentation and measured datasets Because of the stringent requirements accompanying the optimistic scenario expressed above, only sites equipped to continuously monitor DNI, precipitable water and turbidity can be useful candidates for the present performance study. Five sites meeting this criterion have been selected: Barrow (Alaska), Bondville (Illinois), Billings (Oklahoma), Mauna Loa (Hawaii), and Solar Village (Saudi Arabia). Further information on these sites is provided in Table 2, showing the large range of geographic locations, altitudes, and climates this dataset covers. All the selected sites are automatically monitoring broadband and spectral solar radiation on a routine basis. Direct irradiance is normally measured with an Eppley NIP, whereas spectral radiation is normally measured with a collocated Cimel CE-318 sunphotometer, as part of the AERONET world network (Holben et al., 1998). Processed data for the latter instrument are available from the AERONET website (http: / / aeronet.gsfc.nasa.gov). Rather than using the regular AERONET data for Billings, the dataset used here for this site has been obtained during the Intensive Operational Period (IOP) that took place at the Cloud And Radiation Test (CART) bed of the Southern Great Plains (SGP) site of the Atmospheric Radiation Measurement (ARM) Program during the Fall of 1997 (hereafter, IOP97). Although radiative measurements are measured at the CART site on a routine basis, the specic IOP97 datasets are of exceptional quality for three important reasons: (i) a standardclass active cavity radiometer (ACR), with only 0.5%

(5) The extraterrestrial irradiance is currently known within 2 about 3 W/ m or 0.2%, taking into account slight variations ( 0.1%) due to the 11-year suns cycle and larger day-to-day variations due to short-term activity. This small uncertainty is only second-order compared to most other terms in the right-hand side of Eq. (5). The error analysis above can be used to evaluate the uncertainty in Ebn for two extreme cases: (i) the optimistic scenario, which occurs when collocated pressure, ozone, and sunphotometric or spectroradiometric data (providing precipitable water and turbidity) are available and no time interpolation is necessary; and (ii) the pessimistic scenario, when no such data are directly available, and time / space interpolation is necessary in the estimation of all inputs. Sample results for this exercise are displayed in Table 1 for three air masses: 1, 2 and 5, and for typical values of the absolute errors in the input data, Dp, Du o , Du n , Dw and Db. A site with very low tropospheric NO 2 concentration is assumed for simplicity, so that u n reduces to a typical 24 stratospheric NO 2 amount (2 3 10 atm cm). To evaluate the partial derivatives in Eq. (5), Choudhurys model has been used for the Rayleigh, mixed gas, ozone and water vapor extinctions, Yangs model for the aerosol extinction, and MLWT2 for NO 2 . The results in Table 1 indicate that the uncertainty in

C. A. Gueymard / Solar Energy 74 (2003) 381395

Table 1 Total uncertainty in direct irradiance, DEbn / Ebn (%), resulting from the combination of individual errors in the different input values, per Eq. (5); Du n is xed at 0.0002 atm cm in all cases
Case Unit Optimistic m 51 m 52 m 55 Pessimistic m 51 m 52 m 55 DE0n / E0n tRg / p (mb 2 1 ) to / u o (atm cm)2 1 tn / u n (atm cm)2 1 tw / w (cm 2 1 ) ta / b p (mb) uo (atm cm) un (atm cm) w (cm)

Dp (mb)

Duo (atm cm)

Dw (cm)

Db

DEbn / Ebn (%)

0.002 0.002 0.002

2 7.48E 2 05 2 1.16E 2 04 2 1.62E 2 04

2 0.0372 2 0.0484 2 0.0687

2 3.207 2 6.295 2 14.901

2 0.0225 2 0.0277 2 0.0364

2 1.321 2 2.130 2 2.956

1000 1000 1000

0.3 0.3 0.3

2E 2 04 2E 2 04 2E 2 04

1.5 1.5 1.5

0.1 0.1 0.1

5 5 5

0.01 0.01 0.01

0.15 0.15 0.15

0.005 0.005 0.005

0.8 2.3 8.0

0.002 0.002 0.002

2 7.48E 2 05 2 1.16E 2 04 2 1.62E 2 04

2 0.0372 2 0.0484 2 0.0687

2 3.207 2 6.295 2 14.901

2 0.0225 2 0.0277 2 0.0364

2 1.321 2 2.130 2 2.956

1000 1000 1000

0.3 0.3 0.3

2E 2 04 2E 2 04 2E 2 04

1.5 1.5 1.5

0.1 0.1 0.1

10 10 10

0.025 0.025 0.025

0.75 0.75 0.75

0.05 0.05 0.05

6.8 21.7 75.2

385

386

C. A. Gueymard / Solar Energy 74 (2003) 381395

Table 2 Experimental sites and their instrumentation. Latitudes and longitudes are in degrees, west longitudes are negative. Pressure and total vertical ozone are measured on site if a dot ( ) is indicated. Instrument key: ACR, active cavity radiometer; NIP, normal incidence pyrheliometer; SPM, sunphotometer Site Latitude Longitude Elevation (m) 10 290 290 213 3398 650 Period Pressure Ozone SPM Pyrheliometer DNI timestep (min) 1 0.5 0.5 3 1 5

Barrow Billings-A Billings-C Bondville Mauna Loa Solar village

71.31 36.61 36.61 40.06 19.54 24.91

2 156.67 2 97.49 2 97.49 2 88.37 2 155.58 46.41

78 / 99 910 / 97 910 / 97 4 / 006 / 01 16 / 99 2 / 992 / 02

Cimel AATS-6 Cimel Cimel Cimel Cimel

NIP ACR ACR NIP NIP NIP

uncertainty, was deployed, rather than just the regular thermopile instruments; (ii) datasets from two multi-wavelength sunphotometers were availablea Cimel CE-318 (Halthore et al., 1999) and a NASA AATS-6 (Matsumoto et al., 1987); and (iii) different methods of measuring precipitable water were continuously available during this 2-week period, namely radiosondes, two microwave radiometers (MWR), a global positioning system (GPS), a rotating shadowband spectroradiometer (RSS), a multi-

Fig. 2. Precipitable water measurements by different instruments at the ARM-CART site of Billings, Oklahoma during a clear day of IOP97.

lter rotating shadowband radiometer (MFRSR), and the two multi-wavelength sunphotometers already mentioned. This exceptional redundancy of data allowed precise intercomparisons of water vapor retrievals (Michalsky et al., 1999; Schmid et al., 2001; Revercomb et al., 2003). Based on these various IOP97 datasets (available online at http: / / www.arm.gov), Fig. 2 shows the time series of measured precipitable water from some of the instruments involved on September 28, 1997, chosen here as a typical day. The strong diurnal variation in precipitable water during this clear day is remarkable. Similar daily gradients have been found frequently during this IOP or other periods, as well as at other locations (Halthore et al., 1997), thus conrming the earlier statement that highresolution precipitable water data are needed to obtain the highest performance with DNI models. Based on both the intercomparison presented here and detailed discussions proposed elsewhere (Halthore et al., 1997; Schmid et al., 2001; Bokoye et al., 2003), it is concluded that the determinations of precipitable water obtained from the Cimel instrument with the algorithm of Bruegge et al. (1992), as used routinely by AERONET to retrieve w, should be uniformly decreased by 10% for best results. This correction has been applied to all the Cimelbased data used here, i.e. for all benchmark datasets except Billings-A (see below). Only AERONET data that passed the automatic cloudscreening and Level 2 quality-control processes (Smirnov et al., 2000) have been used here. They have then been subjected to a supplementary manual inspection, which removed a few outliers. Precipitable water and aerosol data for all sites under scrutiny here, except Billings, have been obtained from the AERONET website. Because of the availability of two separate sunphotometric datasets at Billings, two different performance tests were performed there, and will be referred to as Billings-A for the AATS-6 instrument, and Billings-C for the Cimel instrument. Spectral aerosol optical depth (AOD) is retrieved at ve wavelengths (0.38, 0.451, 0.525, 0.864, and 1.021 mm) by the AATS-6 instrument and at seven wavelengths (0.34, 0.38, 0.44, 0.50, 0.67, 0.87, and 1.02 mm) by

C. A. Gueymard / Solar Energy 74 (2003) 381395

387

Table 3 Cumulative statistics for the measured data from all sites. The stratospheric and tropospheric NO 2 columnar amounts are uniformly xed at 0.15 and 0 matm cm, respectively; the number of experimental data points is indicated in the last column Site Barrow m w (cm) u o (atm cm) a b Ebn (W/ m 2 ) Billings-A (AATS6) m w (cm) u o (atm cm) a b Ebn (W/ m 2 ) Billings-C (Cimel) m w (cm) u o (atm cm) a b Ebn (W/ m 2 ) Bondville m w (cm) u o (atm cm) a b Ebn (W/ m 2 ) Mauna Loa m w (cm) u o (atm cm) a b Ebn (W/ m 2 ) Solar Village m w (cm) u o (atm cm) a b Ebn (W/ m 2 ) All m w (cm) u o (atm cm) a b Ebn (W/ m 2 ) 2.631 1.853 0.283 1.314 0.038 796.6 2.140 1.769 0.275 1.322 0.027 811.2 1.553 1.243 0.030 0.053 0.035 173.8 3.020 2.052 0.285 1.310 0.034 676.4 2.893 1.928 0.289 1.329 0.035 675.3 1.101 0.315 0.010 0.054 0.013 121.7 2676 2.730 2.602 0.275 1.327 0.033 754.7 2.090 2.512 0.276 1.319 0.040 796.8 2.438 1.505 0.327 1.298 0.071 728.2 2.712 0.227 0.263 1.303 0.010 990.9 2.271 0.868 0.265 1.291 0.048 849.3 1.962 2.529 0.273 1.330 0.027 777.1 1.697 2.581 0.276 1.326 0.031 804.3 2.310 1.161 0.319 1.298 0.057 755.5 2.716 0.196 0.278 1.303 0.008 994.5 1.746 0.542 0.264 1.282 0.028 870.6 1.854 0.887 0.004 0.045 0.014 148.4 76 0.860 0.847 0.004 0.058 0.021 119.4 1021 1.060 1.040 0.034 0.058 0.053 165.9 871 1.112 0.138 0.029 0.058 0.008 101.1 294 1.071 0.665 0.016 0.060 0.046 175.6 4960 Variable Average Median St. Dev. N 22

the Cimel sunphotometer. For the present study, the a and b coefcients have been obtained by a Angstrom least-squares t of these experimental AODs to Angstroms equation after linearization: ln ta l 5 ln b 2 a ln l (6)

where ta l is the AOD at wavelength l, as retrieved from the sunphotometer. All channels between 0.38 and 1.021 mm have been used in these ts, i.e. ve channels for the AATS-6 and six channels for the Cimel. Only data for which 1.2 , a , 1.4 have been retained, to approach the Angstrom-recommended value a 5 1.3, in conformity with the assumptions detailed in Part 1. This restriction eliminated a considerable number of data points at Barrow, where a is very frequently . 2 due to a predominance of small continental particles, and at Solar Village, where a is usually , 1 because of the predominance of large sand particles. A discussion of the performance of the Cimel and AATS-6 instruments, with particular reference to the IOP97 data, is provided elsewhere (Schmid et al., 1999). Depending on site, DNI data were available at 0.5- to 5-min intervals (Table 2). The AATS-6 measurement interval was 0.2 min, so that the DNI data and the AATS-6 data could be matched within about 0.1 min for Billings-A. For all other sites, data from the Cimel instrument were recorded at a much lower frequency (normally 15 min), and the DNI data were sampled to retain only those within 0.5 min of the Cimel data. These close matches guarantee that the performance results are normally not affected by undetected changes in solar position or atmospheric conditions. As indicated in Table 2, pressure was measured at all sites except Solar Village. As discussed in Section 2.1, pressure was estimated there from altitude and latitude (Gueymard, 1993b). At Barrow and Mauna Loa, ozone was routinely measured with spectrophotometers of the WOUDC network. For all the other sites, where no in situ measurements were available, the ozone vertical column was obtained from TOMS overhead data (http: / / jwocky.gsfc.nasa.gov / teacher / ozone]overhead.html). Finally, a xed value of 0.15 matm cm was considered for the amount of stratospheric NO 2 in the vertical column. No tropospheric NO 2 was considered because all sites are away from urban areaseven though occasional pollution plumes are always a possibility, particularly over Bondville. Table 3 groups the measurement statistics for the six experimental datasets under scrutiny.

4. Performance results The same performance statistics that were employed in Table 9 of Part 1 (i.e. MBE, RMSE, k and R ) are used

388

C. A. Gueymard / Solar Energy 74 (2003) 381395

Table 4 Performance statistics of all models at Barrow. MBE and RMSE are expressed in percent of the average measured DNI (see Table 2); k and R are the slope and correlation coefcient of the linear t between modelled and measured values, respectively Model Bird CEM Choudhury CPCR2 Dogniaux Iqbal C King and Buckius METSTAT MLWT1 MLWT2 MMAC Molineaux MRM Perrin Power Psiloglou PSIM REST Rodgers Santamouris Yang Average 682.3 711.8 712.7 690.9 701.5 688.4 737.0 710.4 707.2 692.7 669.2 718.5 377.3 732.0 700.8 436.2 697.6 702.6 740.2 675.7 711.8 Median 684.8 713.5 710.2 689.0 691.0 690.9 740.9 713.0 700.4 686.1 667.7 714.2 357.4 736.5 691.1 419.4 700.2 700.3 740.8 677.9 711.9 St. Dev. 128.9 138.1 130.9 122.0 115.3 130.2 119.8 126.5 124.5 127.2 132.5 114.7 58.0 126.5 130.9 131.9 116.4 125.6 127.1 141.5 126.5 Min. 507.8 520.5 536.2 527.2 552.6 512.3 572.5 539.3 541.3 521.7 488.9 574.9 298.6 553.1 517.9 251.2 540.4 534.1 570.6 484.7 542.2 Max. 842.1 891.0 886.9 850.4 861.8 849.7 887.6 870.7 876.6 866.3 845.7 870.4 490.2 892.6 878.8 643.0 856.1 868.1 903.8 851.7 875.6 MBE (%) 0.9 5.2 5.4 2.1 3.7 1.8 8.9 5.0 4.6 2.4 2 1.1 6.2 2 44.2 8.2 3.6 2 35.5 3.1 3.9 9.4 2 0.1 5.2 RMSE (%) 2.8 6.0 5.7 2.7 4.0 3.3 9.2 5.5 4.7 2.8 2.4 6.6 45.7 8.4 4.2 35.8 3.5 4.3 9.7 3.9 5.6 k 1.010 1.055 1.054 1.021 1.034 1.019 1.086 1.050 1.045 1.025 0.992 1.058 0.553 1.081 1.037 0.657 1.029 1.038 1.092 1.003 1.052 R 0.990 0.994 0.997 0.995 0.993 0.990 0.986 0.992 0.997 0.998 0.993 0.984 0.760 0.994 0.996 0.898 0.994 0.995 0.991 0.982 0.994

here, where MBE, RMSE, k and R are the mean bias error, the root mean square error, and the slope and regression coefcient of the linear t between predicted and measured

values, respectively. These results are displayed in Tables 49 for the six experimental datasets separately, and in Table 10 for all datasets combined.

Table 5 Same as Table 4 but for Billings-A Model Bird CEM Choudhury CPCR2 Dogniaux Iqbal C King and Buckius METSTAT MLWT1 MLWT2 MMAC Molineaux MRM Perrin Power Psiloglou PSIM REST Rodgers Santamouris Yang Average 756.5 790.8 790.0 762.0 781.4 769.0 803.0 783.8 778.7 765.8 748.6 797.6 396.7 801.9 750.7 519.3 760.3 777.3 818.6 754.7 786.5 Median 778.5 822.1 814.9 783.4 802.2 791.7 826.0 806.8 799.2 788.4 776.3 813.0 385.0 831.2 752.8 577.9 781.8 798.2 841.1 778.5 808.4 St. Dev. 142.5 162.5 149.1 138.7 131.8 142.7 136.3 140.3 143.7 147.5 152.4 123.1 72.0 148.2 162.5 170.7 135.2 142.6 144.3 155.8 140.9 Min. 309.2 233.5 309.6 325.3 360.3 320.8 340.7 336.1 332.5 310.1 259.4 441.6 283.2 292.1 277.4 87.9 329.2 322.8 267.1 269.8 343.0 Max. 928.3 975.0 965.6 929.9 941.3 940.5 957.8 949.5 953.8 943.8 927.9 958.1 889.2 964.6 969.5 724.5 922.2 949.4 986.6 942.7 956.4 MBE (%) 0.2 4.8 4.7 1.0 3.5 1.9 6.4 3.9 3.2 1.5 2 0.8 5.7 2 47.4 6.3 2 0.5 2 31.2 0.7 3.0 8.5 0.0 4.2 RMSE (%) 1.2 5.4 4.7 1.7 4.2 2.3 6.7 4.1 3.3 1.6 1.4 6.7 49.5 6.5 4.0 32.0 2.1 3.1 8.6 1.3 4.4 k 1.001 1.049 1.045 1.007 1.030 1.017 1.059 1.035 1.029 1.014 0.993 1.048 0.519 1.060 0.998 0.703 1.004 1.027 1.080 1.002 1.039 R 0.998 0.996 0.999 0.996 0.986 0.997 0.985 0.994 0.997 0.999 0.998 0.959 0.646 0.993 0.983 0.884 0.993 0.997 0.990 0.998 0.995

C. A. Gueymard / Solar Energy 74 (2003) 381395 Table 6 Same as Table 4 but for Billings-C Model Bird CEM Choudhury CPCR2 Dogniaux Iqbal C King and Buckius METSTAT MLWT1 MLWT2 MMAC Molineaux MRM Perrin Power Psiloglou PSIM REST Rodgers Santamouris Yang Average 790.2 833.3 829.2 796.2 813.2 803.0 836.3 817.9 815.5 804.4 788.7 825.2 421.2 840.7 794.7 579.9 795.1 813.1 853.6 791.6 821.2 Median 802.4 850.8 841.5 808.5 822.3 815.5 842.5 831.7 822.8 816.1 806.1 839.4 408.0 855.0 779.9 625.3 805.0 823.7 864.9 805.2 833.9 St. Dev. 114.1 121.6 118.5 112.6 109.5 113.3 105.7 111.9 119.2 120.6 119.6 108.6 83.4 108.7 139.0 132.6 108.6 116.3 113.3 125.7 114.7 Min. 579.0 591.3 600.5 586.0 608.6 591.9 632.8 607.5 587.9 574.2 555.0 613.3 267.5 623.6 551.3 276.2 591.8 595.5 641.1 559.5 608.8 Max. 962.2 1011.3 1006.0 965.5 984.3 973.5 994.2 984.3 997.4 987.4 966.7 985.8 578.3 999.5 1036.1 739.5 954.1 990.4 1023.0 981.5 994.5 MBE (%) 2 0.8 4.6 4.1 2 0.1 2.1 0.8 5.0 2.6 2.3 1.0 2 1.0 3.6 2 47.1 5.5 2 0.3 2 27.2 2 0.2 2.0 7.1 2 0.7 3.1 RMSE (%) 1.5 4.7 4.1 1.2 2.6 1.5 5.3 3.0 2.4 1.2 1.4 4.0 47.5 5.7 4.5 28.3 1.7 2.2 7.2 1.5 3.2 k 0.991 1.045 1.040 0.998 1.018 1.006 1.046 1.024 1.023 1.009 0.990 1.033 0.531 1.052 1.000 0.733 0.996 1.019 1.069 0.995 1.029 R

389

0.997 0.998 0.997 0.997 0.991 0.995 0.980 0.993 0.999 0.999 0.998 0.987 0.926 0.986 0.966 0.854 0.993 0.997 0.989 0.996 0.995

Despite the large range of climatic conditions covered here, all results are in close agreement from one site to the other. It should be noticed, however, that the best results are obtained at Billings (where an ACR and other labora-

tory-grade instrumentation were used) and that the largest differences for most models are obtained at Bondville (where turbidity is frequently and noticeably larger than at other sites).

Table 7 Same as Table 4 but for Bondville Model Bird CEM Choudhury CPCR2 Dogniaux Iqbal C King and Buckius METSTAT MLWT1 MLWT2 MMAC Molineaux MRM Perrin Power Psiloglou PSIM REST Rodgers Santamouris Yang Average 690.2 741.4 743.9 712.4 747.7 700.2 754.9 723.6 741.7 729.5 704.2 737.4 504.4 751.5 722.5 544.3 714.0 726.8 758.4 685.0 730.8 Median 715.9 773.7 776.4 739.0 787.6 727.2 792.5 754.1 773.3 756.8 732.5 769.2 508.8 785.5 741.8 547.7 738.3 755.3 793.0 711.6 757.4 St. Dev. 175.6 188.1 184.9 170.8 171.0 176.0 174.4 179.5 178.5 178.1 182.4 175.7 121.3 178.6 177.7 158.0 171.3 177.4 183.3 190.5 179.1 Min. 109.0 119.0 117.4 165.9 161.5 115.8 170.8 115.4 168.0 166.5 114.0 127.1 209.2 124.7 176.7 246.7 165.1 160.6 126.1 81.9 148.7 Max. 963.9 1018.2 1015.4 973.0 1002.1 974.2 1003.7 990.0 1008.8 999.7 977.5 989.0 756.6 1006.8 1053.7 791.9 967.2 999.0 1025.4 985.2 1004.5 MBE (%) 2 5.2 1.8 2.2 2 2.2 2.7 2 3.8 3.7 2 0.6 1.9 0.2 2 3.3 1.3 2 30.7 3.2 2 0.8 2 25.3 2 2.0 2 0.2 4.1 2 5.9 0.4 RMSE (%) 6.5 4.9 4.8 3.9 5.4 5.5 5.4 4.0 4.2 3.5 5.2 4.4 33.9 5.0 7.2 28.9 4.1 3.7 6.1 7.9 3.9 k 0.953 1.023 1.026 0.980 1.026 0.966 1.037 0.998 1.021 1.005 0.973 1.014 0.687 1.034 0.994 0.748 0.982 1.001 1.044 0.950 1.007 R 0.983 0.986 0.987 0.989 0.980 0.984 0.987 0.987 0.989 0.990 0.985 0.985 0.756 0.989 0.956 0.798 0.988 0.989 0.987 0.974 0.988

390 Table 8 Same as Table 4 but for Mauna Loa Model Bird CEM Choudhury CPCR2 Dogniaux Iqbal C King and Buckius METSTAT MLWT1 MLWT2 MMAC Molineaux MRM Perrin Power Psiloglou PSIM REST Rodgers Santamouris Yang Average 944.4 1012.2 1010.3 960.5 1072.2 978.2 1008.8 974.1 1016.2 995.3 981.3 994.9 747.2 994.8 947.2 570.0 887.8 992.0 1017.6 964.0 996.9

C. A. Gueymard / Solar Energy 74 (2003) 381395

Median 948.9 1013.9 1011.2 962.8 1075.7 979.8 1018.4 977.9 1020.2 998.2 982.2 992.4 749.1 1002.1 926.5 534.3 882.6 993.9 1021.8 970.5 999.2

St. Dev. 102.1 92.0 93.9 98.6 69.8 93.8 84.7 98.1 95.7 101.9 96.4 80.2 98.2 93.4 168.2 170.0 108.8 102.9 79.3 112.1 109.0

Min. 445.2 587.5 598.3 489.5 721.2 553.3 532.0 475.8 547.2 521.7 560.5 608.9 473.0 517.9 521.4 290.7 446.9 497.8 615.3 411.8 491.0

Max. 1150.8 1201.2 1209.3 1171.1 1214.5 1174.2 1189.5 1177.8 1220.4 1213.6 1186.7 1157.7 997.3 1191.3 1268.2 891.8 1119.6 1208.0 1160.9 1187.9 1235.3

MBE (%) 2 4.7 2.2 2.0 2 3.1 8.2 2 1.3 1.8 2 1.7 2.6 0.4 2 1.0 0.4 2 24.6 0.4 2 4.4 2 42.5 2 10.4 0.1 2.7 2 2.7 0.6

RMSE (%) 4.9 2.7 2.5 3.2 8.9 1.9 2.8 2.0 2.8 1.2 1.8 2.6 25.6 1.3 9.6 43.6 10.7 1.1 3.8 3.3 1.5

k 0.954 1.020 1.019 0.969 1.078 0.987 1.016 0.983 1.025 1.004 0.990 1.002 0.754 1.003 0.962 0.584 0.898 1.001 1.024 0.974 1.007

R 0.990 0.983 0.984 0.995 0.806 0.991 0.965 0.994 0.992 0.995 0.988 0.949 0.751 0.991 0.856 0.690 0.968 0.994 0.930 0.984 0.993

A slight difference may be noticed between the results for Billings-A and Billings-C. The time resolution effect should have favoured the Billings-A dataset because of its far larger number of data points lowering the associated

random error (expressed by RMSE). Therefore, the slightly larger RMSE exhibited by most models at Billings-A compared to Billings-C can be only attributed to the small difference in turbidity and precipitable water values mea-

Table 9 Same as Table 4 but for Solar Village Model Bird CEM Choudhury CPCR2 Dogniaux Iqbal C King and Buckius METSTAT MLWT1 MLWT2 MMAC Molineaux MRM Perrin Power Psiloglou PSIM REST Rodgers Santamouris Yang Average 821.9 881.0 885.2 841.1 895.2 839.2 884.7 856.1 879.1 865.9 843.3 869.1 621.2 881.5 858.3 608.9 826.6 863.6 894.0 827.5 867.8 Median 838.4 901.8 905.0 857.0 912.7 854.9 908.2 875.7 898.0 880.9 862.0 891.2 636.1 900.5 845.7 676.7 830.6 881.5 918.1 845.7 882.7 St. Dev. 176.0 179.4 176.8 171.6 169.5 173.3 170.9 175.5 177.2 179.0 178.8 162.0 138.3 171.2 202.7 171.4 168.5 178.3 170.3 193.4 182.0 Min. 280.8 333.6 317.8 317.9 334.3 303.9 329.6 309.0 330.4 329.9 315.1 310.6 314.8 326.3 281.4 272.0 318.8 326.8 342.6 251.6 321.2 Max. 1063.8 1115.4 1117.5 1074.6 1118.0 1076.7 1111.1 1089.3 1117.6 1110.0 1083.3 1075.2 837.4 1107.3 1164.3 822.6 1056.2 1105.2 1107.5 1093.6 1120.9 MBE (%) 2 3.2 3.7 4.2 2 1.0 5.4 2 1.2 4.2 0.8 3.5 2.0 2 0.7 2.3 2 26.8 3.8 1.1 2 28.3 2 2.7 1.7 5.3 2 2.6 2.2 RMSE (%) 3.4 4.0 4.4 1.5 6.1 1.7 4.5 1.4 3.7 2.2 1.5 3.3 27.9 4.0 6.6 31.9 3.2 2.0 5.5 3.5 2.6 k 0.969 1.037 1.041 0.990 1.050 0.988 1.039 1.008 1.034 1.020 0.994 1.019 0.732 1.035 1.015 0.717 0.973 1.017 1.049 0.979 1.022 R 0.998 0.998 0.997 0.999 0.987 0.998 0.994 0.999 0.998 0.999 0.998 0.990 0.946 0.996 0.963 0.741 0.996 0.999 0.994 0.992 0.999

C. A. Gueymard / Solar Energy 74 (2003) 381395 Table 10 Same as Table 4 but for all sites combined Model Bird CEM Choudhury CPCR2 Dogniaux Iqbal C King and Buckius METSTAT MLWT1 MLWT2 MMAC Molineaux MRM Perrin Power Psiloglou PSIM REST Rodgers Santamouris Yang Average 779.9 825.2 825.1 791.6 832.4 795.9 834.3 809.3 819.0 804.8 786.2 824.2 494.0 830.4 786.2 539.2 777.3 809.9 845.8 781.7 817.0 Median 800.6 851.7 846.4 809.0 839.0 815.4 853.8 830.8 834.6 821.2 808.2 842.0 450.8 854.9 793.2 563.0 794.5 825.7 866.9 803.5 833.3 St. Dev. 167.6 183.5 176.2 164.4 177.4 170.9 163.9 166.8 175.0 175.8 179.6 157.0 162.0 169.1 186.8 169.5 152.3 171.7 169.0 183.4 171.9 Min. 109.0 119.0 117.4 165.9 161.5 115.8 170.8 115.4 168.0 166.5 114.0 127.1 209.2 124.7 176.7 87.9 165.1 160.6 126.1 81.9 148.7 Max. 1150.8 1201.2 1209.3 1171.1 1214.5 1174.2 1189.5 1177.8 1220.4 1213.6 1186.7 1157.7 997.3 1191.3 1268.2 891.8 1119.6 1208.0 1160.9 1187.9 1235.3 MBE (%) 2 2.1 3.6 3.6 2 0.6 4.5 2 0.1 4.7 1.6 2.8 1.0 2 1.3 3.5 2 38.0 4.2 2 1.3 2 32.3 2 2.4 1.7 6.2 2 1.9 2.6 RMSE (%) 3.9 4.6 4.3 2.7 6.1 3.0 5.6 3.5 3.4 2.0 2.6 5.3 40.7 5.2 6.7 35.2 6.1 2.8 7.0 3.9 3.6 k 0.978 1.036 1.034 0.991 1.043 0.998 1.042 1.013 1.027 1.010 0.989 1.028 0.624 1.039 0.989 0.681 0.970 1.015 1.057 0.984 1.024 R

391

0.988 0.993 0.994 0.993 0.981 0.990 0.985 0.988 0.996 0.997 0.995 0.974 0.755 0.987 0.960 0.781 0.962 0.994 0.983 0.988 0.992

sured by the two sunphotometers, or to the difference in solar positions (Billings-A has relatively more points with large air masses than Billings-C). This illustrates the difculty of assessing the performance of radiation models using measured data, even in the present case of the most optimistic scenario possible. A comparison between Table 9 in Part 1 and Tables 410 here shows that most models seem to perform better when experimentally tested than when theoretically tested. This can be explained by the far larger span of solar positions and atmospheric conditions contemplated by the theoretical tests, which thus can be considered more stringent. Nevertheless, the best performers appear to be the same in both series of tests. From Table 10, the seven best models from Table 9 in Part 1 (CPCR2, METSTAT, MLWT1, MLWT2, MMAC, REST, and Yang) also produce an experimental MBE lower than 63% and an RMSE lower than 4%. However, three more models also meet this criterion: Bird, Iqbal C, and Santamouris. But these three models did not perform as well as the former seven in the individual transmittance tests (Tables 48 of Part 1), thus conrming that the present experimental conditions are less severe than the theoretical conditions of Part 1. The four recommended models from the theoretical assessment (CPCR2, MLWT2, REST and Yang) can also be recommended from this experimental assessment. The absolute differences between their predicted irradiance and the measured irradiance are shown in Fig. 3 for the 2752 data points of the complete IOP97 dataset at Billings,

selected here for its superior quality. Results for MLWT2, the best performer of this study, show a atter trend than those of CPCR2, which was the best performer in a previous study (Gueymard, 1993a). REST shows a trend that is intermediate between those of CPCR2 and MLWT2, whereas Yangs model has the same trend as CPCR2, but with more bias. Figs. 47 present scatterplots for the same four best models, but when considering the whole experimental dataset from all sites. For each model, the apparent mean bias error and associated standard deviation has also been calculated in bins of 50 W/ m 2 . This tool is indicative of the performance of the models for a variety of DNI regimes. As noted above for Billings, MLWT2 shows the lower error and the atter response, with a small, nearly constant bias of only 510 W/ m 2 over the range 300 1200 W/ m 2 . (Note that there are not many data points below 300 W/ m 2 , and that the performance of all models appear to degrade below this value; this might be explained by the larger air masses involved, and therefore the possible non-respect of the optimistic scenario described above. For this reason, the results for DNI , 300 W/ m 2 will not be considered in what follows.) CPCR2 also shows a relatively small and at bias over a wide range of DNI values, but tend to underestimate (by 2 30 W/ m 2 ) above 950 W/ m 2 . REST and Yang both show a tendency to slight and variable overestimation (by 520 W/ m 2 and 1035 W/ m 2 , respectively) overall. As indicated above, METSTAT and MMAC also demonstrate very good experimental performance. They can

392

C. A. Gueymard / Solar Energy 74 (2003) 381395

Fig. 3. Absolute differences between predicted DNI by four recommended models and measured data at the ARM-CART site of Billings, Oklahoma during IOP97. (a) CPCR2 and MLWT2 models; (b) REST and Yang models.

thus be recommendedbut under low turbidity conditions only, because of their limitations under turbid conditions that were discussed in Part 1. The results in Tables 410 also demonstrate that the best performing models can predict DNI within the experimental uncertainty inherent to eld pyrheliometersto the essential condition that accurate and frequent data of both precipitable water and turbidity are available. This nding is consistent with that in a previous investigation (Gueymard, 1993a). Nevertheless, this nding is more qualied and signicant now because of the greater accuracy and completeness of the experimental datasets now available, which reached the point where they can respect the requirements of the optimistic scenario described above.

5. Conclusion Using the same 21 irradiance models as in Part 1 of this investigation (Gueymard, 2003a), the main objective here was to determine if and how they could predict DNI within the experimental uncertainty of modern eld pyrheliometers. A sensitivity study and a detailed error analysis have shown that the uncertainty in the predicted DNI is very sensitive to errors in the two main inputsturbidity and precipitable waterand increases sharply with air mass. In those frequent applications where DNI is predicted from time / space interpolations or extrapolations of turbidity and precipitable water, it is highly likely that the uncertainty in the individual predictions is considerably higher than

C. A. Gueymard / Solar Energy 74 (2003) 381395

393

Fig. 4. Scatter plot (top panel) and error plot (bottom panel) for the predictions of CPCR2 at all sites vs. measured data. The bottom panel shows the apparent mean absolute prediction error and the associated standard deviation in bins of 50 W/ m 2 .

Fig. 5. Same as Fig. 4 but for predictions by the MLWT2 model.

previously acknowledged. It is therefore emphasized that, to obtain signicant results, any experimental performance assessment study must be conducted only at sites where accurate measurements of both irradiance and essential ancillary data are available. A series of experimental tests, using six datasets from ve sites, has been conducted according to the recommended practice just mentioned. First-class DNI measurements, obtained with high sampling rates (nearly 5000 data points total), were compared to the predictions of all 21 models, which were fed with collocated, high-quality sunphotometric retrievals of turbidity and precipitable water, as well as locally measured pressure and ozone wherever possible. Even though the ve sites have widely different climates, it is found that the models perform consistently from one site to the other. Their performance is also better (RMS , 4%) than with the theoretical tests (in Part 1), which thus appear more stringent than experimental tests due to the simulation of some more extreme or ideal atmospheric conditions. However, the experimental tests are found to globally conrm the results of the theoretical tests. From a practical standpoint, it can be concluded that appropriate theoretical tests can be used with condence to assess the performance of broadband radiation models, considering the far larger level of

Fig. 6. Same as Fig. 4 but for predictions by the REST model.

394

C. A. Gueymard / Solar Energy 74 (2003) 381395

Acknowledgements Daryl Myers of NREL and Ellsworth Dutton of NOAA were instrumental in kindly providing the radiation data for Solar Village, Barrow and Mauna Loa. The AERONET staff and participants are thanked for their successful effort in establishing and maintaining the Barrow, Bondville, Mauna Loa, and Solar Village sites whose optical depth data were advantageously used in this investigation. The ARM datasets were obtained from the Atmospheric Radiation Measurement (ARM) Program sponsored by the U.S. Department of Energy, Ofce of Science, Ofce of Biological and Environmental Research, Environmental Sciences Division.

References
Battles, F.J., Olmo, F.J., Tovar, J., Alados-Arboledas, L., 2000. Comparison of cloudless sky parameterizations of solar irradiance at various Spanish midlatitude locations. Theor. Appl. Climatol. 66, 8193. Bokoye A.I., Royer A., ONeill N.T., Cliche P., McArthur L.J. B., Teillet P.M., Fedosejevs G., Theriault J.M., 2003. Multi-sensor analysis of integrated atmospheric water vapour over Canada and Alaska. J. Geophys. Res. (in press). Bruegge, C.T., Conel, J.E., Green, R.O., Margolis, J.S., Holm, R.G., Toon, G., 1992. Water vapor column abundance retrieval during FIFE. J. Geophys. Res. 97D, 1875918768. Choudhury, B., 1982. A parameterized model for global insulation under partially cloudy skies. Solar Energy 29, 479486. Davies, J.A., McKay, D.C., Luciani, G., Abdel-Wahab, M., 1988. Validation of Models For Estimating Solar Radiation On Horizontal Surfaces. Iea Task Ix Final Report. Atmospheric Environment Service, Downsview. Gueymard, C., 1993a. Critical analysis and performance assessment of clear sky solar irradiance models using theoretical and measured data. Solar Energy 51, 121138. Gueymard, C., 1993b. Mathematically integrable parameterization of clear-sky beam and global irradiances and its use in daily irradiation applications. Solar Energy 50, 385397, Erratum: 51, 525 (1993). Gueymard, C., 1995. SMARTS2, Simple Model of the Atmospheric Radiative Transfer of Sunshine: Algorithms and performance assessment. Rep. FSEC-PF-270-95, Florida Solar Energy Center, Cocoa, FL. Gueymard, C., 2001. Parameterized transmittance model for direct beam and circumsolar spectral irradiance. Solar Energy 71, 325346. Gueymard, C., 2003a. Direct solar transmittance and irradiance predictions with broadband models. Part I: detailed theoretical performance assessment. Solar Energy (in press). Gueymard, C., 2003b. SMARTS code, version 2.9.2 Users Manual. Solar Consulting Services, Bailey, CO. Available from http: / / rredc.nrel.gov / solar / models / SMARTS / smarts index] .html. Halthore, R.N., Eck, T.F., Holben, B.N., Markham, B.L., 1997. Sun photometric measurements of atmospheric water vapor

Fig. 7. Same as Fig. 4 but for predictions by Yangs model.

difculty, effort, time and expenses involved in conducting careful experimental tests. In summary, four models (CPCR2, MLWT2, REST and Yang, in alphabetical order; including the corrections implemented here for the latter, see Part 1) can be recommended because of their consistently high performance in all theoretical and experimental tests. Remarkably, they can predict DNI under a variety of atmospheric conditions within the uncertainty of modern and wellmaintained pyrheliometers, provided that good quality inputs of precipitable water and turbidity are used, and that only low air masses are considered. Overall, the best results (very low, constant mean bias of only 510 W/ m 2 over all the DNI range) are obtained with MLWT2, as a result of the added capabilities introduced by its sophisticated multilayered-spectrum weighting scheme (see Appendix A of Part 1). No further improvements in current high-performance models appear necessary until more accurate experimental data become available. Finally, regular users of those models that have been found here to perform less favourably can now precisely evaluate the uncertainty in their predictions. To gain accuracy, they might elect to either replace a decient transmittance equation by a more robust one, or to change the whole model completely for one in the recommended list.

C. A. Gueymard / Solar Energy 74 (2003) 381395 column abundance in the 940-nm band. J. Geophys. Res. 102D, 43434352. Halthore, R.N., Schwartz, S.E., Liu, Y., Daum, P.H., Holben, B.N., Michalsky, J.J., 1999. Sun and sky radiometric measurements at the CART ARM SGP site. In: Proc. 9th ARM Science Team Meeting, San Antonio, TX, Atmospheric Radiation Measurement Program. US Department of Energy. Holben, B.N. et al., 1998. AERONETA federated instrument network and data archive for aerosol characterization. Remote Sens. Environ. 66, 116. Kambezidis, H.D., Adamopoulos, A.D., Gueymard, C., 2001. Total NO 2 column amount over Athens, Greece in 199697. Atmos. Res. 57, 18. Matsumoto, T., Russell, P.B., Mina, C., Van Ark, W., Banta, V., 1987. Airborne tracking sunphotometer. J. Atmos. Ocean. Technol. 4, 336339. Michalsky, J.J., Min, Q.L., Kiedron, P.W., Slater, D.W., Barnard, J.C., 1999. Column water vapor retrievals based on rotating shadowband spectroradiometer (rss) direct solar measurements. In: Proc. 9th ARM Science Team Meeting, San Antonio, TX, Atmospheric Radiation Measurement Program. US Department of Energy. Olmo, F.J., Vida, J., Foyo-Moreno, I., Tovar, J., Alados-Arboledas, L., 2001. Performance reduction of solar irradiance

395

parametric models due to limitations in required aerosol data: case of the CPCR2 model. Theor. Appl. Climatol. 69, 253263. Revercomb, H.E. et al., 2003. The ARM programs water vapor intensive observation periods. Bull. Amer. Meteorol. Soc. 84, 217236. Schmid, B., Michalsky, J., Halthore, R., Beauharnois, M., Harrison, L., Livingston, J., Russell, P., Holben, B., Eck, T., Smirnov, A., 1999. Comparison of aerosol optical depth from four solar radiometers during the Fall 1997 ARM Intensive Operational Period. Geophys. Res. Lett. 26, 27252728. Schmid, B., Michalsky, J.J., Slater, D.W., Barnard, J.C., Halthore, R.N., Liljegren, J.C., Holben, B.N., Eck, T.F., Livingston, J.M., Russell, P.B., Ingold, T., Slutsker, I., 2001. Comparison of columnar water-vapor measurements from solar transmittance methods. Appl. Opt. 40, 18861896. Schroeder, R., Davies, J.A., 1987. Signicance of nitrogen dioxide in estimating aerosol optical depth and size distributions. Atmos. Ocean 25, 107114. Smirnov, A., Holben, B.N., Eck, T.F., Dubovik, O., Slutsker, I., 2000. Cloud screening and quality control algorithms for the AERONET database. Rem. Sens. Environ. 73, 337349. Yang, K., Huang, G.W., Tamai, N., 2001. A hybrid model for estimating global solar radiation. Solar Energy 70, 1322.

You might also like