• Breitkreuz, H., , M. Schroedter-Homscheid, , T. Holzer-Popp, , and S. Dech, 2009: Short-range direct and diffuse irradiance forecasts for solar energy applications based on aerosol chemical transport and numerical weather modeling. J. Appl. Meteor. Climatol., 48, 17661779.

    • Search Google Scholar
    • Export Citation
  • Dare, R. A., 2004: The BMRC Bulk Explicit Microphysics (BEM) scheme. Australian Bureau of Meteorology Tech. Rep. 99, 110 pp.

  • Grant, I., , D. Jones, , W. Wang, , R. Fawcett, , and D. Barratt, 2008: Meteorological and remotely sensed datasets for hydrological modelling: A contribution to the Australian Water Availability Project. Extended Abstracts, CAHMDA-III Int. Workshop on Hydrologic Prediction: Modelling, Observation and Data Assimilation, Melbourne, VIC, Australia, Australian Research Council, 23–51.

  • International Organization for Standardization/International Electrotechnical Commission, 2008: Uncertainty of measurement—Part 3: Guide to the expression of uncertainty in measurement (GUM:1995). ISO/IEC Guide 98-3:2008(E), 130 pp.

  • Iqbal, M., 1983: An Introduction to Solar Radiation. Academic Press, 390 pp.

  • Lacis, A. A., , and J. E. Hansen, 1974: A parameterization for the absorption of solar radiation in the earth’s atmosphere. J. Atmos. Sci., 31, 118133.

    • Search Google Scholar
    • Export Citation
  • Lemus, L., , L. Rikus, , C. Martin, , and R. Platt, 1997: Global cloud liquid water path simulations. J. Climate, 10, 5264.

  • Lorenz, E., , J. Hurka, , D. Heinemann, , and H. G. Boyer, 2009: Irradiance forecasting for the power prediction of grid-connected photovoltaic systems. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 2, 210.

    • Search Google Scholar
    • Export Citation
  • McArthur, L. J. B., 2005: World Climate Research Programme Baseline Surface Radiation Network (BSRN) operations manual, version 2.1. WCRP-121 and WMO Tech. Rep. WMO/TD 1274, 188 pp. [Available online at http://www.wmo.int/pages/prog/gcos/documents/gruanmanuals/WCRP/WCRP21_TD1274_BSRN.pdf.]

  • Ohmura, A., and Coauthors, 1998: Baseline Surface Radiation Network (BSRN/WRMC): New precision radiometry for climate research. Bull. Amer. Meteor. Soc., 79, 21152136.

    • Search Google Scholar
    • Export Citation
  • Perez, R., , K. Moore, , S. Wilcox, , D. Renne, , and A. Zelenka, 2007: Forecasting solar radiation—Preliminary evaluation of an approach based upon the national forecast database. Sol. Energy, 81, 809812.

    • Search Google Scholar
    • Export Citation
  • Puri, K., , G. S. Dietachmayer, , G. A. Mills, , N. E. Davidson, , R. A. Bowen, , and L. W. Logan, 1998: The new BMRC Limited Area Prediction System, LAPS. Aust. Meteor. Mag., 47, 203223.

    • Search Google Scholar
    • Export Citation
  • Ramaswamy, V., , and S. M. Freidenreich, 1992: A study of broadband parameterizations of the solar radiative interactions with water vapor and water drops. J. Geophys. Res., 97, 11 48711 512.

    • Search Google Scholar
    • Export Citation
  • Reikard, G., 2009: Predicting solar radiation at high resolutions: A comparison of time series forecasts. Sol. Energy, 83, 342349.

  • Rikus, L., 1997: Application of a scheme for validating clouds in an operational global NWP model. Mon. Wea. Rev., 125, 16151637.

  • Seaman, R., , W. Bourke, , P. Steinle, , T. Hart, , G. Embery, , M. Naughton, , and L. Rikus, 1995: Evolution of the Bureau of Meteorology’s Global Assimilation and Prediction system. Part 1: Analysis and initialisation. Aust. Meteor. Mag., 44, 118.

    • Search Google Scholar
    • Export Citation
  • Tiedke, M., 1989: A comprehensive mass flux scheme for cumulus parameterization in large-scale models. Mon. Wea. Rev., 117, 17791800.

    • Search Google Scholar
    • Export Citation
  • Vincent, C. L., , W. P. Bourke, , J. D. Kepert, , M. Chattopadhay, , Y. Ma, , P. J. Steinle, , and C. I. W. Tingwell, 2008: Verification of a high-resolution mesoscale NWP system. Aust. Meteor. Mag., 57, 213233.

    • Search Google Scholar
    • Export Citation
  • Welch, R. M., , S. K. Cox, , and J. M. Davis, 1980: Solar Radiation and Clouds. Meteor. Monogr., No. 39, Amer. Meteor. Soc., 96 pp.

  • Weymouth, G. T., , and J. F. Le Marshall, 1999: An operational system to estimate global solar exposure over the Australian region from satellite observations 1. Method and the initial climatology. Aust. Meteor. Mag., 48, 181195.

    • Search Google Scholar
    • Export Citation
  • Weymouth, G. T., , and J. F. Le Marshall, 2001: Estimation of daily surface solar exposure using GMS-5 stretched-VISSR observations: The system and the basics. Aust. Meteor. Mag., 50, 263278.

    • Search Google Scholar
    • Export Citation
  • Wittman, M., , H. Breitkreuz, , M. Schroedter-Homscheidt, , and M. Eck, 2008: Case studies on the use of solar irradiance forecast for optimized operation strategies of solar thermal power plants. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 1, 1827.

    • Search Google Scholar
    • Export Citation
  • World Meteorological Organization, 1975: International cloud atlas, volume I: Manual on the observation of clouds and other meteors. WMO Tech. Rep. 407, 155 pp.

  • World Meteorological Organization, 1987: International cloud atlas, volume II [photographs]. WMO Tech. Rep. 407, 206 pp.

  • World Meteorological Organization, 2012: Guide to meteorological instruments and methods of observation: 2008 edition, updated in 2010. WMO Tech. Rep. 8, 716 pp. [Available online at ftp://ftp.wmo.int/Documents/MediaPublic/Publications/WMO8_CIMOguide/WMO8_Ed2008_Up2010_en.pdf.]

  • Zamora, R., , E. G. Dutton, , M. Trainer, , S. A. McKeen, , J. M. Wilczak, , and Y.-T. Hou, 2005: The accuracy of solar irradiance calculations used in mesoscale numerical weather prediction. Mon. Wea. Rev., 133, 783792.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    Locations of the solar exposure observation sites listed in Table 1. Cape Grim is located on the island of Tasmania. The Great Dividing Range is the mountain range that runs along the southeast coast of the continent.

  • View in gallery

    Plots of daily solar exposure averaged over continental Australia for both MALAPS forecasts. The suffixes d1 and d2 refer to first- and second-day forecasts, respectively.

  • View in gallery

    Spatial analysis of monthly averaged solar exposure for (top) January, (middle) May, and (bottom) November. Shown are (left) PDF distributions for MALAPS and satellite data, color plots of relative percentage difference for MALAPS (center left) first-day and (center right) second-day forecasts, and (right) color plots of average solar exposure for the satellite data.

  • View in gallery

    As in Fig. 3, but for (top) 27 Jan, (middle) 7 Nov, and (bottom) 9 Nov.

  • View in gallery

    Scatterplots of daily solar exposure at selected sites within Australia for different seasons. The MALAPS second-day forecast is shown.

  • View in gallery

    Plots of relative percentage error of daily solar exposure at selected sites within Australia for January and May. The MALAPS first- and second-day forecasts are shown, along with the satellite estimates.

  • View in gallery

    (top) Analysis of hourly mean forecast solar exposure and cloud cover for the second-day forecast for Darwin on 13 Jan. Site-based solar exposure values are superimposed on the top-left plot. (bottom) The observed cloud properties.

  • View in gallery

    As in Fig. 7, but for 16 Jan.

  • View in gallery

    As in Fig. 7, but for Melbourne Airport on 11 Jan.

  • View in gallery

    As in Fig. 9, but for 13 Jan.

  • View in gallery

    Analysis of hourly mean forecast solar exposure and cloud cover for the second-day forecast for Alice Springs on (top) 7, (middle) 8, and (bottom) 9 May. Site-based solar exposure values are superimposed on the plot. There were no observed cloud features.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 20 20 4
PDF Downloads 6 6 1

Testing and Diagnosing the Ability of the Bureau of Meteorology’s Numerical Weather Prediction Systems to Support Prediction of Solar Energy Production

View More View Less
  • 1 Centre for Australian Weather and Climate Research, Bureau of Meteorology, Melbourne, Victoria, Australia
© Get Permissions
Full access

Abstract

The ability of the Australian Bureau of Meteorology’s numerical weather prediction (NWP) systems to predict solar exposure (or insolation) was tested, with the aim of predicting large-scale solar energy several days in advance. The bureau’s Limited Area Prediction System (LAPS) and Mesoscale Assimilation model (MALAPS) were examined for the 2008 calendar year. Comparisons were made with estimates of solar exposure obtained from satellites for the whole Australian continent, as well as site-based exposure observations taken at eight locations across Australia. Monthly-averaged forecast solar exposure over Australia showed good agreement with satellite estimates; the day-to-day exposure values showed some consistent biases, however. Differences in forecast solar exposure were attributed to incorrect representation of convective cloud in the tropics during summer as well as clouds formed by orographic lifting over mountainous areas in southeastern Australia. Comparison with site-based exposure observations was conducted on a daily and hourly basis. The site-based exposure measurements were consistent with the findings from the analysis against satellite data. Hourly analysis at selected sites confirmed that models predicted the solar exposure accurately through low-level clouds (e.g., cumulus), provided that the forecast cloud coverage was accurate. The NWP models struggle to predict solar exposure through middle and high clouds formed by ice crystals (e.g., altocumulus). Sites located in central Australia showed that the monthly-averaged errors in daily solar exposure forecast by the NWP systems were within 5%–10%, up to two days in advance. These errors increased to 20%–30% in the tropics and coastal areas.

Corresponding author address: Paul Gregory, Bureau of Meteorology, GPO Box 1289, Melbourne 3001, VIC Australia. E-mail: p.gregory@bom.gov.au

Abstract

The ability of the Australian Bureau of Meteorology’s numerical weather prediction (NWP) systems to predict solar exposure (or insolation) was tested, with the aim of predicting large-scale solar energy several days in advance. The bureau’s Limited Area Prediction System (LAPS) and Mesoscale Assimilation model (MALAPS) were examined for the 2008 calendar year. Comparisons were made with estimates of solar exposure obtained from satellites for the whole Australian continent, as well as site-based exposure observations taken at eight locations across Australia. Monthly-averaged forecast solar exposure over Australia showed good agreement with satellite estimates; the day-to-day exposure values showed some consistent biases, however. Differences in forecast solar exposure were attributed to incorrect representation of convective cloud in the tropics during summer as well as clouds formed by orographic lifting over mountainous areas in southeastern Australia. Comparison with site-based exposure observations was conducted on a daily and hourly basis. The site-based exposure measurements were consistent with the findings from the analysis against satellite data. Hourly analysis at selected sites confirmed that models predicted the solar exposure accurately through low-level clouds (e.g., cumulus), provided that the forecast cloud coverage was accurate. The NWP models struggle to predict solar exposure through middle and high clouds formed by ice crystals (e.g., altocumulus). Sites located in central Australia showed that the monthly-averaged errors in daily solar exposure forecast by the NWP systems were within 5%–10%, up to two days in advance. These errors increased to 20%–30% in the tropics and coastal areas.

Corresponding author address: Paul Gregory, Bureau of Meteorology, GPO Box 1289, Melbourne 3001, VIC Australia. E-mail: p.gregory@bom.gov.au

1. Introduction

The increased use of solar energy for power generation has created a need for accurate solar irradiance forecasting. These forecasts are required at local operational scales and on large power network scales. In the first instance, the forecasts need to have temporal resolution on the order of 1 h or less to provide plant operators with information required to keep the plant running efficiently during a day with fluctuating solar irradiance. In the second instance, managers of a power generation grid require estimates of daily solar exposure (or insolation) up to several days in advance to balance the energy demand from the grid with other power sources.

Improved forecasts of solar exposure can improve weather forecasts for other applications. Solar irradiance is a key component of the surface energy balance and is dependent on predicted water vapor (including cloud coverage), ozone, atmospheric aerosols, and pollution. Recent studies have used operational numerical weather prediction (NWP) models, dedicated aerosol forecast models, or a combination of both to produce solar exposure forecasts on daily time scales. Breitkreuz et al. (2009) examined the capability of mesoscale forecasts to predict clear-sky conditions. Their studies focused on aerosol chemical transport, which is a significant atmospheric parameter that determines solar exposure on clear-sky days. Breitkreuz et al. (2009) developed Aerosol-Based Forecasts of Solar Irradiance for Energy Applications (AFSOL) and tested it for the period of July–November 2003. Comparisons with European Centre for Medium-Range Weather Forecasting (ECMWF) products showed that AFSOL reduced the bias error from −26.3% to 11.2% and relative RSME from 31.2% to 18.8% for clear-sky days. For cloudy situations, the AFSOL system was less accurate.

Lorenz et al. (2009) analyzed the ECMWF irradiance forecasts against sites in Germany to develop a network-scale forecasting tool. These results were then processed to produce hourly forecasts of photovoltaic power generation at specific sites. For clear-sky days (total cloud cover < 0.03) the ECMWF forecast was replaced with a clear-sky model that directly modeled aerosols. The best forecasts were obtained when a linear combination of the ECMWF forecast and the clear-sky model with a bias correction were used. For irradiance forecasts of single sites, the relative RMSE of this approach was 36.9% for the first forecast day, increasing to 46.3% for the third forecast day.

Work by Zamora et al. (2005) compared solar irradiance forecasts derived from mesoscale NWP models with observations. They focused on clear-sky irradiance forecasts with high ozone and aerosol concentrations. The fifth-generation Pennsylvania State University–National Center for Atmospheric Research Mesoscale Model (MM5) solar irradiance forecasts were validated against three sites featuring high surface ozone levels in Texas, central California, and New England. Zamora et al. (2005) suggested that the Lacis–Hansen-type parameterization is adequate for regions of the country not impacted by high aerosol loadings and can provide estimates of solar irradiance within 3%–4% as long as the aerosol optical depth (AOD) is less than or equal to 0.1. Significant biases of clear-sky irradiance can occur for values of AOD that are greater than 0.1, however. Sensitivity tests suggest that clear-sky solar irradiance errors can have a significant effect on numerical model performance.

Perez et al. (2007) developed a solar radiation forecast model using sky-cover predictions from the National Digital Forecast Database combined with ground-measured and satellite-derived irradiances in Albany, New York. A curve fit of global hourly irradiance (GHI) against sky cover was obtained, which was then analyzed against ground- and satellite-based GHI measurements. The relative mean bias error for the best-fit model was less than ±10%, and relative RMSE was less than 40% throughout the 72-h forecast range.

Other methods have been used to provide forecasts at shorter time scales. Wittman et al. (2008) used satellite-based nowcasting of solar irradiance and assimilation of aerosols into air-quality models as a basis for day-ahead irradiance forecasts. This technique was developed for control and maintenance of solar thermal power plants. Reikard (2009) examined forecasting over short time horizons (<60 min) for operational planning and support using predictive models such as autoregressive integrated moving averages, transfer functions, and unobserved-components models.

The goal of the research presented here was to evaluate the capability of the Bureau of Meteorology’s NWP codes to support solar energy prediction at the power-network scale. The aim was to show whether accurate forecasts of daily solar exposure (within 10%) could be provided by the bureau’s current operational forecast models.

2. The bureau’s NWP models

The operational NWP suite consists of a global assimilation system (GASP) with a resolution of 0.8° that uses an optimal interpolation scheme over a 6-h window to produce global analyses and forecasts. It provides boundary conditions for the Australian regional assimilation system [Limited Area Prediction System (LAPS)], which is also based on the optimal interpolation scheme. A mesoscale assimilation system (MALAPS) that is nested in LAPS became operational in December of 2007. Details of the GASP, LAPS, and MALAPS systems can be found in Seaman et al. (1995), Puri et al. (1998), and Vincent et al. (2008), respectively. The LAPS model has a resolution of 0.375° and provides 72-h forecasts, whereas the MALAPS model has a resolution of 0.1° and provides 48-h forecasts. Both LAPS and MALAPS have 61 pressure levels, with 41 levels below 100 hPa. Results for both models have been processed for the entire 2008 calendar year. The LAPS and MALAPS models were replaced operationally in August of 2010 by the new Australian Community Climate and Earth-System Simulator (ACCESS) model.

The radiative transfer parameterization in the bureau’s model is based on Lacis and Hansen (1974) with modifications according to Ramaswamy and Freidenreich (1992). The vertical profiles of optically active atmospheric constituents (e.g., water vapor, ozone, aerosols, and cloud) are used to calculate the absorption, albedo, and transmission for each layer defined by the model’s level distribution. The radiation fluxes are computed through each layer for two spectral bands—one for the ozone absorption region and one that uses a 12-term exponential sum fit to the spectral absorption coefficient for water vapor. The radiation flux calculations depend on transmission and reflection coefficients for both direct and diffuse radiation that are generated by using the absorption and scattering characteristics of each layer and assuming a two-stream approximation as used in Lacis and Hansen (1974). Rayleigh scattering is treated as a modification to the surface albedo, and multiple scattering is included.

The (MA)LAPS models use a cloud-fraction parameterization that is based on Rikus (1997). The parameterization for cloud condensation amount is based on Lemus et al. (1997). The model’s mass-flux convection scheme is based on Tiedke (1989) and is not explicitly linked to the cloud parameterizations, but it does modify the moisture and temperature distributions and therefore affects the cloud fields implicitly. The ice/water ratio within a cloud is a temperature-dependent function that is also used to define the dependence of relative humidity on ice/water. Shallow convection is part of the mass-flux scheme that provides moisture transport in the lower levels. The associated cloud fraction is diagnosed from a function of static stability and relative humidity. An explicit cloud microphysics scheme that is based on Dare (2004) is used to simulate large-scale precipitation processes; the resultant fields were not suitable for use in the radiative transfer calculations, however.

Cloud radiative properties (optical depth, cloud fraction, phase, etc.) are diagnosed at each hourly radiation time step from the model thermodynamic fields. These cloud radiative quantities, together with other radiative parameters such as the solar zenith angle, remain constant over the period of each radiation time step. The solar zenith angle is chosen to reproduce the mean solar irradiance at the top of the atmosphere for each profile. This gives the correct result for clear skies but introduces cloudy-sky errors that are related to the relative temporal and spatial scales of the modeled and real cloud as the model thermodynamic fields and cloud microphysics are evolving with each model time step. Because the radiative transfer calculations are expensive (~30% of the model run time with hourly radiation) and operational NWP models are required to run in tight time windows, the choice of hourly radiation is a compromise. More-frequent calculations would (potentially) improve the solar forecasts but would make NWP forecasts too late to be useful.

Daily solar exposure is computed by summing the calculated hourly solar irradiance over daylight hours, thereby converting it to daily solar exposure. The operational forecast starting at 1200 UTC encompasses two complete solar days (from ~1800 to ~1200 UTC) and these have been evaluated separately as the first- and second-day forecasts for solar exposure. The LAPS model allows the computation of an additional third-day solar forecast. The forecasts based at 0000 UTC were ignored because they only encompassed one complete solar day for the 48-h forecast (MALAPS) or two solar days for the 72-h forecast (LAPS).

3. Validation data

a. Satellite data

Satellites can infer solar irradiance by measuring the outgoing radiation at the top of the atmosphere for a spectral band in the visible part of the solar spectrum. Additional algorithms are used to diagnose the effects of cloud and so on. The benefit of the satellite-derived data is that they are produced over the entire continent and are the only estimate available for most regions. The validation data are sourced from an updated version of the bureau’s surface solar exposure product. This is derived using hourly geostationary satellite data using a semiempirical model developed by Weymouth and Le Marshall (1999, 2001). This model is validated against data from a small number of surface radiation sites. The data referenced in this paper were sourced from the Multifunctional Transport Satellite (MTSAT)-1R. The satellite images have a 1.25-km resolution at the subsatellite point and are typically produced hourly. The data are acquired in real time by the bureau and are geolocated, calibrated, and regridded to a 0.01° grid and then are spatially averaged to the final 0.05° grid. Instantaneous surface solar irradiance at each grid point is calculated for the time of each satellite image with a semiempirical model that parameterizes the important aspects of the radiative transfer in clear and cloudy atmospheres in two spectral bands that cover visible and near-infrared wavelengths, respectively. The hourly surface irradiance estimates for each day are integrated to give the daily solar exposure for each point. Because the satellite product has a horizontal resolution of 0.05°, regridding of the satellite data to the model grid is required for direct comparison. This was performed using the Climate Data Analysis Tools product available from the Lawrence Livermore National Laboratory’s Program for Climate Model Diagnosis and Intercomparison. Linear interpolation was used. Further details of the satellite data can be found in appendix A.

b. Site data

The Bureau of Meteorology’s Solar and Terrestrial Network started in 1993 and is based on the World Climate Research Programme Baseline Surface Radiation Network (BSRN) protocols (Ohmura et al. 1998), and three of the nine bureau stations contribute data to the BSRN data archive. Relevant irradiance signals from a pyrheliometer, shaded and exposed pyranometers, and a shaded pyrgeometer are sampled at 1 Hz and then are used to derive 1-min statistics for direct, diffuse, global, and terrestrial irradiance. The primary global irradiance is made up of the sum of direct irradiance translated to a horizontal plane and the diffuse irradiance (called the component sum), with the direct and diffuse zero signal corrected for the net-zeros solar irradiance balance of the thermopiles before generating the direct and diffuse irradiance. When the direct irradiance minute statistics are not available, the global irradiance derived from the global pyranometer signal substitutes for the component sum value. Uncertainties that are based on the International Organization for Standardization/International Electrotechnical Commission (2008) method are derived for every minute mean irradiance using each minute’s statistics and the metadata associated with the measurement (cleaning frequency, solar zenith angle, time between calibrations of the direct pyrheliometer, zero signal correction, instruments used, etc.). These minute statistics are then used to generate the standard solar exposure products released by the bureau—namely, half-hourly and daily exposures referenced to true solar time (±30 s) and the associated uncertainty for each exposure. The statistical analysis in this paper used the daily exposure data; the half-hourly data were used for illustrative purposes for the figures in section 5g, which examines hourly forecast data at specific sites.

Cloud amount was measured in octas, with cloud types and heights recorded by bureau observers according to guidelines outlined by the World Meteorological Organization (WMO) Commission for Instruments and Methods of Observations (WMO 2012, part I: chapter 15). Cloud classifications were assigned according to WMO definitions given in WMO (1975, 1987).

Further details of the site data (as well as examples of uncertainty calculations) can be found in appendix B. The list and locations of the sites are given in Table 1 and Fig. 1. Table 2 gives observed values of mean global daily solar exposure and 95% uncertainty for each month at each site.

Table 1.

Locations of sites that record ground-based solar exposure measurements.

Table 1.
Fig. 1.
Fig. 1.

Locations of the solar exposure observation sites listed in Table 1. Cape Grim is located on the island of Tasmania. The Great Dividing Range is the mountain range that runs along the southeast coast of the continent.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

Table 2.

Observed values of mean global daily solar exposure (Exp) and 95% uncertainty (Unc) (MJ m−2) for each month at each site for the 2008 calendar year.

Table 2.

4. Results: Australia wide

Results for NWP forecasts are presented in this section and are compared with the equivalent satellite results. These results are for solar exposure that has been spatially averaged across the whole of Australia.

a. Monthly results

Results for the monthly solar exposure from the forecast models for the 2008 calendar year are presented in Tables 3 and 4 for the MALAPS and LAPS models, respectively. These statistics were computed for each month from daily solar exposure values from the NWP forecast and satellite estimate, respectively. The predicted daily exposure was then spatially averaged.

Table 3.

Monthly validation statistics for the MALAPS model solar exposure predictions against the satellite solar exposure estimates across the entire Australian continent for the 2008 calendar year.

Table 3.
Table 4.

Monthly validation statistics for the LAPS model solar exposure predictions against the satellite solar exposure estimates across the entire Australian continent for the 2008 calendar year.

Table 4.

For two data vectors of length N, with the forecasts denoted as Fi and the observations as Oi, the multiplicative bias is defined as
e1
the mean absolute error (MAE) is defined as
e2
and the relative absolute error (RAE) is defined as
e3

The tables show the correlation between the daily time series for each month, along with multiplicative bias error, MAE, and RAE expressed as a percentage. These metrics should be regarded as “differences” because the satellite data are not a direct measurement of irradiance but rather are an estimate.

The results for both models that are given in Tables 3 and 4 show that the forecasts tend to overestimate solar exposure levels for most of the year; there is a definite shift that occurs in late autumn, however, when the model underestimates solar exposure (with the maximum bias occurring in May). The correlation for both models is worst in July. Increasing the forecast range for the MALAPS models improves forecasts in spring and early summer but degrades the forecast in late autumn and winter. Increasing the forecast range for the LAPS model shows similar behavior with respect to the bias; the relative differences now show more consistent change, however, with the LAPS third-day forecast giving worse agreement than does the second-day forecast.

The mean absolute differences for the MALAPS model are still within the maximum summer uncertainty value for the satellite estimate of 1.5 MJ m−2 (although they are close to this value in February). The values in winter are still within the satellite uncertainty value of 0.8 MJ m−2. Mean absolute differences for the LAPS model were smaller throughout the warmer seasons but were higher in the winter months. The difference in May and June for the LAPS model appears to exceed the uncertainty of the satellite estimate, especially for the second- and third-day forecasts.

Further analysis presented in subsequent sections shows that this seasonal change in behavior between the models and the satellite estimates is due to seasonal biases of the satellite estimates at certain times of the year.

b. Results for selected months of 2008

Plots of daily solar exposure averaged over the whole of Australia for selected months are shown in Fig. 2 for the MALAPS model. The months selected for detailed analysis are January, May, July, and November, which are representative of summer, autumn, winter, and spring conditions, respectively.

Fig. 2.
Fig. 2.

Plots of daily solar exposure averaged over continental Australia for both MALAPS forecasts. The suffixes d1 and d2 refer to first- and second-day forecasts, respectively.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

The January results show a consistent bias between the NWP and satellite data of order 5%–10%. During May, this bias has reversed sign, and the magnitude of the NWP data is less than that of the corresponding satellite data. Once again, the difference is on the order of 5%–10% at various days, although toward the end of this month the agreement is very good. During July, the tendency of the NWP to overestimate solar exposure has returned; it is not as consistent as that observed during January, however.

The behavior for November is similar to that for January: although there is very good agreement for some days (particularly 9, 14, and 23–27 November), there are also some days for which the NWP overestimates the satellite data by 10%–20%. Over these four particular months, the behavior of the first- and second-day forecasts is very similar.

Results for the LAPS model are excluded here for brevity; little difference was observed between the two NWP systems, however, despite the large differences in resolution. In addition, the LAPS results show little variation with each additional forecast day.

The results show that the existing NWP systems can provide qualitative predictions of daily mean solar exposure, up to three days in advance. The plots of Fig. 2 show that some months have a systematic bias that is present throughout the entire month (e.g., May), whereas other months (e.g., November) show good agreement for days with high values of average daily solar exposure but poorer agreement for days with lower values. The presence of major cloud bands (which is responsible for day-to-day exposure variation) is shown to be predicted as the NWP forecasts can replicate the daily trends in solar exposure throughout the month. The degree of attenuation caused by the cloud bands is still subject to some variation relative to the satellite-derived data, however. Tables 3 and 4 show that the relative differences of the NWP forecasts of average solar exposure in comparison with the satellite estimates vary between 2% and 5% for the MALAPS model (2%–7% for the LAPS model). The question of whether the model or the satellite data are closest to reality is considered further in section 5, which compares both methods with surface data.

c. Spatial analysis: Monthly results

Figure 2 showed the spatially averaged value of daily accumulated solar exposure calculated across Australia for each day, eliminating any spatial variation in solar exposure over the continent for each day. Solar irradiance varies with latitude and is also dependent on local cloud cover. Additional information to determine the source of differences in spatially averaged solar exposure can be achieved by calculating the probability density function (PDF) for the monthly mean daily accumulated solar exposure over the continent. The PDF enables the spatial distribution of the NWP results to be compared with the satellite data by showing the relative amount of each average solar exposure value present for a given month. For brevity, the results presented here are limited to the MALAPS system.

Figure 3 shows computed PDFs and color plots of relative percentage difference of the MALAPS forecasts. Color plots of satellite estimates of solar exposure are also shown. For January, the color plots for satellite solar exposure show lower values of solar exposure in the tropics and along the eastern seaboard. Monthly averaged relative differences for both the first- and second-day MALAPS forecasts are within ±6% over most of the continent. The forecast overpredicts solar exposure in the tropics by more than 20% while differences of 10%–20% are observed along the eastern coast along the Great Dividing Range. Relative to the first-day forecast, the second-day forecast shows slightly better performance in the interior of the continent, with worse performance in the far north and along the eastern seaboard. The corresponding PDFs show a small shift in the peak toward cloudier conditions with increasing forecast day. The PDF of the satellite data agrees well with the model peaks, but the tails of the PDFs for low solar exposure values are very different. The satellite PDF has more data points with solar exposure values of less than 25 MJ m−2.

Fig. 3.
Fig. 3.

Spatial analysis of monthly averaged solar exposure for (top) January, (middle) May, and (bottom) November. Shown are (left) PDF distributions for MALAPS and satellite data, color plots of relative percentage difference for MALAPS (center left) first-day and (center right) second-day forecasts, and (right) color plots of average solar exposure for the satellite data.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

The results for May are significantly different. The PDF shapes are much rounder and broader, which is partly due to more cloud diminishing the solar exposure values over the month; the changing of the solar zenith angle with season also contributes to a broader PDF for clear skies. The NWP and satellite PDFs show only slight differences. The color plots show a very different spatial pattern of differences than is seen for January. Large differences in the tropics are no longer present; overprediction of solar exposure occurs in the mountainous areas in the south of the continent and Tasmania, however.

For November, large differences are again observed. The peak of the satellite PDF has a very different shape, as does the tail for lower solar exposure values. The spatial difference pattern is again different, with relative differences of 15%–20% now observed in central Australia and throughout the southeast of the continent.

d. Spatial analysis: Selected daily results

The spatial analysis from section 4c was repeated for specific days in each month to try to identify any common features between days that featured particularly good or bad agreement between the NWP and satellite results. Table 5 lists the selected days along with the computed differences in average solar exposure.

Table 5.

Average solar exposure (MJ m−2) across Australia at a specific day.

Table 5.

Spatial analysis of the selected days is shown in Fig. 4. This analysis shows how accurately the NWP systems can predict cloud formations on a day-to-day basis, which is the most important aspect of being able to forecast solar exposure accurately. Taking a monthly average of solar exposure smooths out any daily discrepancies between the NWP and satellite results. Therefore, the analysis of solar exposure color plots and PDFs for specific days is an important step.

Fig. 4.
Fig. 4.

As in Fig. 3, but for (top) 27 Jan, (middle) 7 Nov, and (bottom) 9 Nov.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

An example of a “good” day is 27 January. In this instance, most of the continent is cloud free, although there is some lighter cloud in the tropical regions. Relative differences across most of the continent were small; large differences of up to 60% were observed in cloudy regions in the north, however.

Similar cloud coverage is shown on 7 November for both the NWP and satellite results, which in this instance are dominated by a large cloud band through central Australia. The forecast cloud is optically thinner than the satellite estimates with differences of more than 60% in the south of the continent and Tasmania. This discrepancy in cloud depth is noticeable in the enhancement of the left tails of the PDFs. Although for this day the NWP forecast of cloud coverage was fairly accurate on a continental scale, differences in predicted cloud depth still caused large differences across much of Australia.

An example of a good spring day is 9 November, as the continent is almost clear of cloud. Both the forecast and satellite observations show cloud distributed throughout the tropics and along the northeast coastline, with some isolated cloud in southeastern Australia. Relative differences in solar exposure of up to 60% are present in regions where insufficient cloud was forecast, with differences of up to −100% in regions where too much cloud was present.

The results of Figs. 3 and 4 suggest that the NWP models systematically predict too much solar exposure at low latitudes in Australia. This problem is mainly confined to the summer months, and it is hypothesized that it is caused by the NWP algorithm treating the clouds present in this region as being too transparent to incoming solar irradiance—that is, their optical depth is too thin. These discrepancies in the poor prediction of minimum values of solar exposure were also observed along the Great Dividing Range and in southeastern Australia. These systematic differences create some opportunity for a bias correction.

On the basis of these observations and the known systematic errors in the NWP model formulation, it is hypothesized that most of the differences in tropical regions are due to an inability to simulate convective cloud correctly. In most tropical regions, convective clouds are often produced locally; that is, they are not part of larger-scale cloud bands and weather systems.

In addition, those clouds created along the Great Dividing Range are strongly influenced by geographical features that create clouds through orographic lifting. These geographical features are not fully resolved by the model grid.

5. Results: Ground-based observations

The bureau’s radiation monitoring network allows validation of the NWP results presented earlier. Annual, seasonal, and monthly statistics of daily surface solar exposure for each site are presented, along with time series for accumulated daily solar exposure for selected months at each site. This analysis uses the observed daily accumulation of solar irradiance. Separate analysis of clear-sky days is also presented. Hourly analysis for selected days at various sites is also presented that uses the half-hourly site data. The hourly analysis allows the observed direct and diffuse solar exposure and the observed cloud fields to be compared with the NWP forecasts for solar exposure and cloud coverage. This detailed analysis makes it possible to test the hypothesis proposed earlier in sections 4c and 4d: namely, that the current NWP algorithms treat very cold cloud as too transparent to incoming solar irradiance. The gridded NWP and satellite data were extracted at each site using a weighted average of the inverse distance of the four closest grid points to each site.

a. Annual results

Correlation and bias statistics at each site for the entire 2008 calendar year are shown in Tables 6 and 7 for the MALAPS and LAPS models, respectively. These statistics are computed for the daily time series over the entire year. There is little difference in the performance of both models. The common feature is the poor correlation observed in the tropics at Darwin and Broome. The tendency of both models to overpredict solar irradiance is evidenced in the positive bias errors at all sites. Correlation and bias do not change significantly with increased forecast length; the RAE error is shown to increase for both models, however.

Table 6.

Validation statistics for the MALAPS model solar exposure predictions at various sites in Australia for the 2008 calendar year.

Table 6.
Table 7.

Validation statistics for the LAPS model solar exposure predictions at various sites in Australia for the 2008 calendar year.

Table 7.

b. Seasonal results

A seasonal breakdown of the results is presented for the MALAPS model in Table 8. Scatterplots are shown in Fig. 5 for the MALAPS second-day forecast. For brevity, results are only shown for the sites at Darwin, Alice Springs, and Melbourne. These locations correspond to tropical, desert, and midlatitude locations, respectively. These seasonal statistics show the improvement in NWP forecasts in the tropics outside of summer, highlighting the inability of the NWP to predict cloud cover created by local convection and instability in the monsoon season. The MALAPS model achieves excellent results at both Alice Springs and Melbourne in autumn, where seasonal climates feature clear skies and a stable atmosphere. Winter and spring in Melbourne usually feature the frequent passing of cold fronts and increased atmospheric instability, which gives more scope for errors in the NWP forecast.

Table 8.

Validation statistics for the MALAPS model solar exposure predictions at various sites in Australia for the 2008 calendar year.

Table 8.
Fig. 5.
Fig. 5.

Scatterplots of daily solar exposure at selected sites within Australia for different seasons. The MALAPS second-day forecast is shown.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

c. Annual clear-sky results

The annual and seasonal analyses were repeated for clear-sky days only. The clear-sky-day criterion was that no observed or forecast cloud was present at the site between sunrise and sunset. Such a strict criterion meant that there would be few days with completely clear sky in areas such as the tropics in summer (or the midlatitudes in winter).

This analysis was performed to isolate the solar irradiance calculations from any cloud effects, although aerosols could still be an issue. It also allowed assessment of the model’s ability to forecast clear-sky conditions, which are crucial for energy load balancing of networks that include solar power plants. Note that the Cape Grim station is not included in this analysis because it has no associated observed cloud data.

Additional metrics were used to measure clear-sky forecasting skills. The probability of detection (POD) is defined as
e4
and the false-alarm rate (FAR) is defined as
e5
Tables 9 and 10 show validation statistics for clear-sky exposure across all sites in Australia for the MALAPS and LAPS models, respectively. The results show good values of correlation and bias errors for these days. The relative absolute errors were roughly 3% at most sites, although tropical locations showed slightly higher values of up to 6%. Correlation values were also poorer at tropical sites. Bias errors were between 3% and 4% at most sites.
Table 9.

Validation statistics for the MALAPS model clear-sky solar exposure predictions at various sites in Australia for the 2008 calendar year.

Table 9.
Table 10.

Validation statistics for the LAPS model clear-sky solar exposure predictions at various sites in Australia for the 2008 calendar year.

Table 10.

The second-day forecast results were superior at the midlatitude sites and were worse at sites in the tropics. This trend was consistent across both models. The third-day forecast for the LAPS model was the worst forecast in terms of relative-error magnitude, although this was still below 5% across all sites. It was evident from these results that neither model was able to produce a correct forecast of clear-sky conditions at Melbourne during the entire year.

Tables 11 and 12 show POD and FAR for the clear-sky forecasts at each site for the 2008 calendar year for the MALAPS and LAPS models, respectively. The MALAPS model only showed useful skill at forecasting clear-sky conditions in Alice Springs and Broome. Other tropical sites showed slightly worse results; the midlatitude forecasts showed no skill, however, with POD of less than 0.15. For the LAPS model, performance at Alice Springs was considerably worse than for the MALAPS results. All other sites showed similar performance. There was little difference between results between the first- and second-day forecasts, although the third-day LAPS forecasts did show considerable reduction in skill at most sites.

Table 11.

Validation statistics for the MALAPS forecasts of clear-sky conditions at various sites in Australia for the 2008 calendar year.

Table 11.
Table 12.

Validation statistics for the LAPS forecasts of clear-sky conditions at various sites in Australia for the 2008 calendar year.

Table 12.

Table 13 shows the contingency tables for clear-sky forecasts at each site for the 2008 calendar year for the MALAPS model. These data help to quantify the number of actual clear and cloudy days observed at each site and the associated forecast accuracy. It is apparent that the midlatitude sites at Adelaide, Wagga Wagga, and Melbourne Airport have similar totals of observed clear-sky days to the tropical sites of Rockhampton and Darwin, yet the forecast performance at the midlatitude sites is very bad. The MALAPS model was unable to forecast any of the 38 observed clear-sky days at Melbourne Airport. The contingency tables also highlight the difference between Broome and the other tropical sites of Darwin and Rockhampton. Although Broome is at a lower latitude than Rockhampton, it is situated on the west coast of the continent. Rockhampton is located on the east coast and receives much higher annual rainfall (800 vs 532 mm) because of predominantly onshore winds from the Pacific Ocean for most of the year. Broome’s climate is drier because it is downwind of inland grasslands and deserts in Australia’s interior. Darwin’s climate is even wetter than Rockhampton (over 1700 mm of annual rainfall).

Table 13.

Contingency tables for the MALAPS forecasts of clear-sky conditions at various sites in Australia for the 2008 calendar year. Note that statistics are only computed for 365 days because no forecast data were available for 25 Jul.

Table 13.

d. Seasonal clear-sky results

A seasonal breakdown was again performed for the MALAPS model, with the sites Wagga Wagga, Alice Springs, and Darwin chosen to represent results at midlatitude, desert, and tropical locations, respectively (Table 14). The MALAPS model shows very different performance at each site for each season. At Wagga Wagga, the seasons with the lowest errors for forecast clear-sky exposure are summer and spring. For Alice Springs, the best seasons are summer and autumn, and for Darwin the best seasons are autumn and winter, or the “dry” seasons. As expected, the model is unable to forecast any clear-sky days in Darwin during the “wet” seasons (only one clear-sky day was observed at Darwin during the wet seasons).

Table 14.

Seasonal validation statistics for the MALAPS model clear-sky solar exposure predictions at various sites in Australia for the 2008 calendar year.

Table 14.

The strong variation in clear-sky irradiance prediction at Wagga Wagga and Alice Springs is thought to be due to aerosol factors such as summer dust storms or bushfires, or high pollen levels in the spring. An examination of observer visibility data (in kilometers) at these locations showed no seasonal trends across the 2008 calendar year, however. A more in-depth analysis using satellite products on a day-by-day basis may provide better answers in the future.

e. Monthly results

Table 15 shows monthly validation statistics for solar exposure at the three selected sites. The table shows the strong seasonal signal in the model performance at each site. The results at Darwin during the wet season are very poor, with poor values of correlation (including some negative values) and RAE values of almost 40%. This is to be expected at a location that is dominated by tropical storms and heavy cloud cover during this time. During the dry season the model performance improves significantly, and RAE values drop to less than 10%.

Table 15.

Monthly validation statistics for the MALAPS model solar exposure predictions at various sites in Australia for the 2008 calendar year.

Table 15.

Results for Alice Springs showed far superior correlation values throughout the year. The annual trends in model performance differ slightly. The worst results generally occurred in the warmer months (from November to March); the results for January were excellent, however, with values of RAE of less than 5%. The model showed consistently better performance in the cooler months, with errors being typically less than 10%. The exception occurred in June, when a bad forecast on 7 June (which produced a daily RAE value of more than 400%) skewed the rest of the data, which otherwise showed excellent agreement for the remainder of the month.

Melbourne showed the opposite seasonal behavior to Darwin. This was expected because of the geographic location of Melbourne, which creates a climate that features hot, dry summers and cool, cloudy winters. Forecast performance is therefore better in summer than in winter. The errors relative to the other two sites are far worse. The best months (generally in the late summer) feature RAE of less than 15%, which grow to over 30% in the winter.

At all three stations, there was no discernible performance difference between the first-day and second-day forecasts. All sites showed positive bias errors for the entire year.

f. Daily results

Figure 6 shows relative errors in daily solar exposure at selected sites for January and May. Errors between the site and NWP data from the first-day and second-day MALAPS forecasts are plotted along with satellite-derived data.

Fig. 6.
Fig. 6.

Plots of relative percentage error of daily solar exposure at selected sites within Australia for January and May. The MALAPS first- and second-day forecasts are shown, along with the satellite estimates.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

The January results for Alice Springs show that the errors in forecast daily solar exposure were within ±5% for most of the month, although this error grows to ±10% from 20 January onward, with the last day of the month showing a noticeable forecast “bust” with an error of roughly 70%. There is very little difference between the first- and second-day forecasts, with 28 January showing the only noticeable difference. The errors in the satellite estimate are slightly smaller than the NWP errors.

At lower latitudes, the results for Darwin show much greater variability in solar exposure, which in turn creates larger relative errors. Large positive bias errors are observed for the NWP forecasts, generally in the range of 25%–50%, although on some days the NWP overpredicts daily solar exposure by over 100%. Differences between the first- and second-day forecasts are small in the context of these large error magnitudes. The large positive bias errors observed at Darwin are consistent with the earlier analysis, which shows that the cloud produced by the NWP systems in the tropics is too optically thin. Errors in the satellite estimate of solar exposure are far smaller and are generally less than 25%.

Performance of the NWP forecasts at Melbourne is mixed. Errors during the first half of January are within ±10%; there are seven days in the second half of the month on which the relative errors are greater than 50%, however. Performance of the NWP forecasts relative to the satellite product is roughly equivalent.

As the seasons move from summer to winter, the performance of the NWP and satellite systems relative to the ground-based site observations changes. This was evidenced in the seasonal statistics discussed previously. The desert site of Alice Springs shows that the NWP forecast gives superior estimates of solar exposure for most days in May, with relative errors usually less than 5%. Poorer performance of the satellite estimate could be due to the use of bias corrections extrapolated from the previous year.

As the sun moves lower in the sky at Darwin as the winter solstice approaches, the magnitude of the daily solar exposure falls, as do the relative errors of the NWP forecasts. The majority of the errors lie within ±10% of the site measurement. The positive bias in the NWP forecasts that was present in January is not as consistent for this month, and there is a noticeable degradation in performance of the second-day forecasts.

Error magnitudes in Melbourne have increased as winter approaches. Although solar exposure magnitude has decreased, the frequent passing of cold fronts creates high variability in solar exposure. The satellite data show inferior performance to both MALAPS forecasts. Recall from section 4 that the NWP models showed the largest monthly bias during May (relative to the satellite estimates). The results presented here suggest that this bias error is due to errors in the satellite product rather than to a systematic deficiency in the NWP model.

The trend for the MALAPS data to perform poorly at coastal locations can be explained by the model’s inability to cope with strong cloud gradients along the coast. Often sharp boundaries in cloud fraction can form along the coast because of differences in temperature and heating rate between the land and ocean. The nature of the discretization schemes used in NWP can make it difficult to represent these strong gradients without an excessively fine grid.

g. Hourly results

As mentioned previously in section 3b, the ground-based observations also record cloud fraction and cloud height at half-hourly intervals. The use of these data enables comparison with the cloud fields present in the NWP model at any particular site. These data allow the errors in solar radiation to be attributed to 1) the NWP giving poor forecasts of cloud coverage, 2) the NWP giving poor computations of optical thickness for the clouds present, or 3) a combination of both factors.

1) Darwin: 12–16 January

Table 16 shows the values of accumulated solar exposure at Darwin for the MALAPS second-day forecast and site observations over the period of 12–16 January, which is an example of the NWP models overpredicting solar exposure at tropical locations during the summer. The MALAPS second-day forecast of solar exposure for 12 January was very accurate relative to the site observation. The forecast predicted a partly cloudy day, which was what happened according to the site data. The observed clouds for that day were low-level cumulus with cloud fractions of less than 0.5, with high-level cirrus in fractions between 0.6 and 0.9.

Table 16.

Comparison of accumulated solar exposure (MJ m−2) between the second-day MALAPS forecast and site data for Darwin, Melbourne, and Alice Springs for selected dates in 2008.

Table 16.

Figure 7 shows that the site solar exposure on 13 January was completely diffuse, which indicates a completely overcast day, whereas the NWP predicted a mostly sunny day, with almost complete high-cloud cover during the entire day along with lower amounts of middle-cloud cover. Low-cloud cover was forecast to dissipate in the afternoon.

Fig. 7.
Fig. 7.

(top) Analysis of hourly mean forecast solar exposure and cloud cover for the second-day forecast for Darwin on 13 Jan. Site-based solar exposure values are superimposed on the top-left plot. (bottom) The observed cloud properties.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

The observed cloud cover that day featured almost complete coverage of altostratus, with patches of nimbostratus, and altocumulus. Low-level cloud such as cumulus and stratocumulus was present in smaller amounts. Although the forecast for high cloud was fairly accurate, there was almost no reduction of the predicted solar exposure. This would suggest that the high-level clouds within the MALAPS model are too optically thin. The overprediction of solar exposure continued on 14 January; the errors in predicted solar exposure for this day are also due to an incorrect cloud-cover forecast, however.

The performance of the NWP forecast on 15 January improves, as this day had mostly clear skies with small amounts of cumulus. The forecast of 16 January (Fig. 8) is very accurate, as this was day with a few patches of high cloud and small cloud fraction (<0.5) of cumulus cloud. The model overpredicted the amount of high cloud and underpredicted the amount of low cloud, but the overall solar exposure amount was very accurate.

Fig. 8.
Fig. 8.

As in Fig. 7, but for 16 Jan.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

2) Melbourne: 11–13 January

This analysis is repeated at Melbourne airport for the period of 11–13 January. The values of the daily accumulated solar exposure for these days are given in Table 16.

The results for 11 January at Melbourne airport (Fig. 9) show that the forecast produces unrealistic cloud fields. The observational data show some high- and midlevel cloud around midday, with some low cloud (cumulus and stratus) in the afternoon. The NWP second-day forecast predicts too much cloud, with almost 100% cloud cover present for most of the day. The agreement for 12 January is much better. The forecast predicts a day with mostly clear skies. The observed cloud pattern showed patches of stratocumulus in the morning but clear skies in the afternoon.

Fig. 9.
Fig. 9.

As in Fig. 7, but for Melbourne Airport on 11 Jan.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

On 13 January, cloudy skies returned at Melbourne airport. Figure 10 shows that the second-day forecast for this day predicted heavier cloud cover, with 80%–100% low-cloud cover until late in the afternoon. The site observations showed overcast skies comprising cumulus and stratocumulus, which broke up at around midday. The predicted solar exposure is therefore too low around midday, but it is more accurate in the morning and afternoon. This example shows that the model estimates of optical depth for low clouds are fairly accurate.

Fig. 10.
Fig. 10.

As in Fig. 9, but for 13 Jan.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

3) Alice Springs: 7–9 May

The final hourly analysis was carried out at Alice Springs. Figure 6 and Table 16 show that the forecasts produces poor results for 8 May, but the overall agreement for the rest of the month is within 10%. Hourly plots for the days noted in Table 16 are shown in Fig. 11.

Fig. 11.
Fig. 11.

Analysis of hourly mean forecast solar exposure and cloud cover for the second-day forecast for Alice Springs on (top) 7, (middle) 8, and (bottom) 9 May. Site-based solar exposure values are superimposed on the plot. There were no observed cloud features.

Citation: Journal of Applied Meteorology and Climatology 51, 9; 10.1175/JAMC-D-10-05027.1

Clear skies were observed on 7 May, whereas the NWP forecast predicts large amounts of high-cloud cover during the day. It is evident that the computed high clouds produce small levels of solar attenuation as there is little to no difference in the computed solar exposure at Alice Spring, especially in the morning and afternoon when there were considerable amounts of forecast high cloud.

The observed solar exposure data for 8 May show that the day featured a large amount of diffuse solar exposure, suggesting a cloudy day. The observations recorded clear skies. It is highly likely that the diffuse solar exposure was created by aerosols. The forecast data predicted almost 100% high-cloud cover; this produces little decrease in the forecast solar exposure, however.

The forecast data for 9 May predicted only small amounts of cloud coverage. Therefore, the computed solar exposure is almost equal to the observed solar exposure for this day.

It is evident from Fig. 11 that, although the forecast cloud cover for these days was very different, there was almost no difference in the forecast solar exposure. This is further evidence that the optical depth of high clouds is too thin.

6. Discussion and conclusions

Spatial analysis of the satellite and NWP data from the MALAPS model suggested that the computed clouds in the tropics and over southeastern Australia were too optically thin and did not produce enough attenuation of solar irradiance. Clouds in the tropics during summer are often dominated by local convective clouds that occur during the day. Clouds over southeastern Australia are often created by orographic lifting around the Great Dividing Range, and many of the differences are seen in the vicinity of these ranges. It is possible that the satellite algorithm may also be incorrect over high topography because of the presence of snow and so on. This issue also needs further investigation.

Previous analysis of the cloud scheme by Rikus (1997) showed that an earlier version of the model could reproduce climatological cloud fields well but struggled with daily distributions. This global study used satellite measurements of IR brightness temperature to estimate cloud coverage and height. Results in Australia showed underestimation of mid- and high-level cloud associated with the Australian monsoon; for most of the year a strong cold bias was observed over Australia, however. This was attributed to excess model cloud as well as to sensitivity of the Geostationary Meteorological Satellite (GMS) 11-μm band to cold upper-level cloud. This early model version had insufficient resolution to accurately simulate orographic lifting. Values of cloud optical thickness were not validated, either.

Site analysis at various locations in Australia showed that the MALAPS model was able to predict the solar exposure accurately through low-level convective clouds (e.g., cumulus and stratocumulus), provided that the forecast cloud coverage was accurate. Results obtained at Darwin, Melbourne, and Alice Springs suggest that the NWP systems struggle to predict solar exposure through high clouds (e.g., cirrus, altocumulus, and altostratus). These clouds are possibly remnants from convective rain events that occurred earlier in the day. This site analysis also showed the presence of a seasonal bias in the satellite solar exposure product.

This problem with high-cloud cover is almost certainly due to the temperature-dependent parameterization of cloud condensate amount for very cold cloud. High-level (and higher midlevel) cloud is at relatively low temperature. The parameterization of cloud condensate is a diagnostic that is based on an empirical relationship between cloud temperature and condensate content. Cloud fraction is diagnosed as a quadratic function of the large-scale relative humidity. Recall from section 2 that the model’s mass-flux convection scheme only modifies the moisture and temperature distributions and therefore affects the cloud fields indirectly.

Because the cloud scheme is diagnostic (i.e., it has no “memory” from previous time steps), the model is unable to determine whether very cold cirrus cloud around the tropopause is thick anvil produced by recent local convection or is subvisual cloud that has spread out (and thinned out) over the region over longer time periods. To avoid complications due to overemphasizing this subvisual cirrus, the condensate parameterization is designed to produce very small condensate amounts at colder temperatures, which essentially removes the radiative effects of very cold cirrus. This does not discount the possibility that the model’s parameterization for ice crystal cloud may also have problems. Higher clouds are formed by ice crystals, and not by water vapor/droplets, and therefore they have different radiation scattering and absorption properties. Ice crystals in clouds can form a variety of crystal forms, which makes modeling of ice cloud optical properties difficult.

Further analysis showed that the LAPS and MALAPS models could compute daily solar exposure amounts on clear-sky days to within 3%; the forecasting skill of clear-sky days was poor, however, with only forecasts at Alice Springs and Broome showing some skill.

The analysis of the NWP systems has shown that it struggles to produce accurate forecasts of solar exposure in some areas of Australia and, in particular, in tropical regions during the summer, as well as in regions along the southeast coast of Australia close to the Great Dividing Range.

Note, however, that these areas are poor locations for any large-scale solar power plant, because the factors that create poor NWP predictions of solar exposure (i.e., cloud formations created by local effects such as convection or orographic lifting) are the same factors that would discourage a large solar power plant from being built in these regions.

In the midlatitudes of Australia, the NWP has shown the ability to provide good estimates of solar exposure. Annual MALAPS results at Alice Springs, Adelaide, and Wagga Wagga all showed values of correlation of greater than 0.8 with absolute relative percent error between 10% and 16%. Therefore, the current NWP system can provide useful forecasts of solar exposure at likely solar power plant locations.

As discussed previously, the LAPS and MALAPS models were replaced operationally in 2010 by the new ACCESS model, which is based on the Unified Model from the Met Office. Preliminary results using the ACCESS model have shown improvements in predicting solar irradiance attenuation through clouds and, in particular, improved treatment of transmission through high clouds. ACCESS also uses a prognostic cloud scheme.

Further work will involve testing the ACCESS model’s ability for forecast solar exposure. Additional cloud analysis would involve comparison of daily satellite cloud fields with the NWP cloud fields. This would help to determine how accurately the NWP code can forecast particular cloud types. Satellite products could also be used to detect significant aerosol concentrations in the atmosphere.

Acknowledgments

The authors gratefully acknowledge support from the Australian Department of Resources, Energy and Tourism (DRET) under the Wind Energy Forecasting Initiative for research. The National Energy Market Corporation (NEMCO) was also a stakeholder in the project. The authors are indebted to Bruce Forgan and Ian Grant for providing all of the information regarding the site and satellite data. The authors are very grateful to Ian Muirhead and Zhian Sun for their detailed and constructive reviews and comments.

APPENDIX A

Satellite Data

The semiempirical model accounts for Rayleigh scattering, surface albedos, water vapor and ozone absorption, and, in the case of cloudy skies, cloud albedo and absorption. Total column water vapor amount, a required input to the model, is taken from a LAPS model forecast for near-real-time processing. Aerosols are ignored in the model, but their impact is accounted for by tuning model output against surface observations. Surface albedos are derived from cloud-free satellite data and are assumed to remain constant for 4–6 weeks. The albedos are a function of local view angle and sun angle. The data from each satellite image are used to calculate cloud albedos in the satellite measurement band (0.55–0.90 μm) from which broadband cloud albedos are derived. The relationship between cloud albedos and cloud absorption in the visible and infrared segments of the satellite band is based on data from Welch et al. (1980). Water vapor transmittances are calculated using methods outlined in Iqbal (1983). Ozone absorption is computed from total column ozone data, which are prescribed as a fixed function of latitude according to the formula
ea1
where ϕ is latitude. Equation (A1) gives ozone amounts in Dobson units. Estimates of temperature and moisture are computed using raw cloudy radiances with a perturbation form of the radiative transfer equations.

The model output is tuned using the method described in Weymouth and Le Marshall (1999, 2001). The algorithms can be bias corrected by comparing the surface site-based data, which are unfortunately limited over the Australian continent. To prevent overestimation of irradiance with increasing cloud albedo, regression corrections to the model irradiance and surface data were developed. These corrections implicitly correct for sensor drift and relate normalized deviations from surface data and average cloud albedo. The mean absolute error of model daily solar exposure from site measurements generally cycles annually from approximately 0.8 MJ m−2 in winter to 1.5 MJ m−2 in summer, as compared with an annual exposure range of approximately 5–30 MJ m−2 (Grant et al. 2008).

The “tuning” is not of the model itself, but is a removal of the bias in the output satellite daily exposure on a monthly basis by reference to the surface station observations. The model and its coefficients are static, whereas the bias correction is dynamic but once done should normally not need changing.

APPENDIX B

Site Data

Each site uses CM-11 pyranometers manufactured by Kipp and Zonen. The calibration of the pyrheliometers, pyranometers, and pyrgeometers is performed in situ at the stations after an initial characterization at the bureau’s Regional Radiation Centre using a variety of calibration methods as specified by the WMO Commission for Instruments and Methods of Observations (WMO 2012, part I: chapter 7) and the BSRN operations manual (McArthur 2005). The pyranometer characterization ensures that the pyranometers used at a station have characteristics that are matched as closely as possible to reduce uncertainty during the various in situ calibration processes and to minimize the influence of radiance distribution on the diffuse calibration value. All of the resultant calibrations and measurements are part of traceable hierarchies to the WMO World Radiometric Reference and the World Standard Group (WMO 2012, part I: chapter 7).

An example of the calculation of monthly averaged uncertainty is given below. Values of daily exposure and uncertainty estimates (for both 95% and 66% confidence levels) for January 2008 at Melbourne Airport are given in Table B1. The mean and standard deviation of the daily exposure E are equal to
eb1
and the sum of the daily squared U66 values and its standard deviation are equal to
eb2
where number of samples N = 30.
Table B1.

Values of daily solar exposure and daily 95% uncertainty (MJ m−2) for Melbourne Airport in January 2008.

Table B1.
The standard uncertainty due to the standard deviation in mean exposure is equal to
eb3
Therefore the total standard uncertainty for mean exposure is equal to
eb4
For a 95% probability, the value of the (two sided) Student’s t inverse cumulative distribution with the degrees of freedom ν = 29 is
eb5
so that the total 95% uncertainty for the mean daily exposure at Melbourne Airport for January 2008 is equal to
eb6
This is the value that is given in Table 2.

REFERENCES

  • Breitkreuz, H., , M. Schroedter-Homscheid, , T. Holzer-Popp, , and S. Dech, 2009: Short-range direct and diffuse irradiance forecasts for solar energy applications based on aerosol chemical transport and numerical weather modeling. J. Appl. Meteor. Climatol., 48, 17661779.

    • Search Google Scholar
    • Export Citation
  • Dare, R. A., 2004: The BMRC Bulk Explicit Microphysics (BEM) scheme. Australian Bureau of Meteorology Tech. Rep. 99, 110 pp.

  • Grant, I., , D. Jones, , W. Wang, , R. Fawcett, , and D. Barratt, 2008: Meteorological and remotely sensed datasets for hydrological modelling: A contribution to the Australian Water Availability Project. Extended Abstracts, CAHMDA-III Int. Workshop on Hydrologic Prediction: Modelling, Observation and Data Assimilation, Melbourne, VIC, Australia, Australian Research Council, 23–51.

  • International Organization for Standardization/International Electrotechnical Commission, 2008: Uncertainty of measurement—Part 3: Guide to the expression of uncertainty in measurement (GUM:1995). ISO/IEC Guide 98-3:2008(E), 130 pp.

  • Iqbal, M., 1983: An Introduction to Solar Radiation. Academic Press, 390 pp.

  • Lacis, A. A., , and J. E. Hansen, 1974: A parameterization for the absorption of solar radiation in the earth’s atmosphere. J. Atmos. Sci., 31, 118133.

    • Search Google Scholar
    • Export Citation
  • Lemus, L., , L. Rikus, , C. Martin, , and R. Platt, 1997: Global cloud liquid water path simulations. J. Climate, 10, 5264.

  • Lorenz, E., , J. Hurka, , D. Heinemann, , and H. G. Boyer, 2009: Irradiance forecasting for the power prediction of grid-connected photovoltaic systems. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 2, 210.

    • Search Google Scholar
    • Export Citation
  • McArthur, L. J. B., 2005: World Climate Research Programme Baseline Surface Radiation Network (BSRN) operations manual, version 2.1. WCRP-121 and WMO Tech. Rep. WMO/TD 1274, 188 pp. [Available online at http://www.wmo.int/pages/prog/gcos/documents/gruanmanuals/WCRP/WCRP21_TD1274_BSRN.pdf.]

  • Ohmura, A., and Coauthors, 1998: Baseline Surface Radiation Network (BSRN/WRMC): New precision radiometry for climate research. Bull. Amer. Meteor. Soc., 79, 21152136.

    • Search Google Scholar
    • Export Citation
  • Perez, R., , K. Moore, , S. Wilcox, , D. Renne, , and A. Zelenka, 2007: Forecasting solar radiation—Preliminary evaluation of an approach based upon the national forecast database. Sol. Energy, 81, 809812.

    • Search Google Scholar
    • Export Citation
  • Puri, K., , G. S. Dietachmayer, , G. A. Mills, , N. E. Davidson, , R. A. Bowen, , and L. W. Logan, 1998: The new BMRC Limited Area Prediction System, LAPS. Aust. Meteor. Mag., 47, 203223.

    • Search Google Scholar
    • Export Citation
  • Ramaswamy, V., , and S. M. Freidenreich, 1992: A study of broadband parameterizations of the solar radiative interactions with water vapor and water drops. J. Geophys. Res., 97, 11 48711 512.

    • Search Google Scholar
    • Export Citation
  • Reikard, G., 2009: Predicting solar radiation at high resolutions: A comparison of time series forecasts. Sol. Energy, 83, 342349.

  • Rikus, L., 1997: Application of a scheme for validating clouds in an operational global NWP model. Mon. Wea. Rev., 125, 16151637.

  • Seaman, R., , W. Bourke, , P. Steinle, , T. Hart, , G. Embery, , M. Naughton, , and L. Rikus, 1995: Evolution of the Bureau of Meteorology’s Global Assimilation and Prediction system. Part 1: Analysis and initialisation. Aust. Meteor. Mag., 44, 118.

    • Search Google Scholar
    • Export Citation
  • Tiedke, M., 1989: A comprehensive mass flux scheme for cumulus parameterization in large-scale models. Mon. Wea. Rev., 117, 17791800.

    • Search Google Scholar
    • Export Citation
  • Vincent, C. L., , W. P. Bourke, , J. D. Kepert, , M. Chattopadhay, , Y. Ma, , P. J. Steinle, , and C. I. W. Tingwell, 2008: Verification of a high-resolution mesoscale NWP system. Aust. Meteor. Mag., 57, 213233.

    • Search Google Scholar
    • Export Citation
  • Welch, R. M., , S. K. Cox, , and J. M. Davis, 1980: Solar Radiation and Clouds. Meteor. Monogr., No. 39, Amer. Meteor. Soc., 96 pp.

  • Weymouth, G. T., , and J. F. Le Marshall, 1999: An operational system to estimate global solar exposure over the Australian region from satellite observations 1. Method and the initial climatology. Aust. Meteor. Mag., 48, 181195.

    • Search Google Scholar
    • Export Citation
  • Weymouth, G. T., , and J. F. Le Marshall, 2001: Estimation of daily surface solar exposure using GMS-5 stretched-VISSR observations: The system and the basics. Aust. Meteor. Mag., 50, 263278.

    • Search Google Scholar
    • Export Citation
  • Wittman, M., , H. Breitkreuz, , M. Schroedter-Homscheidt, , and M. Eck, 2008: Case studies on the use of solar irradiance forecast for optimized operation strategies of solar thermal power plants. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 1, 1827.

    • Search Google Scholar
    • Export Citation
  • World Meteorological Organization, 1975: International cloud atlas, volume I: Manual on the observation of clouds and other meteors. WMO Tech. Rep. 407, 155 pp.

  • World Meteorological Organization, 1987: International cloud atlas, volume II [photographs]. WMO Tech. Rep. 407, 206 pp.

  • World Meteorological Organization, 2012: Guide to meteorological instruments and methods of observation: 2008 edition, updated in 2010. WMO Tech. Rep. 8, 716 pp. [Available online at ftp://ftp.wmo.int/Documents/MediaPublic/Publications/WMO8_CIMOguide/WMO8_Ed2008_Up2010_en.pdf.]

  • Zamora, R., , E. G. Dutton, , M. Trainer, , S. A. McKeen, , J. M. Wilczak, , and Y.-T. Hou, 2005: The accuracy of solar irradiance calculations used in mesoscale numerical weather prediction. Mon. Wea. Rev., 133, 783792.

    • Search Google Scholar
    • Export Citation
Save