ABSTRACT

The potential of geostationary satellites for world-wide solar resource assessment is generally recognized. However, most studies evaluating the value of the satellite’s approach have tended to focus on the accuracy of satellite-derived irradiance values, rather than on the strengths that satellite techniques bring to solar resource assessment. This paper presents strong evidence that, when assessed against realistic benchmarks relevant to the needs of the solar energy community, satellite-based resource assessment can provide considerably more information, more accurately than traditional approaches.

1. INTRODUCTION

The well-documented shortcomings of the satellite approach include a short term global irradiance error (RMSE) on the order of at least 20% with respect to a ground measuring station, poor performance in winter when ground is snow covered, and an inability to precisely detect turbidity changes [1].

Although these shortcomings are real, we make the case that they represent a one-sided measure of remote sensing true capabilities for solar radiation resource assessment -- at least as far as solar energy applications are concerned.

Most solar energy applications are concerned with two types of solar radiation resource information: (1) climatological information including tables, maps and historical/typical data sets; and (2) time/site specific data including past, real time, and, eventually, forecasted data. This paper presents initial evidence that the strength of the satellite resource is considerable for both types of data, when gauged against the available alternatives.

2. CLIMATOLOGICAL RESOURCE ASSESSMENT

When a utility, a local government or a project developer wants to know about the solar energy resource in their area of interest, they generally have two choices:

1. To consult existing data sets (e.g., SOLMET/ERSATZ, NSRDB, TMY2 [2,3]), and maps and tables drawn from these data sets
2. If they can afford it, to conduct a measurement campaign involving one or more instruments distributed over a network.

The reason why one needs localized climatological solar resource information is to be able to perform site-specific simulations, assessments and feasibility studies. It is therefore important that the solar resource information properly account for site-specificity and regional/local differences.

In this paper, we compare climatological information derived from satellite remote sensing with the two alternatives mentioned above.

Our comparison benchmark is the regional distribution of summer global irradiance – the only season for which we had a long enough satellite data set as of this writing. The test region covers the eastern United States from Virginia to Vermont and from western Pennsylvania to southern Maine.

2.1 Experimental data

Climatological Data Set: We chose the TMY2 data set [3] as the representative of existing climatological data sets. This is now widely used for engineering applications. The TMY2 set is a reflection of climatological trends for the 1961-to-1990 time frame.

Instrumented Network: Our test ground measurement network consists of a combination of the ASRC solar radiation monitoring station and the rotating-shadowband-based southern New York/Massachusetts network [4]. Data in this network were subjected to both visual and automatic QC and, inasmuch as possible, corrected a posteriori for calibration drifts (see [5]).The climatological map drawn from this network is a six year average (1991-1995).

Satellite: Since this study makes use of GOES-8 data that was not distributed until mid-1995, the climatological satellite maps are only based on two years worth of data. The satellite methodology was described in [5]. Briefly, the data consist of intermediate resolution (»10 km pixel) GOES-8 images [6]. The model itself is an empirical model based on the linear relationship between the satellite signal (corrected for non-linear sensor response and solar elevation) and the relative clearness expressed as a fraction of a standard clear sky model. The model accounts for ground albedo variations through space and time, and was recently modified to account for atmospheric scattering effects.

This model is an adaptation of Cano’s empirical model [7] which has been validated in Europe [8]. In the northeastern US, the model was validated against twenty ground truth stations, including the network stations mentioned above and nine other rotating shadowband radiometers located from Virginia to central New York State [9,10]. Gauged against 43,000 ground truth station-hours, the average satellite RMS error for global irradiance is 81 W/m², that is 23% of average global irradiance, while the average mean bias error is 4 W/m2, that is 1.2% of measured global irradiance.

2.2 Satellite vs. TMY2:

The summer global irradiance maps derived from satellite and TMY2 data are compared in Fig.1. We used kriging to generate the isopleths. Both maps were calibrated so they would agree with measurements at our Albany solar radiation station.

The contrast in spatial resolution and dynamic range between the two maps is striking. The satellite map offers a well defined micro-climatic differentiation capability unmatched by the TMY alternative. These differences are larger than one would expect from the variability attributable to their different time references (1961-1990 for TMY2 and 1995-1996 for the satellite).

Most features of the satellite map are traceable to logical micro-climatic factors such as orography and coastal effects. See for instance the signature of the Hudson River in eastern New York (more solar radiation), of the Green Mountains in Vermont or the Allegheny mountains in western New York/Pennsylvania (less solar radiation). Higher global radiation characterizes the cold waters of the great lakes and the Atlantic Ocean – note the large enhancement of radiation on the shores of the Delmarva Peninsula and the New Jersey coast.

Figure 1: Comparing summer global irradiance maps derived from satellite and from TMY2(NSRDB)

Quantitatively, the satellite map shows an irradiance dynamic range of 35% between values in the Vermont mountains and the Delmarva peninsula.

The TMY2-derived map features a broad trend that is consistent with the satellite’s, but its medium and small scale features are not. Most noticeable is the relative minimum in southern New England, which appears unjustified by local orography or other climatic features. A most important difference between the two maps is their dynamic range, which, here, reaches only 12% -- three time less than above.

With 27 sites, the TMY2 set paints a very uniform picture of radiation distribution in the region, which may not be fully reflective of its solar resource diversity. Two possible causes for this are: (1) the non-optimal location of stations and (2) the dampening effect of meteorological models which were used to generate most NSRDB-TMY2 solar radiation data.

2.3 Satellite vs. Ground Measurements:

The summer global irradiance maps derived from satellite and ground measurements are compared in Fig. 2a and b. As above, we used kriging to generate the isopleths.

Figure 2: Comparing satellite derived and ground based summer global irradiance maps

The satellite map is a close-up of Fig. 1, focusing on the ground network area. Its main features are the coastal effects (e.g., notice the signature of Long Island, which promotes afternoon convective cloud formation); the Hudson Valley enhancement; and the orographic minimums (Catskill Mountains in the west and Green Mountains in the north). The dynamic range for this area is 20% from southern Vermont to the southern shore of Long Island.

The ground-based map exhibits an overall north-south gradient that is consistent with the satellite map, and its dynamic range of 14% is comparable. However its local features are very different, and seem to be driven by instrumentation rather than climate: cosine response differences, remaining calibration uncertainties and instrument maintenance schedules are possibly the cause of localized features which do not appear to be linked to micro-climatic factors. As we suspected that the non ideal sampling of ground stations may be partly responsible for the pattern in Fig 2-b, we also produced a satellite map derived only from observations at the instrument sites. This map, shown in Fig. 2-c, is not as detailed as the map in Figure 2-a, but its overall pattern is more in line with climate/terrain-based expectations

3. OPERATIONAL RESOURCE ASSESSMENT

Time/site specific data are needed for applications were the time/location element is important. They are necessary, for instance, for the determination/control of the impact of a dispersed solar power generation base on utility or local (e.g., buildings, substations) electric loads [e.g., see 11].

The determining criterion for these data is their short term RMS error. We said above that the satellite hourly RMSE was of the order of 23%, which may sound quite large if left unqualified. We propose here to qualify this parameter in a realistic context.

This context is illustrated in Fig. 3 (from [5]): when the hourly RMS error between neighboring ground stations is plotted as a function of their distance, one observes a sharp rise near the origin: within the first ten kilometers, the RMSE reaches 17%. In geostatistical terms, this quasi-discontinuity at zero distance is known as the "nugget effect". Here, the nugget effect reflects (1) measurement errors and instrument-to instrument variability, which is likely to amount to 3-5% [1], and (2) the discontinuous and fractal nature [12] of cloud patterns on an hourly time scale.

Hence, qualifying satellite accuracy is highly dependent on the user’s considered spatial resolution – note that this argument is not unlike the classical fractal example showing that the length of a coastline is a function of the length of the yardstick used to measure it [12].

Figure 3: Relative extrapolation RMSE as a function of station distance

Given the present scarcity of solar radiation measurement stations, and the reality of today’s solar resource assessment, a ground resolution of the order of one intermediate resolution satellite pixel (10 km) would constitute a momentous improvement. However, the data in Fig. 3 show that, even if perfect instruments were deployed over such a high resolution grid -- amounting to 80,000 radiometers for the US alone -- the user could not expect to achieve an accuracy better than 13% for points located between stations.

3.1 Concept of effective Accuracy

It is tempting to introduce the concept of effective accuracy, and to define an "effective" satellite RMSE, function of its ground resolution context. In the above example, the effective satellite RMSE would only be 10% (i.e., its 23% pinpoint accuracy, minus 13% highest achievable accuracy on a 10Km network grid).

More pragmatically, let us consider the example of the Albany metropolitan area (see Fig. 4). This figure shows a sample intermediate resolution GOES-8 image overlaid on a local administrative map. The Albany solar radiation monitoring station is located at the ASRC, about 9 km from downtown Albany. Although this station features WMO class 1 instruments and has a stringent daily maintenance schedule, one could not expect an RMS accuracy better than 15% when extrapolating the data for a site/time specific application in downtown Albany. For the Albany county as a whole, where the average distance from the Station is 25 km, the accuracy achievable by extrapolating from our station would only be 21%. For nearby Rensselaer county, home of much of the Albany metropolitan area, but at an average distance of 42 km from our station, the achievable accuracy would only be 25%.

Thus, for downtown Albany, Albany county, and Rensselaer county, the effective satellite RMSE would respectively be 8%, 2% and 0%.

Figure 4: GOES 8 Intermediate resolution image close-up around Albany

4. DISCUSSION

4.1 Main Conclusions

We have shown that solar resource assessment from geostationary satellites constitutes a powerful alternative to ground network and meteorological data sets for both climatological and operational data.

Climatologically, the satellite provides a detailed view of regional solar resource diversity that is unmatched by the traditional approaches.

Operationally, the satellite is far ahead of today’s alternatives. We introduce the measure of effective RMSE, to quantify the satellite’s advantage in terms of ground resolution needs and ground station deployment feasibility.

4.2 This Assessment is Conservative

These observations were reached with a simple satellite model that leaves ample room for performance improvement. For instance, we made no attempt to account for solar position in the determination of the pixel, or group of pixels, relevant to a point on the ground. In our comparison with ground stations, we only considered the closest pixel. No attempt was made either to fine-tune satellite navigation, using ground features such as water-land interface – our day-to-day observations show that an uncertainty of half a pixel or more is not uncommon. These two factors alone introduce a potentially correctable error of 5-10 km in the determination of the exact satellite model location. Referring to Fig.3, these two accountable factors amount to about half of the satellite’s RMSE. The RMSE of the satellite could thus approach 10-15% pinpoint accuracy.

4.3 Need for ground Measurements

The question then arises: Are ground instruments needed for solar resource monitoring?

Our tentative answer is two-fold:

1. A small number of high accuracy ground stations are needed for satellite model ground truth and real time calibration, particularly to complement the satellite’s noted weaknesses: snow covered grounds and clear sky turbidity levels. We do not sense, however that there is a need for high density networks.
2. Instruments are needed when ground resolution requirements are very high, as for the monitoring of an experimental solar system for the purpose of characterizing its operational efficiencies. In this case, solar radiation sensors located at the project site would be appropriate.

ACNOWLEDGEMENT

This work is supported by NYSERDA (contract No. 4126-ERT-TERER-95, J. Harvey, Officer) and NREL (subcontract No. XAH-515-22-201). Many thanks to Ascension Technology, to Joe Michalsky, to the Niagara Mohawk Power Corporation and to AWS Scientific for providing us with irradiance data.

1. J. Schmetz J.(1989): Towards a Surface Radiation Climatology: Retrieval of Downward Irradiances from Satellites, Atmospheric Research, 23, 287-321
2. E. Maxwell, W. Marion, D. R. Myers, M. Rymes and S. Wilcox, (1995): The National Solar Radiation Data base, NREL/TP-463-5784 (1994), NREL, Golden, CO
3. W. Marion and S. Wilcox, (1994): Solar Radiation Data Manual for Flat Plate and Concentrating Collectors. NREL/TP-463-5607 1994, NREL, Golden, CO
4. New York State Energy Research and Development Authority’s Southern New York State Solar Radiation Network (1991-1995), including ATI’s Waltham station, NYSERDA, Albany, NY
5. R. Perez, R. Seals and A. Zelenka, (1996): Production of Site/Time Hourly Irradiances – satellite remote Sensing vs. Network Interpolation. Proc. ASES-96 Annual Meeting, Asheville, NC, 6 pp., Solar Energy (in press)
6. Internet Data Distribution System. UCAR, Boulder, CO
7. D. Cano, J.M. Monget, M. Albuisson, H. Guillard, N. Regas, and L. Wald (1986): A method for the determination of the global solar radiation from meteorological satellite data", Solar Energy 37, 31-39.
8. L. Diabate, G. Moussu, and L. Wald (1989): Description of an operational tool for determining global solar radiation at ground using geostationary satellite images, Solar Energy 42, 201-207
9. E. Kern, Private Communication, (1995-96): Data Acquisition at selected USEPA Photovoltaic Project sites.
10. J. Michalsky, Private Communication (1995-96): Data acquisition at selected USDOE Quantitative Links network sites
11. R. Perez, R. Seals and C. Herig, (1996): Determination of Photovoltaic’s Effective Capacity. Interim Report No. 2. NREL Contract XR-1-11168-1
12. B. B. Mandelbrot (1982): The Fractal Geometry of Nature. W.H. Freeman and Co., New York