Abstract. Introduction

Transcription

1 Evaluation of the ISCST2 model with measurements of SO^ concentrations in the greater Cape Town region D.A. Dracoulides, R.K. Dutkiewicz Energy Research Institute, University of Cape Town, P.O. Box 207, 7800 Plumstead, Cape Town, South Africa Abstract This paper presents the evaluation of the Industrial Source Complex Short Term 2 (ISCST2) model with SOi measured concentrations at three monitoring stations in the urban Cape Town area. The model performance was quantified by utilising several statistical measures, as proposed in the Woods Hole EPA/AMS Workshop, as well as by additional parameters, such as the systematic and unsystematic mean square errors (MSE<., MSE J, the index of agreement (D) and the fractional bias (FB). It was found that the Ih and 24h predicted concentrations were within the acceptable accuracy limits only at one monitoring station. However, the Ih maximum predicted concentrations, independent of time of occurrence during the day, were in better agreement with the observations at all three sites. The model performance was poor under very stable atmospheric conditions (classes: E and F), as well as under low wind speeds (< 3 m s"^) at all three sites. The statistical measures and the bootstrap analysis revealed that the use of complicated schemes to derive the meteorological parameters, from one-point meteorological observations, did not improve the model performance. Introduction Cape Town is a large metropolitan area with heavy industrial activity and contains commercial and residential areas with different fuel consumption patterns, such as the informal housing areas. An emission inventory based on the fuel consumption by each sector was created, in order to constitute the SO? emissions input into the ISCST2 model. On a monthly basis tons of SOi, tons of NO% and tons of paniculate matter were emitted in this area during the year Figure 1 illustrates the locations of the main emitting sources, as well as the tliree monitoring stations at Bellville (SI), Cape Town's

2 108 Urban Pollution CBD (S2) and Goodwood (S3), which were used for the model evaluation. A total of 187 point and 263 area sources were allocated to the greater Cape Town (GCT) area. kg/m < ,500 1,500-3,500 o 3,500-8,000 o > 8, km Figure 1: The Greater Cape Town (GCT) region with the main industrial sources of SCX and the three monitoring sites: SI, S2 and S3. Meteorological parameters The meteorological parameters needed for the model runs were calculated from routinely available meteorological data collected from Cape Town airport weather station (see Fig 1). A total of 1272 and 936 hours were selected from the years 1991 and 1992 respectively. The 1991 data set consisted of 16 summer and 37 winter days. All hours of the 1992 set belonged to the winter period. Three schemes for determining the mixing height and three schemes for determining the atmospheric stability class were utilised for the model, and the accuracy for each combination was assessed. All schemes were adopted to utilise one-point meteorological measurements. The methods for the atmospheric stability computation were the Pasquill scheme, the inverse Monin-Obukhov (1/L) length and the modified Kazanski-Monin (p/) parameter (Sutherland et au). The Monin-Obukhov length boundary values

3 Urban Pollution 109 between the successive stability classes were obtained from Tagliazuka and Nanr for a roughness length of 0.5m. The friction velocity (u*) was estimated, using the earth's energy balance to calculate the potential temperature (6*) and iteratively solving the friction velocity and the M-O length equations (Venkatranf). The mixing height was estimated utilising midday and midnight sounding at Cape Town airport. The first method to calculate the heights was based on the pressure equation. The second utilised the Holzworth"* procedure, which is also used by the U.S. EPA meteorological preprocessors. The third method adopted several empirical equations, according to the atmospheric condition and was based on the heat exchange between the earth's surface and the atmosphere (Ludwig et al/; van Ulden and Holtslag^; Venkatranf; San Jose^). Combining the above-mentioned methods, nine data sets were produced for the ISCST2 meteorological input, and the model performance was assessed for each set. Model performance evaluation The overall evaluation of the ISCST2 model was based on the comparison of hourly predictions of SO-,, with concentrations measured at the three monitoring stations. For each monitoring station nine data sets of predictions were produced, according to the methods used to calculate the meteorological parameters. The evaluation of the ISCST2 model included paired and unpaired analyses. The measures employed were: the mean predicted (T ) and observed (O) values, the slope and intercept of a linear regression, the root mean square error (RMSE), the systematic and unsystematic mean square errors (MSE^, MSEJ, the normalised mean square error (NMSE), the index of agreement (D) and the fractional bias (FB)^'*. The uncertainty of parameters, such as the fractional bias, index of agreement and mean difference (ct ), was estimated by employing the blocked bootstrap resampling procedure (Efron^). Examination of the paired and unpaired statistics for all the meteorological combinations did not reveal any significant differences in the model performance. The bootstrap procedure was also applied for each set to the mean difference, index of agreement and fractional bias. Comparison of the 95% confidence intervals of these statistical measures did not reveal any significant improvement in the model predictions. However, the Holzworth procedure for the mixing height combined with the Pasquill or M-O length method for the stability class produced slightly better results than the rest of the combinations. The paired and unpaired statistical measures for the Pasquill-Holzworth data set are depicted in Table 1. The ISCST2 clearly performed better at site SI. At the remaining sites S2 and S3, the model under-predicted the average measured concentration by about 72% and 58% respectively.

4 110 Urban Pollution The regression parameters (intercept and slope) suggest a linear under-prediction at all three sites. However, these measures should be interpreted with caution since they are not independent from one another. Ideally, the intercept should be 0.0 and the slope 1.0. This condition was more closely met at site SI. The RMSE at site SI is the highest, which reveals that extreme differences (P;-Oj) exist in the data set. At sites S2 and S3 the RMSE, which is a high estimate of the average error, indicates that the model systematically under-predicts the observed concentrations. This is reinforced by the high systematic RMSE^ which is a linear function of the differences (P^-O;). In contrast, the errors produced at site SI are random. Thus, a simple parameterisation of the model or a more accurate emission inventory would be unlikely to improve the model performance. Table 1 Summary of paired statistics for the Ih measured and predicted concentrations at Bellville (SI), Cape Town CBD (S2) and Goodwood (S3) sites. Bellville (SI) CBD (S2) Goodwood (S3) Obs. Pred. Obs. Pred. Obs. Pred Sample size Range" Mean* Standard deviation (STD)'' Average P/O Mean difference'* Intercept Slope Fractional bias (FB) Index of agreement (D) NMSE RMSE* RMSE, MSEu/MSE MSEJMSE % 8% % 86% % 53% * The units of range, mean, STD, mean difference and RMSE are f. ig/nf The fractional bias (FB) and the normalised mean square error (NMSE) for site SI indicate that the model generally reproduces the Ih observed concentrations within a factor of two. The high systematic mean square error (MSEJ at sites S2 and S3 indicate that the model's poor performance could be a result of improper input information, such as the emission strengths, source characterization (point-area) or the meteorological parameters being unrepresentative for that particular area. The central city lies in a bowl which is formed by Table Mountain and other hills in the area. It is well known that within this bowl, the direction and vigour of the advection of pollutants could

5 Urban Pollution 111 be different to the ones measured at Cape Town airport^'. The high SO? readings could also be due to air pollution trapping, as a result of air recirculation regimes formed at the lees of Table Mountain, or the oscillation of pollutants to and from the shore, caused by land and sea-breezes. Evaluation according to atmospheric stability and wind velocity The results of the model performance, according to atmospheric stability, are summarised for site SI in Table 2. Three stability groups were used. The first was the unstable and comprised the P-G classes A to C, the second was the neutral class (D) and the third consisted of the stable classes E and F. The results of this analysis show that the model under-predicts under unstable and neutral conditions and over-predicts at stable conditions. The index of agreement indicates that the model is approximately 12% more accurate under unstable conditions than under stable atmospheric conditions. This finding is also reinforced by the fractional bias of and the high percentage of the unsystematic mean square error of 89%. For the stable conditions the FB is 0.83 andthemsejmse is 51%. Table 2 Summary of paired statistics according to atmospheric stability for the monitoring station at Bellville (SI) (unstable A-C, neutral D, and stable E-F). A-C (Unstable) D (Neutral) E-F (Stable) Obs. Pred. Obs. Pred. Obs. Pred. Sample size Range" : Mean" Standard deviation (STD)* Average P/0 Mean difference^ Intercept Slope Fractional bias (FB) Index of agreement (D) NMSE RMSE* RMSE,, RMSEg MSE./MSE MSEJMSE % 11% % 50% % 49% * The units of range, mean, STD, mean difference and RMSE are j Lg/m" A similar analysis for the monitoring stations S2 and S1 revealed that the model under-estimates the concentrations under all stabilities and produces the highest predictions under stable conditions. The index of agreement at site S2 indicates that, similarly to the Bellville station, the model predicts about 10% more

6 112 Urban Pollution accurately under unstable rather than under stable atmospheric conditions. When the same data set is grouped according to wind velocity (< 3 m s"\ 3-6 m s"\ > 6 m s~*), the poorest agreement between predictions and observations is evident under low wind speeds (< 3 m s~*). At Bellville monitoring station (SI), the model predictions over-estimate the observed concentrations under low wind speeds. Similarly, predictions at the other two sites (S2, S3) are highest in the low speed group. Nevertheless, consistent with the Ih analysis, the observed concentrations at sites S2 and S3 are under-estimated for all wind categories. Maximum concentrations The maximum concentrations are very important for regulatory applications of dispersion models. The measured and predicted Ih maximum concentrations independent of the time of occurrence during the day were selected for each day, in order to assess the simulation of the magnitudes of the measured concentrations. Table 4 summarises the statistics for the maximum concentrations. The model reproduces the range and, in particular, the upper boundary of the concentrations approximately within a factor of two at all three sites. Predictions at CBD (S2) monitoring station systematically under-estimate the maximum concentrations, as suggested by the systematic fraction of the mean square error (MSE<,/MSE=67%). The fractional bias of indicates that, on average, the predicted maximum concentrations are within a factor of two from the measured ones. A better performance is evident at the Goodwood site (S3) (see Table 3). In contrast to the general under-prediction shown by the one hour concentrations at this location, the predicted maxima are, on average, 21% higher than the observed. The unsystematic fraction of the MSB reveals that the model performs closer to its potential accuracy at the Goodwood (S3) location, since 81% of the mean square error is unsystematic. The closest fractional bias (FB) to the desired value of 0.0 is also the one at site S3. At Bellville monitoring station (SI) the observed maxima are over-estimated approximately by a factor of two (Table 3). The mean maximum predicted concentration is 106 ig/m\ when the mean observed is 56 ig/nf. The systematic fraction of the MSB is high (42%). This indicates a possible systematic cause for the over-prediction, such as the emission strengths. The fractional bias is within the ±0.67 range at all three monitoring stations. When the FB is out of these limits, predictions are considered to be over- or under-estimated by a factor greater than two. In general, the model seems to perform poorly at Cape Town's CBD site (SI), since the maximum concentrations are under-estimated. The magnitude of over-prediction at the Bellville site (SI) could possibly be explained by the lack of detailed emission information available for the Bellville area. The portion of the SO? emissions

7 Urban Pollution 113 in the Bellville magisterial district, which was allocated to point sources was approximately 25% of the total, whilst it was 40% for Goodwood and 55% for Cape Town. Table 3 Summary of statistics for the one hour maximum observed and predicted concentrations independent of time of occurrence during the day. CBD (S2) Bellville (SI) Goodwooci(S3) Obs. Pred. Obs. Pred. Obs. Pred. Sample size Range* Mean* Standard deviation (STD)* Average P/0 Mean difference* Intercept Slope Fractional bias (FB) Index of agreement (D) NMSE RMSE* RMSE,, RMSE, MSEu/MSE MSEJMSE % 67% , % 42% E; % 19% The above-mentioned findings are also reinforced by the bootstrap analysis of the fractional bias (FB) and mean difference of the maximum concentrations. The 95% confidence intervals for Bellville are above the zero line which reveal a systematic over-prediction. At Goodwood site (S3) the bounds of the FB and mean difference overlap with zero indicating that 95% of the time the model predicts within the acceptable accuracy limits. At site SI, consistent with the Ih concentrations, the bootstrap analysis revealed an under-prediction. However, the difference was not greater than a factor of two. CONCLUSIONS The objective of this work has been to test and evaluate the performance of the ISCST2 model, using observed SO? concentrations at three locations in the Greater Cape Town region. It was found that the model performed poorly at the second monitoring station (S2), situated in Cape Town's Central Business District (CBD). The meteorological conditions of that area are complicated, due to the sheltering effects of Table Mountain and the land and sea-breezes of Table Bay. Therefore, the Cape Town airport atmospheric measurements may not provide representative wind data. A second possible explanation is the model's lack of

8 114 Urban Pollution a fumigation algorithm, in order to simulate the observed trends of the concentrations along the shorelines and thirdly, the lack of treatment in the model of physical processes, such as locally produced turbulence and the heat-island effect, typically found in urban areas. The overal model performance at locations with uncomplicated terrain and away from Table mountain and the shorelines was within the acceptable limits. The maximum concentrations independent of time of occurrence during the day were predicted within a factor of two at all three sites. Under stable atmospheric conditions and low wind speeds the ISCST2 model performed poorly at the three stations. As a result of this study, it was proposed, firstly, to examine further the model's accuracy by utilising more monitoring sites, as well as different locations for the derivation of the meteorological parameters. Secondly, the compilation of a more accurate emission inventory data was recomended, in order to implement the building down-wash option of the ISCST2 model and thirdly to intercompare the results against other models which can simulate the fumigation process and the heat-island effect. REFERENCES 1 Sutherland, R.A., Hansen, F. V. & Bach W.D. A quantitative method for estimating Pasquill stability classes from wind speed and sensible heat flux density. Boundary Layer Meteorology, 1986, 37, Tagliazucca, M. and Nanni, T. An atmospheric diffusion classification scheme based on the Kazanski-Monin stability parameter Atm. Env., 1983, 17, Venkatram, A. Dispersion in the Stable Boundary Layer, in: Lectures on air pollution modeling, eds Venkatram, A. and Wyngaard J.C. American Meteorological Society, Boston, Holzworth, C. G. Mixing heights wind speeds and potential for urban air pollution throughout the contiguous United States. Publication No. AP-101. U.S. Environmental Protection Agency, Research Triangle Park, North Carolina, Ludwig, F.L., Johnson, W.B., Moon, A.E. & Mancuso, R.L. A practical multipurpose urban diffusion model for carbon monoxide, Final Report Coordinating Research Council (CRC) Contract CAPA-3-68 and Nat. Air Poll. Control Admin. Contract CPA , Van Ulden, A.P. and Holtslag, A.M. Estimation of atmospheric boundary layer parameters for diffusion applications. Bull. Amer. Met. Soc., 1985,24, Jose, R.S. A simple approach to evaluate mixed layer depth. Environmental Software, 1991, 6, Fox, D.G. Judging air quality model performance. Bulletin American Meteorological Society, 1981, 62, Hanna, S.R. Air quality model evaluation and uncertainty. JAPCA, 1988, 38, No 4, Rao, K.S., Ku J. and Rao, S.T. A comparison study of three urban air pollution models. Atm. Env., 1989, 23, No 4, Efron, B. The Jackknife, the Bootstrap and other resampling plans. CBMS-NFS 38, Society for Industrial and Applied Mathematics, Philadelphia, Keen, C.S. Air pollution survey of greater Cape Town. Vol 4: Meteorological Aspects. Energy Research Institute, University of Cape Town, 1979.