Model evaluation and intercomparison study in Po basin

In the frame of COST 728 Action "Enhancing mesoscale meteorological modeling capabilities for air pollution and dispersion applications", a mesoscale meteorological and chemical transport model evaluation and intercomparison study was carried out for two two-month periods: August and September 2003 and January and February 2004. The domain of interest was northern Italy, the focus being on the Po Basin. Aim of the study is to evaluate model performance both in terms of meteorology and chemistry and to explain the discrepancies between different models and between models and observations. The modeling systems used in the study are the PSU/NCAR meteorological model (MM5) offline coupled with the Comprehensive Air Quality Model with Extensions (CAMx), MM5 off-line coupled with the Community Multiscale Air Quality model and the online coupled meteorology-chemistry model WRF/Chem. The three modeling systems will be hereafter referred to as M1, M2 and M3 respectively. All models ran in 5km horizontal resolution (innermost domain). MM5 in M1 has 31 vertical layers with lowest layer depth of 9.5m whereas MM5 in M2 and WRF M3 have 23 vertical layers with lowest layer depth of 70m. The model output is evaluated against a network of 55 meteorological stations and 51 air quality stations providing quality-checked data. In this abstract, results concerning the meteorology model output and observations are shown.

Both simulation periods had high levels of air pollution. For example, ozone daily peak levels exceeded 200mgm-3 in certain days in many stations during August-September 2003. These levels of air pollution occurred largely because of unfavorable weather conditions: high-pressure systems over Europe resulting in low wind speed (1.7ms-1 on average), high solar irradiance and high temperatures. In January and February 2004 the mean observed wind speed was also 1.7ms-1. PM10 concentrations exceeded average values of 50mgm-3 in many stations.

A number of statistical measures were used to quantify model performance. Among them is the hit rate, which is defined as the percentage of the model values, which lie within a certain interval around the corresponding measured values. This interval is specified as ±1ms-1 for observed wind speed up to 10ms-1 and ±2.5ms-1 for observed wind speed above 10ms-1. For temperature and dewpoint temperature the interval is ±2¨¬C.

Regarding the results in August-September 2003 period models M1 and M3 overestimate wind speed. The model bias is 0.9ms-1 and 0.5ms-1 for M1 and M3 respectively. The bias of M2 on the other hand is not significantly different from the ideal value (zero) at the 95% level of confidence. The three models have similar values of product-moment correlation, around 30%. The hit rate of M1 is 41%, considerably lower than M2, which has 54% and M3, which has 49%. Temperature is underestimated by all models. The underestimation is larger for M1, which has a bias of -1.8¨¬C whereas the bias values of M2 and M3 are -1.1¨¬C and -1.2¨¬C respectively. The temperature hit rates were around 40% and the product-moment coefficients were around 90% for all models. Considering dewpoint, M1 had a bias of 0.2¨¬C whereas M2 bias was -1.4¨¬C and M3 bias was -1.6¨¬C. The dewpoint hit rates were slightly lower than the temperature hit rates.

In January and February 2004 more challenging weather conditions led to poorer model performance. Models M1 and M3 bias is 1.7ms-1 (predictions twice as large as the observed value) and model M2 bias is 1.0ms-1. The hit rate was reduced by about 10% for all models compared to the August-September values. Temperature predictions on the other hand improved in the winter period. A bias of -0.5¨¬C was found for M1, 0.0¨¬C for M2 and 0.9¨¬C for M3. Despite the ideal performance of M2 in terms of bias, the hit rate was 47% a value similar to the other models' (51% for M1 and 43% for M3). The temperature product moment correlation, which was particularly high for the August-September period is now somewhat lower, around 65% for all models, thus reflecting a more irregular daily temperature variation during winter. Dewpoint was underestimated by all models whose bias values were -1.4¨¬C, -3.1¨¬C and -3.2¨¬C for M1, M2 and M3 respectively. The hit rate values are 39% for M1, 25% for M2 and 24% for M3.

One important issue with the model performance is the overestimation in wind speed, especially during the winter period. This is expected to contribute towards underestimated concentrations of all pollutants in the chemistry runs. One should also note that although both models M1 and M2 derive the meteorology using MM5, they have considerable differences in their performance. M2 predicts better wind speed and temperature and M1 predicts better dewpoint temperature. The two models use different initial and boundary conditions, different nesting techniques and different vertical interpolation methods for wind speed. Considering dewpoint temperature, it is hypothesized that it is the selection of different physics schemes, which lead to quite different results. The effect of different model setups will be investigated in the future in order to explain the aforementioned differences in the model output. Further plans include the examination of the estimated and observed vertical structure of the atmosphere by using sounding data.

In addition to the meteorological aspect of this study, the chemistry model output will examined in order to evaluate the models and compare between them. Output from one more modeling system, the Local Area Model Italy offline coupled with CAMx will be added. This will provide the opportunity to examine the sensitivity of chemistry produced by M1 (which also uses CAMx) to two different meteorological drivers.