The modeling period was selected as June 1–3, 2006 during a summer synoptic condition. A moderate easterly synoptic flow associated with a high pressure over the subtropical North Atlantic Ocean prevailed during this period. The coarse meteorological data is chosen from two sources i.e., the NCEP FNL analyses, and the EDAS Eta analyses. The FNL analyses is prepared by combining observations with short-range forecasts from a global model and the data are available on the surface and 26 vertical levels from 1,000 mb to 10 mb at a spatial resolution of roughly 100 km at 6 h intervals. EDAS is a regional analyses for North America based on the Eta regional model. These data are available at 3 h intervals on Eta 212 grid at a spatial resolution of 40 km on 26 vertical levels from 1,000 mb to 50 mb. To obtain high-resolution meteorological fields a simulation is conducted with the Advanced Research WRF (ARW) meso-scale model (V2.2) [18]. The model consists of fully compressible non-hydrostatic equations and the prognostic variables include the three-dimensional wind, perturbation quantities of potential temperature, geo-potential, surface pressure, turbulent kinetic energy and scalars (water vapour mixing ratio, cloud water). The model vertical coordinate is terrain-following hydrostatic pressure and the horizontal grid is Arakawa C-grid staggering. A 3^rd order Runge-Kutta time integration is used in the model.

The details of the meso-scale simulation are given in [17] and are briefly described here. The model is designed with three nested grids (36, 12 and 4 km) and with 34 vertical levels. The outer domain covered the South-central US and the surrounding Atlantic Ocean (Figure 1). The inner finer grid covered the Mississippi Gulf Coast off Louisiana above the Gulf of Mexico. The 36 km grid (54 by 40 points) is centered at 30.8 N, 85.3 E and the finest 4 km grid (187 by 118 points) is centered at 30.96 N, 87.5 E and covers the Mississippi Gulf Coast (Figure 1). The second and third nests are two way interactive. The model physics options used are Kain-Fritsch scheme [19] for convective parameterization, WRF Single Moment Class 3 (WSM3) simple ice scheme for explicit moisture [20], Yonsei University non-local scheme for boundary layer [21], standard five-layer soil model [22], Dudhia scheme for short wave radiation [23] and the Rapid Radiative Transfer Model [24] for long-wave radiation processes. The model is initialized at 00 UTC 1 June and integrated for 48 hours using FNL data for initial and boundary conditions. Four dimensional data assimilation (FDDA) [25,26,27] grid nudging is performed for the first 12 h period on all grids for temperature, mixing ratio and wind on the model grids using the NCEP ADP (Atmospheric Data Project) surface and upper air observations. The WRF model is run in the data assimilative mode for the first 12 h period (after enhancing the initial / boundary conditions with incorporation of surface / upper air observations), and then in pure forecast mode thereafter. The nudging coefficients used are 2.5×10⁻⁵ for temperature and wind and 1.0×10⁻⁵ for mixing ratio. The USGS topography and vegetation data (25 categories) and FAO Soils data (17 categories) with resolutions 5m, 2m and 30 sec (0.925 km) were used to define the lower boundary conditions. Various options used in WRF model are given in Table 1.

The Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model [13] developed by the Air Resources Laboratory is used to simulate the dispersion of airborne pollutant releases. It computes both simple trajectories and complex dispersion and deposition simulations using puff and or particle approaches. The dispersion computation consists of three components: particle transport by the mean wind, a turbulent transport component, and the computation of air concentration. Pollutant particles are released at the source location and passively follow the wind field. The mean particle trajectory is the integration of the particle position vector in space and time. In particle mode, the turbulent component of the motion defines the dispersion of the pollutant cloud and it is computed by adding a random component to the mean advection velocity in each of the three-dimensional wind component directions. The vertical turbulence is computed from the wind and temperature profiles and the horizontal turbulence is computed from the velocity deformation. The meteorological fields needed in the model are u,v,w (horizontal, vertical wind components), T(temperature), Z (height) or P (pressure), surface pressure (Po) and the optional fields moisture and vertical motion. These gridded three dimensional fields are linearly interpolated in space and time to the particle’s position.

The advection of a particle or puff is computed from the grid scale three-dimensional velocity vectors obtained from the meso-scale model. The horizontal turbulent velocity components at any given time are computed from the turbulent velocity components at the previous time, an autocorrelation coefficient that depends upon the time step, the Lagrangian time scale, and a computer generated random component. The lagrangian time scales T_Lw (vertical) = 100 s and T_Lu (horizontal) = 10,800 s are assumed to be constant. These values result in a random walk vertical dispersion for most of the longer time steps. Turbulent mixing is calculated using a diffusivity approach based upon vertical stability estimates and the horizontal wind field deformation. The stability estimates are based on surface fluxes when available or the wind and temperature profiles otherwise. Pollutant concentrations are estimated as the integrated mass of individual particles as they pass through the concentration grid which is a matrix of cells, each with a volume defined by its dimensions.

Three simulations were conducted with the HYSPLIT model using i) FNL analysis ii) Eta analysis and iii) WRF model outputs. Results are compared with observations to study the impact of the spatial and temporal resolution of the mean wind on the dispersion pattern. Dispersion simulation is done over a range of 100 km around the sources. A horizontal grid of 2° × 2° with resolution of 0.01° × 0.01° (approximately 1 km × 1 km) and with eight vertical levels 25, 50, 100, 200, 500, 1,000, 2,000 and 5,000 m above ground level (AGL) is considered in HYSPLIT. The emissions data is taken from the Mississippi Department of Environmental Quality (MDEQ) and comprises normal annual average for the respective sources. From a cluster of elevated point sources located along the MS coast, four major sources are considered in the present study. These are Mississippi Power Company-Plant Jack Watson (MPJ), Chevron Products Company- Pascagoula Oil Refinery (CR), Mississippi Power Company- Plant Victor Daniels (MPV) and Dupont Delisle Facility (DDF) (Table 2).

The dispersion calculations are made for SO₂ species and no seasonal or diurnal variations in the emissions are considered in the present study. Also the plume rise due to plume effluent velocity and plume temperature is not considered in the present study. The point sources considered have exit velocities since power plant plumes are certainly buoyant. The buoyant plumes rise to higher heights before being subjected to downwind transport and dispersion. The plume rise for these buoyant plumes is expected to impact the trajectory paths and concentration results since there is considerable vertical variation in winds and temperature with height. A detailed calculation of plume rise could be done in future work using the next version of HYSPLIT which incorporates the complete plume rise equations. The pollutant plume is treated as top-hat puffs in the horizontal and particle in the vertical. A total of 500 particles or puffs are released during one release cycle with a maximum of 10,000 particles permitted to be carried at any time during the simulation (Table 3). The vertical turbulence mixing is computed using a diffusivity approach based upon vertical stability estimates. Although there are several methods for the calculation of horizontal diffusivity such as isotropic similarity based on turbulent fluxes or temperature/ wind profiles (i.e., gradient Richardson number), in the present study the standard velocity deformation method is used as it is relatively simple and requires only the wind field which is commonly available in all the three data sets used in the study. Ground level concentrations are computed as averages for the lowest 50 m AGL within each horizontal grid cell.

Calculations from dispersion models are limited by the nature of the input meteorological data such as the resolution and representation of local terrain features in the low-level wind pattern [8]. The surface wind from the three input data sets are analysed first. Horizontal wind fields from the three meteorological data sets are shown in Figure 2 at 12 UTC (06 LT) and 02 UTC (20 LT) corresponding to the night and day time conditions. Surface wind in FNL data has coarse resolution. It is easterly over the sea region and northeasterly over the land region. Slight shift in wind flow to the land is seen in the daytime (Figures 2a,b). Wind field in the EDAS analysis seems to be represented better and more resolved indicating spatial variations than the FNL data. In a few grids near to the coast wind is northerly in the night and southerly in the day time indicating the diurnal variation in the flow pattern near the MS coast (Figures 2c,d). Simulated wind flow from WRF model (Figures 2e,f) is well resolved and more clearly represents the diurnal pattern of winds than the FNL and the EDAS data. It depicts the topographic variations in the flow expected from the presence of complex coast line of Louisiana.

Results of the meso-scale model simulation are given in more detail in [17]. Here a few results from that simulation relevant to the dispersion study are presented. During the morning time the surface wind over Mississippi and Louisiana is northerly. It gradually becomes north-westerly offshore and gains in strength especially in parts of Mississippi and Louisiana coastal plains indicating prevalence of land breeze circulation. The surface wind at the coast turns southerly around 11:00 LT indicating development of sea breeze. Simulated sea breeze penetrates more inland in the subsequent times (Figures 2e,f). The maximum wind speed increases to about 5 ms⁻¹ near the coast at 16:00 LT, the extent of the sea breeze was up to 50 km at this time. The flow is onshore and extended deep inland in the night. The simulated vertical cross-sections of wind and potential temperature at 15:00 LT at 89.75° E are shown in Figure 3. Circulation near the coast line at distance point (=100 km from south) shows development of a sea-breeze front with convergence, ascending winds at the leading edge, return flow aloft and subsidence behind the front. Horizontal circulation associated with sea breeze is seen to prevail up to a height of about 1,000 m. Temperature contours across the coast indicate stably stratified layers in the lower atmospheric region in the morning time on the sea and land regions, which gradually transforms to unstable stratification over the land in the afternoon time at 15:00 LT (Figure 3). A shallow boundary layer near the coast and deep boundary layer in the inland region are seen. The shallow mixing layer extends horizontally up to 30 to 40 km inland. Simulated PBL height across the coast also shows a shallow PBL near the coast and a deep boundary layer inland (Figure 4). Sea breeze extends up to 80 km inland and mixed layer depth across the coast varies from 200 m to 800 m. Diurnal variation in modeled and observed near surface (10 m) wind speed, wind direction and air temperature (at 2m) along with corresponding values from FNL and EDAS data for the period, June 01–03, 2006 are shown in Figure 5 at Pascagoula Mesonet tower located at 30.36 N and −88.52 E roughly at a distance of 5 km from the coast. Both observations and model values at Pascagoula indicate increase in wind speed, decrease in air temperature and shift in wind direction around the noon time indicating sea-breeze onset. A shift in wind direction from 275 (northwesterly) to 200 (southerly/ southwesterly) is found both in the model values and observations. The model could reproduce the observed trends of the surface variables (Figure 5). The data from EDAS and FNL follow the trend in observations to some extent. Although the EDAS and FNL data follow the order of the parameters, the diurnal cycle is better represented in the WRF simulated fields.

Simulated forward trajectories from the release locations are noticed to vary with each case of meteorological data (Figure 6). Trajectories with FNL data move in the northeast direction and no significant back turning is found. Trajectories in the cases of EDAS and WRF data sets also move in north east direction but show looping to the coast indicating effect of onshore and off-shore flows on the movement of air parcels. The looping of the trajectories is clearer from the trajectories drawn with WRF wind fields. Trajectories in these two cases also show vertical growth while moving inland due to variations in the vertical mixing across the coast.

Dispersion simulation in the present work is evaluated following the procedures outlined in [7,8]. A visual examination of the plume distribution pattern is made followed by statistical analysis. Figures 7, 8 and 9 show contours of model calculated 2 h average ground level concentrations in the case of simulations with FNL, EDAS and WRF meteorological fields respectively. Each of these figures shows the concentration distributions for the periods ending at 10, 18, 28 and 42 h respectively after the beginning of calculation. All the plots contain the same concentration contour intervals. Visual comparison shows the spatial pattern of concentration field is distinctly different in the three cases of meteorological data. Calculations with FNL data show relatively uniform distribution of the plume around the sources and the individual plumes from the sources are difficult to be identified. Also diurnal alteration in the plume direction was not significant in this case. Orientation of the plume was mostly to the west and followed the large scale wind flow pattern in the case of simulation with FNL data. Simulations with EDAS meteorological fields showed diurnal alteration in the plume direction to some extent but was largely influenced by the synoptic flow.

Plume pattern in the case of simulation with WRF meteorological fields clearly showed the diurnal alternation in plume direction according to diurnally varying wind flow pattern. Wide spatial and temporal variability of the plume is noticed in the calculations with WRF fields. Also the individual plumes from each of the sources could be clearly seen in the downwind direction in the cases of WRF and EDAS fields. The plume direction was to the east at early morning time (Figure 9A), to the southeast in the noon (Figure 9B), to the north east (Figure 9C) in the evening / night conditions and to the east-northeast in the afternoon next day (Figure 9D) in the case of simulation with WRF fields. In the simulation with WRF fields, the concentration pattern during offshore wind condition shows the ranges (2.15e-7 – 8.25e-8), (1.48e-6 – 5.62e-7) extended to larger downwind distances over the oceanic region. Similarly the concentration pattern during the night condition (Figure 9C) showed the range (2.15e-7 – 8.25e-8) extended to large distance over land.

The concentration distribution shows that the concentration maximum occurred near the source in the cases with FNL and EDAS data fields whereas it occurred away from the source in the case of simulation with WRF. The iso-concentration contours 3.83e-6, 1.47e-6, 5.62e-7, 2.15e-7, 8.25e-8 are seen to extend to larger downwind distances in the simulation with WRF fields as compared to the simulation with EDAS data. Statistics of the concentration distribution in dispersion simulations of different meteorological data are given in table 4. The maximum, mean and standard deviation of the concentration for the four sampling times (08–10 UTC June 01, 16–18 UTC June 01, 02–04 UTC June 02, 20–22 UTC June 02) are less in the cases of simulation with EDAS and WRF meteorological fields than with FNL fields. The mean and range of concentration values are high in the simulation case with WRF fields. The relatively uniform concentration distribution in FNL plume prediction is perhaps due to the reason that the horizontal dispersion rate in the method followed depends on the spatial resolution of the meteorological data. This is related to the under sampling of the flow field as seen from the wind vectors (Figures 2a,b) which change dramatically between the grid points along the coast line due to unresolved smaller scale features. The relatively coarse grid FNL data suggests that the horizontal deformation values will be much larger than in the WRF or EDAS data which are not under-sampled and have a more consistent pattern in the wind vector change between grid points. Thus the high resolution WRF data has improved the horizontal dispersion over the relatively coarse FNL and EDAS meteorological data fields.

Figure 10 shows observed and model concentrations sampled at Pascagoula monitoring station. It is to be noted that the SO₂ observations include both the background and longer range contributions from other nearby sources as well which are not included in the model calculations. The model values show contributions only above the background. Hence the observations are generally higher than the model values. However the observations are used here merely to compare the trends in the simulated values. Model calculations in all the three cases show diurnal variation in concentration. Calculations in the initial 6 h period are below observations in all the three cases due to grid instauration.

Model concentrations with WRF data are closer to observations up to 40 h and deteriorate thereafter. Statistical parameters of correlation and fractional bias are calculated between observed and calculated concentrations. Correlation (R) is used to represent the scatter among paired measured and predicted values and a fractional bias (normalized bias) (FB) is used to indicate over-prediction or under-prediction [7]. Calculated concentrations in all cases show an increase in bias towards the tail of simulation. Simulated concentrations with Eta analysis are an order less than those with FNL and WRF fields. Calculations with FNL fields are overestimated during day time and underestimated during night time. This is seen in the positive fractional bias during 8 h daytime period and negative fractional bias during 8 h night time period (Table 5). An overall (24 h) high correlation and positive bias is noticed with FNL data. The higher correlation in FNL plume prediction is probably because of the closeness of the wind speed, direction and other meteorological parameters in FNL data with the observations over 6 h intervals and because the FNL values remain constant in the dispersion simulation over every 6 h period unlike the EDAS and WRF data. However the WRF plume provides a more realistic representation of the plume diurnal cycle.

Higher concentrations observed in coastal regions are generally associated with sea-breeze circulation and shallow mixing during thermal internal boundary layer formation. Downwind concentrations in the three simulations for the afternoon time (18 LT) are shown in Figure 11 when the plume is spread completely over land. While the downwind concentration falls monotonously with FNL and Eta EDAS data some difference is noticed in the case with WRF fields. Maximum concentrations are noticed at distances of 10 – 40 km in the afternoon time coinciding with sea-breeze time and the consequent shallow mixing layer formation and during the night conditions. Figure 9C shows the concentration contours 5.6e-7, 2.15e-7, 8.25e-8 extend to large downwind distances in the north east direction during the night time in the dispersion simulation with WRF outputs unlike in the other two cases.