The western coast of Norway is known for its picturesque mountain landscapes with sea inlets (fjords) penetrating deep into coastal valleys. Similar, if not as dramatic, settings with coastal valleys opening into sea inlets and bays are frequently accommodating harbour cities in other parts of the globe as well. Therefore, as we believe, the methodology and experience described in this study might be of interest to the research and urban management communities worldwide.

Bergen is the second largest city in Norway. This coastal city is located at 60.4∘ N and 5.3∘ E. The mountains around the city have peak elevations between 284 and 643 m a.s.l. They protect the valley from storms, significantly reducing the surface layer wind speed (Jonassen et al., 2013). The northern location with its weak solar irradiation during wintertime and cold air pooling in lower parts of the relief causes frequently observed but highly local temperature inversions during periods with persistently calm and clear weather. These inversions can last through several days. The resulting weak turbulent mixing of the lower valley atmosphere fosters the accumulation of locally emitted pollutants despite only moderate emissions. A climatology of temperature inversions and air pollution events in Bergen was previously reported in Wolf et al. (2014).

Bergen has more than 275 000 inhabitants. More than 75 000 of them reside in the central districts, located in the elongated central Bergen valley, which is the focus of this study. The valley opens toward a sea inlet (Byfjorden) in the northwest. It widens towards a large brackish water lake and more residential areas in the southwest. Figure 1 shows the relief of the studied area, the model simulation domain, as well as the location of the major air pollution sources in the city.

A reader may like to know that the air quality in Bergen has been monitored continuously since 2002 with an increasing number of measurement stations. In addition, a routine air quality forecast exists within the “Bedre Byluft” national project for monitoring and prediction of air pollution in Norwegian cities. The results from the measurements are available under NILU (2019). The predictions can be found under Miljødirektoratet (2019). Both measurements and predictions are regularly summarised (Bergen Kommune, 2018; Tarrasón et al., 2017). The forecast system has recently been changed. It was based on the AirQuis/EPISODE air quality model using meteorological input from a 1 km mesoscale numerical weather prediction (NWP) model specifically run for this application. The new forecast system is based on the “Nasjonalt Bergningsverktøy” (NBV) tool. The NBV uses uEMEP as a dispersion model and the control run of the standard METCoOp ensemble prediction system (MEPS) for Norway with a 2.5 km spatial resolution to calculate physics in the dynamical spectral model as meteorological input (Denby and Süld, 2015). With this, the previous AirQuis/EPISODE model at least to some degree included the valley topography into the dispersion calculations via a higher resolution of the meteorological input and a meteorological pre-processor that created a divergence-free wind field in the valley (Baklanov et al., 2007; The World Bank, 2009).

This study demonstrates the high-resolution modelling methodology using the comprehensive local data context. The model simulations require high-resolution topographic data, emission inventories for the different sources of pollution and vertical profiles of the large-scale (geostrophic) wind and temperature. All data sources are summarised in Table 1. Their detailed description is presented below.

The added value of the high-resolution model simulations is created by their ability to resolve the local relief features, which control the air flow and the turbulent dispersion. We run the PALM model with 10 m spatial resolution. To obtain the adequate relief, we used a topographic laser-based dataset, as described in Wolf-Grosse et al. (2017a). These topographic data were processed to ensure consistency and eliminate artefacts in the airborne laser data. The processed data constitute a digital elevation model (DEM) for the Bergen municipality. However, this DEM does not cover the entire simulation domain, as we run the simulations over a significantly larger area to resolve multiple local flows steered by the topographic relief. We extended the DEM, adopting digital surface model (DSM) data. The DSM does not include buildings (the DEM does), but outside the urbanised central districts, and sufficiently far away from the focus area, we assume that it is sufficient to resolve the major features of the relief and coastal line.

We selected the typical air pollution scenarios for this study based on joint air quality and meteorological data analysis. Routine air pollution measurements since 2003 have been available from two of the local air pollution stations, namely the Danmarksplass (DP) and Rådhuset (RH) stations in the central Bergen districts (see Fig. 1). The DP station represents an area affected by heavy road traffic, whereas the RH station serves as an urban background reference. We jointly analysed the air quality and meteorological records for several pollution episodes, when high concentrations of NO2 and PM2.5 were observed. As local meteorological data, we used measurements from the automatic weather stations located in the Florida neighbourhood (on top of the Geophysical Institute), the Ulriken summit and in the Jekteviken and Skolten harbour areas. For the large-scale weather conditions, we used data from the retrospective meteorological ERA-Interim analysis (Dee et al., 2011) and from the local boundary layer temperature profiles observed by a microwave radiometer MTP-5HE, which the Nansen Center has been operating on top of the Geophysical Institute since 2011.

The central Bergen districts are affected by emissions from three major polluting sources: vehicles on the road network; ships in the Bergen harbour; and wood-burning fireplaces in private residences. There are no major industrial sources. The road network is contributing to both emissions of NO2 and PM2.5. To specify these emissions, we used the gridded traffic density per road segment. It was available from registers at the national road authority for the main streets and from a traffic model run by the national road authority for the side streets.

Emissions from ships at berth in the harbour are contributing to both emissions of NO2 and PM2.5. The Bergen harbour is located in the historic city centre, and therefore ships docked at berth in the harbour can directly contribute to the local air pollution. Here, we considered only emissions from supply vessels for the offshore oil industry, during their periods at berth in Bergen harbour. Due to their often-large number and their large sizes, these ships are major contributors to the emissions from the harbour compared to other relevant ship types like the frequent, but small, short-range public transportation ferries or the less frequent, but larger, freight or passenger ships. In summertime, emissions from cruise ships can exceed the emissions from the offshore supply vessels in the harbour. In recent years, the cruise traffic seemed to expand also into other seasons. The impact of these supply vessels, that are docked right in the city centre of Bergen for extensive periods of time, on local air pollution is still highly disputed by the local population and therefore a matter of public interest and concern. The BOH provided us with data to specify the ships' location in the harbour between January 2015 and March 2016. For the emission rates per ship, we used the emission factors from the certification documentation of two representative ships.

Domestic heating with wood-burning fireplaces only contributes to emissions of PM2.5. We assessed these emissions through accounting for estates with registered active fireplaces. These data were provided by the Bergen fire department. For the emission rates per fireplace, we used typical emission factors for the existing mixture of new, clean burning wood ovens and older ones with higher emissions. All three emission sources are graphically summarised in Fig. 1.

The high-resolution air quality assessment requires a proper atmospheric model which can resolve the most energetic turbulence eddies in fully three-dimensional simulations. This ability to resolve the turbulent motions and interactions across a multitude of scales distinguishes the turbulence-resolving, or at least turbulence-permitting, LES models from more traditional meteorological and cloud-resolving models. The traditional meteorological models have the whole turbulence spectrum collapsed in their closure schemes, even if those models are running at 1 km or sometimes higher resolutions. The turbulence closure schemes are designed for weakly stratified, horizontally homogenous ABLs over a flat terrain. Their application to a strongly stable stratified ABL commonly results in excessive vertical mixing and hence in erroneous pollutant transport. The results of the GABLS intercomparison exercises are instructive to one working with this issue (Cuxart et al., 2006; Holtslag et al., 2013; Vignon et al., 2017).

We used the PALM (version 4.0, revision 1550) code in this study. The PALM code is developed by the PALM group at the Leibniz University of Hannover, Germany (Maronga et al., 2015). This model solves the primitive hydro- and thermodynamic equations for incompressible Boussinesq fluids. We ran the model for dry atmospheric conditions. This choice simplifies the model setup and is motivated by the fact that the worst air quality was always observed during fair-weather conditions, under prolonged occurrence of temperature inversions in the Bergen valley. Thus, the model was initiated only with temperature and wind profiles. The initial wind profiles also serve as geostrophic wind profiles or forcing, which drives the PALM simulations via the geostrophic wind term.

We ran PALM over a geographical domain centred on Bergen municipality. The domain spans 12.79 km in the zonal (east–west) and 17.27 km in the meridional (north–south) directions (Fig. 1a). The surface geometry was set by the DEM-DSM data. The lateral boundary conditions in the model runs were periodic. The domain includes a 1000 m wide buffer zone at the outer boundaries of the model domain. This buffer was necessary to linearly interpolate the surface geometry, making it periodic in both lateral directions. We set the model grid resolution to 10 m. The grid is vertically stretched by 1 % for each subsequent grid level above 750 m. In total, there were 128 vertical grid levels reaching 1450 m a.s.l., which is more than 2 times the height of the highest mountain peak within the model domain. The surface boundary conditions were different for the land and water surfaces. We used the Neumann (constant flux) condition for the land grid cells, as they have low heat capacity and quickly adjust the skin temperature. For the water grid cells, we used the Dirichlet (constant temperature) condition, as water has a large heat capacity and retains an almost constant skin temperature over the simulation time window.

Turbulent mixing is an irreversible process. It changes the temperature stratification and the wind profiles in the model also during the model spin-up. We counteract this difficulty by introducing a nudging routine. It relaxes the horizontally averaged temperature profile towards the initial temperature profile. This nudging is enabled starting from the second grid level above the local surface. The relaxation timescale is τ=43 200 s at elevations z<400 m a.s.l. This timescale linearly decreases to τ=1800 s at z>600 m. Nudging is therefore very weak in the lower parts of the atmosphere, allowing the temperature profile inside the Bergen valley to be determined dynamically through the boundary conditions. In the free atmosphere above the mountains, we applied much stronger nudging, retaining the initial temperature profile. A simpler setup of the PALM model has been tested successfully in its application to the stably stratified ABL inside the Bergen valley (Wolf-Grosse et al., 2017a).

In order to simulate the dispersion of pollutants from the relevant sources, we used constant emission rates per grid cell in the model. The emission rates were set separately for each of the three emission sources. Hence, NO2 and PM2.5 tracers from the harbour, chimneys and roads are simulated independently. In the applied model version of PALM, all pollutants are treated as passive tracers. We argue that this is a reasonable choice for the wintertime air quality assessment in a local high-latitude domain. The background near-surface ozone (O3) that could serve as a source of additional NO2 from existing NO is greatly depleted during the winter months. Sunlight that could lead to a photolytic conversion of NO2 to NO and O3 is largely absent. The nucleation of NO2 into nitrate particles is slow compared to the transport and mixing processes. The particle growth beyond the PM2.5 limits is also a slow process, whereas the particle wet scavenging is minimal under the clear-sky conditions considered in this study. Gravitational settling of particles for the size range below 2.5 µm is slow.

The high-resolution environmental assessment and modelling is still a costly and computationally demanding exercise. Therefore, we selected the most relevant and impactful scenarios of high air pollution using joint statistical analysis of the long-term air quality and meteorological observations (Wolf-Grosse et al., 2017b; Wolf et al., 2014; Wolf and Esau, 2014). Here, we will not repeat the details of that analysis. Nevertheless, a brief summary might be useful for new readers. Studies of the extreme deterioration of the air quality may frequently benefit from the fact that the high concentrations of pollutants are reached after several hours (or even days) of persistently calm clear-sky weather. Such weather conditions are limited to a few specific sets of local values of the meteorological parameters. The high concentrations are mostly observed under southeasterly winds over the Bergen valley. Due to a complex interaction between the locally forced circulation and the large-scale winds, the wind direction in the city is mostly southeasterly, too. That means the air moves from the city towards the Bergen fjord and harbour. The local wind therefore transports air pollution from the locations with the most intensive emission and towards the fjord. The efficiency of this transport is dependent on a convergence zone that can, dependent on the interplay between the local and larger-scale drivers, be located either over the fjord or over the city areas in proximity to the harbour. Different inhabited areas might therefore be affected by emissions from some or all local emission sources.

The scenario ws01_wd01_ft01 is the baseline scenario for the air quality simulations (see Table 2). This scenario describes the most typical meteorological conditions during episodes with high air pollution in Bergen. The baseline scenario maintains a low-level (surface) temperature inversion in the lowermost 100 m a.s.l. almost everywhere over land in the Bergen simulation domain. The weak winds and wind channelling by the valley relief could be noted over the lowermost 400 m. Figure 4 shows the relief and vertical temperature and wind profiles in two selected areas of interest and five selected sub-areas for more detailed analysis. Area (a) represents the harbour area in the city centre, while area (b) represents an area with dense population and heavy traffic in the southwest of the city centre. Area (a) is strongly affected by the sea–land temperature contrast. The enhanced mixing over the water surface dilutes the temperature inversion there. Area (b) includes the heavily trafficked road junctions at DP and its direct surroundings (sub-area 4). It also includes the Bergen meteorological observatory with a large amount of meteorological instrumentation collocated in sub-area 3.

The wind speed and directions in the baseline scenario clearly indicate a low-level air transport in the Bergen valley towards its opening into Byfjorden – the sea inlet where the harbour is located. This channelled local circulation is likely enhanced by the land–sea temperature contrast as it has been described already through statistical analysis in Wolf et al. (2014) and dynamical analysis in Wolf-Grosse et al. (2017a). The latter study also addressed the distinct rotation in the wind direction from down-valley near the ground to up-valley at around z=300 m and down-valley again above that altitude. This rotation with altitude is believed to be caused by an interaction between the large-scale meteorological circulation, the local topographic steering and the local forcing through the land–sea temperature contrast.

The temperature inversion in the simulated baseline scenario was most pronounced in sub-area 4. We assume this sub-area to be the most representative of the interior valley. The other sub-areas show more or less pronounced effects of the cold air passing over warm water bodies. The simulated inversion profiles show multiple inversion layers that are not resolved in the measured vertical temperature observations (resolution at 50 m). The maximum heights of the inversion profiles in the simulations are somewhat shallower than what is typically observed. A more thorough discussion of the simulated and observed inversion profiles in Bergen can be found in Wolf-Grosse et al. (2017).

Due to its high spatial resolution, the PALM simulations created a detailed geographical dispersion pattern of the NO2 and PM2.5 concentrations in and around the city. Figure 5 shows the NO2 concentration pattern created by the road traffic emission in the baseline scenario. The underlying geographical map (land and water surfaces) is given in gray shading as a Google Maps^® picture. The use of geoinformation tools to visualise the model simulations may facilitate the use of the simulated quantitative information in decision-making processes. The familiar map design simplifies orientation of and identification with the complicated pollution pattern.

The simulations revealed that emissions from road traffic are the dominant emission source for NO2 over the populated parts of the Bergen valley during high-air-pollution episodes. This was also confirmed with a detailed analysis of the timing of peak pollution levels from the local measurements at the two reference stations together with considerations of the local mean measured wind directions during high-air-pollution conditions (not shown). The baseline scenario indicates that the NO2 concentrations are rather high (>100 µg m-3) not only in direct proximity to the major roads but also in many adjacent urban areas. The exact shape and structure of the valley topography imprint strongly on the local dispersion conditions. Downwind of the major road transecting through the valley, areas with a channelled flow appear. These are visible as streaks of elevated pollution concentrations sometimes as high as 150 µg m-3. These streaks are separated by areas with pollutant concentrations below 50 µg m-3. The atmospheric channelling follows the areas with the lowest topographic height. In addition, the small water bodies (lakes and fjords) appear as areas with relatively lower air pollutant concentrations due to their comparatively higher surface temperature and the subsequent enhanced ventilation of the lowest air layers. This fine-scale structure of the dispersion pattern is especially relevant, since the simulations indicated elevated concentrations also in areas without continuous measurements with a sufficient temporal resolution. Some urban areas might be affected by pollution transport over several kilometres and accumulation of emitted substances strongly different from the emission pattern.

The dominant emission sources for PM2.5 during high-air-pollution episodes in the central Bergen valley turned out to be the wood-burning fireplaces. This was visible in the PALM simulations as well as in the detailed analysis of the available air pollution observations at the two reference stations (not shown). Therefore, one could expect the highest concentrations of the PM2.5 in the densely populated areas. Figure 6 shows the pattern of the PM2.5 concentrations in the baseline scenario. Distinct to the NO2 pattern, the PM2.5 pattern is more evenly distributed over the entire city. The concentrations peak in the areas of the near-surface flow convergence in the lower parts of the relief. Since the wind is weak, areas with higher density of fireplaces are clearly identifiable (Fig. 1).

A major advantage of the high-resolution modelling is related to the model's ability to simulate separately the impact and patterns of the different emission sources. As for NO2, the most significant emission sources in Bergen are related to the ships in the harbour and the road traffic. This is likely a typical situation for many coastal cities around the globe. We have a reasonably good estimation of the absolute ship and traffic (road vehicles) emission rates and their spatial distribution to run independent simulations of their pathways and the NO2 concentration patterns. The combined assessment of the absolute and relative contributions from these two local air pollution sources is provided in Fig. 7. The road traffic emission dominates the urban air pollution almost everywhere in the city. The absolute emission rate per ship in port is, however, quite high. Assuming that within the assessed time, each vehicle is moved by 10 km, the emission from each ship would correspond to the emission from 1377 typical vehicles in Bergen. Using the concrete counts of the cars passing the central area, it gives us that 16 ships at berth in the harbour emit about 127 % of the total car emission in the considered area. Such a significantly higher emission has, however, smaller contribution to the street-level concentrations of air pollutants. The ships emit at higher elevations. As the vertical mixing is strongly reduced, their emission does not reach the ground before they are either transported offshore over the Byfjorden or diluted by the horizontal air movements.

There are three major sources of the PM2.5 emission in Bergen, namely ships, road traffic and wood-burning fireplaces. Figure 8 shows that fireplaces dominate in their contribution to the local pollutant loadings, even after the rescaling to provide more reasonable concentrations. The emissions per ship are approximately 34 times that from a single wood-burning fireplace after applying the scaling factor of 0.1. However, the fireplaces are emitting at lower heights above the ground and clustered in the most populated area. The ships emit the PM2.5 in the harbour area where the emitted pollution is, in many weather conditions, effectively transported offshore and diluted over the unpopulated fjord area. The PM2.5 concentrations from the road traffic are overall low.

At this point, however, the high uncertainty of PM2.5 emissions should be emphasised, as emissions by road, tire and break abrasion have been neglected in this study, in addition to the necessary correction of the emission strength. Deposition of small- and larger-sized particulate matter (PM) may in addition play a crucial role and correct the simulated pattern to some degree. PM deposited on the ground can lead to high peak PM concentrations, when moist urban surfaces are drying off after several days with fair weather. This is especially relevant for major roads, where car-induced turbulence can lead to a resuspension of dust greater than what the local wind conditions would generate.

The baseline weather scenario represents the most typical meteorological conditions leading to build-up of the air pollution in the city. The concrete observed weather conditions, however, vary and may differ from the considered scenario. Therefore, the air quality assessment needs to characterise sensitivity of the pollution distribution patterns to imposed perturbations of the meteorological parameters. The strong topographic steering in the valley restricts deviations in the simulated patterns produced by the sensitivity runs. Overall, the precursor sensitivity simulations showed very similar geographical concentration patterns but with varying accumulation strengths. The baseline and sensitivity scenarios are listed in Table 2. Only the weather scenarios with the most notable differences in the local dispersion conditions will be discussed below for shortness of presentation.

The stronger off-shore (easterly) wind (scenarios ws03_wd02_ ft01 and ws03_wd02_ft02) deflect the pollutant plumes over the fjord water (Fig. 9 for NO2 and Fig. 10 for PM2.5), reducing the pollutant transport out of the inhabited city centre and the harbour area. The waterfront takes most of the impact. The concentration patterns in the valley interior remain largely unchanged. We note that the pollution patterns became more fragmented in those scenarios, indicating somewhat reduced stagnation and accumulation of pollutants in the valley.

The most influential weather scenario with respect to the local dispersion conditions is ws03_wd02_ft03 (see Figs. 11 and 12). A reversal of the flow at some elevation can be recognised from the up-valley transport of the emissions from ships in the harbour in Fig. 11, while the wind vectors at the lower (z=55 m) level still suggests a down-valley flow closer to the ground. This reversal was already visible in the vertical profiles in Fig. 4 but at the higher levels, so that it was invisible in the plume dispersion at the surface (Fig. 7). The flow reversal is also vertically more extended in this simulation (not shown). The reason for this is a change of the interplay between the locally forced circulation due to the land–sea temperature contrast, the topographic steering and the large-scale meteorological circulation. The weaker convergence over the fjord due to the lower water surface temperatures and the stronger large-scale winds weakens the down-valley circulation, reducing venting of the pollutants from inhabited areas to the water surface in Byfjorden. Hence, the populated areas close to the waterfront experience increased accumulation of pollutants, especially visible for the PM2.5 concentrations in Fig. 12. This result is counterintuitive, as a stronger large-scale wind allows for the increased accumulation. At the same time, the highest concentrations of NO2 are reduced, as the large road junctions in the interior of the valley are now affected by stronger winds and hence dispersion of the pollutants.