High Resolution Simulation of the Variability of Surface Energy Balance Fluxes Across Central London with Urban Zones for Energy Partitioning

Abstract

The parameterization of surface heat-flux variability in urban areas relies on adequate representation of surface characteristics. Given the horizontal resolutions (e.g. $\approx$0.1–1 km) currently used in numerical weather prediction (NWP) models, properties of the urban surface (e.g. vegetated/built surfaces, street-canyon geometries) often have large spatial variability. Here, a new approach based on Urban Zones to characterize Energy partitioning (UZE) is tested within a NWP model (Weather Research and Forecasting model; WRF v3.2.1) for Greater London. The urban land-surface scheme is the Noah/Single-Layer Urban Canopy Model (SLUCM). Detailed surface information (horizontal resolution 1 km) in central London shows that the UZE offers better characterization of surface properties and their variability compared to default WRF-SLUCM input parameters. In situ observations of the surface energy fluxes and near-surface meteorological variables are used to select the radiation and turbulence parameterization schemes and to evaluate the land-surface scheme and choice of surface parameters. For radiative fluxes, improved performance (e.g. $>$25 W m$^{-2}$ root-mean-square error reduction for the net radiation) is attained with UZE parameters compared to the WRF v3.2.1 default for all three methods from the simplest to the most detailed. The UZE-based spatial fluxes reproduce a priori expectations of greater energy storage and less evaporation in the dense city centre compared to the residential surroundings. Problems in Noah/SLUCM partitioning of energy between the daytime turbulent fluxes are identified with the overestimation of the turbulent sensible heat and underestimation of the turbulent latent heat fluxes.

Appendix 1: Sensitivity to Physical Parameterizations

To determine appropriate online model set-up the first vertical level SEB fluxes, and meteorological variables are compared to observations (Sect. 3.1) for three sensitivity runs (Goddard–YSU/Dudhia–MYJ/Dudhia–YSU; see Table 1 and Sect. 3.3.2); all are performed with the UZE classification.

The analysis of the radiative components (Fig. 11a) clearly shows a positive bias in the K ${\downarrow }$ estimation from the Goddard parameterization ($MBE \approx 26$ W m$^{-2},$ $RMSE \approx 41$ W m$^{-2}$; Supplementary Table S.1a) compared to Dudhia ($MBE = -1.4$ W m$^{-2}$, $RMSE \le 24$ W m$^{-2}$). With a similar treatment of the urban surface in all three runs, differences in the outgoing components can be directly linked to biases in radiative forcing. The K ${\downarrow }$ overestimation from Goddard leads to an overestimation of K ${\uparrow }$ ($MBE = 1.6$ W m$^{-2}$), while the Dudhia runs underestimate K ${\uparrow }$ ($MBE = -1.8$ W m$^{-2}$).

Fig. 11

Observed (symbols) and modelled sensitivity (lines) of: a incoming and outgoing shortwave $(K{\downarrow }, K{\uparrow })$ and longwave $(L{\downarrow }, L{\uparrow })$ fluxes, b net all-wave radiation $(Q^*)$, c turbulent sensible heat flux $(Q_\mathrm{H})$ and d turbulent latent heat flux $(Q_\mathrm{E})$

Since all three runs are performed with the same longwave parameterization (RRTM; Table 1) difference in L ${\downarrow }$ are smaller and likely attributable to the change of PBL schemes. The YSU option provides the optimum $RMSE$ and $MBE$ for L ${\downarrow }.$ All runs show a daytime underestimation of L ${\downarrow }.$ Analysis of the outgoing fluxes suggests that for all runs the surface emits long wave too much during the day (L ${\uparrow }$ overestimation; $\approx$0600–1600) and not enough at night (L ${\uparrow }$ underestimation; 1900–0500).

The net all-wave radiation flux ($Q*,$ Fig. 11b) has a lower $RMSE$ (25 W m$^{-2})$ when the Goddard scheme is used over the diurnal period and better daytime performance than Dudhia ($RMSE \approx 29$ W m$^{-2}$). However, this is a direct consequence of biases in the surface parameterization (Sect. 4.2.2). The Goddard scheme overestimation of K ${\downarrow }$ compensates with an overestimation of L ${\uparrow }$ and underestimation of L ${\downarrow }$ (Fig. 11a). This demonstrates a problem of online evaluation exercises where errors arising from one parameterization scheme might compensate for another. However, when one is fixed/improved the other error will remain. In all three runs there is an over estimation of $Q*$ at night (under estimation of longwave emissions): there is a negative bias from Dudhia ($MBE \approx -9$ W m$^{-2}$) and a positive one for Goddard ($MBE \approx 10$ W m$^{-2}$).

As for the offline runs (Fig. 4), there is an overestimation of daytime $Q_\mathrm{H}$ for all three settings and the run performing best for $Q*$ (Goddard) has the largest bias ($MBE$ $\approx$10 W m$^{-2}).$ This behaviour of the scheme is linked to the trade-off between modelled $Q*$ and $Q_\mathrm{H}$ reported in both Loridan et al. (2010) and Loridan and Grimmond (2012b). Of the three the Dudhia–YSU has the lowest $RMSE$ (83 W m$^{-2}$) and $MBE$ (4.8 W m$^{-2}$). For $Q_\mathrm{E}$ (Fig. 11d) the three are almost equal with the Goddard run providing the largest daytime maximum (greater $Q*$ available).

The impact on modelled meteorological variables at the first vertical level reveals a distinct cold bias in temperature from the Dudhia–MYJ run ($MBE = -1.9^{\circ }$C; Supplementary Table S.1c). This is reduced when using the YSU ($MBE= -0.9$ $^{\circ }$C; Fig. 12a). The best performance is obtained with Goddard–YSU ($MBE= -0.2$ $^{\circ }$C) but this is associated with the K ${\downarrow }$ overestimation (Fig. 11a); it is not for the correct reasons. All three runs underestimate early morning temperatures (before 0600), come very close to the observed temperatures after sunrise (0700–1100) prior to overpredicting the peak midday values, followed by a too rapid cool down in the afternoon/evening (after 1600). This tendency can easily be related to the misrepresentation of the longwave radiation budget (Fig. 11a) with an overestimation of peak daytime/lack of nighttime L ${\uparrow }$ values and a transition from overprediction to underprediction occurring around 0600/0700 and 1600.

Fig. 12

Observed (symbols) and model sensitivity (lines) of: a air temperature, b water vapour mixing ratio, c wind direction and d wind speed at first vertical level. Measurement at KSS and KSK sites from Vaisala WXT510 weather transmitters

Although all three runs simulate the general trend in humidity, they all predict conditions that are too dry around mid-day and too humid in the afternoon/evening (Fig. 12b). Switches in biases again coincide with those in L ${\uparrow }$ and air temperature (i.e. 0600/0700 and 1600). All runs fail to reproduce the change in wind direction before sunrise ($\approx$0400, Fig. 12c) but follow the general trend well afterwards (i.e. from 0700 onwards). The best statistics are obtained from the Goddard run (Supplementary Table S.1c), which is closer to observations before 1400. All three runs also tend to overestimate wind speed (Fig. 11d) and the smallest bias is obtained with the MYJ run ($MBE = 0.4$ m s$^{-1}$; Supplementary Table S.1c). This also agrees with the wind-speed overestimation reported by Salamanca et al. (2011). Note that the variability between the two sites’ observations can be seen from the difference in measured wind speed (and direction) between 0200 and 0500.