Methodology

1. Numerical model and downscaling method for extended simulation period

Mesoscale simulations for this project were carried out using the limited-area configuration of the Global Environmental Multiscale (GEM) atmospheric model (GEM-LAM hereafter). In general, the GEM model works by first solving a set of dynamical equations directly on the model grid. Physical processes including atmospheric radiation, fluxes from different land-surface components, boundary-layer turbulent mixing as well as clouds and precipitation are not directly resolved at the grid scale. These subgrid-scale processes are accounted for in the model by supplementing the solutions of the dynamical equations with parameterized tendencies associated with the pertinent physical processes.

Large-scale atmospheric features in the meteorological fields simulated with limited-area models are susceptible to deviations from the generally coarse-resolution driving fields over time, particularly for large continental-scale spatial domains. A major scientific challenge for this project was, therefore, to determine the appropriate strategy to address the issue of large-scale deviations for multi-year simulations.

In order to minimize the impact of large-scale deviations associated with a large spatial domain over extended-range simulations, the problem may be separated into multiple periods of sufficiently short time-frames. Construction of a final continuous time-series of any meteorological variable in this approach, however, suffers from abrupt changes due to temporal blending. Furthermore, individual shorter integration requires time for spin-up of clouds that are not analyzed in the regional analysis files from the Meteorological Service of Canada (MSC). The spin-up issue would therefore increase the computational cost of the project.

Dividing the problem into multiple mesoscale simulations over smaller domains each running for extended time-periods (from weeks to months) followed by spatial blending of the end results, on the other hand, results in spatial discontinuities in the meteorological fields, particularly along the lateral boundaries of the smaller domains. Furthermore, existing literature shows that the nested simulation domains cannot be arbitrarily small in order to permit proper development of small scales.

Based on the aforementioned adverse implications associated with temporal and spatial blending, a continuous temporal integration over the entire spatial domain appears to be the most suitable approach for this project, provided a mechanism is put in place to restrict large-scale deviations in the simulated fields. Large-scale atmospheric deviations are controlled by spectrally nudging the model outputs to the driving fields. Spectral nudging of the atmospheric large scales, as implemented in this project, is found to effectively control any undesirable deviation without considerable suppression of the small scales. The research conducted during this project have shown that simple temporal interpolation to derive the reference fields, in between two analysis hours, can lead to mesoscale variance deficiency in the spectrally nudged simulated fields. Two different strategies have been proposed and examined during this project to deal with such variance deficiencies. One option is to use a time-varying nudging coefficient that puts the maximum weight on the analyses fields only when the simulation time is very close to the time for which valid analyses fields are available. The other approach, which has been demonstrated to be more effective, is to produce hourly estimates of analyses by running 6-hour forecasts initialized with the analyses and assuming a linear growth of forecast error within the first 6 hours of model simulation. The latter approach is found to effectively eliminate any variance deficiency associated with the temporally interpolated analysis fields. Furthermore, different spectral nudging approaches, including the appropriate nudging length scales as well as the vertical profiles and temporal relaxations for nudging, have been investigated to determine the optimal nudging strategy. Analysis conducted during the course of this project has shown that specific humidity is well constrained during extended-range simulations when only temperature and wind speed are controlled by nudging. As a result, only the simulated temperature and horizontal wind speed fields were selected for nudging in this project. Further details regarding the spectral nudging approach developed for this project are provided in (Husain et al., 2014).

Although controlling the evolution of the atmospheric large scales generally improves the outputs of high-resolution mesoscale simulations within the surface layer, the prognostically evolving surface fields can nevertheless deviate from their expected values leading to significant inaccuracies in the predicted surface-layer meteorology. A forcing strategy based on grid nudging of the different surface fields, including surface temperature, soil-moisture, and snow conditions, towards their expected values obtained from a high-resolution offline surface scheme was therefore developed to limit any considerable deviation. The offline surface scheme used to downscale the surface fields (surface temperature, soil moisture, snow depth, and snow density) is known as the Surface Prediction System (SPS) which is based on the ISBA (Interactions between Soil, Biosphere, and Atmosphere) land-surface scheme. The standard implementation of the SPS scheme was considerably modified in the course of this project to include weighted blending of the driving forecasts fields to remove abrupt changes during switching between the driving forecast fields. Furthermore, a large-scale relaxation scheme for the evolving soil moisture field towards its regional analysis counterpart was implemented to restrict intermittent large-scale deviations. Additional details on the land-surface component of this project are provided in (Separovic et al., 2014).

2. Simulation Strategy

The basic simulation followed a two-stage strategy. First, the MSC's regional analysis fields, available every 6 hours (0000, 0600, 1200 and 1800 UTC, where UTC denotes the Coordinated Universal Time), were used to initialize and drive a GEM-LAM simulation involving 15 km horizontal grid spacing over the entire simulation domain for the entire time period (2008-2010). The 15-km GEM-LAM (LAM-15 hereafter) simulation included large-scale nudging of horizontal wind speed and temperature toward the operational regional analysis fields. The relevant surface fields within the LAM-15 simulation were also nudged towards to the SPS-generated reference fields. The purpose of the LAM-15- simulation was to produce three-dimensional meteorological fields with large-scale features closely resembling those embedded in the driving analysis fields, but available more frequently (every 20 min) to force the second-stage 2-km GEM-LAM (LAM-2 hereafter) simulation. The LAM-2 simulation also involved atmospheric and surface nudging towards the appropriate reference fields to produce the final desired outputs.

3. Verification of Strategy

Outputs of the LAM-15 and LAM-2 simulations were extensively analyzed to verify the validity of the strategy developed during the course of this project. The verifications that were conducted during this project are provided below.

  1. Similarity between the large-scales of the driving and the simulated field were compared to determine the impact of different atmospheric nudging configurations and to identify the most appropriate nudging strategy. Similarity of large scales is compared for both LAM-15 and LAM-2 simulations that were forced with the operational regional analysis and LAM-15 outputs, respectively.
  2. The ratios of spectral variance between the driving and analysis fields for the different length scales and at different vertical levels were compared to study the impact of the different nudging configurations. It helped to identify the appropriate nudging length scales. Spectral variance ratio was analyzed for both LAM-15 and LAM-2 simulation outputs. It was also useful in determining the impact of surface nudging.
  3. Simulated fields are compared at the screen level (2-m temperature and dew point, and 10-m wind speed) against those obtained from the ground-based stations spread all across Canada. In addition to the entire domain, screen-level statistical scores (bias, root-mean-square error, standard error) were compared for individual regions separately. The results have shown that both LAM-15 and LAM-2 simulations, coupled with atmospheric and surface nudging, resulted in improved screen-level temperature compared to the operational regional forecast while wind speed scores were also found to be equivalent. Over complex terrain, e.g., over British Columbia, LAM-2 simulations were found to result in improved scores for wind speed.
  4. Simulated fields were also compared against the limited wind turbine data that were available. For those limited number of stations, the simulated fields demonstrated improved statistical score compared to the operational forecasts. Moreover, the results for the optimal LAM-2 simulation were found to clearly outperform the optimally configured LAM-15 simulation.

4. Conclusion

A dynamical downscaling strategy based on high-resolution mesoscale simulations over a large continental-scale spatial domain and an extended time-period has been developed within the course of this project. Continuous temporal integration over the entire domain – as opposed to extended integrations over smaller spatial domains or multiple simulations with shorter time periods over larger domains followed by spatiotemporal blending –was found to be the most suitable approach for accomplishing the high-resolution downscaling objectives of the project. The developed scheme was employed during the project to generate the multi-year time series of meteorological variables for CanWEA as a contribution to the broader PCWIS.

In order to improve the impact of spectral-nudging, two novel methods have been developed that reduces or eliminates variance deficiency in the simulated fields. This includes the concept of time-varying nudging increment and the method of computing frequent analysis estimates. Extensive sensitivity studies have been carried out to identify the optimal nudging configuration in terms of the shape of nudging vertical profile, nudging length-scales and the type of temporal relaxation.

Large-scale spectral nudging of horizontal wind speed and temperature was found to adequately control large-scale deviations in specific humidity. In order to overcome potential intermittent deviations, the surface fields were nudged towards a reliable reference dataset obtained from the modified SPS external surface model. Results show that compared to the coarse-resolution regional analysis fields, the SPS fields when used as reference for surface nudging clearly led to improved screen-level scores for both air temperature and dew point temperature. Nudging of the surface fields was however found to be neutral for the screen-level wind speed. Increasing the strength of surface nudging was found to improve screen-level scores further.

Meteorological fields obtained through high resolution LAM-2 simulations following the nudging strategy adopted in this project is able to maintain large-scale similarity with the driving LAM-15 fields, while adding substantially increased spatial variance for the smaller scales (less than 200 km). In terms of screen-level scores, LAM-2 simulations were, in general, found to be equivalent compared to the LAM-15 simulations over the entire domain, although over British Columbia and the North – where orography-induced spatial variance is more influential – LAM-2 simulations were found to improve both screen-level temperature and wind speed. Performance of different atmospheric nudging configurations for both LAM-15 and LAM-2 simulations was also evaluated against 80-m wind and temperature data obtained from three wind farm locations. For all three stations, LAM-2 simulation with its optimal nudging configuration was found to deliver better statistical accuracy for both wind speed and temperature over its LAM-15 counterpart.

Due to a number of factors, including semi-Lagrangian advection, horizontal diffusion and orographic filter, the effective model resolution is around 7 – 10 times the actual model horizontal grid spacing. This implies an effective resolution of approximately 14 – 20 km for the time series variables. The orographic filter may affect scales as large as 40 km over complex terrain. Overall, the combined effect of these factors may lead to increased error in the wind time series in the vicinity of complex terrain. Furthermore, due to more limited possibilities of evaluations during spring and fall, it is difficult to ascertain the overall accuracy of the time series outputs for these periods. Users are recommended to be aware of these potential limitations while using the time series data.

5. References

Husain, S.Z., Separovic, L., Yu, W., and Fernig, D. (2014): Extended-range high-resolution dynamical downscaling over a continental-scale spatial domain with atmospheric and surface nudging. Journal of Geophysical Research – Atmospheres, 119(24): 13720-13750.

Separovic, L., Husain, S.Z., Yu, W., and Fernig, D. (2014): High-resolution surface analysis for extended-range downscaling with limited-area atmospheric models. Journal of Geophysical Research – Atmospheres, 119(24): 13651-13682.

Date modified: