WRF evaluation during Storm Ophelia using remote sensing and in-situ measurements at Mace Head, Ireland.

: Storm Ophelia made landfall over Ireland as an extra-tropical storm on the morning of the 16 th October 2017. The storm caused major power outages, lifted roofs, caused coastal flooding in Ireland, and resulted in the loss of three lives. A model’s capability to forecast extreme weather events such as Storm Ophelia is of upmost importance and now with a changing climate, it becomes more important to improve and enhance model forecasting capability. The Weather Research and Forecasting (WRF) model V3.9 has been configured for the Irish domain and this study presents a preliminary evaluation of the Model during Storm Ophelia. Simulated wind speed and direction were compared with hourly remote sensing (lidar) and in-situ (wind speed and wind direction at 10m) observations at the coastal site of Mace Head Atmospheric Research Station on the West coast of Ireland (53.33◦ N, 9.90 49 ◦ W). The model simulation has generally small biases in the simulated wind speed and wind direction during this case study. The model also realistically simulated the magnitude and geographical distribution of the wind speed and wind direction observed during Ophelia.


Introduction
Earth's climate is rapidly changing, with widespread changes in the atmosphere, ocean, cryosphere, and biosphere [1]. Human-induced climate change is already affecting many weather and climate extremes in every region across the globe and there is more evidence of observed changes in extremes such as heatwaves, heavy precipitation, droughts, and tropical cyclones. Increasing likelihoods of extreme weather events is the most noticeable and damaging manifestation of human induced climate change [2]. Hurricanes are among the most devasting natural disasters, with approximately 80-90 tropical cyclones reaching tropical storm intensity around the globe [6]. The number of categories 3, 4, and 5 North Atlantic hurricanes during the first decade of the 21st century was the highest since 1951. The recent increases in activity are linked, in part, to higher sea surface temperatures in the region where Atlantic hurricanes form and move through. An increase in strength and number of intense hurricanes is projected for the end of this century. These projected changes are based on an average of projections from a number of individual high-resolution models and mechanistic considerations [3] [4] [5].
Storm Ophelia was the easternmost major hurricane on record. It caused almost 70 million worth of damage and is considered to be the worst storm to affect Ireland in 50 years. The tropical storm developed from the strongest eastern Atlantic storm in a century and a half and brought wind speeds of 191 km/h to parts of Ireland. It left more than 385,000 homes without power, and effectively shut down the country for 2 days. Three deaths can be directly attributed to Ophelia, all of which occurred in Ireland. Figures 1(a) and 1(b) give an overview of the storm from satellite images. In Figure 1(b), the eye of the hurricane is clearly visible. Current mesoscale numerical models, such as the Weather Research and Forecasting (WRF) model, are capable of simulating intense tropical cyclones with realistic structures [7]. The WRF Model is designed to serve both operational forecasting and atmospheric research needs [8][9] [10]. Evaluation of the Hurricane Weather Research and Forecasting (HWRF) model, designed specifically for hurricane studies [11] against WRF Models showed that the intensity forecasts displayed only marginal improvement of 5-8 % over the operational forecasts [15]. Domain size, location, and resolution are critical [12] [13] and the importance of selecting representative grid points for evaluating the model performance has been shown to be of relevance [14].

WRF model configuration
One of the essential steps in numerical weather forecasting simulations is the initial setup of physics options for the region and time period being considered. The mesoscale model used in this study is the Weather Research and Forecasting Model (WRF). WRF is a multi-scale model which scales from hemispheric scale down to km scale and can accommodate multiple levels of nesting. The WRF model is used extensively within academia, governments, and industry, and has an ability to simulate a wide range of synoptic scale, mesoscale, and microscale atmospheric phenomena. The model is equipped with multiple parameterization schemes (e.g., microphysics, land-surface interaction, radiation, planetary boundary layer, etc.) [7]. WRF can produce simulations based on actual atmospheric conditions (i.e., from observations and analyses) or idealized conditions. In this study simulations were carried out using 3 nested resolutions, 25km, 5km, and 1km. WRF Version 3.9 was used for model simulations. The centre of domain is located at 52.32N, 1.51W with a resolution of 25 km, which covers Ireland, UK, and most European 3 of 14 regions. The vertical structure is divided into 30 levels, of which 8 levels are below 1 km altitude [10]. The terrain, land use and soil data are interpolated into model grids from the USGS global elevation with resolutions of 10 m, 2 m, and 30 m respectively. The European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis (ERA) Interim data files are used for the meteorological initialization and boundary conditions [21]. The ERA-Interim ends on 31st August 2019. The ERA-Interim file needed to be interpolated to the limited area of WRF domain. This process is realised by the WRF Pre-Processing system (WPS). In addition, we use the RRTM shortwave radiation schemes [22], and the Unified Noah land surface model. The sub grid-scale turbulence mixing is turned on without explicit computational mixing. The microphysics and PBL schemes are Lin et al & YSU scheme [23]. The full model configuration used in this study is presented in Table 1. Table 1: WRF 3.9 model configuration used in this study.

Doppler Lidar Measurements
We validated WRF wind profiles using a scanning Doppler wind lidar of type WindCube 200S from manufacturer Leosphere (France). The lidar is located at Mace Head, about 300 m from the shoreline at 21 m above ground level. The scanning lidar delivers radial wind speed along any line of sight, from which horizontal wind speed and direction can be calculated using a specific scan pattern. At intervals of 15 minutes, conical scans at fixed elevation angles, also referred to as velocity-azimuth-display (VAD), were performed consisting of 12 beams at discreet azimuth angles 30° apart. The elevation angle of the conical scans used for this study was 15°. Temporal and vertical resolution of the resulting wind field was 15 minutes and 20 m.
To obtain horizontal wind speed and direction, radial wind values were fitted to a sine-curve as suggested by Browning and Wexler [33]. The least-square fit was applied to each range bin (horizontal cross section of the cone forming a circle). Availability of lidar signal depends on the presence of scatters, which are mostly aerosol particles. Therefore, low signal-to-noise ratios are common at altitudes above the boundary layer and lead to poor fits, which we screened out by applying a threshold of the correlation coefficient R of the fit of 0.98. Besides, in this part of the study we focus on eight WRF height levels below 1300 m above sea level, to avoid outliers affecting the validation. For comparison of WRF and lidar profiles of wind speed and direction, the closest lidar altitude to the WRF levels, as well as the closest lidar time stamp to the full hour were selected.

Evaluation parameters
WRF model simulations from the three domains (D1), (D2) and (D3) were evaluated statistically against the near surface in situ observations measured by the automatic weather station. The investigated parameters were wind speed and wind direction. The capability of the model was evaluated for a 24-h forecast period starting from 0000 UTC to 0000 UTC the next day (the initial 12 h being discarded for model spin-up). Due to the complexity and uncertainty associated with the forecast verification process, it is well known that a single verification statistic cannot depict the quality of a forecast [34]. A number of standard qualitative and quantitative accuracy measures were used to assess the forecast, namely mean bias error (MBE), root mean square error (RMSE), standard deviation of the mean value (σ) and correlation coefficient (r).

WRF Model comparison to in-situ measurements
Our simulation period covers the time when Storm Ophelia reached Ireland, from 16 to 18 October 2017. In this section, model simulations for nested resolutions 25km, 5km, and 1km are compared against hourly in-situ Met Eireann measurements of wind speed and direction at Mace Head. Results using the metrics introduced in the previous section, for the 16 and 17 October are shown in Table 2.  Table 3 shows maximum and medium wind speed, and the time the maximum wind speed was recorded, or forecasted by WRF. The nested WRF 5km resolution captures both, timing, and wind speed very well, forecasting the peak within 1 hour of the maximum hourly average wind speeds measured on 16 October. WRF 25km appears to have a time lag of several hours. While the WRF 1km resolution misses both the intensity and timing of the peak wind speeds.

WRF & Lidar Wind speed (30 vertical levels)
For comparison of WRF and lidar profiles of wind speed, the closest lidar altitude to the WRF levels, as well as the closest lidar time stamp to the full hour were selected.   Comparing lidar wind speed against WRF simulations on the 16 October, we can see that the agreement with wind speed VAD_15 is good (R 2 >0.59 for all WRF resolutions). The agreement with wind speed VAD_75 is poorer, (R 2 >0.41 for all WRF resolutions). The WRF 5km resolution perform' s marginally better than the 25km domain for both VAD_15 and VAD_75. This agrees with the previous in-situ 10m wind speed analysis. Comparing lidar wind speed against WRF simulations on the 17 October, we can see that the agreement with wind speed for Lidar VAD_15 & VAD_75 is excellent (R 2 > 0.93) for all WRF resolutions. Again, this agrees with the in-situ 10m wind speed analysis on the 17 th of October.

WRF & Lidar Wind direction vectors (30 vertical levels)
For comparison of WRF and lidar profiles of wind direction, the closest lidar altitude to the WRF levels, as well as the closest lidar time stamp to the full hour were selected.   We now look at the time-height plots of the relative difference in percent between WRF and lidar calculated as: diff = 100*(X_lidar -X_wrf)/X_lidar. We look at plots for u and v wind direction components as well as horizontal wind speed for days 16 & 17 October for Lidar angle, VAD 15.
From the Met Eireann in-situ measurements, we know that max wind speeds were detected on the 16 th of October at 7:00am. On the 16 th October, we can see from Figure 3, that the 5km WRF domains capture the max wind speeds seen at Mace Head at around 7:00am.       From Figure 6, we can see that on the 17 th October, the 1km WRF domain overestimates wind speed at the lower altitudes of less than 200m, while the 5km domain slightly underestimates the wind speed at less than 200m and indeed underestimates the wind speed at all altitudes in the early morning.
Here we can see that clearly the 25km domain overestimates the wind speed at all altitude, suggesting that the model performs better with a smaller domain over land, rather than a large domain which covers land and sea. The v-vector wind direction component in better agreement than the u-vector component. Results show that as seen with the in-situ measurements, the nested 5km WRF resolution performs best for both Wind Speed and Wind Direction on the 16 h & 17 th October.

WRF Model Grid Point Analysis
The importance of selecting appropriate grid points to compare with observations is now examined in detail. The wind speed from the nearest grid point is not always the most appropriate one for this comparison, nearby ones may be more representative. To see the sensitivity of model simulation to selected grid points, we extracted the data at three grid points. One is centred at Mace Head (-9.9, 53.3), and the other two grid points are near Mace Head. One is moving towards ocean (-10.2,53.3), and the other is moving inland (-9.7,53.3).
 Mace Head centred (original)  Mace Head, subg (inland)  Mace Head, subo(ocean) As Mace Head is located in the boundary of ocean and land there is a transition between two distinct surfaces which need more sub-grid points to resolve the sub-grid scale problem. 16 From table's 7(A-C) we can see that the wind speed and direction of Mace Head Sudg performs best on the16 th October with good correlations, significantly improving (R=0.78 for WRF 1km resolution) as the grid spacing approaches 1 km. There is poor agreement for wind speed, with R 2 smaller than 0.23 for all WRF resolutions at grid point Mace Head centre and Mace Head subgo. MBE, however, was small. Wind direction compared better than wind speed, with R 2 larger than 0.5 for all WRF resolutions, but RMSE and MBE were large.
The comparison on the 17 th October shows excellent correlation for the Mace Head centred grid point, with wind speed correlation R 2 between 0.90 and 0.94 for all resolutions, and small MBE and RMSE. Wind direction also agreed very well, with R 2 between 0.75 and 0.92, and smaller MBE and RMSE than from the run on 16 th October.

WRF Model real-time Analysis & Storm tracking
In terms of local scale meteorology, we developed a unique interactive forecasting evaluation system that provides real-time, in-situ, observational data delivered with a range of model products to allow the viewer to discern model accuracy, and which one is performing best. Figure 9 & Figure 10 below shows the models performance during Ophelia using the real-time analysis system, three distinct models are all compared at their normal 25 km x 25 km resolution and then the nested model simulations are compared at 25 km 2 , 5 km 2 and 1 km 2 . A range of real-time meteorology data at Mace Head with a 1minute temporal resolution is streamed and compared against the WRF model forecast.   From Figure 9 and 10, we can see that as observed from our lidar comparison study, that the WRF 5km, appears to follow the real-time in-situ measurements at 10m best, but fails to capture the peak wind speeds. The WRF 25km showed a delay in capturing the max wind speeds and a time lag in wind direction during Storm Ophelia. While the WRF 1km, underestimates the wind speeds during Ophelia. The importance of WRF horizontal resolution is very apparent.
Finally, WRF Simulations were also evaluated in terms of their ability to track the path of Ophelia, Figure 11 show's Ophelia simulation maps from the 15 th -18 th October 2017. From Figure 11(A-D), you can see the track of Ophelia as it makes its way towards Ireland. The eye of the storm is clearly visible in Figure 11

Conclusions
In this study we analysed WRF simulations during storm Ophelia, looking at key meteorological parameters, wind speed, wind direction and compared the model results against in-situ observations and remote sensing Lidar measurements at Mace Head Atmospheric research station. The model realistically simulated the magnitude and geographical distribution of the wind speed and wind direction observed during Ophelia, with good agreement with LIDAR data for both wind speed and direction on the 16 th October and excellent agreement with wind speed and direction on the 17 th October (R 2 >0.9 for the WRF 25km domain and the 5km domain). The model simulation showed to have generally small biases in the simulated wind speed and wind direction during this case study.
The importance of selecting representative grid points for evaluating the model performance has also been shown to be of relevance at Mace Head. As Mace Head is located in the boundary of ocean and land there is a transition between two distinct surfaces which need more sub-grid points to resolve the sub-grid scale problem. We can see from our study that moving the grid inland (Mace Head Sudg) increases the performance of the model during Ophelia (16 th October) with good correlations (R 2 >0.58) for wind speed, with good correlations (R 2 =0.78 for WRF 1km resolution) as the grid spacing approaches 1 km. Wind direction also correlated well, with R 2 >0.53 for all WRF resolutions, with excellent correlation (R 2 =0.87) as the resolution increases to 1km.
From this study, the correct choice of model resolution becomes apparent, we can see that the WRF 25km has a delay in capturing the max wind speeds and a time lag in wind direction. Results show that as seen with the in-situ measurements, the nested 5km WRF resolution performs best for both Wind Speed and Wind Direction on the 16 h & 17 th October, the model's capability to forecast extreme weather events such as Storm Ophelia is of upmost importance and now with a changing climate, it becomes more important to improve and enhance model forecasting capability.