Further study of the winds recorded on a bridge in Hong Kong for two tropical cyclone events in 2023

1. Introduction

Tropical cyclones could bring about high winds that may adversely affect road traffic (Baker et al.,1992; Zhu et al., 2012; Kim et al., 2020). For the operation of bridges, normally, there are pre-defined wind strengths over which the road traffic may be reduced or even halted. Accurate forecasting of the times and the wind speeds would be crucial for the timely decision of the operation of the bridges, especially in tropical cyclone situations. Apart from affecting the traffic, high winds might also cause wind-induced vibration and have an impact on the structure of the bridge itself (Xu, 2013; Fujino et al., 2013; Zhang et al., 2021; Zhao et al., 2024). Thus, the capability to simulate the wind speeds on a bridge would be of great interest.

Computational fluid dynamics (CFD) simulations and wind tunnel experiments have both been used to study and predict wind speeds on bridges (Zhang et al., 2021). However, wind tunnel testing often faces practical limitations such as scale effects (Wang et al., 1996) and the difficulty in representing realistic meteorological inflow. On the other hand, the use of CFD offers flexible control of boundary conditions and geometry, enabling targeted parametric analyses that are not always feasible in wind tunnel setups.

In Lo et al., 2024, an attempt has been made to forecast the wind speeds at the Cross Bay Link, in Hong Kong, during the passage of two tropical cyclones in which the territory was directly affected by them, namely, Saola in September 2023 and Koinu in October 2023. As the bridge is a rather small-scale structure compared with the spatial resolution (0.125° latitude) of the global numerical weather prediction model, nesting with a mesoscale meteorological model in large eddy simulation mode (namely, Regional Atmospheric Modelling System, RAMS version 6.3, down to a spatial resolution of 40 m with the use of Deardorff parameterization scheme (Deardorff, 1980) and further nesting with a CFD model, PALM (Maronga et al., 2015; Rassch et al., 2001; Maronga et al., 2020) (down to a spatial resolution of a few meters), have been conducted. The study aims at reproducing a real-time run of the nested models, namely, the forecast from a global model has been used instead of re-analysis. The forecast results are compared with anemometer readings recorded on the bridge. It is found that the model chain is able to generally reproduce the wind speed trend for one case (Saola), but the maximum wind speed at the top of the bridge arc is still under-estimated by 5 to 10 m∙s^-1. For Koinu, it is rather challenging to forecast the wind trend due to the rather small size of the cyclone and the difficulty in analyzing the wind structure of the cyclone by global models.

The present study is a follow-up of the previous study in Lo et al., 2024 with two major modifications:

(a) The global re-analysis is used instead of the forecast from a global numerical weather prediction model. In this study, the hourly analysis of the European Centre for Medium-range Weather Forecast (ECMWR) ERA5 with a spatial resolution of 0.125° latitude has been adopted to see if the comparison with anemometer readings could be improved using a global re-analysis instead of a forecast.

(b) The regional model of the Guangdong Meteorological Service (GMS) Tropical Regional Atmospheric Modelling System (TRAMS) with a spatial resolution of 9 km (Zhong et al., 2020). GMS-TRAMS is run in forecast mode with a higher spatial resolution and the assimilation of additional data (e.g., radar wind profilers, Doppler weather radar, etc.) compared to the global models, which may improve the analysis of the structure of the tropical cyclone as it gets closer to the coast of Guangdong. This study also investigates whether GMS-TRAMS could provide better driving data for the mesoscale meteorological model and CFD model.

2. Materials and Methods

The hourly re-analysis of ECMWF ERA5 is used in the first set of experiments. It is nested with RAMS and then a CFD model, PALM, in a setup similar to the previous study (Lo et al., 2024).

For the outermost domain of RAMS, the lateral boundary condition of the normal velocity component is specified by a radiative boundary condition with the Klemp-Wilhelmsen scheme (Klemp et al., 1978), where a typical constant value of 20 m∙s^-1 was adopted as the gravity wave phase velocity. For the other variables, the zero gradient inflow boundary condition and radiative outflow boundary condition were applied. For the inner domains, a two-way nesting technique was applied for all prognostic variables (Clark et al., 1984; Clark et al., 1991). The output of RAMS was then interpolated into the grids of PALM to create a dynamic driver file, which then provides the one-way offline coupling between RAMS and PALM.

Sample re-analysis fields of ERA5 could be found in Fig. 1. For both Saola and Koinu, the re-analysis fields appear to be reasonable, but the winds associated with Koinu seem to be under-estimated, as hurricane force winds have not been analyzed for this typhoon. The re-analysis field is used to provide boundary conditions for RAMS every hour.

Close

Fig. 1.

Reanalysis of Surface isobars and wind barbs from ERA5 at (a) 0000 UTC on 1 September 2023 and (b) 0000 UTC on 8 October 2023.

Export to PPT

The GMS-TRAMS is used in the second set of experiments. Since GMS-TRAMS is run only twice a day (initialization at 00 UTC and 12 UTC), the 00 UTC run is adopted in the present setup. Sample forecast model fields for Saola and Koinu have been given in Figs. 2(a, b), respectively. They are nested with RAMS for dynamic downscaling. In the initialization of GMS-TRAMS, the meteorological observations in the vicinity of southern China have been assimilated, including surface automatic weather stations, radar wind profilers, Doppler weather radar, etc. From Fig. 2(a), it can be seen that the hurricane force winds associated with Saola have been retained even with a forecast time of 12 h. On the other hand, for Koinu (Fig. 2b), the forecast winds appear to be underestimating, as the typhoon at that time is not forecast to have hurricane force winds.

Close

Fig. 2.

Forecast of surface isobars and wind barbs from GMS-TRAMS at (a) 1200 UTC on 1 September 2023 with initialization at 00UTC and (b) 1200 UTC on 8 October 2023 with initialization at 00UTC.

Export to PPT

3. Results

3.1. Results for Saola

The locations of the anemometers on the bridge could be found in Xu (2013). There are mainly two groups of anemometers, namely, those located on the bridge deck and the lower part of the arc, which measure relatively lower wind speed (CB1 to CB4, CB7 to CB8), and the two anemometers at the uppermost part of the arc on the bridge, which measure relatively high wind speed (CB5 and CB6). The correlation coefficients between the measured and the simulated wind speeds for RAMS and PALM can be found in Tables 1 and 2 for the use of ECMWF and GMS-TRAMS as boundary conditions, respectively. The result driven by NCEP is available in Table 1 in Lo et al. (2024). For the former group of anemometers, the use of ECMWF has the highest correlations, whilst the use of GMS-TRAMS has the lowest. For CB5, the highest correlation is obtained with the use of NCEP. This is also reflected in the percentage errors of the 10-min mean wind speeds (Table 3 in Lo et al., 2024, Tables 3 and 4) and the maximum of 10-min mean wind speeds (Table 5 in Lo et al., 2024, Tables 5 and 6) throughout the study. The timing of the maximum of 10-min mean wind speed is also closest to the actual observations for CB5 with the use of NCEP (Table 7 in Lo et al., 2024, Tables 7 and 8). As such, though only the forecast of NCEP is used instead of re-analysis, the highest wind speeds at CB5 are captured well, whereas overall (i.e., considering most anemometers) the use of ECMWF re-analysis would achieve the best results. On the other hand, the use of GMS-TRAMS does not seem to improve the comparison results.

Table 1. Correlation of 10-min mean and hourly mean wind speeds between measurements and simulations for Saola (driven by ERA5).

	CB1	CB3	CB5	CB6	CB7	CB8
RAMS (10-min mean)	0.81	0.68	0.51	0.71	0.65	0.49
PALM (10-min mean)	0.83	0.80	0.57	0.67	0.76	0.53
RAMS (hourly mean)	0.88	0.74	0.73	0.80	0.70	0.57
PALM (hourly mean)	0.90	0.85	0.75	0.78	0.80	0.67

Table 2. Correlation of 10-min mean and hourly mean wind speeds between measurements and simulations for Saola (driven by GMS-TRAMS 00Z).

	CB1	CB3	CB5	CB6	CB7	CB8
RAMS (10-min mean)	0.38	0.20	0.39	0.23	-0.04	0.10
PALM (10-min mean)	0.33	0.19	0.32	0.36	0.18	0.26
RAMS (hourly mean)	0.58	0.38	0.63	0.48	0.13	0.28
PALM (hourly mean)	0.62	0.26	0.54	0.57	0.28	0.39

Table 3. Average of 10-min mean wind speed and percentage error with respect to the measurement for Saola (driven by ERA5).

Wind speed (m∙s-1)	CB1	CB3	CB5	CB6	CB7	CB8
Observation	14.9	16.7	26.3	21.9	16.3	12.6
RAMS	13.0	12.7	15.5	15.6	12.7	12.8
% diff of RAMS	-12%	-24%	-41%	-29%	-22%	2%
PALM	14.2	14.8	17.5	17.4	16.6	14.0
% diff of PALM	-4%	-11%	-33%	-21%	2%	12%

Table 4. Average of 10-min mean wind speed and percentage error with respect to the measurement for Saola (driven by GMS-TRAMS 00Z).

Wind speed (m∙s-1)	CB1	CB3	CB5	CB6	CB7	CB8
Observation	14.9	16.7	26.3	21.9	16.3	12.6
RAMS	14.4	12.8	14.3	14.6	12.5	12.7
% diff of RAMS	-3%	-24%	-46%	-33%	-23%	1%
PALM	16.1	15.8	16.4	16.2	16.6	13.5
% diff of PALM	8%	-5%	-38%	-26%	2%	7%

Table 5. Maximum of 10-min mean wind speed and percentage error with respect to the measurement for Saola (driven by ERA5).

Wind speed (m∙s-1)	CB1	CB3	CB5	CB6	CB7	CB8
Observation	30	29.1	44.9	41.7	31.2	21.7
RAMS	23.3	22.5	26.7	26.7	22.8	22.9
% diff of RAMS	-22%	-23%	-41%	-36%	-27%	5%
PALM	26.3	27.8	30.4	32.3	28.8	24.9
% diff of PALM	-12%	-4%	-32%	-22%	-8%	15%

Table 6. Maximum of 10-min mean wind speed and percentage error with respect to the measurement for Saola (driven by GMS-TRAMS 00Z).

Wind speed (m∙s-1)	CB1	CB3	CB5	CB6	CB7	CB8
observation	30	29.1	44.9	41.7	31.2	21.7
RAMS	20.6	17.1	19.4	20.3	18.0	18.6
% diff of RAMS	-31%	-41%	-57%	-51%	-42%	-14%
PALM	24.7	25.2	24.5	23.8	26.3	19.7
% diff of PALM	-18%	-13%	-45%	-43%	-16%	-9%

Table 7. Time at which the 10-min mean wind speeds reached the maximum for Saola (driven by ERA5).

Maximum wind time	CB1	CB3	CB5	CB6	CB7	CB8
Observation	21:41:00	21:45:00	19:01:00	21:51:00	21:51:00	21:51:00
RAMS	21:38:50	20:10:50	20:11:00	20:11:00	20:11:00	20:11:00
PALM	21:38:00	21:03:20	21:01:40	21:01:40	21:02:40	21:02:40

Table 8. Time at which the 10-min mean wind speeds reached the maximum for Saola (driven by GMS-TRAMS 00Z).

Maximum wind time	CB1	CB3	CB5	CB6	CB7	CB8
Observation	21:41:00	21:45:00	19:01:00	21:51:00	21:51:00	21:51:00
RAMS	22:09:10	18:30:50	18:31:00	23:18:10	23:18:00	23:18:10
PALM	23:20:40	23:24:00	22:28:00	18:09:20	22:25:20	22:25:00

To study the comparisons in more detail, the time series of 10-min mean wind speeds at CB5 and CB6 for the three boundary conditions have been shown in Figs. 3 and 4, respectively. It could be seen that, though the simulated values could differ from the actual observations by 5 to 10 m∙s^-1 for the peaks in wind speeds, the simulation results are closer to the actual observations with the use of NCEP for CB5 and CB6 in comparison with the use of the two other models as boundary conditions. In particular, the use of GMS-TRAMS has achieved the lowest wind speeds among the three. As such, the structure of the tropical cyclone Saola may not be represented well in the GMS-TRAMS model, even though it has the finest resolution among the three sets of model data.

Close

Fig. 3.

10-min mean wind speed at CB5 on 1^st September 2023 (local time) for the following cases: (a) NCEP-driven, (b) ERA5-driven, and (c) GMS-TRAMS-driven.

Export to PPT

Close

Fig. 4.

10-min mean wind speed at CB6 on 1^st September 2023 (local time) for the following cases: (a) NCEP-driven, (b) ERA5-driven, and (c) GMS-TRAMS-driven.

Export to PPT

Another way to examine the simulation results is to look at the wind directions in the models. The results for CB5 have been shown in Fig. 5. In general, the changes in wind direction in the process are captured equally well by the three boundary conditions. However, discrepancies are found at times of maximum wind speeds, namely, between 21:00H and 23:00H local time.

Close

Fig. 5.

Wind direction at CB5 on 1^st September 2023 (local time) for cases: (a) NCEP-driven, (b) ERA5-driven, and (c) GMS-TRAMS-driven.

Export to PPT

To examine the difference in wind speed simulations in more detail, the 2D wind fields at around the height of the highest anemometer (CB5, i.e., about 64 m above mean sea level) for the three boundary conditions have been shown in Figs. 6(a-c), respectively, at the times in which the simulated wind speeds at CB5 are highest. The reverse flow downstream of the buildings at the upper boundary of the domain is captured by all three models. However, downstream of that, the wind speeds with the use of NCEP as boundary condition are the highest, and as such, the wind speed at the anemometer CB5 is also the highest. With the use of NCEP, winds as high as 35 m∙s^-1 are simulated, but they are absent from the other two numerical models at the times of the highest simulated wind speeds of CB5. The vertical cross-section across the bridge (location could be found in Fig. 2 of Xu (2013)) has been shown in Fig. 7 for the three boundary conditions. Both NCEP and ECMWF forecast descending jet upstream of the bridge, which may lead to the higher simulated wind speeds at the anemometer locations CB5 and CB6. On the other hand, the winds are generally much weaker with the use of GMS-TRAMS as a boundary. The GMS-TRAMS model appears to have issues in representing the structure of the cyclone, both horizontally and vertically.

Close

Fig. 6.

Horizontal wind fields at around 64 m above mean sea level. (a) NCEP-driven, (b) ERA5-driven, and (c) GMS-TRAMS-driven. The selected time corresponds to when the simulated wind speeds at CB5 reach their maximum.

Export to PPT

Close

Fig. 7.

Vertical cross-section of wind fields across the bridge. (a) NCEP-driven, (b) ERA5-driven, and (c) GMS-TRAMS-driven. The selected time corresponds to when the simulated wind speeds at CB5 reach their maximum.

Export to PPT

3.2 Results for Koinu

The correlation coefficients between the simulated and the observed wind speeds have been given in Table 5 in Lo et al. (2024), Tables 9 and 10, respectively. They are mostly negative and the use of re-analysis and regional models do not seem to improve the comparison. The 10-min mean wind speed and the maximum have also been under-estimated, as given in Tables 4 and 6 in Lo et al. (2024), Tables 10-15 for the three boundary conditions. In the various forecasts, the times of occurrence of the maximum 10-min mean wind speeds have significant discrepancies with the actual observations (Tables 1 and 16). As such, the simulation results for Koinu are not so satisfactory, even with the trial of different external models as boundary conditions.

Table 9. Correlation of 10-min mean and hourly mean wind speeds between measurements and simulations for Koinu (driven by ERA5).

	CB1	CB3	CB5	CB6	CB7	CB8
RAMS (10-min mean)	-0.57	-0.35	-0.34	-0.26	-0.16	-0.26
PALM (10-min mean)	-0.43	-0.19	-0.20	-0.05	0.00	-0.33
RAMS (hourly mean)	-0.93	-0.82	-0.81	-0.68	-0.58	-0.60
PALM (hourly mean)	-0.93	-0.82	-0.55	-0.52	-0.59	-0.70

Table 10. Correlation of 10-min mean and hourly mean wind speeds between measurement and simulations for Koinu (driven by GMS-TRAMS 00Z).

	CB1	CB3	CB5	CB6	CB7	CB8
RAMS (10-min mean)	0.05	0.37	0.43	0.34	0.30	0.24
PALM (10-min mean)	-0.31	-0.19	0.27	0.51	0.08	-0.05
RAMS (hourly mean)	0.55	0.66	0.79	0.50	0.52	0.41
PALM (hourly mean)	-0.74	-0.76	0.35	0.89	-0.44	-0.08

Table 11. Average of 10-min mean wind speed and percentage error with respect to the measurement for Koinu (driven by ERA5).

Wind speed (m∙s-1)	CB1	CB3	CB5	CB6	CB7	CB8
Observation	12.0	13.2	18.4	17.7	13.4	8.9
RAMS	8.5	7.7	8.6	8.8	7.6	7.7
% diff of RAMS	-29%	-42%	-53%	-50%	-43%	-14%
PALM	9.1	8.5	10.0	9.5	9.6	8.0
% diff of PALM	-25%	-36%	-46%	-47%	-28%	-10%

Table 12. Maximum of 10-min mean wind speed and percentage error with respect to the measurement for Koinu (driven by ERA5).

Wind speed (m∙s-1)	CB1	CB3	CB5	CB6	CB7	CB8
Observation	16.6	18.8	32.7	25.6	19.5	14.5
RAMS	12.7	11.8	13.4	13.8	12.0	12.2
% diff of RAMS	-23%	-37%	-59%	-46%	-38%	-16%
PALM	13.6	14.1	15.0	14.3	16.5	13.1
% diff of PALM	-18%	-25%	-54%	-44%	-15%	-10%

Table 13. Average of 10-min mean wind speed and percentage error with respect to the measurement for Koinu (driven by GMS-TRAMS 00Z).

Wind speed (m∙s-1)	CB1	CB3	CB5	CB6	CB7	CB8
Observation	12.0	13.0	13.2	8.9	18.4	17.7
RAMS	8.2	8.2	7.3	7.4	8.1	8.2
% diff of RAMS	-32%	-37%	-44%	-17%	-56%	-54%
PALM	8.1	9.1	7.3	7.9	9.3	8.9
% diff of PALM	-32%	-30%	-44%	-11%	-50%	-49%

Table 14. Maximum of 10-min mean wind speed and percentage error with respect to the measurement for Koinu (driven by GMS-TRAMS 00Z).

Wind speed (m∙s-1)	CB1	CB3	CB5	CB6	CB7	CB8
Observation	16.6	18.8	32.7	25.6	19.5	14.5
RAMS	10.8	9.0	10.1	10.6	8.9	9.2
% diff of RAMS	-35%	-52%	-69%	-59%	-54%	-37%
PALM	12.2	11.6	12.7	12.0	12.3	9.9
% diff of PALM	-27%	-38%	-61%	-53%	-37%	-32%

Table 15. Time at which the 10-min mean wind speeds reached the maximum for Koinu (driven by ERA5).

Maximum wind time	CB1	CB3	CB5	CB6	CB7	CB8
Observation	20:05:00	20:08:00	22:27:00	20:05:00	20:41:00	20:41:00
RAMS	18:02:00	18:02:30	18:02:30	18:02:40	18:02:30	18:02:40
PALM	17:43:00	16:23:40	16:24:00	16:24:00	16:24:00	16:24:00

Table 16. Time at which the 10-min mean wind speeds reached the maximum for Koinu (driven by GMS-TRAMS 00Z).

Maximum wind time	CB1	CB3	CB5	CB6	CB7	CB8
Observation	20:05:00	20:08:00	22:27:00	20:05:00	20:41:00	20:41:00
RAMS	16:08:40	21:44:30	21:44:20	21:44:50	21:44:20	21:44:30
PALM	16:59:00	16:34:00	21:03:40	20:15:20	16:09:40	16:13:00

To illustrate the relatively lower correlation, the time series of 10-min mean wind speeds for CB5 and CB6 has been shown in Figs. 8 and 9, respectively, with the use of the three boundary conditions. It could be seen that, in the period of the high wind speeds (between 20:00H and 23:00H local time), the simulated wind speeds remain generally much lower. This contributes to the negative correlation coefficient as the trend of the actual and simulated wind speed is opposite within the simulation period. The simulated wind directions (Figs. 10 and 11) are also generally different from the actual observations. Koinu is a very compact typhoon, and it appears that its structure could not be represented well even with the use of the state-of-the-art global reanalysis data or a high-resolution mesoscale model with data assimilation. Possibilities of further improving the simulation results are being explored.

Close

Fig. 8.

Time series of 10-min mean wind speed for CB5 on 8^th October 2023 in local time for cases: (a) driven by NCEP; (b) driven by ERA5, and (c) driven by GMS-TRAMS.

Export to PPT

Close

Fig. 9.

Time series of 10-min mean wind speed for CB6 on 8^th October 2023 in local time for cases: (a) driven by NCEP; (b) driven by ERA5 and (c) driven by GMS-TRAMS.

Export to PPT

Close

Fig. 10.

Time series of wind direction for CB5 on 8^th October 2023 in local time for cases: (a) driven by NCEP; (b) driven by ERA5, and (c) driven by GMS-TRAMS.

Export to PPT

Close

Fig. 11.

Time series of wind direction for CB6 on 8^th October 2023 in local time for cases: (a) driven by NCEP; (b) driven by ERA5, and (c) driven by GMS-TRAMS.

Export to PPT

4. Conclusions

The simulation of anemometer readings on a bridge in Hong Kong during the passage of Tropical Cyclone Saola and Koinu is studied further with the use of another two sets of boundary conditions, namely, hourly re-analysis of ECMWF and forecast with higher spatial resolution for the regional model GMS-TRAMS. It is found that, for Saola, the use of ECMWF re-analysis generally improves the comparison with actual observations, but the highest wind speeds at the anemometers over the bridge arc are still captured better with the use of NCEP forecast. The performance of the use of GMS-TRAMS turns out to be the worst. However, no significant improvements are found for the case of Koinu. In particular, negative correlation coefficients are found in the comparison between actual and simulated winds, and the wind directions are not captured well by the models. It turns out that Koinu is such a compact tropical cyclone that its representation in the numerical models still requires further improvements.

Accurate forecasting of wind speeds would be crucial in the timely operation of road and railway traffic in Hong Kong. Further possibilities of improving the comparison results, such as the relatively large discrepancy (5 to 10 m∙s^-1) for maximum wind speeds of Saola, and the negative correlation with actual observations for the case of Koinu, would be further studied with other possible boundary conditions and data assimilation methods.