Observations and Simulations of the Winds at a Bridge in Hong Kong during Two Tropical Cyclone Events in 2023

Abstract

The impact of tropical cyclones on the operation of bridges and railways are mostly dependent on the forecast wind speed trend (upward, downward or staying steady) at present in Hong Kong. There are requests to forecast the exceedance of wind speed thresholds for such operations with a lead time of many hours ahead. This study considers the technical feasibility of forecasting the wind speed along a recently built bridge in Hong Kong with a coupled mesoscale numerical weather prediction (NWP)–computational fluid dynamics (CFD) model. This bridge features numerous anemometers where the coupled model can be verified. It is found that these two tropical cyclone cases are very challenging, especially in representing the wind structure of the cyclone because of its relatively small circulation. As such, the timing of the maximum wind could be 2 to 4 h earlier than the actual observation. The maximum wind speed from CFD modeling could be higher than that from the NWP model alone by 5 to 10 m/s, which shows that CFD modeling could add value in the forecasting of wind speed exceedance, although the maximum simulated value is still lower than the actual observation by as much as 10 m/s. As a result, while the general wind speed trend may be forecast, the exceedance of definite values of wind speed limits is still practically rather challenging, given the present two cases of tropical cyclones.

1. Introduction

Situated along the southern coast of China, Hong Kong is very often affected by tropical cyclones every summer. On average, there are about four to seven tropical cyclones occurring within 500 km of Hong Kong each year. Hong Kong is also a highly urbanized city. There are many bridges over waters, as well as many open sections of railways. The Hong Kong Observatory (HKO) is responsible for providing weather support for the operation of the bridges and some of the railways, especially in tropical cyclone situations. There are wind speed limits designed for the operation of the bridges and the open sections of the railways. The observatory may be asked to give a trend for the future wind speed given the current wind situation, i.e., to predict whether there will be an increasing trend or decreasing trend in the wind speed, or whether the wind speed will remain roughly the same, in order for the bridge operator to consider closing the bridge. The timely closing of the bridge is of the utmost importance as wind-induced accidents, including the overturning of vehicles, have been reported worldwide [1,2,3]. For a recent summary of the literature on vehicle safety assessments along lengthy bridges exposed to high winds, readers can refer to [4].

There have been enquiries about the technical feasibility of forecasting the exceedance of wind speed limits of a bridge or the open section of a railway when Hong Kong is affected by tropical cyclones—for example, with an advance assessment of several hours ahead. The exact forecast of the timing of the exceedance of/failure to reach the wind speed limit is rather challenging, but there could be requests for such services. In recent years, with the increasing computational power at an affordable cost, it has been possible to perform high-resolution numerical weather prediction (NWP) modeling coupled with computational fluid dynamics (CFD) modeling at a high spatial scale, e.g., in the order of meters, along bridges or open sections of the railway in order to study the technical feasibility of supporting bridge/road/railway operation with such an approach.

The feasibility of using NWP-CFD coupled models to predict urban winds due to tropical cyclones has also been demonstrated recently [5,6,7,8,9,10,11,12]. The NWP models are necessary to provide realistic boundary conditions to drive the high-resolution CFD simulations, which can resolve the complex flow patterns around buildings and other structures in urban areas. For example, Schulman and DesAutels conducted a CFD simulation with the k-epsilon turbulence model and boundary conditions from the mesoscale model MM5 to investigate the pressure forces acting on a steel storage building during Hurricane Katrina [5]. They found that the simulated pressure exceeded the damage threshold determined by structural engineers and thus explained the observed damage. Takemi et al. studied the wind gusts in an urban district of Osaka City, Japan during Typhoon Jebi [8] by merging a mesoscale weather research forecast (WRF) model and large-eddy simulation (LES) with the Smagorinsky turbulent model [13]. The maximum simulated wind gust was 60–70 ms⁻¹ and was comparable to the measured values. Huang et al. used a WRF-CFD coupled model to study the wind speeds and pressure fields surrounding a tall building located in a densely built area of Hong Kong during Typhoon Kammuri [11]. The simulated wind-induced vibration of the tall building was also in good agreement with the measurement data.

The bridge under consideration in this study, Cross Bay Link, opened in 2022. It is the first marine viaduct in Hong Kong that contains a carriageway, cycle track, and footway. It has butterfly-shaped arc structures in its middle and is the longest steel arch bridge in Hong Kong, with a main span of about 200 m. The highest point of the bridge arch is about 70 m above sea level, and the width of the bridge is around 34 m (for more details about Cross Bay Link, readers can refer to https://www.cblsbtko.hk/?lang=en (accessed on 26 August 2024)). There are eight anemometers along the bridge, and high-resolution NWP coupled with CFD modeling can be performed for comparison with actual observations, to illustrate the challenges in forecasting the wind speed trend in tropical cyclone situations in a typical coastal city like Hong Kong.

This study will use two tropical cyclone cases in 2023, namely Super Typhoon Saola and Severe Typhoon Koinu, as an illustration of this technical feasibility. The synoptic surface weather charts for Saola at 08:00 H on 1 September 2023 in local time (+8 UTC; local time is assumed henceforth, unless otherwise specified) and Koinu at 08:00 H on 8 October 2023 are given in Figure 1a,b, respectively. In both cases, the tropical cyclones are represented by tightly packed close isobars situated near the coast of Southern China. To illustrate their sizes, satellite imagery from Himawari-9 of the Japan Meteorological Agency for Saola and Koinu at the corresponding time is shown in Figure 1c,d. The yellow lines represent a distance of 2°. From the synoptic weather charts and satellite images, Saola appears rather compact, with the radius of overcast cloud being less than 2°, and Koinu is even smaller.

One way to quantify the size of a tropical cyclone is through its radius of maximum wind (RMW). Lau et al. estimated that the RMW for Saola decreased from about 22 km at 08:00 H on 1 September 2023 to 13 km at 18: 00H on the same day, based on the Doppler velocity field [14]. He et al. estimated Koinu’s RMW to be around 20 km when it was located just to the south of Hong Kong on 8 October 2023, based on images from synthetic aperture radar [15]. In contrast, Lu et al. estimated the mean and median values of the RMW for tropical cyclones over the Western North Pacific to be about around 47 and 48 km, respectively [16]. Thus, the RMW for both Saola and Koinu was less than half of the typical value for a tropical cyclone. Due to their small sizes, the locations and structures of these tropical cyclones could not be accurately captured by global models with a resolution of around 10 km. Even slight differences in their locations and wind structures could lead to significant changes in the local wind conditions. As such, forecast centers issuing storm-related warnings were rather challenged ([17] for Saola and [18] for Koinu). It was found that the high sensitivity of the local wind conditions to the precise structures and locations of these two tropical cyclones restricted the applicability of the current approach of NWP-CFD in forecasting the wind speeds on the bridge under consideration. In the case of Saola, bridge observations confirmed some predictive capabilities in the general trend of the wind speeds and maximum wind along the bridge deck.

2. Actual Observations

There are eight anemometers along the bridge. Their locations with a scale bar can be found in Figure 2a, where CB1, CB2, CB3, CB4, CB5, CB6, CB7, and CB8 are approximately 30, 40, 20, 20, 70, 70, 20, and 20 m above the mean sea level, respectively. A vertical sectional view of part of the bridge is shown in Figure 2b for reference. The anemometers CB5 and CB6 are at the top of the bridge arc, while the anemometers CB1, CB3, CB4, CB7, and CB8 are closer to the bridge deck, and they might better represent the winds experienced by vehicles, cyclists, or passengers. Taking CB1 as a reference point, the horizontal distances of CB2, CB3, CB4, CB5, CB6, and CB7 from CB1 are about 210, 370, 370, 420, 420, 490, and 490 m, respectively. In the following discussions, for simplicity of presentation, only the time series of the wind data of CB1, CB3, CB5, CB6, CB7, and CB8 are given.

Figure 3 shows the measured 10 min mean wind speeds for the eight anemometers in the case of Saola at (a) around the time when the simulated wind speeds were the highest (for comparison with the simulation result later) and (b) the time when the measured wind speeds at CB5 and CB6 were the highest. Regarding Figure 3a, the wind speeds for anemometers CB5 and CB6 were unfortunately unavailable as they did not pass the quality check. The wind speeds at the other six anemometers were all below 20 m/s. As shown in Figure 3b, the wind speeds at anemometers CB5 and CB6 were the highest, reaching around 40 m/s. The wind speeds at the other six anemometers slightly increased, mostly in the range of 20 to 28 m/s.

Figure 4 shows the measured 10 min mean wind speeds for the eight anemometers in the case of Koinu at (a) around the time when the simulated wind speeds were the highest and (b) the time when the measured wind speed at CB5 was the highest. As shown in Figure 4a, the anemometers CB5 and CB6 recorded wind speeds of around 15 m/s, while the other six anemometers recorded wind speeds of around 10 m/s only. According to Figure 4b, the anemometer CB5 had the highest wind speed of around 26 m/s. The anemometers CB1, CB2, CB3, CB6, and CB7 had wind speeds ranging between 10 and 16 m/s. The wind speeds from the anemometers appear to be rather uneven.

3. NWP Plus CFD Coupling

The Regional Atmospheric Modelling System (RAMS) version 6.3 (https://vandenheever.atmos.colostate.edu/vdhpage/rams.php (accessed on 26 August 2024)) is used as the mesoscale NWP model. The RAMS is nested within the global model of the National Centers for Environmental Prediction (NCEP), i.e., forecast and analysis data from the global model of the NCEP were used to determine the initial and boundary conditions of the RAMS. Five nested grids are used, and the spatial resolutions are 25 km, 5 km, 1 km, 200 m, and 40 m, respectively. All grids are centered at anemometer CB5 on the bridge. The five grids are shown in Figure 5a,b. The outermost grid has a time step of 10 s, and time step ratios of 10 are applied for successive domains. For turbulence parameterization, the Smagorinsky scheme is applied for the two outermost grids and the Deardorff scheme is applied for the remaining three grids [19]. The Leaf-3 land surface model is used [20,21]. The RAMS is suitable for microscale simulation and has been used for the LES of the atmospheric boundary layer in idealized and realistic conditions [22,23,24,25]. No building or bridge was included in the RAMS simulation in the current study.

The PALM is nested inside the innermost RAMS domain, which was formerly an abbreviation for the parallelized large-eddy simulation model and has now become an independent name [26]. The PALM is a parallelized large-eddy simulation (LES) CFD model that solves non-hydrostatic, filtered, incompressible Navier–Stokes equations in Boussineseq-approximated form with subgrid-scale covariance terms parametrized with a 1.5-order closure scheme [19]. The details of the PALM can be found in [26,27,28]. Figure 5c shows the innermost domain of the RAMS and the domain of the CFD model. The offline coupling between the RAMS and PALM is performed with a dynamic input file. The PALM provides Python scripts that can be used to extract and interpolate data from the WRF to generate a dynamic input file containing the initial and boundary conditions for the PALM of different meteorological variables, such as the wind, potential temperature, and mixing ratio. Custom modification was applied to the Python scripts such that they could generate the required dynamic input file using the RAMS simulation data.

The current PALM simulation setup has a spatial resolution down to 4 m with the application of vertical stretching and a stretching ratio of 1.08. The maximum vertical grid size is limited to 12 m. Synthetic turbulence generation is also applied [29,30]. The study employs a fifth-order upwind difference scheme [31] for momentum advection and a third-order Runge–Kutta scheme [32] for time stepping. Adaptive time steps are used, and the average time steps are about 0.05 s and 0.1 s for the cases of Saola and Koinu. The 3D spatial data of the bridge under study were provided by the Lands Department of the Government of the Hong Kong Special Administrative Region. The high resolution of the PALM allows the explicit representation of the bridge and the buildings in its upstream region. With a higher resolution, it is expected that the turbulent flow around the bridge can be better resolved and hence the wind speed can be more accurately simulated.

4. Results for Saola

The model is initialized at 06 UTC on 1 September 2023 and run until 17 UTC on this day. Another model run had been performed with an initialization time of 00 UTC. However, Saola was a rather small cyclone and it was not well captured by the global model of the NCEP (i.e., GFS). As such, there were significant errors in the location and size of Saola, even with a forecast time within 24 h of the observational verification. Chan et al. presented a study on the performance of global models in forecasting the track, intensity, and wind structure of Saola [17]. In particular, the root mean square error of the 24 h forecast position for all global models considered was greater than 50 km, more than three times that of the RMW of around 13 km for Saola when it was near Hong Kong. The wind structure forecast also generally deviates from the observations for all global models. According to the wind speeds measured at anemometer CB5, the time of peak winds predicted by the model was quite different from that observed. As such, the 00 UTC run is not used and the 06 UTC run is used to drive the CFD model. It is somewhat better to use 06 UTC as the initial time, although there are still differences in the timing, strength, and direction of the maximum wind speed at anemometer CB5. No attempt has been made to perform initialization at 12 UTC, because the forecast period would become too short to be of practical value.

Some of the RAMS domain 5 outputs are shown in Figure 6, including the wind (Figure 6a) and eddy dissipation rate (EDR, Figure 6b) at a height of around 60 m above the mean sea level. They were taken at the time when the simulated wind speed at anemometer CB5 is the highest. The winds were generally rather strong, in the order of 25 to 30 m/s. The EDR was well above 0.5 m^2/3s⁻¹, indicating the turbulent nature of the boundary layer flow as Saola interacted with a region of complex terrain.

The simulated winds for the CFD model along the bridge at the same time as in Figure 3a,b are shown in Figure 7a,b, respectively. Comparing Figure 7a with the actual observations in Figure 3a, the simulated wind speeds are generally higher than the measured values for the corresponding time. However, if we compare Figure 7a with the actual observations in Figure 3b, although the times are not exactly the same, the wind distributions appear to be rather similar—namely, wind speeds of 40 to 45 m/s over the highest points of the bridge (anemometers CB5 and CB6) and generally around 20 to 30 m/s at the other anemometers. This suggests that there may be a temporal shift between the simulation results and the actual observations due to the errors in the location and wind structure of the tropical cyclone in the underlying NWP model.

To better illustrate the temporal shift, the time series of the 10 min mean wind speed and wind gust (maximum 3 s gust within 1 min for anemometer measurement and maximum wind speed within 1 min for PALM simulation at 20 s interval) for the six chosen anemometers are shown in Figure 8 and Figure 9, respectively. For the RAMS, the simulation data from its innermost domain are used to generate the time series. Spectral analysis in the frequency domain indicates that the turbulent energy spectrum of the measured data and both sets of simulated data generally follow the −5/3 Kolmogorov power law. The wind speed trends of anemometers CB1, CB3, CB7, and CB8 are similar between the actual observations and both simulation results (including the innermost domain of the RAMS and PALM). The correlations for the 10 min mean wind speeds between the anemometer readings and the simulation results are summarized in the first two rows of

Table 1. Correlation of 10 min mean and hourly mean wind speeds between measurement and simulations for Saola.

	CB1	CB3	CB5	CB6	CB7	CB8
RAMS (10-min mean)	0.49	0.46	0.56	0.58	0.44	0.35
PALM (10-min mean)	0.69	0.65	0.64	0.35	0.69	0.52
RAMS (hourly mean)	0.63	0.58	0.86	0.79	0.52	0.49
PALM (hourly mean)	0.77	0.71	0.86	0.64	0.70	0.62

. In general, the PALM demonstrates an improved correlation with the measurements, apart from CB6. Its correlations with the measurements are above 0.6 for CB1, CB3, CB5, and CB7, indicating that the PALM performed fairly well in capturing the trends of the wind speeds along the Cross Bay Link during the passage of Saola, despite the apparent time shift between the measurements and simulation results. To better demonstrate the general wind trend, the correlation of the hourly mean winds at 1 h intervals is also considered. The result is shown in the last two rows of Table 1. The correlations for the hourly mean wind speeds are higher than those for the 10 min mean wind speeds as smaller-scale temporal fluctuations are excluded. The PALM simulation again shows a better correlation with the measurements than the RAMS simulation for most of the anemometers, reaching 0.7 or above for four out of the six anemometers. The correlations are on par with those obtained in [11], which calculated that the correlations of the hourly mean wind speeds between the measurements and simulations at four weather stations in Hong Kong ranged from 0.66 to 0.78. This indicates the RAMS–PALM coupled model could capture the general wind trend along the bridge considered.

Table 2. Average of 10 min mean wind speed and percentage error with respect to the measurement for Saola.

Wind Speed (m/s)	CB1	CB3	CB5	CB6	CB7	CB8
measurement	14.9	16.7	26.3	21.9	16.3	12.6
RAMS	17.4	17.2	19.6	19.6	17.4	17.4
% diff of RAMS	17%	3%	−26%	−11%	7%	38%
PALM	17.5	16.3	23.7	24.4	20.0	18.2
% diff of PALM	18%	−3%	−10%	11%	23%	44%

and

Table 3. Maximum of 10 min mean wind speed and percentage error with respect to the measurement for Saola.

Wind Speed (m/s)	CB1	CB3	CB5	CB6	CB7	CB8
measurement	30	29.1	44.9	41.7	31.2	21.7
RAMS	29.9	30.2	34.7	34.5	31.0	30.9
% diff of RAMS	0%	4%	−23%	−17%	−1%	42%
PALM	28.4	34.0	38.9	38.8	31.4	27.3
% diff of PALM	−5%	17%	−13%	−7%	1%	26%

show the average and maximum of the 10 min mean wind speed for both the measured and simulated values for different anemometers. The percentage differences between the simulated values and measured values are also included. The mean and maximum wind speeds are very close to each other for anemometers CB1, CB3, CB7, and CB8 in the RAMS simulation, while the difference among these anemometers is more noticeable in the measured data and PALM simulation. This suggests that the difference in the winds for these anemometers is induced by the bridge and buildings upstream, which were not resolved in the RAMS simulation. The 10 min mean wind speeds from the RAMS tend to under-read compared to the actual observations at anemometers CB5 and CB6 (Figure 8c,d). The PALM adds value over the RAMS by providing simulated 10 min mean wind speeds that are around 5 m/s higher than the RAMS but are still below the actual observations by 5 to 10 m/s. For the other anemometers, both the RAMS and PALM values are, in general, similar to or higher than the actual observations.

While the maximum wind speed could provide an estimate of the potential wind hazards, its exact value might not be the most important in operation. In operation, various traffic control approaches would be considered once the wind speed exceeds certain critical values. Thus, if the simulation could correctly predict that the wind speeds on the bridge would exceed the critical value with a few hours of lead time, it would provide sufficient response time for the bridge operator to carry out the required traffic control measures. Previous studies on the critical wind speeds for different types of vehicles generally suggest a range of 14 m/s to 20 m/s [1,33,34,35]. Even taking a more conservative value of 20 m/s, the exceedance of this critical value is well captured by the CFD simulation for all anemometers. The times for exceedance are, in general, around or before the actual occurrence, as shown in Figure 8. A similar conclusion can also be drawn for the gust prediction, as shown in Figure 9.

Table 4. Time at which the 10 min mean wind speeds reached the maximum for Saola.

Maximum Wind Time	CB1	CB3	CB5	CB6	CB7	CB8
measurement	21:41:00	21:45:00	19:01:00	21:51:00	21:51:00	21:51:00
RAMS	20:09:50	20:11:40	20:11:40	20:11:40	20:11:40	20:11:40
PALM	21:08:40	21:02:40	20:07:00	20:07:00	21:35:40	21:02:20

shows the comparison of the time at which the anemometers reach their maximum values between the observations and simulations. For the observations, the times for the maximum wind speeds are generally around 22H, apart from CB5. For CB5, fluctuations are recorded at around 19H, leading to an earlier time for the maximum winds, but a local maximum is also attained at around 22H. The times for the RAMS simulation to reach the maximum values are around 20H for all anemometers. For the PALM simulation, it gives maximum winds at around 20H for CB5 and CB6, similar to that of the RAMS. Its times for the maximum winds are around 21H for the other anemometers. In general, the maximum values have a time difference of around 1–2 h between the observations and numerical simulations, and the difference is even larger in the model run initialized at 00 UTC. This time discrepancy is related to the inadequate NCEP model prediction of Saola’s track and wind structure, even with a forecast lead time of around 12 h only. It is reflected in the wind direction at CB5 when comparing the actual and the simulated values. The comparison is shown in Figure 10. The measured and simulated wind direction of CB5 are both generally northerly to begin with. However, the PALM exhibits a sharp change in wind direction to easterly at around 21:30H, while the actual measurement shows a much more gradual shift from generally northerly at around 20H to easterly winds at around 23H. Comparing this to the time series of the wind speeds in Figure 8, both the measured and simulated wind speeds show a generally decreasing trend after the wind direction veers to the east, which could be attributed to the local terrain effect due to the higher topography along the two sides of the bridge.

As an illustration of the simulated wind field for the PALM, the wind distribution at a height of about 70 m above the mean sea level is shown in Figure 11a. It can be seen that the PALM manages to illustrate the fine details of the wind pattern, e.g., the lower wind speed and even the reverse flow by a small hill near the northern boundary of the simulation domain. A vertical cross-section of the wind is obtained across the bridge. The location of the cross-section is shown in Figure 2a. An example of the cross-section is shown in Figure 11b. Due to the blockage of the bridge deck and the arc, the wind speeds are lower downstream of these super-structures of the bridge. On the other hand, the wind speeds are much higher if there are no physical blockages.

5. Results for Koinu

Following similar reasoning as in the case of Saola, the model initialized at 06 UTC on 8 October 2023 is considered in the present work, running until 16 UTC on this day. The 00 UTC run is found to have too large a discrepancy in the timing and strength of the maximum wind speed at anemometer CB5 at the bridge, although the discrepancy in the case of 06 UTC is not too much smaller. The 12 UTC run is considered to have limited practical application value. Using a model run with earlier initialization to drive the coupled NWP–CFD model is not possible with Koinu again due to the relatively small circulation of this cyclone (referring to Figure 1b), which is poorly resolved by the global NCEP model.

The RAMS-simulated winds at about 60 m above the mean sea level at the time of the maximum simulated wind at anemometer CB5 are shown in Figure 12a. A higher wind speed has been simulated at the location of the anemometer downstream of a gap flow in the northerly to northeasterly winds. The EDR pattern is shown in Figure 12b. Compared to the case of Saola, the turbulence is much weaker this time. Flares of higher turbulence (severe turbulence, coloured red) appear downstream of the hills, but they do not manage to reach the location of the bridge.

The PALM-simulated winds at the anemometer locations along the bridge at the same time as in Figure 4 are shown in Figure 13. Compared to the actual observations, the differences are rather large. A shift in time is again observed, and it is even larger for Koinu. Nevertheless, the wind speeds are comparable for Figure 4b and Figure 13a. The uneven wind distribution over the bridge is rather well captured by the PALM model. The anemometer CB5 has the highest wind speed of around 18–20 m/s, albeit weaker than the actual observation. The anemometer CB4 has the lowest speed wind, below 10 m/s. The other anemometers have wind speeds in the region of 10 to 15 m/s.

The time series of the 10 min mean winds and gusts at the anemometer locations CB1, CB3, CB5, CB6, CB7, and CB8, for both the actual and simulated data, are shown in Figure 14 and Figure 15, respectively. Compared to the case of Saola, the simulation results of Koinu do not appear to be satisfactory. The correlations of the 10 min mean wind speed between the observations and simulations are even negative, as shown in

Table 5. Correlation of 10 min mean wind speed between measurements and simulations for Koinu.

	CB1	CB3	CB5	CB6	CB7	CB8
RAMS	−0.37	−0.27	−0.19	−0.1	−0.03	−0.3
PALM	−0.35	−0.18	−0.14	−0.12	−0.13	−0.12

, indicating that the PALM and RAMS fail to capture the general wind trend in the case of Koinu.

Table 6. Average of 10 min mean wind speed and percentage error with respect to the measurement for Koinu.

Wind Speed (m/s)	CB1	CB3	CB5	CB6	CB7	CB8
measurement	12.0	13.2	18.4	17.7	13.4	8.9
RAMS	9.7	9.2	10.4	10.4	9.0	9.1
% diff of RAMS	−19%	−30%	−44%	−41%	−32%	1%
PALM	8.9	8.9	11.5	10.9	9.6	10.0
% diff of PALM	−26%	−32%	−38%	−38%	−28%	12%

and

Table 7. Maximum of 10 min mean wind speed and percentage error with respect to the measurement for Koinu.

Wind Speed (m/s)	CB1	CB3	CB5	CB6	CB7	CB8
measurement	16.6	18.8	32.7	25.6	19.5	14.5
RAMS	13.6	13.5	15.6	15.8	14.2	14.3
% diff of RAMS	−18%	−28%	−52%	−38%	−27%	−1%
PALM	16.7	16.9	18.5	17.4	20.5	15.6
% diff of PALM	0%	−10%	−43%	−32%	5%	8%

show a comparison of the average and maximum of the 10 min mean wind speeds between the measured and simulated values. In terms of the average of the 10 min mean wind speed, the simulated values are, in general, lower than the actual observations, apart from CB8. For the maximum of the 10 min mean wind speed, the values in the PALM simulation are below those of the actual wind data for the highest anemometers CB5 and CB6, by 5 to 10 m/s, although the PALM already has value over the RAMS output by having an addition of around 5 m/s in the maximum simulated wind speed. For the other anemometers, the PALM simulations are comparable to the actual observations.

Regarding the timing of the maximum wind, both the RAMS and PALM outputs are earlier than the actual observations by around 4 h. This difference appears to be too large to provide reliable forecast information. It is again related to the difficulty of the NCEP model in forecasting the track and wind structure of Koinu, which had very small circulation (see Figure 1b). The difference could be reflected in the actual and the forecast wind direction; see Figure 16. The wind direction of the PALM becomes easterly too quickly, followed by the RAMS. On the other hand, the real wind direction remains northeasterly throughout the period under consideration.

Figure 17 shows the HKO provisional best track and NCEP model forecast track for Saola, initialized at 2023090106UTC (a), and Koinu, initialized at 2023100806UTC (b). It appears that the errors in the locations of the tropical cyclones are comparable in the two cases, generally in the order of 10–20 km. Thus, the main cause of the worse performance of the CFD simulation in Koinu is not the larger error in the track but rather the poorer representation of the wind structure of Koinu. This might not be surprising as Koinu was a more compact tropical cyclone by comparison, as depicted in Figure 1. It is more challenging for global NWP models to capture such a structure correctly.

An example of the simulated wind pattern is shown in Figure 18. Again, the PALM manages to provide the fine details of the wind pattern, e.g., the weaker wind downstream of a hill near the northern boundary of the model simulation domain; see Figure 18a. The edge effect of the bridge’s structure also appears in the vertical cross-section of the wind field; see Figure 18b. Near the butterfly arc and the bridge deck, the wind speeds are simulated to be higher, reaching a gale forecast (red wind barbs).

6. Summary and Discussion

The technical possibility of accurately forecasting the exceedance of wind speed limits, in addition to the wind speed trend, is studied for a bridge in Hong Kong with a full set of anemometer data during two recent tropical cyclone events, namely Saola and Koinu in 2023. In both cases, accurately forecasting the timing of the maximum wind speed appears to be rather difficult, even with the use of the state-of-the-art NWP coupled CFD model simulation with a spatial resolution of down to 4 m horizontally. This is mostly attributed to the limitations of the global models in accurate forecasting the tracks and the structures of the cyclones. Saola and Koinu are found to be rather difficult to forecast due to their rather small cyclonic circulation. There could be around a 1 to 2 h difference in the timing of the maximum wind at the bridge between the actual and the simulated data for Saola. The difference is even larger and up to 4 h in the case of Koinu. Moreover, the wind direction is not simulated well by the models. Nevertheless, the general trend of the wind speeds is reasonably reproduced by the CFD model in the Saola case as the correlations for the 10 min mean wind speed are around 0.6 or above for most of the anemometers. However, negative correlations are observed for Koinu. In terms of the critical wind speed for traffic control, the CFD simulation manages to predict the exceedance of the critical wind speed, and the timings for exceedance are, in general, around or before the actual occurrence for Saola. This suggests that CFD simulation could provide some guidance for operational traffic control if such results are available in real time. However, its practical use for Koinu is rather limited due to the large difference in the timing of the maximum winds.

With the use of the CFD model, it is possible to forecast a maximum wind speed that is higher by 5 to 10 m/s over the mesoscale meteorological model for the two anemometers at the top of the bridge arc, albeit still lower than the actual data by around 10 m/s in terms of the 10 min mean wind speeds. For the other anemometers at an elevation closer to the bridge deck, which better represent the wind condition to be experienced by the vehicles on the bridge, the maximum winds predicted by the CFD model are generally similar to the observations. This opens up the technical possibility of forecasting the maximum wind speed with higher accuracy.

Moreover, the CFD model manages to forecast well the uneven distribution of the wind at the eight anemometers along the bridge. Thus, the technical feasibility of forecasting the wind distribution and identifying the portions that are more vulnerable to wind hazards along bridges/railways may be explored, although the forecast accuracy is still far from being practically useful in real-time applications. Nevertheless, this might provide valuable insights during the design phases of bridges/railways.

The applicability of the current approach is mainly limited by the ability of the global models to capture the track and wind structure of the tropical cyclone. Instead of using the NCEP’s GFS to drive the coupled model, another simulation using the Global and Regional Assimilation Prediction System [36,37] mesoscale model (GRAPES-Meso) to drive the coupled model was performed for the case of Saola. GRAPES-Meso has a horizontal resolution of 3 km and might better resolve the horizontal structure of a tropical cyclone. The model run initialized at 2023090100Z was used. The time series for the 10 min mean wind speeds are plotted in Figure 19. The result indicates that the times to reach the maximum winds for the PALM only differ by around 30 min–1 h from the measurements at anemometers CB1, CB3, CB7, and CB8 (

Table 8. Time at which the 10 min mean wind speeds reached the maximum for Saola. The coupled model is driven by GRAPES-Meso.

Maximum Wind Time	CB1	CB3	CB5	CB6	CB7	CB8
measurement	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	18:31:00	18:31:00	18:31:00
PALM	22:09:40	22:55:40	22:28:00	18:09:20	22:25:20	22:25:00

), and the simulated maximum winds are only slightly smaller than the measurements (

Table 9. Maximum of 10 min mean wind speed and percentage error with respect to the measurement for Saola. The coupled model is driven by GRAPES-Meso.

Wind Speed (m/s)	CB1	CB3	CB5	CB6	CB7	CB8
measurement	30	29.1	44.9	41.7	31.2	21.7
RAMS	20.6	17.1	19.4	19.5	17.0	17.2
% diff of RAMS	−31%	−41%	−57%	−53%	−45%	−21%
PALM	24.2	24.7	24.5	23.8	26.3	19.7
% diff of PALM	−19%	−15%	−45%	−43%	−16%	−9%

). This suggests that simulations driven by GRAPES-Meso could provide guidance for traffic control as these anemometers are closer to the bridge deck and better represent the wind conditions experienced by the driver. However, the simulated wind speeds at CB5 and CB6 are smaller in this case; further investigation and tuning should be performed to improve the wind speed simulation for these two anemometers.

As the simulation result of the coupling model is largely influenced by the underlying NWP model, one possible improvement is to leverage ensembles to provide the initial and boundary conditions for the mesoscale–CFD coupled model. With this approach, probabilistic guidance on the exceedance of the wind speed limit could be generated. However, such an approach would be very computationally demanding.

Recently, Lim et al. suggested a data-driven approach for typhoon-induced strong wind prediction on long-span bridges [38]. By developing data-driven predictive models, the wind speed forecasting process could potentially be accelerated compared to computationally intensive CFD simulations. Building upon this concept, one possible extension of the current work could be to simulate a large sample of time series of wind speeds along the bridge for different tropical cyclone track and intensity scenarios. These simulated wind speed time series could then be used as training data to develop a machine learning model capable of predicting the wind speed time series on the bridge given a forecast tropical cyclone track. This would enable the generation of ensemble-based wind speed predictions in operation, without the need to run computationally expensive CFD simulations in real time.

Further tuning, including the assimilation of real wind data for the better representation of the wind structure of the typhoon, could be explored. More cases of tropical cyclones could be considered in future studies, at least for those with larger circulation and better captured by NWP models. An even higher spatial resolution and the even more realistic representation of the surrounding environment could also be considered.