Ultra-High Resolution Large-Eddy Simulation of Typhoon Yagi (2024) over Urban Haikou

Abstract

About 16% of typhoons making landfall in China strike Hainan Island, where near-surface extreme winds in dense urban areas exhibit a strong spatiotemporal heterogeneity that is difficult to capture with current observations and mesoscale models. Focusing on Haikou during Super Typhoon Yagi (2024)—the strongest autumn typhoon to hit China since 1949—we developed a multiscale ERA5–WRF–PALM framework to conduct 30 m resolution large-eddy simulations. PALM results are in reasonable agreement with most of the five automatic weather stations, while performance is weaker at the most sheltered park site. Mean near-surface wind speeds increased by 20–50% relative to normal conditions, showing a coastal–urban gradient: maps of weighted cumulative exposure to strong winds (≥Beaufort force 8) show much longer and more intense events along open coasts than within built-up urban cores. Urban morphology exerted nonlinear effects: wind speeds followed a U-shaped relation with both the open-space ratio and mean building height, with suppression zones at ~0.5–0.7 openness and ~20–40 m height, while clusters of super-tall buildings induced Venturi-like acceleration of 2–3 m s⁻¹. Spatiotemporal analysis revealed banded swaths of high winds, with open areas and islands sustaining longer, broader extremes, and dense street grids experiencing shorter, localized events. Methodologically, this study provides a rare, systematically evaluated application of a multiscale ERA5–WRF–PALM framework to a real typhoon case at 30 m resolution in a tropical coastal city. These findings clarify typhoon–city interactions, quantify morphological regulation of extreme winds, and support risk assessment, urban planning, and wind-resilient design in coastal megacities.

1. Introduction

Under global climate change, the intensity and frequency of tropical cyclones (TCs; called typhoons in the western North Pacific) are undergoing notable shifts, posing an increasingly severe threat to coastal regions [1,2,3]. Coastal cities, with highly concentrated populations and economic activities, are particularly vulnerable [4]; extreme winds at landfall frequently cause serious loss of life and property. High-density clusters of tall buildings substantially modify the near-surface boundary-layer wind field, producing strong spatial heterogeneity in local extreme winds [5], and even doubling the local mean wind speed relative to open terrain in some cases [6]. However, for super typhoons (e.g., Rammasun in 2014 or Katrina in 2005), large uncertainties remain in the interaction mechanisms between boundary-layer turbulence and urban surfaces [7]. Hainan, China’s only tropical island province, experiences frequent typhoon disasters with pronounced terrain–track modulation; the frequency of typhoon-induced extreme precipitation events increased during 2000–2020 [8]. At the national level, the Hainan Free Trade Port is a major strategic initiative, further highlighting the constraints that meteorological resilience imposes on critical infrastructure and urban functions [9,10].

Beyond the meteorological perspective, typhoon resilience in coastal cities has been extensively examined in the fields of disaster management and urban planning. Previous research has shown that typhoon hazards can interact with land-use change to reshape the resilience of socio-ecological systems [11] and that indicator-based frameworks can be used to assess local resilience to typhoon disasters in Nansha, Guangzhou [12]. Other studies have highlighted the role of disaster policies and networked governance in promoting community resilience [13] and have extended this perspective to infrastructure by using carbon emissions as an indicator when evaluating the resilience and sustainability of bridge systems under wind disasters [14]. Collectively, these studies show that typhoon resilience depends on hazard, urban form and governance; yet, typhoon winds are usually represented by coarse statistics or empirical proxies rather than building-resolving flow fields, leaving a gap that our high-resolution modeling seeks to fill [11,12,13,14].

At present, mesoscale models such as ERA5 and WRF are limited by kilometer-scale grids and parameterizations and thus struggle to resolve building-scale flow structures, often underestimating gust extremes within cities [15,16,17,18]. By contrast, large-eddy simulation (LES) models (e.g., PALM) can explicitly resolve urban turbulence on tens-of-meters grids (including the O (30 m) resolution used here) [17], providing far finer characterization of near-surface winds than traditional RANS approaches [18]. For the typhoon boundary layer, high-fidelity LES frameworks that have undergone systematic evaluation can, within O (5 km) subdomains, reproduce near-surface strong wind shear, low-level jets, and energy-spectrum characteristics and have been used to examine the applicability of planetary boundary layer (PBL) schemes [19]. LES under idealized and realistic urban morphologies further shows that the roughness jump induced by landfall rapidly spawns an internal boundary layer and redistributes turbulent kinetic energy, thereby amplifying near-shore gust extremes and altering the turbulence-scale structure [18,20]; in real cases, neighborhood-scale LES has successfully reproduced spatial contrasts in extreme wind pressure and window-failure risk around buildings [21]. In recent years, multiscale nesting between WRF and LES has attracted much attention [22]. For example, WRF–PALM simulations indicate that districts with dense high-rise buildings accentuate the nonuniformity of near-surface extreme winds [5]; introducing data assimilation into meso-to-micro coupling increased the consistency index for urban near-surface wind forecasts from 0.28 to 0.72 [23]. Combining mesoscale and microscale models offers clear advantages for simulating typhoon winds in urban areas and compensates for the deficiencies of any single-scale model. However, applications of such multiscale WRF–LES couplings to real coastal cities at O (30 m) resolution with systematic evaluation remain scarce, which motivates the present study.

During Typhoons Saola and Koinu, Hong Kong employed NWP–PALM coupling to verify bridge-site gust thresholds and revealed meter-scale wind-field differences [24]; in Tokyo, building-resolving LES evaluated strong urban winds during Typhoon Melor [7]; and in Miami, WRF simulations during Irma showed the weak sensitivity of near-surface urban winds to the choice of PBL scheme but high sensitivity to the urban canopy model (UCM) configuration [25]. Haikou is a low-latitude island city where coastline and dense built-up belts are juxtaposed, making 30 m LES particularly necessary to finely resolve the extreme winds of Typhoon Yagi. On 6 September 2024, Yagi made landfall along the coast of Wenchang, Hainan, as a super typhoon, with a maximum sustained wind near the center of 62 m s⁻¹ (≥Beaufort force 17) and a minimum central pressure of 915 hPa; the event has been characterized as the strongest autumn typhoon to make landfall in China since 1949 [26,27]. Gusts of force 11–15 were recorded over northern Hainan on land and nearshore, and an extreme gust of 66.7 m s⁻¹ (≥force 17) was measured in Longlou, Wenchang [28,29]; provincial assessments indicate that the destructive power far exceeded that of Rammasun (2014), making Yagi the second strongest typhoon to make landfall in Hainan [30]. The characteristic intensity of Typhoon Yagi thus provides an ideal case to test a multiscale modeling system. At building-resolving scales, embedded LES and multiscale coupling have achieved observational verification of urban wind fields and impacts in complex megacities [31].

In this study, we develop and apply a multiscale ERA5–WRF–PALM coupled modeling system to explicitly simulate the near-surface urban wind field in Haikou under the influence of super Typhoon Yagi at a 30 m resolution. ERA5 global reanalysis provides the background, WRF performs dynamical downscaling, and PALM finally resolves turbulence within the urban canopy and building-induced flow features. We therefore pursue three objectives: (1) to evaluate the system’s performance in reproducing the spatiotemporal evolution of urban winds under a super typhoon; (2) to quantify the nonlinear relationships between urban morphological parameters (e.g., open-space ratio, building height) and near-surface extreme winds; and (3) to elucidate differential responses across urban functional zones and their underlying physical mechanisms. The results not only deepen our understanding of typhoon–urban boundary-layer interactions but also provide key scientific evidence for wind-hazard risk assessment and climate-adaptation planning and design in densely built coastal cities.

2. Data and Methods

2.1. Observational Data and Meteorological Station Configuration

The study area is Haikou, the core city in northern Hainan Island, situated in a low-latitude tropical coastal setting dominated by coastal plains. Figure 1 shows the spatial distribution of land-surface types and building heights across the main urban area and its surroundings. The basemap, at meter-scale resolution, distinguishes open ground, water bodies, and built-up areas; the maximum building height reaches ~330 m. High-rise buildings are primarily concentrated on Haidian Island and in the Guomao (CBD) center, while the old town and coastal belts are characterized by lower building heights and extensive open surfaces and water bodies. The five instrumented sites—M1008, M1043, M1039, M1021, M1009—cover major functional zones and typical underlying-surface types. The inset depicts the track of Typhoon Yagi (2024) over northern Hainan and the location of the study domain. Because of power outages, communication failures, and instrument damage during the typhoon, many originally planned stations suffered data gaps; only these five sites maintained complete, continuous records. The boundaries of key functional zones (e.g., Qilou Old Street, Wanlv Park, Haidian Island, Guomao) are marked by colored rectangles to provide spatial references for subsequent analyses of wind-field features and spatial heterogeneity.

2.2. ERA5/WRF/PALM Model Configuration

In this study, “ERA5–WRF–PALM” denotes a three-stage multiscale workflow rather than a new standalone numerical model. We build a multiscale coupled modeling system driven by ERA5 reanalysis, with stepwise refinement from the regional to the urban scale via WRF mesoscale modeling and PALM large-eddy simulation (LES). The WRF-ARW v4.3 configuration employs triple nesting at 9 km, 3 km, and 1 km horizontal resolutions with 50 sigma levels. Physics options include WSM6 microphysics, RRTMG radiation, the YSU planetary boundary layer scheme, and the Noah land-surface model; cumulus parameterization is activated only on the coarser nests. The PALM LES uses a 30 m horizontal grid (400 × 270 points). The vertical grid starts at 10 m near the surface and is stretched by a factor of 1.11 up to 400 m, yielding 55 levels in total. Buildings are represented explicitly; the surface roughness length is set to 0.1 m. PALM is one-way coupled to WRF to obtain boundary conditions, and a turbulence recycling zone is configured to generate a fully developed inflow. The workflow uses ERA5 fields as large-scale forcing, near-surface wind observations for evaluation, and building-footprint and land-cover data to construct PALM static inputs. The main user-defined choices are the WRF domain configuration (extent, nesting and vertical levels), the planetary boundary-layer and land-surface schemes, and the PALM grid spacing, turbulence closure and surface roughness.

Observational data are taken from the five automatic weather stations in Haikou that span diverse urban functions and underlying surfaces; due to equipment failures during the typhoon, only these five maintained complete records. Urban building data are sourced from the Amap (Gaode) Open Platform (2021 edition) and were processed for PALM static boundary inputs. The simulation window is 12:00 UTC 5 September to 00:00 UTC 8 September 2024, covering the full influence period of Typhoon Yagi.

The analysis adopts a multiscale statistical evaluation strategy, including time-series comparisons, spatial correlation assessment, and extreme-value diagnostics. The correlation coefficient, root-mean-square error (RMSE), and systematic bias are computed to evaluate model performance, with emphasis on how urban structure modulates the wind field and on the spatial patterns of extreme winds.

For urban morphology and block definitions, the PALM 30 m grid is aggregated into regular blocks by 20 × 20 grid points (≈600 m × 600 m), which serve as the basic units for statistics and binning. For the resolved building surfaces, a uniform aerodynamic roughness length of z₀ = 0.1 m is prescribed, following previous PALM studies of explicitly resolved urban geometry, where building-scale roughness dominates over sub-grid roughness elements. Two indicators are computed for each block: (i) the open-space ratio (per block), defined as the fraction of grid cells labeled as open ground (cells with no building height assigned and not classified as water); and (ii) the mean building height (per block), i.e., the average of building heights over building cells only. Blocks without buildings are recorded as missing. For practical use, the ERA5–WRF–PALM workflow can be summarized in five steps: (1) prepare nested WRF domains and generate initial and boundary conditions from ERA5; (2) run WRF to downscale the typhoon circulation to 3 km and 1 km; (3) derive PALM static input files from building-footprint and land-cover data on the 30 m grid; (4) drive PALM with fields from the innermost WRF domain to simulate 10 m winds over urban Haikou; and (5) post-process model outputs together with station observations for time-series, spatial and extreme-value evaluation.

2.3. Time Standardization and Sequence Alignment

All data are standardized to UTC and to right-edge interval averages. In the observations, an hourly value at HH:00 represents the average over (HH–1:00, HH:00]. WRF outputs are aggregated to hourly means and then shifted forward by 1 h to match the observational semantics. PALM 5 min outputs are first aggregated to hourly means and then advanced by 1 h for alignment. For quality control, periods with <80% valid sample coverage are flagged as missing and excluded.

Based on the evolution of Typhoon Yagi, the typhoon period is defined as 00:00 UTC 6 September to 00:00 UTC 7 September 2024, and the pre-typhoon period as 00:00 UTC 5 September to 00:00 UTC 6 September 2024.

3. Results

3.1. Model Performance Evaluation

Figure 2 compares the temporal evolution of the 10 m mean wind speed over Hainan during Typhoon Yagi. Values rose rapidly to a peak of 13.5 m s⁻¹ around 06:00 UTC on 6 September, followed by a rapid decay. Both ERA5 and WRF captured the overall evolution, with R² = 0.49 and 0.86, respectively, relative to observations; however, both systems systematically overestimated winds during the peak and post-peak phases. WRF produced a transient spike exceeding 24 m s⁻¹ near 12:00 UTC on 6 September, while ERA5 exhibited a dip-and-rebound feature shortly after the peak. In the late stage of the event, both models overpredicted low-wind conditions, indicating systematic biases in the amplitude and timing of extremes during the typhoon.

Figure 3 evaluates the spatial performance of ERA5 and WRF for wind speed over Hainan. For spatial correlation, island-wide means were 0.72 for ERA5 and 0.64 for WRF; both performed best over the central–northern island (max up to 0.96), with negative correlations in parts of the south and mountainous regions (min down to –0.44). Mean RMSE values were comparable—5.27 m s⁻¹ (ERA5) and 5.33 m s⁻¹ (WRF)—with larger errors over urban and inland areas (peaks > 12 m s⁻¹) and smaller errors along the coast. Systematic bias maps show overall positive biases (overestimation) for both systems, with mean biases of 4.29 and 4.49 m s⁻¹; positive biases dominate the central and northwestern island, with negative pockets along parts of the coast. Detailed station-level time series and inter-model differences are provided in Appendix A (Figure A1 and Figure A2), including high-resolution (10 min) validation of PALM at the five representative sites. These patterns indicate that both ERA5 and WRF reproduce the broad structure of near-surface winds but tend to overestimate magnitudes over inland and urban areas during the peak and decay phases, and the correlation map highlights where the multiscale system is most and least reliable for potential applications.

3.2. Wind Field Characteristics

Figure 4 compares the probability distributions of U/V wind components at the five stations during the typhoon. Observations show predominantly negative U and positive V, indicating a prevailing southeasterly–southerly flow. PALM reproduces the modal positions of U/V but tends to overestimate spread and extremes, yielding more concentrated distributions with heavier tails, particularly at the high-wind sites M1008 and M1039. ERA5 and WRF align on the principal modes but show broader distributions and multimodality at some sites. At lower-wind stations M1009 and M1021, the primary modes in ERA5/WRF diverge more from observations, highlighting the limitations of reanalysis and mesoscale models in capturing local wind-direction variability.

Figure 5 shows wind roses for the five stations. During the event, all sites are dominated by winds from the southeast to south, consistent with the landfall circulation. The observed roses exhibit a strongly concentrated primary lobe (SE, SSE) with only limited secondary lobes. ERA5 and WRF reproduce this primary lobe and its azimuthal range well, with substantial overlap between their southeast–south sectors and the observed distribution. The 30 m PALM simulations also capture the dominant southeasterly sector at most stations, but the corresponding wind roses are more dispersed. This behavior is characteristic of high-resolution LES: PALM explicitly resolves local turbulence, gusts and flow-channeling effects induced by buildings within the urban canopy, whereas these fine, transient structures are smoothed out in the coarser ERA5 and WRF fields. Although the high-resolution fields appear noisier, they also contain a great deal of additional, physically meaningful signals. This behavior is consistent with the high-resolution directional roses shown in Appendix A (Figure A4).

3.3. Urban Structure Impact on Wind Speed

Figure 6 maps the over-threshold wind exposure for the northern and southern subregions, defined as the hourly accumulation of the product of Beaufort level exceeding (≥8) and its duration (unit: BFT·h). High values cluster over coasts, river mouths, and open urban edges, whereas values are generally low inside the urban core, yielding pronounced spatial contrasts. In the north, hotspots concentrate along the northeastern coast and peninsulas, with maxima > 100 BFT·h, indicating prolonged extreme winds in wind-facing open sectors. In the south, elevated belts occur along the southwestern coast and open tracts, with patchy peaks (>140 BFT·h) near estuaries and shorelines. In contrast, values in the dense urban core—including compact residential and commercial districts—are typically <20 BFT·h, reflecting strong attenuation of extremes by dense development. This pattern underscores how urban layout and surface types modulate extreme-wind risk. A fuller analysis of total wind exposure (no threshold) is provided in Appendix B (Figure A5).

Figure 7 reveals a robust U-shaped nonlinear relationship between the near-surface wind speed and the open-space ratio, confirmed by highly consistent quadratic fits (R² = 0.91) both before and during the typhoon. The shape reflects regime shifts in the dominant aerodynamic mechanisms across building densities. In open sectors (open-space ratio > 0.8), surface drag is minimal and high-level momentum downmixes efficiently, yielding the highest winds. As openness declines to moderate levels (~0.5–0.7), collective shielding by building arrays becomes most effective, inducing skimming flow over rooftops and creating a wind-suppression band with the lowest speeds. With further densification (<0.2), overall shielding gives way to canyon-jetting (Venturi-like) acceleration in narrow passages; enhanced vertical mixing transports momentum downward, causing a rebound in near-surface speed. The typhoon does not alter this rule but acts as a strong dynamical forcing that uplifts the entire U-curve by ~1.8–2.0 m s⁻¹, with uneven amplification: the suppression band shows relative resilience, while the open-exposed and dense-accelerated ends respond strongly and emerge as risk hotspots. Because the multiscale ERA5–WRF–PALM chain exhibits an island-mean high bias of about 4–5 m s⁻¹ (Section 4), the vertical shift of ~1.8–2.0 m s⁻¹ in the fitted U-curves should be interpreted as a relative intensification under typhoon forcing within the same model configuration, rather than as a bias-corrected estimate of absolute wind-speed change. Partitioned analyses for four representative districts are provided in Appendix B (Figure A6).

Figure 8 documents a similarly U-shaped relation with building height (R² = 0.58–0.59). Typhoon passage elevates the curve by ~1.7 m s⁻¹ without altering the basic shape. The pattern arises from the relative position of the 10 m observation height to building configurations. When buildings are low (<20 m), 10 m sits near or above rooftops, where little blockage is experienced and the boundary-layer flow is directly sampled—hence the higher speeds. For 20–70 m, the 10 m level is fully embedded in the canopy, with skimming flow over roofs and vortical structures within the canopy, forming a sheltered zone with minimum speeds. Beyond ~80 m, very tall buildings stir the flow and induce separation, recirculation, and downwash, re-energizing canopy winds. In Haikou, most buildings fall within the 20–70 m range, implying a natural wind-shielding advantage. Error bars are smallest for mid-heights, consistent with uniform skimming; they grow toward both ends—especially in the super-tall regime—reflecting strong local variability due to canyon jets and downwash. Further diagnostics on district-specific height effects appear in Appendix B (Figure A7); the roles of neighboring building heights and the heterogeneous response of open areas are discussed in Appendix B (Figure A8 and Figure A9).

3.4. Spatiotemporal Evolution

Figure 9 presents Hovmöller latitude–time sections of 10 m winds for four representative districts (Qilou Old Street, Wanlv Park, Haidian Island, Guomao), depicting the spatiotemporal evolution during Typhoon Yagi. The abscissa is UTM Y, the ordinate is hours since start, and color denotes the 10 m wind. All four districts exhibit banded wind-enhancement swaths during the typhoon’s peak (~hours 20–30). Wanlv Park (b) and Haidian Island (c) show broader high-wind zones and longer persistence (>15 m s⁻¹) during the strongest phase, with pronounced maxima over the eastern cross-section or islet segments—evidence that topography and surface type extend the duration and footprint of extremes. The maximum for Qilou Old Street (a) is concentrated its on the eastern side, with a narrow spatial extent and short duration. Guomao (d) features multiple pulses and larger amplitude variability, with a more dispersed spatial distribution—indicating distinct responses between dense urban cores and open/island settings. Haidian Island (c) displays a “high–low–high” pattern across the section, suggesting local enhancement and shielding due to the island shape and coastline effects. Wanlv Park and Guomao illustrate a rise-then-fall temporal evolution with migrating spatial maxima. Overall, the Hovmöller diagrams reveal propagation pathways and zone-dependent persistence: open and island areas sustain long-lived, widespread extremes, whereas dense districts experience short-lived, localized peaks—highlighting the strong modulation of extreme winds by urban surfaces and spatial layout.

4. Discussion

By building an ERA5–WRF–PALM multiscale coupled modeling system, this study reproduced the near-surface wind field over Haikou during the passage of super Typhoon Yagi at an ultra-high resolution of 30 m. Although the system consistently captured the “rise–peak–decay” evolution of wind speed during landfall, the simulations showed a systematic high bias, with island-wide errors of RMSE ≈ 5.3 m s⁻¹ and bias ≈ 4–5 m s⁻¹. This bias does not originate from a single component, but rather from uncertainty accumulation along the entire modeling chain. The root causes originally trace back to the driving data: ERA5 reanalysis exhibits inherent limitations in representing the intensity and fine structure of tropical cyclones—especially during rapid intensification [32]. Subsequently, at the mesoscale WRF stage, the responses of the planetary boundary layer (PBL) and urban canopy schemes to frictional drag and turbulent kinetic energy under strong winds exhibit scheme dependence and systematic errors [33,34].

With respect to parameterizations, typhoon extremes challenge the applicability of existing boundary-layer theory. For example, Monin–Obukhov similarity theory, which underpins most surface-layer schemes, may break down under the highly non-stationary, strongly sheared typhoon boundary layer [35]. In addition, accurate representation of drag is a central difficulty. Numerous studies confirm that the sea-surface drag coefficient saturates or even decreases at very high wind speeds [36,37]; whether an analogous “equivalent drag coefficient” exists over urban canopies under typhoon intensities remains an open question. Current urban parameterizations, largely developed for fair-weather conditions, have uncertain validity in typhoon regimes and may underestimate aggregate urban friction [38]. These errors are then propagated into the microscale PALM model and, via wall functions whose validity is also uncertain at extreme Reynolds numbers, may further amplify simulated wind speeds [39]. Meanwhile, simplifications in the urban representation—static building geometries neglecting vegetation and small-scale structures—constitute additional sources of uncertainty.

There are also intrinsic limitations in observational validation. Owing to typhoon damage, only five automatic weather stations provided complete data, limiting spatial representativeness. Although these stations span shoreline, island, and urban-core settings and suffice for process-level checks, they are insufficient to fully verify the fine-scale spatial patterns produced by PALM (e.g., leeward wakes behind buildings or street-canyon hotspots). Hence, the high-resolution wind maps in this study should be interpreted as high-likelihood hazard scenarios inferred from the best-available physics, rather than as the fully validated ground truth.

Despite these limitations, the revealed U-shaped response of near-surface wind speed to urban surface morphology aligns closely with classical urban-flow theory. Wind speeds are highest in high-openness and low-rise areas due to weaker morphological blockage and more efficient downward transport of typhoon boundary-layer momentum. In moderately tall and dense districts (≈40–70 m), wind speeds exhibit a pronounced minimum consistent with skimming-flow dynamics—flow passes primarily over rooftops, rendering the pedestrian level a relatively sheltered zone [38,40,41,42]. As building height increases further, wind speeds rise again, pointing to the combined actions of canyon-jetting (Venturi-like) acceleration between tall buildings and strong downwash around structures. Under typhoon forcing, these nonlinear effects are substantially amplified [18,22].

These findings have important practical implications for rapidly urbanizing tropical and subtropical coasts. At the urban-planning scale, wind-sensitive design should be implemented [43]. Open coastal and estuarine sectors act as amplifiers of extremes; in transition zones between these areas and dense built-up belts, planners should avoid broad, straight corridors aligned with the prevailing typhoon winds. At the building-design scale, the documented strong heterogeneity of the urban wind field implies that traditional wind-load provisions based on uniform profiles may underestimate site-specific risks; site-specific microscale wind assessments should become a necessary step for critical infrastructure and super-tall buildings. Although this case study focuses on Haikou and Typhoon Yagi, the ERA5–WRF–PALM workflow itself is generic and can be applied to other coastal or inland cities where (i) reliable reanalysis or NWP forcing, (ii) basic near-surface wind observations, and (iii) building-footprint and land-cover data at a tens-of-meters resolution are available. In new applications, the WRF domain setup, land-use fields and PALM static inputs need to be adapted to local conditions, but no fundamental change in the modeling framework is required. Finally, the study’s limitations indicate pathways for improvement: at the provincial level, multi-model ensembles and data assimilation should be advanced and both surface and upper-level wind measurements should be densified, and in model development, urban-canopy parameterizations tailored to typhoon conditions should be pursued and machine learning should be integrated to improve forecasts of extreme winds. From a resilience perspective, the building-resolving wind maps and morphology–wind response curves derived here provide the physical hazard information required by composite resilience indices and local resilience assessments that have been developed for typhoon-prone cities [11,12]. Combined with social and infrastructural indicators, they can help identify districts where extreme winds, exposure and limited recovery capacity coincide, and should therefore be prioritized for wind-resilience interventions in Haikou and similar coastal cities.

5. Conclusions

Drawing on ERA5–WRF–PALM simulations, multi-source observations, spatial statistics, and sensitivity analyses, we summarize the coupling between the temporal evolution, directional structure, and extremes of the near-surface wind field during Typhoon Yagi in Haikou and its links to urban morphology (open-space ratio, building height, and neighborhood structure) as follows:

(1)

Multiscale modeling performance.

ERA5/WRF reproduced the typhoon evolution (R² = 0.49/0.86) but exhibited a systematic high bias of 4–5 m s⁻¹, mainly due to limitations in boundary-layer parameterizations and land–sea roughness treatments. PALM showed higher sensitivity to extremes at strong-wind sites (e.g., M1039, r = 0.77), refining local turbulence features, yet its directional concentration and late-stage decay still require improvement. The multiscale nesting effectively combines mesoscale circulation positioning with microscale refinement, offering a feasible pathway for urban typhoon wind modeling.

(2)

Spatial pattern (“coastal high, urban-core low”).

PALM simulations reveal high wind exposure over coasts, estuaries and peninsulas (Beaufort ≥ 8; 100–400 BFT·h versus < 30 BFT·h in dense cores). NE and SW coasts form continuous high-exposure belts, aligning with wind-facing open terrain and low-density districts. The contrast reflects the combined effects of roughness jumps and ventilation corridors; dense urban fabrics systematically shorten the persistence of extremes, offering quantitative guidance for risk zoning. In practice, these results suggest that coastal estuaries and the transition zones between open coastal sectors and dense built-up belts should be treated as priority areas for wind-risk zoning and wind-resilient urban design in Haikou.

(3)

Open-space ratio: robust U-shape.

The near-surface wind speed exhibits a stable U-shaped relation with the open-space ratio (sample-size-weighted quadratic fits, R² ≈ 0.91/0.91 for pre-/during typhoon). Moderate openness (0.4–0.7) yields the lowest and most stable winds (6–9 m s⁻¹); both low (<0.2) and high (>0.7) openness produce higher winds. Typhoon forcing lifts winds by 1–4 m s⁻¹ overall, with the largest increases at both ends (peaks up to 17.5 m s⁻¹). Responses vary by district: in historic blocks, they fluctuate weakly; in open areas, they amplify strongly; and in island zones, they show homogenization. Moderately open blocks exert damping and stabilizing effects. This provides a quantitative basis for optimizing the arrangement and size of open spaces in typhoon-prone neighborhoods, avoiding layouts that are either too closed or excessively open.

(4)

Building height: U-shaped nonlinearity.

The wind speed decreases then increases with building height (weighted quadratic R² = 0.58/0.59). Low-rise (<20 m) districts average ~14 m s⁻¹ during the typhoon; 40–70 m dense mid-rise districts form a velocity trough (7.5–9 m s⁻¹); and super-tall (≥90 m) areas rebound to 9–10 m s⁻¹. The U-shape reflects differential flow modulation: shielding dominates in dense mid-rise areas, while canyon-jetting intensifies in super-tall clusters. Under typhoon forcing, nonlinearity is amplified, informing layout optimization of high-rise districts. When the prevailing wind aligns parallel to the major street axis, the right branch (high-openness/high-height end) steepens due to stronger acceleration; with perpendicular/oblique incidence, upwind blockage widens the trough and flattens the curve. Thus, the urban wind response depends not only on the static form but also on coupling between the typhoon wind direction and street fabric. The identified U-shaped response of wind to building height indicates that height limits and street orientation in new high-rise clusters should be evaluated with building-resolving wind simulations under dominant typhoon tracks.

(5)

Spatiotemporal evolution.

All four districts exhibit synchronized wind enhancement around hours 20–30, but with distinct modes. Open (Wanlv Park) and island (Haidian) areas experience long-lived, expansive extremes; Haidian peaks approach 30 m s⁻¹ with a “high–low–high” spatial pattern. The historic core (Qilou) has narrow, short-lived high-wind bands, while the modern CBD (Guomao) shows multi-core fluctuations. These contrasts reflect strong terrain and surface modulation of extreme winds and support mitigation strategies differentiated by urban function. In operational terms, this calls for district-specific design standards and emergency planning that reflect the distinct wind-exposure characteristics of each functional zone.

Overall, while the study delineates clear physical regularities, its quantitative outcomes rest on limited station validation data. Broader observations will be required to further verify and refine the fine-scale spatial patterns indicated by the model.