Historical rainstorm in Hong Kong on 7–8 September 2023: Diagnosis, forecasting and nowcasting

Abstract

On 7–8 September 2023, Hong Kong was hit by a historical and record‐breaking rainstorm associated with the remnant of Tropical Cyclone Haikui (2311). The hourly rainfall recorded at the Hong Kong Observatory Headquarters once reached 158.1 mm, the highest since record began in 1884. The 24‐h rainfall even exceeded 600 mm in some parts of the territory. The historical rainstorm resulted in heavy flooding and landslides, bringing significant societal impact to Hong Kong. This paper aims to review this unprecedented heavy rain event from the aspects of diagnosis, forecasting and nowcasting. Early indicators of such events over Hong Kong with substantial lead time are limited from the dynamics and thermodynamics consideration, the numerical weather prediction models, given the present technology. The only indication may come from the climatologically extreme total precipitable water. While recent research of developing a regional risk‐based alerting system on the higher impact event of flooding associated with heavy rain might have potential to enhance the weather service, and emerging AI model showed some promising post‐simulations, predicting historical and record‐breaking rainstorms remains a challenge for operational weather forecasting and warning services.

INTRODUCTION

Hong Kong is often affected by rainstorms every summer (Wu et al., 2015). To warn the public about heavy rain, and to ensure readiness within the essential services to deal with emergencies, the Hong Kong Observatory (HKO) has operated the color-coded rainstorm warning system, namely Amber, Red and Black Signals, to signify at the hourly rainfall thresholds of 30, 50 and 70 mm, respectively (Lam, 2004). In the overnight period of 7–8 September 2023, a historical and record-breaking rainstorm occurred over Hong Kong. The hourly rainfall recorded at HKO Headquarters once reached 158.1 mm, the highest since record began in 1884. The 24-h rainfall exceeded 600 mm in some parts of the territory (Figure S1, Supporting Information). The prolonged torrential rain to Hong Kong necessitated the issuance of the Black Rainstorm Warning Signal for 16 h and 35 min, setting the longest record since the introduction of Hong Kong's rainstorm warning system in 1992.

This paper aims at documenting this exceptionally heavy rain event. A diagnosis of the dynamics and thermodynamics aspect is reviewed. Forecasting aspect is examined from the longer-term perspective by numerical weather prediction (NWP) models and artificial intelligence (AI) model, and nowcasting timescale for several hours ahead from a couple of algorithms.

For the occurrence of heavy rain in Hong Kong, a study by Lee (1993) has identified a set of favorable meteorological factors, including surface and upper-air synoptic patterns, thermal and dynamical mechanisms as well as mesoscale features. This paper will first focus on the diagnosis aspect, examining factors favorable for the occurrence of heavy rain. Even from a retrospective view point for this event, an early assessment of the occurrence of historical rainstorm may be limited, except from the consideration of precipitable water vapor. In the forecasting side, signals from NWP models were mixed, generally not able to capture the occurrence of historical rainstorm. There are some skills in the nowcasting tools, but their lead time may not be long enough for earlier preparation of the prevention of flooding and landslides.

SYNOPTIC PATTERN

The major triggering system for the heavy rain is the remnant of Tropical Cyclone Haikui (2311), which made landfall over the eastern coast of Guangdong in the morning of 5 September 2023 and continued heading west over the inland areas while quickly weakening and dissipated. This remnant continued to move across Guangdong slowly (Figure S2), resulting in prolonged rainstorm from late 7 September to 8 September in Hong Kong. The general overall of the synoptic pattern is given in Figure S3. The slow westwards movement of the remnant was driven by the easterlies along the southern flank of the mid-level anticyclonic circulation. Meanwhile, cyclonic circulations associated with the remnant could be found in surface isobaric chart, and further up on 850 and 500 hPa streamline analysis, as in Figures S3b and S3c. There was also an upper-level anticyclone, bringing significant divergence on 200 hPa (Figure S3a).

Another factor that might contribute towards the heavy rainstorm, though may not be readily established and quantified to have significant contribution, was the arrival and prevalence of the weak northeast monsoon over southern China, as in Figure S3d. The weak east to northeasterly flow, together with the warm and moist south to southwesterly winds to the south of the remnant of Haikui, resulted in a surface trough (Figure S3d) and may enhance the surface convergence. One point not so certain is whether the cooler east to northeasterly flow may help elevate the warm and moist air from the south, leading to persistent heavy rain, in the form of a weak frontal system. This is not yet confirmed in high-resolution numerical simulation of the event and further numerical study by the research community would be needed.

The synoptic pattern suggested that heavy rain could be expected over the coast of southern China. However, such synoptic features may be found in other typical rainy cases in the region as well (Lee, 1993). It is not possible from merely the synoptic pattern itself to judge that a historical rainstorm is going to happen.

OBSERVATION FROM WEATHER RADAR

Quasi-stationary south–north oriented rainbands persisted over Hong Kong for an extended period of time, thus leaded to large accumulated rainfall. From the sequence of weather radar pictures from 7 to 8 September (Figure 1), the intense rainbands continued to develop and persistently affected Hong Kong. Besides, it seems that there is confluent flow right over the Pearl River Estuary, as indicated from the arrows showing the movements of the individual rainbands.

[IMAGE OMITTED. SEE PDF]

Converging flow is indeed observed from the surface and lower troposphere. From the surface observations (Figure S4a), there is clear signature of the converging easterly airstream and southeasterly airstream near Hong Kong. At 850 hPa (Figure S4b), wind profiler measurements clearly identified the remnant of Haikui, located to the northwest of Hong Kong, which persisted throughout the whole heavy rain episode. This feature is also apparent from the cyclonic movement of weather radar echoes in that region. Southwesterly flow is suggested over the western part of Hong Kong from the winds and the orientation of the weather radar echoes.

DYNAMICS CONSIDERATION

A sequence of the surface winds over the south China coastal waters is provided by satellite scatterometer (Figa-Saldaña et al., 2002), in Figure S5. On 7 September (Figure S5a), the convergence of wind is rather clear, and it is consistent with other surface observations over southern China. A southwesterly jet established to the south of Hong Kong on the following morning (Figure S5b). The low-level jet is expected to play a major role in triggering heavy rain. It also shows up clearly from the local wind profiler measurement (Figure S6). However, the winds are not particularly strong, only reaching about 20 knots below 1500 m above mean sea level.

From dynamics consideration, some factors favoring heavy rain could be identified, for example, converging flow and low-level jet. However, these factors are rather common in the heavy rain occurring in southern China in the season, and their occurrence could not be relied upon for an early assessment or alerting of the occurrence of historical rainstorm.

THERMODYNAMICS AND WATER VAPOR CONSIDERATIONS

The total precipitable water (TPW) product from the geostationary satellite GEO-KOMPSAT-2A (GK2A) (Kim et al., 2021) provides some new angles on the heavy rain occurrence. As seen from a sequence of the TPW images in Figure 2, a region of high TPW over 70 mm (light purple in Figure 2a) first occurs to the east of Hong Kong. It then moves across Hong Kong on 7–8 September. Eventually it reaches the region to the west of Hong Kong and becomes disorganized.

[IMAGE OMITTED. SEE PDF]

This remote sensing product is only available recently at HKO. Its accuracy is supported from the radiosonde observations in Hong Kong. In fact, at the 0000 UTC (i.e., 0800 Hong Kong Time, HKT = UTC + 8 h) of 7 September radiosonde, TPW already exceeds 70 mm. This situation persists through next morning (Table 1). Such a level of TPW is extremely high and reach the 99.9th percentile from the climatological statistics of radiosondes in Hong Kong from 2001 to 2020. This feature could be observed as early as the morning of 7 September. For this historical rainstorm event, high TPW appears to be a major and the only factor that suggests exceptionally heavy rain is on the way, though the actual timing and the rainfall amount are not known, which is an important information for an early alert.

TABLE 1 Various quantities derived from the 0000 UTC (i.e., 0800 HKT), 8 September 2023 radiosonde observation at King's Park (45004), Hong Kong.

Quantities	Values	Percentile
Total precipitable water	78.2 mm	99.9
500 hPa relative humidity	100%	98.7
700 hPa relative humidity	100%	98.6
925 hPa relative humidity	93.0%	75.7
1 km surface wind shear	8.8 × 10−3 s−1	67.6
3 km surface wind shear	3.1 × 10−3 s−1	83.0
Storm relative helicity	94.2 m2·s−2	83.8
850 hPa wind speed	12.5 m·s−1	88.5
925 hPa dew point	21.7°C	87.8
Surface dew point	24.9°C	76.4
Convective available potential energy	490 J·kg−1	42.7
Lifted index	−1.9	54.9
K index	39.8	97.2

For the other factors obtained from the radiosonde sounding (Table 1), for instance, dynamical factors (such as vertical wind shear, storm relative helicity and low-level wind speed) and thermodynamic factors (such as relative humidity, dew points and lifted index) are only ranked moderate in comparison with climatology. The convective available potential energy is also not particularly high. They provided no much signal on the occurrence of this historical rainstorm.

FORECASTING FROM NUMERICAL WEATHER PREDICTION MODELS AND NOVEL AI MODEL

Daily rainfall patterns near Hong Kong for 8 September as forecasted by various numerical weather prediction (NWP) models are given in Figure S7a. The actual rainfall pattern in the region over that period of 24 h is given in Figure S7b, with the region over 100 mm rainfall enclosed in red. The major NWP models failed to provide an indication of heavy rain over Hong Kong, not to say historical rainstorm. The regional model Tropical Regional Atmospheric Modelling System (TRAMS; Zhong et al., 2020), at the rightmost column of Figure S7a, is the only exception suggesting heavy rain would be in excess of 100 mm in Hong Kong. However, this model is rather new and its performance in heavy rain events is yet to be fully assessed.

We further experimented Weather Research and Forecasting (WRF) model simulations of rainfall for this rainstorm event with different configurations of initial conditions, as described in Table 2. Figure 3 shows the 24-h accumulated precipitation over the Pearl River Estuary by the different WRF simulations, together with the rainfall estimated by Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG). Note that those WRF simulations initialized by NCEP operational Global Forecast System (GFS) data or NCEP Final operational global analysis (FNL) (Figure 3a–c) forecasted similar rainfall patterns and failed to predict the intense rainfall over Hong Kong. While those WRF runs using ERA5 as initial conditions (Figure 3d–f) shared a similar forecast with an east–west oriented band of rainfall along the coast line. Unfortunately, all six WRF runs were not able to capture the rainstorm event in and near Hong Kong (Figure 3g). The unsatisfactory WRF simulations could be attributed to the larger-scale initial conditions. It is evident from Figure S7a that both ECMWF and NECP fail to predict the rainstorm in Hong Kong.

TABLE 2 Description of six WRF simulations.

WRF simulation	Initial condition
WRF_GFS	NCEP operational Global Forecast System (GFS) 0.25° × 0.25° grid data
WRF_GFS + GNSS	Forecast of WRF_GFS's at 1000 UTC (1800 HKT) on 7 September 2023 used as the first guess, and TPW data observed at 1000 UTC from 18 Global Navigation Satellite System (GNSS) stations in Hong Kong assimilated
WRF_FNL	NCEP Final (FNL) operational global analysis 0.25° × 0.25° grid data
WRF_ERA5	ERA5 0.25° × 0.25° grid data
WRF_ERA5_W	ERA5 0.25° × 0.25° grid data, but low-level wind speed at 1000–850 hPa increased by 100%
WRF_ERA5_WR	ERA5 0.25° × 0.25° grid data, but low-level wind speed at 1000–850 hPa increased by 100%, and relative humidity at 1000–850 hPa increased by 20%

[IMAGE OMITTED. SEE PDF]

A comparison between Figure 3a,b clearly showed the contribution of assimilating GNSS TPW data. While all TPW data assimilated were located in Hong Kong, the simulated rainfall amounts for inland Guangdong (to the north and west of Hong Kong) increased considerably. This demonstrated that the assimilation impact can be propagated to the areas far away from Hong Kong. Statistically, the area-averaged 24-h accumulated rainfall amount increased from 8.9 to 28.7 mm upon assimilation of TPW. The point maximum 24-h accumulated rainfall also increased from 214.9 mm (at 23.1°N, 112.9°E in Figure 3a) to 396.7 mm (at 22.8°N, 112.8°E in Figure 3b).

Furthermore, the sensitivity of the WRF model's rainfall forecast to low-level wind speed and moisture content is experimented (Figure 3d–f). Modifying either the low-level wind speed or both the wind speed and relative humidity resulted in similar rainfall patterns and amounts. These findings reveal that the WRF experiments, even with enhanced low-level wind speed and moisture contents, is not able to capture the extreme rainfall during this rainstorm event.

Rainfall forecasts from a version of the WRF ARW (Advanced Research WRF) model (Skamarock et al., 2008) implemented at the Central Weather Administration (CWA) in Taiwan was also examine after the event, a system known as TWRF (Chen et al., 2021; Hsiao et al., 2010, 2012, 2015, 2020). On forecasting this record-breaking rain storm, TWRF initialized at 0000 UTC of 6 September gave promising results as showed in Figure 4. TWRF was able to predict the west movement of the low-level vortex associated with the remnant of Haikui (Figure 4a–d), and in general the associated rain bands. The total precipitation forecasted for the 12-h window ending at 0000 UTC (i.e., 0800 HKT) of 8 September (Figure 4c), which was the peak rainfall period, generally aligned with the actual observations, although the heaviest rainfall area was slightly off to the west of Hong Kong. The predicted maximum accumulative rainfall by this run of TWRF reached an extreme value of 508 mm during the 12-h window (Figure 4c). This prediction could better alert operational forecasters about 1 day before regarding the possibility of extreme rainfall in the vicinity of Hong Kong. As a case study, the capability of TRAMS and TWRF reconfirm the value of high-resolution regional NWP models in terms of heavy rain forecast.

[IMAGE OMITTED. SEE PDF]

Post simulation using AI models, such as a new internal preview version of Pangu-Weather (Bi et al., 2023) which has been enhanced to higher spatial resolution of 0.1 degrees and trained to forecast 6-hourly precipitation as well, have also been studied. The results are given in Figure S8a–f, where Pangu-Weather was initialized with the operational analysis of ECMWF. The Pangu-Weather model at multiple consecutive runs suggests that there could be heavy rain in the vicinity of Hong Kong and more realistically predicts that those heavier rainfall occurring over the seas south of Hong Kong. It generally outperformed those traditional global NWP models as shown in Figure S7a. Moreover, some of its runs could predict 24-h accumulated rainfall more than 200 mm near Hong Kong, a few days ahead of the event (e.g., Figure S8e as initialized at 0000 UTC of 5 September 2023). Nevertheless, the forecast rainfall is still much weaker than the actual maximum rainfall amount and its rainfall pattern is still not fully consistent with the actual observations (Figure S7b). Given that Pangu-Weather is a global model with a much longer lead time than most regional models, and more AI models have been emerging recently, it is hope that more research inputs on a combination of AI global models and higher-resolution regional models could help tackle the long-lasting issue on heavy rain prediction by NWP.

NOWCASTING: CONVENTIONAL, OPERATIONAL METHOD

HKO uses optical flow method in the operational nowcasting system, namely “Real-time Optical flow by Variational methods for Echoes of Radar” (ROVER-A), for heavy rain (Woo & Wong, 2017). While early indications of heavy rain over Hong Kong a few hours ahead are not possible, ROVER-A could provide some indications of rather heavy rain over various parts of Hong Kong as time gets closer to the onset of historical rainstorm. An example is the time instance of 1330 UTC (i.e., 2130 HKT), 7 September 2023, the predicted hourly rainfall amount and pattern as shown in Figure S9a are quite consistent with the actual observations (Figure S9b), though the lead time is rather limited. Woo and Wong (2017) also demonstrated that the critical success index (CSI) (Schaefer, 1990) for predicting heavy rain of 30 mm·h⁻¹ or above by ROVER-A is around 20% for 1-h lead time, and then dropped below 10% for forecast lead time of more than 2 h.

For accumulated rainfall amount, an example is given in Figure S9c, namely, 6-h accumulation rainfall as forecast at 1430 UTC (i.e., 2230 HKT), 7 September 2023. Over 300 mm of rainfall is forecast over the central part of Hong Kong, and the forecast rainfall pattern and amount are rather consistent with the actual (Figure S9d). However, as in the hourly rainfall case, the lead time is limited.

NOWCASTING: AI METHOD

The NowcastNet AI model, developed by Zhang et al. (2023) for predictions of extreme precipitation events, has also been studied. The model takes nine radar images (covering a span of 90 min) as inputs, and generates 18 rainfall prediction images (spanning 3 h) at a 10-min interval. To evaluate the performance of the AI nowcasting model, we utilized radar images collected by HKO. The radar station is located at 22.358°N and 114.218°E. And each radar image has 960 × 960 pixels in latitude and longitude, each pixel covering an area of 533.33 m × 533.33 m. An example of radar images at 15:00:20 UTC (i.e., 23:00:20 HKT) on 7 September 2023 is shown in Figure S10.

The AI nowcasting model was evaluated in terms of area-mean precipitation, CSI and cumulative distribution function (CDF) over the Hong Kong region (black rectangle in Figure S10) at different starting times at 1300 UTC (i.e., 2100 HKT) (Figure 5a–c), 1400 UTC (i.e., 2200 HKT) (Figures 5d–f) and 1500 UTC (i.e., 2300 HKT) (Figure 5g–i). In terms of the forecasted area-mean precipitation rate starting at 1300 UTC (i.e., 2100 HKT) (Figure 5a), the AI model only managed to predict the first 30 min, whereas significant deviation was observed between the forecasted and actual conditions after T + 30 min, with the discrepancy becoming more pronounced by T + 60 min (i.e., around 1400 UTC (i.e., 2200 HKT)). Later forecasts starting at 1400 UTC (i.e., 2200 HKT) and 1500 UTC (i.e., 2300 HKT) (Figure 5d–g) had progressively better agreement with observations up to about T + 90 min, but those forecasts would have arrived too late for adequate warnings. The CSI of the AI nowcasting model showed a rapid decrease after the first hour. The CSI values were higher for the rain rate threshold of 8 mm·h⁻¹ compared to those for 16 and 32 mm·h⁻¹. For forecast starting at 1300 UTC (i.e., 2100 HKT), the AI model overestimates the low rain rates (0–40 mm·h⁻¹) and underestimate for the high rain rates. Later forecasts could better capture those lower rain rates yet still underestimated those rain rate on the very high end.

[IMAGE OMITTED. SEE PDF]

EXPERIMENTAL REGIONAL ALERTS FOR FLOOD RISKS ASSOCIATED WITH HEAVY RAIN

Tam et al. (2023) recently proposed an experimental risk-based alerting system that predicts flood risks associated with heavy rain over five main regions (Figure S11a) in Hong Kong based on regional statistical relationships obtained from historical flood reports which were classified as high and medium impact events and the corresponding rainfall intensity. The framework essentially follows the impact-based forecasting approach promoted by the World Meteorological Organization (WMO, 2015). The color-coded risk matrix in Figure S11b is constructed in terms of impact levels and occurrence likelihood. Predicting impact level is then transformed to predicting the corresponding rainfall intensity based on the statistical relationships, and probabilistic rainfall nowcast was the major input to assess the flood risks.

The table in Figure S12 summarizes predicted regional risk level (either medium or high risks) evolution for the alert issuance, from the evening on 7 September when the intense rainband gradually developed at the vicinity of the territory until small hours on 8 September under the influence of widespread heavy rain. Medium risk or even high risk were first attained in regions I, III and IV at around 1315 UTC (i.e., 2115 HKT). Later, more regions attained persistent high-risk signals before 1500 UTC (i.e., 2300 HKT) with high-impact flooding events reported over these regions. Regional flood risk alerts could potentially give advanced signal for the occurrence of higher-impact events in regional scale, and even an earlier indication for the possibility of territory-wide medium to high-impact events as compared to the actual issuance of Red and Black Rainstorm Warnings at 1350 UTC (i.e., 2150 HKT) and 1505 UTC (i.e., 2305 HKT), respectively.

CONCLUSIONS

The historical rainstorm in Hong Kong on 7–8 September 2023 has significant societal impact on Hong Kong. It would be critical to provide an earlier indication of its occurrence, preferably far ahead of the nowcasting time scale. The extensive types of observations taken from surface weather stations to radar and satellite-based measurements shown in the previous sections match with the conceptual model of heavy rain formation (Figure S13), that is, the aligning setting of surface convergence, low- to mid-level cyclonic flow and significant upper divergence. The WRF model simulations also demonstrate the significant contribution of TPW to the rainfall over Guangdong region, despite the spatial distribution of simulated rainfall deviated from the actual possibly due to the limitations from its larger-scale initial conditions. However, from the dynamics and thermodynamics consideration, and the current generation NWP models' forecast, such an early indication of exceptionally heavy rain may not be possible given the present technology. The only indication may come from the climatologically extreme TPW, but this single piece of information is far from sufficient to give operational forecasters the confidence of earlier forecasting and early warning of extremely heavy rain event. The forecasting of heavy rain is still a very challenging issue for operational weather forecasting and warning services.

On the nowcasting time scale, recent research efforts on developing a regional risk-based alerting system on the higher-impact event of flooding associated with heavy rain shines a light of hope to enhance the weather service with an earlier indication for the possibility of a high-impact event. Another major lesson to learn from this case is the application of AI models, where on time scale of nowcasting for a couple of hours ahead, they do not appear to perform much better than conventional methods. At most both the AI and the conventional nowcasting methods are on par with similar forecasting skills. Meanwhile, an emerging AI model on medium-range forecasting seems to outperform traditional NWP models. High-resolution regional models could also hint for a very heavy rain event nearby. It is hoped that the present paper could stimulate further research to improve the rainfall forecast, in particular this historical record-breaking case, which form the basis to provide better support on early warning with a longer lead time.

AUTHOR CONTRIBUTIONS

Hiu Ching Tam: Investigation. Yu-Heng He: Formal analysis. Pak Wai Chan: Conceptualization; formal analysis; investigation. Shiwei Yu: Formal analysis. Hui Su: Formal analysis; writing – original draft. Huisi Mo: Investigation. Ling-Feng Hsiao: Formal analysis; investigation; methodology. Yangzhao Gong: Investigation.

ACKNOWLEDGEMENTS

We thank Huawei Cloud for offering an internal preview version of Pangu-Weather model, which comes with precipitation output, to post-simulate this rainstorm case.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

References

Bi, K., Xie, L., Zhang, H., Chen, X., Gu, X. & Tian, Q. (2023) Accurate medium‐range global weather forecasting with 3D neural networks. Nature, 619, 533–538. Available from: [DOI: https://dx.doi.org/10.1038/s41586-023-06185-3]