Flood forecasting by means of dynamical downscaling of global NWPs coupling with a hydrologic model at Nong Son-Thanh My River basins

Floods are among the most frequent hydro-meteorological hazards, particularly in Vietnam which is affected by monsoon and tropical climate regimes. Almost every year people in Vietnam suffer floods; many lives were lost each time, and the costs for the damaged and lost properties were considerable (Tran et al. 2009; Navrud et al. 2012; Chau et al. 2015; Chinh et al. 2016; Vu & Ranzi 2017). Hillslope flows initiated by severe precipitation events caused recent floods in Vietnam (Ngo et al. 2020; Satofuka 2020; Trinh et al. 2020). The precipitation preceding such disasters can lead to flood only a few hours following rainfall observation by satellites, radars, or ground stations. This is especially true in regions with steep mountainous terrain that typically have hydropower plants and dams (Soula et al. 1998; Fenn et al. 2005; Schad et al. 2012). Fenn et al. (2005) found evidence of a flood event rising to a peak in 1.5 h, with out-of-bank flows lasting around 5 h in Boscastle in the UK. According to Soula et al. (1998), floods can reach maximum activity in only 3.5 h following the first observation of rain cells by a ground station and S-Band radar. In Vietnam, Schad et al. (2012) reported that the water level of the Chieng Khoi reservoir rose at the rapid rate of 2.5 cm h⁻¹ and exceeded the critical level of 11.9 m, initiating overflow of the spillover after several hours.

Traditional flood detection and forecasting are commonly based on the information of rainfall and other variables (e.g. river discharges, water levels, flood stages) either measured at certain locations of the watershed and river reaches or simulated by hydrological-hydraulic models (e.g. de Roo et al. 2003; Thielen et al. 2009). These methods provide forecast lead times that vary according to the size of the basin and are relatively short for such small, steep catchments because the lead times are basically determined based on the lag time between rainfall and runoff concentration (Werner et al. 2005). They allow a short period of time for people in downstream areas to take precautionary actions and for local decision-makers to prepare for potential evacuation or other measures. In light of these observations, flood forecasting and early warning systems are needed to provide information on the starting time, duration, and magnitude of possible floods with longer lead time in the future.

An accurate and reliable forecast and warning system could enhance and improve the effectiveness of emergency responses (Nam et al. 2014; Chen et al. 2015; Hussain et al. 2021). Nevertheless, the rapid response of flood following the observation of rainfall limits the amount of lead time available for the implementation of emergency management measures such as the evacuation of potential flooding areas. The improvement in lead time for the runoff-river level forecasts is dependent on the ability to gain information ahead of the runoff event's occurrence, which in turn, can improve the amount of lead time that is available for emergency management measures. The knowledge of future rainfall and other climatic conditions at different lead times, ranging from the next few hours to the coming 2–3 days, is fundamental to providing the necessary lead times for flood forecasting. Appropriate lead times cannot be covered by observations of rainfall only after it has fallen to the ground. When very short forecast horizons on the order of 3 h or less are beneficial, extrapolative, and trend-based meteorological techniques are used. The use of such techniques is known as ‘nowcasting’, resulting in short-range quantitative precipitation forecasting (Barr et al. 2020; Shehu & Haberlandt 2021). Beyond this 3-h horizon, numerical weather prediction (NWP) models have to be used for quantitative precipitation forecasting. Currently, atmospheric models can provide forecast results with a lead time of more than 10 days. Despite this, a 1–3-day forecast is recommended due to the larger uncertainty associated with longer lead time forecasting (Adamowski 2008). NWP models are widely researched and developed in many places including the United States, Japan, and Europe. Along with the rapid development of science and technology in recent years, especially in the field of information technology and computer science, the level of detail and the creation of global NWP models has been increasing. Among the many available models, the GFS (global forecasting system) of America and the GSM (global spectral model) of Japan stand out. GFS is the global spectrum model of the US National Centers for Environmental Predictions (NCEP). GFS began its operational use at the National Meteorological Center, the predecessor of NCEP until 1988. The model is regularly improved and upgraded. To date, the model has had two configurations including one with a horizontal resolution of 35 km and 64 vertical layers (for forecasts of 7.5 days ∼180 h), and one with a horizontal resolution of 70 km and 64 vertical layers for forecasts of about 16 days ∼ 360 h (Han et al. 2017). The GSM, developed by the Japan Meteorological Agency (JMA), employs primitive equations to express the resolvable motions and states of the atmosphere. The model originally had a horizontal resolution of T63 and 16 vertical layers up to 10 hPa with a sigma-coordinate (Mizuta et al. 2006; Yamaguchi et al. 2012). Multiple upgrades and developments have led to improvements such as a spectral resolution of TL959, a highest altitude of 0.01 hPa, and assimilated 4D-Var with 100 vertical layers together with common and vertical initialization processes. Currently, GSM provides 84-h forecasting time for operations at 00, 06, 12, and 18. GSM is currently one of the highest-resolution global models (about 20 km). However, free versions are only available at ∼ 60 km resolution, which is equivalent to the resolution of GFS (JMA 2013). These two NWP datasets can be downloaded directly from their associated organizations.

In this study, the dynamical downscaling technique is coupled with a hydrologic model for the detection and forecast of real-time flood events with a lead time ranging from 1 to 3 days. The objectives of this study are: (1) proposing a new approach to forecast real-time flood events and longer lead time; (2) testing the flood forecasting system over Thanh My and Nong Son watersheds in central Vietnam. This approach is demonstrated by applying regional NWP and physically based hydrologic models to the Thanh My and Nong Son watersheds in central Vietnam with system inputs provided from different global NWPs (GFS and GSM). The present study seeks to overcome the research challenge of flood detection and forecasting and help improve the flood warning and forecasting practices in Vietnam.

River tributaries in central Vietnam originate from elevated mountain ranges, lying along the border between Vietnam and Laos. They flow through narrow floodplains and finally empty into the East Sea of Vietnam. These rivers are very short in length and have tributaries with steep bed slopes. Rapid changes in land use for agricultural expansion and economic development have enhanced the acceleration of catchment responses. About 72% of the study region is mountainous or hilly and mainly covered by forested areas (Chau et al. 2015). The region has often experienced large rainfall and flooding during the wet seasons, from September to December, accounting for 85% of the average annual water availability. The dry season, from January to August, accounts for 15% of average annual water availability. River flows measured in floodplains in the region show that six out of the seven major floods in the last 50 years occurred in the period from 1995 to 2010 (Nam et al. 2015). Proposing appropriate technologies for flood prediction with greater forecast horizons is indeed essential to reduce the pressure on the existing flood defense system and assist in better preparation for flood damage reductions.

There are four main components forming the flood forecast system. The first component is retrieving and decoding data from global NWPs; the second is the dynamical downscaling of global NWPs outputs by means of a regional-scale model; the third is merging the atmospheric output into the hydrological model; the fourth is the implementation of the hydrological forecast models; and the last is flood warning platform that continuously provides information of atmospheric and flood conditions with a lead time of 1–3 days for selected locations. It is noted that this system automatically updates forecasting every 12 h based on retrieval of GFS and GSM atmospheric data.

Due to its successful application in Vietnam, the Weather Research and Forecasting (WRF) model is recommended for application in this study despite the availability of similar mesoscale NWP models (Raghavan et al. 2012; Ho et al. 2019; Trinh et al. 2020, 2021). Since its first release in 2000, the WRF model has performed a wide range of meteorological applications across scales from tens of meters to thousands of kilometers (Skamarock et al. 2005). With respect to real-time operational weather forecasting, the WRF model offers a flexible and computationally efficient platform, while offering advances in physics, numerical methods, and data assimilation, as contributed by many research community developers. As a result, the WRF model is currently in operational use at NCEP and other meteorological modeling centers around the world.

The mesoscale WRF model setup for Thanh My and Nong Son watersheds employed the outputs from the GFS and GSM models. These outputs were used as the initial and boundary conditions of the WRF model. In this study, the WRF model was configured with three nested domains, as shown in Figure 1. The outer domain (D1) covers the whole of Vietnam, Cambodia, Laos, Thailand and parts of China, and the Philippines, having a spatial resolution of 54 km (39 × 49 horizontal grid points), similar to the spatial resolution of the boundary datasets. The domain (D2) in between the outer and inner domains covers central Vietnam with a resolution of 18 km (45 × 60 horizontal grid points). The innermost and smallest domain (D3) covers the study area with a spatial resolution of 6 km (60 × 51 horizontal grid points).

A key step in setting up the mesoscale WRF model was the selection of parameterization schemes in the model configuration. An optimal physical parameterization scheme is crucial for extreme rainfall prediction. There are available parameterization schemes which have been implemented for various weather patterns in Vietnam (Raghavan et al. 2012; Minh et al. 2018; Ho et al. 2019; Trinh et al. 2021). In this study, the model configuration was examined with 17 parameterization schemes as suggested in the literature involving nine different microphysics schemes, six cumulus parameterizations, five planetary boundary layers schemes, and four radiation physics parameterizations, as presented in Table 1. In addition, the WRF model provides forecast atmospheric data hourly with 18 meteorological levels.

Dynamical downscaling is applied to generate a high-resolution quantitative rainfall forecast over the targeted watersheds. Thus, the mesoscale WRF model calibration and validation were performed against observed rainfall data. The observation data were derived from two different sources, the Vietnam Gridded Precipitation (VnGP) and observations from five rain gauges. VnGP was developed based on daily rainfall observation across Vietnam from 1980 to 2010 with a resolution of 0.1° (Nguyen-Xuan et al. 2016). Rain gauge observations in the area have been available since 1979. In the present study, WRF-simulated daily basin average precipitation was compared to observation data.

Model calibration was initiated through the selection of an optimal parameterization scheme from the 17 considered. Model performance was assessed and evaluated for the downscaled variables of the D3 domain using the Nash–Sutcliffe efficiency (NSE). Two extreme events were selected for model calibration and validation. Calibration was performed with an event from 1 October to 31 October 2008. An event from 24 October to 10 November 2009 was employed for model validation. These two events were selected due to their impacts to the study region as mentioned in Nam et al. (2011, 2014). Table 2 shows model performance statistics for the downscaled rainfall at domain D3 from two global datasets examined with 17 parameterization schemes and 1-, 2-, and 3-day forecast lead times. The overall score exhibits that GSM-based downscaling results outperform those of the GFS is global forecasting system (GFS), especially the forecasts with 1-day lead time. The statistics also indicate that using the same parameterization scheme results in significant variance for the selected global NWP datasets. It is found that the optimal parameterization scheme for the GFS dataset consists of WSM3 (Hong et al. 2004) as the microphysics processes, Grell–Freitas (Fowler et al. 2016) as cumulus parameterization, BouLac (Bougeault & Lacarrere 1989) as the planetary boundary layer, and New Goddard (Chou & Suarez 1999) as the radiation physics. Meanwhile, the parameterization scheme selected for the GSM dataset consists of Goddard (Tao & Simpson 1993) as the microphysics processes, Tiedtke (Tiedtke 1989) as the cumulus parameterization, BouLac (Bougeault & Lacarrere 1989) as the planetary boundary layer, and Dudhia (Dudhia 1989) as the radiation physics. These model configurations are summarized in Table 3.

It is noted that this result is a real-time precipitation forecast with a 1- to 3-day lead time, and the precipitation forecast updated the initial and boundary conditions every 12 h. The boundary condition was continuously updated using retrieved GFS and GSM forecast data, while the initial condition was updated by the previous time step result. Therefore, the calibration and validation are encouraging and the selected configurations are reliable for further forecast applications including atmospheric and flood forecasts.

The Song Tranh 2 and Dak Mi 4 dams were constructed in 2006 and 2007 and started operation in 2010 and 2012, respectively. Song Tranh 2's dam and power plant are located in the main stream of the Nong Son watershed, while Dak Mi 4's main dam is in the Thanh My watershed and its power plant is located in the Nong Son watershed. Dak Mi 4's dam is roughly 3 km from its power plant. Summary information about the two dams is provided in Table 5.

In this study, existing operating rules for the Song Tranh 2 and Dak Mi 4 reservoirs are applied. The dam operation subprogram requires customization for each of the main reservoirs in the selected watersheds (Song Tranh 2, and Dak Mi 4 Reservoirs). The operation rules were initially implemented following available documentation but had to be further refined by model simulations. Representation of the operation rules for these dams is crucial to accurately simulate flow conditions using the WEHY model.

Following the successful implementation and validation of the atmospheric and hydrologic components of the proposed system, the real-time forecasting system is tested using several extreme events that were reported in Loi et al. (2019) and Nguyen et al. (2021). The selected area covers complex topography, climate, and land use activities. It is subject to heavy rainfall and frequent flooding, with an average of three to five flood events per year. Flood events often occur consecutively in a short time. The rise and fall of floods are quick in the upper and middle areas but slow in the downstream areas. According to Luu & von Meding (2018), historic floods recorded during the 19 years from 1997 to 2015 left 726 dead and resulted in economic damage of 614.6 million USD in the downstream areas. Four main flood events, occurring between 2012 and 2015, were selected to test the real-time forecasting system. They were selected based on the availability of historical flood records and atmospheric data, and the timing of the construction of the two main dams.

It is noted that ΔW% and Δpeak% are volume and peak error, calculated as a percentage as given by the following equations.

WEHY-WRF-GSM showed higher volume and peak errors, particularly in the 2012 and 2013 events. Despite these errors, the forecasting results are encouraging due to their acceptable correlation coefficients for most of the events. The reliability and accuracy of these forecasting results can be improved by applying assimilation and bias correction methods, particularly in the selected locations (Trinh et al. 2022b). For the reliability of the 2–3 days precipitation and flood forecasts, it is necessary to update these forecasts by upper air sounding observations, ground weather observations, remote-sensed observations (radar or satellite), and real-time ground precipitation and flow gauge observations.

This study proposed a forecasting system that coupled a dynamical downscaling technique and hydrologic models to forecast real-time flood events with a lead time ranging from 1 to 3 days. The WEHY-WRF models were selected as the hydrologic-atmospheric components for the proposed system. The WEHY-WRF models were successfully implemented and validated before testing a real-time forecasting system over the Nong Son-Thanh My watershed. There are four significant historical flood records selected for evaluation of the real-time rainfall-flood testing system. Overall, the comparison between the model simulations and corresponding observations show the rainfall and flood of the WEHY-WRF-GFS forecast match quite well with observation data. The rainfall and flood forecast results based on WEHY-WRF-GSM did not match as well with the historical record as the WEHY-WRF-GFS results; however, these results are quite encouraging based on their acceptable correlation coefficients and encouraging NSE. Some limitations remain in the proposed approach, including missing bias correction and data assimilation. In the future, improvements can be made through data assimilation of rainfall and flood forecast, which will merge the WEHY-WRF prediction outputs with the observed hydrometric, automatic hydro-meteorological data, and real-time quantitative radar data. In addition, future work includes developing an automatically updating forecast every 6 h and continuously providing information on atmospheric, and flood with a lead time of 2–3 days (depending on requirements as well as forecasting objectives) for selected locations. The flood forecast also can be coupled with the hydraulic-sedimentation models to estimate flood inundation and sediment transport during real-time extreme flood events (Tu et al. 2020). The flood forecasting and warning system could no doubt assist management authorities during flooding events. An accurate and reliable forecast could enhance and improve the effectiveness of emergency responses. In addition, reliable forecast data will assist with decision-making focused on reducing the effects of flood risks on society in an economically and environmentally sustainable manner.

This research was funded by the Vietnam National Foundation for Science and Technology Development (NAFOSTED), Ministry of Science and Technology, under grant number 105.06-2019.326.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.