Creating precise quantitative precipitation forecasts is essential for reducing losses and damages. This study aimed to identify the best microphysics and cumulus schemes for forecasting monsoon rainfall over the Kelani River basin, Sri Lanka, using the WRF-ARW model. Four extreme rainfall events from the 2020 and 2021 monsoon seasons were simulated with various microphysics and cumulus parameterizations to find the optimal combinations. These combinations were then tested for their ability to forecast two monsoon events with a 24-h lead time. Simulated and forecasted rainfalls were compared with observations from 15 gauging stations. Results indicate that WSM3 and WSM6 microphysics schemes with the Betts–Miller–Janjic (BMJ) cumulus scheme are optimal for simulating rainfall, with WSM3_BMJ being the most suitable for forecasting. The findings of this study provide valuable initial data for research in regions with similar environmental conditions, offering insights into the suitability of various physics schemes for simulating and forecasting monsoon rainfall, particularly under extreme conditions. Furthermore, given the prevalence of monsoons in many tropical and subtropical climates, these results will be instrumental in enhancing the use of numerical weather prediction models for forecasting monsoon rainfall on a global scale.

  • The influence of the WRF model physics schemes for forecasting monsoon rainfall over the Kelani River basin was explored.

  • Four microphysics and two cumulus physics schemes were tested.

  • Field data were used to assess the simulations and forecasts.

  • WSM3_BMJ and WSM6_BMJ were optimal in simulating the rainfall events.

  • WSM3_BMJ gave better results in forecasting rainfall events.

Seasonal rainfall fluctuations over time can have a significant socioeconomic influence. Droughts and crop failure are linked to seasonal rainfall deficits, while catastrophic floods and damage to houses are linked to seasonal rainfall excesses. Moreover, rainfall plays a key role in maintaining and enhancing both groundwater and surface water resources, ensuring their availability and quality for diverse needs. Rainfall replenishes aquifers, maintaining groundwater levels essential for drinking, irrigation, and industry. It also sustains rivers, lakes, and reservoirs, ensuring a sustainable surface water supply (Noori et al. 2021; Maghrebi et al. 2023; Noori et al. 2023). Hence, rainfall prediction offers crucial information for a wide range of real-world uses, including the management of water resources, food security, farm management, disaster prevention, energy supply and industry. Moreover, due to climate change, rapid urbanization and economic growth, the world is becoming more susceptible to unforeseen hydro-meteorological risks, which highlights the need for rainfall prediction (Ngailo et al. 2018; Sarker et al. 2024; Singhal et al. 2024).

Sri Lanka is a tropical country, and air temperature barely minimally varies year-round, with the exception of the highlands, meaning that rainfall is the primary determinant of climate. As it is located in the Indian Ocean between the Bay of Bengal and the Gulf of Mannar, it receives unexpected heavy rains due to the pressure variations in the Bay of Bengal with high winds (Figure 1(a)) (Wijetunge 2014). The geographical settings and geomorphic and climatological conditions of the country cause floods, the most frequent natural disaster that strikes not only the economy but also the social status of the affected areas as well as of the country (Manawadu & Wijeratne 2021). Hence, a flood early warning system is needed to give warnings, take action to mitigate flooding, evacuate residents who may be impacted, and perform other related tasks. To build an accurate flood forecasting system, precise rainfall predictions are necessary. In addition to that, in Sri Lanka, as in most of South and Southeast Asia, rice agriculture is vital. Therefore, the beginning of the rainy season is quite important, and farmers must decide what tactics to use when the rains arrive later than expected (Jacobi 2014).
Figure 1

(a) Location of Sri Lanka and the Kelani River basin and (b) Kelani River basin including locations of the selected 15 rainfall gauging stations.

Figure 1

(a) Location of Sri Lanka and the Kelani River basin and (b) Kelani River basin including locations of the selected 15 rainfall gauging stations.

Close modal

Rainfall forecasts are typically addressed through two methods: either statistical models based on correlations between predictands and predictors (Jin et al. 2015) or experimenting with general circulation models (GCMs) (Hawthorne et al. 2013). GCMs have demonstrated their relative ability to simulate many aspects of climate variability at large or global scales but not at local scales. Two approaches are frequently used to close the gap between the global and local scales; that is, statistical downscaling (Lanzante et al. 2018) and dynamic downscaling (Xu et al. 2018). Numerous regional numerical weather prediction (NWP) models, including the Weather Research and Forecast model (WRF), the Pennsylvania State University–National Center for Atmospheric Research (PSU/NCAR) mesoscale model (MM5), and the regional atmospheric modeling system (RAMS), have been made available as dynamic downscaling tools to produce quantitative precipitation forecast (QPF) products with higher temporal and spatial resolution (Liu et al. 2012).

WRF is one of the most popular and advanced mesoscale NWP models available, with a variety of physical schemes that make it easier to predict weather variables with a high degree of accuracy (Avolio & Federico 2018). It is frequently used for researching and predicting a variety of high-impact meteorological occurrences, including precipitation (Politi et al. 2018), tropical cyclones (Routray et al. 2016) and thunderstorms (Osuri et al. 2017). It has been shown through simulations and real-time forecasting that the WRF model has good properties for predicting a wide variety of weather (Zhang 2004). Das et al. (2008) studied the performance of four NWP models to investigate the 2006 monsoon season rainfall patterns over the Indian region. The best one-day forecast was produced by WRF.

The primary goal of NWP is to improve the description of the behavior of the physical processes in the atmosphere that influence weather occurrences and utilize that information in the model to predict the condition of the atmosphere at a future date (Patel et al. 2019). These physical operations are referred to as parameterization schemes. The WRF model employs five different types of parameterization schemes, namely, microphysics (MP), radiation, cumulus, land surface model, and planetary boundary layer. Hence, for the model to evolve more effectively, the optimal physics solution for a certain area must be determined through a physical sensitivity investigation (Das et al. 2015).

Numerous investigations have been conducted globally to assess the physical schemes of the WRF model in order to simulate rainfall (Rama Rao et al. 2012; Darshika & Premalal 2015; Sikder & Hossain 2016). Cumulus physics (CP) and MP schemes in regional models were considered crucial for resolving micro-scale weather systems and determining the distribution and intensity of precipitation projections (Liu et al. 2018). Due to the intricate topographical features, particularly its central mountainous region, numerical simulations using the WRF model are more rigorous in Sri Lanka (Gimhan et al. 2022). In addition, owing to its geographical location, Sri Lanka is impacted by the monsoon system originating from the Indian Ocean and the Bay of Bengal, which causes a yearly, methodical movement of heavy rainfall throughout the area. The southwest monsoon (SWM: May to September) and the northeast monsoon (NEM: December to February) control seasonal variations in rainfall. SWM originates from the Indian Ocean while NEM brings rain from the Bay of Bengal (Geethalakshmi et al. 2009). Rainfall forecasting is crucial for river basins, including the Kelani River basin, for flood management, water resource management, disaster preparedness, agriculture, and hydropower generation. By minimizing flood risks and ensuring residents' safety, effective forecasting also enhances agricultural and industrial activities, promoting economic sustainability in the region (Gao et al. 2022; Sarker et al. 2023). Therefore, more research is required as there is no clear conclusion on physics systems to forecast monsoon rainfall in Sri Lanka.

With this background, this paper seeks to evaluate the sensitivity of both sets of CP and MP schemes of the WRF- Advanced research WRF (ARW) model in forecasting SWM and NEM over the Kelani River basin, Sri Lanka. The Kelani River basin is home to more than 20% of the Sri Lankan population (Mahagamage et al. 2020). Forecasting rainfall in the Kelani River basin has significant societal and economic benefits as both the commercial capital (Colombo) and the administrative capital (Sri Jayawardanapura) of Sri Lanka are in the lower Kelani River basin (Gunasekara 2008). The landscape of the Kelani River basin is shaped by the interplay of natural features and human activities. Effective management and conservation are vital for maintaining ecological balance and promoting sustainable development. Thus, rainfall forecasting supports erosion control, sediment management, habitat protection, infrastructure planning, and land use planning (Sarker 2021). Furthermore, since monsoons are a common occurrence in many tropical and subtropical climates, the results of this study will be helpful in utilizing NWP models to predict monsoon rainfall in a variety of global locations. The paper is structured as follows: in Section 2, the methodology, including the study area, rainfall event selection, model configuration, and evaluation of model performance, are summarized. The interpretation of the results is given in Section 3. Results are discussed in Section 4 and conclusions are summarized in Section 5.

Kelani River is the fourth longest river in Sri Lanka, which originates in the central hills and flows mainly to the west, passing 192 km, until it meets with the sea at the northern boundary of Colombo city. The Kelani River basin is located at the coordinates between northern latitudes of 6° 46′ and 7° 05′ and eastern longitudes of 79° 52′ and 80° 13′, as shown in Figure 1(b). The basin receives about 3,450 mm of average annual rainfall. The river carries a peak flow of about 800–1,500 m3/s to the Indian Ocean during monsoons. Floods in the river basin mainly occur during southwest monsoons that arrive in late May (Silva 2016).

Topographically, the river basin consists of two types of landscapes. The upper basin is a mountainous region and the lower basin is a flat coastal plain (Figure 1(b)). The valley slope is steep and the coastal plain is extremely flat. In the last 30 km of the river, the bed slope flattens considerably and river valleys open out, revealing wide flood plains. While the lower basin is extensively populated, the upper basin is mostly covered with dense flora like tea, rubber, grass, and woodland.

Numerical model set-up

The WRF model is a system designed for modeling atmospheric conditions for use in research and NWP (Ntwali et al. 2016). The WRF model was initially developed at the NCAR in collaboration with other research institutes in the United States (National Center for Atmospheric Research 2023). ARW is a WRF system configuration that uses the ARW dynamics solver and additional compatible parts to generate a simulation. As such, it is a subset of the WRF system that includes initialization procedures, physics schemes, numeric/dynamics choices, and a data assimilation package in addition to the particular solver. The non-hydrostatic, compressible Euler equations are integrated by the ARW dynamics solver. Following the Ooyama (1990) approach, these equations are expressed in flux form using variables that have conservation qualities. Furthermore, as explained by Laprise (1992), they are formed with hydrostatic pressure as an independent variable. Further details on the WRF-ARW model and its governing equations can be found in Skamarock et al. (2021).

Figure 2 displays the general system workflow of the WRF-ARW model. As shown in the figure, the process begins with the input of external data. The geogrid in the preprocessing system defines the model domains and interpolates static geographical data onto the grids. The ungrib tool extracts meteorological fields from General Regularly-distributed Information in Binary form (GRIB)-formatted files. The metgrid tool then horizontally interpolates these meteorological fields to the model grids defined by geogrid. The interpolated output from the metgrid is ingested by the WRF real program. Then WRF generates the model forecasts. Finally, WRF post-processing offers several options to manage and visualize data, although any computing language or software can be used to handle the output WRF files, which are in the netCDF format.
Figure 2

WRF-ARW modeling system workflow.

Figure 2

WRF-ARW modeling system workflow.

Close modal
The WRF-ARW model (version 4.1.3), which has been widely used in rainfall forecasting simulations, was used in this study. Considering a healthy balance between accuracy and computational cost, nesting was used to obtain a finer spatial resolution in this study. Three domains (Figure 3) with a horizontal resolution of 27 km (118 × 118), 9 km (118 × 118), and 3 km (64 × 64) along with 45 vertical levels were used. Considering the significant effects of low-pressure systems in the Bay of Bengal having heavy precipitation in Sri Lanka, the largest domain was selected to cover the entire Bay of Bengal. The smallest domain was selected to represent the Kelani River basin.
Figure 3

WRF model configuration.

Figure 3

WRF model configuration.

Close modal

The National Centers for Environmental Prediction (NCEP) Final (FNL) high-resolution analysis dataset was used as the initial and lateral boundary conditions to identify the optimum physics combination for the river basin due to the better accuracy, as it is known that around 10% more observational datasets are ingested by NCEP-FNL data than by Global Forecasting System (GFS) (Mukherjee & Ramakrishnan 2021; Li et al. 2024). Analysis data are on 0.25° × 0.25° grids available at 6 hourly intervals. The NCEP GFS analysis dataset was used for the forecasting simulations. The forecast grids are on a 0.25 × 0.25 global latitude–longitude grids including forecast time steps at a 3 hourly interval (National Center for Atmospheric Research 2023). The model was left to solve its nested domain numerically utilizing the forcing from the parent domain after the initial and lateral boundary conditions were forced onto the outer domain. The ‘Mandatory static data’ set was used as the terrestrial data for the simulation (National Center for Atmospheric Research 2023). Eight days of selected monsoon rainfall events were simulated. The first day was designated as the spin-up period.

Selection of parameterization

The WRF model has a wide range of physics scheme options that have a variety of features that can be adjusted to fit the model's operating environment. For reliable results, the best physics combinations must be chosen. In the present study, to find appropriate schemes, a comprehensive study of the literature, of the research conducted in tropical areas was conducted. Among the physics combinations the MP and CP schemes were changed that are considered most susceptible to the NWP over tropical regions (Liu et al. 2018). In the current study, Kessler (Kessler, 1969), Purdue Lin (Lin et al., 1983), WRF Single-Moment3-class Microphysics scheme (WSM3) (Han & Hong 2018), and WRF Single-Moment6-class Microphysics scheme (WSM6) (Hong & Lim 2006) were used for MP processes while the Kain-Fritsch (KF) scheme (Kain 1994) and Betts-Miller-Janjic (BMJ) scheme (Janjić 1994) were used for CP processes. The remaining schemes were set as WRF default schemes within the selected combinations. MP and CP combinations incorporated in this study are shown in Table 1.

Table 1

Selected physics combinations for the study

MP schemeCP schemeAbbreviation
Kessler KF K_KF 
Kessler BMJ K_BMJ 
Lin KF L_KF 
Lin BMJ L_BMJ 
WSM3 KF W3_KF 
WSM3 BMJ W3_BMJ 
WSM6 KF W6_KF 
WSM6 BMJ W6_BMJ 
MP schemeCP schemeAbbreviation
Kessler KF K_KF 
Kessler BMJ K_BMJ 
Lin KF L_KF 
Lin BMJ L_BMJ 
WSM3 KF W3_KF 
WSM3 BMJ W3_BMJ 
WSM6 KF W6_KF 
WSM6 BMJ W6_BMJ 

Rainfall events selection

Daily precipitation data of the selected months at 15 rainfall gauging stations (Figure 1(b)) were received through the Department of Meteorology, Sri Lanka. Rainfall events were selected based on the highest recorded total rainfall over the catchment. Thiessen polygons were constructed using the ArcGIS software over the catchment based on the rainfall gauging stations, to cater for the contribution of each station.

Simulation periods, Event 1 – SWM – 10–16 May 2021 and Event 2 – NEM – 5–11 November 2021, were selected to identify the optimum physics combination for the river basin. To validate the optimality of that physics combination under the monsoon, Event 3 – SWM – 11–17 May 2020 and Event 4 – NEM – 4–10 November 2020, were selected. Finally, events 1 and 2 were forecasted incorporating the selected physics combinations to evaluate the performance of the WRF model in forecasting the rainfall during the two monsoons.

Evaluation of model performance

The WRF model rainfall simulation results were compared with the observed data gathered at the 15 rainfall gauging stations described in Section 2.2. The Thiessen polygon method was used to get the average rainfall over the basin. For the accuracy evaluation, both continuous and categorical verification indices were employed. Categorical indices provide a general assessment of the model performance based on the correctness of occurrences of rainfall while continuous indices provide a quantitative value for the simulation error. Every index in this study was computed for every experiment and ranked based on how well the model performed. Every event was then ranked jointly using all of the categorical and continuous indices. Ultimately, the optimal physics schemes were recommended based on the rainfall season over the Kelani River basin.

Categorical indices

The selected categorical indices are the probability of detection (POD), frequency bias index (FBI), false alarm ratio (FAR), and critical success index (CSI) (Liu et al. 2012; Praja & Wati 2019). The categorical indices were calculated based on the rainfall contingency table. In the current study, a threshold level of 10 mm was used. The POD indicates the fraction of the observed rains that were correctly simulated. The FAR shows the percentage of the simulated rainfall that didn't occur. The CSI shows the relationship between the observed and simulated rainfall. The FBI indicates if the WRF tends to overestimate (FBI > 1) or underestimate (FBI < 1) the amount of rainfall that occurred.

Continuous indices

In this study, the selected continuous indices include root mean square error (RMSE), correlation coefficient (CC), and standard deviation (SD) (Gimhan et al. 2022; Maddah & Mostamandi 2024). These continuous scores were calculated to build the Taylor diagram (Taylor 2001), which graphically depicts the degree to which a pattern or group of patterns matches observations.

Identification of the optimum physics combination

Event 1 (May 2021)

Figure 4(a) illustrates the graphical representation of the daily cumulative rainfall resulting from rainfall Event 1. W6_BMJ, W3_BMJ, K_KF, and K_BMJ captured the peak of the rainfall event with an underestimation. It is noticeable that K_KF, W3_KF, and K_BMJ showed two peaks during the rainfall event. Combination L_KF significantly deviated from the observed rainfall curve.
Figure 4

Simulated May 2021 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Figure 4

Simulated May 2021 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Close modal

The computed continuous statistical indices (Figure 4(c)) were displayed incorporating the Taylor diagram, as depicted in Figure 4(b). In Taylor diagrams, WRF models that match the reference well are located closest to the ‘ref’ point on the x-axis. These WRF models feature low RMSE and a comparatively good CC. WRF models closest to the dashed arc have the proper SD, meaning that the model's variations fall within the appropriate amplitude range. Taylor's diagram showed that W3_BMJ is located closest to the ‘ref’ mark as well as to the dashed line. Among combinations, W3_BMJ exhibited the strongest correlation (0.84) and the lowest RMSE (29.14). W6_BMJ produced the second-best performance in comparison to the remaining combinations. It had the second-best CC (0.75) and the second-lowest RMSE (31.95). L_KF displayed the lowest CC (0.43) and the highest RMSE (49.69), showing the poor relationship between the two variables, which could be seen in the Taylor diagram as well. Hence, L_KF's performance showed the poorest outcomes. It was evident that in the simulation of Event 1, the BMJ CP scheme performed better than the KF scheme. Furthermore, the current rainfall event was better replicated using the two WSM MP schemes together with the BMJ CP scheme.

Figure 4(d) exhibits the results obtained for categorical verifications for rainfall Event 1. Results showed that L_BMJ, W3_KF, W3_BMJ, W6_KF, and W6_BMJ displayed the perfect score for FAR (0). W3_KF, L_BMJ, and W3_BMJ scored nearly the best scores for all the categorical indices and may be considered the top-performing models. L_KF scored poorly in every categorical index. Considering the second-best-performed models in categorical indices, W6_BMJ and W6_KF simulated the current rainfall event more accurately than the other schemes.

Event 2 (November 2021)

A graphical representation of the daily cumulative rainfall resulting from Event 2 (Figure 5(a)) showed that only combinations W6_BMJ, W3_BMJ, and K_BMJ captured the peak rainfall during this event. With the exception of W3_KF and W6_KF, every combination displayed curves that were lower than the actual amounts of rainfall. Furthermore, W3_KF overestimated the rainfall magnitudes except on the 8 November 2021.
Figure 5

Simulated November 2021 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Figure 5

Simulated November 2021 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Close modal

The Taylor diagram constructed for rainfall Event 2 (Figure 5(b)) showed that W3_BMJ attained the best performance in the continuous verification. It had the best CC (0.75) and the RMSE (20.78) was also lower than other combinations (Figure 5(c)). The second-best performance could be seen in W6_BMJ. In general, it was apparent that the estimated continuous indices were barely satisfactory. Moreover, L_BMJ, W3_KF, L_KF, and W6_KF resulted in negative correlation values, showing a poor linear relationship between the simulated and the observed rainfall data. The poorest performance was recorded by W3_KF.

Categorical verification results of rainfall Event 2 (Figure 5(d)) showed that K_KF was unable to capture any rainfall higher than 10 mm. Even though K_BMJ scored zero FAR value, it scarcely scored other categorical indices. W6_KF recorded the best FBI (1) while W3_KF recorded the best POD (1). W6_BMJ, W3_BMJ, L_KF, L_BMJ and K_BMJ had the best FAR (1). Considering the overall categorical indices, L_BMJ, W3_BMJ, and W6_BMJ recorded the best results compared to other combinations. K_KF showed the poorest performance in reproducing the current rainfall event.

Validation

Event 3 (May 2020)

Two peaks occurred during this rainfall event, and only W3_BMJ and W6_BMJ captured both peaks and showed satisfactory similarity with the observed data curve compared to the rest of the combinations (Figure 6(a)). By eyeball analysis, L_KF, W6_KF, and K_KF captured a great quantity of rainfall in the later part of the event. K_BMJ was unable to seize either of the peaks and showed underestimated rainfall magnitudes throughout the event. W6_BMJ, W3_BMJ, and L_BMJ exhibited underestimated results throughout the simulation period.
Figure 6

Simulated May 2020 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Figure 6

Simulated May 2020 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Close modal

Upon reviewing the Taylor diagram (Figure 6(b)), it was clear that W6_BMJ could be identified as the best combination for simulating Event 3. It obtained the highest CC (0.75) (Figure 6(c)). W3_BMJ recorded the next best results among the remaining combinations with a comparably higher CC (0.68) and a lower RMSE (22.44). Even though W3_KF had the best RMSE (22.29), its CC (0.51) showed poor results. So it could be recognized as the third-best combination. L_BMJ had the best SD (4.87); however, its RMSE (28.70) was high. Thus its rank was lower. In the case of K_KF, it recorded the highest RMSE (41.58) as well as a lower CC (0.20). Hence, K_KF had the poorest results in simulating the current rainfall event. It was apparent that WSM MP schemes in combination with the BMJ CP scheme showed favorable results compared to other schemes.

Figure 6(d) exhibits the categorical verification results for rainfall Event 3. K_BMJ and W6_BMJ displayed the best score for FAR (0). W3_BMJ had the prime score for FBI (1) while K_KF, L_KF, W3_KF, and W6_KF exhibited the best score for POD (1). W6_BMJ showed an agreeable performance in each categorical index. W3_BMJ could be represented as the next best performance in each category among the remaining combinations. The poorest performance could be seen in K_BMJ. Nevertheless, it scored the best in the FAR (0), showing that the model did not falsely capture any rainfall with a magnitude less than 10 mm. Still, the model scored poorly in the rest of the indices.

Event 4 (November 2020)

As shown in Figure 7(a) the current rainfall event comprised two rainfall peaks and W3_BMJ, W6_BMJ, L_BMJ, and W6_KF seized the peak rainfall that occurred on 5 November. Moreover, W3_BMJ and W6_BMJ also captured the second rainfall peak that occurred on 8 November. It should be noted that both the combinations above underestimated the magnitudes of the two peaks.
Figure 7

Simulated November 2020 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Figure 7

Simulated November 2020 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Close modal

When reviewing the Taylor diagram constructed for rainfall Event 4 (Figure 7(b)), it was apparent that W6_BMJ showed very close agreement with the observations. As shown in Figure 7(c), it had the best CC (0.80) and the best RMSE (14.07) as well. W3_BMJ and L_BMJ also showed better continuous index values. The remaining combinations had negative CC values, indicating the simulated rainfall and the observed rainfall moved in opposite directions. These combinations had higher RMSE values replicating inadequate performance in capturing the rainfall event. The poorest performance could be seen in the simulation incorporating the K_KF scheme. It is important to note that the present rainfall event was more accurately reproduced by all MPs using BMJ CP compared to the MP combinations with KF CP.

The findings of categorical verifications undertaken for event 4 are shown in Figure 7(d). None of the combinations scored the best POD (1), FAR (0) or CSI (1) scores. L_BMJ, W3_KF, W3_BMJ, and W6_BMJ scored the prime score for FBI (1) showing that the simulations were unbiased, where the event was forecasted exactly as often as observed. Considering the overall categorical indices, W3_BMJ, W6_BMJ, and L_BMJ displayed positive aptness in simulation among others. The W6_KF and L_KF exhibited the second-best performance. The poorest categorical index values were recorded by K_KF. It should be noted that, overall, the categorical results were generally insufficient in the current rainfall event.

Forecasting

Upon reviewing the results of the simulations conducted for the four events, it was concluded that, compared to the other methods, W3_BMJ and W6_BMJ were better able to replicate the rainfall events. Hence, W3_BMJ and W6_BMJ were employed to examine the potentiality of the WRF model to forecast SWM and NEM rainfall events. First and second rainfall events were forecasted with a lead time of 24-h.

Event 1 (May 2021)

The forecasted daily cumulative rainfall totals for the rainfall event 1 are represented graphically in Figure 8(a). The highest rainfall on 13 May was only partially captured by W3_BMJ. Nevertheless, the combination highly underestimated the magnitude of the peak rainfall. W6_BMJ peaked on 12 May. The observed rainfall curve had an upward trend until 13 May when it hit the maximum rainfall. After that, it displayed a descending trend. W6_BMJ exhibited a similar feature, albeit reaching its maximum values on different days.
Figure 8

Forecasted May 2021 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Figure 8

Forecasted May 2021 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Close modal

The computed continuous indices, as shown in Figure 8(c), demonstrated that the W3_BMJ had the best CC (0.70) and RMSE (34.56). Still, there was a significant difference between the simulation's SD (30.18) and the data. It was also evident in the Taylor diagram that is displayed in Figure 8(b). Taylor's diagram demonstrated that compared to W3_BMJ, W6_BMJ had a lower performance in forecasting the rainfall event. When comparing the simulated Event 1 results with the forecasted results, it is clear that the predicted results were inferior to the simulated ones.

The findings of categorical verifications, as shown in Figure 8(d), showed that W3_BMJ received the highest POD score (1), indicating that the model accurately captured observed rainfalls over 10 mm and that W6_BMJ received the highest FAR score (0), indicating that the model did not incorrectly estimate rainfalls below 10 mm. However, for every verification criterion, W3_BMJ outperformed W6_BMJ in forecasting rainfall Event 1.

Event 2 (2021 November)

Graphical representation of the predicted daily cumulative rainfall totals for the rainfall Event 2 (Figure 9(a)) showed that none of the combinations could adequately represent the observed curve's form or the highest rainfall that fell on 8 November. W3_BMJ peaked on the 9th, while W6_BMJ peaked on the 7th. It was evident that both combinations underestimated the peak rainfall amount.
Figure 9

Forecasted November 2021 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Figure 9

Forecasted November 2021 rainfall event: (a) daily cumulative rainfall values, (b) Taylor diagram, (c) continuous indices, and (d) categorical indices.

Close modal

The computed continuous statistical indices of the rainfall Event 2, as presented in Figure 9(c), showed that W3_BMJ demonstrated better results compared to W6_BMJ in every index. It was observable in the Taylor diagram displayed in Figure 9(b) as well. In comparison to W6_BMJ, W3_BMJ was closest to the reference point in the figure, indicating that the WRF model more closely matched the observed data. Furthermore, W3_BMJ was closest to the dashed arcs, indicating that the amplitude range of the model's variations was suitable.

According to Figure 9(d), which depicts the categorical verification findings of the forecasting of Event 2, W3_BMJ scored the perfect score in FAR (0) and W6_BMJ scored the prime score in FBI (1). According to the results, W3_BMJ performed better in the categorical verification. It was clear that W3_BMJ outperformed W6_BMJ in forecasting rainfall Event 2 in each verification criterion. In general, it was apparent that the verification results were barely satisfactory.

The most crucial factor in QPF using the WRF model is the choice of physical parameterization schemes that the model simulations employ, as precipitation formation is strongly influenced by CP and MP parameterization. The inclusion or removal of certain hydrometeors and the conversion mechanisms between them vary between these schemes. MP schemes are required to address the behaviors of clouds and precipitation (Huang et al. 2020). No matter what kinds of combinations with which MP schemes are established for a given CP scheme, the rainfall simulations of each combination differ significantly. This suggests that the vertical structure is a key component of the qualities of MP. More specifically, the MP approaches for treating rainfall simulations are connected to the various mixing ratios of rainwater, ice crystal, cloud water, and snow crystal (Yin et al. 2019). Four MP schemes were evaluated in the current study: Lin, WSM3, WSM6, and Kessler. In the Lin MP scheme, six classes of hydrometeors are included: water vapor, cloud water, rain, cloud ice, snow, and graupel. This complex WRF MP system is best suited for research investigations (Lin et al. 1983). The WSM3 forecasts three types of hydrometers, rainwater mixing ratio, cloud water, and water vapor. This system does not include supercooled water or slow melting rates, but it is computationally efficient for the inclusion of ice processes (Han & Hong 2018). The WSM6 method comprises six distinct arrays for vapor, rain, snow, cloud ice, cloud water, and glaciers (Hong & Lim 2006). The Kessler scheme is a simple warm cloud plan that incorporates rain, cloud water, and water vapor. The creation, fall, and evaporation of rain, the buildup and auto conversion of cloud water, and the condensation-based creation of cloud water comprise the microphysical process (Kessler 1969). Out of these four MP methods, the Lin and WSM6 approach considers the transition of mixed-phase clouds as well as the transformation of ice phases. Nevertheless, the forecast skills of these schemes are not necessarily superior to others. This indicates that to provide a more accurate description of the microphysical processes, MP schemes must work in conjunction with other parameterization schemes.

The CP schemes indicate the influence of convection on environmental variables (Budakoti et al. 2019). The current study evaluated two CP schemes, namely KF and BMJ. Under the KF scheme, when condensates in the updraft above a threshold amount, they are transformed into precipitation. Convective available potential energy (CAPE) and downdrafts are both taken into account in the KF convective scheme. However, the deep convection processes are not well described in the KF scheme (Kain & Fritsch 1993). The BMJ CP scheme modifies temperature and humidity at the grid scale by nudging. There is no overt updraft or downdraft in this design and no cloud detrainment. The modified moist-adiabatic profile serves as the convective adjustment reference profile in the BMJ system, which accounts for both the shallow and deep convection processes (Janjić 1994).

In the current study, the QPF of Sri Lanka's two main monsoon seasons was assessed using the eight MP and CP combinations above, using the WRF-ARW model. Looking at the summary of each rainfall event, as shown in Table 2, it can be seen that both the WSM3 and WSM6 MP schemes with the BMJ CP scheme were able to simulate the rainfall events more accurately compared to other schemes. Similar results could be observed in the studies of Sikder & Hossain (2016) and Rama Rao et al. (2012).

Table 2

Summary of ranks of tested combinations in four rainfall event simulations

Verification CombinationContinuous verificationCategorical verificationOverall rankContinuous verificationCategorical verificationOverall rank
Rainfall event Event 1 Event 3 
K_KF 
K_BMJ 
L_KF 
L_BMJ 
W3_KF 
W3_BMJ 
W6_KF 
W6_BMJ 
Rainfall event Event 2 Event 4 
K_KF 
K_BMJ 
L_KF 
L_BMJ 
W3_KF 
W3_BMJ 
W6_KF 
W6_BMJ 
Verification CombinationContinuous verificationCategorical verificationOverall rankContinuous verificationCategorical verificationOverall rank
Rainfall event Event 1 Event 3 
K_KF 
K_BMJ 
L_KF 
L_BMJ 
W3_KF 
W3_BMJ 
W6_KF 
W6_BMJ 
Rainfall event Event 2 Event 4 
K_KF 
K_BMJ 
L_KF 
L_BMJ 
W3_KF 
W3_BMJ 
W6_KF 
W6_BMJ 

Considering the WRF single moment MP schemes, it includes the revised ice MP, which includes processes such as ice nucleation, accretion and number concentration, among other processes, to produce a more realistic production of cloud ice. The two most notable aspects of this approach are the new assumption that the concentration of ice crystals is a function of ice amount and the practical representation of ice microphysical processes through the ice nuclei number concentration as a function of temperature. In the study of Hong et al. (2004), they concluded that the WRF single moment MP schemes, which provide a better depiction of the ice-cloud/radiation feedback, combined with the sedimentation of cloud ice, display a considerable improvement in the high amount of clouds, surface precipitation and large-scale mean temperature. The WSM3 scheme is simple, while WSM6 is more complex. When taking into account the two SWM rainfall events, W3_BMJ is the optimal combination in simulating Event 1, while in the case of Event 3, W6_BMJ can be identified as the best scheme. Of the two NEM events, W3_BMJ reproduced Event 2 the best, and W6_BMJ placed first in terms of replicating Event 4. On the other hand, both combinations underestimated the maximum precipitation projection. Nevertheless, W6_BMJ displayed more understated outcomes than W3_BMJ. This can be explained by incorporating the findings of Wu et al. (2023). In that study, it was found that the convection initiation happened relatively early in the WSM3 experiments when fewer hydrometeor categories were taken into account, and vice versa in the WSM6 experiments due to the longer time needed for the freezing latent heat to be released. The slow condensation rate in the current study may be why WSM6 underestimated the occurrence of severe rainfall compared to WSM3.

In the case of CP schemes, since the reference profiles the BMJ system utilizes are more similar to the observed one, it has performed better in reproducing more realistic rainfall forecasts in every rainfall event in comparison to KF. The study's findings suggest that the CP scheme of KF, which is predicated on specific CAPE assumptions, could not be as applicable in the event of a monsoon depression when dynamic forcing plays a more significant role in the rising of moist air. For the instance of monsoon depression, the BMJ approach, which is predicated on the convective adjustment assumption, seems to be more applicable.

Finally, events 1 and 2 were forecasted incorporating the best performing two combinations, W3_BMJ and W6_BMJ. As shown in Table 3, in both events, W3_BMJ performed better than W6_BMJ.

Table 3

Summary of ranks of tested combinations in forecasting the two rainfall events

Verification CombinationContinuous verificationCategorical verificationOverall rankContinuous verificationCategorical verificationOverall rank
Rainfall event Event 1
 
Event 2
 
W3_BMJ 
W6_BMJ 
Verification CombinationContinuous verificationCategorical verificationOverall rankContinuous verificationCategorical verificationOverall rank
Rainfall event Event 1
 
Event 2
 
W3_BMJ 
W6_BMJ 

Results elucidated that both events were forecasted poorly compared to the previous simulations. In the forecasting simulations, the boundary conditions incorporated were NCEP GFS, which are less accurate in forecasting compared to NCEP FNL, which was used in the previous simulations. Although they come from the same data assimilation and forecasting technology, FNL and GFS are related but distinct products. The quantity of ‘real’ data incorporated into the starting conditions for GFS and FNL varies slightly, despite the fact that they both come from the same data sources. Using NCEP's massive computational resources, running a global NWP model is still time-consuming. So, in order to obtain a forecast rather than a hindcast, they must launch GFS early enough. To incorporate all available observational data, FNL, the final analysis, is slightly behind GFS. About 10% more observations are typically ingested by FNL than by GFS (Mukherjee & Ramakrishnan 2021; Li et al. 2024).

In order to increase the accuracy of precipitation forecasts, some researchers suggest using ensemble predictions to get optimum findings, whereby ensemble members are altered by establishing weight from forecasting results (Fathi et al. 2019). Furthermore, many researchers modify specific physical parameters in specific WRF physical schemes in order to customize the simulation scenarios (Di et al. 2018). Furthermore, the background field directly impacts how accurate the model simulation is. Consequently, the forecasting accuracy can be increased by adjusting the background field using data assimilation techniques such as three-dimensional variation assimilation.

This study highlighted and shortlisted effective and ineffective physics options for further numerical simulations and examinations to improve monsoon rainfall forecasting. Monsoon rainfall forecasting is crucial for environmental protection and climate change adaptation. It improves land, water, and disaster management, helps develop strategies for changing precipitation patterns, enhances preparedness for extreme weather events, supports sustainable water use, and protects habitats from flooding or drought, aiding biodiversity conservation (Sarker 2022; 2023). Therefore, the findings of this research are significant for optimizing the WRF model not only for the tropical island of Sri Lanka but also for similar regions worldwide with comparable environmental conditions.

In this study, we assessed how different MP and CP parameterization techniques affected the QPF over the Kelani River basin during the two main rainfall seasons, SWM and NEM, using the WRF-ARW model. The main conclusions are as follows:

While the position and magnitude of precipitation varied between model runs, overall the WRF model could replicate the distribution of precipitation.

The physical parameterization schemes employed in the model impact the precipitation predicted by the WRF model, and selecting the right combination is crucial to accurately modeling monsoon precipitation over the Kelani River basin.

WSM3 and WSM6 MP schemes with BMJ CP schemes are applicable to simulate monsoon rainfall over the Kelani River basin.

W3_BMJ showed better performance in forecasting rainfall events. However, when compared to the simulation of the same events, the performance was lacking.

This study focused on the influence of CP and MP physics schemes on monsoon rainfall forecasts due to their significant impact on precipitation. To enhance prediction accuracy, incorporating various parameterization schemes, such as those for radiation, boundary layer, and land surface physics, is recommended. Conducting more simulations of extreme rainfall across different seasons would help identify optimal physics combinations for year-round forecasting. Further sensitivity testing is needed to evaluate the impact of global model data, like National Oceanic and Atmospheric Administration (NOAA) and European Centre for Medium-Range Weather Forecasts (ECMWF), on WRF-ARW model predictions, given their dependance on lateral boundary and initial conditions.

Authors gratefully acknowledge the scholarship provided by the Postgraduate Programme in Civil Engineering, University of Peradeniya, Sri Lanka to the first author and the Computer Engineering Department of the University of Peradeniya for giving access to their high-performance computing (HPC) for conducting research simulations. Moreover, the authors are grateful to the Department of Meteorology, Sri Lanka for providing rainfall data throughout the study.

Funds to conduct this research were provided by the Faculty of Engineering, University of Peradeniya, Sri Lanka.

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

The authors declare there is no conflict.

Darshika
D. W. T. T.
&
Premalal
K. H. M. S.
2015
Simulate heavy rainfall during 19th to 28th December 2014 using WRF for different atmospheric physics
.
Sri Lanka Journal of Meteorology
1
,
32
40
.
Das
S.
,
Ashrit
R.
,
Iyengar
G. R.
,
Mohandas
S.
,
Das Gupta
M.
,
George
J. P.
,
Rajagopal
E. N.
&
Dutta
S. K.
2008
Skills of different mesoscale models over Indian region during monsoon season: forecast errors
.
Journal of Earth System Science
117
,
603
620
.
Das
M. K.
,
Chowdhury
M. A. M.
&
Das
S.
2015
Sensitivity study with physical parameterization schemes for simulation of mesoscale convective systems associated with squall events
.
International Journal of Earth and Atmospheric Science
2
(
2
),
20
36
.
Fathi
M.
,
Azadi
M.
,
Kamali
G.
&
Meshkatee
A. H.
2019
Improving precipitation forecasts over Iran using a weighted average ensemble technique
.
Journal of Earth System Science
128
,
1
17
.
Geethalakshmi
V.
,
Yatagai
A.
,
Palanisamy
K.
&
Umetsu
C.
2009
Impact of ENSO and the Indian ocean dipole on the north-east monsoon rainfall of Tamil Nadu State in India
.
Hydrological Processes: An International Journal
23
(
4
),
633
647
.
Gimhan
P. G. S.
,
Neluwala
P.
,
Acierto
R. A.
&
Raghavan
S. V.
2022
Assessment of WRF microphysics and cumulus parameterizations in simulating heavy rainfall events over Badulu Oya catchment, Sri Lanka
.
Journal of Water and Climate Change
13
(
12
),
4213
4233
.
Gunasekara
I. A.
2008
Flood Hazard Mapping in Lower Reach of Kelani River
.
FLOOD
.
Available from: https://wrd.unwomen.org (accessed 23 July 2023)
.
Hawthorne
S.
,
Wang
Q. J.
,
Schepen
A.
&
Robertson
D.
2013
Effective use of general circulation model outputs for forecasting monthly rainfalls to long lead times
.
Water Resources Research
49
(
9
),
5427
5436
.
Hong
S. Y.
&
Lim
J. O. J.
2006
The WRF single-moment 6-class microphysics scheme (WSM6)
.
Asia-Pacific Journal of Atmospheric Sciences
42
(
2
),
129
151
.
Jacobi
J.
2014
A Changing Climate in Sri Lanka: Shifts, Perceptions, and Potential Adaptive Actions
.
Vanderbilt University
.
Available from: https://www.proquest.com (accessed 23 July 2023)
.
Jin
L.
,
Zhu
J.
,
Huang
Y.
,
Zhao
H. S.
,
Lin
K. P.
&
Jin
J.
2015
A nonlinear statistical ensemble model for short-range rainfall prediction
.
Theoretical and Applied Climatology
119
,
791
807
.
Kain, J. S. 2004 The Kain-Fritsch convective parameterization: an update. Journal of Applied Meteorology 43 (1), 170–181.
Kessler
E.
1969
On the Distribution and Continuity of Water Substance in Atmospheric Circulations
.
American Meteorological Society
,
Boston, MA
, pp.
1
84
.
Lanzante
J. R.
,
Dixon
K. W.
,
Nath
M. J.
,
Whitlock
C. E.
&
Adams-Smith
D.
2018
Some pitfalls in statistical downscaling of future climate
.
Bulletin of the American Meteorological Society
99
(
4
),
791
803
.
Li
Z.
,
Peng
J.
,
Zhang
L.
&
Guan
J.
2024
Exploring the differences in kinetic energy spectra between the NCEP-FNL and ERA5 datasets
.
Journal of the Atmospheric Sciences
81 (2), 363–380
.
Lin
Y. L.
,
Farley
R. D.
&
Orville
H. D.
1983
Bulk parameterization of the snow field in a cloud model
.
Journal of Applied Meteorology and Climatology
22
(
6
),
1065
1092
.
Liu
D.
,
Yang
B.
,
Zhang
Y.
,
Qian
Y.
,
Huang
A.
,
Zhou
Y.
&
Zhang
L.
2018
Combined impacts of convection and microphysics parameterizations on the simulations of precipitation and cloud properties over Asia
.
Atmospheric Research
212
,
172
185
.
Maghrebi
M.
,
Noori
R.
,
Mehr
A. D.
,
Lak
R.
,
Darougheh
F.
,
Razmgir
R.
,
Farnoush
H.
,
Taherpour
H.
,
Moghaddam
S. M. R. A.
,
Araghi
A.
&
Kløve
B.
2023
Spatiotemporal changes in Iranian rivers’ discharge
.
Elementa: Science of the Anthropocene
11
(
1
),
00002
.
Manawadu
L.
&
Wijeratne
V. P. I. S.
2021
Anthropogenic drivers and impacts of urban flooding – A case study in Lower Kelani River Basin, Colombo Sri Lanka
.
International Journal of Disaster Risk Reduction
57
,
102076
.
Mukherjee
P.
&
Ramakrishnan
B.
2021
Comparative analysis of WRF simulated surface winds with satellite observations over the Arabian Sea (No. EMS2021-199)
. In
Copernicus Meetings
.
National Center for Atmospheric Research
2023
WRF USERS PAGE
. .
Ngailo
T. J.
,
Shaban
N.
,
Reuder
J.
,
Mesquita
M. D.
,
Rutalebwa
E.
,
Mugume
I.
&
Sangalungembe
C.
2018
Assessing Weather Research and Forecasting (WRF) Model parameterization schemes skill to simulate extreme rainfall events over Dar es Salaam on 21 December 2011
.
Journal of Geoscience and Environment Protection
6
(
01
),
36
.
Noori
R.
,
Maghrebi
M.
,
Mirchi
A.
,
Tang
Q.
,
Bhattarai
R.
,
Sadegh
M.
,
Noury
M.
,
Torabi Haghighi
A.
,
Kløve
B.
&
Madani
K.
2021
Anthropogenic depletion of Iran's aquifers
.
Proceedings of the National Academy of Sciences
118
(
25
),
e2024221118
.
Noori
R.
,
Maghrebi
M.
,
Jessen
S.
,
Bateni
S. M.
,
Heggy
E.
,
Javadi
S.
,
Noury
M.
,
Pistre
S.
,
Abolfathi
S.
&
AghaKouchak
A.
2023
Decline in Iran's groundwater recharge
.
Nature Communications
14
(
1
),
6674
.
Ntwali
D.
,
Ogwang
B.
&
Ongoma
V.
2016
The impacts of topography on spatial and temporal rainfall distribution over Rwanda based on WRF model
.
Atmospheric and Climate Sciences
6
(
2
),
145
157
.
Ooyama
K. V.
1990
A thermodynamic foundation for modeling the moist atmosphere
.
Journal of Atmospheric Sciences
47
(
21
),
2580
2593
.
Patel
P.
,
Ghosh
S.
,
Kaginalkar
A.
,
Islam
S.
&
Karmakar
S.
2019
Performance evaluation of WRF for extreme flood forecasts in a coastal urban environment
.
Atmospheric Research
223
,
39
48
.
Politi
N.
,
Nastos
P. T.
,
Sfetsos
A.
,
Vlachogiannis
D.
&
Dalezios
N. R.
2018
Evaluation of the AWR-WRF model configuration at high resolution over the domain of Greece
.
Atmospheric Research
208
,
229
245
.
Praja
A. S.
&
Wati
T.
2019
A preliminary results of assessment of BMKG-WRF numerical model daily rainfall forecasts performance using categorical verification
.
IOP Conference Series: Earth and Environmental Science
,
284
(
1
).
012013
.
Rama Rao
Y. V.
,
Alves
L.
,
Seulall
B.
,
Mitchell
Z.
,
Samaroo
K.
&
Cummings
G.
2012
Evaluation of the weather research and forecasting (WRF) model over Guyana
.
Natural Hazards
61
,
1243
1261
.
Routray
A.
,
Mohanty
U. C.
,
Osuri
K. K.
,
Kar
S. C.
&
Niyogi
D.
2016
Impact of satellite radiance data on simulations of Bay of Bengal tropical cyclones using the WRF-3DVAR modeling system
.
IEEE Transactions on Geoscience and Remote Sensing
54
(
4
),
2285
2303
.
Sarker
S.
2021
Investigating Topologic and Geometric Properties of Synthetic and Natural River Networks Under Changing Climate
.
Available from: https://stars.library.ucf.edu/etd2020/965 (accessed 3 July 2024)
.
Sarker
T.
,
Fan
P.
,
Messina
J. P.
,
Mujahid
N.
,
Aldrian
E.
&
Chen
J.
2024
Impact of Urban built-up volume on Urban environment: A case of Jakarta
.
Sustainable Cities and Society
105
,
105346
.
Silva
G. D.
2016
Event based flood inundation mapping under the impact of climate change: A case study in lower Kelani River basin, Sri Lanka
.
Hydrology: Current Research
7
(
1
). .
Singhal
A.
,
Jaseem
M.
,
Divya
,
Sarker
S.
,
Prajapati
P.
,
Singh
A.
&
Jha
S. K.
2024
Identifying potential locations of hydrologic monitoring stations based on topographical and hydrological information
.
Water Resources Management
38
(
1
),
369
384
.
Skamarock
W. C.
,
Klemp
J. B.
,
Dudhia
J.
,
Gill
D. O.
,
Liu
Z.
,
Berner
J.
,
Wang
W.
,
Powers
J. G.
,
Duda
M. G.
,
Barker
D.
&
Huang
X. Y.
2021
A Description of the Advanced Research WRF Model Version 4.3; No. NCAR/TN556+ STR
. Available from: https://opensky.ucar.edu/islandora/object/technotes%3A588/datastream/PDF/view (accessed 20 January 2023).
Taylor
K. E.
2001
Summarizing multiple aspects of model performance in a single diagram
.
Journal of Geophysical Research: Atmospheres
106
(
D7
),
7183
7192
.
Wijetunge
J. J.
2014
Disaster Risk Assessment and Mitigation Strategy for Tropical Cyclone Induced Storm Surge Hazard and Coastal Impacts of Climate Change in Sri Lanka
.
Available from: https://academicworks.cuny.edu (accessed 23 July 2023)
.
Xu
Z.
,
Rhoades
A. M.
,
Johansen
H.
,
Ullrich
P. A.
&
Collins
W. D.
2018
An intercomparison of GCM and RCM dynamical downscaling for characterizing the hydroclimatology of California and Nevada
.
Journal of Hydrometeorology
19
(
9
),
1485
1506
.
Zhang
G. C.
2004
Progress of weather research and forecast (WRF) model and application in the United States
.
Meteorological Monthly
30
(
12
),
27
31
.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY 4.0), which permits copying, adaptation and redistribution, provided the original work is properly cited (http://creativecommons.org/licenses/by/4.0/).