1. Introduction
Presently, waterlogging incidents occur frequently in numerous cities worldwide [
1], which has attracted a great deal of attention from both domestic and international scholars. Urban waterlogging has become one of the most significant natural disasters in cities [
2], which is a significant barrier to the sustainable development of cities. Increases in heavy urban rainfall and waterlogging in recent years [
3] pose a serious threat to urban residents’ property and personal safety [
4,
5]. This is demonstrated by the impact of events such as the “721” heavy rainfall in Beijing in 2012, the “720” heavy rainfall in Zhengzhou in 2021, and the “Storm Issa” subtropical cyclone’s large-scale waterlogging events in South Africa in 2022.
A large number of waterlogging simulation studies and waterlogging scenario predictions could provide theoretical support for the emergency management and orderly response of heavy urban rainfall waterlogging disaster events [
6] to combat waterlogging more effectively. Numerous academics from both China and other countries have conducted the following study to predict the outcomes of urban waterlogging disasters. The researchers simulate various urban waterlogging conditions in the study area based on PCSWMM and other software [
7] for varying rainfall return periods, rainfall duration, and peak coefficients. They identify locations with a high risk of waterlogging [
8] and then apply the research findings to other regions, such as Wales, United Kingdom; Wallonia, Belgium; and Fujian and Shenzhen, China [
7,
8,
9,
10,
11,
12,
13]. Alternatively, others estimate future waterlogging scenarios using the correlation between flooding and other factors. For example, Y. Kexin et al. [
14] conducted estimations based on the close relationship between the causes of waterlogging prediction changes and land cover using the LULC prediction model combined with a hydrodynamic model to predict future waterlogging trends. Y. Yue et al. [
15] used topography, evaporation, infiltration, pipeline discharge, river flood level, and other factors to evaluate the correlation between rainfall and waterlogging levels and to develop a multifactor correlation model. The simulation results provide a basis for scientific research and urban planning to reduce the risk of waterlogging.
Existing research, however, focuses primarily on the simulation and disaster risk analysis of waterlogging scenarios under the existing atmospheric conditions of cities, and there are relatively few studies on the simulation and prediction of waterlogging disaster events under different future climate scenarios with climate change. Utilizing Chicago rain patterns to generate rainfall sequences, Mei Chao et al. [
16] constructed a numerical model of urban waterlogging in order to simulate waterlogging in the central urban area of Shanghai. Wang Shijing et al. [
17] designed rainfall patterns based on 24 h rainfall in the local rainstorm and flood atlas and simulated scenarios with various rainfall return periods. P. Chen et al. [
9] determined the average rainfall intensity at any point in the rainfall process by analyzing historical rainfall data and incorporating the rainfall intensity formula into the rainfall model as a parameter input in the numerical model of waterlogging. The majority of the rainfall data used in current studies for the prediction of future waterlogging scenarios are rainfall data series generated or predicted under current climate conditions, which do not fully account for rainfall conditions under future climate change conditions and could result in inaccurate estimation results. As a significant factor influencing rainfall and waterlogging, the climate is better suited to the realities of the future when fully combined with the results of climate change scenarios for urban waterlogging. Therefore, this paper will use CMIP6 rainfall model data to provide a reasonable portrayal of precipitation changes and provide scientific support for assessing climate change risks and effectively responding to climate-change-induced waterlogging disasters.
As the most recent stage with the greatest number of models and the richest data since the inception of the CMIP program, the CMIP6 model has significantly improved the simulation effect of average precipitation and extreme precipitation events [
18,
19,
20,
21] and is widely used in climate simulation and evaluation across a variety of spatial scales. Y. Farhad et al. [
22] analyzed the performance of CMIP6 on the historical operating precipitation estimation results of Iran’s climate zone. They discovered that each CMIP6 model had its own adaptive specific climate zones, such as the model that performed well in a humid zone but poorly in ultra-arid and arid zones. H. Zahra [
23] examined trends in extreme precipitation in the Middle East and North Africa (MENA) during historical periods (1985–2014) and future periods (2021–2050), and the results suggest that extreme precipitation and its intensity will continue to increase despite the decrease in precipitation in the Middle East and North Africa region. J. d. M. Felipe et al. [
24] estimated the changes in CMIP6’s worst-case scenario for the medium term (2046–2065) and future (2081–2100), predicting more severe, frequent, and persistent extreme precipitation events in all regions of Brazil and more pronounced changes in heavy precipitation and severe drought in north-central and southern Brazil.
In the rainfall forecast for China, Yang Yang et al. [
25] estimated precipitation in Asia over the next 30 years, revealing that flooding in western China could increase, and precipitation in southwest and northern China could decrease during the warm season. Drought disasters could occur in northern South Asia and southern Central Asia during the cold season. A further evaluation by Xiang Junwen [
26] and others determined that the simulation effect of CMIP6 in precipitation and temperature in northwest China was improved, particularly in the semi-arid area of the arid region, where the precipitation simulation ability was significantly enhanced compared to CMIP5 [
27]. Zhao Mengxia [
28] et al. simulated the spatial distribution of precipitation in the upper reaches of the Yellow River using the average data of CMIP6 multi-model collections and estimated that the annual precipitation in the upper reaches of the Yellow River would fluctuate and increase in the future. Global climate models are an important tool for studying past climate change mechanisms and predicting future climate change; however, due to the complexity of the global climate system and the stability of climate models, global climate models have certain deficiencies in East Asia’s climate simulation capabilities [
29,
30,
31]. Therefore, before using climate models to conduct climate change mechanism studies and climate forecasts, it is necessary to test and evaluate their simulation capabilities, which not only provide a foundation for improving model performance but also aid in identifying the sources of uncertainty in model results [
32].
This paper uses the most recent global climate models and datasets to predict future urban rainfall and flooding scenarios that are more closely aligned with realistic future scenarios, and it assesses the precipitation modeling capabilities of China’s major regions using the ten most recent climate models released by CMIP6. The SWMM model was used to predict future urban waterlogged scenarios based on varying climate scenarios. The CMIP6-based urban rainfall forecasts provide decision makers and planners with important information for urban planning, infrastructure design, and climate change adaptation, while the predicted flooding scenarios enable city management to better understand and assess the potential future flood risk and to take steps to mitigate the potential impact of flooding. This ensures the sustainable development of cities and the protection of residents and provides a theoretical foundation for emergency management and the scientific response to urban waterlogging events.
3. Results and Discussion
3.1. Initial Evaluation of Precipitation Patterns
Aiming at the degree of deviation between the model historical observation rainfall value and the model simulated rainfall value, Xiang Junwen, LIN Wenqing [
26,
43], and others used the Taylor diagram method to evaluate the simulation performance of extreme rainfall in China and the middle and high latitudes of Asia in multiple climate models of the CMIP6 database. The results indicate that the model EC-Earth has the best simulation performance. Zhou Tianjun and Wang Yu et al. [
27,
44] evaluated the simulation capability of multiple CMIP6 models on China’s historical extreme precipitation and discovered that GFDL-ESM4 and other models performed better. Hu Yiyang and Li Xiaolei et al. [
45,
46] compared the historical measured values of precipitation from multiple models with the model simulation values and discovered that the errors and precision of MPI-ESM1-2 models in China were relatively small and high. Only the high-resolution model of MPI-ESM1 maintains a good rainfall simulation effect as the resolution increases.
In conclusion, the pertinent literature was compiled, and the selection was based on various parameters, including the simulation of extreme precipitation, applicability to the study area, precipitation patterns, and enhanced performance relative to CMIP5, as indicated in
Table 2 below.
As the first large sample dataset of GCMs (global climate models) in China, the FGOALS-g3 model exhibits a high degree of adaptability to the region. The models ACCESS-CM2, IPSL-CM6A-LR, MPI-ESM1-2-LR, and MPI-ESM1-2-HR have low error rates when simulating historical precipitation in several regions of China. The simulation performance of the models GFDL-ESM4 and GFDL-CM4 is superior to that of the comparable model in CMIP5. EC-Earth3, EC-Earth3-Veg, GFDL-ESM4, GFDL-CM4, and MRI-ESM2-0 have better simulation capability for extreme rainfall. In conclusion, the pertinent literature was compiled to acquire the preliminary screening findings of the CMIP6 rainfall model in the Chinese context, and the model-specific information is presented in
Table 3 below.
3.2. Comparison of Models
Initially, each model’s historical monthly precipitation data was evaluated using box plots.
Figure 2 below shows the results of the box plot analysis.
In the historical monthly rainfall data, there were large anomalies in both the measured and predicted values of each model, indicating that extreme rainfall events were frequent. In the simulation of the anomalies, i.e., extreme rainfall, the simulated values of GFDL-ESM4, EC-Earth3, and MPI-ESM1-2-HR were the best. The fluctuation range of the simulated values of GFDL-CM4 and GFDL-ESM4 is in good agreement with the measured fluctuation range. With the exception of MPI-ESM1-2-LR, the mean values of the simulated values and measured values for each model are near to 35 mm, with GFDL- ESM4, EC-Earth3, and GFDL-CM4 being the closest models. Except for models MPI-ESM1-2-LR and MPI-ESM1-2-HR, the mean values of the simulated historical rainfall values are lower than the mean values of the measured values, indicating that the simulated rainfall data of the models are generally lower than the measured data.
Using the simulated and observed monthly precipitation data for each model from 1994 to 2014, a Taylor diagram depicting the link between multi-year monthly precipitation and the observed values for each model was then generated, As shown in
Figure 3 below.
The results of the Taylor diagram indicate that the simulation results of the CMIP6 model for monthly rainfall are superior to those of the wet and dry seasons. For monthly rainfall, the RMSEs of simulation versus observation are all within 1, indicating that the simulation results are in some agreement with the observation results, but there are still errors. The ratio of standard deviation between most models and observations is close to 1, revealing that the amplitude of the rainfall variation simulated by the models is comparable to that observed and that the discrepancies across the models are minimal. Additionally, the pattern distribution on the Taylor diagram is more concentrated, indicating a degree of pattern constancy. The correlation coefficients of the models ranged from 0.25 to 0.6, with MPI-ESM1-2-HR achieving the highest value of 0.6, indicating that the model has a strong ability to simulate rainfall. Taking into account the three indicators of correlation coefficient, standard deviation, and root mean square error, MPI-ESM1-2-HR, EC-Earth3, and GFDL-ESM4 performed the best, and these three models have the greatest ability to simulate rainfall. The outcomes of these three models’ simulations are closer to the observed results, and their simulation capacities are greater.
Finally, the Taylor score (TS) and the interannual variability score (IVS) were determined for each model according to the necessary equations, and the best-performing rainfall model was synthesized by rating them independently.
According to
Table 4 above, the top-performing models in terms of interannual variability score are FGOALS-g3, GFDL-ESM4, and MPI-ESM1-2-HR, while the top-ranked models in terms of the Taylor score are EC-Earth3, GFDL-ESM4, and MPI-ESM1-2-HR. GFDL-ESM4, MPI-ESM1-2-HR, and EC-Earth3 are the final selection of the best-performing models overall in the research area when the two comparators are combined.
3.3. SWMM Model Validation and Sensitivity Analysis
There are both deterministic and uncertain parameters in the SWMM model. Deterministic parameters are primarily physical parameters. They can be derived via field measurements or by utilizing the spatial processing function of ArcGIS. Accurate values for uncertainty parameters are often not available due to a variety of conditions, such as the manning coefficient of impervious area and the manning coefficient of pervious area, the depth of depression storage on the impervious area and the depth of depression storage on pervious area, the maximum infiltration rate, the minimum infiltration rate, and the decay constant. When determining the initial values of the uncertainty parameters, reference can be made to the SWMM model application manual, outdoor drainage design standards, and recommendations from the relevant literature.
Table 5 details the SWMM model uncertainty parameters for this investigation.
The SWMM model incorporates continuity errors for surface runoff and flow rate. In general, if these errors fall within a range of less than 10%, the results of the model simulation can be considered reasonable, indicating the model’s validity. Using the Chicago Rainfall Model calibration parameters, the SWMM calculations are acceptable if the calculated precipitation export flows are within 20% of the design flows and the relative deviations are less than 20%.
In this investigation, the SWMM model’s continuity error was evaluated. The results indicate that the surface runoff error is −0.01% and that the flow rate error is −0.94%, both of which are less than 10%, indicating that the model is accurate. Moreover, the model’s relative deviation is within 20%, indicating that the calculated results are reasonable. As a result, the model can be applied to the analysis of urban storm extreme weather scenarios in order to provide accurate information regarding the performance of urban drainage systems.
Next, the total runoff sensitivity of each parameter in the SWMM hydrological module was calculated using a sensitivity formula based on a modified Morris screening method to obtain the sensitivity of each parameter (
Table 6).
As demonstrated by the simulation results, the parameters that are most sensitive to runoff are destore-imperv, conduit roughness, and %zero-imperv, in that order.
3.4. Rainfall Forecasts under Future Scenarios
3.4.1. Annual Rainfall
Under different scenarios, the predicted results of each model have certain differences, so it is necessary to analyze under the uncertain conditions of the forecast. This is essential for setting standards for climate-resilient rainwater drainage systems in the future context of climate change. This study examines the interannual variance of Beijing’s total yearly precipitation and creates a map of this variation.
In terms of total annual precipitation for the period 2020–2050, the three future precipitation models are inconsistent. However, they are not below their historical averages. This suggests that there may be more rainy days or an increase in the average daily precipitation. The annual total rainfall that approaches or exceeds the historical maximum (about 800 mm) happens most frequently in the SSP1–2.6 scenario, and two, three, and four times in the forecast findings of GFDL-ESM4, MPI-ESM1–2-HR, and EC-Earth3, respectively.
As shown in
Figure 4, the uncertainty in the predicted future annual total rainfall increases, and the variability in the characteristics with historical rainfall is reflected in the significant increase in interannual extremes, with the highest annual rainfall of more than 1200 mm and the lowest annual rainfall of less than 200 mm predicted by EC-Earth3 under the SSP2–4.5 scenario for the future period.
Figure 5 below shows the future simulated rainfall compared to the historical measured maximum daily rainfall:
During the historical period of 1994–2014, the maximum daily precipitation observed was 160.274 mm. However, in all three models, it is predicted that future precipitation events exceeding the historical maximum daily precipitation will occur by the year 2050. Specifically, the EC-Earth3 model predicts the occurrence of the largest daily precipitation event under the SSP5–8.5 scenario, with a 24 h precipitation amount of 225.77 mm.
Figure 6 demonstrates the number of events that exceeded the historical maximum daily precipitation:
In the statistical analysis of the number of extreme rainfall events exceeding the maximum historical daily rainfall in three scenarios for each model, it was found that the EC-Earth3 model simulated the most extreme rainfall events, reaching up to two events in the SSP5–8.5 scenario. Meanwhile, the MPI-ESM1-2-HR model simulated the fewest extreme rainfall events, with only one event in the SSP2–4.5 scenario. Overall, it is predicted that one to two extreme rainfall events will exceed the maximum historical daily rainfall in the next 30 years.
The statistics for the quantity of future rainstorm events are depicted in
Figure 7:
According to the classification of rainfall levels, rainfall exceeding 100 mm per day is considered heavy rain. Historical rainfall data shows that there were three heavy rainfall events during the period of 1994–2014. However, future simulation data indicate a significant increase in heavy rainfall events in both the EC-Earth3 and GFDL-ESM4 models, particularly the EC-Earth3 model, which predicts 14 heavy rainfall events in the SSP2–4.5 scenario. The simulation results of the MPI-ESM1-2-HR model also indicate that heavy rainfall events will continue to occur in the future but with a lower frequency than historical rainfall. These differences suggest that the prediction of future rainfall may be subject to many uncertainties.
This prediction indicates that there may be more extreme rainfall events in the future, which could negatively impact the ecological environment, urban infrastructure, and human society. Therefore, measures should be taken to address climate change, including reducing greenhouse gas emissions, strengthening natural disaster prevention and adaptation, etc.
3.4.2. Daily Rainfall
On the basis of the results of the three climate models selected in the previous section and the revised results of the downscaling under the three emission scenarios, nine different sets of 24 h rainfall series were generated in Beijing urban areas using the distribution coefficients of the rainfall time series as input conditions for the hydrological model of the region. As illustrated in
Figure 8 below.
Compared to the total daily rainfall of the status quo, the three predicted results of future models for a 2-year rainfall of 24 h duration have a total rainfall of approximately 125–170 mm, while the uncertainty is greater for a 50-year rainfall with a total rainfall of approximately 270–380 mm. Consequently, rainfall uncertainty grows as the return period increases. In contrast, the effect of emission scenarios on total precipitation is less than the return period.
The models GFDL-ESM4 and EC-Earth3 both predict higher future rainfall than the status quo total daily rainfall, with the amount of growth being approximately 40%. The model MPI-ESM1-2-HR predicts a lower trend for total future daily rainfall than GFDL-ESM4 and EC-Earth3. Nonetheless, the predicted values for the SSP2–4.5 and SSP5–8.5 scenarios are still close to the status quo and slightly lower. However, the predicted values for the SSP2–4.5 and SSP5–8.5 scenarios are still close to and slightly higher than the status quo values, indicating that the predicted total daily precipitation under future climate change conditions is generally higher than the status quo; i.e., there will be more precipitation in the future.
In general, the results of the experiments on future rainfall fit well with previous studies. For instance, Wu Jian et al. [
47] predicted future precipitation in the Yangtze River basin in China and determined that future precipitation and temperature in the Yangtze River basin will be higher than in historical periods under all scenarios. Xiang Junwen et al. [
26] modeled future precipitation in major regions of China and demonstrated that annual and extreme precipitation in future periods in China will increase significantly. The rate of annual and extreme precipitation increases progressively as the scenario evolves.
3.5. Estimation of Waterlogging Situation under Future Scenarios
3.5.1. Simulation of Surface Runoff under Future Scenarios
Three scenarios with varied discharges of SSP126, SSP245, and SSP585 result in runoff quantities that are distinct among the three models for the same return period. As shown in
Figure 9, GFDL-ESM4 generates more runoff with an increasing discharge without significant changes in total rainfall, with an increase of approximately 1%. Model EC-Earth3 generates approximately 1% less runoff with increasing discharge in sequence for the SSP126, SSP245, and SSP585. Finally, MPI-ESM1-2-HR generates approximately 1% less runoff with increasing discharge in sequence from scenario SSP126 to SSP245.
Figure 10 depicts the peak runoff for each catchment sub-area at various return periods and under various scenarios. It can be seen that the areas of high peak runoff in the region are concentrated in a few catchment sub-areas, whereas peak runoff is negligible for the majority of catchment sub-areas. In general, the peak runoff increases dramatically with the return period, with the 50-year return period in the highest peak catchment areas being around twice as high as the 2-year return period. In certain regions with low peaks, the multipliers are greater. This means that in future climate scenarios, runoff volumes from heavy and even extreme rainfall will have much higher peaks, and economic and safety costs resulting from waterlogging will increase exponentially.
Those catchment sub-areas with low peak flows may be more vulnerable to severe and intense rainfall because they cannot handle significant volumes of runoff in a relatively short period of time. Thus, urban planning and construction should take into account the peak and return periods of runoff from these various places, and steps should be taken to decrease the risk of flooding. Green infrastructure, for instance, can be utilized to increase the infiltration capacity of surface water and lower the rate of runoff, and approaches such as green roofs on buildings and rainwater collecting systems can be employed to reduce runoff generation. In addition, plans for disaster prevention and control, as well as contingency plans, can be prepared in order to respond swiftly and minimize damage in the event of flooding. Future climate scenarios will raise the risk of internal flooding, necessitating a variety of preventative measures to maintain the safe and sustainable development of metropolitan areas.
3.5.2. Simulation of Node Overflows in Future Scenarios
Using the SSP2–4.5 scenario as an example, the historical measured data under the four return periods of 50 years, 10 years, 5 years, and 2 years were calculated and analyzed to produce a line graph of the rainfall trend. The line graphs,
Figure 11, show that the rainfall trends for the four return periods are similar, with a tiny peak of around 5 mm 4–5 h following the beginning of the rainfall, followed by a high of approximately 20 mm after 1020 min.
At 2 yr and 5 yr return periods, the simulated data of all three models are lower than the measured historical data, with significant variances at higher precipitation levels, when the model GFDL-ESM4 is closest to the historical data. When the return period is raised to 50 years, the simulated precipitation values of MPI-ESM1-2-HR are most similar to actual values.
According to the “Outdoor Drainage Design Standard” (GB 50014-2021), the standard design return period for rainwater drains in large urban areas is between three and five years. In this study, the 24 h rainfall process is simulated with a 5-year return period, and it is known from the analysis of a single rainfall event that the peak rainfall occurs at 17:00 and the overflow occurs at 17:05, which is when the most severe overflow is observed at the single-day node.
Figure 12 below illustrates the node overflows for each scenario.
The interface of the SWMM model for nodal overflows at 17:05 is depicted in
Figure 12a, with white nodes indicating no overflows and red nodes showing overflows for which severity increases as the color intensity increases. Thirteen nodes overflow under the future emission scenarios of SSP1–2.6 compared to the status quo, of which seven nodes overflow due to future climate change, which would not overflow under the status quo, and are therefore referred to as potential overflow points (red circles in
Figure 12b), representing approximately 1.5% of all nodes in the study area.
Figure 12c,d illustrate that the SSP2–4.5 and SSP5–8.5 scenarios generate 13 and 14 potential overflow sites, with a probability of 2.7% and 2.9% of potential overflow points, respectively. As the climate scenario worsens, it is evident that nodal overflows will become increasingly severe. This condition can have severe effects on the urban environment, such as generating flooding, polluting water sources, and disrupting traffic operations. Consequently, appropriate efforts must be made to improve the management and maintenance of urban drainage systems to increase their ability to withstand extreme climate scenarios.
The overall examination of the three models demonstrates that the effect of increasing scenario emissions on nodal overflows is less than that of the return period. The effect of scenario emissions on overflow start time, overflow duration, and overflow volume is more stable for both GFDL-ESM4 and EC-Earth3. For GFDL-ESM4, an increase in scenario emissions worsens the degree of node overflow, i.e., earlier overflow start times, increased overflow duration, and increased overflow volume, whereas, for EC-Earth3, the effect slows down. The MPI-ESM1-2-HR model is unstable in relation to overflows caused by higher scenario emissions, exhibiting a tendency to increase and subsequently drop.
The simulation findings for the three modes indicate that GFDL-ESM4 and EC-Earth3 are relatively similar, with an overflow volume difference of less than 1 × 105 ltr for the same scenario and return period. MPI-ESM1-2-HR exhibits a lower amount of overflow in the case region than GFDL-ESM4 and EC-Earth3, and there are instances in which GFDL-ESM4 and EC-Earth3 exhibit overflow but MPI-ESM1-2-HR does not for the identical return period and emission scenario.
In general, node overflows are greater in future scenarios than in present ones, and the more severe the scenario emissions, the greater the node overflows will be. This suggests that urban rainfall under future climate change conditions will result in more nodal overflows and that the corresponding level of urban waterlogging will also increase. This is consistent with the findings of previous researchers, such as K. X. Yang [
14], who predicted that flooding in Haining, China, would be worse than the status quo over the next decade using the LULC prediction model, and X. T. Wang [
48], who discovered that the depth of flooding in Shanghai is anticipated to increase as the return period increases. In addition, the growth in node overflows suggests that the present gray stormwater infrastructure is unable to battle future climate change rainfall scenarios and that design standards must be upgraded to some extent, such as by increasing the diameter and depth of stormwater pipes or by the introduction of green infrastructures.