Assessment of the impact of climate change on urban flooding: A case study of Beijing, China

Urban flooding is one of the major challenges facing the world in the 21st century. Future flood risk is exacerbated by climate change, urbanization and ageing infrastructure (O'Donnell & Thorne 2020). Climate change has added heavier rainstorms, as well as severe and frequent floods that are hard to predict (Tingsanchali 2012). Affected by intensified urbanization, population growth and ageing of urban drainage systems, the risk of urban flooding is increasingly affected by climate change (Zeng et al. 2021). The design of a drainage system is usually based on the statistical data of historical precipitation in a certain period, without considering the potential change of extreme precipitation value in the design return period (Zhou et al. 2018). Arnbjerg-Nielsen (2012) indicated that the design storm intensity in Denmark is expected to increase by 10–50% over a return period of 2–100 years. The recent Intergovernmental Panel on Climate Change (IPCC) special report on the impact of 1.5 °C of global warming estimated that global warming could reach 1.5 °C between 2030 and 2052. The final reflection of the climate system will be in the form of an increase in the intensity and frequency of extreme precipitation. In China, flood disasters have occurred in many cities and caused huge economic losses. For example, the rainstorm in Beijing in July 2012 caused 79 deaths and a loss of $1.86 billion. The rainstorm in Shanghai on September 13th, 2013 caused huge direct impact and indirect losses. The rainstorm in Wuhan on July 23, 2015 caused many places in the city to be flooded, and its maximum rainfall exceeded 100 mm. During the ‘7.20’ extreme rainstorm in Zhengzhou, serious mountain torrents and waterlogging disasters occurred in the city, resulting in road damage, traffic disruption and dam break of a large number of reservoirs. Therefore, it is very important to understand the future climate change for urban flood disaster management and the design of flood control facilities.

The research on the drainage capacity of the urban drainage system adopts the historical precipitation process and rarely considers the impact of climate change on the intensity of short-duration rainstorms. For example, Hou et al. (2020) adopted the designed storm intensity formula of the Xixian New Area to study the effects of different return periods, peak coefficients and durations on flood inundation. Palla & Gnecco (2015) studied the effects of low impact development (LID) as source control measures on hydrological processes in urban catchments under different precipitation return periods. Wang et al. (2019) studied the changes of urban flood resilience under different return periods. The rainstorm intensity formula is an important basis for reflecting the rainfall pattern, guiding the design of urban drainage and waterlogging prevention projects and the construction of related facilities. If the impact of climate change on the intensity–duration–frequency (IDF) curve can be considered in the design of the drainage system, the drainage system will be able to accommodate greater extreme precipitation, and that will increase its reliability and stability (Kourtis & Tsihrintzis 2022; Kourtis et al. 2022).

The current way to quantify climate change is to use the future precipitation predicted by global climate models (GCMs) (Shen et al. 2018). Climate data published by the Coupled Model Intercomparison Project (CMIP) is widely adopted (Touzé Peiffer et al. 2020). The most recent CMIP6 has 55 GCMs (Iqbal et al. 2021). Although CMIP6 can reflect the changes in the precipitation in the future, these data also have great uncertainty. Numerous studies indicated that CMIP6 overestimates or underestimates future precipitation. For example, Iqbal et al. (2021) indicated that some models of CMIP6 have large deviations of 25–75% for annual precipitation in northern Mainland South-East Asia and show an underestimation of −25 to −50% in coastal areas. CMIP6 can be used to assess the impact of climate change after bias correction. Therefore, CMIP6 needs to be bias corrected before use. Most of the researches are about the correction of the monthly precipitation. The monthly precipitation cannot meet the requirements of the drainage model for the resolution of the precipitation. The resolution of the precipitation must be daily or less.

In order to solve the problems mentioned in the above analysis, the objectives of this study are as follows: (1) evaluating the simulation accuracy of CMIP6 on daily precipitation; (2) proposing a new statistical downscaling method for correction of daily precipitation; (3) the response of this urban drainage system to different design storm scenarios in the future is analyzed; (4) the effect of LID on rainfall runoff was analyzed under different precipitation conditions to provide guidance for flood risk management in small-scale urban developments.

In this study, the daily precipitation from 1951 to 2012 for observation are from China meteorological forcing dataset (1979–2018) (Yang et al. 2010; Kun & Jie 2019; He et al. 2020) (http://data.tpdc.ac.cn/). The name and code of the weather station are ‘Beijing’ and ‘54,511’. The distance between the weather station and the study area is 6.5 km. The daily precipitation data of CMIP6 in the historical period from 1951 to 2012 is used to evaluate the simulation accuracy. The model with the best accuracy will be used for bias correction. Daily precipitation for the future period of CMIP6 is adopted to assess climate change from 2023 to 2062 (https://esgf-node.llnl.gov/search/cmip6/). Compared with CMIP5, CMIP6 has made great improvements in the parameterization scheme of dynamics and resolution of the model, and most of the models adopt the bidirectional coupling of atmosphere and chemistry (Eyring et al. 2016). In the design of future scenarios, compared with CMIP5, CMIP6 adopts Scenario Model Intercomparison Project (ScenarioMIP), which is produced with integrated assessment models (IAMs) based on new future pathways of societal development, the Shared Socioeconomic Pathways (SSPs) and related to the RCPs (O'Neill et al. 2016). CMIP6 applies a new combined scenario (SSP-RCP) and increases three new emissions paths (RCP1.9, RCP3.4, RCP7.0). These scenarios take into account changes in the global economy and population, as well as emissions of greenhouse gas. Two scenarios for SSP Tier 1 are adopted in this study, which are, respectively, SSP2-4.5 and SSP5-8.5. SSP2-4.5 represents the medium part of the range of future forcing pathways and updates the RCP4.5 pathway. SSP5-8.5 represents the high end of the range of future pathways and updates the RCP8.5 pathway. Nominal Resolution of CMIP6 is 100 and 250 km, Variant Label of which is ‘r1i1p1f1’. The spatial resolution and land model of each GCMs are presented in Table 1.

Due to uncertainties of the simulation for CMIP6 in different regions, it is necessary to evaluate the applicability of the data before use. The evaluation methods used in this study include correlation coefficient (CC) (Ding et al. 2019), percent bias (PBIAS), root mean square error (RMSE), coefficient of variation (CV), Mann–Kendall Trend Test (MK) (Ding et al. 2019), Sen's slope estimator (SSE), Taylor diagram (Taylor 2005) and extreme precipitation index (ETCCDI) (Ding et al. 2019). The ETCCDI indices used in this study are shown in Table 2. In Table 2, RR denotes daily precipitation and PRCP denotes precipitation.

The range of CC is [−1,1]. The closer CC is to 1, the closer PBIAS and RMSE values are to 0 and the closer TSS and CRI values are to 1, indicating that the model performs better.

The model is interpolated to the weather station by bilinear interpolation. Since most models adopt the calendar as ‘noleap’, leap years are not considered, the study averaged the precipitation on February 28 and March 1 in leap years to interpolate the data on February 29, so that the daily data of model are equal to the number of observations. Many studies have pointed out that the performance of Multiple Model Ensembles (MMEs) is better than that of a single model, which can reduce the PBIAS. Therefore, this paper adopts the Equal Weight Ensemble Mean (EWEM) (Cos et al. 2022) in MMEs to obtain EWEM data. The purpose of the processing is to compare with a single model, evaluate the simulation effect of the ensemble model on precipitation and obtain models with higher precision.

The correction effect of the above methods on monthly precipitation is better than that of daily precipitation. Most areas in China have less precipitation from November to March of the following year, especially from December to February, the daily precipitation of most sites is zero. For the QDM method, when the daily precipitation of the model in the historical period is zero, the proportional coefficient is inf or NaN. When the corrected value of EDCDF is less than zero, if the absolute value of which is greater than the precipitation of the future period, the corrected value will be negative. If the empirical CDF is used for QM, frequent interpolation and extrapolation are required, which is not satisfactory (Li et al. 2010). The parametric distribution function is used to fit the empirical CDF of daily precipitation. It is different from the fitting of extreme values. After fitting with different distribution functions, it is found that the accuracy of the Gamma distribution fitting is the best, but the Gamma distribution requires that the sample is greater than zero. A large number of zeros are included in the daily precipitation, so that the parameters of the theoretical distribution cannot be calculated, and only the moment method can be used, but the fitting accuracy of which is lower. Many studies firstly correct the monthly precipitation and then correct the daily precipitation proportionally according to the ratio of the monthly precipitation before and after the correction. This method will cause the time and position of the extreme value of the daily precipitation to be changed, which increases the error. Therefore, in order to solve the correction problem of modeled daily precipitation, this paper proposes a QDM method of daily precipitation based on frequency (DFQDM). Firstly, the frequency of wet days with daily precipitation from January to December is corrected. Secondly, the 95th percentile of daily precipitation in each month is calculated, and the daily precipitation is divided into two sections according to the percentile value. Based on the mixed Gamma distribution, the cumulative distribution of each segment is calculated. Then the quantile delta map is adopted to correct the daily precipitation. Finally, the corrected precipitation is synthesized into a complete time series. The specific steps of DFQDM method are as follows:

The simulated performance of different GCMs for annual precipitation was first evaluated in 1951–2012. It can be seen from Table 3 that the MEAN of annual precipitation is overestimated and underestimated by GCMs. The absolute error (AE) between MEAN of CanESM5 and that of the observation is the smallest, which is 43.48 mm. The AE of CMCC-CM2-SR5 and TaiESM1 are larger, which are 457.37 and 430.16 mm, respectively, and CMCC-CM2-SR5 is nearly double the observation. The PBIAS and RMSE of CanESM5 are the smallest, which are −7.33% and 239.66 mm, respectively. The PBIAS of CMCC-CM2-SR5 is 77.16%, which is the largest and severely overestimated precipitation. The CC of GCMs is all close to 0, which indicates that the model has a poor simulation of the changing trend of annual precipitation. The CV of the observation is 0.35. The CV of MRI-ESM2-0 is 0.31, which is close to the observation, and the CV of TaiESM1 is far from the observation, which is 0.17. For the MK, the observation shows a downward trend, the MK and SSE of ACCESS-ESM1-5 are close to the observation, MK and SSE of which are −1.45 and −1.57, respectively. CMCC-ESM2, MPI-ESM1-2-HR, FGOALS-g3 and MIROC6 showed a large opposite trend compared with the observation. In summary, CanESM5 performs well for the simulation of annual precipitation, and its error is the lowest and the change of trend is more consistent with the observation.

For annual precipitation, the PBIAS of GCMs is positive or negative. If overestimated and underestimated models are ensemble averaged, the PBIAS for the precipitation will become very low. To obtain a better simulation of precipitation, GCMs were processed by the EWEM in this study. In Table 3, GCMs processed by EWWM are No.1 and No.5, No.1 and No.7, No.3 and No.5, No.3 and No.7, No.4 and No.11, No.5 and No.12, No.6 and No.11, No.7 and No.12, No.10 and No.11, respectively. The ensemble GCMs is named by appending the number of model to EWEM. Table 4 shows the error of annual precipitation of EWEM. PBIAS of all models is ensembled as EWEMALL. Affected by the overestimation of precipitation in most models, the mean of EWEMALL is quite different from the observation. EWEM0107 and EWEM1011 have the lowest PBIAS, which are, respectively, 0.1 and 0.86%; For CC, CV, MK and SSE, EWEM1011 performs best, and the change trends of other ensemble models are opposite to the observation.

Table 5 shows the evaluation results of monthly precipitation. The AE between ACCESS-ESM1-5, CanESM5, EWEM0105, EWEM0107, EWEM0305, EWEM0307, EWEM0411, EWEM0512, EWEM0611, EWEM0712, EWEM1011 and the mean of the observation are the smallest, which are, respectively, 4.48, −3.62, 2.19, 0.05, 4.33, 2.19, 1.49, 3.02, 1.52, 0.88, and 0.43 mm. The AE of EWEM0107 is the smallest, and the PBIAS of the above models are also the smallest. The RMSE of EWEMALL is the smallest, which is 53.51 mm. The RMSE of CMCC-CM2-SR5 is the largest, which is 90.37 mm. The CC of EWEMALL is the largest, which is 0.74, and that of CanESM5 is 0.59. The CC of the model is around 0.65. The variation of the model for monthly precipitation is consistent with the observation. For the CV, the variation of MRI-ESM2-0 relative to the overall mean is the closest to the observed value; the SSE of all models is 0, which indicates that the monthly precipitation does not have a significant trend of change.

The models with lower CRMSE are, respectively, EWEM0105 and EWEM0512 for October and November.

The models that perform better for SD are ACCESS-CM2, EWEM0305, MRI-ESM2-0, EWEM0512, EWEM0712, MPI-ESM1-2-HR, ACCESS-ESM1-5, CESM2-WACCM, CMCC-ESM2, MPI-ESM1-2-HR, EWEM0611 and NorESM2-MM. The models with larger TSS are ACCESS-ESM1-5, EWEM0105, EWEM0107, TaiESM1, TaiESM1, CMCC-ESM2, TaiESM1, CMCC-ESM2, CMCC-ESM2, CESM2-WACCM, MIROC6 and NorESM2-MM.

To evaluate the correction effect of DFQDM on daily precipitation, this study selected nine extreme precipitation indices of ETCCDI as evaluation indicators. The changes in indices before and after the correction are shown in Tables 6 and 7. Except for RX5day, the correction effect of other indices is significant. The PBIAS of CWD is significantly reduced, which indicates that the correction of the frequency of wet and dry days by DFQDM is reasonable. After the correction, the PBIAS of the maximum continuous precipitation days and the observation are significantly reduced. It can be seen from RX1day, RX5day, SDII, R95PTOT that DFQDM has a better effect on the correction of the intensity of extreme precipitation. For RX1day, the corrected PBIAS of all models excepting NorESM2-MM and TaiESM1 are controlled within 6%. For RX5day, the PBIAS of most models is controlled at around 10%, SDII represents the intensity of mean precipitation for wet days, and is underestimated in many models. The reason may be that CDF fitted by the mixed Gamma distribution underestimates the observation at the same frequency. The study will further improve the distribution function in the future. The absolute value of PBIAS of SDII is controlled within 10%. The PRCPTOT is controlled within 5%, which indicates that DFQDM has a good correction effect on the daily precipitation and the frequency of wet days. This method can be used to correct the climate model. In summary, by comparing the accuracy of the model before and after correction, the corrected CanESM5 is selected in the study to reflect the change in urban future precipitation.

This rainfall process is added to SWMM to simulate the changes in total flood volume (TFV) of nodes and drainage capacity of conduits. The simulation time of SWMM is 120 min. This model has been calibrated and can accurately describe the change of water level in the study area. The calibrated parameters of this model are shown in Table 8.

The drainage system in the study area was insufficient to cope with rainfall with a return period of more than 10-years in the historical period, and the PON was more than 30%. Under the SSP2-4.5 scenario, the PON has reached 32.48% during the 5-year return period. Under the SSP5-8.5 scenario, the PON has reached 31.15% during the 3-year return period. The area of hardened pavement in this study area has a high proportion, which leads to a decrease in the infiltration capacity of rainfall. A large amount of rainwater is drained into rainfall wells, which increases the load on the conduit operation. In the future, the study area should transform the conduit network with a smaller diameter in low-lying areas. However, expanding existing drainage systems has proven expensive and unsustainable, especially in changing environments (Zhu et al. 2019). LID can effectively reduce the impact of the decline of surface permeability caused by urban development and has been adopted in many countries (Yu et al. 2022). At the same time, LID measures should be constructed to increase the interception rate of rainwater at this initial stage. This overflow will be effectively mitigated. LID can reduce the pollutants produced by the scouring of the initial rainwater on the road surface and improve the natural landscape. LID also has some disadvantages, such as (1) increase in the cost of maintenance; (2) if the design is unreasonable, it will reduce the removal effect of pollutants; (3) with the increase of time, the performance of LID will decrease.

In order to reduce the risk of flooding in the study area, the LID was applied and its effect on the control of precipitation at the source was evaluated without considering the reduction of performance. Based on the topographic features and future planning of this study area, five LIDs were adopted in the study area. The LIDs are, respectively, Permeable Pavement (PP), Bio-retention Cells (BRC), Green Roofs (GR), Rain Gardens (RG), Vegetative Swales (VS). In SWMM, LID is represented its coverage by setting the number and area in the subcatchment. The proportion of LID-controlled area for different land uses in the subcatchment is shown in Table 9.

The simulation of an urban drainage system requires high resolution of precipitation for the climate model, the correction of daily precipitation is a problem to be solved. Considering the correction of frequency and bias of daily precipitation, a new statistical downscaling method called DFQDM is proposed. CMIP6 is used to calculate the rainstorm intensity under different return periods in the future. The impact of climate change and LID construction on urban flooding is analyzed. The main conclusions of this study are as follows:

• (1)

  For the annual precipitation, CanESM5 has better accuracy for the prediction of precipitation, and the minimum absolute error between CanESM5 and the mean value of observation value is 43.48 mm. The PBIAS and RMSE of CanESM5 are also the smallest, which are −7.33% and 239.66 mm, respectively, and the change trend is more consistent with the observed value. CMCC-CM2-SR5 seriously overestimates the annual precipitation. This ensemble model performs well except for EWEMALL. For the monthly precipitation, the CC of the climate model is about 0.65, which is close to the variation trend of the observation.

• (2)

  The correction effect of DFQDM on daily precipitation is significant. The PBIAS for annual precipitation was significantly reduced. The PBIAS of this daily extreme precipitation index are all controlled within 6%, and the PBIAS of RX5day for most climate models are controlled around 10%. SDII of many climate models is underestimated, and the correction of the average precipitation intensity on this wet day will be improved in the future research. Finally, CanESM5 is selected as the precipitation data for future climate change analysis.

• (3)

  The simulation results of this SWMM show that the relationship between TFV and return period follows a logarithmic curve. The TFV of SSP5-8.5 and SSP2-4.5 increased by 236.59 and 85.1% in the 2-year return period and 45.43 and 20.8% in the 100-year return period, respectively. In different return periods, this LID reduces the runoff of this drainage system by more than 60%, and its control effect on waterlogging risk is significant.

This study was supported by the National Natural Science Foundation of China (No. 52179027).

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.