Impacts of climate change on streamflow of Qinglong River, China Open Access

BASINS

better assessment science integrating point and non-point sources

CDF

cumulative distribution function

CMIP5

Coupled Model Intercomparison Project 5

DEM

digital elevation model

GCM

global climate model

GenScn

GENeration and analysis of model simulation SCenNarios

GHG

greenhouse gas

HEC-HMS

Hydrologic Engineering Center–Hydrologic Modeling System

HSPF

Hydrological Simulation Program-Fortran

IPCC

Intergovernmental Panel on Climate Change

MME

multi-model ensemble

NASA

National Aeronautics and Space Administration

NIMA

National Mapping Agency

NSE

Nash–Sutcliffe efficiency coefficient

PEST

Parameter estimation

QM

quantile mapping

RCP 2.6

Representative Concentration Pathway 2.6

RCP 4.5

Representative Concentration Pathway 4.5

RCP 6.0

Representative Concentration Pathway 6.0

RCP 8.5

Representative Concentration Pathway 8.5

R

correlation coefficient

RMSE

root mean square error

SRTM

Shuttle Radar Topography Mission

S-value

the Taylor skill scores

SWAT

Soil and Water Assessment Tool

TF

transfer function

T_max

daily maximum temperature

T_min

daily maximum temperature

USEPA

United States Environmental Protection Agency

The global surface temperature between 2011 and 2020 was 1.09 °C higher than the period from 1850 to 1900. It is anticipated that by the mid-2030s, the temperature rise will reach or exceed 1.5 °C (Lee et al. 2021; Intergovernmental Panel on Climate Change 2022). In this context, anthropogenic warming is accelerating the hydrological cycle (Huntington 2006; Wang et al. 2018; Zhang et al. 2019; Su & Sun 2021), leading to changes in global precipitation patterns. These alterations are making precipitation distribution more extreme and influencing the frequency of hydrological extreme events (Kundzewicz 2008; Putnam & Broecker 2017; Manfreda et al. 2018; Duan & Duan 2020). The fifth report of the United Nations Intergovernmental Panel on Climate Change (IPCC) estimated greenhouse gas (GHG) concentrations for the next nearly 100 years, converting them into increased radiative forcing (W/m²). Various agencies collaborated to develop four emission scenarios (Moss et al. 2010; van Vuuren et al. 2011), namely the Representative Concentration Pathway 2.6 (RCP 2.6), the Representative Concentration Pathway 4.5 (RCP 4.5), the Representative Concentration Pathway 6.0 (RCP 6.0), and the Representative Concentration Pathway 8.5 (RCP 8.5) (Intergovernmental Panel on Climate Change 2015). Applying these climate change scenarios provided climate models with a precise platform to visually analyze the direct impacts of climate change on hydrological water resources (Thomson et al. 2011; Auer et al. 2021).

The above-mentioned climate change scenarios are widely employed to investigate the impact of climate change on streamflow. Moazami Goudarzi et al. predicted higher future temperatures in the Maharu basin under scenarios RCP 4.5 and RCP 8.5, noting more significant precipitation changes in the wet season than in the dry season. Their results, obtained using the Soil and Water Assessment Tool (SWAT), revealed a year-by-year increase in annual streamflow for the watershed under both RCP4.5 and RCP8.5 scenarios (Moazami Goudarzi et al. 2020). Zhao et al. predicted the characteristics of streamflow changes in the Xijiang River Basin under RCP4.5 and RCP8.5 scenarios. Their findings indicated that compared with the baseline phase, the future intra-annual streamflow distribution would become more homogeneous under the RCP4.5 scenario and more heterogeneous under the RCP8.5 scenario (Zhao et al. 2020). Badora et al. assessed future hydrological changes in the Bystra River upland catchment using the SWAT model based on data from regional climate models. They discovered that the future monthly average total streamflow in the watershed under most climate change scenarios would be lower than in the reference period (Badora et al. 2022). Shafkat Ahsan et al. evaluated the future streamflow changes in the Jhelum basin using SWAT in combination with several GCMs. The results indicated reductions in watershed streamflow of approximately 23–37% under the RCP4.5 scenario and 19–46% under the RCP8.5 scenario (Ahsan et al. 2022). Danyang Gao et al. predicted streamflow changes in the Yangtze River under scenarios RCP2.6, RCP4.5, and RCP8.5. The results demonstrated that the streamflow would increase by 9.19, 10.23, and 13.44% during 2041–2100 under the three scenarios, respectively (Gao et al. 2020). Qihui Chen et al. found that annual streamflow in the Jinsha River Basin is expected to increase from 2017 to 2050 under RCP4.5 and RCP8.5 scenarios (Chen et al. 2020). Most studies were conducted using a combination of GCMs and SWAT models. However, some studies lack reasonably corrected climate data for GCMs, and the number of GCMs employed was limited. Additionally, many studies predicting future streamflow lack diversity in temporal scale, often focusing on only one or two scales. Consequently, these studies lack a comprehensive analysis of climate factors at different time scales. Hydrological models like SWAT encounter challenges in simulating streamflow on smaller time scales.

Considering these points, this study integrates 31 GCMs and the HSPF model to analyze changes in streamflow under the RCP4.5 and RCP8.5 scenarios. The HSPF model facilitates the simulation of hourly-scale and daily-scale streamflow (Choi et al. 2008). The analysis encompasses streamflow results at various scales, including hourly, daily, monthly, seasonal, and annual scales. Given the variability in simulated streamflow output among different GCMs (Koirala et al. 2014), the study specifically examines extreme values and the multi-model ensemble (MME) mean of the 31 GCMs across different time scales. To enhance the accuracy of climate model data, quantile mapping (QM) methods are employed to mitigate prediction errors.

In this study, we focused on the Qinglong River, situated in Hebei and Liaoning provinces in Northern China, to investigate its future water regulation and management (refer to section 2.1 for details). The impact of various climate factors on watershed streamflow varies depending on the actual situation of different watersheds (Tan et al. 2017; Wang et al. 2020). The objectives of this research are to predict variations in the Qinglong River streamflow from the present to the year 2100 across multiple temporal scales, including annual averages, monthly averages, daily peaks, and hourly peaks, under different climate change scenarios (RCP4.5 and RCP8.5). Utilizing the predicted data, such as precipitation and streamflows, we aimed to forecast and assess the conditions of water resources, water supply, and river flooding.

The Qinglong River watershed is situated in the East Asian monsoon climate zone, characterized by a multi-year average temperature of 10.1 °C. The lowest temperatures are typically recorded in January, with a monthly average of −6.8 °C and a minimum of −29.2 °C. Conversely, the highest temperatures occur in July, with a monthly average of 24.7 °C and a maximum of 39.4 °C. The watershed experiences an average annual water surface evaporation of 1,089 mm, with the maximum monthly evaporation reaching 239.6 mm. Annual precipitation in the watershed averages between 500 and 700 mm, with 70% of this concentrated in July and August. Consequently, 80% of the annual streamflow and river flooding typically occur during this period, posing a significant concern for watershed management.

According to the flowchart of this study, several methods need to be emphasized here for introduction.

In this study, the predicted downscaling meteorological data (e.g. precipitation, temperature) were acquired from CMIP 5. Systematic deviations in climate model simulations compared to observations can significantly influence streamflow simulation results. Consequently, it is necessary to correct errors in climate model simulation results. Here, the observed precipitation data for the Qinglong River from 1961 to 2020 were utilized to calibrate the predicted precipitation data.

In this study, the QM method was selected for calibrating the downscaling precipitation data from climate models. The QM method offers advantages in non-independent error corrections, capable of revising both numerical and frequency distributions of precipitation. This approach not only corrects the mean value and inter-annual variability of the predicted data but also mitigates bias in the predicted data for extreme events. QM is a calibration method based on frequency distributions, primarily achieved through calculating the cumulative distribution function and the transfer function (TF). There are two primary methods for establishing TFs. The first involves constructing a TF based on the theoretical probability distribution function, assuming that the simulated data and measured data adhere to a known probability distribution function (Piani et al. 2010). The second method constructs a TF based on an empirical probability distribution (Themeßl et al. 2012).

In the formula, represents the average observed flow during the simulation period. The closer the RSR value is to 0, the closer the simulated value is to the actual measured value.

In this study, the HSPF model was selected to simulate and predict the future flows of the Qinglong River. HSPF is a semi-distributed watershed model developed by the U.S. Environmental Protection Agency (USEPA) in 1980 (Kinerson et al. 2009). It has the capability to model the hydrology and water quality of a watershed. HSPF has been integrated into the Better Assessment Science Integrating Point and Non-point Sources (BASINS) system, becoming the core watershed model in BASINS (Donigian & Imhoff 2006). The HSPF model comprises three main application components. PERLND, IMPLND, and RCHRES correspond to Pervious Land Segment, Impervious Land Segment, and free-flowing reach or mixed reservoirs, respectively. Through these application components, the HSPF model can simulate various processes, including streamflow, soil loss, pollutant transport, river hydraulics, and others. The HSPF model combines the advantages of both distributed and centralized hydrological models, offering the ability to autonomously adjust the size of the hydrological response unit, thereby providing greater flexibility and improving the accuracy of watershed hydrological simulations (Liu & Tong 2011; Albek et al. 2019). This makes it well suited for the process simulation of regional watershed hydrology.

When utilizing HSPF to simulate streamflow, the initial steps are carried out within the BASINS system. In this system, the watershed model is first established using river network data, meteorological data, measured streamflow data, and other relevant information. Meteorological data and measured streamflow data are subsequently converted into WDM files using the WDMtil module. Following these procedures, the data are transferred to the HSPF model within the BASINS system for calculation and simulation. During this transfer process, the UCI file, which serves as the project file for HSPF, is automatically generated. Throughout the HSPF model, UCI files can be edited, run, and viewed, encompassing aspects such as the temporal range of the simulated data, settings of model parameters, and more.

The calibration of the HSPF model primarily involves adjusting relevant parameters in the PERLND, IMPLND, and RCHRES modules. Following HSPF simulation, the outputs can be processed using the post-processing tool GENeration and analysis of model simulation SCenNarios (GenScn) for visualization. GenScn facilitates the analysis and collation of simulation results (Kittle 1998). Calibration and validation of the HSPF model entail parameter tuning to attain optimal values, and this optimal parameter set is subsequently used for further simulations.

In this study, the sensitivity of HSPF model parameters was initially analyzed using the built-in parameter estimation (PEST) auto-calibration procedure. Eleven model parameters were ultimately selected for calibration based on sensitivity analysis results and insights from relevant studies (Zhang et al. 2018; Liu et al. 2021), as detailed in Table 1. Notably, the Support Vector Machine (SVM) surrogate model and general likelihood uncertainty estimation algorithm were employed to determine the optimized intervals and values of HSPF parameters. The daily flow series at the entrance section of the Taolinkou Reservoir from 2011 to 2013 served as the measured flows for calibration and validation.

In this study, predictions were made for precipitation, temperature, and the corresponding daily simulated HSPF streamflows from 31 climate models. Subsequently, annual and monthly average parameters (precipitation, temperatures, streamflow volume) were synthesized using the MME means method (Ahmed et al. 2020; Tegegne et al. 2020). It is important to note that the maximum and minimum values of parameters (precipitation, streamflow volume, streamflow) were calculated based on the results of a single climate model and each HSPF simulation.

Furthermore, the baseline phase was set as the period from 2001 to 2020 (P0), and the prediction phase was divided into short-term (P1, 2021–2030), medium-term (P2, 2031–2050), and long-term (P3, 2051–2100). The prediction results were subjected to linear fitting using the least squares method to identify their trends.

For HSPF simulation, various watershed data were essential, encompassing topographic data, land use data, river network data, meteorological data, and actual hydrological data.

The topographic data, sourced from the Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM), are a joint effort by the United States National Aeronautics and Space Administration (NASA) and the National Mapping Agency (NIMA) of the Department of Defense (Yang et al. 2011).

Land use data, inclusive of information on land use, soil types, and spatial distribution, were retrieved from the Geospatial Data Cloud Platform managed by the Computer Network Information Centre of the Chinese Academy of Sciences (http://www.gscloud.cn/). Data on the river network in the study watershed were acquired through the Hydrological Analysis module of ArcGIS software. Meteorological data were sourced from the China Meteorological Science Data Sharing Centre (http://data.cma.cn). This meteorological data included daily precipitation, temperature, relative humidity, evaporation, sunshine, wind speed, and other relevant information spanning from 1957 to 2020. The daily measured streamflow (2011–2013) and annual streamflow (2001–2020) in the studied watershed were obtained from the Taolinkou Reservoir Administration in Hebei province, China, as detailed in Table 2.

In this study, the initial predicted precipitation and temperature data were derived from the Coupled Model Intercomparison Project 5 (CMIP5) climate model dataset processed by the NWAI-Weather Generator (WG) statistical downscaling model (Liu & Zuo 2012). Following spatial and temporal downscaling, the precipitation and temperature data from CMIP5 underwent effective error correction (Li et al. 2019; Yao et al. 2020). The relevant data for 31 global climate models (GCMs) under two climate scenarios, RCP 4.5 and RCP 8.5, were obtained from the website (https://esgf-data.dkrz.de/search/cmip5-dkrz). The list of 31 GCMs utilized in this study is provided in Table 3.

The correction results for the initially predicted precipitation data of the 31 climate models are presented in Table 4. It is evident that the RMSE values of downscaled precipitation data from all 31 climate models showed improvement after correction. Notably, the average RMSEs of ACCESS1-0, ACCESS1-3, and BCC-CSM1-1 were reduced by more than 20%, leading to a reduction of errors in the predicted precipitation by more than 10%.

The improvement in NSE and the value of RSR is presented in Table 5. The NSE value for each model has increased by at least 0.2, and the RSR value is close to 0. This indicates that the QM method brings the simulated data closer to the observed data, significantly reducing the error.

The specific characteristics of the data, including skewness, coefficient of variation, and so on, are listed in the appendix. The daily precipitation under both the RCP4.5 and RCP8.5 scenarios demonstrates a distribution with an average kurtosis of 79.518 and 118.731, respectively, and skewness of 8.870 and 8.904, respectively. These values suggest that the data distribution is more peaked and skewed to the right compared to a normal distribution. Additionally, the coefficient of variation fluctuates around 4.4, pointing to substantial variability in the data, with values in the dataset being relatively dispersed.

Firstly, it is observed that the maximum daily precipitation under RCP4.5 and RCP8.5 scenarios is 41.87 and 48.85 mm, occurring on 18 July 2045 and 7 August 2044, respectively, as depicted in Figure 3 and detailed in Table 6. The maximum daily precipitation during the historical phase took place on 22 July 2012, measuring 65.53 mm. Secondly, the historical multi-year average daily precipitation is 0.93 mm. Future multi-year average daily precipitation is projected to increase by 111% (short-term), 135% (medium-term), and 137% (long-term) for RCP4.5 scenarios, and by 124% (short-term), 129% (medium-term), and 154% (long-term) for RCP8.5 scenarios.

Based on the results from 31 GCMs (refer to Tables 7 and 8), the maximum daily precipitation under the RCP4.5 scenario is 412.7 mm for the ACCESS1-3 model. For the RCP8.5 scenario, the maximum daily precipitation is 367.8 mm for the HadGEM2-AO model. Notably, the daily precipitation maxima for most GCMs under both the RCP4.5 and RCP8.5 scenarios occur in the long-term and surpass the historical maximum daily precipitation.

Monthly precipitation holds significant importance for water supply, particularly in agriculture. As depicted in Figures 4 and 5, the highest average seasonal precipitation during the historical period is observed in the summer (June–August) at 499.95 mm, while the lowest occurs in the winter (December–February) at 11.05 mm. Under the RCP4.5 scenario, the maximum average seasonal precipitation transpires in the summer (long-term) at 587.16 mm, marking a 17.44% increase from the historical maximum. The minimum value occurs in winter (short-term) at 14.19 mm, reflecting a 28.37% rise from the historical minimum. Similarly, under the RCP8.5 scenario, the peak average seasonal precipitation is in the summer (long-term) at 621.24 mm, indicating a 24.26% increase from the historical maximum. The minimum value occurs in winter (short-term) at 13.66 mm, representing a 23.60% increase from the historical minimum. There are fluctuating trends during precipitation in spring (March–May) and autumn (September–November).

The measured annual precipitation for the historical phase (2001–2020) and the modeled annual precipitation for the future phase (2021–2100) are depicted in Figure 6. To begin, the annual precipitation during the historical phase exhibits a trend rate of −7.64 mm/10a, with a multi-year average annual precipitation of 717 mm (Table 9). Additionally, under the RCP4.5 scenario, the maximum increase rate in future annual precipitation occurs in the medium-term, at 25.74 mm/10a. Under the RCP8.5 scenario, the maximum increase rate in future annual precipitation takes place in the long term, at 21.94 mm/10a (Table 9). The multi-year average annual precipitation under the RCP4.5 scenario experiences a 0.31% increase (short-term), 11.26% increase (medium-term), and 13.42% increase (long-term) compared to the historical period. Similarly, the multi-year average annual precipitation under the RCP8.5 scenario witnesses a 5.88% increase (short-term), 7.70% increase (medium-term), and 20.36% increase (long-term) compared to the historical period.

Firstly, compared to the average value of the baseline phase (2001–2020), the projected increases in T_max and T_min under the RCP4.5 scenario reach 1.67 and 1.46 °C, respectively, at the end of the century. Meanwhile, the increases in T_max and T_min under the RCP8.5 scenario reach 4.10 and 4.16 °C, respectively, by the end of the century (Figure 7). Furthermore, the results for specific changes in temperature over time are shown in Table 10. Moreover, the increasing rates of T_max and T_min gradually decrease during 2021–2100 under the RCP4.5 scenario, while they gradually increase under the RCP8.5 scenario. The increase rate of T_min is greater than that of T_max in both the medium-term and long-term. In general, T_max and T_min show an increasing trend during 2021–2100 under both scenarios, with T_min increasing more noticeably.

It is evident that the largest increase of 2.6 °C compared to the baseline phase is expected to occur in the summer of the P3 period under the RCP8.5 scenario for T_max. Overall, in both scenarios, T_max and T_min show a decrease and then an increase in spring (March to May). This is more pronounced in March, while T_min decreases and then increases in January and February during the winter (December to February) months, with T_max exhibiting a different pattern.

By analyzing the hourly and daily streamflow data generated by HSPF, we can assess the potential for future flash floods in the Qinglong River.

Initially, hourly streamflow projections for the period 2021–2100 under both RCP4.5 and RCP8.5 scenarios were obtained by configuring the HSPF output time scale to hourly intervals. After thorough calculation and comparison, the mean and maximum hourly streamflow values for 31 GCMs under the respective scenarios during different periods are presented in Tables 12 and 13. The MME mean results of the 31 GCMs are illustrated in Figure 11.

For the years 2021–2100, the maximum values of hourly streamflow peak at 402.1 and 289.9 million m³ under RCP4.5 and RCP8.5 scenarios, occurring on 18 July 2069 and 23 August 2088. Conversely, the minimum values of maximum hourly streamflow are recorded in the short term, at 43.7 and 38.9 million m³. Figure 11 displays the MME mean of hourly streamflow for the 31 GCMs, revealing that the maximum future hourly streamflow is 32.39 million m³ under the RCP4.5 scenario, occurring on 18 July 2040, and is 34.34 million m³ under the RCP8.5 scenario, occurring on 19 July 2084.

Due to the absence of measured hourly streamflow data, the maximum hourly streamflow values for RCP4.5 and RCP8.5 scenarios were further converted into streamflow per second and compared with historical flood data for the watershed. The results are shown in Table 14. As of now, since the completion of the Taolinkou Reservoir in 1998, the maximum flood peak of the Qinglong River has been 2,011 m³/s. It is noteworthy that the maximum flood peak discharges recorded in the Qinglong River before the completion of the downstream reservoir were 8,760 m³/s (Table 14). Referring to relevant information (Wu 2009), the design flood standard for the Taolinkou Reservoir is outlined in Table 15, indicating that the maximum design flood corresponds to a 100-year flood with a peak flow of 14,340 m/s.

Analyzing the results from all 31 GCMs, the projected future peak flood flow under the RCP4.5 scenario is expected to range from 138.57% (short-term) to 1,275.05% (long-term) of the historical 8,760 m³/s. For the RCP8.5 scenario, the projected peak flood flow is expected to range from 123.36% (short-term) to 919.05% (long-term). However, the peak flood flow obtained from the MME mean is projected to be 83.36% (short-term) to 101.15% (medium-term) of the historical 8,760 m³/s under the RCP4.5 scenario and 84.70% (short-term) to 108.89% (long-term) under the RCP8.5 scenario.

Next, the daily streamflow was simulated by configuring the HSPF output time scale to daily. The average daily streamflow and maximum daily streamflow values for the 31 GCMs are presented in Tables 16 and 17, respectively. The detailed statistical characteristics of daily streamflow under the RCP4.5 and RCP8.5 scenarios, including skewness, coefficient of variation, confidence intervals, etc., can be found in the appendix.

The daily streamflow under both the RCP4.5 and RCP8.5 scenarios shows a distribution with an average kurtosis of 183.056 and 160.786, respectively, and skewness of 10.884 and 10.396, respectively. These findings indicate a more peaked data distribution that is skewed to the right, suggesting a higher frequency of extreme values compared to a normal distribution. Moreover, the coefficient of variation fluctuates around 4.3, indicating considerable variability in the data. The values in the dataset are relatively dispersed, similar to the pattern observed in daily precipitation.

According to the results, the maximum values of the maximum daily streamflow for 2021–2100 are 392 million m³ under the RCP4.5 scenario and 260.3 million m³ under the RCP8.5 scenario. Both are simulated using meteorological data from the ACCESS1-3 model, occurring on 18 July 2069 and 23 August 2088, respectively. The MME mean of daily streamflow and the MME mean of daily precipitation are illustrated in Figure 12. The results indicate that the maximum daily streamflow values of the MME mean under the RCP4.5 and RCP8.5 scenarios are 30.5012 and 31.8099 million m³, respectively, occurring on 18 July 2040 and 18 July 2081.

Table 18 lists the comparison of historical and simulated data. According to the available measured data, the historical maximum daily streamflow was 172.37 million m³ on 3 August 2012 (Figure 10 and Table 18). The average daily streamflow for 2001–2016 was 1.0053 million m³.