Impact of Climate Change on Reservoir Operation during the Dry Season in the Pearl River Basin

Abstract

Climate change has far-reaching impacts that have created new challenges for water resource management. As an important measure to coordinate the relationship between society, economy, and environment, reservoir scheduling can reduce the future impact of climate change. It is, therefore, important to investigate the impacts of scheduling on reservoir operation. In this study, a reservoir system in the Pearl River Basin was selected to explore these impacts. Results show that the basin temperature significantly and abruptly increased in 2000, and precipitation and streamflow changed abruptly in 1983 and 1992. Historically, climate change has increased power generation, increased the risks to water supply security and ecological protection, and altered the relationship between power generation and ecological protection objectives. Based on 28 global climate models, the rank scoring method, Delta statistical downscaling, and the SWAT model, three emission scenarios (RCP2.6, RCP4.5, and RCP8.5) in CMCC-CM were evaluated to assess the climate change impact. In the future, the temperature will continue to exhibit an increasing trend, and the amount and distribution of streamflow will be altered. Although climate change will increase power generation in the dry season, it will also bring about new challenges for ecological protection and water supply security. Accordingly, the Datengxia Reservoir operating rules may require significant amendments.

1. Introduction

Over the past century, the global climate has significantly changed [1,2], becoming a common concern for the scientific community, governments, and the public [3,4]. The sixth assessment report of the Intergovernmental Panel on Climate Change (IPCC) has revealed that the global average temperature from 2011 to 2020 increased by 1.09 °C (0.95–1.2 °C) when compared to that from 1850 to 1900, with a warming amplitude of 1.59 °C (1.34–1.83 °C) being observed on land. China, which is an important region in terms of climate change sensitivity, showed a significantly higher warming rate from 1951 to 2020 than the global average. Climate change has altered precipitation patterns, increased extreme weather events, augmented sea levels, and altered the biosphere [5]. Water resources are the basis for human survival and development, and climate change intensification has led to uncertainties regarding the already rampant global water crisis [6]. The frequency and intensity of extreme weather events are bound to increase in the future [7], thereby affecting the operation of existing water infrastructure [8,9]. Climate change also alters the water cycle by influencing regional precipitation mechanisms [10], enhancing or weakening evaporation [11], and accelerating the hydrological cycle [12], thus affecting the total amount of water resources in river basins, altering the temporal and spatial distributions of these resources, and providing additional challenges for watershed water resource management [13].

As a means of water resource management, reservoir operation plays an important role in balancing water resource utilization and ecological protection [14]. Under increasingly severe global climate change [5], reservoir operation is challenged with complex problems, making it an important research focus area. Booker and O’Neil [15] studied the relationship that leakage and evaporation losses have with reservoir storage capacity; they found that these losses are more profound under uncertain runoff conditions. Meanwhile, Burn and Simonovic [16] studied the potential impact of climate change on the operation and storage of the Shellmouth Reservoir in Canada; they reported that reservoir operation is highly sensitive to runoff data. Park et al. evaluated the impact of climate change on the runoff and storage of irrigation reservoirs [16], and found that irrigation reservoirs must reduce discharge in August and September. Vicuna et al. studied the California hydroelectric power system under six climate change scenarios [17], revealing that climate change is decreasing electricity production and other income streams due to a reduction in regional precipitation and runoff and an increase in temperature. Additionally, Mujumdar et al. studied the Hirakud Reservoir in India and found that its performance and annual power generation was decreasing under the influence of climate change [18]. Wanders et al. studied the Tekeze hydropower reservoir under future climate scenarios and showed that the storage potential of the reservoir will increase in the future [19]. Moreover, Ehsani et al. estimated the climate change impacts on reservoirs based on a general reservoir operation scheme, while reporting that the importance of dams in the region for providing water security will increase.

Hence, climate change will continue to impact reservoir operations in complex and varied ways [1,20]. For example, rainfall patterns have already been altered [10,21], which has led to changes in streamflow inputs to reservoirs and affected the storage and release strategy of reservoirs. Moreover, temperature directly affects evaporation [6] and further impacts the water level and storage capacity of reservoirs, thereby posing significant challenges to reservoir operations [22]. Traditional reservoir scheduling strategies are often based on historical climate data and hydrological observations [23]; however, historical data may not accurately reflect future conditions under climate change. Thus, in the future, reservoir dispatching methods under different climate scenarios will have to undergo changes to face new challenges [24,25]. With the increasing impact of climate change on water resources [26], reservoir management must respond flexibly to different climate scenarios and take appropriate measures to ensure the sustainable use of water resources [27,28,29]. By simulating reservoir dispatching under different climate scenarios [30], we can better understand future hydrological process, the supply and demand of water resources [31], and the impact of reservoir dispatching to provide a scientific basis for water resource managers, optimize decisions, and ensure sustainable utilization of water resources. Through scientific and reasonable reservoir management strategies, the relationship between supply and demand can be balanced to ensure sustainability and stability of the water supply.

However, research related to the impact of climate change on reservoir operation often focuses on a single aspect, such as water supply, storage capacity, hydroelectric power generation, or flood control. Although reservoirs generally have multiple functions during operation, few studies have evaluated the impact of climate change on the reservoir group operation system by considering multiple objectives, such as power generation, ecology, and water supply. Furthermore, most studies have focused on climate changes that have either occurred in the past or will occur in the future [24,32,33], while few papers have incorporated both periods simultaneously to assess the impact of climate change on reservoir operation. Climate change has occurred in history and is bound to further intensify in the future; thus, considering both historical and future trends in climate change can provide a more complete picture of the overall impact. The Pearl River plays an important role in China’s development, and a group of reservoirs has been built in the Pearl River Basin (PRB). However, few studies have evaluated reservoir operation in the PRB and even fewer have elucidated the impact of climate change on reservoir operation in the PRB.

This study considered the reservoir group in the PRB as the research object and methods associated with meteorology, hydrology, ecology, etc. were applied to: (a) identify climate change characteristics in the PRB using historical and future data series; (b) quantify the impact of climate change on streamflow based on the SWAT model; (c) analyze the impacts of climate change on reservoir operation while considering water supply, power generation, and ecological protection. We comprehensively applied the trend and mutation analysis method, hydrological model, multi-objective optimal dispatching model of reservoir groups, multi-objective optimization algorithm, multi-objective decision-making method, and suitability evaluation index system of the climate model to study the key issues in the adaptive dispatching of reservoir groups in the PRB under climate change conditions. This study holds great significance for maintaining the power generation efficiency of cascade reservoirs, ensuring flood control safety, and protecting the ecological environment in the PRB, and the results provide theoretical support for reservoir operation under the conditions of future climate change and offer a reference for future studies aiming to elucidate the impact mechanism of climate change on water resources and their management.

2. Materials and Methods

2.1. Study Area and Data

The Pearl River Basin

The Pearl River is one of the seven major rivers in China, has a total length of 2400 km, and passes through six provinces (Guangdong, Guangxi, Yunnan, Hunan, Guizhou, and Jiangxi), Hong Kong, and Macao (Figure 1). The PRB can be divided into four different areas, including the Xijiang River Basin, Beijiang River Basin, Dongjiang River Basin, and the Pearl River Delta [34], forming a unique river system pattern named the “three rivers converge and eight gates diverge”. Affected by geographical, climatic, and topographical conditions, the precipitation distribution in the PRB is extremely uneven throughout the year [35]. Precipitation during the wet season (April–September) accounts for more than 70% of the annual precipitation, and extreme rainfall with high intensity, long duration, and high frequency tends to occur during the flood season. In the dry season, saltwater intrusion caused by reduced streamflow into the delta and the strengthening of tidal dynamics pose challenges to the water supply in the estuary [36]. In addition, cities affected by saltwater intrusion, such as Macao, Zhuhai, and Zhongshan in the Guangdong–Hong Kong–Macao Greater Bay Area, mainly have river-based water sources, and the proportion of river water intake to the water supply exceeds 85%. In addition to the substantial increase in water consumption and the increase in requirements for water supply security, the pressure to guarantee water supply safety and security during the dry season is also steadily increasing.

To ensure flood control and water supply in the PRB, different reservoir groups have been constructed, including Tianyi, Guangzhao, Longtan, Baise, and Datengxia. The characteristic parameter values of the reservoirs are listed in

Table 1. Characteristic parameter values of the reservoirs.

Characteristic	Unit	Reservoir
		Guangzhao	Tianyi	Longtan	Baise	Datengxia
Construction status	–	Completed	Completed	Completed	Completed	Under construction
Total storage capacity	Billion m3	32.45	102.57	188.09	56.60	34.79
Storage capacity for flood control in the wet season	Billion m3	–	–	50.00	16.40	15.00
Crest elevation	m	750.5	791	382	234	64
Water level for normal uses	m	745	780	375	228	61
Water level limit for flood control in the wet season	m	745	773.1	359.3	214	47.6
Dead water level	m	691	731	330	203	47.6
Installation capacity	MW	1040	1224	4900	540	1600
Annual power generation	Billion kW·h	2.75	5.23	15.67	1.70	6.06

. These reservoirs have played an important role in resisting flood and drought disasters, increasing power generation, ensuring navigation, and protecting the ecological environment. Based on these reservoirs, the Pearl River Water Resources Commission of the Ministry of Water Resources (PRWRC) has successfully conducted 19 years of water scheduling operations in the dry season since 2005.

2.2. Data

The datasets used in this study included meteorological, hydrological, spatial, and GCM data.

2.2.1. Meteorological and Hydrological Data

Daily meteorological data from 1956 to 2019 collected from 70 meteorological stations in and around the PRB were included. Meteorological data on precipitation, maximum and minimum temperatures, wind speed, and relative humidity were obtained from the National Meteorological Information Center (http://data.cma.cn accessed on 6 November 2021). Hydrology station and reservoir inflow data were obtained from the hydrological yearbook and Water Regime Daily, which are openly accessible on the website of the PRWRC.

The meteorological data used in this study underwent quality control before release [37]. Meteorological and hydrological data quality control checks were performed using RClimDex (http://etccdi.pacificclimate.org/software.shtml accessed on 12 December 2021), and any potential outliers (such as typing errors, missing data, rejected values, and minimum values exceeding maximum values) were manually checked and corrected. Errors were corrected by replacing unreasonable values with the mean value of the meteorological elements on the same date for the previous two years and next two years.

2.2.2. Spatial Data

Digital elevation model (DEM) (ASTER GDEMV2) data were obtained from the Geospatial Data Cloud (http://www.gscloud.cn/ accessed on 10 December 2021), which was jointly developed by MRTI in Japan and NASA in the United States. The DEM resolution was selected by comprehensively considering the simulation accuracy and computational efficiency, and a 90-m DEM was adopted. Land use and soil data were obtained from the National Earth System Science Data Center (http://www.geodata.cn/ accessed on 12 December 2021).

2.2.3. GCM Data

GCMs (

Table 2. Twenty eight global climate models (GCMs) of the Coupled Model Intercomparison Project Phase 5 (CMIP5).

GCM	ID	Originating Group(s)	Country
BCC-CSM1.1	1	Beijing Climatic Centre	China
BCC-CSM1.1.M	2
BNU-ESM	3	Beijing Normal University
CanESM2	4	Canadian Centre for Climate Modelling and Analysis	Canada
CCSM4	5	National Centre for Atmospheric Research (NCAR)	USA
CMCC-CM	6	Centro Euro-Mediterraneo per Ⅰ Cambiamenti Climatici	Italy
CSIRO-ACCESS-1	7	Commonwealth Scientific and Industrial Research	Australia
CSIRO-ACCESS-3	8	Organization (CSIRO) and Bureau of Meteorology (BOM)
CSIRO-Mk3.6.0	9	Commonwealth Scientific and Industrial Research Organization (CSIRO), in collaboration with the Queensland Climate Change Centre of Excellence
EC-EARTH	10	EC-EARTH consortium	Europe
FIO-ESM	11	The First Institute of Oceanography, SOA	China
GISS-E2-R	12	NASA Goddard Institute for Space Studies	USA
GISS-E2-R-CC	13
INMCM4	14	Institute for Numerical Mathematics	Russia
IPSL-CM5A-LR	15	Institut Piere-Simon Laplace	France
IPSL-CM5B-LR	16
MIROC5	17	National Institute for Environmental Studies, and Japan Agency for Marine-Earth Science and Technology	Japan
MIROC5-ESM	18	Japan Agency for Marine-Earth Science and Technology, Atmosphere and Ocean Research Institute
MIROC5-ESM-CHEM	19	The University of Tokyo and National Institute for Environmental Studies
MPI-ESM-LR	20	Max Planck Institute for Meteorology	Germany
MPI-ESM-MR	21
MPI-ESM-P	22
MRI-CGCM3	23	Meteorological Research Institute	Japan
NorESMI-M	24	Norwegian Climate Centre	Norway
NorESMI-ME	25
CESM1(BGC)	26	Community Earth System Model Contributors NSF-DOE-NCAR	USA
CESM1(CAM5)	27
CESM1(WACCM)	28

) under three representative concentration pathways (RCP2.6, RCP4.5, and RCP8.5) were used to extract future climate scenarios for the PRB, and 28 GCMs were selected based on data availability. CMIP5, a version of the AR5 GCM, was used; it is a collaborative framework designed to improve knowledge of climate change.

2.3. Methodology

Based on trend and abrupt change analyses of temperature and precipitation, the historical series was divided into base and mutation periods. The SWAT model was constructed based on the hydrometeorological element series in the base period; subsequently, the streamflow for the mutation period was obtained using the SWAT model. A multi-objective operational model of the reservoirs was built to compare changes in the objectives and reservoir operation in the PRB during the dry season. The rank scoring method was introduced to select the most suitable GCM from the 28 GCMs. Delta statistical downscaling was employed to simulate the temperature, precipitation, and streamflow under three emission scenarios (RCP8.5, RCP4.5, and RCP2.6) over the next 50 years. Finally, the impacts of the three scenarios on reservoir operation were compared (Figure 2).

2.3.1. Trend Analysis

Traditional methods, including Senslope and Mann-Kendall (MK) [37,38], were used to analyze trends in the annual average temperature and precipitation in the PRB from 1956 to 2019. The annual average streamflow series of the Wuzhou Station over the past 64 years was also analyzed. In addition, an innovative trend analysis (ITA) method was introduced to enhance confidence in the results. The ITA method displays results in the form of images; thus, it is simple and intuitive, does not require a data series to follow a certain assumption distribution, and avoids the impact of series autocorrelation on the results [39].

2.3.2. Abrupt Change Analysis

In addition to the tendency of continuous change, meteorological series also show abrupt changes from one stable state to another at certain points. Different abrupt change tests are based on different mathematical principles, series lengths, and numerical distribution requirements. Thus, the results obtained using different mutation test methods can vary even for the same series [40]. Therefore, this study comprehensively used the cumulative anomaly, MK, Pettitt, and Moving-T methods to identify abrupt changes in temperature, precipitation, and streamflow series in the PRB.

2.3.3. SWAT Model

The SWAT model has a modular structure, with each hydrological cycle process corresponding to a sub-module [41]; this structure is convenient for model expansion and application. Overall, the model is used to perform three functions: sub-watershed hydrological cycle, river flow calculation, and reservoir water balance processes. According to the simulation requirements, the SWAT model divides the watershed into several sub-basins and simulates large-scale spatial variables, such as watershed soil, land use, and management measures. The model classifies molecular watersheds according to the threshold for the minimum area of the sub-catchments. According to different land uses and soil types, each sub-catchment is further divided into one or more hydrological response units (HRUs). In the model calculation process, each HRU responds to the transport and loss of water, sand, nutrients, and pesticides, accumulates the corresponding values of each sub-basin, and calculates them for tributaries (travelling to the main stream and to the outlet section of the basin) by determining the river confluence point.

Many parameters are involved in the SWAT model [42]; hence, it is difficult to adjust each parameter. The parameter sensitivity analysis module solves this problem. Through calibration of these sensitive parameters, the simulation accuracy is improved, and the model parameters conform to the actual distribution of the basin as much as possible to increase the applicability of the model and the accuracy of the simulation.

The simulation effect of the model reflects its applicability to a region. The root mean square error (RMSE), mean absolute error (MAE), relative error (RE), determination coefficient (R²), and Nash coefficient (NSE) are typically used to evaluate the applicability of the SWAT model. The RE test model simulates the water balance state. By calculating the RE between the predicted and actual values, the simulation effect of the model was found to be optimal when the RE tended toward zero. R² and NS are used to test the reasonable degree of fitting of the simulation process and are important indicators for characterizing the correlation and effectiveness of the forecast model. If the correlation and NSE tend toward 1, the annual simulation results of the model are optimal. The RE was calculated as per Equation (1):

$$\mathrm{Re}=\frac{Q_p-Q_0}{Q_0}\times 100\%$$

(1)

where $Q_p$ is the simulated streamflow (m³/s) and $Q_0$ is the measured streamflow (m³/s). If RE is >0, the simulated streamflow value of the model is too large; if RE is <0, the simulated streamflow of the model is small; if RE = 0, the simulated streamflow is equal to the measured streamflow.

R² was obtained using the linear regression method in Excel. If R² = 1 the measured value is consistent with the simulated value; if R² < 1, the smaller the value, the lower the degree of coincidence between the measured and simulated values.

The formula for calculating the NSE is presented as Equation (2):

$$NSE=1-\frac{{\displaystyle {\sum }_{i=1}^n{(Q_0-Q_p)}^2}}{{\displaystyle {\sum }_{i=1}^n{(Q_0-\overline{Q})}^2}}$$

(2)

where $\overline{Q}$ is the average measured streamflow, $Q_0$ is the measured streamflow, and $Q_p$ is the simulated streamflow. When NSE is <0, the simulated value is less reasonable than the measured value.

2.3.4. Multi-Objective Operation Model of Reservoirs

This study comprehensively considered the construction scale, regulation capacity, control effect, and operation status of the reservoir, combined with the actual dispatching experience of the basin. The reservoir groups in Guangzhao, Tianyi, Longtan, Baise, and Datengxia were treated as the research objects. The streamflow series observed by the control hydrological station, Wuzhou Station, was used to calculate the objective. A generalized diagram of the reservoir group system in the PRB is shown in Figure 3.

Reservoirs in the PRB have multiple objectives, including water supply, power generation, and ecological protection during the dry season. The water supply objective was to minimize the ratio of water deficiency, and the power generation objective was to maximize power generation. With regard to ecological protection, there are four ways to consider the ecological flow of rivers. In this study, we focused on climate change that occurs in the average state of a ≥30-year climate series as well as reservoir regulation as an important adaptive measure that will effectively eliminate harm and provide benefits in the distant future. In addition, considering the complete description of the ecological demand of the PRB, ecological flow should be maintained by reducing the degree of flow change as much as possible and minimizing the degree of ecological deviation to maintain species stability, population structure, and river ecosystems. The calculation formulas used for these objectives are presented in Equations (3)–(5):

(1)

Water supply:

$$f_1=\min {\displaystyle \sum _{t=1}^T(\frac{D_t-q_t}{D_t}{)}^2}$$

(3)

where T represents the total period, D_t represents water demand at Wuzhou Station in period t, and q_t represents the streamflow at the Wuzhou Station.

(2)

Power generation:

$$f_2=\max E=\max \left\{{\displaystyle \sum _{g=1}^5{\displaystyle \sum _{t=1}^T\left(N_{g,t}\times \mathsf{\Delta }t\right)}}\right\}$$

(4)

where E is the total power generation (kW·h), g represents the reservoir number, and N_g,t represents the power output (kW).

(3)

Ecological protection:

$$f_3=\min ECO=\min {\displaystyle \sum _{t=1}^T{\left(q_t-q_{et}\right)}^2}$$

(5)

where ECO is the degree of ecological deviation and q_et represents ecological flow. The distribution flow method was used to quantify the ecological flow.

To ensure the safety and stability of the reservoir, constraints were applied, as represented by Equations (6)–(10):

(1)

Water balance constraint:

$$V_{i,t+1}=V_{i,t}+\left(I_{i,t}-Q_{i,t}\right)\times \mathsf{\Delta }t-E{P}_{i,t}$$

(6)

where V_{i, j+1} and V_i,j are the reservoir storage of the ith reservoir during period t + 1 and t, respectively; I_i,t and Q_i,t represent reservoir inflow and outflow, respectively; and EP_i,t is the sum of evaporation and leakage.

(2)

Water level constraint:

$$Z_{i,t}^{\min }\le Z_{i,t}\le Z_{i,t}^{\max }$$

(7)

where Z_i,t is the water level of the ith reservoir during period t, and $Z_{i,t}^{\min }$ and $Z_{i,t}^{\max }$ represent the minimum and maximum water levels, respectively.

(3)

Discharge constraints:

$$Q_{\min }^i\le Q_{i,t}\le Q_{\max }^i$$

(8)

where Q_i,t is the outflow of the ith reservoir during period t, and $Q_{i,t}^{\min }$ and $Q_{i,t}^{\max }$ represent the minimum and maximum outflows, respectively.

(4)

Hydraulic connection between cascade reservoirs:

$$I_{i + 1,t}=Q_{i,t}+\mathsf{\Delta }{q}_{i+1,t}$$

(9)

where I_{i + 1,t} is the inflow of the i + 1th reservoir during period t, and $\mathsf{\Delta }{q}_{i + 1,t}$ is the interval discharge between the ith and i + 1th reservoirs.

(5)

Power output constraints:

$$N_{i,t}^{\min }\le N_{i,t}\le N_{i,t}^{\max }$$

(10)

where $N_{i,t}^{\min }$ and $N_{i,t}^{\max }$ represent the minimum and maximum power outputs, respectively.

To determine the reservoir operating rules, we extracted the scheduling experience of the reservoir water level and flow rate, among other factors. Moreover, the scheduling rules and formulated parameters were considered during the reservoir design and incorporated into the model as constraints.

2.3.5. Rank Scoring Method

GCMs are important tools for analyzing climate change and studying adaptive strategies [43]. In the process of climate simulation, initial conditions, boundary conditions, scenario setting, observational data, model parameters, and structure all lead to uncertainty in the simulation results, particularly at the regional scale, where the uncertainty is more prominent. To predict future climate conditions relatively reliably, climate models should not only accurately simulate the characteristics of the climate system in historical periods but also accurately depict the changes in the climate system. Therefore, before using GCMs to predict future climate change, it is crucial to compare the effects of different climate models that fit the regional historical climate characteristics and changes to obtain a relatively accurate prediction of regional climate change over the next few decades.

Based on the monthly precipitation and temperature series in the historical periods of the 28 GCMs in CMIP5 (Table 2), the rank-scoring method was introduced to select the best-fitting GCM in the PRB. The eight evaluation indices listed in

Table 3. Evaluation index and weight of rank sum method.

Serial Number	Evaluation Index		Indicator Symbol	Weight
1	Mean relative error		${\mathrm{RE}}_{\overline{x}}$	1.0
2	Variance relative error		${\mathrm{RE}}_{\delta }$	1.0
3	Root mean square error		$\mathrm{NRMSE}$	1.0
4	Spatial and temporal distribution eigenvalues	Intra-year distribution correlation coefficient	$Cor{r}_M$	1.0
5		Spatial autocorrelation coefficient	$Cor{r}_S$	1.0
6	Trend indicator	ITA	S	1/3
7		MK	Z	1/3
8		Senslope	b	1/3

were used to evaluate the effectiveness of each GCM, and bilinear interpolation was used to convert the original GCM data to the meteorological station scale. The calculation method for each indicator has been provided previously [44].

After calculating the evaluation index and assigning a rank score of 0–9 to each index, the final rank sum was calculated according to the given weight and sorted according to the rank sum. The GCM with the smallest rank sum was selected as the optimal model for simulating precipitation and temperature in the PRB. The rank sum was calculated using Equation (11):

$$\mathrm{RS}={\displaystyle \sum }_{i=1}^8{\omega }_i\times {\mathrm{RS}}_i={\displaystyle \sum }_{i=1}^8{\omega }_i\times \frac{x_i-x_{min}}{x_{max}-x_{min}}\times 9$$

(11)

where RS is the rank sum of the GCMs, ${\omega }_i$ is the weight of the ith index, RS_i is the rank score of the ith index, x_i is the ith index value, and x_min and x_max represent the minimum and maximum ith index values, respectively.

2.3.6. Delta Downscaling Method

The Delta downscaling method uses the change characteristics of future climate elements predicted by the GCM to reconstruct future scenarios by superimposing the change characteristics on the measured climate element sequence in the base period. Specifically, precipitation in the future scenario can be obtained by calculating the rate of precipitation change in each future period and multiplying it by the measured precipitation in the historical period. The air temperature was obtained by calculating the absolute change in air temperature and superimposing the absolute change based on the measured air temperature series in the base period. Fifty years (1956–2005) of GCM historical simulations were used as the base period, and a 50-year period (2021–2070) was used for the future time frames. The associated calculations are represented by Equations (12) and (13):

$$P_f=P_O\times \frac{P_{Gf}}{P_{GO}}$$

(12)

where P_f is the future precipitation series, P_Gf is the future precipitation series predicted by the GCM, P_GO is the annual monthly average precipitation in the base period simulated by the GCM, and P_O is the annual monthly average precipitation observed in the base period.

$$T_f=T_O+(T_{Gf}-T_{GO})$$

(13)

where T_f is the reconstructed future air temperature series, T_Gf is the future air temperature series predicted by the GCM, T_GO is the average annual air temperature in the base period simulated by the GCM, and T_O is the average annual air temperature observed in the base period.

3. Results and Discussion

3.1. Identification of Historical Climate Change in the PRB

3.1.1. Trend Analysis

The results of the indicators using the ITA, Senslope, and MK trend analysis methods based on the annual mean low temperature, high temperature, and annual mean precipitation series in the PRB from 1956–2019 are shown in

Table 4. Results of the indicators based on three trend analysis methods.

Series	MK	Senslope	ITA
Tmin	6.425 *	0.020	0.020 *
Tmax	2.914 *	0.010	0.011 *
Precipitation	0.875	0.918	0.752

Note: * trend exceeded the 95% significance level.

. T_min and T_max showed significant increases, whereas precipitation exhibited a non-significant increase (Table 4). Changes in T_min were more obvious than those in T_max, and the rate of increase in T_min was approximately twice that in T_max. However, the warming rate of the PRB was lower than that of the whole of China (0.026 °C/year). According to the ITA (Figure 4), the low, median, and high categories of T_max all showed increasing trends (<5%), which exhibited a weakening–strengthening–weakening pattern with an increase in T_max.

The three categories of T_min also showed an increasing trend that gradually strengthened with an increase in T_min, where the high category of T_min was approximately 5%. The low and median categories of precipitation were concentrated between the no-trend and −5% lines, showing a slightly decreasing trend, whereas the trend in the high category was the opposite, showing an increasing trend.

3.1.2. Abrupt Change Analysis

The Moving-T test and MK method were used to identify the abrupt change points in the temperature and precipitation series (Figure 5a–e). For both T_max and T_min, the statistical series of the Moving-T test was significant (α = 0.05) in 2000, and two curves referred to as UF and UB (i.e., indicators of the MK method) also intersected in 2000. Therefore, an abrupt change occurred in the temperature series of the PRB in 2000. The Moving-T results for precipitation revealed mutations in 1983 and 1992 (Figure 5c). Furthermore, the MK test confirmed that precipitation exhibited an abrupt change in 1983 and 1992, as UF and UB intersected within the α = 0.05 confidence interval. After the abrupt change in 1983, annual rainfall decreased by 95 mm (6.4% of annual average rainfall). After the abrupt change in 1992, precipitation returned to its original level.

This study focused on the impact of climate change on hydrological cycles. Therefore, based on the Pearson correlation coefficient, a correlation analysis between T_max, T_min, precipitation, and the streamflow series was conducted. The results showed that only precipitation was significantly correlated with streamflow (r = 0.80). Therefore, based on the abrupt change analysis results for precipitation, the entire series (1956–2019) was divided into three stages: Stage A (1956–1983), Stage B (1983–1992), and Stage C (1992–2019).

When focusing on the dry season (Figure 6), the precipitation of all three stages from November to the following February was <100 mm. Precipitation generally first decreased and then increased over the month, with the lowest in December and the highest in March. However, the difference between the two stages gradually narrowed, with precipitation from January–March exceeding the benchmark period of Stage A.

Except for February, precipitation in Stage C was generally higher than that in Stage A during the dry season. Precipitation in Stage C in November and February decreased by approximately 20% compared to Stage A, while precipitation in November and January increased by approximately 20%. It was apparent that Stage B had a significant proportion of differences in each month compared to Stages C and A.

3.2. Streamflow Simulation

Based on the ITA and MK methods, streamflow at the Wuzhou Station showed an insignificant increasing trend; according to the moving-T test and MK results, abrupt change points in the streamflow were observed in 1983, 1992, and 2002. Subsequently, the SWAT model was used to simulate the streamflow of the PRB, and Stage A (1956–1983) was used for calibration and validation. Specifically, the daily meteorological and hydrological data from 1956–1978 were used to calibrate the SWAT model; the first two years were set as the warm-up period. The verification period ranged from 1979 to 1983. SWAT-CUP was introduced to conduct parameter sensitivity analysis, and the SUFI-2 algorithm was used to optimize the parameters. The SWAT model was run 500 times; five out of the 28 parameters exhibited significant results: CN2, CANMX, CH_k2, Sol_Z, and RCHRG_DP.

The results of the streamflow simulation compared to the observed series at the Wuzhou Station are shown in Figure 7. The variation processes of the measured and simulated flows were similar (Figure 7a,b), and the peak flow and dry season flow determined using the SWAT model were consistent with the observed series. Scatter plots (Figure 7c,d) of the two periods showed that the numerical points were evenly distributed on both sides of the 45° diagonal, indicating that the measured and SWAT-simulated flows were consistent overall. In addition, the results of the five evaluation indicators calculated based on the whole year and dry season series are listed in

Table 5. Evaluation of the simulation results of Wuzhou streamflow in the Pearl River Basin.

Series	Period	MAE	RMSE	Re/%	NSE	R2
Whole year	Calibration	1583	2354	14.9	0.86	0.91
	Verification	1462	2090	10.7	0.87	0.89
Dry season	Calibration	639	2670	6.8	0.68	0.87
	Verification	647	948	1.9	0.80	0.87

. The RE in the calibration and verification period was <15%, while the NSE and R² were >0.85. The MAE and RMSE of the whole year series were approximately 1500 and 2000, respectively. The MAE of the dry season was <650, while the RMSE of the dry season was 2670 in the calibration period and 948 in the verification period. Figure 7e presents a violin plot for the calibration and verification periods. The median value of the SWAT model results was slightly larger than that of the observed series. Meanwhile, the observed series was more concentrated, particularly for streamflows <1000 m³/s. Additionally, the distributions of the simulated series and observed series were similar. Thus, the SWAT model effectively simulated daily streamflow in the PRB.

Streamflow in the three stages during the dry season is shown in Figure 6b. Streamflow generally decreased initially and then increased over the month. In addition, the changes in monthly streamflow in Stages B and C were similar to those in precipitation, with the changes in February and March in Stage C being an exception.

3.3. Impact of Climate Change on Reservoir Operation

The results of the multi-objective analysis of the PRB reservoir are shown in Figure 8. Figure 8a compares the Pareto frontier of the three different stages, indicating that compared to the objective of Stage B, Stage C was more similar to Stage A overall. In particular, power generation ranged from 16.39 to 16.59 billion kW·h in the dry season in Stage A, while the mean value of the series was 16.49 billion kW·h. The mean power generation in Stage B increased to 16.51 billion kW·h, and the distribution was more concentrated; the mean value reached 16.54 billion kW·h in Stage C. The ecological protection objective ranged from 0.201 to 0.218 billion in Stage A, and the average of the series was 0.204 billion; the average increased by approximately 7.3% in Stage B. The water supply objective varied from 0 to 0.123 in Stage A, with mean values of 0.008, 0.019, and 0.021 in Stages A, B, and C, respectively. It was evident that both the average value and range of the series increased over time. Furthermore, Figure 8b–d show that the power generation and ecological protection objectives had a strong competitive relationship and that the correlation of different objectives changed in different stages. According to the Pearson correlation analysis, the statistical value based on power generation and the ecological protection objective is r, which was 0.64 in Stage A, 0.69 in Stage B, and 0.46 in Stage C. In conclusion, historical climate change increased power generation, water supply risk, and ecological protection objectives and changed the relationship between power generation and ecological protection objectives.

The optimal solution was selected based on the TOPSIS method, and the results of the reservoir operation at different stages are shown in Figure 9. In Stage A, in the Guangzhao, Longtan, and Baise reservoirs, the water level continuously decreased over a month, whereas the Tianyi and Datengxia reservoirs released water first, then increased the water level in February, and decreased it to below or near that for flood control in the wet season. In Stage B, the water levels in Guangzhao, Tianyi, Baise, and Datengxia were lower than those in Stage A from December to February, whereas Longtan stored more water in January and February. In Stage C, the water levels of Guangzhao, Tianyi, and Baise were higher than those in Stage A from December to February, whereas the water level of Longtan was between those of Stages A and B, and the operation rules of Datengxia differed significantly. Among the five reservoirs, Longtan and Base were relatively similar across the three different stages, while Datengxia exhibited the maximum water level fluctuation.

3.4. Comparison of GCMs in the PRB

Based on monthly simulation data for 28 GCMs in CMIP5 and 70 meteorological station observation data points in and around the PRB, the rank scoring method was introduced to determine the best-fitting CCM models in the PRB. Eight evaluation indices, including the mean relative error, variance relative error, root mean square error, intra-year distribution correlation coefficient, spatial autocorrelation coefficient, and ITA, MK, and Senslope indicators, were calculated and then weighted to obtain the rank score; the lower the rank score, the better the performance of the GCM.

From the temperature series (Figure 10a), the rank score for the top four GCMs (NorESMI-M, MIROC5-ESM-CHEM, BCC-CSM1.1. M, CMCC-CM) was only ~0.5, while the rank score of the 18th GCM (MIROC5-ESM) reached over 3.5 (7-times higher). The results for precipitation ranged from 2.4 to 3.6. The 6th GCM (CMCC-CM) had a lower rank score, followed by EC-EARTH, MIROC5, INMCM4, and CSIRO-ACCESS-1 (Figure 10b). Therefore, considering the temperature and precipitation results, CMCC-CM was selected as the most suitable climate model for the PRB among the 28 GCMs.

3.5. Future Climate Change Simulation in the PRB

Using the Delta method, the change characteristics of future climate elements predicted by the CMCC-CM model were used to reconstruct the future scenario by superimposing the change characteristics on the measured climate element sequence in the base period. Climate change over the next 50 years under the three emission scenarios, RCP8.5, RCP4.5, and RCP2.6, was simulated (Figure 11).

The simulation results showed that the temperature in the PRB will increase significantly, with the temperature under the RCP8.5 emission scenario increasing most obviously and the increase under the RCP2.6 emission scenario being relatively minimal over the next 50 years. Specifically, compared to the historical scenarios (1956–2005), the average maximum temperatures under RCP2.6, RCP4.5, and RCP8.5 scenarios increase by approximately 0.8, 0.93, and 1.23 °C, respectively, and the minimum temperatures increase by approximately 1.09, 1.44, and 1.82 °C, respectively. Furthermore, the standard deviation of the maximum temperature changes from 0.41 to 0.66 in the three different emission scenarios; for the minimum temperature, it changes from 0.39 to 0.54, 0.58, and 0.77 under RCP2.6, RCP4.5, and RCP8.5 scenarios, respectively. Thus, the numerical value and instability of temperature is bound to significantly increase in the PRB. Figure 11b,d indicate that the median temperature and range under the three scenarios were significantly larger in the future than those observed for the historical scenarios.

Regarding precipitation, compared to the historical series, the average showed no apparent change, decreasing by 2.46 mm under RCP2.6 and increasing by ~8.23 mm and 22.05 mm under RCP4.5 and RCP8.5, respectively. The median value decreases by 5.85 mm under RCP2.6 and increases by 5.70 mm under RCP8.5; however, under RCP4.5, precipitation increases by ~44.16 mm (i.e., by nearly eight times). Moreover, the stability of temperature significantly decreases in the PRB, while the standard deviation of precipitation increases by 35.20, 50.77, and 58.65 under RCP2.6, RCP4.5, and RCP8.5, respectively. Figure 11f shows that the future precipitation distribution in the three scenarios was similar to that observed in the historical scenarios.

3.6. Multi-Objective Reservoir Operation under Climate Change

By inputting the precipitation and temperature series into the SWAT model, the annual streamflow and monthly flow during the dry season in the next 50 years under the RCP2.6, RCP4.5, and RCP8.5 emission scenarios were calculated (Figure 12). Although streamflow showed future periodic changes in abundance and dryness, the annual streamflows under the three different scenarios were distinguishable (Figure 12a). Specifically, average streamflow changes by approximately 10% compared to the historical series under RCP2.6 and RCP4.5 and increases by >25% under RCP8.5. The medium value showed a similar regularity, with a change of >20% under RCP8.5 and approximately 5% under the other scenarios. The standard deviation increases by 60, 11, and 115% under RCP2.6, RCP4.5, and RCP8.5, respectively. Hence, flood control should be a significant area of focus during the 2035–2055 period, and droughts may occur around 2025 and 2060.

In addition to the significant changes occurring in annual runoff, considerable variations were observed in the annual streamflow distribution. The monthly discharge during the dry season is shown in Figure 12b. In the future, the monthly streamflow in November and December significantly decreases, and the degree of change is in the order of RCP2.6 > RCP4.5 > RCP8.5. In January and February, the streamflow changes slightly; however, the differences among the three scenarios were not significant. Meanwhile, in March, the streamflow increases, and the results for the three scenarios varied significantly. Climate change imposes challenges to ecological protection and water supply security as reservoirs must release more water to compensate for the reduced runoff caused by climate change, which will undoubtedly affect the efficiency of the reservoir group throughout the dry season.

Under the different scenarios, the streamflow volume and distribution varied significantly, and the responses of the different targets met by the reservoir operation varied. The relationships among power generation, water supply, and ecological protection under the historical scenario and the RCP2.6, RCP4.5, and RCP8.5 emission scenarios are shown in Figure 13. Compared to the historical scenario (Figure 8b and Figure 13a), the average value of power generation increases by 3 × 10⁷, 1.8 × 10⁸, and 1.3 × 10⁸ kW·h, and the degree of ecological deviation increases by 5.2, 3.3, and 8.1%, respectively, under RCP2.6, RCP4.5, and RCP8.5, whereas, the water supply increases in all three scenarios. If the optimal value is concentrated within the Pareto frontier, power generation increases by 0.2, 1.5, and 0.9%, and the degree of ecological deviation increases by 3, 3.2, and 6.5%, respectively, under RCP2.6, RCP4.5, and RCP8.5. Reduction in the planting of consumption crops, such as maize, soybean, mung bean, and cassava, in the dry season as substitutes for rice [45] and developing water management policies such as water conservation [1] are necessary steps to preserve the local ecology. Furthermore, there is a need to build water storage projects.

The relationships among the different objectives are shown in Figure 13b–d. Compared to the historical scenario, the competitive relationship between ecological goals and power generation objectives weakens, particularly for RCP8.5. According to the Pearson correlation analysis, the statistical values based on the power generation and ecological protection objectives, r, change from 0.64 to 0.53, 0.61, and 0.41 under RCP2.6, RCP4.5, and RCP8.5, respectively. In conclusion, future climate change will increase power generation; however, it will also increase the risks to water supply and ecological protection. The relationship between power generation and ecological protection objectives will certainly change.

The water level changes in Tianyi, Guangzhao, Longtan, and Baise during the dry season are approximately the same under the three scenarios (Figure 14). There is a significant difference in water levels between February and March during the Tianyi dry season and between December, January, and February in Guangzhao under the three different scenarios. As the most downstream reservoir, the water level of Datengxia varies greatly in different months under the three scenarios. The change trends under RCP4.5 and RCP8.5 were similar.

4. Conclusions

In this study, we explored the impact of future climate change on water resources in the PRB and the response of the reservoir system to climate change in historical and future series over the next 50 years. The main conclusions were as follows:

(1)

The temperature in the PRB showed a significant increasing trend, with the low temperature increasing twice as rapidly as the high temperature, whereas the precipitation and streamflow trends were not significant. The temperature abruptly changed in 2000, and the precipitation and streamflow changed in 1983 and 1992.

(2)

The SWAT model was proven to simulate the daily runoff of the Pearl River; the RE of the calibration and validation was <15%, and the NSE and R² were >0.8.

(3)

Climate change in the historical series increased power generation; compared to that in Stage A, power generation increased by 0.02 and 0.05 billion kW·h in Stages B and C, respectively. However, risks to water supply security and ecological protection also increased. Moreover, climate change has altered the relationship between power generation and ecological protection objectives.

(4)

CMCC-CM was selected as the optimal model among 28 GCMs. In the future, the temperature will show an increasing trend, and precipitation will increase under RCP8.5. The streamflow will change not only in amount but also in distribution. The mean annual streamflow showed a slight decrease (<10%) under RCP2.6 and a slight increase (<10%) under RCP4.5, whereas, under RCP8.5, it increased by >25%. In addition, the streamflow in November and December significantly decreased, whereas it increased in March.

(5)

Future climate change will present new challenges to ecological protection and water supply security. Focusing on the optimal value within the Pareto frontier, power generation increases by 0.2, 1.5, and 0.9%, respectively, and the degree of ecological deviation increases by 3, 3.2, and 6.5%, respectively, under RCP8.5, RCP4.5, and RCP2.6; the water supply in the three scenarios increases. The regulation rules of Tianyi, Guangzhao, Longtan, and Baise in the dry season will be approximately the same in the future; however, the Datengxia Reservoir operating rules may require significant adjustments. Thus, more attention should be directed toward these rules.

In this study, the GCM most suitable for the PRB was selected for analysis through model screening. However, the uncertainty of the GCMs is an important source of uncertainty in future climate and runoff predictions, and a supplementary analysis is required. In addition, the research scale in the scheduling model is 10-d, hence, further refinement is required.