Impact of Hydroclimate Change on the Management for the Multipurpose Reservoir: A Case Study in Meishan (China)

China holds the largest amount of reservoirs in the world, while more than 80% of them were constructed 50–70years ago and are approaching a critical stage of their designed lifetime. Before deciding the future of a reservoir, it is essential to ﬁnd out whether it could still satisfy its original purpose in the context of hydroclimate change under global warming. Here, we present a case study of the Meishan reservoir in east-central China, which was primarily built for irrigation and ﬂood control in the 1950s. We evaluate the impacts of rainfall change on the hedging and releasing rules over the historical period (1969–2008) by instrumental data and future period (2061–2100) based on simulations in a regional rivalry-mitigated scenario from the Coupled Model Intercomparison Project Phase 6. The main conclusions are as follows: (1) the annual total rainfall has a remarkable increasing trend from 2015 to 2100 and the annual precipitation variability exceeds the envelope range during the past 50-year period. The increased precipitation amount mainly occurs in spring (March to May). (2) The optimal regulation cycle is from September to August and from July to June for both historical and future periods. The limiting level during the nonﬂooded season is lower than the operating water level for more than ﬁve months in the historical period, which limits the ability of reservoir regulation and utilization of water resources. However, the water supply is no longer aﬀected by ﬂood control in 2061–2100 because of the redistribution of annual precipitation. (3) The projected irrigation and residential water demands of the Meishan reservoir are stable; thus, the improvement of the total economic beneﬁt will mainly depend on power generation. This case provides a practical guide for many reservoirs serving water supply for small cities in eastern China, where the size of the population and cultivated land area is stagnant and the climate is getting wetter.


Introduction
Reservoir and dam system stores additional water in the rainy season and releases it in drier months, playing a fundamental role in achieving sustainable uses of the water resource. In practice, the storage or empty space to be maintained in a reservoir during different times of the year is determined by the rule curve, which is predefined in the design stage based on long-term (30 years or more) water inflow and demand statistics. Generally, the rule curve is fixed during the lifespan (∼ a hundred years) of reservoirs [1,2]. Almost 50% of hydropower reservoirs worldwide were originally commissioned more than 40 years ago [3], meaning that their rule curves are calculated from climate data in the mid-twentieth century or even earlier times. However, many observed changes in the climate system during the past decades are unprecedented over centuries, including precipitation and temperature-induced evaporation [4], which have already affected the regional runoff [5]. For the rest of the twenty-first century, the precipitation is projected to increase over high latitudes in parts of the monsoon regions and in wet regions over the tropics, but decrease in dry regions including large parts of the subtropics under 1.5°C and 2°C global warming scenarios. e changing patterns of precipitation minus evaporation and runoff are similar to precipitation [4]. Consequently, the predefined reservoir rule curve in the mid-twentieth century may not remain as effective operation guidance in current and future situations. Moreover, aged reservoirs are about to reach the designed service life [3]. Maintaining or demolishing these reservoirs in the future requires an overall assessment of the energy-related benefits and ecological environment impacts, in which the most important issue is whether the reservoirs could still satisfy their original purpose in the context of hydroclimate change under global warming.
Studies on evaluating the impacts of climate change on reservoir management are carried out all over the world, for example, in northwest America [6], southeast Canada [7], eastern India [8], and northeast ailand [9,10]. Among countries owing dam-reservoir systems, China counts more than 40% of the total in the world [11]. us, a large number of cases are conducted in China, located in the Yangtze River basin [12][13][14][15], Xin'an River basin [16], Hai River basin [17], Yuan River basin [18], and Zhu River basin [19]. Some studies [12][13][14][16][17][18][19] quantify the rainfall or streamflow change in the reservoir watershed and project the water budget in the future based on multiple theoretical models, but few of them provide the updated rule curves that could be directly used to guide the gate opening schedule of a dam. Others [6][7][8][9][10]15] develop advanced optimization techniques to calculate rule curves but their future projections are extracted from a single climate model or based on previousgeneration scenarios applied in the Coupled Model Intercomparison Project (CMIP) Phases 3 and 5. Now, the outputs of CMIP6 use a set of new plausible scenarios [4] and the multimodel ensemble is proved to be more reliable than a single model [20], which provides us more possible future pathways to assess the alteration of the rule curve under the climate change.
Over 80% of reservoirs in China were constructed between the 1950s and 1970s, most of which are not operated reliably and effectively now [21]. So far, however, the research in China mainly focuses on the electricity generation change of large or recently constructed hydropower reservoirs [12-14, 16, 18] and mega reservoirs serving the urban agglomerations [15,17,19]. ere has been a little quantitative analysis of the rule curve change for the aged reservoirs. Here, we pay particular attention to the Meishan dam and reservoir system built during 1956-1958. It is located in Anhui Province (east-central China), where the province mean temperature and precipitation are projected to increase 2.3°C (with reference to 1850-1900) and 4.0% (with reference to 1985-2005) under 2°C global warming scenario [22]. Moreover, annual runoff changes in Anhui are basically consistent with precipitation, but the increased rates are different among seasons, that is, higher in spring and winter than in summer and autumn [23]. us, there is an urgent need to update the original regulating rules of Meishan reservoir developed 65 years ago. In this study, we use instrumental data for the past decades and the latest CMIP6 scenario simulations till 2100 to evaluate the potential impacts of climate change on the hedging and releasing strategies. A description of the study area, climate datasets, and statistical methods is given in Section 2.
Regulating rule curves for historical and future periods are presented in Section 3. A summary of the main conclusions and the practical implications of the study case is discussed in Section 4.

Study Area.
Meishan reservoir and dam project spans the Shi River in Anhui Province, eastern China (Figure 1(a)). Shi River is born in the Ta-pieh Mountains and joins the Huai River from the south side between Guan River and the west branch of Pi River (Figure 1(b)). e length of Shi River is 216 km and the catchment area is 6880 km 2 . ere are 11 branches upstream of the dam site, which is 130 km away from the confluence of the rivers Shi and Huai. e upstream watershed controlled by the dam is calculated from the holefilled Shuttle Radar Topography Mission (SRTM) 90 m digital elevation data [24] by using a Matlab TopoToolbox [25]. e watershed area is 1970 km 2 , which is mainly covered by forest and cropland (Figure 1(c)). e subtropical humid monsoon climate dominates the catchment area, where the annual mean temperature, precipitation, evaporation, and aridity are 15.5°C, 1381.5 mm, 1173.6 mm, and 0.8, respectively [26]. Meishan reservoir is a multipurpose facility and its main characteristics [27][28][29] are summarized in Table 1.
e primary purposes are flood control and irrigation. e flood level is designed to accommodate an estimated 100-year recurrence interval event and it will provide the irrigation water needed for the agricultural sector to 25.5 × 10 2 km 2 farmland in five downstream counties (Figure 1(b)), that is, Jinzhai, Huoqiu, Liuan, Gushi, and Shangcheng [30]. Moreover, it releases the ecological flow to maintain the stability of the downstream ecosystem (e.g., biological protection, sediment deposition, and pollution control). However, hydropower generation is not mandatory and the turbines only work during irrigation season. e original capacity of the Meishan hydropower station was 40 MW, and it was increased to 52 MW after a retrofit project in 2012 [28].

Hydroclimate Observation and the Model for Predicting
Inflow.
e monthly inflow data at the dam site from 1969 to 2008 (Figure 2(a)) are provided by the Meishan hydrologic station, which could be used to directly calculate the water supply in the historical period. However, the water inflow is not an output variable of the CMIP6 models and we need to derive water inflow from other variables to calculate the water supply in the future period. Prior studies demonstrated that precipitation is the dominant factor of runoff in the upper-middle reaches of Huai River and its branches, while evaporation makes a minor contribution [31,32].
erefore, we test their relationships with water inflow. e rainfall data  are from a grid precipitation product [33] with a spatial resolution of 0.5°, and we use the data in the grid centered on 115.75°E, 31.75°N (Figure 2(b)) since it overlaps most of the reservoir area. e evaporation data are from the dataset of daily climate data from Chinese surface stations [34], and records of the nearest station    Figure 2(c)). Several regression models are calculated using different combinations of predictors and validated by the leave-one-out procedure ( Table 2), and the result shows that the 2nd degree polynomial fitting by precipitation has the highest predicted variance of 85.99%. is nonlinear relationship is also observed in other reaches of the Huai River as that rains could induce more surface runoff after soil saturation [35]. However, introducing the evaporation parameter would reduce the model skill, such that only precipitation is used to predict future inflow change.

CMIP6 Data and Ensemble
Method. Projected precipitation change is needed to generate the reservoir rule curve for the future. Here, we firstly describe the future scenario selection and then introduce the method to calculate weighted mean precipitation from multiple model outputs. CMIP aims to better understand climate changes in a multimodel context. e model performance is assessed by a set of experiments with idealized and historical radiative forcings. Future climate change simulations are based on designed scenarios [36]. CMIP has developed in phases with expanded number and complexity of climate models. In CMIP5, four Representative Concentration Pathway (RCP) scenarios (RCP2.6, RCP4.5, RCP6, and RCP8.5) represent different assumptions about atmospheric greenhouse gas concentrations based on population, economic growth, energy, and land use over the twentieth century has been developed [37]. ey are labeled after a possible radiative forcing anomaly value (relative to the preindustrial level) in the year 2100. e latest CMIP6 has a new set of paths socalled SSP (shared socioeconomic pathways) scenarios, indicating different future socio-economic projections and political environments [38]. ere are five high-priority SSP scenarios defined in terms of challenges to adaptation and mitigation, from an increasingly sustainable world (SSP1) to a high fossil fuel dependent world (SSP5). Combining the five SSP pathways with the different radiative forcings yields  a scenario matrix and four of them are selected as standard scenarios (SSP585, SSP370, SSP245, and SSP126). It is worth noting that SSP585, SSP245, and SSP126 can be understood as an update or remake of the RCP8.5, RCP4.5, and RCP2.6, while SSP370 is a newly introduced scenario closing the gap between RCP6.0 and RCP8.5. Specifically, SSP370 represents a regional rivalry world that degrades the international priority for addressing environmental concerns and the corresponding forcing by 2100 is 3.7 W/m 2 [38]. On a global scale, the mean Earth surface temperature will increase 3.6°C over 2081-2100 in SSP370 compared to 1850-1900 and this scenario represents the upper limit of plausible future [4]. erefore, hydroclimate change under the SSP370 pathway in the Meishan reservoir and potential operation change are investigated in this study. A total of 34 models containing precipitation simulation for SSP370 are used and their information is summarized in Table 3 [39]. All CMIP6 model data are downloaded from the Climate Explorer developed by the Royal Netherlands Meteorological Institute [40], including the SSP370 experiment from 2015 to 2100 and the historical run (forced with changing conditions that are consistent with observations) from 1850 to 2014, the latter is used to evaluate the model performance in our study area by comparing with instrumental data.
It is fundamentally impossible to describe all true climate processes in a single model and the uncertainties introduced by choices in the model design cannot be addressed by varying its parametrizations. Multimodel ensemble strategy is a way to exploit the advantage and to avoid the weakness of individual models due to substantial reduction of the systematic errors in individual models, and multimodel means often agree better with observations than individual model results [20]. In this study, we compute the multimodel mean by both equally and unequally weighted methods and select the best one to predict future precipitation. Note that some of the models have multiple ensemble members with different initial conditions. erefore, there are three equally weighted multimodel ensemble methods: (1) calculating the mean of all members; (2) calculating the mean of all models with only one member per model; (3) first calculating the means of individual models by corresponding members and then calculating the mean of all models. Moreover, the unequal weights are determined through the best subset regression constrained by maximizing adjusted R-square. All model outputs are resized to 192 × 144 horizontal resolution to calculate the ensemble mean.

Reservoir Rule Curve Calculation.
e reservoir rule curve guides the releasing schedule to impound excess water in wet periods and ensure water demand in dry periods because the water demand change is not consistent with the inflow change. us, except for water inflow data described in Section 2.2, water demand data are also required to calculate the rule curve. Meishan reservoir project is the main water source of the Shi River irrigation area. e irrigation release flow is 190 m 3 /s and the average agricultural water consumption is 10.63 × 10 8 m 3 per year under the designed assurance level of 80%. In addition, the annual urban water demand is 400 × 10 4 m 3 and the downstream ecosystem flow is 5.47 m 3 /s [28]. e total annual water consumption of the aforementioned water-using sectors is 12.36 × 10 8 m 3 . According to the long-term statistics, the month-by-month water demands of the reservoir are shown in Figure 3(a).
To calculate the rule curve, the first step is identifying the regulation cycle. A complete operation cycle ranges from one day to multiple years, determined by the coefficient of reservoir storage capacity β (proportion of regulating storage to multiyear average inflow). e range of β for daily, seasonally, yearly, and multiyear (or carryover) reservoirs is β ≤ 2%, 2%β ≤ 8%, 8%β ≤ 20%, and β > 20%, respectively [41]. In this study, β is equal to 68% (9.57 × 10 8 m 3 / 14.05 × 10 8 m 3 , Table 1) such that the reservoir needs to be carryover regulated. Generally, the horizontal axis of the operating rule curve represents time (month or 10-day period unit), and the vertical axis indicates the reservoir water level or storage capacity. ere are three time-varying curves between normal pool level and dead water level. e first is the flood limited curve, which constrains the reservoir water level in the flood season. e other two are the lower water supply curve (or upper rule curve) and limited water supply curve (or lower rule curve), guiding the water refill and release strategies. We select 3-6 representative hydrological years, in which the water inflow is close to the water demand, to calculate operating rule curves, as well as associated annual and carryover storage, and the step-by-step computation follows the engineering standard guidance from reference [42]. e elevation-volume relationship [43] is plotted in Figure 3

Results
Generally, the regulating year is different from the calendar year, and the former usually starts from the beginning of the flood season and lasts for 12 months. Under the predefined operation schedule of Meishan reservoir, the regulating year is from June to May [43]. However, the reservoir was commissioned more than 60 years ago and the catchment area experienced significant climate change since then. In this section, we first explore how the optimal regulating rules change in the past decades (1969-2008, historical period). en, we present the future precipitation projection over 2061-2100 in the SSP370 scenario and associated rule curve change.

Reservoir Rule Curve of the Historical Time.
e annual cycle of water inflow and demand is shown in Figure 3(a), and the flood season is from May to September. e irrigation water demand peak (in June) is a little earlier than the water inflow peak (in July) during 1969-2008; consequently, the predefined operation schedule (June to May) frequently suffers from water deficit in the early months. Moreover, the water demand cycle is largely overlapped with the water inflow cycle such that it is difficult to identify the appropriate regulating cycle. erefore, we plot all possible  6 Advances in Meteorology cases together to make a comparison (Figure 4). For each case, six representative hydrological years (Table 4) are selected to calculate rule curves. Although the total inflow water amount is adjusted to match the water demand, there exist years that the water supply is damaged (negative storage anomaly) due to the uneven seasonal distribution of water inflow. In these cases, the carryover storage needs to be used although it is not a dry year. erefore, the percentage of damage years is used to screen the appropriate regulating year cycle. On this basis, the best choice is from September to August, which only has one damage year (16.7%). e second optional choice is from July to June (two months later than the beginning of the flood season), which has three damage years (50%). e remaining annual cycles have at least four damage years (>66.7%) while all years are damaged from April to March and May to April. Figure 5 shows the reservoir rule curves starting from July and September, respectively. However, the limiting level during the nonflooded season is 126.0 m, which is lower than the operating water level for seven and five months in the above two cases. Such issues, to some extent, limit the ability of reservoir regulation and utilization of water resources.

Local Precipitation Change in SSP370 Scenario.
Firstly, we evaluate the ability of a set of CMIP6 models to capture the observed changes by Taylor diagram [44] ( Figure 6). For a good simulation, the correlation coefficient with the observation and the ratio of their variance should be close to one, while the centered root mean square error (RMSE) should be close to zero. In this study, we test the performance of multiple models in only one grid; thus, the statistics are calculated based on the historical time series instead of the average spatial pattern used in other studies. e statistics for the majority of models are clustered in the intervals in which the correlation coefficients are between 0.35 and 0.55, the ratios of variance are between 45% and 90%, and the centered RMSEs are between 0.8 and 1. e result indicates that most models have a reasonable performance in simulating temporal change of precipitation in the study site, while none of them shows outstanding skills. e three equally weighted multimodel ensemble means also give results that are similar to the best individual model. Moreover, we build the best subset regression ensemble by seven CMIP6 models whose correlation coefficients are larger than 0.5 and it finally maintains five input variables. e regression equation is as follows: where Y is the ensemble predicted precipitation, X is the model simulated precipitation, and the subscript number represents the model in Table 3. is unequally weighted multimodel ensemble has the adjusted R-square of 0.43 (equal to the correlation coefficient of 0.66), which has 11% more explained variance than the best individual model (IPSL-CM6A-LR, correlation coefficient � 0.57, RMSE � 0.82) and three equally weighted ensemble methods. Based on the best subset regression, the future precipitation is projected to 2100 (Figure 7). Like the temperature, the annual total rainfall also has a remarkable increasing trend. e precipitation amount of the wettest year reaches 2760 mm in the 2090s, while it is only 831 mm in the driest year in the mid-twenty-first century, indicating that the annual precipitation variability exceeds the envelope range during the past 50-year period (Figure 7(a)). e mean annual precipitation is 2022 mm over 2061-2100, which is 500 mm more than that of 1961-2010. e increased precipitation amount mainly occurs in spring (March to May), while July to September experiences a reduction (Figure 7(b)).

Reservoir Rule Curve of the 2061-2100 Period.
Since the monthly distribution of rainfall changes in the future, the reservoir regulating strategy also needs to be adjusted. We assume that the water demand is the same as the current time and plot all possible regulating cycles for 2061-2100 in Figure 8. It is worth noting that the standard deviation of the weighted ensemble mean is reduced by a factor of ��� � 1/N √ since the ensemble members are sequentially independent and possess an identical standard deviation. erefore, the range of representative hydrological years of the multimodel ensemble is also reduced. To eliminate this effect, we pick out three representative hydrological years for seven selected Advances in Meteorology 7 models (Table 5) to generate the rule curves ( Figure 8). Similar to the historical period, the number of damage years is least (48%) from September to August. e second optimal cycle is from July to June and November to October, which both have 70% damage years. e cycles with 100% damage years are increased from two to four. e selected rule curves are provided in Figure 9. Compared with the historical period, the annual storage is increased for the July-June scheme while decreased for the September-August scheme. Due to the redistribution of annual precipitation,     Table 3.

Precipitation (mm)
Time (   the operating water level is reduced during the nonflooded season and it is lower than the flood control level, indicating that the water supply is no longer affected by flood control in the future. As a result, the efficiency of water utilization is enhanced.

Discussion and Conclusions
In this study, we evaluate the impacts of hydroclimate change on the regulating rule curves of the Meishan reservoir constructed in the 1950s, for the historical period (1969-2008) by instrumental data and future period (2061-2100) by CMIP6 simulations. We find that the optimal regulation cycle is from September to August and from July to June for both historical and future periods. e limiting level during the nonflooded season is lower than the operating water level for more than five months in the historical period, which limits the ability of reservoir regulation and utilization of water resources. However, the water supply is no longer affected by flood control in 2061-2100 because of the redistribution of annual precipitation.
erefore, the predefined operating strategy from June to May is not an efficient scheme for either historical or future scenarios. Operation log of the Meishan reservoir records that the water level is lower than the dead water level in nearly 25% of months from 1965 to 2015, and sometimes even lower than the elevation of irrigation water intake of 95.27 m [27]. Because the operation dead water level is lower than design, the utilization rate of water resources is reduced and the conflicts among water demand sectors are inevitable within the hydrological year. In addition, the associated facilities, such as dam, dike, gate, and plant, will deteriorate more quickly under an abnormal operating condition.
Many large reservoirs in China are designed to provide residential water to metropolises. e population in these cities will continue to grow under the urbanization process in future China [45]. erefore, the water supply requirement of the large reservoirs is expected to increase year by year. On the contrary, the Meishan reservoir affords water supply for some small counties, which experience a net outflow of population. For example, the population statistics for Jinzhai County, which is the first residential area downstream of the Meishan reservoir, has dropped by 3.49% while the proportion of people aged 65 or above has increased by 6.33% from 2010 to 2020 [46]. erefore, the demand for urban water is not expected to increase. Furthermore, the irrigation water demand in the future is likely to stay the same as the current level because the cultivated land area in Anhui Province is stable [47]. Without additional water pressure, the improvement of the total economic benefit of the Meishan reservoir will mainly depend on the increase in the benefit of power generation, which will also help China to reach the carbon peak and neutrality targets. Studies have shown that precipitation in many regions of China will increase under global warming scenarios [48]. e case of Meishan reservoir provides a possible reference for the planning and operation of medium-size  Advances in Meteorology 11 reservoirs in these regions.
is study also presents a framework to calculate rule curves from multisource datasets (i.e., geographic raster, construction planning document, and climate data) by using a set of software packages. Rule curves for other similar reservoirs could be calculated following this framework, and the individual dataset or software could be substituted by state-of-the-art products.
Data Availability e shaded relief map in Figures 1(a) and 1(b) is from GEBCO https://download.gebco.net/. e land cover data for 2015 in Figure 1(c) is from http://maps.elie.ucl.ac.be/CCI/ viewer/download.php. e grid monthly precipitation data for China are from http://poles.tpdc.ac.cn/zh-hans/data/ 653191c3-df99-479b-95bd-a18bee30f2bd/. Requests for access to the inflow data of Meishan reservoir should be made to the corresponding author.

Conflicts of Interest
e authors declare that they have no conflicts of interest.