Annual shifts of flow regime alteration: new insights from the Chaishitan Reservoir in China

Reservoir regulation is variable for flow regime alterations and mainly depends on operational objectives and hydro-meteorological conditions. In this study, the flow regime metrics (i.e., magnitude, variability and frequency, duration, timing and rate of change) are adopted to describe variations in both long-term inflow and outflow series of the Chaishitan Reservoir in China. Deviations between the inflow and outflow metrics are calculated to assess the flow regime alterations at annual scale. Further, dimensions of both time and flow regimes are reduced by multivariate statistical analysis, and the regulation patterns and their annual shifts are identified. Results show that: four regulation patterns are identified from 2004 to 2015. The regulation is gradually enhanced over time with typical features of different hydrological years. In dry years, the pattern is slightly regulated flow regimes with slightly discharging stored water and flattening outflow, moderate stability and intermittency. In normal years, the pattern is slightly regulated flow regimes with extremely increasing flow magnitude in the pre-nonflood season, high stability and slight intermittency. In wet years, the pattern is moderately regulated flow regimes with moderately decreasing flow magnitude in the flood season but extremely increasing flow magnitude in the nonflood season, slight stability and high intermittency.

regime metrics are considered to be time-consuming and highly dimensional 25 , they can capture all hydrograph changes. Moreover, multivariate statistical analysis methods are widely adopted to identify the representative characteristics from numerous pieces of flow regime metrics and without losing metric information 29 . Some typical methods include Analysis of Similarity, principal component analysis, cluster analysis and set pair analysis 22,30 . This paper hypothesizes that the impact patterns of reservoir regulations on flow regime characteristics are different across different years. This hypothesis is based on the fact that climate variability, upstream inflow dynamics and operational objectives, as well as reservoir regulation rules, are time-varying 31,32 . Note that most previous flow regime assessments have mainly focused on the annual average steady-state variations of flow regimes and ignored their temporal shifts 6,28,33,34 . The reservoir regulation usually belongs to a certain pattern, such as "run-of-river" for flood control, "storage" for water supply, "peaking" for hydroelectric generation 10,33 , and so forth. Many studies have been implemented to investigate the regulation patterns at the basin 35 , continental 10,20 and global 27 scales. The temporal shifts of dam regulation are still poorly understood, which can cause understanding deviations in the dam regulation impacts, and restrict scientific dam regulations.
Therefore, this paper elaborates on the annual shifts of flow regime alteration by reservoir regulation. The investigation focuses on the Chaishitan Reservoir ( Fig. 1), which is the leading reservoir of the upstream cascade development of the Pearl River and is one of the major water projects of the 13th Five-Year Plan of China (2016)(2017)(2018)(2019)(2020). This reservoir plays a critical role in protecting the environment of plateau lakes and controlling stony desertification, which are of great significances for the socio-economic development and ecological construction of the upstream provinces of the Pearl River Basin. The principal objectives of this study are to: (1) describe the long-term flow regime characteristics of both reservoir inflow and outflow in terms of the magnitude, frequency, variation, duration, timing, and rate of changes; (2) assess their inter-annual variations based on the deviation of synchronous inflow and outflow regime characteristics; and (3) identify representative impact Therefore, the regulation of the Chaishitan Reservoir alternately decreases or increases the mean daily runoff in individual years and the flood seasons at the slight or moderate grade over time, and increases obviously in the nonflood seasons, particularly in the pre-nonflood seasons; as a result, the regulation decreases flow variability except in the pre-nonflood seasons, and it decreases the frequencies of low flow events whereas it increases the frequencies of high flow events, particularly for extreme high flow events. It also extends the flow event durations, increases the mean rates of both positive and negative changes in flow before 2011 and then turns to decrease the rates. However, the impacts on flow timing are without obvious rules during the study period. The most impactful Impact pattern classifications. The impact grade series of all flow regime metrics can be composited into four, five, six, seven and eight independent principal components, which explain 76.45%, 83.90%, 89.30%, 93.73% and 96.24%, respectively, of the entire impact grade series (Fig. 3a). All the component compositions are individually adopted for cluster analysis; the cluster performance results are shown in Fig. 3b and Table 1. The four-cluster classification for the four principal components is the most robust with the largest Goodman-Kruskal index (GKI: 0.76) and the smallest C index (CI: 0.10). All of the class sizes are also no less than two.
The annual variations in flow regime in four years (2004,2005,2006 and 2013) belong to Class 1, accounting for 33.3% of the total period. The regulation of the Chaishitan Reservoir slightly increases the mean daily flow magnitude in the flood season, the nonflood season and the whole year. For the flow variability and frequency,  the regulation moderately decreases the Coefficient of Variation (CV) of the mean daily flow in the flood season, the nonflood season and the whole year, and it decreases the frequencies of both low and high flow events, except FDF10 (Fig. 4). Moreover, the regulation slightly increases the durations of all flow events (except DDF25), moderately increases the zero-flow days, delays the Julian dates of annual maximum daily flows, and slightly increases the mean rates of both positive and negative changes in flow events. Thus, the impact pattern of the Chaishitan  527 mm to 960 mm, with the mean of 720 mm being close to that of dry year (p = 75%, 697 mm). The regulation mainly increases the flow magnitude and its variability in the pre-nonflood season to relieve the severe drying of the downstream rivers 10,33 . The main function is also the frequent discharge of the stored water in the nonflood season for downstream water consumption, particularly for the agricultural irrigations in the spring. The regulation is also similar to that of the "run-of-river reservoir" 10,33 but is strengthened in the flow magnitude in the pre-nonflood season and the flow stability compared to the metric variations of Class 1. During the last subperiod from 2014 to 2015, the main impact patterns are in Class 4 (MRF:MDM-SS-HI). The annual precipitation amounts range from 938 mm to 1137 mm, which are close to or much greater than that of wet year (p = 25%, 940 mm). The regulation mainly decreases the flow magnitude in the flood season, but increases the flow magnitude in the nonflood season, thereby resulting in more stable flow with lower frequency and longer durations of both the high and low flow events 2,10,33 . This function is generally accepted as a conventional water quantity regulation for most reservoirs in the world, i.e., storing water and reducing the peak flow in the flood season for the safety of flood control ("storage reservoir"), while discharging water in the nonflood season for downstream water consumption ("run-of-rive reservoir") 2 .
Therefore, because of the annual variations in both upstream precipitation and inflow, the regulation of the Chaishitan Reservoir shows remarkable annual shifts to guarantee the downstream water consumption of residential needs and agricultural irrigations (e.g., Kunming city), and to ensure the safety of flood control. Compared to the inflow regime metrics, the outflow magnitudes are increased slightly in most years, particularly in the pre-nonflood season (January~March) and flood season (April~September), and are decreased at the moderate impact grade around 2014. The outflow variation is interchanged between variability and stability, and the flow event durations are extended whose increments are gradually shifted from the moderate grade to the slight grade. The impacts of outflow intermittency are changed from the slight grade to the high grade, and the Julian dates of both the annual maximum and minimum daily outflows are delayed, being shifted from the slight grade to the high grade.

Discussion and Conclusions
This paper tests the hypothesis that the impact patterns of reservoir regulation are variable, rather than stable over time. The results of the case study of the Chaishitan Reservoir show that: (1). The regulation of the Chaishitan Reservoir slightly or moderately increases the mean daily flow magnitude in the flood season of most years, and even obviously increases that in the nonflood season over time, particularly in the pre-nonflood season. These phenomena occur because Chaishitan is a multiyear regulating reservoir of large storage capacity, which is strongly advantageous for changing the interannual distributions of outflow, i.e., to store runoff in wet years (e.g., 2004, 2005) and release in dry years (e.g., 2010~2012). Thus, the daily flow regimes gradually become stable, except in the pre-nonflood season, and the frequencies of low flow events are decreased, whereas those of high flow events are increased, particularly for extreme high flow events at the extremely impact grade. Moreover, the flow event durations are slightly extended, and the mean rates of both positive and negative changes of flow events increased before 2011 and then decreased. However, the impacts on flow timing do not exhibit obvious temporal patterns during the entire period. The most impactful metrics are the mean daily runoff magnitude in the pre-nonflood season (January-March), the frequencies of low flow events and extreme low flow events, the zero-flow days and the Colwell's Constancy of mean daily runoff.
(2) Four annual impact patterns are identified for the regulation of the Chaishitan Reservoirs and obviously annual shifts are detected in the whole period, which are divided into three subperiods. The regulation is gradually enhanced with typical features of different hydrological years in a comprehensive manner. In dry years, the main regulation characteristic is the slightly regulated flow regime with slightly discharging the stored water and reduction of the peak outflow in the whole year, moderate stability and moderate intermittency. In normal years, the main regulation characteristic is the slightly regulated flow regime with extremely increasing flow magnitude in the pre-nonflood season, high stability and slight intermittency. In wet years, the main regulation pattern is the moderately regulated flow regime with moderately decreasing the flow magnitude in the flood season but extremely increasing flow magnitude in the nonflood season, along with slight stability and high intermittency.
The regulation characteristics are closely related to the annual variations of upstream precipitation or inflow, operational objectives, and so forth. The annual shifts are typically used to ensure the downstream water requirements of both production and residence, as well as the safety of flood control. The main regulation of the reservoir still focuses on the water resource utilization, that causes the flow regime to be stable and intermittent, with more regulation-induced flow events of longer durations. All of the current regulations promote the local developments of agriculture and residence. However, the potential environmental consequences should be not ignored, e.g., sedimentation in the reservoir caused by the more stable flow, aquatic organisms experiencing changes from turbulent flows to quiescent water pools, and the diversity loss of the aquatic river system in the downstream caused by the high intermittency of low flow events (particularly the zero-flow days). The future regulation of the Chaishitan Reservoir should further consider the river ecosystem health and environmental deterioration to relieve its negative impacts. Some proposed regulation strategies include weakening the interannual regulation capacity, inducing more low flow events with short duration in the nonflood season, and increasing the peak flow in the flood season. Additionally, future studies should also be implemented to consider the responses of river ecosystem to flow alteration, multi-objective optimization for interannual regulations, reservoir ecosystem services evaluation and environmental flow assessment.

Material and Methods
Study area and data source. The Chaishitan Reservoir is located in the Nanpan River, upstream of the Pearl River in Yunnan Province, South China (Fig. 1) (Table 2). In the Chaishitan catchment, the total area is 4,469 km 2 and the major land covers are forest (46.17%) and paddy (46.02%). The climate characteristic belongs to the temperate climate region with dry winter and warm summer, with an annual average precipitation and temperature of 778 mm and 21.4 °C, respectively. There are two inflow tributaries, i.e., the Maitian River and Nanpan River.
Long-term observed inflow and outflow series are collected for our study. The daily runoff series from 2004 to 2015 are collected from the Maitian hydrological station in the Maitian tributary, the Tianshen hydrological station in the upper Nanpan River and the Chaishitan hydrological station in the reservoir downstream, respectively. The controlled areas of the Maitian and Tianshen stations are 439.7 km 2 and 3,786.7 km 2 , the sum of which covers the vast majority of the Chaishitan catchment (94.6%). Therefore, the sum of the daily runoff series at the Maitian and Tianshen hydrological stations is considered to represent the inflow series of the Chaishitan Reservoir, and the runoff series at Chaishitan station is the outflow series. The inflow ranges from 0.00 m 3 /s to 666.19 m 3 /s (July 4 th , 2015) with a mean of 33.33 m 3 /s, and the outflow ranges from 0.00 m 3 /s to 469.50 m 3 /s (August 26 th , 2015) with a mean of 32.85 m 3 /s. The entire period also contains wet, normal and dry years.

Methods
Flow regime metrics. The entire flow regime is usually fully characterized by five basic components, including magnitude, variability and frequency, duration, timing, and rate of changes; these components have been widely used to characterize long-term hydrographs in a detailed manner 18 . Twenty-two metrics are selected in our study to cover all of the basic components, i.e., four magnitude metrics, eight variability and frequency metrics, five duration metrics, three timing metrics and two metrics regarding the rating of change 24,36 (Table 3) Impact assessment of individual metrics. The discrepancies between inflow and outflow metrics are adopted to assess the impacts of reservoir regulation. The relative deviation is used for the impact assessments of magnitude, frequency and variability, rating of change, and the absolute deviation is used for the impact assessments of duration and timing. The equations are given as follows.   regulation positively impacts the flow regime metric, whereas if it is less than 0.0, then the regulation negatively impacts the flow regime metric; when the deviation is equal to 0.0, the regulation has no impacts. To clearly present the regulation impacts, the deviation is further categorized into five impact grades based on percentiles or days:no (i.e., −5% ≤ AD ≤ 5%, −3 days ≤ RD ≤ 3 days), slight (i.e., −25% ≤ AD < −5% or 5% < AD ≤ 25%, −15 days ≤ RD < −3 days or 3 days < RD ≤ 15 days), moderate (i.e., −50% ≤ AD < −25% or 25% < AD ≤ 50%, −30 days ≤ RD < −15 days or 15 days < RD ≤ 30 days), high (i.e., −75% ≤ AD < −50% or 50% < AD ≤ 75%, -91 days ≤ RD < −30 days or 30 days < RD ≤ 91 days) and extreme (i.e., AD < −75% or AD > 75%, RD < −91 days or RD > 91 days) 22,34 . Dimensionality reduction. High dimensionality and multicollinearity are key obstacles to flow regime analysis when a large number of metrics are used. As classical statistical methods, principal component analysis has strong advantages in merging numerous correlated metrics into a few independent composite components without losing metric information, and cluster analysis is quite competitive in identifying some typical characteristics from numerous chaotic information 19,34 .
In this study, the annual impact grades of all flow regime metrics are applied to further identify the impact patterns of reservoir regulation on the entire flow regimes, and their annual shifts. Principal component analysis is adopted to reduce the multicollinearity among all 22 flow regime metrics. If the threshold of cumulative variance of all metrics is no less than 75% of the total variance, then the first m principal factors are selected for further analysis. Subsequently, cluster analysis is adopted to cluster the impact grades of 12 years into some temporally  representative variations. Therefore, the impact patterns of dam regulation on the entire flow regimes are identified, and the annual shifts are the alternate variations between different impact patterns over time. The Euclidean distance 37 is used to calculate the similarity of the principal components between different years, and then Ward's method 38 is used for hierarchical clustering of the principal component series. Moreover, the Goodman-Kruskal index (GKI) 39 , the C index (CI) 40 , and the minimum cluster size are used to assess the cluster performance and determine the final class number. The most robust clustering is determined by the greatest value of GKI and the lowest value of CI 22 . All class sizes should be no less than two to ensure the representativeness and integrity of the individual classes. Both principal component analysis and cluster analysis are implemented in the R platform (version 3.1.1) 41 ; specifically, they are based on the princomp function in the stats package 41 , the hcluster function in the amap package 42 and the cluster.stats function in the fpc package 43 .