Impact of Three Gorges Reservoir Operation on Water Level at Jiujiang Station and Poyang Lake in the Yangtze River

Abstract

The variation in water level at Jiujiang Station (JJS) directly affects flow exchange between the Yangtze River and the Poyang Lake. Quantitative research on the influencing factors of water level changes at JJS is of great importance for water supply and eco-environment protection in the Poyang Lake region. In this study, the Mann-Kendall method was used to test the trend of water level variation, and the impacts of riverbed incision and flow volume changes on water level at JJS were macroscopically analyzed using the observed monthly flow data series from 1981 to 2021. Furthermore, Long Short-Term Memory (LSTM) neural network model was used to simulate the impacts of outflow discharge of Three Gorges Reservoir (TGR) and flow discharge of the interval basin between TGR and JJS on water level at JJS; the partial dependence plot was adopted to analyze the impact of single feature variable variation on the simulation results. The results show that, after the TGR was put into operation in 2003, the water level changes at JJS mainly occurred during the impoundment period, the annual average storage of TGR was decreased 6.9 billion m³, and the annual average runoff volume at JJS was decreased 11.5 billion m³, which resulted in the average water levels at JJS being decreased 1.74 m and 2.11 m in September and October, respectively. The annual average runoff of JJS was increased 4.5 billion m³ with TGR replenishment of 1.8 billion m³ from December to March of the following year. Impacted by riverbed incision, the water levels at JJS were decreased 0.59 m and 0.99 m in September and October and increased 0.63 m from December to March. Every additional 5000 m³/s (1000 m³/s) of TGR outflow discharge could increase 1.0 m (0.16 m) the water level at JJS in September and October (from December to March of the following year).

1. Introduction

Poyang Lake, being the largest freshwater river-connected lake in China, plays a vital role as a regulating reservoir in the middle and lower reaches of the Yangtze River [1]. It exhibits a distinctive hydrological pattern during high- and low-water-level periods, which is a crucial factor in regulating flow of the Yangtze River and maintaining the ecological balance in this region [2]. Under the influences of global climate changes and human activities, the interaction between the Yangtze River and Poyang Lake has been intensified in recent years [3]. Poyang Lake is now facing problems such as an early and prolonged low-water-level period, shortage of water resources, ecological degradation of wetland [4], etc., which severely affect the water supply and ecological safety of the lake area and restrict the social and economic development in this region [5]. These variations and interactions have become a focus of hydrological research both domestically and internationally [6].

The Poyang Lake is connected to the Yangtze River, and its water level and water surface area are affected by both the inflows from its catchment and the water level of the Yangtze River [7]. From April to June, the water level of Poyang Lake rises due to the increase in inflow from its catchment area and remains high from July to August because of the hydrological barrier formed by the elevated water level of the Yangtze River [8]. Subsequently, starting from September, the water rapidly flows out of the lake as the water level of the Yangtze River decreases, leading Poyang Lake into the dry season until March of the following year [9].

The construction and operation of Three Gorges Reservoir (TGR) and other hydraulic facilities in the upper Yangtze River basin, together with the changes in river channel erosion in the middle and lower reaches of the Yangtze River, have significantly altered the interactions and flow exchange dynamics between the Yangtze River and Poyang Lake [10]. Since the TGR started operation in 2003, it has remarkably regulated the flow discharges and water levels in the middle and lower reaches of the Yangtze River [11]. During the TGR impoundment period from September to October, the flow discharges and water levels in the middle and lower reaches decrease notably, reducing the backwater effect of the Yangtze River on Poyang Lake and increasing the amount of lake water flowing into the Yangtze River [12]. Liu et al. suggested that the recent water level decline of Poyang Lake was mainly due to the weakened backwater effect of the Yangtze River [13]. Jiang et al. pointed out that the TGR plays a dominant role in changing the downstream flow regime [14]. During the TGR seasonal release operation from December to June, the flow discharge and water level of the main stream in the Yangtze River is increasing and consequently raising the water level and surface area of the Poyang Lake [15].

The water level of the Yangtze River is also influenced by the flow discharge and water level relationship curve [16]. Over the past two decades, alterations in the riverbed topography have had a significant impact on this relationship, especially in the middle and lower reaches of the Yangtze River [17]. As a result, the shift in the flow discharge and water level relationship curve of the Yangtze River may serve as another factor contributing to the change in the backwater effect on Poyang Lake, in addition to the direct flow regulation by TGR [18].

Besides the variation in the water level of the Yangtze River, sand mining in the Poyang Lake is another key factor affecting the flow discharge conditions and contributing to the reduction in the water surface area [19]. Since 2000, the increasing demand for construction sand has led to more intensive sand mining activities in the lake area, significantly changing its bathymetry [20]. Jiang et al. [21] reported that, by the end of 2010, the total amount of sand extracted from the Poyang Lake was equivalent to 6.5% of the lake’s storage capacity at a water level of 18 m. These sand mining activities are mainly concentrated in the northern part of the lake, where they have widened and deepened the originally narrow drainage channels [22]. As a result, the drainage capacity of the lake has increased, facilitating a more efficient outflow into the Yangtze River through the expanded channels [23]. Some studies have suggested that sand mining is a major contributor to the significant decline in the lake water level in recent decades [24].

The operation of the TGR and other major reservoirs in the upper reaches of the Yangtze River, along with the changes in hydrologic regime caused by river channel improvements and sand dredging in rivers and lakes, have changed the spatial–temporal distribution of water resources in the Yangtze River basin [25]. The interception of floods and reduction in peak flow discharge during the flood season, centralized reservoir refill operation during the impoundment period, and organized replenishment during the drawdown period have significantly affected the flow exchange dynamics between the Yangtze River and Poyang Lake.

JJS is an important hydrologic station located near the outlet of Poyang Lake that divides the middle and lower reaches of the Yangtze River. The water level change at JJS is an important basis for analyzing the relationship between the Yangtze River and Poyang Lake. The uneven spatial–temporal distribution of TGR inflow runoff, TGR regulation, and riverbed incision are the main factors contributing to the water level change at JJS. However, due to the complexity of the flow discharge from the TGR–JJS interval basin and the insufficiency of dataset and hydraulic models, current studies mainly focus on the quantitative analysis of the influence of TGR regulation on flow discharge at JJS to judge its impact on the flow exchange relationship between the Yangtze River and Poyang Lake [26]. Only a few studies have conducted a quantitative analysis of the response relationship between reservoir regulation and water level changes at JJS.

The aims of this study are to quantitatively analyze the impact of TGR operation on water level at JJS and Poyang Lake and to provide a basis for the TGR scheduling decision making. This paper will analyze the water level changes at JJS before and after the operation of the TGR in 2003, calculate the contributions of the flow volume changes during the TGR impoundment and drawdown periods, assess the riverbed incision to the water level change at JJS, and establish the relationship between the TGR outflow discharge and the water level at JJS using the Long Short-Term Memory (LSTM) neural network. The rest of this paper is organized as follows. The study basin and materials are introduced in Section 2. Section 3 presents the methodologies. Section 4 analyzes and discusses the results. Conclusions are given in Section 5.

2. Study Area and Materials

The Poyang Lake basin, covering an area of 162,200 km², encompasses the cities of Jiujiang, Nanchang, and Shangrao in Jiangxi Province, China, and is situated on the south bank of the middle reaches of the Yangtze River. Figure 1 presents a schematic map of the Yangtze River and Poyang Lake basins as well as the main rivers. As illustrated in Figure 1b, there are five rivers, namely Xiushui, Ganjiang, Fuhe, Xingjiang, and Raohe Rivers, flowing into the Poyang Lake. Figure 1c shows the location of Jiujiang Station and the flow exchanges between the Yangtze River and Poyang Lake, which depends on the water level fluctuations and differences.

JJS is the hydrological station closest to the mouth of Poyang Lake in the main stream of the Yangtze River, and its flow discharge and water level relationship can reflect the hydrological regime of the main stream of the Yangtze River near the mouth of Poyang Lake. The observed water level and flow discharge data series at the Yichang Station (outlet of TGR) and Jiujiang Station (JJS) from 1981 to 2021, the flow discharge between TGR–JJS interval basin, and TGR outflow discharge from 2003 to 2021 are adopted to quantitatively analyze the influencing factors of the water level change at JJS, so as to determine flow exchange between the main stream of the Yangtze River and Poyang Lake. The observed flow discharge data series of Yichang Station from 1981 to 2002 is not regulated by the TGR and the observed TGR outflow discharge from 2003 to 2021 is used. The flow discharge of TGR–JJS interval basin is the difference between the flow discharge at JJS and outflow discharge of TGR.

Under the influence of the interaction flow exchange between the Yangtze River and Poyang Lake, the relationship of daily flow discharge and water level at JJS in 2023 is rather dispersed, as depicted in Figure 2. In the dry season, the relationship between the flow discharge and water level is relatively statable. During the flood season, the relationship of flow discharge and water level forms a loop rating curve due to unsteady flow condition and river bed erosion, which implies that the water level at JJS is affected not only by the upstream flow discharge but also by the downstream backwater effect.

3. Methodologies

3.1. Mann-Kendall Test

The Mann-Kendall test is a kind of nonparametric statistics test method that has been widely used for testing the trend of hydrological variables. The data samples are neither required to follow certain distribution nor inferred by few abnormal values and, hence, are suitable for the study of hydrological time sequence trends [27]. The statistics are defined under the assumption of a random and independent time sequence:

$$U{F}_k=(s_k-E(s_k\left)\right)/\sqrt{Var\left(s_k\right)},(k=\mathrm{1,2},\cdots ,n)$$

(1)

where $E\left(s_k\right)$ and $Var\left(s_k\right)$ are the mean value and variance of cumulative number $s_k$, which are calculated as follows:

$$E(s_k)={\displaystyle \frac{n(n-1)}{4}}$$

(2)

$$Var(s_k)={\displaystyle \frac{n(n-1)(2n+5)}{72}}$$

(3)

The opposite sequence $(x_n,x_{n-1},\cdots ,x_1)$ of the time sequence $x$ is generated and the above calculation process is repeated, making $U{B}_k=-U{F}_k$, $k=n,n-1,\cdots ,1$, and $U{B}_1=0$. The sequence shows a rising trend apparently if the value of $U{F}_k$ is greater than 0 and shows a declining trend if smaller than 0. When it is beyond the critical line, it indicates an obvious rising or declining trend. The range beyond the critical line is identified as the time domain of mutation. If the two curves $U{F}_k$ and $U{B}_k$ have an intersection that is between the critical line, the time moment corresponding to this intersection is the start time of mutation [28].

3.2. LSTM Neural Network Model

The LSTM neural network is a type of recurrent neural network (RNN) designed to model sequential data. Unlike traditional RNN, LSTM is capable of learning long-term dependencies by utilizing a memory cell, which can store information over long periods of time. This is achieved through a set of gates: input, forget, and output gates, which regulate the flow of information and allow the model to retain or discard information as needed.

The LSTM model is particularly effective in tasks involving time-series data, natural language processing, and other applications where the sequence of events or data points is important, such as flood forecasting, speech recognition, and stock market prediction. The key computational formulas for the LSTM neural network are as follows, involving the calculations for the input gate, forget gate, and output gate.

The forget gate determines what information will be discarded from the cell state, which is calculated by the following formula:

$$f_t=\sigma (W_f·[h_{t-1},x_t]+b_f)$$

(4)

where $f_t$ is the output of the forget gate, $\sigma$ is the sigmoid activation function, $W_f$ is the weight matrix, $h_{t-1}$ is the previous time step’s hidden state, $x_t$ is the current input, and $b_f$ is the bias term.

The input gate determines what new information will be stored in the cell state, and the calculation formula is:

$$i_t=\sigma (W_i·[h_{t-1},x_t]+b_i)$$

(5)

$${\tilde{C}}_t=tanh(W_C·[h_{t-1},x_t]+b_C)$$

(6)

where $i_t$ is the output of the input gate, ${\tilde{C}}_t$ is the candidate cell state, $W_i$ and $W_C$ are the weight matrices, $b_i$ and $b_C$ are the bias terms, and $tanh$ is the hyperbolic tangent activation function.

The cell state carries the long-term memory of the model and is updated through the forget and input gates, and the update formula is:

$$C_t=f_t\ast C_{t-1}+i_t\ast {\tilde{C}}_t$$

(7)

where $C_t$ is the current cell state and $C_{t-1}$ is the previous time step’s cell state.

The output gate determines the current hidden state, which serves as the output of the network, and the calculation formula is:

$$o_t=\sigma (W_o·[h_{t-1},x_t]+b_o)$$

(8)

$$h_t=o_t\ast tanh\left(C_t\right)$$

(9)

where $o_t$ is the output of the output gate, $h_t$ is the current hidden state, $W_o$ is the weight matrix, and $b_o$ is the bias matrix. These formulas work together to enable the LSTM to effectively handle time-series data, overcoming the gradient vanishing problem present in traditional RNN, and excel in learning long-term dependencies [29].

3.3. Partial Dependence Plot

Partial dependence plot (PDP) is adopted to analyze the marginal impact imposed by one or two features on the predicted results [30]. The partial dependence function used for regression analysis is defined as follows:

$$f_s(z_s)=E_{z_c}[\widehat{f}(z_s,z_c)]=\int \widehat{f}(z_s,z_c)p_c(z_c)d{z}_c$$

(10)

where $z_s$ is the feature to be plotted by partial dependence function, $z_c$ are other features used by $\widehat{f}$, $c$ is the complementary set of $s$, and the combination of feature vector $z_s$ and $z_c$ compose the total feature space $z$. $p_c(z_c)$ is the marginal probability distribution $p_c(z_c)=\int p(z)d{z}_s$ of $z_c$.

Using the training data series, a partial function is calculated by the following formula [31]:

$$\overline{f_s}(z_s)={\displaystyle \frac{1}{n}}\sum _{i=1}^n\widehat{f}(z_s,z_{i,c})$$

(11)

3.4. Evaluation Metrics

Several evaluation metrics, i.e., Nash–Sutcliffe efficiency coefficient (NSE), total runoff relative error (RE), and mean flow absolute error (MAE), are used to evaluate the performance of the LSTM model, which are calculated by the following formulae:

$$NSE=1-{\displaystyle \frac{{\sum }_{i=1}^N{(Q_{o,i}-Q_{f,i})}^2}{{\sum }_{i=1}^N{(Q_{o,i}-\overline{Q_o})}^2}}$$

(12)

$$RE={\displaystyle \frac{{\sum }_{i=1}^N{Q}_{f,i}-{\sum }_{i=1}^N{Q}_{o,i}}{{\sum }_{i=1}^N{Q}_{o,i}}}\times 100\%$$

(13)

$$MAE={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N\left|Q_{o,i}-Q_{f,i}\right|$$

(14)

where $N$ is the number of the sample size and $Q_o$ and $Q_f$ denote the observed and simulated flow discharges, respectively.

The NSE ranges from negative infinity to 1, with a large NSE value indicating a better fit between the observed and simulated flow values. The small MAE and RE values indicate high simulation accuracy.

4. Result Analysis and Discussion

4.1. Hydrological Change Trends at JJS

(1)

Change trend of water level

A statistical analysis of the water level characteristics at JJS from 1981 to 2021 was conducted with the dataset grouped into intervals of 5 years. As depicted in Figure 3, among these intervals, the annual average of monthly water level at JJS from 2006 to 2010 was the lowest, with 10.7 m, whereas, from 1981 to 1985, it was the highest, with 12.3 m. By examining the trend of the ratio between the mean and median water levels, it was found that the water level at JJS exhibited a downward tendency from 1981 to 2010 and then experienced a slight increase after 2010.

Table 1. The annual average of monthly water level at JJS before and after the operation of TGR in 2003 (m).

Month	Sept.	Oct.	Dec.	Jan.	Feb.	Mar.
1981–2002	15.10	13.42	8.56	7.48	7.70	9.07
2003–2021	13.35	11.31	7.77	7.46	7.57	9.19
Difference	−1.74	−2.11	−0.78	−0.02	−0.13	0.12

shows the annual average of monthly water level at JJS in September, October, and from December to March of the following year before and after the operation of TGR in 2003. Compared to 1981–2002, the monthly average water level at JJS during 2003–2021 was decreased −1.74 m and −2.11 m in September and October, reduced by −0.02 m and −0.13 m in January and February, and risen 0.12 m in March.

Figure 4 shows the Mann-Kendall test curves of monthly water levels at JJS in September, October, and December. As illustrated in Figure 4, the Mann-Kendall test method was utilized to compute the curves of $U{F}_k$ and $U{B}_k$ for the water level at JJS in September, October, and December, respectively. The water level at JJS $U{F}_k$ has been less than 0, signifying a downward trend, in September since 1984. After 2014, the $U{F}_k$ curve fell below the critical line, clearly demonstrating a significant decline in the water level at JJS. Regarding the water level at JJS in October, $U{F}_k$ has been below 0 since 1984, indicating a declining trend. After 2002, the $U{F}_k$ curve crossed the critical line and even dropped below the September minimum, suggesting that the degree of decline in October was more pronounced than that in September. The water level at JJS in December also displayed a downward trend but did not cross the critical line, indicating that the water level decreases in December were not conspicuous.

(2)

Change trend of flow discharge

Figure 5 shows a box plot of the annual average of monthly flow discharges at JJS from 1981 to 2021 with an interval of five years. The average and median flow discharge at JJS from 2006 to 2010 were 20,961 m³/s and 17,000 m³/s, respectively, which were the lowest values from 1981 to 2021. In comparison with the water level characteristics from 2006 to 2010, it is observed that dry years occurred in the upper Yangtze River basin.

Table 2. Monthly average flow volume at JJS before and after the operation of TGR (billion m³).

Month	Sept.	Oct.	Dec.	Jan.	Feb.	Mar.
Z	−1.62	−2.4	2.01	4.59	3.58	2.29
1981–2002	91.0	72.8	30.9	25.3	23.9	33.9
2003–2021	80.1	60.7	32.5	31.0	28.5	39.9
Difference	−10.9	−12.1	1.6	5.7	4.7	6.0

presents the annual average of monthly flow volume of JJS before and after the operation of TGR in 2003. The monthly average flow volume in September and October from 2003 to 2021 was decreased by 10.9 billion m³ and 12.1 billion m³, respectively, while, from December to March of the following year, the monthly average flow volume was increased by 4.5 billion m³, with the largest increasement occurring in January. The Mann-Kendall standard normal statistics were Z = −1.62 in September and Z = −2.4 in October, both indicating a downward trend of the runoff. From December to March of the following year, the standard normal statistic Z values were all greater than 2, suggesting an upward trend of the flow volume, which was most prominent in January.

4.2. Inflow Volume Changes of TGR

(1)

Inflow volume changes of TGR

Table 3. The annual average of monthly inflow volume of TGR before and after operation in 2003 (unit: billion m³).

Month	Sept.	Oct.	Dec.	Jan.	Feb.	Mar.	Annual
Z	−1.73	−1.46	2.07	5.14	4.59	4.27	−1.1
1981–2002	63.7	46.3	15.8	11.8	9.8	12.2	434.5
2003–2021	59.5	42.4	16.4	14.5	11.9	15.5	414.3
Difference	−4.2	−3.9	0.7	2.7	2.2	3.3	−20.2

presents the monthly average inflow volume before and after TGR operation in 2003. From 1981 to 2002, the annual average inflow volume of TGR was 434.5 billion m³. Specifically, monthly average flow volumes in September, October, and December were 63.7, 46.3, and 15.8 billion m³, respectively. After TGR operation in 2003, the annual average flow volume dropped by 20.2 billion m³ and decreased 4.2 and 3.9 billion m³ in September and October, while the flow volume increased by 0.7, 2.7, 2.2, and 3.3 billion m³ from December to March of the following year, respectively.

The Mann-Kendall standard normal statistic Z values in September and October were less than Z = −1, indicating a decline in both the annual and monthly (September and October) inflow volume of TGR. From December to March of the following year, the values exceeded Z = 1, suggesting an increase in inflow volume during these months, which aligns with previous analysis. Considering the annual inflow volume, the successive operation of upstream reservoirs in the upper Yangtze River only regulated the annual flow process without altering the total runoff volume. The overall upstream inflow volume exhibited a downward trend. In September and October, the inflow volume of TGR decreased more significantly due to the refill operation of upstream reservoirs in the end-of-flood season. From December to March of the following year, the increasement of TGR inflow volume was mainly due to the centralized replenishment of the upstream reservoir group during the drawdown period.

4.3. Flow Volume Changes at JJS

According to the principle of water balance, the flow volume at JJS includes TGR inflow volume, TGR outflow volume, and the flow volume from the TGR–JJS interval basin.

Table 4. The annual average of monthly flow volume changes at JJS from different regions after TGR operation in 2003 (billion m³).

Month	Sept.	Oct.	Dec.	Jan.	Feb.	Mar.
Flow volume change at JJS	−10.898	−12.088	1.600	5.755	4.678	5.962
TGR inflow volume change	−6.656	−7.074	0.746	1.674	2.804	2.082
TGR outflow volume change	−4.200	−3.900	0.700	2.700	2.200	3.300
Interval flow volume change	−0.042	−1.114	0.154	1.381	−0.326	0.580

lists the annual average of monthly flow volume changes at JJS from different regions after TGR operation in 2003. During the TGR impoundment period in September, the flow volume at JJS was reduced 10.898 billion m³, in which TGR inflow runoff volume was reduced −6.656 billion m³ due to refill operation of the upper stream reservoir group, TGR impoundment storage was 4.2 billion m³, and TGR–JJS interval basin flow volume was decreased 0.042 billion m³. Meanwhile, in the drawdown period, the TGR outflow volume was increased 0.7, 2.7, 2.2, and 3.3 billion m³ from December to March of the following year, respectively, and the flow volume at JJS was increased in the whole drawdown period.

4.4. Water Level Changes at JJS

(1)

Factors of flow discharge and riverbed incision at JJS

The water level change at JJS is mainly affected by flow discharges, riverbed incision, and other factors, among which the flow discharges at JJS come from TGR outflow discharge and interval basin flow discharge. The differences in the flow volume and water level relationship curve at JJS before and after TGR operation in 2003 are used to distinguish the influences of flow discharge and riverbed incision on water level at JJS.

Figure 6 presents a schematic diagram of the flow volume and water level relationship before and after TGR operation in 2003. These are presumed to represent the two water level curves before and after TGR operation in 2003. The monthly average flow volume at JJS before 2003 is denoted as $V_2$ and the corresponding water level is denoted $H_3$. After 2003, the flow volume at JJS is denoted as $V_1$ and the corresponding water level is denoted as $H_1$. $∆H=∆H_1+∆H_2$ represents the water level variations resulting from the combined effects of flow volume reduction and riverbed incision before and after 2003, in which $∆H_1=H_3-H_4$ indicates the water level changes without the influence of riverbed incision; $∆H_2=H_4-H_1$ represents the riverbed incision under the same $V_1$ conditions. The above analysis elucidates the trajectory of water level decline at JJS.

Based on the aforementioned principles,

Table 5. Impact of riverbed incision and flow volume on water level changes at JJS (m).

Month	Sept.	Oct.	Dec.	Jan.	Feb.	Mar.
Riverbed incision	−0.52	−0.66	−0.99	−1.02	−1.05	−0.89
Flow volume change	−0.94	−1.28	0.23	0.82	0.68	0.8
Total change	−1.46	−1.95	−0.77	−0.19	−0.37	−0.09

shows the impact of riverbed incision and flow volume on water level changes at JJS. Among the factors contributing to the water level reduction at JJS, riverbed incision accounts for 35% (a reduction of 0.52 m) and flow volume accounts for 65% (a reduction of 0.94 m) in September from 2003 to 2021. In October, riverbed incision contributes 34% (a reduction of 0.66 m) and flow volume contributes 66% (a reduction of 1.28 m). From December to March of the following year, the average riverbed incision is 1.0 m, and the increase in flow volume led to an average elevation of 0.6 m and the largest elevation of 0.82 m in January.

According to the flow runoff changes listed in Table 5, the influences of different flow volume changes on water level at JJS are shown in Figure 7. In September and October, the water storage of TGR reduced the water level at JJS by 0.66 m on average, the decrease in the inflow volume from the upstream TGR reduced the water level at JJS by 0.39 m on average, and the flow volume from TGR–JJS interval basin reduced the water level at JJS by 0.06 m. From December to March of the following year, the TGR replenishment increased the water level at JJS by 0.26 m on average, the increase in the TGR inflow volume raised the water level by 0.31 m on average, and the flow runoff from TGR–JJS interval basin basically had no influence on the water level at JJS.

(2)

Factors of TGR operation

The response of the water level variations at JJS to the TGR outflow discharges during the impoundment and drawdown periods is quantitatively analyzed. The LSTM neural network model is used to simulate the water levels at JJS in September and October, as well as from December to March of the following year, respectively. The monthly data series of TGR outflow and TGR–JJS interval flow discharges are used as LSTM model inputs, while the water level at JJS is used as the output variable.

The LSTM model is constructed using TensorFlow, which incorporates a single hidden layer, with each layer consisting of 128 neurons, and employs the tanh function as the activation function. To avoid fitting distortion resulting from significant disparities in the data, the input and output data are first normalized using the min–max scaling method available in scikit-learn. Subsequently, the entire dataset is divided into a training set and a verification set. Specifically, 30% of the data are randomly selected as the verification set, and the remaining data constitute the training set.

Table 6. Simulation results of water level at JJS in TGR impoundment and drawdown periods.

Evaluation Indicators	Sept. to Oct.		Dec. to Mar.
	Training Period	Test Period	Training Period	Test Period
NSE	0.99	0.98	0.98	0.96
MAE (m)	0.16	0.23	0.19	0.24
RMSE (m)	0.24	0.31	0.27	0.32

presents the simulation results of water level at JJS in the TGR impoundment period from September to October and drawdown period from December to May of the following year. The Nash–Sutcliffe efficiency coefficient (NSE) values are 0.98 and 0.98, and the mean absolute errors (MAE) are 0.16 m and 0.23 m in training and test periods, respectively, from September to October. Evidently, both the TGR outflow discharge and interval basin flow discharge can effectively regulate the water level at JJS, thus providing a means for quantitatively analyzing the influence of TGR outflow discharge on water level at JJS.

Figure 8 illustrates the observed and simulated water levels at JJS by the LSTM model in TGR impoundment and drawdown periods. The annual average TGR outflow discharges in September and October are variated between 7100 and 24,100 m³/s and, from December to March of the following year, are ranged 3400~9500 m³/s from 2003 to 2021. In September and October, the TGR outflow discharge is fixed between 10,000 and 25,000 m³/s, with an increasement of 5000 m³/s, and interval basin flow is set between 5000 and 23,000 m³/s.

The LSTM model is used to estimate the increase in water level at JJS under the conditions of additional 5000 m³/s of the TGR outflow discharge and unchanged interval basin flow. As shown in Figure 9, when the interval basin flow is between 5000 and 12,000 m³/s, the increase in the TGR outflow discharge can raise the water level at JJS by 1.08~1.18 m; when the interval basin flow discharge is larger than 15,000 m³/s, the increase in the TGR outflow discharge can raise the water level at JJS by 0.90 m on average. On the whole, when the TGR outflow discharge increases 5000 m³/s, the water level at JJS will rise about 1.0 m in the impoundment period.

From December to March of the following year, the TGR outflow discharge is often between 5000 and 9000 m³/s, with an increasement of 1000 m³/s, and interval basin inflow is set as 3000~15,000 m³/s. The LSTM model is used to calculate the water level variation at JJS under the conditions of every additional 1000 m³/s of the TGR outflow discharge and unchanged interval basin flow. As shown in Figure 10, when the interval flow discharge is less than 7000 m³/s, every additional 1000 m³/s of the TGR outflow discharge can increase water level at JJS by 0.14~0.15 m; when the interval basin flow discharge is greater than 7000 m³/s, the TGR outflow discharge can raise the water level at JJS by 0.16 m. On the whole, the water level at JJS will rise 0.16 m when the TGR outflow discharge increases 1000 m³/s in the drawdown period.

Partial dependence plot (PDP) is adopted to analyze the impact of single feature variable variation on the simulation results. Partial dependence plot can show whether the relationship between the target and features is linear, monotonic, or complicated. Figure 11 shows the partial dependence plots of the TGR outflow discharge, TGR–JJS interval basin inflow, and the water level at JJS in September and October after the TGR operation in 2003. As shown in Figure 11, when the TGR outflow discharge is larger than 20,000 m³/s, the water level at JJS is not obviously raised and, when the interval basin flow discharge is greater than 10,000 m³/s, the water level at JJS is not obviously raised either in September and October. As shown in Figure 12, the water level variation at JJS as the TGR outflow discharge increases is not much more obvious than that caused by interval basin flow discharge from December to March of the following year. When the interval basin flow discharge is greater than the TGR outflow discharge, the increasement of the TGR outflow discharge will not obviously raise the water level at JJS.

5. Conclusions

In this study, the Mann-Kendall method was used to test the change trend of water level and flow discharge at JJS; the TGR outflow discharge and TGR–JJS interval basin flow discharge from 1981 to 2021 were used to estimate the contributions of riverbed incisions and flow runoff variations to the water level change at JJS. The LSTM model was used to quantitatively study the impact of TGR outflow and interval basin flow discharges on water level at JJS. The main conclusions are the following:

(1)

After the operation of TGR in 2003, the annual average of monthly water levels at JJS are decreased by −1.74 m, −2.11 m, and −0.78 m in September, October, and December, respectively, and are slightly decreased by −0.02 m and −0.18 m in January and February, with a little rise of 0.12 m in March.

(2)

The annual average monthly water levels at JJS are decreased 0.59 m and 1.11 m, which are caused by riverbed incision and TGR outflow discharge during the impoundment period, respectively. From December to March of the following year, the water levels at JJS are reduced 0.99 m due to riverbed incision and rise 0.63 m with the increasement of TGR outflow discharge in the drawdown period.

(3)

Every additional 5000 m³/s of the TGR outflow discharge can increase about 1.0 m the water level at JJS in the impoundment period from September to October, while, in the drawdown period from December to March of the following year, every additional 1000 m³/s of the TGR outflow discharge can increase 0.16 m the water level at JJS.