Capturing the attenuating influence of the Three Gorges Dam on downstream sediment loads using artificial neural networks

In recent decades, the natural flow regimes and sediment transport processes in rivers have been dramatically reshaped by anthropogenic interventions, with dam construction standing out as a leading cause of transformation. River dams have a profound effect on the hydrological balance and sediment conveyance capacity of downstream stretches, altering the river's natural state (Graf 2006; Milliman & Farnsworth 2011; Lyu et al. 2020), as highlighted by the Yangtze River's experience, where a series of dams have collectively accounted for 22% of its annual flow storage (Yang et al. 2010; 2011).

The Three Gorges Dam (TGD), completed in 2003, stands as one of the world's largest hydroelectric projects and has significantly influenced the Yangtze River's hydrology and sediment load (Li et al. 2018; Mei et al. 2018; Lyu et al. 2019). Located in the upper reaches of the Yangtze River, the TGD has a total reservoir capacity of 39.3 billion m³ and a power generation capacity of 22,500 MW. While its impact on downstream runoff has been considered minimal (Yu et al. 2013), sediment transport alterations have garnered significant attention (Chen et al. 2008; 2010; Dai et al. 2011). The dam's operation has led to a substantial reduction in downstream sediment discharge, with annual sediment load at the Yichang (YC) station (immediately downstream of the TGD) decreasing from an average of 530 million tons before the dam's construction to about 60 million tons in recent years (Dai & Liu 2013; CWRC 2017).

Despite extensive research documenting pre- and post-construction sediment discharge changes (Dai et al. 2011; Dai & Liu 2013; CWRC 2017; Lyu et al. 2020), a detailed, quantitative assessment of the TGD's operational impact on individual downstream sections remains a research gap. Most existing studies either focus on specific river reaches or regions or overlook the influence of factors such as dam operations and tributary inflows, failing to comprehensively capture the dynamics of regulated river systems.

Artificial neural networks (ANNs) have emerged as a powerful tool for modeling complex, nonlinear systems in hydrology and sediment transport. ANNs are data-driven, machine-learning techniques that have demonstrated remarkable prowess in tackling complex, nonlinear problems across various disciplines, including hydrology and hydraulics (Nagy & Watanabe 2002; Cigizoglu 2003a; 2003b; Kisi 2004, 2005; Mustafa & Isa 2014; Alizadeh et al. 2017; Bouzeria et al. 2017; Lamchuan et al. 2020). Several studies have applied ANNs to predict suspended sediment concentration and load in rivers (Jain 2001; Cigizoglu & Kisi 2006; Raghuwanshi et al. 2006). In the context of dam impacts, ANNs have been used to forecast sediment loads in regulated rivers (Mustafa & Isa 2014; Bouzeria et al. 2017) and to assess the effects of dam operations on downstream sediment transport (Afan et al. 2015; Hassan et al. 2015). However, the application of ANNs specifically to quantify the attenuating influence of large dams like the TGD on downstream sediment loads remains limited.

This study aims to fill this knowledge gap by quantifying the attenuating influence of the TGD on sediment loads along the middle and lower Yangtze River reaches downstream (from YC to Datong (DT)) using ANN models. Our objectives encompass: (1) analyzing actual water and sediment observations from key Yangtze stations; (2) outlining sediment transport theory, underscoring the need for ANNs due to the problem's inherent complexity; (3) detailing the ANN methodology, including design, training protocols, and validation measures; (4) implementing these models to explore the interplay between water-sediment discharge in the middle-lower Yangtze, and precisely quantify TGD-induced variations' downstream reach impacts; and (5) discussing the implications, limitations, and future avenues of study, alongside contrasting sediment transport predictions from different input scenarios for the YC station, highlighting the value of ANNs over traditional models in capturing the nuances of post-dam sediment dynamics.

This approach contributes to a deeper understanding of dam management implications and sediment transport dynamics, offering insights applicable beyond the Yangtze River basin in East Asia. By leveraging the power of ANNs to capture complex, nonlinear relationships, this study provides a novel perspective on quantifying the spatial variability of large dam impacts on river systems, potentially informing future dam operation strategies and river management practices in other major river systems worldwide, particularly in regions experiencing rapid hydropower development.

Daily average runoff and sediment load data were obtained from the Changjiang Water Resources Commission (CWRC) for the seven hydrometric stations along the Yangtze River mainstream. Additionally, water and sediment inputs from Dongting Lake (represented by the Qilishan station), the Han River (represented by the Xiantao station), and Poyang Lake (represented by the Hukou station) were incorporated for LS, HK, and DT stations, respectively. The CWRC rigorously validates and controls the integrity and reliability of these data, which have been extensively utilized in scientific research (Chen et al. 2008; Li et al. 2016; CWRC 2017).

Figure 3(b) depicts the MK trends of annual sediment concentration at the eight major stations along the Yangtze River. The MK test indicates a significant decreasing trend at the 95% confidence level after 2003 for all stations. A critical point, signifying an abrupt drop, occurred in 2003 for all stations except JJ. For the JJ station, the lack of data before 1996 and systematic error inherent to the MK method likely contribute to the multiple intersections of the test statistic curves during the 1998–2002 period.

In summary, the MK analysis reveals the following temporal trends at the eight major stations along the Yangtze River: (1) an insignificant decreasing trend in annual water discharge at all stations; (2) a significant decreasing trend in annual sediment concentration after 2003 at all stations, with an abrupt drop observed in 2003 at seven out of the eight stations. Qualitatively, the impact of the TGD on sediment concentration gradually diminishes from YC to DT.

By utilizing Equations (2),  (6), and  (7), the quantitative analysis of the effects of runoff and sediment discharge at the YC gauge station, and ⁠, on the downstream stations can be performed if the derivative functions in these equations are determined. Thus, the derivative function plays a crucial role.

Traditional hydrodynamic models often rely on explicit formulations of these governing equations to predict sediment transport downstream. However, the complex and nonlinear nature of the problem poses challenges for such models, leading to potential inaccuracies or oversimplifications. Therefore, an alternative approach capable of capturing the inherent nonlinearities is desirable.

To clarify, these governing equations are presented to illustrate the complexity of sediment transport processes, highlighting the challenges in traditional modeling approaches. They are not used for baseline modeling in this study but rather to motivate the use of data-driven ANN models as follows.

To comprehensively analyze the impact of the TGD on downstream hydrological stations while overcoming the limitations of traditional models, ANN models are employed in this study.

Unlike traditional models that rely on predefined equations and assumptions, ANNs construct implicit models directly from data, bypassing potential biases stemming from inadequate assumptions or misconceptions. By training on extensive measured data, ANNs can effectively capture the inherent nonlinearities and complex interactions among variables, offering a powerful alternative to deterministic approaches, particularly in cases where the underlying processes are not well understood or difficult to represent analytically. ANNs can be classified based on the number of layers (single-layer, two-layer, multi-layer), and the direction of information flow (feed-forward and feedback). Multi-layer feed-forward ANNs, widely employed in function-fitting tasks (Aroui et al. 2007; Benardos & Vosniakos 2007; Zhang et al. 2008), are utilized in this study.

Seven ANN models were established for hydrological stations from ZC to DT, focusing on examining water and sediment transport between adjacent sections of the Yangtze River. Input data comprised daily average water and sediment input upstream of each station, with outputs being daily average water discharge and sediment concentration at the respective stations. Training and testing utilized complementary sample sets, with 70% for training and 30% for testing.

The ANN models used are multi-layer perceptrons with three hidden layers, each containing 64 neurons. We used the ReLU activation function and Adam optimizer with a learning rate of 0.001. The models were trained for 200 epochs with a batch size of 32.

The structure of the ZC ANN model and the loss function during training are illustrated in Figures 4(b) and 4(c), respectively. The loss function reflects differences between expected and true values, diminishing during training. Similar methodologies were applied to the other six models.

After the training and testing phases, the derivative functions between adjacent sections from YC to DT were determined. Consequently, the impact of variations in water and sediment input (⁠ and ⁠, respectively) from the upstream sections on variations in sediment discharge (⁠⁠) can be examined using the established ANN models and the measured water and sediment discharge data from 2003 to 2018 as a reference. The impacts are expressed as dimensionless percentages. Similarly, the impacts of variations in water and sediment input (⁠ and ⁠, respectively) from the upstream sections on variations in water discharge in section i were also analyzed. Furthermore, the relationship between the partial derivative of sediment discharge in the sections of interest and the water and sediment input from upstream adjacent sections can be estimated, expressed as and ⁠, respectively.

Similarly, and can also be obtained, as shown in Figures 6(c) and 6(d). The results demonstrate that variations in water discharge in the sections of interest are more closely related to variations in water input from upstream adjacent sections. Variations in sediment input from upstream sections have almost no impact on water discharge in downstream sections.

The results clearly demonstrate the significant impact of the TGD operations on sediment discharge along the middle and lower reaches of the Yangtze River. The abrupt reduction in sediment concentration observed at the YC station after 2003 propagated downstream, with the magnitude of the impact diminishing gradually with increasing distance from the dam. This finding corroborates previous studies (Li et al. 2016; Lyu et al. 2020) that reported a substantial decline in sediment load following the TGD's impoundment. Notably, the ANN models revealed that variations in sediment discharge at a given section were predominantly influenced by changes in sediment input from the immediately upstream section, as evidenced by the high values of (>0.35). Conversely, the effect of variations in water discharge from upstream was relatively minor, except for sections downstream of SS. This observation suggests that the TGD's impact on sediment transport was primarily due to its trapping efficiency, rather than alterations in the hydrological regime.

The diminishing influence of the TGD on sediment discharge farther downstream can be attributed to the contributions from major tributaries, such as the Han River at HK and Poyang Lake at DT. These tributaries introduce additional sediment loads, effectively diluting the impact of the TGD's sediment trapping. Moreover, the dynamic nature of sediment transport processes, including erosion, deposition, and resuspension, likely plays a role in modulating the downstream propagation of the TGD's impact. The findings of this study not only contribute to a deeper understanding of the TGD's implications on sediment dynamics but also provide insights applicable to other large-scale hydraulic engineering projects and their operation management. The quantitative assessment of the attenuating influence of the TGD on downstream sediment loads offers a new perspective for developing scientifically sound water and sediment regulation strategies.

It is well-known that several measured water and sediment discharge series for the YC station were employed at different stages to predict sediment transport downstream when using classical hydrodynamic mathematical models, which has become an important means of solving practical problems for the TGD. Specifically, the classical 60 series (1961–1970) was used at the design and demonstration stage of the TGD to predict the effect after 2003. Subsequently, with the construction of reservoirs upstream, a new 90 series, which considered the influence of upstream cascade reservoirs (such as the Xiangjiaba, Xiluodu, Baihetan, and Wudongde reservoirs) based on the 1991–2000 period, was proposed and widely used by researchers to predict water and sediment transport after 2012. However, deviations between the water and sediment discharge input for these series and the actual measured data inevitably exist, as listed in Table 1.

While the ANN models effectively quantified the impact of the TGD on downstream sediment discharge, it is essential to recognize potential limitations. The models were trained and validated using post-TGD data, which may not fully capture the pre-impoundment conditions and sediment dynamics. Additionally, the models do not explicitly account for the complex physical processes governing sediment transport, such as bed evolution, armoring, and tributary contributions. Future research could explore the integration of ANN models with physics-based models to enhance their predictive capabilities and provide insights into the underlying mechanisms.

This study employed ANN models to quantitatively estimate the impacts of the TGD on runoff and sediment discharge variations along the middle and lower reaches of the Yangtze River from YC to DT stations. The results demonstrate the significant influence of the TGD on downstream sediment transport, with an abrupt reduction in sediment concentration observed at YC after 2003 propagating downstream, albeit with diminishing magnitude farther away from the dam. The ANN models revealed that variations in sediment discharge at a given section were predominantly influenced by changes in sediment input from the immediately upstream section, suggesting the TGD's impact was primarily due to its sediment trapping efficiency rather than alterations in the hydrological regime. The diminishing impact farther downstream can be attributed to contributions from major tributaries and the dynamic nature of sediment transport processes. Notably, the analysis highlighted the robustness of the ANN approach in capturing the underlying relationships between input and output variables, even in the presence of uncertainties or discrepancies in the boundary condition data. While the post-TGD data used for model training and validation may not fully capture pre-impoundment conditions, integrating ANN models with physics-based models could enhance predictive capabilities and provide insights into the underlying mechanisms governing sediment dynamics in regulated river systems.

This work was funded by the National Natural Science Foundation of China (Grant Nos 52379083 and 52409103), the Scientific Program of State Key Laboratory of Simulation and Regulation of Water Cycle in River Basin, China (Grant No. SKL2022TS12), the Basic Scientific Research Project of China Institute of Water Resources and Hydropower Research (Grant Nos SE0199A102021 and SE0145B042021), the National Key Hydraulic Engineering Construction Funds (Grant No. 12630100100020J005), and the Open Research Fund of Key Laboratory of Sediment Science and Northern River Training, the Ministry of Water Resources, China Institute of Water Resources and Hydropower Research (Grant No. IWHR-SEDI-2023-02).

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.