Exploration of Runoff Simulation Based on Seasonal Precipitation Characteristics and Its Impact on Hydropower Generation

Abstract

Accurate and robust runoff simulation is crucial for effective reservoir regulation. Although it is clear that enhancing runoff simulation or optimizing reservoir operation strategies can improve the management of hydropower resources, the specific impact of enhanced simulated runoff on reservoir operation under optimized regulation has not been thoroughly examined. To investigate how high-precision runoff simulation influences reservoir performance, this study proposed a unidirectional coupling framework of the distributed hydrological model CREST and the LSTM model, incorporating the seasonal characteristics of the satellite-based precipitation product CHIRPS. The influence of simulated runoff on hydropower generation was examined from two perspectives: metrics’ accuracy and process-based analysis. The results showed that, following the unidirectional coupling, the Coupling scheme achieved improvements in NSE and R² by 6% and 4%, respectively, while RMSE decreased by 24%. Additionally, it accurately captured the seasonal variations and amplitude of runoff at the annual scale, and was able to reliably detect the periodic signals within runoff across various scales. After reservoir optimization operation, the simulated runoff derived from the Coupling scheme produced hydropower and surplus water values close to those obtained from observed runoff, with errors of 1.09% and −21.64%, respectively. Moreover, the Coupling scheme corrected the prominent peaks in hydropower generation seen in the CREST model across multiple periods, demonstrating a stronger capability for temporal runoff simulation closely aligned with observed runoff in terms of temporal structure.

1. Introduction

Hydropower resources possess the greatest and most stable potential among renewable energy sources, and whether they can be efficiently utilized largely depends on the spatio-temporal distribution of water resources and the effective allocation of water resources through human activities [1,2]. Unlike wind and photovoltaic power generation, which exhibit randomness, volatility, and intermittency due to weather variability, hydropower resources primarily rely on reservoirs for hydroelectric generation [3]. With the successive completion and operation of large-scale reservoir clusters connected in parallel-series configurations across China’s river basins, a unified basin-wide regulation framework for hydropower and water resources has been preliminarily established. Consequently, the water cycle has evolved from a purely natural system into a dual “natural–artificial” water cycle [4,5].

Currently, the primary approaches to enhancing hydropower resource utilization involve optimized regulation of hydraulic projects and coordinated basin management. Both methodologies aim to achieve the goal by coordinating the operation schemes among hydraulic facilities within the river basin, that is, to improve the efficiency of externally constructed models [6,7,8]. For instance, Yang et al. selected the valuable variables in multi-objective cascade reservoir regulation as inputs, applied Gaussian radial basis functions to study multi-objective cascade reservoir operation, and finally utilized the Pareto-archived dynamically dimensioned search algorithm to optimize reservoir operation parameters, resulting in a Pareto front balancing water supply and hydropower generation objectives [6]. Giuliani et al. proposed a novel method combining direct policy search with multi-objective evolutionary algorithms to address high-dimensional state and water resource regulation problems involving multiple conflicting objectives in reservoir operations [9]. To overcome the poor performance of traditional Progressive Optimality Algorithms (POA) when faced with the “curse of dimensionality” in the operations of large-scale cascade reservoirs, Niu et al. proposed an artificial intelligence-based Response Surface Progressive Optimality Algorithm (RSPOA) for optimizing multi-reservoir systems, and the results indicated that the proposed algorithm achieved approximately a 79.2% reduction in computational time [7]. Although the aforementioned studies indicated that model optimization can effectively enhance hydropower resource utilization, it also increased computational costs and the risk of overfitting, potentially compromising system robustness and interpretability, and thus weakening adaptability to uncertainties and unexpected conditions in practical applications. Additionally, these studies also have overlooked the fact that runoff—an essential input for reservoir optimization operation model—possesses inherent stochastic and periodic uncertainties that directly influence model computation and solution accuracy. Therefore, accurately capturing the precipitation-runoff process is equally critical for reservoir operation. For instance, Sibtain et al. conducted a two-stage signal decomposition of runoff and used it as the input of Long Short-Term Memory (LSTM) to improve the simulation and prediction of runoff [10]. Ma et al. proposed a novel hybrid model based on an improved deep residual shrinkage network combined with an optimized Gated Recurrent Unit–Long Short-Term Memory (GRU-LSTM) model, applied to water level forecast across different temporal scales [11]. Li et al. decomposed reservoir inflow into precipitation-runoff and flow-routing processes, utilizing a unidirectional coupling of the hydrological model HEC-HMS with a LSTM network to enhance inflow predictions [12]. Based on runoff correction using improved VDM and LSTM methods, Zhang et al. proposed employing the CVaR method to dynamically quantify and optimize risks related to power shortages and excessive surplus water in cascade hydropower operation strategies, and the results indicated that the proposed optimization framework improved reservoir decision-making accuracy by an average of 0.61% during flood seasons and reduced surplus water by 0.73% [13]. The aforementioned studies mostly emphasized improving the accuracy of variables related to hydrological processes but did not further validate the impacts of enhanced accuracy on reservoir operations. Additionally, although some studies have examined the influence of input data accuracy on hydropower generation, their primary focus remained on accuracy enhancement, neglecting the analysis of processes. Moreover, it can be observed that improvements in runoff accuracy derived from complex simulation models or enhancements in hydropower generation through sophisticated operation models remain limited. Therefore, incorporating process accuracy analysis on the basis of improving the accuracy of metrics has significant scientific value and practical implications for conducting high-precision runoff simulations and investigating their impacts on hydropower resources after reservoir operations.

This study focused on the Yalong River Basin (YRB), one of China’s thirteen major hydropower bases. In response to existing challenges in hydropower resource management, the seasonal characteristics of precipitation were introduced to construct a unidirectional coupled runoff simulation framework based on the Coupled Routing and Excess STorage model (CREST) and LSTM. The Progressive Optimality Algorithm (POA) was employed to solve the multi-stage decision-making optimization problem in reservoir operation, with the aim of enhancing hydropower resource management.

2. Overview and Data of the Study Area

2.1. Overview of the Study Area

The Yalong River originates from the southern slopes of the Bayan Har Mountains in Qinghai Province, located in the southeastern edge of the Qinghai–Tibet Plateau. It is a secondary tributary of the Yangtze River (Figure 1). The main stream of the Yalong River is approximately 1571 km long, with a basin area of 136 × 10³ km. The multi-annual average runoff at the basin outlet is 1890 m³/s, and the multi-annual average precipitation ranges from 600 to 800 mm. With a natural drop of 3830 m, the river possesses abundant hydropower resources, ranking third among China’s thirteen nationally planned major hydropower bases. At present, a total of 22 cascade hydropower stations have been planned along the main stream of the Yalong River. Among them, Lianghekou, Yangfanggou, Jinping I, Jinping II, Guandi, Ertan, and Tongzilin have already been constructed, while others such as Kala and Yagen I are under construction [14,15].

2.2. Data Information

(1)

Reservoir data

The reservoirs that have been built and put into operation in YRB from upstream to downstream are Lianghekou, Yangfanggou, Jinping I, Jinping II, Guandi, Ertan, and Tongzilin. Among these, Guandi, Jinping I, Jinping II, and Tongzilin were successively commissioned between 2013 and 2016, while Yangfanggou and Lianghekou began operation in 2022 and 2024, respectively. Although the Tongzilin Reservoir is located at the lowermost level of the basin, its relatively recent commissioning means that the available observed natural runoff series and reservoir operation data are limited. In contrast, the Ertan Reservoir, which is situated immediately upstream of Tongzilin, was commissioned in 2000, and a hydrological station built in the 1950s is located not far upstream. As a result, Ertan not only has a long natural runoff record but also has an extended period of reservoir operational data. Based on these considerations, the Ertan Reservoir was selected as the research subject in this study.

The Ertan Reservoir is located in the lower reaches of the Yalong River and is the first hydropower station developed as part of the Yalong River hydropower base cascade. The maximum dam height of the station is 240 m, with a normal reservoir water level of 1200 m above sea level. It has a total storage capacity of 5.8 × 10⁹ m³ and a regulation storage capacity of 3.37 × 10⁹ m³. The total installed capacity is 3.3 × 10⁹ W, with a guaranteed output of 1.0 × 10⁹ W and an average annual power generation of 170 × 10⁸ kW∙h. The project is mainly for power generation and also has some flood control tasks. The proportion of other comprehensive utilization benefits is little.

(2)

Meteorological and hydrological data

In this study, precipitation data were obtained from the CHIRPS (Climate Hazards Center InfraRed Precipitation with Station data), jointly developed by the United States Geological Survey (USGS) and the Climate Hazards Center (CHC) at the University of California, Santa Barbara. This satellite dataset integrates 0.05° resolution satellite imagery with ground-based meteorological station data, effectively eliminating systematic biases commonly found in climatological data to generate high-resolution gridded precipitation time series for trend analysis and seasonal drought monitoring [16]. CHIRPS provides multi-temporal datasets covering the global extent from 50° S to 50° N and 180° W to 180° E, available in spatial resolutions of 0.05° and 0.25°, and in various formats such as NetCDF, TIFF, BIL, and PNG. In this study, CHIRPS-2.0 daily data were used, covering the period from 1 January 1981 to 31 December 2013, with a spatial resolution of 0.25°. Potential evapotranspiration (PET) data were sourced from the Global Daily Reference Evapotranspiration dataset, which is available through the FEWS NET (Famine Early Warning Systems Network) Global Data Portal (https://earlywarning.usgs.gov/fews/, accessed on 23 April 2018).

The Ertan Reservoir, located 15 km upstream of the Tongzilin Reservoir, was commissioned in 2000. Not far upstream of the reservoir is the Ertan Hydrological Station, which was established in the 1950s and provides a long-term hydrological time series. This dataset includes a representative combination of wet, normal, and dry years, making it suitable for hydrological analysis. Therefore, the data from the Ertan Hydrological Station were used in this study to represent the inflow to the Ertan Reservoir, with daily runoff records obtained from the Hydrological Yearbook of the Yangtze River Basin. In addition, the data required for hydrological model construction—including the digital elevation model (DEM), flow direction, flow accumulation, and river network—were all sourced from USGS HydroSHEDS (https://www.hydrosheds.org/, accessed on 26 May 2024).

3. Method

The primary objective of this study is to investigate the accuracy of simulated runoff incorporating seasonal precipitation characteristics and its impact on hydropower generation. To achieve this goal, a comprehensive framework was proposed: the seasonal precipitation characteristic was incorporated into runoff simulation based on a unidirectional coupling of the CREST model with an LSTM model. Subsequently, an optimization operation model based on the POA was developed according to the characteristics of reservoirs in the YRB to investigate the impact of runoff simulation on hydropower generation.

3.1. Hydrological Model, Machine Learning and Reservoir Operation

(1)

Hydrological model CREST

The Coupled Routing and Excess STorage (CREST) hydrological model is a grid-based distributed model jointly developed by the University of Oklahoma and NASA project teams. It is characterized by its ability to integrate satellite remote sensing data and observational data for hydrological analysis at local, regional, and global scales [17]. Version 2.1 of the CREST model was employed in this study, with its detailed framework illustrated in Figure 2. It primarily represents the key hydrological processes of the watershed water cycle, including post-precipitation infiltration, surface runoff routing, changes in water storage, and evapotranspiration. Version 2.1 adopts a dual-source mechanism, in which the runoff is divided into surface runoff and subsurface runoff, while groundwater runoff is not considered. The model applies a three-mechanism, grid-based routing calculation. The movement laws of surface water, subsurface runoff, and channel flow differ: surface water mainly follows the kinematic wave equation, subsurface runoff mainly follows Darcy’s law, and channel flow mainly follows the Saint-Venant equations. Each mechanism has its own routing formulation, and the model generalizes these three mechanisms using three linear reservoirs. To account for the heterogeneity of variables within the coarse model grid, sub-grid variability of hydrological fluxes is introduced for improved representation. In addition, version 2.1 features enhanced operational stability, an updated runoff routing module, support for multi-temporal scale configurations, and the incorporation of automatic calibration based on SCE-UA parameters. At present, the CREST model has been applied to studies at the global scale and has demonstrated commendable simulation accuracy [18].

(2)

Machine Learning—LSTM

Long Short-Term Memory (LSTM), a special type of recurrent neural network (RNN), was first proposed by Hochreiter and Schmidhuber in 1997 to address the problems of gradient vanishing and gradient explosion that often occur when conventional RNNs process long time series. By introducing a “gate” mechanism—including the forget gate, input gate, and output gate—LSTM effectively addresses the issues of gradient vanishing and exploding that traditional recurrent neural networks (RNNs) often encounter when handling long sequences. This enables LSTM to capture long-term dependencies more efficiently [19]. After multiple iterations of development, LSTM has evolved into a relatively mature framework and has been widely applied across various fields [20,21]. However, LSTM still has problems such as high computing costs and poor performance in short-term data at present.

Let the current time step be $t$, the input be $x_t$, the hidden state from the previous time step be $h_{t-1}$, and the cell state be $c_{t-1}$. The main equations involved in the LSTM process are as follows:

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

(1)

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

(2)

$${\tilde{c}}_t=\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}(W_c·\left[h_{t-1},x_t\right]+b_c)$$

(3)

$$c_t=f_t⨀c_{t-1}+i_t⨀{\tilde{c}}_t$$

(4)

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

(5)

$$h_t=o_t⨀\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left(c_t\right)$$

(6)

In these equations, $f_t$ denotes the output of the forget gate (a value between 0 and 1), $i_t$ denotes the output of the input gate (a value between 0 and 1) that determines the importance of the current input information, ${\tilde{c}}_t$ is the candidate cell state activated by the $tanh$ function with values ranging from −1 to 1, and $c_t$ represents the cell state at the current time step. The symbol $⨀$ indicates element-wise multiplication (Hadamard product). $o_t$ represents the proportion of information extracted from the cell state to form the current output, while $h_t$ denotes the output at the current time step, which also serves as the input for the next time step. $\sigma$ is the sigmoid activation function, and $W_{*}$ and $b_{*}$ denote the weight matrices and bias terms of each gate.

(3)

Reservoir optimal operation

By selecting the optimal regulation strategy that meets predefined objectives and constraints, and by planning the storage and release of reservoir inflow based on actual inflow conditions and hydrological forecasts, economic and social benefits can be maximized. The POA was employed in this study to solve the multi-stage decision-making optimization problem in reservoir operation [22,23]. The method decomposed the multi-stage optimization problem into a series of two-stage subproblems. By iteratively adjusting the decision variables at each stage, it ensured that the decisions of adjacent stages satisfied optimality conditions, ultimately achieving overall optimality. In this study, the objective function was to maximize the hydropower generation of the Ertan Reservoir:

$$Max E=\sum _{t=1}^T{P}_t∆t$$

(7)

$$P_t=K{Q}_t{H}_t$$

(8)

In the equation, $E$ represents the total hydropower generation over the study period; $∆t$ is the time step during the study period; $T$ is the total number of time steps; $P_t$ is the power output of the hydropower station at time $t$; $Q_t$ and $H_t$ represent the discharge and water head of the hydropower station at time $t$, respectively; and $K$ is the comprehensive power generation coefficient of the hydropower station.

The main constraints of reservoir operation include water balance, reservoir water level and storage capacity constraints, outflow discharge limits, power output restrictions, and initial boundary conditions, as detailed below:

$$V_t=V_{t-1}+(Q_t^{in}-Q_t^{out})∆t$$

(9)

$$V_t^{min}\le V_t\le V_t^{max}$$

(10)

$$Q_t^{min}\le Q_t\le Q_t^{max}$$

(11)

$$Q_t^{in}=Q_t^{interval}+Q_t^{out}$$

(12)

$$N_t^{min}\le N_t\le N_t^{max}$$

(13)

$$\left\{\begin{array}{c}{Z}_1=Z_{initial}\\ Z_{t+1}=Z_{final}\end{array}\right.$$

(14)

In the equation, $V_t$ and $V_{t-1}$ represent the initial storage volumes of the hydropower station at time periods $t$ and $t-1$, respectively; $Q_t^{in}$ and $Q_t^{out}$ are the average inflow and outflow discharges; $V_t^{min}$ and $V_t^{max}$ are the minimum and maximum allowable storage volumes during the time period, where the former usually corresponds to dead storage and the latter to the storage volume at flood control level or normal water level; $Q_t^{min}$ and $Q_t^{max}$ are the minimum and maximum release discharges of the reservoir; $N_t^{min}$ and $N_t^{max}$ are the minimum and maximum power outputs during the time period; and $Z_{initial}$ and $Z_{final}$ denote the reservoir water levels at the beginning and end of the scheduling period, respectively.

3.2. Evaluation Methods and Metrics

(1)

Moving Average over Shifting Horizon, MASH

The Moving Average over Shifting Horizon (MASH) method is a time series analysis tool designed to identify seasonal and interannual variations. This method smooths the original series by first averaging daily values across consecutive years, and then calculating average values over consecutive days according to a sliding window width [24]. Specifically, smoothing consecutive daily data within the same year reveals potential trends in seasonal characteristics, while applying a sliding window operation to data from the same day across different years allows the identification of evolutionary trends at the interannual scale [25,26].

The results of the MASH trend analysis are influenced by two important parameters, Y and w, where Y and 2w + 1 represent the horizontal sliding length and the vertical sliding length, respectively. When Y is large, the number of horizontal sliding levels in MASH decreases, causing the hydrological time series to become more consistent in interannual variations. When w is large, the intra-annual variability trends in each time series become excessively weakened, resulting in the loss of seasonal variations.

(2)

Ensemble Empirical Mode Decomposition, EEMD

Ensemble Empirical Mode Decomposition (EEMD) is a signal-processing method designed to improve upon traditional Empirical Mode Decomposition (EMD), addressing the mode-mixing problem encountered when dealing with non-stationary and nonlinear signals. EEMD repeatedly adds white noise to the original signal and performs multiple EMD decompositions, then averages the resulting modes, thereby enhancing the robustness and consistency of the decomposition [27]. Typically, EEMD can decompose a signal into multiple Intrinsic Mode Functions (IMFs) and a residual component. Each IMF exhibits linear or nonlinear characteristics and often represents either the direct current component of the signal or its underlying trend. EEMD is suitable for analyzing nonlinear and non-stationary signals, such as hydrological, meteorological, and vibration data, effectively reducing noise interference while preserving the true characteristics of the signal. It is widely applied in areas including time series prediction and fault diagnosis [28,29,30].

3.3. Scheme Setting

Based on a comprehensive consideration of basin size, DEM, precipitation, and other relevant data, the study area was divided into 250 grids, each with a spatial resolution of 0.25° × 0.25°. The hydrological model was calibrated over the period 1981~1995 and validated over 1996~2013, with the final optimized parameters presented in

Table 1. Calibration Parameters of the CREST Model.

Parameters	Upper Limit	Lower Limit	Implication	Optimal Value
RainFact	1.2	0.5	precipitation adjustment coefficient	0.68
Ksat	3000	0	Soil saturated hydraulic conductivity	1878.9
WM	200	80	Average water storage capacity	144
B	1.5	0.05	Variable infiltration capacity curve exponent	0.66
IM	0.2	0	Percentage of impermeable area	0.06
KE	1.5	0.1	potential evapotranspiration adjustment coefficient	0.52
coeM	150	1	Hillslope runoff velocity coefficient	106.68
expM	2	0.1	Hillslope runoff velocity exponent	0.39
coeR	3	1	Velocity conversion coefficient of hillslope runoff to channel flow	1.4
coeS	1	0.001	Velocity conversion coefficient of hillslope runoff to interflow	0.54
KS	1	0	Linear reservoir surface runoff coefficient	0.11
KI	1	0	Linear reservoir subsurface outflow coefficient	0.11

. To ensure stable and efficient training of the LSTM model, weights and biases were initialized using the Xavier uniform initialization method, the Adam optimizer with an adaptive learning rate was employed, and the Nash–Sutcliffe Efficiency (NSE) was used as the loss function [31]. Hyperparameters were tuned using random search within the following ranges: number of hidden units (16~256), dropout rate (0.1~0.5), batch size (32~256), number of epochs (50~200), and number of LSTM layers (1~2). The initial learning rate was set to 0.001, adjusted to 0.01 during iterations 1 to 30, reduced to 0.005 between iterations 31 and 40, and reverted to 0.001 after iteration 40. To ensure the stability and reproducibility of the results, each combination of hyperparameters was trained five times under different random seeds, and the average performance was taken as the final result. Ultimately, the model was developed using Python 3.9 and the Keras library built into TensorFlow. A single-layer structure with 29 nodes was constructed to train the model with a batch size of 256, an input sequence length of 82, the dropout rate of 0.2, and a total of 50 epochs. In addition, the three-fold cross-validation was employed to more accurately assess the generalization capability of the model [32]. Specifically, the period from 1996 to 2013 was used as the validation set, while the period from 1981 to 1995 was evenly divided into three subsets, with each round selecting two subsets for training and the remaining one for testing. Building upon the unidirectional coupling of the established CREST and the LSTM, seasonal precipitation characteristics derived from CHIRPS data were further incorporated. The extraction of these seasonal features followed several key steps: first, daily CHIRPS data were extracted at each grid point; second, the study area was divided into upstream, midstream, and downstream regions to calculate areal precipitation; finally, based on this spatial division, daily averages were computed for each seasonal period on an annual basis. The simulated runoff was then used as input flow to drive reservoir optimal operation, ultimately producing reservoir performance outcomes. Additionally, it should be noted that the observed hydropower generation mentioned later refers to the generation obtained by applying the reservoir optimization operation developed in this study to the observed runoff.

To compare the enhancement effect of the framework proposed in this study on runoff simulation and its impact on reservoir operation, the uncoupled CREST hydrological model and LSTM model were separately used as reference groups, as detailed in

Table 2. Scheme setting under the framework.

Scheme Name	The Models and Meteorological Driving Data Contained in the Scheme	Whether It Is Unidirectionally Coupled with LSTM
CREST	CREST/Precipitation and evaporation, etc.	No
LSTM	LSTM/Precipitation and evaporation, etc.	No
Coupling	VIC and LSTM/Precipitation, evapotranspiration and seasonal characteristics of precipitation (upper, middle and lower reaches	Yes

. In addition, four evaluation metrics—NSE, coefficient of determination (R²), bias, and root mean square error (RMSE)—were employed to assess the numerical accuracy of runoff simulation for each scheme.

4. Result and Analysis

4.1. Accuracy Analysis of Precipitation Runoff Simulation

First, based on precipitation data from observed station, a temporal and spatial accuracy analysis of CHIRPS was conducted, with the results shown in Figure 3. The observed precipitation data were taken from the China Ground Climate Data Daily Value Dataset (V3.0) provided by the China Meteorological Data Service Center (http://data.cma.cn/, accessed on 23 April 2018). This precipitation dataset has undergone rigorous quality control inspections. It is important to note that in the YRB and surrounding areas—especially in the high-altitude upstream regions—observation stations are sparse. Therefore, the spatial distribution derived from interpolating station data using the Kriging method cannot accurately represent the actual multi-annual daily average precipitation distribution in the YRB. For this reason, the interpolated spatial distribution of observed stations was not used as the reference. Instead, the multi-annual daily average precipitation at each observation station was directly calculated and presented as point values across the basin.

As shown in Figure 3, the spatial distribution of CHIRPS precipitation was close to that of observations, without obvious smoothing. The precipitation value of CHIRPS was the largest in the downstream of the basin, which is 3.63 mm, and then gradually decreased uniformly towards the upstream, with a minimum of 1.01 mm. The maximum and minimum precipitation values of CHIRPS were respectively higher than and lower than the corresponding observed precipitation values. In terms of temporal accuracy, the correlation coefficient between CHIRPS and observed daily precipitation was 0.53, the relative error was 5.63%, and the root mean square error was 7.37 mm/d. Overall, CHIRPS precipitation showed good correlation with observed data and can be used for hydrological simulations in the YRB.

Runoff simulation was conducted under each scheme, with the results presented in

Table 3. The metrics of each model at different periods.

Period	Scheme	NSE	R2	Bias (%)	RMSE (m3/s)
Calibration/Training	CREST	0.88	0.94	1.47	524.88
	LSTM	0.93	0.96	−1.45	391.59
	Coupling	0.94	0.97	−3.11	368.43
Validation	CREST	0.86	0.93	3.01	519.95
	LSTM	0.88	0.94	−2.61	479.69
	Coupling	0.9	0.96	−4.64	443.19
test	LSTM	0.88	0.94	−2.32	457.88
	Coupling	0.89	0.95	−1.57	437.45
The entire research period	CREST	0.87	0.93	2.04	523
	LSTM	0.91	0.95	1.87	427.42
	Coupling	0.92	0.97	−3.67	398.67

and Figure 4. As shown in Table 3, all schemes yielded satisfactory simulation performance, with NSE and R² equal to or greater than 0.87 and 0.93, respectively, and bias and RMSE not exceeding 4.64% and 523 m³/s. From the perspective of different study periods, the metrics of all schemes during the validation and test periods decreased or remained the same compared with the calibration period, while LSTM and Coupling schemes both maintained relatively strong performance. Overall, Coupling achieved higher simulation accuracy than LSTM, which in turn outperformed CREST. As illustrated in Figure 4, all schemes exhibited significant underestimation in the years 1981, 1993, 1997, 2006, 2011, and during the 1999~2000 period. The CREST slightly underestimated runoff during dry periods and showed delays in simulating recession flows. In contrast, LSTM and Coupling slightly overestimated runoff during the dry seasons between 2007 and 2011 but were better able to capture the timing of rising and falling limbs of the hydrograph. Runoff during the study period was ranked by frequency on an annual scale, and divided into dry, normal, and wet years using the 20% and 80% thresholds. From these, 2011, 2004, and 1998 were selected as representative years for further performance analysis of each scheme. It can be seen that Coupling demonstrated the best overall performance across all representative years, accurately capturing non-flood season flows within the year. LSTM ranked second, though it slightly underestimated low-flow periods in wet years. CREST, while strongest in capturing peak runoff, showed the poorest overall performance, mainly due to false peaks and underestimation of low-flow periods. However, in all representative years, the simulation of peak runoff by the different schemes was unsatisfactory, especially in wet years.

4.2. Accuracy Analysis of Runoff Simulation

The observed runoff and the runoff simulated under each scheme were analyzed using the MASH method to examine seasonal characteristics at both interannual and intra-annual scales. Variations at the interannual scale were derived from the data characteristics of the same calendar day across different years, while variations at the intra-annual scale were identified from the data characteristics of different days within the same year. After conducting a sensitivity analysis on the four sets of runoff data, the sliding window parameters were set to Y = 20 years and w = 10 days, and the results were shown in Figure 5. The MASH method produced 14 sliding levels, with the first level covering the period from 1981 to 2000, and the fourteenth level covering the period from 1994 to 2013. In terms of intra-annual distribution, the simulated runoff from all schemes were similar to observed runoff, and generally exhibited a bimodal pattern, with peak values occurring in July and September. However, the LSTM, CREST, and Coupling all underestimated the peak flows to varying degrees. Additionally, CREST slightly underestimated the low flows around February, while LSTM and Coupling slightly overestimated the low flows during the same period. In terms of interannual distribution, all schemes were able to reasonably capture the overall trends in runoff variation. Among them, the runoff simulated by LSTM and Coupling were highly similar and most closely aligned with the observed data. Specifically, both LSTM and Coupling successfully reproduced the interannual variability in runoff from summer to late winter, but significantly reduced the interannual amplitude of runoff in June. In addition, neither scheme was able to capture the interannual variation trend during the spring season. Although the CREST, which showed the greatest deviation from the observed runoff distribution, did capture the interannual trends in spring and late winter, it also considerably underestimated the amplitude of runoff in June, as well as in autumn and winter.

To further clarify the internal characteristics of the simulated runoff produced by each scheme, EEMD was applied to the monthly time series to extract signal components, enabling an exploration of localized runoff features at different temporal scales from the perspective of physical oscillation modes, and the results were shown in Figure 6. Each IMF represented the variation characteristics of observed and simulated runoff at a specific temporal scale. Specifically, IMF1 and IMF2, which exhibited high frequency and low amplitude, typically corresponded to rapid response processes driven by precipitation. IMFs 3 through 5, characterized by medium frequencies, reflected seasonal fluctuations or interactions between groundwater and surface water. IMF6 and IMF7 revealed the long-term trends and seasonal regulatory effects. Dominant periods of each IMF component were extracted using Fast Fourier Transform. The results showed that IMF1 and IMF2 of the observed runoff corresponded to dominant periods of approximately 6 and 12 months, respectively. The periods of IMF3, IMF4, and IMF5 were approximately 31 months, 102 months, and 204 months, indicating the presence of long-term climate modulation in the study area. IMF6 had a dominant period of approximately 408 months, reflecting the influence of long-term climate change on the runoff. Moreover, IMF2 and IMF5 were the main contributors to the overall energy, accounting for 16.99% and 12.51%, respectively, suggesting that precipitation variability and seasonal climate fluctuations played dominant roles in the region’s runoff variability.

All schemes exhibited consistency with the observed data in terms of the long-term trend periods represented by IMF6 and IMF7, indicating similar capabilities in simulating long-term trends, and all showed a trend of first increasing and then decreasing on the annual scale. CREST performed best at high frequencies, with the amplitudes of IMF1 and IMF2 being largely consistent with observations, both dominated by a 12-month cycle. However, the principal energy contribution terms of its simulated runoff were IMF7 and IMF6, accounting for 64.18% and 13.75% respectively, indicating that this scheme failed to capture the dominant scales of observed runoff variability. With the exception of IMF3, whose dominant cycle is 29 months, the dominant cycles of the other IMFs under LSTM are consistent with the observed runoff. Nevertheless, the principal energy contribution terms of the simulated runoff under this scheme are the same as those of CREST. Coupling successfully reproduced the dominant periods of IMF1 through IMF5, with the dominant cycles of IMF3 to IMF5 being largely consistent with the observations at 29, 102, and 204 months, respectively, demonstrating a strong ability to capture variations across multiple time scales. Under the Coupling scheme, the principal energy contributions are IMF2 and IMF6, accounting for 32.54% and 30.77%, respectively. Overall, all schemes are able to capture the long-term variation trend of runoff at the annual scale, while Coupling, LSTM, and CREST can each accurately simulate the seasonal and daily variation cycles of runoff. However, except for Coupling, the other two schemes fail to identify that the main variation in observed runoff is driven by rainfall fluctuations and the seasonality of climate.

4.3. Performance Analysis of Reservoir Operation

Evaluate the impact of the simulated runoff obtained under each scheme on reservoir regulation results, with particular attention given to its adaptability to reservoir operation. Figure 7a,b illustrate the variations in monthly hydropower generation and surplus water, respectively, resulting from the optimized operation based on the simulated runoff of each scheme. It was observed that the hydropower generation process based on observed runoff exhibited relatively stable seasonal variation, characterized by distinct cycles of high and low peaks, with a multi-annual average generation of 164.62 × 10⁸ kW·h. In comparison, the result of CREST presented pronounced generation peaks during several periods, especially in the early flood season, where the simulated hydropower output significantly exceeded the observed values, with a multi-annual average generation of 168.48 × 10⁸ kW·h. Moreover, CREST underestimated runoff during the dry season, leading to significantly lower hydropower generation than the observed data, and even resulting in periods of continuous power shortage. The above results indicated that although the total power generation derived from the runoff simulated by CREST was close to the observed value, the intra-annual distribution of power generation differed substantially from the observations. Specifically, overestimation during flood periods and underestimation during dry periods have a considerable impact on hydropower generation. The hydropower generation process of LSTM was smoother, and the multi-annul average value of it was 166.66 × 10⁸ kW·h. Although its output was slightly lower than the observed values, it generally followed the same overall trend. The Coupling most closely matched the temporal structure of the observed data. It provided sufficient hydropower support during wet season and maintained reasonable output during dry periods, demonstrating strong temporal simulation capability. Moreover, the multi-annual average power generation of Coupling was also the closest to the observed value, which was 166.43 × 10⁸ kW·h. A further analysis was conducted on the process of surplus water during reservoir operation to assess the effectiveness under different schemes. In terms of the total annual surplus water volume, the operation based on observed runoff resulted in a total spillage of 9.84 × 10⁸ m³, while the operation based on simulated runoff under CREST was 9.83 × 10⁸ m³, which was consistent with the value derived from the observed runoff. The total volume of surplus water under LSTM and Coupling were 3.73 × 10⁵ m³ and 3.45 × 10⁵ m³, respectively, both of which were lower than the value derived from the observed runoff.

Additionally, compared with LSTM and Coupling, CREST exhibited a higher surplus water during the flood season, whereas LSTM and Coupling showed lower for most periods. Considering results of power generation and surplus water together, it can be seen that, in most years, the power generation derived from the observed runoff and from the runoff simulated by each scheme reached full-capacity levels between July and September, meaning that flows during the flood-peak period far exceeded turbine capacity and result in substantial surplus water. Therefore, when observed runoff was used as the input to the reservoir optimal operation, large surplus water occurs between July and September, whereas LSTM and Coupling—which underestimated peak flows during the flood season—yielded a smaller value, indicating that the errors in the runoff simulated by LSTM and Coupling did not propagate through the regulation scheme to affect the generation results.

In summary, the runoff simulated by CREST resulted in the weakest performance following reservoir regulation, mainly reflected in the process of power generation and surplus water within the year. In comparison, the LSTM and Coupling schemes produced results that were more consistent with those derived from the operation based on observed runoff. Notably, the Coupling scheme showed the highest degree of agreement with the hydropower generation and surplus water processes derived from observed runoff, indicating superior capability in capturing the temporal dynamics of reservoir operation. Additionally, the underestimation of peak runoff by LSTM and Coupling does not propagate through the dispatch scheme to affect the power generation results.

5. Discussion

Although substantial progress has been made in both runoff simulation and reservoir operation optimization, the synergistic effects between these two components within practical regulation frameworks have not yet been fully explored [7,13,15]. Traditional research on runoff simulation has largely focused on enhancing prediction accuracy, while studies on reservoir operation have emphasized the development of more sophisticated and computationally efficient optimization algorithms [13]. However, in real-world applications, the performance of reservoir regulation depends heavily on the accuracy of runoff inputs. This is particularly critical during short-term scheduling and periods of transition between wet and dry seasons, where simulation errors can directly affect the feasibility of operational strategies and the efficiency of hydropower utilization. Accordingly, our study does not aim to optimize either simulation or regulation in isolation, but instead focuses on assessing the practical effectiveness of an integrated simulation–operation framework.

Specifically, to reduce simulation errors during the transition between wet and dry seasons, this study incorporated seasonal characteristics of precipitation into the runoff simulation process. Furthermore, while keeping the reservoir regulation algorithm unchanged, we systematically compared the impacts of simulated runoff with varying levels of accuracy on reservoir operation outcomes. The results indicated that improvements in simulation accuracy did not translate linearly into enhanced operational performance, particularly under extreme hydrological conditions. This suggests that relying solely on optimization algorithms or focusing exclusively on simulation accuracy metrics is insufficient for significantly improving hydropower utilization efficiency. Instead, greater attention should be given to the process-based characteristics of high-quality simulation data. In summary, this study underscored that hydrological simulation and regulation optimization were not independent processes. A comprehensive analysis that simultaneously improved simulation accuracy and examined both runoff and reservoir operation dynamics was essential to fully realize the potential for more efficient utilization of hydropower resources.

It should also be noted that in the optimization regulation of this study, the observed runoff after the reservoir operation resulted in different hydropower generation compared to actual generation recorded under real-world operations. This discrepancy is primarily attributed to the fact that the Ertan Reservoir project is primarily designed for power generation, but also serves multiple additional functions. As the main objective of this study was to investigate the impact of various simulated runoff inputs on reservoir-based hydropower generation, other secondary objectives were not incorporated into the optimization framework.

6. Conclusions

This study focused on the Yalong River Basin, a region rich in hydropower resources. By incorporating the seasonal characteristics of precipitation, a unidirectional coupled runoff simulation framework was developed based on the distributed hydrological model CREST and the LSTM neural network, and the Progressive Optimality Algorithm was employed to solve the decision-making optimization problem in reservoir regulation. This framework enhanced the accuracy of runoff simulation and enabled an in-depth investigation of the impacts of different simulation schemes on reservoir operation under optimized regulation.

From the perspective of runoff simulation metrics, the LSTM outperformed CREST in all aspects. Moreover, after incorporating the seasonal characteristics of precipitation based on the unidirectional coupling approach, the accuracy of the simulated runoff was significantly improved. Compared with the uncoupled CREST model, the Coupling scheme increased the NSE and R2 by 6% and 4%, respectively, while reducing the RMSE by 24%. From the process-based perspective, CREST tended to underestimate both low-flow during the dry periods and peak-flow during wet periods, and it also reduced the interannual variability of autumn and winter runoff. Furthermore, it failed to accurately capture the rapid runoff responses driven by precipitation events. While LSTM also underestimated peak flows, it more effectively reproduced low-flow during the dry season. The Coupling, after unidirectional coupling, filtered out the low-flow simulation bias, although it was slightly less effective than LSTM in capturing peak flow magnitudes. Furthermore, both LSTM and Coupling were capable of reproducing the interannual variation and amplitude of seasonal runoff, and can demonstrate improved accuracy in detecting the periodic signals embedded within runoff processes at multiple temporal scales.

The hydropower generation and surplus water obtained under Coupling after reservoir optimal operation were relatively close to those based on the observed runoff, with errors of 1.09% and –21.64%, respectively. In addition, Coupling corrected the pronounced hydropower generation peaks observed in multiple periods under CREST and achieved the highest consistency with the observed data in terms of temporal structure, demonstrating strong capability in simulating time-dependent processes. These findings indicated that incorporating seasonal precipitation characteristics within the unidirectional coupling framework that integrated process-driven and data-driven methods can effectively enhance runoff simulation performance and improve the alignment between simulated and observed reservoir operation under optimized regulation.