Hierarchical Temporal-Scale Framework for Real-Time Streamflow Prediction in Reservoir-Regulated Basins

Abstract

Reservoir construction has profoundly altered natural runoff evolution in river basins. Dynamic conflicts among multi-objective operational strategies—such as flood control, water supply, and ecological compensation—across varying temporal scales exacerbate uncertainties in runoff prediction, primarily due to the complex interplay between hydrological rhythm variations and anthropogenic regulation. To address these challenges, this study proposes a hierarchical multi-scale coupling framework. Long short-term memory (LSTM) networks are employed to extract implicit operational patterns from long-term reservoir records at monthly and weekly scales, while short-term decision dynamics are captured through deviations from these established long-term rules. The proposed framework is validated in the Dongjiang River Basin, a key water source for the Guangdong–Hong Kong–Macao Greater Bay Area. Compared to single-scale models, the hierarchical approach improves prediction accuracy with an average Nash–Sutcliffe Efficiency (NSE) increase of 9.4% and reductions in the Root Mean Square Error (RMSE) and the Mean Absolute Error (MAE) of 13.2% and 9.6%, respectively. When coupled with a hydrological model, the framework enhances simulation accuracy in reservoir-regulated basins by up to 37.8%. By integrating multi-source decision variables, the framework captures the feedback mechanisms between natural flow variability and human interventions across temporal scales, providing a transferable strategy to reconcile operational conflicts with ecological flow requirements. Its flexibility supports optimized water allocation in regulated river basins, contributing to enhanced water security for downstream urban agglomerations.

1. Introduction

Streamflow constitutes a key element of the hydrological cycle and underpins effective river basin management, particularly in flood mitigation and water distribution planning [1,2]. To cope with the spatial–temporal variability of water availability, a vast number of reservoirs have been established globally. These infrastructures are designed to modulate river discharge for purposes including flood prevention and resource reallocation. With the growing emphasis on renewable energy, especially hydropower, the global reservoir count is projected to rise further [3]. At present, reservoir regulation affects approximately half of the world’s river systems [4]. The growing prevalence of reservoirs, particularly large multipurpose ones, has led to increased flow uniformity, reduced runoff variability [5], and attenuated peak flows in many basins [6].

However, reservoir operations are often complex and difficult to characterize, presenting significant challenges for accurately understanding and modeling streamflow in reservoir-regulated basins.

Accurate streamflow forecasting is fundamental to optimizing water resource management, enhancing flood resilience, and improving hydropower efficiency [7,8]. Two principal approaches are commonly used for monthly streamflow prediction: data-driven methods, which leverage machine learning (ML) techniques to learn complex relationships from hydrological and climatic variables—such as gridded meteorological datasets [9], satellite observations [10], and climate indices [11]—and process-based models, which simulate hydrological processes based on physical mechanisms and meteorological inputs [12].

Within the ML domain, deep learning (DL) architectures have shown strong capability in capturing nonlinear streamflow dynamics by extracting high-level features from large spatiotemporal datasets [13,14,15]. In contrast, distributed process-based models like the Variable Infiltration Capacity (VIC) and the Soil and Water Assessment Tool (SWAT) explicitly represent spatial heterogeneity in precipitation, soil properties, land cover, and topography through watershed discretization [16,17]. Reservoir operations increasingly influence hydrological processes across many basins, acting as decision-making hubs that reconcile competing objectives—such as flood control, ecological flow maintenance, and water supply reliability—under changing climatic and societal conditions [4,18]. These engineered systems introduce non-stationarity into river flow regimes through complex operational rule curves.

Recent advances in reservoir operation modeling reflect the integration of machine learning and process-based approaches. For instance, the Adaptive Neuro-Fuzzy Inference System (ANFIS) combines the adaptability of neural networks with fuzzy rule-based logic, using inflow and storage data to simulate reservoir release decisions [19]. At the continental scale, improved reservoir parameterization frameworks have enhanced the representation of storage and release dynamics through multi-tiered algorithms that approximate physical constraints and operational rules [20]. Deep learning models, particularly recurrent neural networks (RNNs), have shown strong performance in capturing the complexity of multipurpose reservoir operations across diverse hydrological and climatic regimes [21].

Meanwhile, process-based models such as the Soil and Water Assessment Tool (SWAT) incorporate empirical reservoir outflow estimation methods within distributed hydrological routing schemes to simulate reservoir regulation under varying hydrological conditions [22]. Recent improvements, such as the integration of operation chart-based reservoir modules, have further enhanced SWAT’s ability to simulate water levels and regulated outflows in multipurpose reservoirs [23].

Despite these advancements, both data-driven and physically based models face inherent limitations when applied to complex multi-reservoir systems. Deep learning models are often criticized for their lack of physical interpretability, as they do not explicitly represent runoff generation mechanisms, which can undermine their credibility in decision-making contexts [24]. Moreover, they tend to exhibit poor generalization when exposed to hydrometeorological extremes not present in the training data, raising concerns under non-stationary climate conditions [25]. Process-based models, in turn, encounter challenges related to the asynchronous and conflicting objectives of multi-reservoir operations—such as reconciling sub-daily hydropower demands with seasonal irrigation needs—which are often misaligned with basin-scale hydroclimatic inputs [26]. Incomplete knowledge of reservoir operation rules and underlying physical processes introduces parameter uncertainty and structural equifinality [27], while high-resolution input data requirements increase computational burden, particularly in large-scale applications [28].

To address these challenges, hybrid frameworks have emerged that aim to combine the strengths of both modeling paradigms. For example, Ref. [29] proposed a real-time forecasting model that integrates data assimilation with process-based constraints to improve streamflow and reservoir outflow predictions. However, human decision-making remains a source of irreducible uncertainty, as operators frequently adjust predefined operation curves based on judgment and experience when faced with uncertain inflows and fluctuating water demands [30]. Recent studies have further explored the time-varying information sensitivity of reservoir systems, demonstrating the potential of adaptive and coordinated release strategies to better manage hydrological variability [26,31]. Nonetheless, the influence of reservoir operations on downstream flow regimes remains underexplored, particularly regarding their hydrological and ecological consequences.

As previously discussed, both data-driven and process-based methods exhibit distinct advantages and inherent constraints. To harness their complementary strengths, this study proposes a hybrid modeling framework for streamflow prediction in reservoir-regulated basins. The specific objectives were as follows: (1) to construct data-driven reservoir operation schemes using long short-term memory (LSTM) networks, aiming to forecast reservoir releases and storage dynamics under diverse decision-making scenarios and across multiple temporal scales; (2) to develop an integrated hydrological modeling framework that incorporates the proposed reservoir operation schemes, adaptable to different levels of reservoir data availability (as illustrated in Figure 1).

By coupling data-driven and hydrological models, this study captures both long-term patterns and short-term anomalies in reservoir operations. This enables the modeling of dynamic and often competing water management objectives (e.g., flood control, water supply, and ecological maintenance) across multiple time scales, offering a new approach for streamflow simulation in reservoir-constrained basins.

2. Materials and Methods

2.1. Long Short-Term Memory Model

Long short-term memory (LSTM) [32] is a powerful type of recurrent neural network (RNN). The internal computations of the LSTM cell in this study are summarized in Appendix A (Figure A1). There are three gates (input, forgetting, and output gates) along with an additional memory cell, which compose the LSTM model. The model facilitates efficient parallel training, leading to accelerated convergence. Moreover, it demonstrates robustness in modeling complex relationships, managing noise, and capturing nonlinear patterns. These characteristics bestow upon LSTM models a distinct advantage in various time-series prediction tasks [33], and they have a wide range of applications in hydrology and water resource management [34,35,36].

This study employed Optuna to automatically optimize six crucial hyperparameters: the learning rate, the number of hidden layers, the number of neurons per layer, the batch size, the dropout rate, and the number of training epochs. Optuna leverages the Tree-Structured Parzen Estimator (TPE), a Bayesian optimization algorithm, to iteratively search for the optimal parameter set [37]. By defining a tailored objective function, this method enhances model generalization and training efficiency on the dataset [38,39,40]. To mitigate overfitting, fivefold cross-validation is incorporated during the training phase.

2.2. Hierarchical Temporal-Scale Framework

This section is dedicated to the development of data-driven reservoir operation schemes utilizing LSTM across multiple time scales, with the objective of forecasting daily reservoir releases and storage within hydrological models. Reservoirs play a fundamental role in managing streamflow for multiple societal needs, including water supply, irrigation, flood regulation, hydropower generation, inland navigation, and recreation [41]. To support operational planning, a reservoir regulation chart is commonly employed, which defines standard operating boundaries. These boundaries are typically represented by the upper and lower limits of permissible water levels, referred to as the upper and lower rule curves [42]. Although reservoir operators are generally expected to follow such regulatory guidelines, operational flexibility is often exercised in response to dynamic hydrological conditions and shifting water demands. In practice, decisions may incorporate expert judgment and on-the-ground experience to address uncertainties and non-stationarity in inflow regimes. From an operational perspective, reservoir release decisions are informed by a combination of historical hydrological records, real-time monitoring data (e.g., water levels), and anticipated future conditions such as short-term precipitation or streamflow forecasts. For instance, flood control operations are typically adjusted in response to current reservoir storage levels, whereas water supply scheduling often considers seasonal inflow projections and long-term availability trends. However, from a decision-making perspective, reservoir releases are not only the relationship between past hydrological information and the current operating state (such as water level), but also the changes in future hydrological information (such as precipitation) that link the current release and specific operating modes during a given period. For instance, flood control decisions may be informed by the current reservoir water-level and streamflow forecasts for the upcoming days, whereas water supply strategies may emphasize long-term variations in streamflow.

Reservoir operation decisions are likely to rely more on the distant hydrological information beyond the past or the upcoming week or month, which is important for reservoir operators to prioritize or deal with various, uncertain hydroclimatic conditions. Gaining clarity on the role of the priority in operators’ decisions is critical to understanding realistic operation rules, as well as building more realistic operation models [41]. We capture the variations between the daily scale and the coarse scales (weekly and monthly scales) by constructing a hierarchical data-driven model and determine the priority of different time scales in the decision-making of operators. Specifically, we initially establish coarse-scale (weekly and monthly) LSTM models that are employed to derive generic operation rules for reservoirs. Subsequently, hierarchical models (weekly–daily scale, WD; monthly–daily scale, MD) are established to investigate the release variations ($R_f$) between the reservoir’s daily operations and the operations inferred from the generic operation rules, thereby facilitating the determination of the reservoir’s operational pattern based on the final prediction accuracy (Figure 2).

The calculation of $R_f$ is defined as follows:

$${\widehat{R}}_{*f}=R_d-{\widehat{R}}_{*s}$$

(1)

where ${\widehat{R}}_{\mathrm{*}f}$ is the reservoir monthly/weekly variation value, $R_d$ is the daily release, and ${\widehat{R}}_{*s}$ is the reservoir monthly/weekly release (* present weekly or monthly). A positive $R_f$ value signifies that the daily release exceeds the planned release, whereas a negative value denotes the opposite.

The release fluctuation ratio, $R_f$, reflects short-term deviations from reservoir releases defined under long-term operating objectives, and its primary purpose is to characterize operator behavioral responses to complex influencing factors. Reservoir release decisions are shaped by both inflow and historical release patterns, as they are typically governed by water balance constraints. Moreover, due to the continuous operation of hydraulic infrastructure, such as gates and outlets, reservoir outflows display short-term persistence and autocorrelation.

Building upon these considerations, we constructed the input dataset for the LSTM-based reservoir operation scheme. The input structure consists of four types of vectors: an inflow vector [$I_d$, $I_w,I_m$], a release vector [$R_d$, $R_w,R_m$], a water-level vector [$L_d$, $L_w,L_m$], and a meteorological vector [$M_d$, $M_w,M_m$], where the subscripts d, t, and m denote daily, weekly, and monthly scales, respectively. The meteorological vector includes precipitation, temperature, humidity, and wind speed, representing local hydroclimatic conditions relevant to reservoir dynamics. Specifically, temperature, humidity, and wind speed characterize atmospheric demand, which influences evaporative losses, whereas precipitation serves as an indicator of potential runoff contributions. Runoff is further categorized into local generation and upstream catchment inflows, with the latter represented by the inflow vector [31]. However, effective features can enhance the predictive capacity of data-driven models [43]. Features related to the model’s target (reservoir release) are selected as determined by the Spearman coefficient [44,45] value. The time lag of the time-series data is evaluated through autocorrelation and partial autocorrelation [46], and the lagged variables are employed as additional inputs to strengthen the generalization ability of the model [44]. In addition, each input feature is normalized by the maximum historical storage value observed during the simulation period. This normalization eliminates scale-related biases associated with reservoir size and facilitates more efficient hyperparameter tuning [41].

2.3. Water Allocation and Simulation Model

In this study, the Water Allocation and Simulation (WAS) model was employed as the core platform for simulating interactions between natural and anthropogenic components of the watershed water cycle. The model is particularly well-suited for representing river basins under intensive human regulation [47,48,49]. Compared with conventional hydrological models, WAS integrates a real-time dynamic framework that facilitates the coupling of natural hydrological processes with anthropogenic water uses, including reservoir management, irrigation, and water reuse. The WAS model has been widely applied across China [50,51,52] and is available as a publicly accessible software package (http://new.ewater.net.cn (accessed on 3 January 2025)). Due to the lack of detailed, sector-specific water withdrawal data in the study area, the influence of anthropogenic water use on streamflow could not be isolated or quantified. Therefore, direct water withdrawals within the basin were not explicitly represented in the model in order to isolate the effects of reservoir regulation on streamflow. The required input datasets include land use/land cover (LUCC) information, meteorological forcing data, and reservoir operation records. LUCC data for the year 2015 were obtained from Landsat TM, ETM+, and Landsat 8 imagery, with a spatial resolution of 1 km. The meteorological variables are consistent with those used in the data-driven model component.

2.4. Performance Metrics

This study employed feature selection and time-lag analysis based on methodologies established in prior research [44,45,46] (refer to Figure A1, Figure A2, Figure A3, Figure A4, Figure A5, Figure A6 and Figure A7 and

Table A1. Input feature correlation results for three DRB reservoir discharge models.

Stations	Input Variables	Correlation
FSB	Water level (t), Water level (t − 1), Water level (t − 2)	0.33–0.35
	Min_T (t), Min_T (t − 1), Min_T (t − 2)	0.17–0.18
	Ave_RHU (t), Ave_RHU (t − 1), Ave_RHU (t − 2)	0.17–0.18
	PCP (t), PCP (t − 1), PCP (t − 2)	0.06–0.07
	Inflow (t), Inflow (t − 1), Inflow (t − 2)	0.45–0.48
	Release (t − 1), Release (t − 2)	0.84–0.89
XFJ	Water level (t), Water level (t − 1), Water level (t − 2)	0.21–0.23
	Max_T (t), Max_T (t − 1), Max_T (t − 2)	−0.13– (−0.12)
	PCP (t), PCP (t − 1), PCP (t − 2)	−0.02– (−0.04)
	Inflow (t), Inflow (t − 1), Inflow (t − 2)	$\approx$−0.01
	Release (t − 1), Release (t − 2)	0.83–0.87
BPZ	Water level (t), Water level (t − 1), Water level (t − 2)	0.46–0.48
	Max_T (t), Max_T (t − 1), Max_T (t − 2)	$\approx$−0.05
	Ave_Wind (t), Ave_Wind (t − 1), Ave_Wind (t − 2)	0.01–0.02
	PCP (t), PCP (t − 1), PCP (t − 2)	$\approx$−0.03
	Inflow (t), Inflow (t − 1), Inflow (t − 2)	0.19–0.21
	Release (t − 1), Release (t − 2)	0.91–0.95

). Furthermore, we quantified the magnitude of the discrepancy between predicted values and observed values using the Mean Absolute Error (MAE) and the Root Mean Square Error (RMSE). Furthermore, to evaluate the LSTM-based hybrid model’s performance, a benchmark model was used. The average method [53] states that the forecasts of all future values are equal to the average of the historical data. In this study, we utilized the average of the observed values for the same calendar month as the forecast value for this period.

The Nash–Sutcliffe Efficiency (NSE) (Equation (2)) coefficient ranges from −∞ to 1, reflecting the correspondence between observed and simulated mean values. A value approaching 1 indicates superior simulation performance [23,29]. Pearson’s correlation coefficient (Equation (3)) serves as a statistical measure for evaluating model performance by quantifying the relationship between two variables [44,54].

$$NSE=1-{\displaystyle \frac{{\sum }_{i=1}^N{\left(O_i-P_i\right)}^2}{{\sum }_{i=1}^N{\left(O_i-\overline{O}\right)}^2}}$$

(2)

$$R={\displaystyle \frac{{\sum }_{i=1}^N\left(O_i-\overline{O}\right)\left(P_i-\overline{P}\right)}{\sqrt{{\sum }_{i=1}^N{\left(O_i-\overline{O}\right)}^2}\sqrt{{\sum }_{i=1}^N{\left(P_i-\overline{P}\right)}^2}}}$$

(3)

where $N$ is the number of observation values; O represents the observed variable; $P$ represents the predicted variable; $\overline{O}$ represents the average of the observed values; $\overline{P}$ represents the average of the predict values; and |R| > 0.6 means a strong correlation, while |R| > 0.8 signifies a very strong correlation.

3. Dongjiang River Basin: Study Area and Data Description

3.1. Study Area

The Dongjiang River Basin (DRB), located in South China, spans an area of approximately 35,340 km² (Figure 3), representing about 24.3% of the total area of the Pearl River Basin within Guangdong Province [55]. This region is predominantly influenced by the South Subtropical Moist Monsoon, resulting in high annual precipitation averaging 1700 mm. Precipitation is markedly seasonal, with roughly 80% occurring between April and September [56].

The DRB plays a vital role in supplying freshwater for domestic, industrial, and ecological purposes, particularly to major urban centers such as Guangzhou, Shenzhen, and Hong Kong. It supports water demand for nearly 40 million residents, making it an essential component of the Guangdong–Hong Kong–Macao Greater Bay Area (GHMB) [23]. The Boluo station, the principal hydrological gauge in the basin, has a contributing drainage area of 25,930 km² [57].

3.2. Reservoir Information and Operation Data

While there are a large number of reservoirs in the DRB, most of them possess limited capacity or are located away from the rivers, thus having a trivial effect on the streamflow of the downstream areas. As shown in Figure 3, this study identifies and incorporates three reservoirs, namely, Fengshuba (FSB), Xinfengjiang (XFJ), and Baipenzhu (BPZ), arranged sequentially from upstream to downstream. These three reservoirs have a total capacity of 17.05 km³, accounting for approximately 41.3% of the Dongjiang River’s long-term average annual runoff [58].

Table 1. Basic parameters of the reservoirs.

Reservoir	Catchment Area/km2	Capacity/km3	Operation Year	Data Period
FSB	5150	1.93	1983	2011–2019
XFJ	5734	13.9	1960	2011–2019
BPZ	856	1.22	1987	2011–2019

provides basic parameters for these multifunctional reservoirs. We collected the daily water level, release, and inflow data for FSB, XFJ, and BPZ, which were then fed into the data-driven reservoir operation scheme for training, testing, and predicting. The historical reservoir operation data were obtained from the Dongjiang River Basin Authority of Guangdong Province.

3.3. Streamflow Data and Meteorological Data

In this study, monthly streamflow records from the Boluo hydrological station (Figure 3) for the periods 1960–2000 and 2011–2019 were employed for hydrological model calibration and validation. Corresponding daily meteorological variables were sourced from the China Surface Climatological Data Daily Value Dataset (v3.0), issued by the National Meteorological Center of China (https://m.data.cma.cn (accessed on 15 January 2025)).

4. Results

4.1. Reservoir Operation Simulations with Data-Driven Operation Scheme on a Single Time Scale

To characterize the general operational patterns of the reservoirs, data-driven models were trained, cross-validated, and tested for FSB, XFJ, and BPZ at both weekly and monthly temporal resolutions, using historical operation data from 1 January 2011 to 31 December 2019. The training period (1 January 2011 to 14 March 2018, accounting for 80% of the data) and the testing period (15 March 2018 to 31 December 2019, 20%) were consistent across all reservoirs and used for LSTM modeling, as illustrated in Figure 4. The corresponding simulation results at weekly and monthly scales are summarized in

Table 2. Metrics of reservoir operation simulations with data-driven model.

Period	Performance Index	Reservoir
		FSB		XFJ		BPZ
		Monthly	Weekly	Monthly	Weekly	Monthly	Weekly
Training	NSE	0.97	0.94	0.97	0.95	0.95	0.82
	R	0.99	0.97	0.98	0.98	0.98	0.9
	MAE (m3/s)	10.81	11.69	5.65	10.4	2.7	6.23
	RMSE (m3/s)	15.13	23.8	16.03	21	7.16	20.65
Testing	NSE	0.95	0.93	0.95	0.93	0.92	0.83
	R	0.99	0.96	0.97	0.96	0.96	0.91
	MAE (m3/s)	16.72	13.08	7.42	13.99	3.34	6.96
	RMSE (m3/s)	27.06	37.91	17.6	25.46	9.28	17.07

. Model parameter settings are summarized in

Table A2. Hyperparameter optimization results of different models.

Model	Station	Number of Hidden Layers	Number of Hidden Size	Batch Size	Epochs	Learning Rate	Dropout Rate
Daily release	FSB	2	196	100	56	0.015	0.001
	XFJ	2	40	111	88	0.003	0.04
	BPZ	2	127	125	100	0.00036	0.056
MDF	FSB	2	77	128	64	0.0015	0.001
	XFJ	2	33	49	95	0.0048	0.027
	BPZ	2	31	46	95	0.0009	0.0086
WDF	FSB	2	40	91	36	0.0022	0.0155
	XFJ	2	125	98	39	0.00137	0
	BPZ	2	154	57	180	0.0002	0.02

. For the FSB reservoir, the Nash–Sutcliffe Efficiency (NSE) during the monthly training and testing periods reached 0.97 and 0.95, respectively. At the weekly scale, the NSE values were 0.94 for training and 0.93 for testing. Overall, the simulation accuracy at the monthly scale slightly outperformed that at the weekly scale. Similar trends were observed for the XFJ and BPZ reservoirs across both temporal scales. However, in the case of FSB, despite comparable NSE values across the periods, significant deviations were observed in the Mean Absolute Error (MAE) and the Root Mean Square Error (RMSE), a pattern not evident in the other two reservoirs.

Further analysis revealed that the FSB reservoir experienced a wide discharge range, with maximum and minimum outflows of 775.1 m³/s and 39.4 m³/s during the training period and 1050.4 m³/s and 24.5 m³/s during the testing period, resulting in a maximum–minimum difference of 735.8 m³/s (training) and 1025.9 m³/s (testing). For the XFJ reservoir, the maximum–minimum differences were 447.1 m³/s (training: 513.4 to 66.35 m³/s) and 445.2 m³/s (testing: 494.7 to 49.5 m³/s). In contrast, the BPZ reservoir exhibited a range of 876.3 m³/s during training (from 880.6 to 4.3 m³/s) and 368.0 m³/s during testing (from 375.4 to 7.4 m³/s). These substantial fluctuations highlight the operational complexity and heterogeneity among the reservoirs.

Given that different metrics reflect different facets of model performance, NSE and Pearson’s correlation coefficient (R) primarily assess the model’s ability to capture overall temporal dynamics, while MAE and RMSE quantify absolute deviations, with RMSE being particularly sensitive to large errors. Assuming short-term stability in watershed hydrometeorological conditions, the historical mean method [53] was employed as a baseline model to evaluate LSTM performance.

As shown in

Table 3. Metrics of reservoir operation simulations with benchmark model.

Period	Performance Index	Reservoir
		FSB		XFJ		BPZ
		Monthly Benchmark	Weekly Benchmark	Monthly Benchmark	Weekly Benchmark	Monthly Benchmark	Weekly Benchmark
Training	NSE	0.26	0.23	0.15	0.16	0.15	0.39
	R	0.51	0.48	0.38	0.4	0.39	0.39
	MAE (m3/s)	50	56.87	57.45	62.98	15.12	6.23
	RMSE (m3/s)	71	83.89	82.67	90.52	30.26	20.65
Testing	NSE	0.39	0.41	0.41	0.37	0.42	0.67
	R	0.63	0.64	0.64	0.61	0.65	0.67
	MAE (m3/s)	60.55	64.94	40.18	48.59	13.44	6.96
	RMSE (m3/s)	97.81	13.07	59.61	73.86	24.32	17.07

, the baseline model consistently underperformed relative to the LSTM-based approach across all reservoirs, with lower overall simulation accuracy and greater variability in RMSE and MAE values. Notably, the LSTM model demonstrated enhanced capability in reproducing reservoir discharge dynamics, particularly under transient hydrological regimes. This comparative analysis confirms that the data-driven operation schemes effectively capture and generalize the underlying regulation patterns of the three reservoirs. Accordingly, the LSTM framework was adopted as the foundation for constructing the hierarchical multi-temporal reservoir operation model, owing to its superior balance between predictive accuracy and robustness.

4.2. Simulation of Reservoir Operational Behavior Across Time Scales

The simulation results for the FSB reservoir’s MD and WD are presented in Figure 5. Figure 5a,b depict the simulated outflow deviations at monthly and weekly scales, respectively. The monthly deviation has a range of (−200, 1000), while the weekly deviation spans (−800, 800), indicating that the FSB reservoir sets overall outflow targets on a monthly basis, with weekly adjustments operating within this framework. Figure 4c reveals that the MD simulation accuracy is superior to that of the WD, with MD outflows aligning more closely with the 1:1 line.

In contrast, Figure 6a,b show that the XFJ reservoir operates with outflow deviations of −300 to 400 at both scales, suggesting a lack of dependence on temporal-scale information. This can be attributed to two factors: (1) The XFJ reservoir has substantial regulation capacity, given its comparable catchment area to the FSB reservoir, but a storage capacity seven times greater. (2) The XFJ reservoir’s operational decisions primarily focus on quantitative control objectives, such as hydropower generation and ecological water replenishment. Figure 6c indicates that the MD release results of the XFJ reservoir show less deviation from the 1:1 line in comparison with the WD results.

Figure 7 indicates that while the BPZ reservoir monthly deviation has a range of (−200, 1200) and the weekly deviation has a range of (−600, 800), the deviation patterns across scales, after excluding extremes, are similar to those of the FSB reservoir. Figure 7c confirms that the MD simulation results for the BPZ reservoir exceed those of the WD.

To examine the effects of temporal resolution, individual daily outflow prediction models were developed for the FSB, XFJ, and BPZ reservoirs using identical input configurations. As presented in

Table 4. Streamflow simulation performance under single-scale and hierarchical MD- and WD-scale models for individual reservoirs.

Res_Name	Performance Metrics	Model
		MD	WD	Daily
FSB	NSE	0.98	0.94	0.92
XFJ		0.97	0.93	0.83
BPZ		0.96	0.85	0.78
FSB	RMSE	20.7	37.5	43.4
XFJ		18.7	26.7	40.2
BPZ		8.9	17.1	20.9
FSB	MAE	5.5	12.7	22.4
XFJ		10.5	16.7	25.5
BPZ		3.5	5.8	8.18
FSB	R	0.99	0.95	0.93
XFJ		0.99	0.98	0.94
BPZ		0.99	0.95	0.93

, both monthly (MD-scale) and weekly (WD-scale) models consistently outperformed the daily-resolution models across all three reservoirs. Specifically, the MD-scale models yielded average improvements of 12.7% in NSE, along with reductions of 18.7 and 12.2 in RMSE and MAE, respectively. The WD-scale models also demonstrated moderate performance gains, with a 6.3% increase in NSE and reductions of 7.7 and 6.96 in RMSE and MAE, respectively. For all three reservoirs, the MD-scale models yielded maximum correlation coefficients of up to 0.99, indicating an excellent agreement between predicted and observed outflows.

A Taylor diagram [59] was utilized to compare the performance of the daily model against the MD and WD models (Figure 8). The results indicate that the MD model consistently outperformed the WD model, which in turn surpassed the daily model across all three reservoirs. Thus, the MD model was selected as the operational forecasting model for the FSB, XFJ, and BPZ reservoirs.

4.3. Streamflow Forecasting for Reservoir-Regulated Basins

To replicate natural flow conditions in the DRB, the WAS model was calibrated without reservoir impacts using monthly natural runoff data from the Boluo station (1960–1982) and validated with data from 1983 to 2000. The calibration and validation time frames were constrained by data availability, as the FSB and BPZ reservoirs became operational in late 1983, and only the XFJ reservoir was operational prior to this period. The results show that the WAS model performed satisfactorily in a no-reservoir scenario, achieving calibration metrics of NSE: 0.75 and R: 0.89 and validation metrics of NSE: 0.64 and R: 0.83 (Figure 9). These results indicate the model’s efficacy in accurately simulating the natural hydrological processes of the DRB.

Following model calibration and validation, we collected monthly observed runoff data for the period 2011–2019 in order to evaluate model performance under natural flow assumptions and to enable comparison with regulated flow conditions. To further assess the representativeness and temporal consistency of the historical hydrological conditions, we compared the monthly runoff climatology of the long-term historical period (1960–2000) with that of the recently observed period (2011–2019). The analysis revealed a strong correlation between the two datasets, with a Pearson correlation coefficient of 0.84, indicating that seasonal runoff patterns during 2011–2019 are largely consistent with those of 1960–2000 (Figure 10). This supports the validity of using the recently observed data as a baseline for comparative scenario analysis. Furthermore, Figure 11 illustrates the distribution of monthly runoff across both periods. Naturalized runoff (1960–2000) exhibits greater intra-annual variability, particularly during the flood season, while regulated runoff (2011–2019) shows more stable patterns, clearly reflecting the influence of reservoir operations in modulating seasonal flow dynamics.

To evaluate the capability of the data-driven model in simulating reservoir operations during this period, we conducted hydrological simulations for the FSB, XFJ, and BPZ reservoirs. Reservoir releases were estimated using a data-driven reservoir operation scheme based on the MD model. Due to the unavailability of daily streamflow observations at the Boluo station, the daily reservoir outflow simulated by the MD model was aggregated into monthly values to serve as reservoir input for the WAS model. Figure 12 presents a comparison between the observed monthly runoff at the Boluo station (black line) with two simulation scenarios: one without reservoir regulation (orange line) and another incorporating MD-based reservoir operations (blue line). The WAS model, when operated without considering reservoir regulation, yielded a relatively low monthly Nash–Sutcliffe Efficiency (NSE) of 0.45 (R = 0.77). In contrast, when the MD-driven reservoir operations were integrated, model performance improved substantially, with the monthly NSE increasing to 0.62 (R = 0.80). These results demonstrate the effectiveness of the proposed stratified modeling framework in capturing reservoir operational dynamics in the DRB. The integration of data-driven reservoir schemes enhances the simulation accuracy and provides a practical reference for runoff forecasting in reservoir-regulated basins.

5. Discussion

This study presents a data-driven reservoir operation modeling framework based on LSTM integrated into a hybrid system to enhance streamflow prediction in reservoir-regulated basins. While previous studies have explored reservoir operation modeling through either physically based approaches or data-driven methods with time-scale awareness, limitations remain in terms of flexibility, scalability, and real-world adaptability. For instance, the enhanced reservoir module developed within SWAT [23] improved outflow simulations by incorporating a rule-based operation chart. Additionally, study [29] applied an LSTM model for streamflow prediction in reservoir-influenced basins, demonstrating the feasibility of deep learning in this context. However, these approaches remain constrained either by predefined operational rules or by single-scale learning structures, which limit their ability to capture complex, dynamic reservoir behavior under variable hydroclimatic and socioeconomic conditions. In contrast, our approach does not rely on explicit rule encoding but learns directly from historical reservoir operation records. This enables the model to implicitly reflect multi-objective decision-making processes without prior assumptions. The resulting framework provides a more flexible and adaptive representation of real-world reservoir operations, particularly under scenarios with limited data availability and evolving management demands.

Similarly, the hierarchical temporal framework proposed in [31] highlighted the importance of modeling reservoir operations across multiple time scales. However, that study primarily focused on identifying large-scale operational patterns across hundreds of reservoirs using national-scale datasets. In contrast, our study targets a sub-basin scale and introduces a practical, multi-temporal modeling strategy grounded in actual operational records. We developed and validated a hierarchical LSTM scheme at monthly, weekly, and daily resolutions which captures inter-scale dependencies and improves model robustness under data-scarce conditions. This localized, high-resolution design enables more realistic simulation of reservoir release dynamics, particularly in catchments where operations are highly sensitive to short-term inflow variability and multi-level decision-making. Furthermore, the application of Optuna for hyperparameter optimization enhances model credibility by providing a more systematic and efficient alternative to traditional trial-and-error approaches [44,60].

Despite these advancements, several challenges remain. Although this study referred to previous research [44,54] to perform correlation analysis of input features, it remains difficult to interpret the specific influence of each input on model outputs due to the black-box nature of neural networks. Data quality and availability continue to affect model performance, and the incorporation of socioeconomic variables is constrained by limited observations, which may lead to deviations between simulated and observed streamflow (as shown in Figure 12). Moreover, only basic meteorological drivers were considered in this study, whereas in practice, the accuracy of reservoir release predictions also depends on the reliability of meteorological forecasts. In particular, for cold-region basins affected by snow- and icemelt, further model enhancements are needed to incorporate glacier and snowmelt runoff into the modeling framework, as these sources can influence reservoir inflows and seasonal runoff dynamics [61].

6. Conclusions

Integrating upstream reservoir operational rules is crucial for accurately simulating runoff in reservoir-regulated watersheds. This study proposes a hierarchical multi-temporal framework that leverages long short-term memory (LSTM) models to extract generalized operational patterns from historical reservoir data. By explicitly accounting for temporal-scale effects on reservoir management, the framework offers valuable insights for runoff modeling in data-scarce regions.

The multi-scale structure of the framework enables effective characterization of daily release patterns in alignment with established reservoir operation plans, facilitating closer emulation of real-world management practices. When coupled with hydrological models, the framework supports the exploration of interactions between natural hydrological processes and human interventions across temporal scales.

Future research should aim to improve the transparency and interpretability of data-driven methods and facilitate their integration with distributed hydrological process representations, thereby strengthening the robustness and predictive accuracy of watershed-scale simulations.