Comparison of Process-Driven SWAT Model and Data-Driven Machine Learning Techniques in Simulating Streamflow: A Case Study in the Fenhe River Basin

Abstract

Hydrological modeling is a crucial tool in hydrology and water resource management for analyzing runoff evolution patterns. In this study, the process-driven soil and water assessment tool (SWAT) model and data-driven machine learning techniques (XGBoost, random forest, LSTM, BILSTM, and GRU) were employed to simulate runoff at monthly and daily intervals in the Fenhe River basin, situated in the middle reaches of the Yellow River, respectively. The SWAT model demonstrated effective performance in simulating runoff at various scales, with the coefficient of determination (R²) exceeding 0.80 and the Nash–Sutcliffe efficiency (NSE) surpassing 0.79. Sensitivity analysis reveals varying degrees of sensitivity among the model parameters. Furthermore, the deep learning techniques (LSTM, BILSTM, and GRU) exhibited superior simulation generalization capabilities compared to the SWAT model across various scales. Additionally, the generalization abilities of traditional machine learning techniques (XGBoost and random forest) were comparable to the SWAT model. This indicates that deep learning techniques demonstrate remarkable stability and generalization capabilities across various scales. This analysis was motivated by the use of external continuous time series data as input and the application of deep learning techniques to internal mechanisms. Moreover, an integrated modeling approach was used to enhance simulation accuracy by combining the SWAT model with machine learning techniques. The results indicate that the integrated modeling approach improves simulation performance across various scales compared to the single-model approach. This research is significant for improving the efficiency of water resource utilization and management in the Fenhe River basin.

1. Introduction

Hydrological modeling, a crucial technique for simulating runoff [1], is vital for understanding runoff evolution patterns. It can be categorized into two main types: process-driven and data-driven models [2]. Improving the precision of runoff simulation is a significant challenge in hydrology [3]. Since the 1970s, the Fen River basin has been affected by human activities, including intensified agricultural irrigation and surface damage from coal mining, resulting in a continuous decline in runoff. Around the year 2000, there was even a complete flow cessation [4]. As the second largest tributary of the Yellow River basin, it supplies sustainable water to a large population for agricultural, industrial, and domestic purposes. It is crucial to explore the evolution of runoff under the influence of human activities.

Traditionally, runoff simulation is performed using hydrometeorological data and empirical formulas. The soil and water assessment tool (SWAT) model is a prominent process-driven hydrological model that employs mathematical and physical equations to replicate natural runoff processes. Numerous studies have demonstrated its effectiveness in runoff modeling [5], sediment transport [6], and pollutant estimation [7]. The SWAT model has been extensively used to model various types of river basins worldwide. Xuan et al. [8] applied the SWAT model to simulate the runoff process in a tributary of the Yarlung Zangbo River basin in data-scarce conditions, demonstrating satisfactory modeling performance and strong ability to mitigate the impact of limited hydrological data on model construction. Additionally, several studies have been conducted to simulate rainfall–runoff processes using the SWAT model in arid and semi-arid watersheds [9]. The results of the SWAT model depend on the selection of sensitive parameters. Cibin et al. [10] present a sensitivity and identifiability analysis of model parameters that influence streamflow generation in SWAT and indicate that streamflow varies in sensitivity to parameters across different climatic settings. The input databases are also vital for the SWAT model. Schuol et al. [11] developed the daily weather generator algorithm (dGen) to serve as the input dataset and demonstrated that using the dGen-simulated daily weather data resulted in a better match with the measured daily weather data. However, process-driven models encounter challenges such as cumbersome construction steps and low computational efficiency. Therefore, to overcome these problems, different artificial intelligence (AI) techniques can be used to estimate runoff based on the non-linear relationships between input and output data.

With the rise of AI techniques and increased accessibility to hydrological data, data-driven models have advanced rapidly. Data-driven machine learning techniques establish correlations between input features and target outcomes using the model’s internal mechanisms, independent of the physical processes governing natural runoff. Machine learning techniques like artificial neural networks (ANNs) and multiple linear regression (MLR) often achieve comparable or superior performance compared to process-driven models. These techniques have been widely used in runoff simulation [12,13], groundwater simulation [14], and soil moisture retrieval [15]. However, traditional machine learning techniques like ANNs struggle with managing prolonged dependencies in sequences [16,17].

Recently, numerous studies have demonstrated the advantages of deep learning (DL) techniques over traditional machine learning (ML) techniques [18]. The recurrent neural networks (RNNs) were specifically designed to analyze time series data using hidden layer loops. Long short-term memory (LSTM) networks [19], a type of RNNs, introduce “forget gate”, “input gate”, and “output gate” mechanisms to address long-term dependencies in time series data. LSTM networks are frequently employed in runoff simulation to analyze patterns using historical runoff and precipitation data [20]. Hu et al. [21] conducted a study in the upper Fenhe River basin to identify runoff and precipitation input features and found that the LSTM model performed better than conventional hydrological models. Feng et al. [22] employed various catchment attributes and meteorology for large-sample studies (CAMELS) data integration strategies with the LSTM model and validated its superior suitability for hydrological modeling. Many studies [23,24] have focused on using the LSTM model for runoff simulation and assessing performance variations across different prediction horizons. These studies indicate that extending the foresight period results in varying degrees of performance decline. However, increasing the number of neurons and training iterations improved forecasting accuracy.

Moreover, the integrated model approach utilized various model combinations to leverage effective information, capitalize on the strengths of different models, and improve simulation accuracy. Ghimire et al. [25] employed a composite LSTM model for runoff simulation, demonstrating significant performance improvements compared to a single LSTM model. Bian et al. [26] developed an integrated LSTM–LightGBM model for runoff prediction in the Shiyang River, showing that the integrated model outperformed the single model. We believe that data-driven models should incorporate the physical mechanisms in process-driven models to ensure scientific rigor. Therefore, an approach for estimating runoff was proposed in this study, which combines SWAT and AI techniques. Subsequently, this approach was verified against observations.

The process-driven SWAT model requires partitioning the watershed into hydrological response units to examine the interrelationships among the watershed’s digital elevation model (DEM), soil types, and land use. In contrast, data-driven models establish correlations between input variables and output outcomes by analyzing historical data. This study focused on the Fenhe River basin in the middle reaches of the Yellow River in China, an area significantly impacted by human activities. This study aimed to evaluate the efficacy of data-driven machine learning techniques compared to the process-driven SWAT model with the following three objectives: (1) to evaluate model performance across various temporal scales; (2) to compare and assess the effectiveness of the SWAT model versus machine learning techniques; and (3) to investigate strategies for integrating models to enhance their generalization capability and resilience.

2. Study Area and Methodology

2.1. Study Area

As illustrated in Figure 1, the Fenhe River basin is situated between 110° E–113° E and 35° N–39° N, within the middle reaches of the Yellow River basin in China, covering approximately 39,500 km². The basin [27] is primarily located in central and northern Shanxi Province, flowing through key cities such as Taiyuan. The geomorphology of the basin comprises mountains, hills, and plains, with higher elevations in the north and lower elevations in the south. The average annual temperature ranges from 8 °C to 14 °C, and average annual precipitation ranges from 400 mm to 550 mm, with significant inter-annual variability. Precipitation is unevenly distributed throughout the year, with around 60% occurring during the flood season, rendering the region vulnerable to drought and flooding [28].

Figure 1 illustrates the basin and the locations of hydrological station, with corresponding details provided in

Table 1. Rainfall and streamflow gauge stations.

Basin	Station	Name	Type	Longitude	Latitude
The Fenhe River	Xinjiang	RF1	Rainfall	111.2° E	35.6° N
	Linfen	RF2	Rainfall	111.5° E	36.2° N
	Huozhou	RF3	Rainfall	111.7° E	36.5° N
	Wenyuhe Reservoir	RF4	Rainfall	112.0° E	37.5° N
	Taiyuan	RF5	Rainfall	112.5° E	37.8° N
	Jinzhong	RF6	Rainfall	112.7° E	37.7° N
	Fenhe Reservoir	RF7	Rainfall	111.9° E	38.1° N
	Jingle	RF8	Rainfall	111.9° E	38.3° N
	Hejin	SF1	Streamflow	110.8° E	35.5° N

. Streamflow data were obtained from the hydrological yearbook, supplemented by daily water level data from the Yellow River Observation Network. The collected streamflow data were sufficient to support predictive modeling. Meteorological data were sourced from the Daily Meteorological Dataset of Basic Meteorological Elements of China National Surface Weather Stations (V3.0), released by the National Tibetan Plateau Data Center (http://data.tpdc.ac.cn (accessed on 12 July 2024)), covering the period from 1951 to 2019. Data extraction was performed using Python3.8, specifically utilizing the numpy and pandas libraries, to obtain basin-scale time series data at the daily scale based on the geographical coordinates of each hydrological station.

2.2. Process-Driven Model: The SWAT Model

The SWAT model [29] is a process-driven, semi-distributed, and time-continuous hydrological model. It was created by the Agricultural Research Service of the United States Department of Agriculture. The operating mechanism of the SWAT model involves sub-basin delineation, analysis of hydrologic response units (HRUs), input of meteorological data, and the initial model simulation. In this study, ArcSWAT version 2012, revision 664, was used for streamflow simulation.

The SWAT model’s input dataset requires a digital elevation model (DEM), land use/land cover (LULC), soil types, and meteorological data before simulating streamflow. The DEM data with a spatial resolution of 30 m, sourced from the shuttle radar topographic mission (SRTM), were acquired from the Geospatial Data Cloud (https://www.gscloud.cn/ (accessed on 12 July 2024)). The LULC data were acquired from GlobelLand30, a global land-cover dataset with a 30 m resolution developed by China and revised up to 2020. This dataset is accessible through the National Platform for Common Geospatial Information Services (https://www.tianditu.gov.cn (accessed on 12 July 2024)). Soil type data were acquired from the Harmonized World Soil Database (HWSD), a dataset created by the Food and Agriculture Organization (FAO) (https://www.fao.org/soils-portal/data-hub/soil-maps-and-databases/harmonized-world-soil-database-v12/en/ (accessed on 12 July 2024)) of the United Nations and other collaborators. The dataset utilized in this study was acquired from the Nanjing Soil Institute of the Second National Land Survey, which has a resolution of 1:5,000,000. Meteorological data were obtained from the China Meteorological Assimilation Driven Dataset (CMADS) [30]. The dataset comprises daily 24 h cumulative precipitation, maximum and minimum temperatures, average solar radiation, average wind speed, and average relative humidity.

As recommended by relevant studies, the present investigation employed thresholds of 10% for land, soil, and slope criteria. Following the successful overlaying of soil, slope, and land-use/land-cover (LULC) datasets, the HRUs were produced by the SWAT model for the Fenhe River basin (Figure 2). The generation of the river network and the division into sub-basins were based on DEM data, leading to the identification of 25 sub-basins comprising 1221 HRUs. Further, surface runoff was computed using the Soil Conservation Service curve number (SCS-CN) method, while evapotranspiration (E) was calculated using the Penman–Monteith (PM) equation. The hydrological cycle processes were calculated by employing the water balance equation:

$$\mathrm{S}{\mathrm{W}}_{\mathrm{t}}=\mathrm{S}{\mathrm{W}}_0+{\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{t}}\left({\mathrm{R}}_{\mathrm{day}}-{\mathrm{Q}}_{\mathrm{surf}}-\mathrm{E}-{\mathrm{W}}_{\mathrm{seep}}-{\mathrm{Q}}_{\mathrm{gw}}\right)$$

(1)

where SW_t and SW₀ denote the concluding and initial states of soil moisture, respectively; R_day signifies the daily precipitation; Q_Surf represents surface runoff; E denotes the daily evapotranspiration; W_seep indicates water infiltrated into the soil; and Q_gw refers to the groundwater discharge. The measurements for all the aforementioned variables are expressed in millimeters (mm).

Depending on the initial setting range, the SWAT model can simulate the runoff process across various scales. Initial simulation outcomes are stored and optimized using the sequential uncertainty fitting version 2 (SUFI-2) algorithm in the SWAT Calibration and Uncertainty Program (SWAT-CUP) [31,32]. This study aimed to model changes in runoff at daily and monthly scales to comprehensively understand the hydrological processes within the watershed. To mitigate the impact of initial conditions and ensure stable SWAT performance, the first year was used as a warm-up period. The study period for the Fenhe River basin included a calibration phase from 2008 to 2015 (with a modeling warm-up phase in 2008) and a validation phase from 2016 to 2018. This time frame ensured data continuity and consistency. In addition to data quality, the accuracy of the SWAT output depends on the careful selection of sensitive parameters. In this study, sensitive parameters for calibrating and validating the SWAT model were selected from an extensive review of the literature [33,34]. The most sensitive parameters were selected by conducting global sensitivity analyses in SWAT-CUP. The initial values for selected parameters were based on practical intervals for each parameter as suggested in the official SWAT documentation.

2.3. Data-Driven Models: Machine Learning Techniques

The SWAT model is constructed based on the hydrologic cycle process, in contrast to ML techniques that rely on input hydrometeorological datasets. In this study, ML techniques and the SWAT model were employed to simulate runoff in the Fenhe River basin to compare their performance. ML techniques mainly have two architectures: traditional ML techniques (such as XGBoost and random forest) and DL techniques. The traditional ML techniques are well-suited for hydrological modeling. Non-linear hydrological problems based on time series, such as rainfall–runoff modeling and river stage–discharge modeling, can be effectively addressed using these models. DL techniques (like LSTM, BILSTM, and GRU) are RNN architectures used extensively in time series forecasting, demonstrating promise in capturing temporal dependencies in hydrological data. The models used in this study are briefly described as follows.

As illustrated in Figure 3, the LSTM network [19] represents a variation of the RNNs architecture, addressing the limitation of traditional RNNs in capturing long-term dependencies. The data are processed by neuron loops within the hidden layer to examine long-distance time-series data, integrating three essential gating mechanisms. These gates control information storage and transmission, effectively addressing the issues of gradient vanishing and gradient explosion found in traditional RNN models.

The LSTM memory cell takes in various information through three specific gates. The first gate is called the forget gate, controlling the elements of the internal state C_t−1 that be forgotten:

f_t = σ(W_f·[h_t−1,x_t] + b_f),

(2)

where f_t is the output vector of the logistic sigmoid function (σ) layer that restricts the calculation to (0,1), indicating the forgotten degree; and W_f; and b_f define the set of trainable parameters for the forget gate.

The next gate is called the input gate, the value of which should be updated:

i_t = σ(W_i·[h_t−1,x_t] + b_i),

(3)

where i_t is the output variable restricting the calculation to (0,1), and W_i and b_i are trainable parameters.

Then the potential vector of cell state is computed by x_t and the last hidden state h_t−1:

$$\tilde{C}=\tanh \left(W_c·\left[{\mathrm{h}}_{t-1},x_t\right]+b_c\right),$$

(4)

where C is a vector restricting the calculation to (0,1), tanh is the hyperbolic tangent, and W_c and b_c are the trainable parameters.

After the update, the previous internal state C_t−1 and the forgotten gate output f_t dot-multiply operation and the internal state $\tilde{C}$:

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

(5)

Additionally, the output gate decides what shall be output by a sigmoid layer:

O_t = σ(W_o·[h_t−1,x_t] + b_o),

(6)

where O_t is the vector restricting the calculation to (0,1). W_o and b_o are trainable parameters defined for the output gate.

Finally, through the fully connected layer, the new hidden state h_t is then calculated by combining Equations (5) and (6):

h_t = O_t × tanh(C_t),

(7)

As illustrated in Figure 4, the gated recurrent unit (GRU) is a modification of the LSTM neural network, characterized by the following two key components: “update gate” and “reset gate”. This design simplifies the model’s complexity while maintaining substantial performance. Specifically, the “update gate” determines the amount of information preserved, while the “reset gate” manages the extent to which prior state information impacts the present state. The GRU model is characterized by a reduced number of parameters, resulting in higher efficiency and a lower risk of overfitting compared to the LSTM model. The bidirectional long short-term memory (BILSTM) model comprises a forward LSTM and a reverse LSTM. The forward LSTM incorporates past information, while the reverse LSTM integrates future information, both processing the input data at time t. This design allows the model to leverage information from both time t − 1 and time t + 1. The advantage lies in the structure of BILSTM neurons, which enables both layers to consider information from before and after, resulting in more accurate model outcomes.

Additionally, this study employed traditional ML techniques in conjunction with DL techniques. The extreme gradient boosting (XGBoost) algorithm, based on gradient boosting decision trees (GBDT), enhances computational efficiency in model computations. The algorithm incrementally constructs an additive model through iterative procedures. In each iteration, a decision tree model is constructed to minimize the model error. The random forest algorithm is developed by aggregating numerous models. This method generates several distinct decision tree models by employing resampling techniques on the initial training dataset. The dataset is used as the test set for the random forest algorithm. The final prediction is determined by either averaging the predictions of the individual trees or by using majority voting.

The hydro-meteorological dataset was pre-processed by cleaning raw data before being input into the model to ensure data correctness and completeness. Interpolation and imputation techniques were combined to handle missing data. For continuous variables, linear interpolation was conducted to fill in missing values, while for categorical variables, the mode imputation method was utilized. Furthermore, due to varying data scales, direct input into the model results in significant training errors. Therefore, the data were standardized to facilitate rapid convergence. Additionally, the sliding window approach was employed to generate sequence samples and calculate the lag step of the input data, enabling to learn historical knowledge and time-dependent correlations.

Prior to running the model, the dataset must undergo feature screening. Feature screening eliminates weakly correlated features to reduce dimensionality and selects strongly correlated features with the target variables to improve model generalization performance. The rainfall data input into the neural network was determined by computing the correlation between the current flow rate and the accumulated rainfall of the preceding period. Recent precipitation refers to precipitation within three days, whereas historical rainfall refers to precipitation events occurring over three days or several weeks. In this study, the basin’s recent cumulative precipitation from 1 to 3 days, as well as historical cumulative precipitation from 6 to 15 days and 30 to 120 days, was estimated and analyzed in relation to current flow. The correlation coefficient between recent cumulative precipitation and current flow was less than 0.2, whereas historical cumulative precipitation showed a high correlation with current flow. To confirm the consistency of the input nodes in the neural network model, the 1–6 day, 9 day, 12 day, 15 day, and 30 day historical cumulative rainfall of the two basins were selected.

As illustrated in Figure 5, the runoff input through autocorrelation and partial autocorrelation analysis of runoff data shows that the correlation coefficient of 7-day historical runoff was less than 0.6. The correlation coefficient had an absolute value of less than 0.2, indicating that the Fenhe River basin was in an area of intensive human activity, disrupting the natural runoff. To establish consistency in the input nodes of the neural network model, the models used 1–3 days of historical runoff. Furthermore, temperature exhibits seasonal variations, impacting the production and sink flow processes, which were fed into the neural network.

Neural networks contain many hyperparameters, with their values set prior to the learning process. The study period for the DL techniques spanned from 2008 to 2015 for training and from 2016 to 2018 for testing. The model parameters underwent training through hyperparameter optimization, a method aimed at determining a set of hyperparameters that minimize the model’s loss function on the specific dataset. In this study, the mean-square-error (MSE) was used as the loss function for the hyperparameter optimization process.

Common hyperparameters include learning rate, training epochs, and the dimensionality of the output space, among others [19]. The learning rate indicates the magnitude of steps taken in the gradient descent algorithm. In this study, the Adam optimization algorithm, a variant of stochastic gradient descent known for its efficiency, was used. Moreover, an epoch, commonly defined as a single iteration over a complete dataset within a neural network, serves to segment the training process into distinct stages. Prolonged training may result in overfitting, where the model learns patterns specific to the training dataset. Therefore, the initial setting for training epochs was established at 200, and the optimal Nash–Sutcliffe efficiency (NSE) was achieved after 50 epochs. Consequently, a total of 50 epochs were selected for the final training phase of the model. Additionally, other hyperparameters such as the dimensionality of the output space were fine-tuned using the grid search method in this study.

2.4. Integrated Modeling Approach

The integrated model approach [35] creates a richer feature space than a single model by using the prediction results of multiple base learners from the first stage as features. These features, combined with the labels of original samples, form a new dataset. Five-fold cross-validation ensures that each base model maintains similar prediction accuracy while demonstrating different learning processes. The overall model’s prediction performance can be improved through the secondary learners’ further learning. This integration strategy helps reduce the potential bias and variance of a single model, thereby enhancing the stability and generalization ability of the entire model system.

The monthly and daily runoff combination model of the Fenhe River basin is illustrated in Figure 6. The initial stage of the integrated model approach is based on the principles of various types of base models and selection of the most effective models. In contrast to the conventional stacking integration model, this approach aims to merge the outcomes of the SWAT model into the novel dataset formed by the predictive results of the base model. To enhance the generalization performance of the model, a decision is made to opt for a simpler linear regression model for the sub-model, and subsequently for the daily runoff combination model.

2.5. Evaluation Metrics

In this study, the model simulation was evaluated using the coefficient of determination (R²), root mean square error (RMSE), and Nash–Sutcliffe model efficiency coefficient (NSE):

$${\mathrm{R}}^2=\frac{({\sum }_{t=1}^n\left(Q_{o,t}-{\overline{Q}}_o\right)\left(Q_{p,t}-{\overline{Q}}_p\right){)}^2}{{\sum }_{t=1}^n{\left(Q_{o,t}-{\overline{Q}}_o\right)}^2{\sum }_{t=1}^n(Q_{p,t}-{\overline{Q}}_p{)}^2}$$

(8)

$$\mathrm{RMSE}=\sqrt{\frac{1}{n}{\sum }_{t=1}^n{\left(Q_{o,t}-Q_{p,t}\right)}^2}$$

(9)

$$\mathrm{NSE}=1-\frac{{\sum }_{t=1}^n{\left(Q_{o,t}-Q_{p,t}\right)}^2}{{\sum }_{t=1}^n{\left(Q_{o,t}-{\overline{Q}}_o\right)}^2}$$

(10)

where ${\overline{Q}}_o$ is the mean value of observed flow for the time period, ${\overline{Q}}_p$ is the mean value of predicted flow in time period, $Q_{o,t}$ is the observed value of flow at time t, $Q_{\mathrm{p},\mathrm{t}}$ is the predicted flow value at time t, and n denotes the length of the runoff sequence.

3. Results

3.1. Estimation of Streamflow Using SWAT Model

As shown in

Table 2. Performance of the SWAT model at the monthly and daily scales.

Time Scale	Period	Observation/(m3·s−1)	Simulation/(m3·s−1)	RMSE/(m3·s−1)	R2	NSE
Monthly	Calibration period	18.65	16.07	7.51	0.82	0.80
	Validation period	28.61	27.02	10.88	0.80	0.79
Daily	Calibration period	18.67	15.56	9.67	0.83	0.80
	Validation period	28.70	30.90	16.97	0.80	0.79

and Figure 7, the monthly and daily-scale runoff simulations conducted in the Fenhe River basin using the SWAT model exhibit a satisfactory performance, with R² greater than 0.80 and NSE exceeding 0.79 throughout the validation period. Nevertheless, the accuracy of monthly and daily-scale models decreased during the validation period compared to the calibration period. This decline can be attributed to the model parameters being set according to the calibration period, while changes in climate and subsurface conditions occurred during the validation period, leading to discrepancies. The rationale behind this phenomenon lies in the fixed nature of parameters in the model validation period, which were established at regular intervals, whereas climate and subsurface conditions change throughout the validation period, leading to a misalignment where the parameters set at regular intervals are not promptly adjusted to account for these variations. Additionally, the temporal scale of simulation in the monthly runoff model exceeds that of the daily runoff model. Consequently, in theory, the runoff value accumulates as the simulation scale extends over time. The scatter plot illustrating the results of the monthly runoff validation period shows a dispersion of data points on either side of the linear regression line. Furthermore, there were fewer data points aligning with the linear regression line, suggesting a discrepancy between the simulated and observed values. The scatter plot analysis of daily runoff indicated that the Hejin station exhibited a stronger linear relationship when the runoff was below 200 m³/s. However, when the runoff exceeded 200 m³/s, the data points were dispersed on either side of the linear regression line, with a concentration toward the upper left quadrant. This distribution suggested that the simulated values surpassed the measured values in the validation period.

3.2. Estimation of Streamflow Using ML Techniques

As illustrated in Figure 8, a horizontal comparative analysis was conducted using ML techniques, and comparison was executed against the SWAT models. The model performance for monthly runoff at Hejin station was evaluated for training and testing periods. The ‘Training period’ and ‘Testing period’ in this figure correspond to the ‘calibration period’ and ‘validation period’ mentioned in the text, respectively. The data-driven ML techniques demonstrated enhanced performance compared to the process-driven SWAT model in both periods. Among the models considered, the GRU model exhibited superior performance, achieving RMSE, R², and NSE values of 4.17 m³/s, 0.952, and 0.933, respectively. During the testing period, the traditional ML techniques (XGBoost and random forest models) demonstrated performance levels closer to the SWAT model, as indicated by RMSE, R², and NSE values ranging from 12.30 to 12.65 m³/s, from 0.814 to 0.817, and from 0.729 to 0.744, respectively. In contrast, the DL techniques (LSTM, BILSTM, and GRU models) demonstrated significantly superior performance compared to the aforementioned models.

The comparison of the SWAT model with ML techniques against the observed flow during the testing period is presented in Figure 9. The GRU model was the most effective, indicating that DL techniques can enhance simulation performance. Moreover, periodic model simulation results showed increased stability and improved accuracy during the training period, but a decline was observed during the testing period. The primary reasons for the decline in simulation results during testing could be the complexity of the models, including their architecture and hyperparameters.

As illustrated in Figure 9, the data-driven ML techniques demonstrated superior performance compared to the process-driven SWAT model in both periods. The linear fit plot illustrates the simulation results during the testing period. At flow rates below 40 m³/s, there was a close agreement between the simulated and measured values of the SWAT model and the traditional ML techniques. However, for flow rates surpassing 40 m³/s, the DL techniques demonstrated superior performance compared to the SWAT and traditional ML techniques. Furthermore, under these conditions of increased flow, there was a notable similarity between the simulated values and those obtained through measurement. This indicated that the DL techniques outperformed the SWAT and traditional ML techniques, particularly for larger flows. However, they tended to underestimate peak flows.

To advance the investigation of runoff evolution across various scales and enhance the simulation of the daily runoff process, this study conducted a comparative analysis of ML techniques horizontally and SWAT models vertically. In comparison to monthly-scale runoff simulations, daily-scale simulations offer extended time series data inputs that can provide the model with ample information. In terms of simulation capability, runoff series generated by DL techniques compared favorably against traditional ML techniques and the SWAT model in the basin, as shown in Figure 10. Among the models considered, the BILSTM model demonstrated the lowest RMSE and the highest R² and NSE (7.15 m³/s, 0.878, and 0.878, respectively) throughout the entire testing period compared to the traditional ML techniques and the SWAT model. The results show a resemblance in the simulation outcomes between daily and monthly scales during both training and testing periods. During the testing period, the traditional ML techniques exhibited comparable performance to the SWAT model, with RMSE, R², and NSE values ranging from 17.55 to 19.43 m³/s, from 0.783 to 0.820, and from 0.740 to 0.788, respectively. The DL techniques exhibit significantly superior performance compared to other models, demonstrating the highest level of performance. This indicates that DL techniques have the capability to improve the efficacy of model simulations.

As illustrated in Figure 11, the data-driven ML techniques demonstrate superior performance compared to the process-driven SWAT model in both periods. To enhance clarity, a linear regression plots was used to demonstrate the simulation outcomes during the testing period. The ML techniques consistently demonstrate superior performance compared to the SWAT model during the testing period. For flow rates below 200 m³/s, both SWAT and ML techniques exhibit comparable results in terms of simulated and measured values. For flow rates surpassing 200 m³/s, the DL techniques demonstrate superior performance compared to the SWAT model and ML techniques, with simulated values closely aligning with the observed values. The SWAT model calculated the sub-basin yield–sink process by partitioning HRUs. The total runoff is calculated by aggregating the sub-basins, with each sub-basin being determined using data from rainfall stations located near the centroid. This characteristic makes the model highly responsive to the spatial arrangement of rainfall stations. The lack of a clear correlation between the sub-basin outlet station for runoff calibration and nearby rainfall stations can result in challenges during model calibration errors in simulated runoff. The utilization of a data-driven model can mitigate the impact of erroneous data from a specific rainfall station on the model’s output and diminish the model’s susceptibility to rainfall data by incorporating extensive time series data, including historical runoff and precipitation records.

3.3. Integration of Data-Driven and Process-Driven Models for Streamflow Prediction

Table 3. Performance of integrated model approach at monthly scale.

Model	Training Period			Testing Period
	RMSE	R2	NSE	RMSE	R2	NSE
Single optimal model	4.17	0.952	0.933	8.53	0.941	0.877
Integrated model approach	3.50	0.953	0.953	6.78	0.937	0.922

displays the simulation results and evaluation indices of monthly runoff at Hejin station using the integrated model approach. The analysis indicates that the monthly runoff simulation results of the integrated modeling approach show a higher degree of proximity to the observed values, generally outperforming the predictions generated by individual models like XGBoost, random forest, and SWAT. During the testing period, the RMSE, R², and NSE of the monthly runoff at Hejin station are 6.78 m³/s, 0.937, and 0.922, respectively. In comparison to the single optimal model (GRU), RMSE reduces by 1.75 m³/s, and NSE improves by 0.045.

The simulation assessment metrics for daily runoff at Hejin station, utilizing the integrated model approach, are presented in

Table 4. Performance of integrated model approach at daily scale.

Model	Training Period			Testing Period
	RMSE	R2	NSE	RMSE	R2	NSE
Single optimal model	7.15	0.878	0.878	12.40	0.901	0.894
Integrated model approach	6.49	0.899	0.899	11.70	0.907	0.906

. The analysis indicates that the evaluation indices for daily runoff at Hejin station, specifically RMSE, R², and NSE, are 11.70 m³/s, 0.907, and 0.906, respectively. RMSE decreases by 0.70 m³/s, and R² and NSE improve by 0.006 and 0.012, in comparison to the single optimal model (BILSTM).

4. Discussion

4.1. Analysis of SWAT Model Sensitivity Parameters

The SWAT model was calibrated using the SUFI-2 algorithm within SWAT-CUP. In addition to data quality, the accuracy of the output depended on the careful selection of sensitive parameters [31]. In this study, the initial values for these selected parameters were determined based on physically feasible intervals recommended in the official SWAT documents [32] and extensive reviews of the literature [34]. A set of 16 flow-related parameters was chosen for global sensitivity analysis, and 200 iterations were performed to obtain the sensitivity parameter analysis results. Global sensitivity analyses in SWAT-CUP were employed to identify the parameters with the highest sensitivity. Insensitive parameters were identified and excluded by sorting t-stat and p-values. The t-stat indicates the sensitivity of a parameter, with a higher absolute t-stat suggesting greater sensitivity. The p-value indicates the significance of the t-stat; a smaller p-value denotes a lower probability of the parameter being identified as sensitive by random chance. The prefix “V_” denotes the replacement of the initial parameter value, while the prefix “R_” signifies the relative change of the initial parameter value.

As shown in

Table 5. Sensitivity analysis for runoff parameters.

Parameter	Description	t-Stat	p-Value	Low	High	Adjusted Value
R_CN2.mgt	SCS runoff curve number	−8.54	0.00	−0.5	0.5	−0.34
R_SOL_Z.sol	Depth from soil surface to subsoil/(mm)	3.11	0.00	−0.5	0.5	−0.34
R_SOL_BD.sol	Moist bulk density/(mg·m−3)	−2.92	0.00	−0.5	0.5	0.49
V_ALPHA_BF.gw	Base flow alpha factor/d	−2.80	0.01	0	1	0.84
V_CANMX.hru	Maximum canopy retention/(mm)	2.62	0.01	0	100	38.25
V_SURLAG.bsn	Surface runoff lag time	1.64	0.10	1	24	11.51
V_TLAPS.sub	Temperature lapse rate/(°C/km)	−1.61	0.11	0	50	44.44
V_OV_N.hru	Manning’s N	−1.33	0.19	0	0.8	0.37
V_GW_REVAP.gw	Groundwater ‘revap’ coefficient	1.19	0.24	0.02	0.2	0.11
R_SOL_AWC.sol	Available water capacity of the soil layer/(mm)	−1.16	0.25	−0.5	0.5	0.08
V_RCHRG_DP.gw	Deep aquifer percolation fraction	−0.94	0.35	0	1	0.03
V_CN_K2.rte	Effective hydraulic conductivity in main channel alluvium/(mm·h−1)	−0.72	0.47	0	150	100.48
V_SFTMP.bsn	Snowfall temperature/(°C)	−0.71	0.48	−5	5	−3.55
V_SMTMP.bsn	Snow melt base temperature/(°C)	−0.63	0.53	−5	5	−1.22
V_ESCO.hru	Soil evaporation compensation factor	0.60	0.55	0.01	1	0.80
V_EPCO.hru	Plant uptake compensation factor	0.60	0.55	0.01	1	0.30

, the CN2 (the SCS runoff curve number) decreased by 34% relative to the initial value, indicating increased soil infiltration and decreased surface runoff. Soil characteristics are essential for both infiltration and surface runoff. Hence, the model demonstrated sensitivity to SOL_Z (the depth from soil surface to subsoil), SOL_AWC (available water capacity of the soil layer), and SOL_K (the saturated hydraulic conductivity). ALPHA_BF (the base flow alpha factor) was adjusted to 0.84, implying that the subsurface geological structure in the region increases the sensitivity of runoff to precipitation. The adjustment of RCHRG_DP (the deep aquifer percolation ratio) to 0.03 was associated with the favorable geological structure that facilitates the extended storage of groundwater. GW_REVAP (the groundwater “revap” coefficient) was adjusted to 0.11, indicating that SFTMP (the snowfall temperature), SMTMP (the snow melt base temperature), and TLAPS (the temperature lapse rate) were significant parameters in the temperature change model. ESCO (the soil evapotranspiration compensation factor) was adjusted to 0.80, allowing the model to accurately account for the increased evapotranspiration demand of the surface soil. EPCO (the plant uptake compensation factor) was adjusted to 0.30, leading to an increase in evapotranspiration generated by the model. According to Abbaspour et al. [36], variations in parameters like CN2, ESCO, and EPCO impact the simulation of peak flow.

The SWAT model demonstrates effectiveness in simulating water–sediment dynamics, runoff patterns, and pollution levels in various basins, including the Tangnaihe basin, a key watershed of the Yellow River, the Han River basin, a tributary of the Yangtze River, and the Huaihe River Basin in China. The utilization of the SWAT model in arid and semi-arid regions is comparatively limited, and there is also a lack of comprehensive discussion on the model’s parameterization algorithm and uncertainty analysis. To address this issue, Zuo et al. [34] effectively developed a distributed hydrological model, SWAT, in the Weihe River basin. The SUFI-2 algorithm was employed to perform sensitivity analysis and uncertainty assessment of model parameters. Subsequently, the model accurately replicated the monthly runoff dynamics in the Weihe River basin for the period spanning from 1961 to 2008. The study demonstrates that hydrologically significant parameters, such as CN2 and ALPHA_BF, exhibit notable sensitivity to the runoff process. Given the geographic location and hydro-meteorological conditions similarities between the Fenhe River basin and the Wei River Basin, the Fenhe River basin was chosen as the focal area for investigation in this study. The Fenhe River basin, situated in the central region of Shanxi Province, serves as a crucial water reservoir supporting various local activities such as agricultural irrigation, industrial production, and residential needs. Simultaneously, the basin serves as a crucial region for economic advancement in Shanxi Province, accommodating numerous urban centers and industrial zones. The sensitivity analysis revealed that the parameters CN2, SOL_K, SOL_BD, and ALPHA_BF exerted a substantial impact on the runoff process. This finding offers a crucial scientific foundation for hydrological simulation and water resource management in the Fenhe River basin.

4.2. Analysis of Factors Influencing Runoff Modeling

Both the process-driven SWAT model and data-driven ML techniques were employed. Generally, DL techniques demonstrated superior performance compared to the SWAT model at both monthly and daily scales. Validation was performed using runoff data obtained from Hejin station that serves as the monitoring station for basin outlet. Meteorological data exert a substantial influence as they significantly impact the basin’s runoff due to climatic conditions. Although the model was trained using nine meteorological stations, the final output was obtained through the weighted sum of input features. Relying solely on meteorological station data may not provide a comprehensive representation of the climatic conditions within the basin, thereby constraining the neural network model’s capacity to accurately replicate the runoff process.

Since the 1990s, there has been a significant decrease in the runoff of the primary tributaries on the Loess Plateau compared to the levels observed in the 1950s and 1960s. This decline can be attributed to the extensive implementation of water conservation and protection projects, as well as the increasing aridification of the basin [4]. The runoff of the basin has exhibited a gradual increase since the 21st century, as illustrated in Figure 12. This phenomenon can be attributed to heightened precipitation levels, the execution of the Fenhe River re-streaming project, and the ecological recharge facilitated by the Wanjiazhai Yellow River Diversion and Transfer Project. Intensive human activities, driven by rapid urbanization and industrialization in the Fenhe River basin, have significantly impacted the basin’s hydrology. This is evidenced by a sudden shift in the double cumulative curve of rainfall runoff observed in 2009.

Regarding extreme flows, both the SWAT model and traditional ML techniques tended to underestimate peak flows. The SWAT model simulates the surface runoff process by dividing HRUs. However, when simulating instantaneous peak flow, the complexity and uncertainty of its physical process simulation may lead to underestimated peak flow. Although ML techniques can learn the complex relationship between inputs and outputs through large datasets, they may not accurately capture dynamic changes in peak flow during extreme events (such as flood peaks) due to sparse data or limitations in the model structure.

4.3. Combined Modeling Approach for Improving Simulation Accuracy

Regardless of whether derived from a process-driven SWAT model or a data-driven machine learning model, achieving a comprehensive understanding of hydrological components through a single-model simulation is often challenging [37,38,39]. Extensive comparisons between these two approaches have been documented in the literature [40,41], with several studies demonstrating superior performance of ML techniques over SWAT models in hydrological process modeling [42,43,44]. This study demonstrates that integrating the SWAT model with ML techniques in hybrid models could enhance the performance of rainfall–runoff modeling. Senent Aparicio et al. [45] demonstrated that the integrated method was crucial for improving the simulation accuracy of instantaneous peak flow (IPF) in the Ladra River basin. The integrated modeling approach aimed to mitigate the limitations of individual models and enhance simulation accuracy. The primary procedure involved selecting multiple base models for pre-training in the initial layer, followed by amalgamating their outputs to form a novel dataset used as input for the subsequent layer of the model. The results indicate that the integrated model demonstrates higher accuracy compared to the best individual model for monthly and daily runoff in basins. This finding suggests that non-linear weight assignment significantly enhances the simulation accuracy of the combined model in the integrated modeling approach.

5. Conclusions

In this study, the process-driven SWAT model and data-driven machine learning (ML) techniques were employed to simulate runoff at both monthly and daily scales. The proposed models were implemented in the Fenhe River basin located in the middle reaches of the Yellow River. The performance of the different models was compared, and the process-driven SWAT model and data-driven machine learning techniques were combined to achieve improved runoff simulations. The main conclusions are summarized as follows:

(1)

The process-driven SWAT model demonstrated excellent performance in simulating runoff across various scales, achieving R² values greater than 0.80 and NSE values greater than 0.79. Sensitivity analysis highlighted CN2 (the SCS runoff curve number) as the most sensitive parameter, providing comprehensive insights into subsurface characteristics that directly impact runoff volume. Higher CN2 values indicate increased impermeability, leading to greater runoff. Parameters like SOL_K, SOL_BD, and ALPHA_BF also showed significant sensitivity, influencing soil infiltration and surface runoff dynamics.

(2)

The basin’s hydro-meteorological dataset served as input for ML techniques, facilitating a comparison with the process-driven SWAT model. Deep learning (DL) techniques outperformed the SWAT model in simulating monthly and daily runoff, whereas traditional ML techniques exhibited results comparable to the SWAT model.

(3)

To address the limitations of individual models, this study integrated the SWAT model with ML techniques using the integrated modeling approach. With the non-linear weight allocation method, the integrated model achieved improved simulation accuracy across various scales. The results underscored the importance of non-linear weight allocation in enhancing the performance of combined process-driven and data-driven models, presenting innovative avenues for integrated hydrological modeling.

While the SWAT model and ML techniques have demonstrated proficiency in simulating runoff with notable fluctuations, they may not effectively represent peak flow characteristics. Therefore, future studies should focus on investigating methods to enhance DL techniques for more accurate estimation of peak flow.