An improved approach for rainfall-runoff simulation under anthropogenic influence

Human activities significantly influence the rainfall-runoff process, leading to pronounced fluctuations in runoff data and presenting substantial challenges for accurate simulation. In this study, the Water Allocation and Simulation (WAS) model was adopted and enhanced to simulate rainfall-runoff under the impact of human activities. The virtual reservoirs were established within the WAS model to effectively represent water consumption associated with human activities, including agricultural irrigation and pond storage. The model was compared with the results of rainfall-runoff simulation under natural conditions. This study method was applied to simulate rainfall-runoff in the Yanglou hydrological station basin, Northern Anhui Province, China, in which the water cycle is strongly influenced by human activities. The results show that the simulation accuracy of the WAS model is superior to that of rainfall-runoff simulation under natural conditions. The WAS model significantly improved the Nash-Sutcliffe efficiency coefficient (NSE) from 0.704 to 0.898 and the correlation coefficient (R 2) improved from 0.869 to 0.952 during the validation period. The WAS model can be practiced in areas with strong human activities. The findings of this study are instrumental in enhancing rainfall-runoff simulation in regions of strong human activities.

Water resources play a vital role in supporting the sustainable development of regional societies and economies. With the continuous expansion of human water use, complex interactions have emerged among economic growth, social activities, and regional water systems, characterized by mutual influence, constraints, and interdependence [1, 2]. In recent decades, human activities have significantly altered rainfall-runoff processes, leading to considerable fluctuations in runoff data [3]. This is particularly evident in plain areas with high population density and numerous ponds, where runoff during normal flow years tends to decrease, while high-flow years exhibit a marked increase in runoff. Conventional hydrological models, such as the Xin’anjiang model [4], often fail to accurately simulate rainfall-runoff processes under anthropogenic influences. Long-term monitoring has confirmed that these hydrological alterations are primarily attributable to intensive human interventions, including water conservancy projects, land-use changes, and modifications of the underlying surface. Consequently, accurately simulating rainfall-runoff processes under anthropogenic influences remains a pivotal challenge in modern hydrology.

Over the last half-century, a common approach has been the development of process-driven hydrological models [5]. These models use parametric methods to numerically solve the governing physical equations that describe the spatial and temporal evolution of rainfall-runoff processes. Currently, established process-driven hydrological models fall into two main categories: lumped conceptual models and physics-based distributed models. Lumped conceptual models, such as the Tank model, Sacramento model, and Xin’anjiang model [6–8], simulate catchment behavior using simplified, integrated representations of hydrological processes. In contrast, physics-based distributed models, such as the Système Hydrologique Europeen (SHE) model, the Topography-based hydrological MODEL (TOPMODEL), and the Soil and Water Assessment Tool (SWAT) model [9, 10, 24]. With the development of science and technology, many high-resolution sophisticated input data (such as digital elevation model (DEM), land use and soil maps, and soil characteristics) can be obtained. Therefore, these models can simulate the hydrological generation process more accurately.

To quantitatively analyze the interaction between the natural water cycle and human activities, researchers have developed various coupled human-hydrological systems models, building upon traditional hydrological modeling frameworks [11]. For example, the SWAT model has been integrated with the Positive Multi-Attribute Utility Programming (PMAUP) model to simulate the coupled dynamics and bidirectional feedback between socio-economic and eco-hydrological systems [12]. From the perspective of social hydrology, the SHE model is used to simulate the potential evolutionary path of the social hydrological system of the Han River in China [13]. The positive and negative feedback interactions between human and natural systems in the region are captured by adding a social module to the LASCAM model [14]. However, in terms of water resources system simulation and allocation model methods, these models mostly adopt the methods of adding the water intake process to the water cycle model and coupling the water cycle model and the allocation model [15]. Because the dynamic changes in water resources cannot be responded to in real-time in the process of water resources allocation, the model is less effective in regions with intensive human activities. Therefore, further research is needed on real-time mutual feed simulation of the natural-social water cycle process. Recent advances have further highlighted the need to incorporate anthropogenic water use into hydrological modeling frameworks [16], particularly in regions with intensive agricultural irrigation and pond networks [17]. The increasing availability of remote sensing data has also enabled more accurate estimation of small water body storage for hydrological applications [18]. However, as noted by Fowler et al. [19], anthropogenic intensification of rainfall extremes, combined with modifications to surface storage, poses new challenges for flood simulation in human-impacted regions.

This study aims to develop an improved approach for rainfall-runoff simulation under strong anthropogenic influences. Based on the theory of natural-social dualistic water cycle and water resources allocation, the water allocation and simulation (WAS) model was adopted [20–22]. The WAS model employs a dynamic feedback simulation approach of water cycle separation, aggregation, modeling, and regulation. It is capable of simulating water supply processes, consumption, and drainage within the social water cycle, and captures the integrated effects of both natural and social water cycles. Furthermore, the model represents dynamic interactions between upstream and downstream water intake and usage, as well as system-wide dynamics of water resources [11]. The main contributions of this study are summarized as follows: First, a multi-level hydrological unit division method was introduced, segmenting units from macro water resource zones down to township levels, which enhances the model capability to quantify changes in river flow under human activities. Second, the virtual reservoirs are established in the WAS model, improving the representation of water consumption processes associated with anthropogenic influences. Finally, the research method was applied to simulate rainfall-runoff in the Yanglou hydrological station basin. Results demonstrate that the WAS model achieves higher simulation accuracy compared to conventional methods under natural conditions. This approach provides a valuable reference for rainfall-runoff simulation in strong human activities areas.

However, to fully contextualize the contribution of this study, a comparative review of representative hydrological models that account for anthropogenic influences is warranted. The SWAT (Soil and Water Assessment Tool) model incorporates reservoir and pond routines that simulate storage-release dynamics based on prescribed operation rules [23]. Similarly, HEC-HMS (Hydrologic Engineering Center-Hydrologic Modeling System) provides reservoir routing and diversion capabilities for flood control and water supply simulations [24]. The VIC (Variable Infiltration Capacity) model, when coupled with MODFLOW or water management modules, has been applied to assess human-water interactions at regional scales [25]. While these models have been successfully implemented in various basins, they exhibit notable limitations in representing distributed, small-scale pond storage dynamics and capturing real-time feedback between water allocation decisions and hydrological responses. Specifically, the reservoir modules in SWAT and HEC-HMS are typically designed for single, well-defined large reservoirs with known stage-storage-discharge relationships, rather than for numerous scattered ponds that collectively exert significant influence on runoff generation in plain areas with intensive human activities. Moreover, the integration of water consumption processes, such as agricultural irrigation withdrawals and domestic water use, with hydrological routing remains loosely coupled in most existing frameworks, limiting their ability to simulate the bidirectional interactions between human water use and natural water cycle [26, 27]. The WAS model addresses these gaps by introducing a virtual reservoir approach that aggregates small ponds within each computational unit based on remote sensing and field survey data, and explicitly couples water allocation rules with hydrological simulation, thereby enabling a more realistic representation of anthropogenic influences on rainfall-runoff processes.

The remainder of the paper is organized as follows. Section 2 provides a detailed introduction to the study area, and the data sources, and describes the WAS model structure, followed by model evaluation criteria. Section 3 introduces the model development and analyzes the simulation results in the study area. Section 4 presents the discussion. Section 5 summarizes the main conclusions.

Study area and data sources

Study area

The Yanglou hydrological station basin is located in Suzhou city, Northern Anhui Province, China, as shown in Figure 1. Yanglou hydrological station is an important control station of the Limin River. It is also an experimental station to explore the relationship between rainfall and runoff in a small area of the Huaibei Plain. Yanglou hydrological station was established in 1966, and the measurement items include water level (stage), discharge, evaporation, precipitation, groundwater, soil moisture, and water quality. There are eight rainfall and groundwater gauging stations within the Yanglou hydrological station basin including Yanglou, Zhengji, Huangkou, Huzhai, Xinzhuangzhai, Shaozhuang, Uotao, and Tangzhai. Huangyang River is the tributary of the Limin River, joining the Limin River in the town of Yanglou. The basin of Yanglou hydrological station is bounded by the abandoned Yellow River embankment in the north and the Longhai Railway in the south. The basin boundary is relatively clear. The total area of the Yanglou hydrological station basin is around 152 km 2.

The meteorological conditions of the Yanglou hydrological station basin are complex, and the annual distribution of precipitation is uneven. The annual average precipitation is 762.8mm, and the annual average evaporation is 921.3 mm. The Yanglou hydrological station is flat with an average slope of 0.2‰. There are five towns in the Yanglou hydrological station basin, with a total population of 142,300. The population density is 783 inhabitants/km 2. The basin has a high concentration of villages, a dense population, and intense human activities. The cultivated area of the Yanglou hydrological station basin is 12813ha. It is the main production area of wheat, corn, cotton, vegetables, and fruits, and plays an important role in the development of the local economy [28].

Data sources

Many data are required to build the WAS model, including precipitation, evaporation, runoff, land use, soil type, geological texture, water consumption data, etc. The hydrological data in this study were derived from the Suzhou Hydrology and Water Resources Bureau of Anhui Province from 1972 to 2020. The land-use data were derived from the GlobeLand30 [29] and Suzhou Statistical Yearbook. Other data were derived from the Suzhou Water Resources Bulletin and the Second Huaihe River Basin Water Resources Evaluation. The main hydrological characteristics of the Yanglou hydrological station basin are shown in Table 1.

Methods

The flowchart below illustrates the methodology of this study (Figure 2) and the following sections discuss the methods involved in the model.

Unit division principle

Unit is an important fundamental component of the water resources system, and the method of dividing the unit directly affects the simulation and regulation of the water resources system. At present, most hydrological models use DEM to extract sub-watersheds or grid units, the soil type and land use type are superimposed, and the hydrological response unit is further divided for hydrological calculation [30]. This applies to watershed or river runoff simulation. However, the management of water resources is mainly based on administrative division. The conventional hydrological modeling approach, which typically relies on natural watershed boundaries, faces challenges in accurately representing water resources allocation and hydrological processes within artificial administrative regions. This mismatch often leads to inadequate simulation of cross-administrative water transfers and human-induced hydrological variability. To address this limitation, the WAS model introduces a three-level unit division method, which integrates both hydrological units and administrative boundaries [15]. In this study, based on the management and administration of water resources in a small watershed, the hydrology units are divided from water resources zoning to townships, allowing the WAS model to more accurately quantify the changes in river flow under the influence of human activities. The division method of the water resources system unit is shown in Figure 3. This three-level unit division method can not only reflect the characteristics of traditional hydrological simulation model unit division but also meet the needs of the management and administration of water resources in a small watershed. The specific steps are as follows: Basic unit formation. Formation of basic units through superposition of water resources zoning and administrative divisions. Basic units are constructed by overlaying regional water resources zoning maps with official administrative division boundaries. This integrated spatial unit captures both the hydrological context and socio-management boundaries, serving as the foundational geographic entity for further subdivision.

Computational unit refinement. Refinement into computational units using DEM. Based on the basic units, a DEM is applied to extract detailed topographic and hydrological features, such as slope, flow direction, and sub-catchment boundaries. Through spatial encryption and disaggregation, these units are further refined into higher-resolution computational units, which more accurately represent the natural drainage structure and surface runoff processes.

Hydrological unit delineation. Delineation of hydrological units by integrating soil type and land use with crop distribution patterns. The computational units are subsequently combined with soil type data and a synthesized land-use map that incorporates detailed crop distribution information. Through spatial superposition and reclassification, this integration yields hydrologically homogeneous units. Each final hydrological unit is characterized by uniform attributes in terms of soil properties, land cover, and crop type, enabling precise simulation of rainfall-runoff processes and human-induced hydrological changes.

Virtual reservoir setting principle

The reservoir is an important node in the regulation of the water resources system. With the increase in urbanization and population, the underlying surface has changed. Due to the large number of ponds in a small area, the rainfall-runoff process is altered, so it is difficult to simulate accurately. The above problem can be solved by considering the ponds in the calculation unit as virtual reservoirs [31].

In this study, virtual reservoirs were established following a systematic five-step procedure. First, all ponds were identified using satellite imagery, with all detections subsequently validated through field surveys. Second, for each pond, the storage capacity was calculated as the product of its surface area and average depth. The total virtual reservoir capacity for each hydrological unit was then obtained by aggregating the storage volumes of all ponds within that unit. Third, key parameters including total storage, dead storage, outflow coefficient, and evaporation coefficient were assigned based on field measurements and model calibration. Fourth, linkages between virtual reservoirs and irrigation demand units were established using data from the Water Resources Bulletin and interviews with local farmers, ensuring that the water supply network reflects actual local water use practices. Fifth, operational rules were defined: during the flood season, the water level of the virtual reservoir cannot exceed the flood control level; during the dry season, stored water is released as needed to meet irrigation demands. The impact of human activity on rainfall-runoff processes can be effectively simulated by setting up virtual reservoirs. The schematic diagram of the setup and role of the virtual reservoir is shown in Figure 4.

Calculation principle

In the past few years, the WAS model has been further developed, which contains a variety of functions, including regional water use and drain, power generation, reservoir operation, economic analysis, ecological flow, and so on [20, 21]. The WAS model can dynamically simulate the natural-social water cycle under natural and human influence. The WAS model consists of four components: runoff simulation module, river confluence module, reclaimed water simulation module, and water resource allocation module. In this study, the reclaimed water simulation module is not involved. In addition, the virtual reservoirs are established in the WAS model. The WAS model structural framework is shown in Figure 5.

The WAS model has five main calculation processes. Firstly, for the surface runoff producing and confluence process, the comprehensive runoff producing method is adopted for the runoff on the permeable area. Secondly, during the flow of soil water, the soil water change is divided into four parts, which are the soil flow, the water soil quantity, the amount of shallow groundwater recharged by the soil, and the soil water evaporation. Thirdly, for the evapotranspiration process, the WAS model is combined with the calculation method of the SIMHYD model [32], and the adjustment coefficient of vegetation evaporation on the underlying surface is considered to calculate the soil evaporation. Then, in the process of groundwater runoff, the linear reservoir method is used to simulate groundwater runoff. Finally, considering the objectives and period of the water resources system simulation and allocation, the surface confluence is calculated using the general SCS model linear reservoir [33].

The objective functions for the WAS model are presented subsequently. They include the two objective functions of load balance objective -minimum water shortage rate and spatial balance-optimal fairness. In this study, this objective function is used in the global multi-objective optimization scheme. The smaller the value of the overall objective function, the better the monthly allocation result can satisfy the global optimality. In fact, in practical applications, the weight values can be adjusted to reflect the focus of watershed management and planning. Lastly, the constraints of the WAS model are presented. They include constraints on water supply capacity, ecological flow, water supply quality, and water balances for rivers, lakes or reservoirs, groundwater, reclaimed water, and units. The calculation principle of the WAS model is detailed in Ref. [15].

To simulate rule-based operation and optimized operation, there are two types of solving algorithms in the WAS model [34]. One is the rule algorithm, which conducts long-term simulations according to the existing operation rules. The other is the non-dominated sorting genetic algorithm with elite strategy (ENSGA-II) [35], which is used to solve multi-objective problems. The ENSGA-II is further developed from the nondominated sorting genetic algorithm (NSGA-II). The NSGA-II is widely used in many engineering and science contexts and has been satisfactorily implemented to solve various water resource management issues [36–38]. In the process of solving the evolutionary algorithm, to avoid falling into local optimization, there is a certain probability of accepting solutions with poor quality. Therefore, the elite strategy is to use the expert experience to intervene in the mating strategy of the genetic algorithm, and select and obtain the optimal genetic seeds, to retain the excellent population for the next-generation. Compared with NSGA-II, the ENSGA-II has a shorter calculation time and faster convergence speed. So, the ENSGA-II has strong robustness and can quickly find the optimal solution.

Model evaluation criteria

In this study, the Nash-Sutcliffe efficiency coefficient (NSE) and correlation coefficient (R 2) (Equations 1, 2) are used as the evaluation standards. The NSE represents the fitting degree between the simulated values and the observed values. The NSE ranges from − ∞ to 1. The larger the NSE values, the better the model performance [39]. The R 2 is equal to the square of Pearson’s correlation coefficient [40]. It indicates the correlation degree between the simulated values and the observed values. The R 2 ranges from 0 to 1. Higher values mean better performance for the model. where Q o and Q s are the observed runoff values and the simulated runoff values (m 3 /s), respectively; Q ¯ o and Q ¯ s are the average value of Q o and the average value of Q s (m 3 /s), respectively. Both NSE and R 2 are dimensionless performance metrics, as the units of the variables are eliminated through the ratio of squared errors.

Model construction

The WAS model was adopted and enhanced in this study based on the existing model framework, with specific modifications for application in the Yanglou basin, including three-level unit division and virtual reservoir configuration. The Yanglou hydrological station basin is divided into 7 hydrological calculation units according to the unit division principle. Through a large number of investigations, the soil type in the study area is sandy soil, and the land use types are divided into 11 types, including wheat, corn, rice, vegetables, dryland, gardens, woodland, grassland, urban land, water, and unused land. In the WAS model, the ponds in each calculation unit are packaged as virtual reservoirs, which are effectively connected by spatial topological relationships. Based on the Suzhou Water Resources Bulletin and Suzhou Statistical Yearbook, this research area is dominated by domestic water and agricultural water. The domestic water is supplied by groundwater, and the agricultural water is supplied by groundwater, ponds, and river. The warm-up period of the model is 1972. The parameter calibration period of the model is from 1973 to 2000, and the validation period of the model is from 2001 to 2020. The data properties of the WAS model in this study area are shown in Figure 6.

In this study, the virtual reservoirs are set according to the dispatching rule of the reservoir. Through field surveys and measurements, the ponds are mainly used for agricultural irrigation. During the non-flood season, most ponds are dry. Due to the small size of the calculation unit, the ponds within each unit are aggregated as a single virtual reservoir. The total storage capacity of each reservoir is equal to the total volume of ponds in the corresponding calculation unit. There is no flood control requirement for ponds in the study area, so the flood control storage capacity of the reservoir is equal to the total storage capacity. The water supply relationship between the reservoir and the calculation unit is set up in the WAS model, which can effectively simulate the water consumption of human activities.

Calibration of model

Table 2 presents the simulation results during the calibration period. According to the two evaluation criteria, the NSE and R 2 of human activities in the WAS model are better than those under natural conditions. The NSE of human activities in the WAS model increased by 15.32% than that of natural conditions. It indicates that the WAS model has a high degree of fitting to human activities. The R 2 of human activities in the WAS model increased by 7.44% compared with the natural conditions. It shows that the correlation degree of human activities in the WAS model is high. Therefore, the calibration accuracy of human activities in the WAS model is superior to that of natural conditions.

The graph presenting comparison between observed and simulated runoff during model calibration is shown in Figure 7 which gives the idea and view of the best match obtained during model calibration. The results show that the agreement between simulation and observation is slightly poor in the calibration period, but the simulation can still reflect a good representation of the overall water balance. Compared with the natural conditions, the simulated runoff of human activities is superior to that of natural conditions. In this study, the Genetic Algorithm was used to optimize the parameters of the WAS model. The population size is 50, the crossover probability is 1, the mutation probability is 0.5, and the maximum number of runs is 100 generations. The important parameters of the optimized WAS model are shown in Table 3.

Validation of model

Table 4 shows the simulation results during the validation period. As detailed in Table 4, the NSE of human activities in the WAS model increased by 27.56% than that of natural conditions. The R 2 of human activities in the WAS model increased by 9.55% compared with the natural conditions. The value of these accuracy parameters indicates that the human activities of the WAS model is performing well in the simulating runoff.

Figure 8 illustrates the comparison between observed runoff and simulated runoff during the validation of the WAS model. The observed and simulated hydrographs of the human activities in the WAS model were found to match well for peak and low runoff as well reasonably. Compared with Figure 7, it was found that the agreement between simulation and observation in the calibration period is worse than that in the validation period. As can be seen in Figure 8 that the general simulation trend of human activities and natural conditions in the WAS model are similar. However, when the rainfall is large, the simulated value of human activities in the WAS model is the closest to the observed value, indicating that the simulation accuracy of human activities in the WAS model is higher. For example, in the comparative analysis of the rainfall-runoff process from 2009 to 2013, the flood peak values simulated by setting the virtual reservoir are less than that under natural conditions. It reveals that the virtual reservoirs can regulate flood flows and better simulate the process of water intake and consumption under human activities.

Table 5 presents the simulation results of multi-year average runoff in the study area. According to the observed runoff and simulated runoff, the runoff error rate of natural conditions in the WAS model is 14.94%. The runoff error rate of human activities in the WAS model is 4.09%. The runoff error rate of human activities in the WAS model is 10.85% lower than that of natural conditions. It indicates that the simulation effect of the WAS model is better.

To achieve a whole process simulation of the natural-social dualistic water cycles, the regional water allocation was validated. The regional water allocation reflects the rationality of water resource allocation. The simulation results of the multi-year average water resource allocation in the Yanglou hydrological station basin are shown in Figure 9. In the multi-year average, the actual water supply was 0.26 × 10 8 m 3, the simulated water supply was 0.24 × 10 8 m 3, and the total error of water resource allocation in the Yanglou hydrological station basin was 7.69%. From a hydrological calculation unit perspective, unit 5 had the largest error at 19.52%, while units 1, 2, and 3 had the smallest error at 0; from a water demand type perspective, all errors are less than 10%. Among them, the agriculture water error is the largest at 9.21%, and the domestic water error is the smallest at 0. The correlation coefficient between the actual water supply and the simulated water allocation in the Yanglou hydrological station basin was 0.990. This indicates that the model has a good simulation effect on water resource al-location and can meet the requirements of water resource management.

Figure 10 details the multi-year average natural-social water resources cycle transformation in the study area. In terms of the social water cycle, the total water supply in the Yanglou hydrological station basin is 0.24 × 10 8 m 3, including 0.16 × 10 8 m 3 of the surface water, and 0.08 × 10 8 m 3 of the groundwater supply. The total water consumption is 0.24 × 10 8 m 3, of which the agricultural water consumption is 0.19 × 10 8 m 3, and the rural domestic water consumption is 0.05 × 10 8 m 3. In terms of the natural water cycle, the precipitation is 1.29 × 10 8 m 3. The evaporation of canopy, soil, and interception are 1.14 × 10 8 m 3, resulting in 0.30 × 10 8 m 3 of the surface water resources, including 0.03 × 10 8 m 3 formed by runoff, 0.11 × 10 8 m 3 formed by return flow, and 0.16 × 10 8 m 3 formed by lateral flow. The total outbound water volume is 0.09 × 10 8 m 3. The results show that the water resources are balanced in the study area.

Impacts of human activities on rainfall-runoff processes

In the past 2 decades, with the acceleration of climate change and urbanization, the rainfall-runoff process will be affected by human activities, resulting in severe fluctuation of runoff data [41, 42]. Figure 11 presents the temporal variations in measured precipitation, groundwater level, and discharge. As detailed in Figure 11a, the temporal distribution of precipitation is markedly uneven, with the majority concentrated between June and October, accounting for approximately 74% of the annual total. Furthermore, comparative analysis of the observed hydrographs from 2003 to 2020 reveals markedly different hydrological responses despite similar annual precipitation totals. This discrepancy underscores the increasing influence of anthropogenic factors on hydrological systems.

This altered hydrological behavior is particularly evident when examining specific years within the study period (Figures 11b–d). For instance, the rainfall during the flood seasons in 2008 and 2012 are 637.5 mm and 606 mm, respectively. The measured discharge of both are relatively small, but the groundwater level has risen significantly. With the increase of rainfall during the flood seasons, the observed discharge significantly increases when the groundwater level rises to a certain extent (Figure 11d). The result indicates that the natural water balance can be altered by building ponds and ditches to intercept, store and divert water flow. As detailed in Figures 11e,f, the flood seasons precipitation is basically the same, but the observed discharge varied greatly. Comparative analysis of observed precipitation and discharge data for 2003 and 2018, the maximum daily precipitation in 2003 was 87.8mm, and the maximum observed discharge was 22.9 m 3 /s. As can be seen from the remote sensing image, there are many villages, ponds, and ditches in the study area. When the intensity of precipitation is low, most of the precipitation converges into the ponds through the ditches. The pond water is mainly used for agricultural irrigation. Therefore, the observed discharge of the main channel is reduced. However, the severe tropical storm Rumbia occurred in 2018. The maximum daily precipitation was 327.5mm, and the maximum observed discharge was 84.7 m 3 /s. When the intensity of precipitation is large, the ponds and ditches quickly reach saturation.

Once this water storage threshold is exceeded, the originally scattered and independent ponds and ditches will be rapidly connected to the main river channel through the channel network, causing a large amount of water to be concentrated and discharged in a short period of time. This further amplifies the peak flood flow, making the runoff response more intense than under natural conditions.

The research findings emphasize that traditional rainfall-runoff models, which assume natural basin conditions, are inadequate for predicting hydrological behavior in intensively managed catchments. Instead, there is a critical need for approaches that explicitly represent human interventions, such as the WAS model presented in this study, which integrates water allocation rules, infrastructure operations, and artificial water transfer processes. By capturing these dynamics, the WAS model provides a more reliable tool for water resources planning and flood forecasting in regions under strong anthropogenic influence.

Model performance and comparative analysis with regional studies

It is noteworthy that the NSE during calibration (0.640) was lower than that during validation (0.898). This inversion can be attributed to two factors. First, the calibration was conducted as a multi-objective optimization, simultaneously minimizing water shortage rate and maximizing allocation fairness, rather than solely maximizing runoff simulation accuracy. This deliberate trade-off resulted in relatively lower NSE in the calibration period while ensuring overall model fidelity. Second, the hydrological system in the Yanglou basin exhibited non-stationarity due to rapid land-use transformation and agricultural intensification during 1973–2000. In contrast, the validation period (2001–2020) witnessed more stable human activity patterns, allowing the calibrated model parameters to achieve superior performance. These explanations demonstrate that the observed inversion does not indicate model inadequacy, but rather reflects the model’s capacity to capture the evolving human-water interactions.

To further contextualize the performance of the WAS model, we compare it with several recent regional studies that investigated human-water interactions under similar challenges. For instance, research in the Keriya River Basin, China [43], high-lighted similar challenges in separating anthropogenic impacts from natural variability, particularly in arid regions where water resources management significantly alters hydrological processes. Our results are consistent with the findings of Miao et al. [44], who concluded that human activities dominate runoff changes in intensively managed basins. The dualistic structure of the WAS model, which explicitly represents both natural hydrological processes and human regulatory mechanisms, provides notable advantages over conventional modeling approaches. While studies in Ghana’s Pra River Basin [45] and Iran’s Central Plateau [46] successfully quantified anthropogenic impacts using various separation methods, they noted limitations in simulating the dynamic interactions between human water use and hydrological responses.

The model performed particularly well in simulating extreme runoff events. This capability is critical given the widespread observation of increasing storm runoff extremes under combined climatic and anthropogenic pressures [47]. Nevertheless, as also reported in studies of the Weihe River Basin [48], the accuracy of the WAS model depends heavily on the quality and resolution of human activity data. This limitation underscores the need for enhanced monitoring of water withdrawals and reservoir operations, as noted in several comparative studies [49].

Beyond performance evaluation, it is equally important to critically examine the applicability boundaries and transferability of the virtual reservoir method. The approach is most suitable for regions where (1) numerous small, ungauged ponds collectively exert significant influence on runoff generation and flood peak attenuation; (2) remote sensing and field survey data are available for pond identification and storage estimation; and (3) water consumption is primarily driven by agricultural irrigation and domestic use. However, the method faces notable limitations in several contexts: it is less suitable for basins dominated by large, rule-curve-operated reservoirs, where individual reservoir representation remains more appropriate; it relies on static pond inventories derived from single-period imagery, which may fail to capture rapid changes due to urbanization or land conversion; and it does not explicitly account for groundwater-surface interactions, necessitating coupling with a groundwater module in basins where seepage or exchange is significant.

Regarding transferability, the method is expected to be directly transferable to plain agricultural basins with similar pond densities, requiring only minor parameter recalibration. It may perform moderately well in hilly agricultural basins (with adjustments to pond distribution and storage estimation), in arid or semi-arid basins with ephemeral ponds (with revised operation rules), and in urbanized basins with detention ponds (with higher temporal resolution). Transferability is low for forested or natural basins, where human interventions are limited, and for basins dominated by large reservoirs, where individual reservoir representation is preferred. Regardless of basin type, successful transfer requires site-specific calibration using local data, adjustment of operation rules to match regional hydrological and agricultural calendars, and validation against independent observations.

This study developed an improved WAS model incorporating a three-level unit division method and a virtual reservoir approach to simulate rainfall-runoff processes under strong anthropogenic influences. The model was applied and validated in the Yanglou hydrological station basin, Northern Anhui Province, China. The main conclusions are as follows: The three-level unit division method, which integrates water resources zoning with administrative boundaries and refines units through DEM-based hydrological analysis and multi-layer land use overlay, enables the WAS model to more accurately quantify river flow changes under human activities.

By aggregating numerous small ungauged ponds into storage units with demand-driven operation rules, the virtual reservoir approach effectively captures the storage-consumption-release continuum and the threshold effect of pond networks on flood peak attenuation and amplification.

The WAS model exhibits superior performance in simulating rainfall-runoff processes under anthropogenic influences compared to scenarios under natural conditions. During calibration, the model achieved increases of 15.32% in NSE and 7.44% in R 2 relative to natural condition simulation. Further improvement was observed during the validation period, where NSE and R 2 rose by 27.56% and 9.55%, respectively. These results demonstrate the model’s improved robustness and reliability in capturing hydrological processes affected by human activities.

The WAS model with virtual reservoirs is most applicable to plain agricultural basins with high-density pond networks and intensive irrigation activities, where remote sensing and field survey data are available for pond identification and storage estimation. The method is less suitable for basins dominated by large rule-curve-operated reservoirs, forested/natural catchments with limited human intervention, or regions where groundwater-surface interactions are dominant without coupling with a groundwater module. Successful transfer to other basins requires site-specific calibration, adjustment of operation rules to match regional hydrological and agricultural calendars, and validation against independent observations.

Based on these findings, the WAS model offers an enhanced framework for simulating rainfall-runoff processes under anthropogenic influences and serves as a valuable tool for applications in highly human-impacted regions. To further advance the approach, future research will focus on three specific directions, which collectively address the need for systematic sensitivity and uncertainty assessment. First, multi-basin verification will apply the model to 3–5 basins with contrasting climatic, topographic, and anthropogenic settings to establish a generalized parameter transfer framework and quantify the method’s performance envelope, and evaluate the sensitivity of model parameters to changing environmental conditions. Second, dynamic land use coupling will integrate time-series remote sensing data with the virtual reservoir module to enable dynamic updating of pond inventories and storage capacities, capturing seasonal and inter-annual changes while providing a basis for uncertainty analysis under evolving landscape conditions. Third, real-time operation integration will couple the WAS model with ensemble weather forecasts to develop an operational tool for flood forecasting and early warning applications, which will facilitate the assessment of prediction uncertainty in operational settings.