The Qingmuguan karst trough valley is located in the southeastern part of the Sichuan Basin, China, at the junction of the Beibei and Shapingba districts in Chongqing, with coordinates of 29∘40′–29∘47′ N, 106∘17′–106∘20′ E. The basin covers an area of 13.4 km2 and is part of the southern extension of the anticline at Wintang Gorge in the Jinyun Mountains, with the anticlinal axis of Qingmuguan located in a parallel valley in eastern Sichuan (Yang et al., 2008). The surface of the anticline is heavily fragmented, and faults are extremely well developed, with large areas of exposed Triassic carbonate rocks. Under the long-term erosion of karst water, a typical karst trough landform developed (Liu et al., 2009). This karst trough landform provides convenient conditions for flood propagation, and the development of karst landforms is extremely common in the karst region of southwestern China, especially in the karst region of Chongqing.

The basin is oriented in a narrow band of slightly curved arcs and is ∼12 km long from north to south. The direction of the mountains in the region is generally consistent with the direction of the tectonic line. The catchment area of the basin is mainly composed of the outlying areas of the Lower Triassic Jialingjiang Formation (T1j), the middle Leikoupo Formation (T2l) carbonate rocks on both sides of the mountain slopes, and part of the Upper Xujiahe Formation (T3xj) quartz-sandstone and mudstone (Yang et al., 2008). Tracer tests show that karst development in the underground river system in the study area is strong, where the karst water-bearing medium is heterogeneous and has high water permeability. A large-scale underground river (Fig. 1) with a length of approximately 7.4 km has developed in the karst trough valley, and the flood peak flow of this underground river lasts for a short time.

The karst landforms in the area are well developed under closed conditions, and precipitation is the main source of recharge for the underground river system. Most of the precipitation, after evapotranspiration and plant retention are deducted, collects along the slope to the depression at the bottom of the trough and joins the underground river through surface karst fissure dispersion infiltration and concentrated injection in the sinkholes. The map in Fig. 1 gives an overview of the Qingmuguan karst basin.

The Qingmuguan basin is located within a subtropical humid monsoon climate zone, with an average temperature of 16.5 ∘C and an average precipitation of 1250 mm concentrated mainly in May–September. An underground river system with a length of 7.4 km has developed in the karst trough valley, and the water supply of the underground river is mainly rainfall recharge (Zhang, 2012). Most of the precipitation is collected along the hill slope and routed into the karst depressions at the bottom of the trough valley, where it joins the underground river through the dispersed infiltration of surface karst fissures and sinkholes (Fig. 1a). An upstream surface river collects in a gentle valley and enters the underground river through the Yankou sinkhole (elevation 524 m). Surface water in the middle and lower reaches of the river system enters the underground river system mainly through cover collapse sinkholes (Gutierrez et al., 2014) or fissures.

The stratigraphic and lithologic characteristics of the basin are dominated largely by carbonate rocks of the Lower Triassic Jialingjiang Group (T1j) and Middle Triassic Leikou Slope Group (T2l) on both sides of the slope, with some quartz sandstone and mudstone outcrops of the Upper Triassic Xujiahe Group (T3xj) (Zhang, 2012). The topography of the basin presents a general anticline (Fig. 1b), where carbonate rocks on the surface are corroded and fragmented and have high permeability. Compared with the core of the anticline, the shale of the anticline is less eroded and forms a good waterproof layer.

To investigate the distribution of karst conduits in the underground river system, we conducted a tracer test in the study area. The tracer was placed in the Yankou sinkhole and recovered in the Jiangjia spring (Fig. 1a and c). According to the tracer test results (Gou et al., 2010), the karst water-bearing medium in the aquifer was anisotropic, the karst conduits in the underground river were extremely well developed, and there was a large single-channel underground river approximately 5 m wide. The response of the underground river to rainfall was very fast, with the peak flow observed at the outlet of the Jiangjia spring 6–8 h after rainfall based on the tracer test results. The flood peak rose quickly, and the duration of the peak flow was short. The underground river system in the study area is dominated by large karst conduits, which are not conducive to water storage in water-bearing media but are very conducive to the propagation of floods.

To build the QMG model to simulate karst flood events, the necessary baseline modelling data had to be collected, including (1) high-resolution DEM data and hydrogeological data (e.g. the thickness of the epikarst zone, rainfall infiltration coefficients of different karst landforms, and the rock permeability coefficient); (2) land-use and soil-type data; and (3) rainfall data in the basin and water flow data for the underground river. The DEM data were downloaded from a free database on the public internet, with an initial spatial resolution of 30×30 m. The spatial resolution of the land use and soil types was 1000×1000 m, and they were also downloaded from the internet. After the applicability of modelling and computational strength, as well as the size of the basin in the study area (13.4 km2), was considered, the spatial resolution of the three types of data was resampled uniformly in the QMG model and downscaled to 15×15 m based on a spatial discrete method by Berry et al. (2010).

The hydrogeological data necessary for modelling were obtained in three simple ways. (1) A basin survey was conducted to obtain the thickness of the epikarst zone, which was achieved by observing the rock formations on hillsides following cutting for road construction. Information was collected regarding the location, general shape and size of karst depressions and sinkholes, which had a significant impact on the compilation of the DEM data and the determination of the convergence process of surface runoff. The sinkholes in the basin are cover collapse sinkholes (Gutierrez et al., 2014) according to the basin survey. There are 3 large sinkholes (more than 3 m in diameter) and 12 small sinkholes (less than 1 m in diameter). The remaining 5 sinkholes are between 1 and 3 m in diameter. The confluence calculation of these sinkholes in the model was based on the results of a previous study (Meng et al., 2009). (2) Empirical equations developed for similar basins were used to obtain the rainfall infiltration coefficients for different karst landforms and the rock permeability coefficient. For example, the rock permeability coefficient was calculated based on an empirical equation from a pumping test in a coal mine in the study area (Li et al., 2019). (3) A tracer experiment was conducted in the study area (Gou et al., 2010) to obtain information on the underground river direction and flow velocity; for instance, underground karst conduits are well developed in the area and form an underground river approximately 5 m wide. There are no hydraulic connections between the underground river system in the area and the adjacent basin, which means that there is no overflow recharge.

Rainfall and flood data are important model inputs and represent the driving factors that allow hydrological models to operate. In the study area, rainfall data were acquired by two rain gauges located in the basin (Fig. 1a). Point rainfall values were then spatially interpolated into basin-level rainfall (for such a small basin area, rainfall results obtained from two rain gauges were considered representative). There were 18 karst flood events from 14 April 2017 to 10 June 2019. We built a rectangular open channel at the underground river outlet and set up a river gauge in it (Fig. 1a) to record the water level and flow data every 15 min.

The hydrological model developed in this study was named the QMG model after the basin for which it was developed and to which it was first applied, i.e. the Qingmuguan basin. The QMG model has a two-layer structure, including a surface part and an underground part. The surface structure is mainly used to perform the calculation of runoff generation and the confluence of the surface river, while the underground structure is used to perform the confluence calculation of the underground river system.

The structure of the QMG model is divided into a two-layer structure, both horizontally and vertically. The horizontal structure of the model is divided into river channel units and slope units. The vertical structure below the surface is divided into a shallow karst aquifer (including soil layers, karst fissures and conduit systems in the epikarst zone) and a deep karst aquifer system (bedrock and the underground river system). This relatively simple model structure means that only a small amount of hydrogeological data is needed in karst regions. Figure 2 shows a flowchart of the modelling and calculation procedures required for the QMG model.

To accurately describe the runoff generation and confluence on a grid scale, these karst subbasins are further divided into many karst hydrological response units (KHRUs) based on the high-resolution (15×15 m) DEM data in the model. The specific steps involved in the division were adopted by referring to studies of hydrological response units (HRUs) in TOPMODEL by Pan (2014). The KHRUs are the smallest basin computing units; the spatial differences in karst development within the units can be effectively ignored, and the use of these units reduces the uncertainty in the model unit classification. Figure 3 shows the spatial structure of the KHRUs.

The right-hand side of Fig. 3 shows a three-dimensional spatial model of KHRUs established in the laboratory to visually reflect the storage and movement of water in the karst water-bearing medium with each spatially anisotropic component and to provide technical support for establishing the hydrological model.

The modelling and operation of the QMG model consists of three main stages: (1) spatial interpolation and the retention of rainfall and evaporation calculations; (2) runoff generation and confluence calculation for the surface river; and (3) confluence calculation for the underground runoff, including the confluence in the shallow karst aquifer and the underground river system.

In total, the QMG model includes 12 parameters, among which flow direction and slope are topographic parameters that can be determined from the DEM without parametric optimization while the remaining 10 parameters require calibration. Other distributed hydrological models with multiple structures usually have many parameters. For example, the Karst–Liuxihe model (Li et al., 2021) has 15 parameters that must be calibrated. In the QMG model, each parameter is normalized as 16xi=x∗i/xi0, where xi is the dimensionless parameter value i after it is normalized, x∗i is the parameter value i in actual physical units, and xi0 is the initial or final value of xi. Through the processing of Eq. (16), the value range of the model parameters is limited to a hypercube Kn=(X|0≤xi≤1, i=1, 2, …, n), where K is a dimensionless value. This normalized treatment ignores the influence of the spatiotemporal variation in the underlying surface attributes on the parameters while also simplifying the classification and number of model parameters to a certain extent. Accordingly, the model parameters can be further divided into rainfall-evaporation parameters, epikarst-zone parameters and underground river parameters. Table 1 lists the parameters of the QMG model.

Because the QMG model has relatively few parameters, it is possible to calibrate them manually, which means that the operation is easy to implement and does not require a special programme for parameter optimization. However, the choice of parameters is subjective, which can lead to great uncertainty in the manual parameter calibration process. To compare the effects of parameter optimization on model performance, this study used both manual parameter calibration and the improved chaotic particle swarm optimization (ICPSO) algorithm for the automatic calibration of model parameters and compared the effects of both on flood simulation.

In general, the structure and parameters of a standard particle swarm optimization (PSO) algorithm are simple, with the initial parameter values obtained at random. For parameter optimization in high-dimensional multipeak hydrological models, the standard PSO is easily limited to local convergence and cannot achieve the optimal effect, while the late evolution of the algorithm may also cause problems, such as premature convergence and stagnant evolution, due to the “inert” aggregation of particles, which seriously affects the efficiency of parameter selection. It is necessary to overcome the above problems and to facilitate a high probability of algorithm convergence to the global optimal solution. In parameter optimization for the QMG model, we improved the standard PSO algorithm by adding chaos theory and developed the ICPSO, where 10 cycles of chaotic disturbances were added to improve the activity of the particles. The inverse mapping equation of the chaotic variable is 17Xij=Xmin+(Xmax-Xmin)⋅ZijZij′=(1-α)Z∗+αZij, where Xij is the optimization variable for the model parameters and (Xmax-Xmin) is the difference between its maximum and its minimum; Zij is the variable before the disturbance is added and Zij′ represents the chaotic variables after a disturbance is added; α (0≤α≤1) is a variable determined by the adaptive algorithm; and Z∗ is the chaotic variable formed when the optimal particle is mapped to the interval [0,1]. The flowchart of the ICPSO for parameter optimization is shown in Fig. 4.

Uncertainties in hydrological model simulation results usually originate from three aspects: input data, model structure and model parameters (Krzysztofowicz, 2014). In the present study, the input data (e.g. rainfall, flood events and some hydrogeological data) were first validated and preprocessed through observations to reduce their uncertainties.

Second, we simplified the structure of the QMG model to reduce the structural uncertainty. As it is a mathematical and physical model, a hydrological model has some uncertainty in flood simulation and prediction because of the errors in system structure and the algorithm (Krzysztofowicz and Kelly, 2000). The model was designed with full consideration of the relationship between the amount of data required to build the model and its performance for flood simulation and prediction in karst regions, and the model's entire framework was integrated through simple structures and easy-to-implement algorithms using the concept of distributed hydrological modelling. Conventionally, the extent of uncertainty increases with the growing complexity of the model structure. We therefore ensured that the structure of the QMG model was simple when it was designed, and the model was divided into surface and underground double-layer structures to reduce its structural uncertainty.

Third, we focused on analysing the uncertainty and sensitivity of the model parameters and their optimization method, for which a multiparametric sensitivity analysis method (Li et al., 2020) was used to analyse the sensitivity of the parameters in the QMG model. The steps in the parameter sensitivity analysis were as follows.

Selection of the appropriate objective function.

The Nash–Sutcliffe coefficient is widely used as an objective function to evaluate the performance of hydrological models (Li et al., 2020, 2021). The coefficient was therefore used to assess the QMG model. Because the most important factor in flood prediction is the peak discharge, it is used in the Nash–Sutcliffe coefficient equation:18NSC=1-∑i=1nQi-Qi′2∑i=1nQi-Q‾2,where NSC is the Nash–Sutcliffe coefficient, Qi [L s-1] represents the observed flow discharges, Qi′ [L s-1] represents the simulated discharges, Q‾ [L s-1] is the average observed discharge and n [h] is the observation period.

Once the model was built, some of the initial conditions had to be set before running it to simulate and forecast floods, such as basin division, the setting of initial soil moisture, and the assumption of the initial parameter range. (1) In the study area, the entire Qingmuguan karst basin was divided into 893 KHRUs, including 65 surface river units, 466 hill slope units and 362 underground river units. The division of these units formed the basis for calculating the process of runoff generation and convergence. (2) The initial soil moisture was set to 0 %–100 % of the saturation moisture content in the basin, and the specific soil moisture before each flood had to be determined by a trial calculation. (3) The waterhead boundary conditions of the groundwater were determined by a tracer test in the basin, where a perennial stable water level adjacent to the groundwater-divide was used as the fixed waterhead boundary. The base flow of the underground river was determined to be 35 L s-1 from the perennial average dry season runoff. (4) The ranges of initial parameters and convergence conditions were assumed before parameter optimization (Fig. 4). (5) Parameter optimization and flood simulation validated the performance of the QMG model in karst basins.

The number of parameters in a distributed hydrological model is generally large, and it is important to perform a sensitivity analysis on each parameter to quantitatively assess the impacts of the different parameters on model performance. In the QMG model, each parameter was placed into one of four categories according to its sensitivity: (i) highly sensitive, (ii) sensitive, (iii) moderately sensitive and (v) insensitive. In the calibration of model parameters, insensitive parameters do not need to be calibrated, which can greatly reduce the number of calculations and improve the model operation efficiency.

The flow process in the calibration period (14 April to 10 May 2017) was adopted to calculate the sensitivity of the model parameters, where Eq. (19) was used, and the parameter sensitivity results are presented in Table 2.

In Table 2, the value of σi (Eq. 19) represents the parameter's sensitivity, and the higher the value, the more sensitive the parameter is. The results in Table 2 show that the rainfall infiltration coefficient, rock permeability coefficient, rock porosity, and the related parameters of soil water content, such as the saturated water content and field capacity, were sensitive parameters. The order of parameter sensitivity was as follows: infiltration coefficient > permeability coefficient > rock porosity > specific yield > saturated water content > field capacity > flow direction > thickness > slope > soil coefficient > channel roughness > evaporation coefficient.

In the QMG model, parameters were classified as highly sensitive, sensitive, moderately sensitive or insensitive according to their influence on the flood simulation results. In Table 4, we divided the sensitivities of model parameters into four levels based on the σi value: (1) highly sensitive parameters, 0.8<σi<1; (2) sensitive parameters, 0.65<σi<0.8; (3) moderately sensitive parameters, 0.45<σi<0.65; and (4) insensitive parameters, 0<σi<0.45. The infiltration coefficient, permeability coefficient, rock porosity and specific yield were highly sensitive parameters. The saturated water content, field capacity and thickness of the epikarst zone were sensitive parameters. The flow direction, slope and soil coefficient were moderately sensitive parameters. The channel roughness and the evaporation coefficient were insensitive parameters.

In total, the QMG model has 12 parameters, of which only eight need to be optimized, which is relatively few from the perspective of distributed models. The parameters of flow direction and slope as well as the insensitive parameters of channel roughness and the evaporation coefficient do not need to be calibrated, which can improve the convergence efficiency of the model parameter optimization.

In the study area, 18 karst floods were recorded at the underground river outlet during the period from 14 April 2017 to 10 June 2019 and used to validate the effects of the QMG model in karst hydrological simulations. The calibration period was from 14 April to 10 May 2017, at the beginning of the flow process, with the remainder of the time being used as the validation period. In the QMG model, the ICPSO algorithm was used to optimize the model parameters. To show the necessity of parameter optimization for the distributed hydrological model, this study specifically compared the flood simulations obtained using the initial parameters of the model (without parameter calibration) and the optimized parameters. Figure 5 shows the iteration process of parameter optimization for the QMG model.

Figure 5 shows that almost all parameters fluctuated widely at the beginning of the optimization. After approximately 15 iterations of optimization calculations, most of the linear fluctuations became significantly less volatile, which indicated that the algorithm was tending towards convergence (possibly only locally). When the number of iterations exceeded 25, all parameters remained essentially unchanged, meaning that the algorithm had converged (at this point, there was global convergence). It took only 25 iterations to reach a definite convergence of the parameter rates with the ICPSO algorithm, which is extremely efficient in terms of the parameter optimization of distributed hydrological models. In previous studies of the parametric optimization for the Karst–Liuxihe model in similar basin areas, 50 automatic parameter optimization iterations were required to reach convergence (Li et al., 2021), demonstrating the effectiveness of the ICPSO algorithm.

To evaluate the effect of parameter optimization, the convergence efficiency of the algorithm and, more importantly, the parameters after calibration were used to simulate floods. Figure 6 shows the flood simulation effects.

Figure 6 shows that the flows simulated by parameter optimization were better than those simulated by the initial model parameters. The simulated flow processes based on the initial parameters were relatively small, with the simulated peak flows in particular being smaller than the observed values, and there were large errors between the two values. In contrast, the simulated flows produced by the QMG model after parameter optimization were very similar to the observed values, which indicated that calibration of the model parameters was necessary and that there was an improvement in parameter optimization through the use of the ICPSO algorithm in this study. In addition, it was found that the flow simulation effect was better in the calibration periods than in the validation periods (Fig. 6).

To compare the flow process simulation results based on the initial model parameters with the optimized parameters, six evaluation indices (Nash–Sutcliffe coefficient, correlation coefficient, relative flow process error, flood peak error, water balance coefficient and peak time error) were applied in this study, and the results are presented in Table 3.

Table 3 shows that the evaluation indices of the flood simulations after parametric optimization were better than those of the initial model parameters. The average values of the initial parameters for these six indices were 0.81, 0.74, 27 %, 31 %, 0.80 and 5 h, respectively. For the optimized parameters, the average values were 0.90, 0.91, 16 %, 14 %, 0.94 and 3 h, respectively. The flood simulation effects after parameter optimization were clearly improved, implying that parameter optimization of the QMG model was necessary and that the ICPSO algorithm was an effective approach for parameter optimization that could greatly improve the convergence efficiency of parameter optimization and ensure that the model performed well in flood simulations.

Following parameter optimization, we simulated the whole flow process (14 April 2017 to 10 June 2019) based on the optimized and initial parameters of the QMG model (Fig. 6), which enabled a visual reflection of the application of the model for the simulation of a long series of flow processes. To reflect the simulation effects of the model for different flood events, we divided the whole flow process into 18 flood events and then used the initial parameters of the model and the optimized parameters to verify the model performance in flood simulations. Figure 7 and Table 4 show the flood simulation effects and their evaluation indices obtained using both the initial and the optimized parameters.

Figure 7 shows that the flood simulation results obtained using the initial parameters were smaller than the observed values and that the model performance in flood simulations improved after parameter optimization. The simulated flood processes were in good agreement with observations and were especially effective for simulating flood peak flows. From the flood simulation indices in Table 4, the average water balance coefficient based on the initial parameters was 0.69, i.e. much less than 1, indicating that the simulated water in the model was unbalanced. After parameter optimization, the average value was 0.92, indicating that parameter optimization had a significant impact on the model water balance calculation.

Table 4 shows that the average values of the six indices (Nash–Sutcliffe coefficient, correlation coefficient, relative flow process error, flood peak error, water balance coefficient and peak time error) for the initial parameters were 0.79, 0.74, 26 %, 25 %, 0.69 and 5 h, respectively, while for the optimized parameters, the average values were 0.92, 0.90, 10 %, 11 %, 0.92 and 2 h, respectively. All evaluation indices improved after parameter optimization, with the average values of the Nash–Sutcliffe coefficient, correlation coefficient and water balance coefficient increasing by 0.13, 0.16 and 0.23, respectively. The average values of the relative flow process error, flood peak error and peak time error decreased by 15 %, 14 % and 3 h, respectively. These reasonable flood simulation results confirmed that parameter optimization by the ICPSO algorithm was necessary and effective for the QMG model.

Compared with the overall flow process simulation shown in Fig. 6, each flood process was better simulated by the QMG model (Fig. 7). This was because the main consideration of the QMG model was the calculation of the flood process and the correlation algorithm of the dry season runoff was not sufficiently described. For example, Eqs. (12)–(15) represent the flood convergence algorithm. As a result, the model is not good at simulating other flow processes, such as dry season runoff, leading to a low accuracy in the overall flow process. The next phase of our research will focus on refining the algorithm related to dry season runoff and improving the comprehensive performance of the model.

The parameter sensitivity results in Table 2 show that the rainfall infiltration coefficient in the QMG model was the most sensitive parameter. This was the key to determining the generation of excess infiltration surface runoff and separating surface runoff from subsurface runoff. If the rainfall infiltration coefficient was greater than the infiltration capacity, excess infiltration surface runoff was generated on the exposed karst landforms; otherwise, all rainfall would infiltrate to meet the water deficit in the vadose zone and then continue to seep into the underground river system, eventually flowing out of the basin through the underground river outlet. The confluence modes of surface runoff and underground runoff were completely different, resulting in a large difference in the simulated flow results. Therefore, the rainfall infiltration coefficient had the greatest impact on the final flood simulation results.

Other highly sensitive parameters, such as the rock permeability coefficient, rock porosity and specific yield, were used as the basis for dividing between slow flow in karst fissures and rapid flow in conduits. The division of slow and rapid flows also had a great impact on the discharge at the outlet of the basin. Slow flow plays an important role in water storage in a karst aquifer and is very important for the replenishment of river base flow in the dry season. Rapid flow in large conduit systems dominates flood runoff and is the main component of the flood water volume in the flood season.

Parameters related to the soil water content, including the saturated water content, field capacity and thickness, were sensitive parameters and had a strong influence on the flood simulation results. This is because the soil moisture content prior to flooding affects how flood flows rise and when peaks occur. If the soil is already very wet or even saturated before flooding, the flood rises quickly to a peak, and the process line of the flood peak flow is sharp and thin. This type of flood process forms easily and can lead to disaster-causing flood events. In contrast, if the soil in the basin is very dry before flooding, the rainfall first counteracts the water shortage of the vadose zone, and after this zone is replenished, the rainfall infiltrates into the underground river. The flood peak of the river basin outlet is therefore delayed.

The flow direction, slope and soil coefficient were moderately sensitive parameters. They had a specific influence on the flood simulation results, but the influence was not as great as that of the highly sensitive and sensitive parameters. The channel roughness and the evaporation coefficient were insensitive parameters. The amount of water lost by evapotranspiration is a very small in part of the total flood water, and it was therefore the least sensitive parameter in the QMG model.

In general, the uncertainty in model simulation is due mainly to three aspects of the model: (i) the uncertainty of its input data, (ii) the uncertainty of its structure and algorithm, and (iii) the uncertainty of its parameters. In the practical application of a hydrological model, these three uncertainties are usually interwoven, which leads to the overall uncertainty of the final simulation results (Krzysztofowicz, 2014). Therefore, the present study focused on the uncertainties in the input data, the model structure and the parameters to reduce the overall uncertainty of the simulation results.

First, the input data – mainly rainfall-runoff data and hydrogeological data – were preprocessed, which substantially reduced their uncertainty. Second, we simplified the structure of the QMG model, which is reflected in the fact that it has only two layers of spatial structure in the horizontal and vertical directions. This relatively simple structure greatly reduced the model structure-related uncertainty. In contrast, the underground structure of our previous Karst–Liuxihe model (Li et al., 2021) has five layers, which leads to great uncertainty. Third, appropriate algorithms for runoff generation and confluence were selected. Different models were designed for different purposes, which led to great differences in the algorithms used. In the QMG model, most of the rainfall-runoff algorithms used have been validated against the research results of others, and some of them were improved to suit karst flood simulation and prediction by the QMG model. For example, the algorithm for the generation of excess infiltration runoff (Eq. 2) represented an improvement over the version used in the Liuxihe model (Chen et al., 2010; Li et al., 2020). Finally, the algorithm for parameter optimization was improved. Considering that the standard PSO algorithm tends to converge locally, this study developed the ICPSO for parameter optimization by adding chaotic perturbation factors. The flood simulation results after parameter optimization were much better than those of the initial model parameters (Figs. 6 and 7 and Tables 2 and 3), which indicates that parameter optimization is necessary for a distributed hydrological model and can reduce the uncertainty of the model parameters.

The QMG model presented in this study uses Visual Basic language programming. The general framework of the model and the algorithm consist of three parts: the modelling approach, the algorithm of rainfall-runoff generation and confluence, and the parameter optimization algorithm. As this model is a free and open-source hydrological modelling programme (QMG model-V1.0), we provide all modelling packages, including the model code, installation package, simulation data package and user manual, free of charge. It is important to note that the model we provide is only for scientific research purposes and should not be used for any commercial purposes. Creative Commons Attribution 4.0 International.

The model installation programme can be downloaded from ZENODO; cite as Li (2021a) and Li (2021b) (registration required). The user manual can be downloaded from 10.5281/zenodo.4964754 (Li, 2021c).