Intelligent watershed hydrological model construction method suitable for regions lacking reservoir operation data
By combining simplified reservoir operation rules with a deep learning module, an intelligent hydrological model solves the problem of runoff simulation in areas lacking reservoir operation data, achieving high-precision hydrological simulation and river prediction, and is suitable for areas lacking reservoir operation data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-04-07
AI Technical Summary
In regulated watersheds where reservoir operation data is scarce, existing hydrological models struggle to accurately simulate runoff, simplified rules lack sufficient simulation accuracy, and complex rules rely on a large amount of measured data, making them difficult to implement.
By combining a conceptual hydrological model that simplifies reservoir operation rules with a deep learning module, an intelligent hydrological model is constructed. The deep learning module is trained using watershed meteorological and attribute data to optimize physical parameters and build an end-to-end model architecture to simulate reservoir outflow.
In the absence of reservoir operation data, the accuracy of runoff simulation has been significantly improved, eliminating the reliance on a large amount of measured data and maintaining the adherence to and interpretability of physical laws.
Smart Images
Figure CN121809223A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hydrological simulation technology, and in particular to a method for constructing intelligent hydrological models for watersheds in areas lacking reservoir operation data. Background Technology
[0002] Reservoir operation and management is a common solution for addressing flood disasters, alleviating water shortages, and addressing energy insufficiency. However, by allocating water resources in time and space, they significantly alter the hydrological rhythms of natural rivers. Therefore, as reservoir operation gradually becomes an integral part of the watershed water cycle, effectively integrating its operation into hydrological models has become a core issue in accurately simulating the hydrological processes of reservoir-regulated watersheds.
[0003] A common approach to representing reservoir operation in hydrological models is to use reservoir operation rules of varying complexity as post-processing modules. These rules transform the hydrological model's output into river runoff regulated by the reservoir. These reservoir operation rules are generally divided into two categories: conceptual operation rules and data-driven operation rules. Conceptual operation rules describe the relationship between inflow, reservoir capacity, and outflow by establishing empirical functions, sometimes further refining the water demand and reservoir capacity design for different objectives. Simplified general reservoir operation rules, as a special type of conceptual rule, use default empirical coefficients, reducing the need for measured reservoir operation data. For example, only the simulated reservoir inflow needs to be used to derive the reservoir outflow, making them particularly suitable for runoff simulation in regulated watersheds where reservoir operation data is scarce. While these simplified operation rules are easy to use, their simplified structure often overlooks the differentiated operation of reservoirs under different scenarios and according to different objectives.
[0004] In contrast, more complex reservoir operation rules (still conceptual rules) rely on long-term measured data on inflow, outflow, and reservoir capacity. By optimizing or calibrating reservoir scheduling functions, they achieve a more accurate simulation of the reservoir's scheduling process. For example, by constructing a parameterized piecewise linear relationship between reservoir capacity, inflow, and outflow, a reservoir management scheme can be approximated. With the advent of the big data and artificial intelligence era, data-driven operation rules are gaining popularity. This method breaks through the limitations of traditional model structures, eliminating the need for pre-defined physical rules. Instead, it learns from a large amount of historical operational data to optimize operational decisions under different scenarios, thus better simulating the actual changes in reservoir operation. Therefore, data-driven operation rules typically outperform conceptual rules. However, due to security and confidentiality considerations, data sharing mechanisms are currently inadequate. Limited by the scarcity of measured operational data, many reservoirs cannot provide sufficient historical records for modeling and optimization. Therefore, this restricts the application of more complex reservoir operation rules in data-scarce regulated watersheds. While data-driven operational rules offer better performance than conceptual rules, they also require a large amount of high-quality reservoir operation data as input.
[0005] In summary, simplified reservoir operation rules, due to structural limitations, struggle to accurately depict actual operational processes, thus restricting their simulation accuracy. While more complex reservoir operation rules and data-driven operation rules offer superior performance, they typically rely on extensive measured data for modeling. Therefore, achieving accurate runoff simulation remains a critical challenge in regulated watersheds where reservoir operation data is scarce. Summary of the Invention
[0006] The purpose of this invention is to provide a method for constructing intelligent hydrological models for watersheds in areas lacking reservoir operation data, so as to more accurately simulate runoff in regulated watersheds where reservoir operation data is scarce.
[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems, in its first aspect, provides a method for constructing a smart hydrological model for watersheds in areas lacking reservoir operation data, comprising the following steps:
[0008] A conceptual hydrological model based on simplified reservoir operation rules is established. The conceptual hydrological model includes a physical hydrological module and a reservoir operation module. The physical hydrological module is used to calculate the inflow runoff based on the required physical parameters. The reservoir operation module is used to calculate the outflow runoff based on the input inflow runoff and the preset reservoir operation rules.
[0009] Collect watershed meteorological data, watershed natural attribute values, reservoir attribute values, and reservoir outflow data to form a sample set, and divide the sample set into a training set, a validation set, and a test set.
[0010] The first deep learning module is constructed to simulate the physical parameters required by the physical hydrology module based on watershed meteorological data, watershed natural attribute values, and reservoir attribute values.
[0011] The physical parameters required by the physical hydrology module in the output of the first deep learning module are input into the physical hydrology module, and the output of the physical hydrology module is input into the reservoir operation module to construct an intelligent hydrology model.
[0012] The intelligent hydrological model is trained using training and validation sets. The loss function is calculated based on the reservoir outflow data from the training and validation sets, and backpropagation is performed to optimize the parameters and hyperparameters of the first deep learning module, thus obtaining the final intelligent hydrological model.
[0013] In some embodiments, to provide a feasible conceptual hydrological model based on simplified reservoir operation rules, the conceptual hydrological model based on simplified reservoir operation rules is an HBV (Hydrologiska Byråns Vattenbalansavdelning) hydrological model that combines simplified reservoir operation rules; the physical hydrological module is an HBV module.
[0014] In some embodiments, to provide a feasible required physical parameter, the required physical parameter includes snowmelt path, soil path, response path, and confluence path.
[0015] In some embodiments, to compensate for the limitations of applying simplified reservoir operation rules to complex environments, the soil path includes parameters controlling the conversion of soil moisture into net rainfall, namely, soil field holding capacity and the contribution rate of precipitation to runoff. The parameters controlling the conversion of soil moisture into net rainfall, namely, soil field holding capacity and the contribution rate of precipitation to runoff, are dynamic parameters, while the remaining required physical parameters are static parameters; the simulated reservoir operation rules are dynamic parameters.
[0016] In some embodiments, to provide feasible watershed meteorological data, watershed natural attributes, and watershed reservoir attributes, the watershed meteorological data includes precipitation data, average temperature data, and potential evapotranspiration data; the watershed natural attributes include drought index data, watershed area data, and heavy precipitation frequency data; and the watershed reservoir attributes include reservoir dam height data, maximum reservoir capacity data, regulation degree data, and main uses data.
[0017] In some embodiments, to further improve the performance of the constructed intelligent hydrological model, a second deep learning module is also constructed. The second deep learning module is used to simulate the reservoir outflow based on the inflow runoff, watershed meteorological data, watershed natural attribute values and reservoir attribute values output by the physical hydrological module.
[0018] The reservoir operation module was replaced with a second deep learning module.
[0019] The intelligent hydrological model is trained using training and validation sets. A loss function is calculated based on the reservoir outflow data from the training and validation sets, and backpropagation is performed to optimize the parameters and hyperparameters of the first deep learning module. The final intelligent hydrological model includes:
[0020] The intelligent hydrological model is trained using training and validation sets. The loss function is calculated based on the reservoir outflow data from the training and validation sets, and backpropagation is performed to optimize the parameters and hyperparameters of the first and second deep learning modules, thus obtaining the final intelligent hydrological model.
[0021] In some embodiments, for ease of subsequent use, the intelligent hydrological model, during training, uses the watershed meteorological data in the training set samples and validation set samples as input values, and incorporates the watershed natural attribute values and reservoir attribute values in the training set samples, and uses the reservoir outflow data in the training set samples and validation set samples as output values.
[0022] When using the intelligent hydrological model, the outflow of water from the reservoir can be obtained by inputting meteorological data from the watershed.
[0023] In some embodiments, to evaluate the hydrological simulation accuracy of the intelligent hydrological model, the method further includes:
[0024] The watershed meteorological data from the test set sample is input into the final intelligent hydrological model to obtain the simulated reservoir outflow. The simulated reservoir outflow is then compared with the reservoir outflow in the test set sample to evaluate the accuracy of the intelligent hydrological model's hydrological simulation.
[0025] In some embodiments, to provide a feasible first deep learning module and / or second deep learning module, both the first deep learning module and / or the second deep learning module employ a long short-term memory artificial neural network (LSTM).
[0026] In the second aspect of the technical solution adopted by the present invention to solve the above-mentioned technical problems, a readable storage medium is provided, which stores the final intelligent hydrological model obtained by the above-mentioned method for constructing a smart hydrological model for watersheds in areas lacking reservoir operation data.
[0027] The beneficial effects of this invention are that, in the solution of this invention, the intelligent hydrological model constructed by the above-mentioned method for constructing intelligent hydrological models for watersheds in areas lacking reservoir operation data improves the accuracy of hydrological simulation in watersheds regulated by reservoirs, while eliminating the dependence on a large amount of measured reservoir operation data required when modeling complex reservoir operation rules based on data-driven methods. This provides a more effective solution for hydrological modeling and river runoff prediction in areas lacking reservoir operation data. Compared with purely physical hydrological models, such as conceptual hydrological models based on simplified reservoir operation rules, the intelligent hydrological model obtained above benefits from the excellent learning ability of deep learning for nonlinear relationships, resulting in a significant improvement in simulation accuracy. Compared with purely deep learning models, this intelligent hydrological model benefits from the nesting of physical hydrological modules within the deep learning model, enabling it to follow basic physical laws such as water conservation and still possess strong physical meaning and interpretability. Attached Figure Description
[0028] Figure 1 This is a schematic flowchart of a method for constructing a smart hydrological model for watersheds in areas lacking reservoir operation data, as described in this embodiment of the invention.
[0029] Figure 2 This is a schematic architecture diagram of the training and verification of the intelligent hydrological model in an embodiment of the present invention.
[0030] Figure 3 This is a schematic architecture diagram of the training and verification of an intelligent hydrological model in another embodiment of the present invention.
[0031] Figure 4 This is a schematic diagram of the location of an example area in one embodiment of the present invention.
[0032] Figure 5 This is a schematic diagram illustrating the performance evaluation of the intelligent hydrological model, the pure deep learning model LSTM, and the conceptual hydrological model HBV in one embodiment of the present invention. Detailed Implementation
[0033] The technical solution of the present invention will now be described in detail with reference to the embodiments and accompanying drawings.
[0034] like Figure 1 As shown in the embodiments of the present invention, in a first aspect, a method for constructing a smart hydrological model for watersheds in areas lacking reservoir operation data is provided, comprising the following steps:
[0035] A conceptual hydrological model based on simplified reservoir operation rules is established. The conceptual hydrological model includes a physical hydrological module and a reservoir operation module. The physical hydrological module is used to calculate the inflow runoff based on the required physical parameters. The reservoir operation module is used to calculate the outflow runoff based on the input inflow runoff and the preset reservoir operation rules.
[0036] Collect watershed meteorological data, watershed natural attribute values, reservoir attribute values, and reservoir outflow data to form a sample set, and divide the sample set into a training set, a validation set, and a test set.
[0037] The first deep learning module is constructed to simulate the physical parameters required by the physical hydrology module based on watershed meteorological data, watershed natural attribute values, and reservoir attribute values.
[0038] The physical parameters required by the physical hydrology module in the output of the first deep learning module are input into the physical hydrology module, and the output of the physical hydrology module is input into the reservoir operation module to construct an intelligent hydrology model.
[0039] The intelligent hydrological model is trained using training and validation sets. The loss function is calculated based on the reservoir outflow data from the training and validation sets, and backpropagation is performed to optimize the parameters and hyperparameters of the first deep learning module, thus obtaining the final intelligent hydrological model.
[0040] It is understood that the intelligent hydrological model constructed in this embodiment combines a conceptual hydrological model based on simplified reservoir operation rules with a first deep learning module, and places it within a differentiable programmable framework. A schematic architecture diagram for its training and verification is shown below. Figure 2As shown, this design enables bidirectional coupling between the first deep learning model and the process-based conceptual hydrological model. Specifically, the framework uses a physical hydrological module as its physical foundation. The first deep learning module learns the nonlinear relationships between watershed meteorological data, watershed natural attribute values, reservoir attribute values, and the physical parameters required by the physical hydrological module. The gradients calculated by the conceptual hydrological model output are then used to update the weights of the first deep learning module, iteratively improving its performance. This supports global backpropagation of gradients in the intelligent hydrological model and joint optimization of the physical hydrological module, reservoir operation module, and deep learning module. Therefore, unlike the loosely coupled methods in previous studies (e.g., first calibrating the hydrological model to obtain simulated output, and then using it as input to retrain the deep learning model), the intelligent hydrological model proposed in this embodiment tightly integrates the physical hydrological module, reservoir operation module, and first deep learning module into a whole, and simultaneously optimizes and trains all three modules using gradient-based optimization methods within a differentiable computational framework, without requiring additional design and intermediate optimization steps. This end-to-end model architecture allows the model to be optimized solely based on the final outflow observations, without requiring difficult-to-obtain intermediate observation data (such as the actual parameters of the conceptual hydrological model). Since the first deep learning module is only used to learn the parameters of the conceptual hydrological model, this intelligent hydrological model retains the structure of the conceptual hydrological model as its basic framework, ensuring it conforms to physical constraints such as water balance. Simultaneously, it eliminates the reliance on large amounts of measured reservoir operation data required for modeling complex data-driven reservoir operation rules, providing a more effective solution for hydrological modeling and river runoff prediction in areas lacking reservoir operation data.
[0041] In some embodiments, to provide a feasible conceptual hydrological model based on simplified reservoir operation rules, the conceptual hydrological model based on simplified reservoir operation rules can be an HBV hydrological model that combines simplified reservoir operation rules; and the physical hydrological module can be an HBV module.
[0042] Understandably, the mathematical expression for the HBV hydrological model is:
[0043] ;
[0044] Where, x hydro θ represents the meteorological driving data of the HBV hydrological model, i.e., the watershed meteorological data; θ represents the required physical parameters and reservoir operation rules; and Q represents the reservoir outflow.
[0045] The HBV hydrological model is a conceptual hydrological model centered on the "precipitation-runoff" process. Through a structure of "climate input-runoff calculation-confluence simulation," it achieves a quantitative description and prediction of the watershed hydrological cycle. Its simplified reservoir operation modules (such as WRS and HNIRS simplified reservoir operation rules) serve as post-processing modules for the HBV hydrological model, converting the inflow simulated by the HBV module into the reservoir outflow. The reservoir operation module only requires the output runoff simulated by the HBV module (i.e., inflow runoff) as input, thus eliminating reliance on measured reservoir operation data.
[0046] The first deep learning module in the above embodiment is used to establish the nonlinear relationship between watershed meteorological data, watershed natural attribute values, reservoir attribute values, and the physical parameters required by the physical hydrology module. The obtained physical parameters are then directly input into the HBV hydrological model to simulate the reservoir outflow in a watershed regulated by a reservoir. This first deep learning module can employ a relatively mature long short-term memory artificial neural network (ENN). para The mathematical description is as follows:
[0047] ;
[0048] Where, x dl This refers to the meteorological driving data used in the first deep learning module, namely, watershed meteorological data, A. basin This represents watershed attribute factors, namely, watershed natural attribute values and reservoir attribute values. The first deep learning module, ENN. para The output is transformed to the range [0,1] by the sigmoid function, and then scaled to the predefined parameter range by the Min-Max method to obtain the final required physical parameters and reservoir operation rules.
[0049] In some embodiments, to provide a feasible required physical parameter, the required physical parameter may include snowmelt path, soil path, response path, and runoff path, etc.
[0050] It is understandable that the physical parameters required for a conceptual hydrological model may vary depending on the actual situation. The physical parameters proposed in the above embodiments are only the more commonly used physical parameters.
[0051] In some embodiments, to compensate for the limitations of applying simplified reservoir operation rules to complex environments, the soil path may include parameters controlling the conversion of soil moisture to net rainfall, such as soil field holding capacity FC and the contribution rate of precipitation to runoff β. The parameters FC and β that control the conversion of soil moisture to net rainfall may be dynamic parameters, while the remaining required physical parameters may be static parameters, and the simulated reservoir operation rules may be dynamic parameters.
[0052] Understandably, the simplified reservoir operation rules employed have limitations in complex environments. Setting the parameters FC and β, which represent the conversion of soil moisture to net rainfall, as dynamic parameters, along with the simulated reservoir operation rules, can better meet the prediction requirements under complex and changing environments. Of course, with a large amount of data, other physical parameters can also be set as dynamic parameters. If the amount of data is small, and to improve training speed and relax the prediction accuracy of the final intelligent hydrological model, all necessary physical parameters and simulated reservoir operation rules can be set as static parameters.
[0053] In some embodiments, to provide feasible watershed meteorological data, watershed natural attributes, and watershed reservoir attributes, the watershed meteorological data may include precipitation data, average temperature data, and potential evapotranspiration data; the watershed natural attributes may include drought index data, watershed area data, and heavy precipitation frequency data; and the watershed reservoir attributes may include reservoir dam height data, maximum reservoir capacity data, regulation degree data, and main uses data.
[0054] It is understood that in this embodiment, the collected watershed reservoir attributes may not require sensitive data such as real-time reservoir regulation data, thereby reducing the collection of actual reservoir operation data.
[0055] In some embodiments, such as Figure 3 As shown, to further improve the performance of the constructed intelligent hydrological model, a second deep learning module may also be constructed. The second deep learning module is used to simulate the reservoir outflow based on the inflow runoff, watershed meteorological data, watershed natural attribute values and reservoir attribute values output by the physical hydrological module.
[0056] The reservoir operation module was replaced with a second deep learning module.
[0057] The intelligent hydrological model is trained using training and validation sets. A loss function is calculated based on the reservoir outflow data from the training and validation sets, and backpropagation is performed to optimize the parameters and hyperparameters of the first deep learning module. The final intelligent hydrological model includes:
[0058] The intelligent hydrological model is trained using training and validation sets. The loss function is calculated based on the reservoir outflow data from the training and validation sets, and backpropagation is performed to optimize the parameters and hyperparameters of the first and second deep learning modules, thus obtaining the final intelligent hydrological model.
[0059] It is understood that the reservoir operation module was replaced by the second deep learning module in the above embodiment. This is because this embodiment focuses on the simulation of the watershed regulated by the reservoir, and the deep learning module can also be used as a replacement for any other internal physical module in the conceptual hydrological model, which will not be described in detail here.
[0060] In this embodiment, the second deep learning module can also employ a relatively mature long short-term memory artificial neural network. Therefore, this second deep learning module (ENN) reservoir The mathematical description is as follows:
[0061] ;
[0062] Among them, Q in,sim The inflow runoff output by the physical hydrology module; Q outflow This indicates the outflow rate of the reservoir.
[0063] Since the reservoir operation module was replaced by a second deep learning module, when training the intelligent hydrological model using training and validation sets, calculating the loss function based on the reservoir outflow data of the training and validation sets, and performing backpropagation to optimize the parameters and hyperparameters of the first deep learning module to obtain the final intelligent hydrological model, it is also necessary to optimize the parameters and hyperparameters of the second deep learning module.
[0064] In some embodiments, for ease of subsequent use, the intelligent hydrological model can use the watershed meteorological data in the training set samples and validation set samples as input values, and input the watershed natural attribute values and reservoir attribute values in the training set samples, and use the reservoir outflow data in the training set samples and validation set samples as output values.
[0065] When using the intelligent hydrological model, the outflow of water from the reservoir can be obtained by inputting meteorological data from the watershed.
[0066] It is understood that the above embodiments clearly define the specific process of training and validating the intelligent hydrological model, enabling the final intelligent hydrological model to obtain reservoir outflow by inputting only watershed meteorological data. That is, during the training processes of the first and second deep learning modules, watershed meteorological data from the training and validation sets are used as input values, along with watershed natural attribute values and reservoir attribute values from the training set samples, and reservoir outflow data from the training and validation sets are used as output values.
[0067] In some embodiments, to evaluate the hydrological simulation accuracy of the intelligent hydrological model, the following may also be included:
[0068] The watershed meteorological data from the test set sample is input into the final intelligent hydrological model to obtain the simulated reservoir outflow. The simulated reservoir outflow is then compared with the reservoir outflow in the test set sample to evaluate the accuracy of the intelligent hydrological model's hydrological simulation.
[0069] It is understandable that using test set samples to evaluate the hydrological simulation accuracy of intelligent hydrological models is a common technique in existing technologies, which will not be elaborated here.
[0070] The above embodiments are illustrated below with specific examples:
[0071] This embodiment takes the Hanjiang River basin upstream of the Danjiangkou Reservoir as an example. It models the regulated water cycle process upstream to simulate the outflow from the Danjiangkou Reservoir. The simulation timescale is daily. The training, validation, and testing periods for the constructed intelligent hydrological model in the basin upstream of Danjiangkou were January 1, 1994 – December 31, 2001; January 1, 2002 – December 31, 2005; and January 1, 2006 – December 31, 2013, respectively. Figure 4 A schematic diagram showing the location of the area in this embodiment is displayed.
[0072] The intelligent hydrological model in this embodiment is constructed using the methods described in the above embodiments.
[0073] The application effect of the intelligent hydrological model in this embodiment will be compared with the commonly used purely physics-based conceptual hydrological model HBV and the purely deep learning model LSTM under the same condition of lacking reservoir operation data. To fully verify the performance of the intelligent hydrological model in this embodiment and the impact of different reservoir operation rules, three watershed intelligent hydrological models were constructed on the framework of the above embodiments, named Hybrid_WRS, Hybrid_HNIRS, and Hybrid_ENN, respectively. The reservoir operation modules used in these models correspond to WRS (simplified reservoir operation rules), HNIRS (simplified reservoir operation rules), and the model using a second deep learning module, respectively. The reservoir scheduling rules are adaptively learned. These three reservoir operation modules require little to no reservoir operation data, making them suitable for performance comparisons under conditions of scarce reservoir operation data.
[0074] Simultaneously, conceptual hydrological models HBV_WRS and HBV_HNIRS, based on the same simplified reservoir operation rules, were established and their performance compared with the intelligent hydrological model in this embodiment. In this embodiment, the intelligent hydrological model, the pure deep learning model LSTM, and the hydrological models HBV_WRS and HBV_HNIRS all use the outflow from the Danjiangkou Reservoir as the target variable for training or calibration. To reduce the uncertainty caused by model parameter initialization, their performance comparisons are based on the results of repeated computations with 50 different random seeds.
[0075] Figure 5The performance evaluation results of the intelligent hydrological model, the pure deep learning model LSTM, and the conceptual hydrological model HBV are presented in the simulation of the outflow from the Danjiangkou Reservoir on the Hanjiang River using 50 random seeds, both overall and at different runoff levels. The comparison results of the NSE and KGE indices show that, regardless of which method is used as the reservoir operation module of the intelligent hydrological model, its overall reservoir outflow simulation performance is superior to that of LSTM and the conceptual hydrological model HBV. Furthermore, compared to the pure deep learning model LSTM and the conceptual hydrological model HBV, the intelligent hydrological model can simulate high, medium, and low outflows with higher accuracy. For the median mean of the high-flow evaluation index HFAB, the intelligent hydrological model improved the accuracy by 69.95% and 34.12% compared to the conceptual hydrological model and the pure deep learning model, respectively; for the medium-flow evaluation index BiasFMS, the intelligent hydrological model improved the accuracy by 51.91% and 22.93% compared to the hydrological model and the pure deep learning model, respectively; and for the low-flow evaluation index LogNSE, the intelligent hydrological model improved the accuracy by 195.00% and 21.44% compared to the conceptual hydrological model and the pure deep learning model, respectively.
[0076] In a second aspect of the embodiments proposed in this invention, a readable storage medium is provided, storing the final intelligent hydrological model obtained by the above-described method for constructing intelligent hydrological models for watersheds in areas lacking reservoir operation data.
Claims
1. A method for constructing a smart hydrological model for watersheds in areas lacking reservoir operation data, characterized in that: Includes the following steps: A conceptual hydrological model based on simplified reservoir operation rules is established. The conceptual hydrological model includes a physical hydrological module and a reservoir operation module. The physical hydrological module is used to calculate the inflow runoff based on the required physical parameters. The reservoir operation module is used to calculate the outflow runoff based on the input inflow runoff and the preset reservoir operation rules. Collect watershed meteorological data, watershed natural attribute values, reservoir attribute values, and reservoir outflow data to form a sample set, and divide the sample set into a training set, a validation set, and a test set. The first deep learning module is constructed to simulate the physical parameters required by the physical hydrology module based on watershed meteorological data, watershed natural attribute values, and reservoir attribute values. The physical parameters required by the physical hydrology module in the output of the first deep learning module are input into the physical hydrology module, and the output of the physical hydrology module is input into the reservoir operation module to construct an intelligent hydrology model. The intelligent hydrological model is trained using training and validation sets. The loss function is calculated based on the reservoir outflow data from the training and validation sets, and backpropagation is performed to optimize the parameters and hyperparameters of the first deep learning module, thus obtaining the final intelligent hydrological model.
2. The method for constructing a smart watershed hydrological model applicable to areas lacking reservoir operation data as described in claim 1, characterized in that, The conceptual hydrological model based on simplified reservoir operation rules is the HBV hydrological model that combines simplified reservoir operation rules; the physical hydrological module is the HBV module.
3. The method for constructing a smart watershed hydrological model applicable to areas lacking reservoir operation data as described in claim 1, characterized in that, The required physical parameters include snowmelt path, soil path, response path, and runoff path.
4. The method for constructing a smart watershed hydrological model applicable to areas lacking reservoir operation data as described in claim 3, characterized in that, The soil pathway includes parameters that control the conversion of soil moisture to net rainfall, namely, soil field holding capacity and the contribution rate of precipitation to runoff. These parameters are dynamic parameters, while the remaining required physical parameters are static parameters. The simulated reservoir operation rules are dynamic parameters.
5. The method for constructing a smart hydrological model for watersheds in areas lacking reservoir operation data as described in claim 1, characterized in that, The basin meteorological data includes precipitation data, average temperature data, and potential evapotranspiration data; the basin natural attributes include drought index data, basin area data, and frequency of heavy precipitation data; the basin reservoir attributes include dam height data, maximum reservoir capacity data, regulation degree data, and main uses data.
6. The method for constructing a smart hydrological model for watersheds in areas lacking reservoir operation data as described in claim 1, characterized in that, It also includes the construction of a second deep learning module, which is used to simulate the reservoir outflow based on the inflow runoff, watershed meteorological data, watershed natural attribute values and reservoir attribute values output by the physical hydrology module; The reservoir operation module was replaced with a second deep learning module. The intelligent hydrological model is trained using training and validation sets. A loss function is calculated based on the reservoir outflow data from the training and validation sets, and backpropagation is performed to optimize the parameters and hyperparameters of the first deep learning module. The final intelligent hydrological model includes: The intelligent hydrological model is trained using training and validation sets. The loss function is calculated based on the reservoir outflow data from the training and validation sets, and backpropagation is performed to optimize the parameters and hyperparameters of the first and second deep learning modules, thus obtaining the final intelligent hydrological model.
7. The method for constructing a smart hydrological model for watersheds in areas lacking reservoir operation data as described in claim 1, characterized in that, During training, the intelligent hydrological model takes the watershed meteorological data in the training set and validation set as input values, and incorporates the watershed natural attribute values and reservoir attribute values in the training set, and takes the reservoir outflow data in the training set and validation set as output values. When using the intelligent hydrological model, the outflow of water from the reservoir can be obtained by inputting meteorological data from the watershed.
8. The method for constructing a smart watershed hydrological model applicable to areas lacking reservoir operation data as described in claim 1, characterized in that, Also includes: The watershed meteorological data from the test set sample is input into the final intelligent hydrological model to obtain the simulated reservoir outflow. The simulated reservoir outflow is then compared with the reservoir outflow in the test set sample to evaluate the accuracy of the intelligent hydrological model's hydrological simulation.
9. The method for constructing a smart watershed hydrological model applicable to areas lacking reservoir operation data, as described in any one of claims 1-8, is characterized in that... Both the first deep learning module and / or the second deep learning module employ long short-term memory artificial neural networks.
10. A readable storage medium, characterized in that, The final intelligent hydrological model is stored in the method for constructing a watershed intelligent hydrological model applicable to areas lacking reservoir operation data as described in any one of claims 1-9.