Distributed hydrological model hydrological parameter gridding method based on deep learning

Abstract

The invention relates to a distributed hydrological model hydrological parameter gridding method based on deep learning, and the method comprises the following steps: collecting the feature data of regional hydrological parameters based on a VIC model, and carrying out the preprocessing of the data, and forming a data set; constructing a multi-layer perceptron (MLP) neural network model, and establishing a mapping relation between an input variable and an optimization parameter; loss changes in the training process are monitored, and the prediction precision and deviation distribution of the model are evaluated through a scatter diagram contrast ratio fixed value and calculated value; original hydrological parameters in the model parameters are updated to hydrological parameters automatically calibrated based on machine learning, the flood process of the research area is simulated and analyzed, and then the optimized parameters are verified through Nash efficiency coefficients. According to the method, a traditional manual parameter adjusting mode is replaced, the complex nonlinear relation between the soil characteristic data and the environment variables is deeply excavated, the model parameters are automatically adjusted, the simulation precision of the hydrological model is improved, and the spatial distribution reasonability and continuity of the hydrological parameters are improved.

Description

Technical Field

[0001] The present invention belongs to the field of hydrological forecasting, and in particular relates to a hydrological parameter gridding method of a distributed hydrological model based on deep learning. Background Art

[0002] Hydrological models are widely used in hydrological research, flood and drought prevention, and water resource assessment and development. Model parameters, as crucial components of hydrological models, have values ​​whose accuracy is closely linked to simulation accuracy. The VIC (Variable Infiltration Capacity) model is a large-scale distributed hydrological model based on the Soil Vegetation Atmospheric Transfer Schemes (SVATS) concept. Simulating hydrological processes involves a large number of physical and environmental parameters, including climate and geographic parameters, vegetation parameters, soil characteristics, and hydrological parameters. However, parameter calibration of the VIC model typically relies on extensive historical data and manual experience, making the adjustment process time-consuming and difficult to guarantee accuracy.

[0003] In recent years, automated calibration methods based on mathematical algorithms, such as optimization algorithms and Bayesian inference, have reduced manual intervention and adjusted model parameters through automated calibration. However, these algorithm-based methods still suffer from high computational complexity, weak generalization capabilities, and a strong reliance on initial parameters. This results in an inability to obtain optimal parameter combinations in complex underlying surfaces and hydrological environments, especially in areas lacking historical data, where high accuracy remains difficult to guarantee. By learning from large-scale data, machine learning algorithms automatically extract features and association patterns, providing more accurate simulation results in data-scarce areas and capable of handling more complex hydrological processes. This effectively improves the automated calibration of hydrological model parameters and potentially achieves stronger generalization capabilities. Summary of the Invention

[0004] The technical problem of the present invention is: by dividing the study area into multiple grids, different hydrological model parameters are set for each grid separately to reflect the heterogeneity within the region; and based on the complex terrain, diverse climate and significant differences in the underlying surface environment, considering the complex nonlinear relationship between soil characteristic data and environmental factors, through a data-driven approach, the complex nonlinear relationships between various parameters are learned from historical hydrological data, and automatic calibration is performed to improve the physical rationality of the spatial distribution of parameters, enhance the model's adaptability to complex underlying surfaces and changeable climatic conditions, and improve the simulation accuracy of distributed hydrological models.

[0005] The purpose of the present invention is to solve the above problems. The present invention provides a method for gridding hydrological parameters of a distributed hydrological model based on deep learning, comprising the following steps: S1. Collect characteristic data of hydrological parameters in different regions based on the VIC model and preprocess the data to form a data set; S2. Build a multi-layer perceptron (MLP) neural network model and establish a mapping relationship between input variables and optimization parameters; S3. Monitor the loss changes during training and compare the fixed and calculated values ​​through scatter plots to evaluate the prediction accuracy and bias distribution of the model. S4. Update the original hydrological parameters in the model parameters to hydrological parameters automatically calibrated based on machine learning, simulate and analyze the flood process in the study area, and then verify the optimized parameters through the Nash efficiency coefficient.

[0006] Furthermore, the characteristic data of the hydrological parameters in step S1 include soil characteristic data and climate characteristic data of the optimized zoning of the study area; the soil characteristic data include hydrological parameters and soil parameters; the hydrological parameters include saturated capacity curve shape parameters and the maximum value of base flow of the bottom soil within 24 hours; the soil parameters include soil saturated hydraulic conductivity change index, soil saturated hydraulic conductivity, soil water diffusion coefficient, critical water content ratio and wilting water content ratio parameters; the climate characteristic data include annual water surface evaporation and annual dryness characteristic data.

[0007] Furthermore, step S1 includes matching the soil characteristic data grid with the climate characteristic values, including the following sub-steps: 1) Obtaining grid longitude and latitude information of soil characteristic data; 2) Calculate the geographic distance between each soil grid cell and the climate data point; 3) Extract climate characteristic data of the climate data point closest to the soil grid location and integrate them with the soil characteristic data to generate a high-resolution parameter dataset.

[0008] Preferably, preprocessing the data set includes normalizing the data set and scaling the data to within an interval to eliminate dimensional differences.

[0009] Preferably, a multi-layer perceptron (MLP) neural network model is constructed, comprising 1 input layer, 2 hidden layers, and 1 output layer; neurons in the hidden layer use a ReLU activation function; and neurons in the output layer use a Sigmoid activation function.

[0010] Furthermore, step S2 includes the following sub-steps: 1) Randomly shuffle the data, divide the dataset into training and validation sets proportionally, set the batch size and maximum number of training iterations, select the optimizer, and automatically adjust the learning rate; 2) For each hydrological parameter to be optimized, the mapping relationship between the parameter and the environmental characteristics is constructed one by one, and independent MLP models are trained for each parameter. Preferably, a mapping relationship between parameters and environmental characteristics is constructed, historical sample values ​​of target parameters are extracted from the preprocessed data set, and other soil characteristic data and climate characteristic data are used as input to train the MLP model to learn the nonlinear relationship between B and environmental variables.

[0011] Furthermore, step S3 includes the following sub-steps: 1) Set the loss function and use the mean absolute error to quantify the deviation between the predicted value and the true value; 2) Denormalize the prediction results of the MLP neural network model in the test and obtain the original scale data. Compare the fixed value and the calculated value through scatter plot to evaluate the model's fitting ability and analyze the optimization parameters, and then output the optimized parameter set of the feature variables; 3) Each hydrological parameter to be optimized corresponds to a trained MLP model. The model is saved in a specified format and can be called at any time. For each grid in the study area, the trained MLP model is called. Combined with the soil and climate characteristics data of the grid, the optimized parameters are calculated and generated one by one, forming an optimized soil parameter dataset for the entire area, which provides subsequent input for VIC model simulation.

[0012] Furthermore, step S4 includes the following sub-steps: 1) Analyze the distribution characteristics and rationality of the optimized hydrological parameters. The rationality analysis includes the changing trend of the spatial distribution of the parameters before and after optimization, and evaluates the high-value and low-value areas of the parameters and the uniformity of the overall distribution. The spatial distribution of the parameters is then simulated in combination with the underlying surface characteristics of topography and land use. 2) The soil data set is updated through the hydrological parameters automatically calibrated by machine learning, input into the VIC model to simulate the flood process in the study area, and the hydrological process of precipitation and runoff is analyzed in combination with the optimized parameters.

[0013] Preferably, step S4 verifies the optimization parameters through the Nash efficiency coefficient, including calculating the Nash efficiency coefficient and the average relative error, evaluating the consistency between the VIC model simulation results and the measured flood process, and then quantifying the effect of the optimization parameters on the simulation accuracy.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1) The present invention provides a distributed hydrological model hydrological parameter gridding method based on deep learning, which replaces the traditional manual parameter adjustment method, deeply explores the complex nonlinear relationship between soil characteristic data and environmental variables, automatically adjusts model parameters, improves the simulation accuracy of the hydrological model, and increases the rationality and continuity of the spatial distribution of hydrological parameters.

[0015] 2) The present invention provides a method for gridding hydrological parameters of a distributed hydrological model based on deep learning. By optimizing the parameters through machine learning and performing parameter gridding processing according to the actual conditions of different regions, the simulation effect, data processing efficiency and adaptability are improved, and the accuracy and automation level of hydrological simulation are enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below with reference to the accompanying drawings and examples.

[0017] Figure 1 This is a flow chart of a method for gridding hydrological parameters of a distributed hydrological model based on deep learning according to an embodiment of the present invention.

[0018] Figure 2 A schematic diagram of the structure of an MLP neural network constructed in an embodiment of the present invention.

[0019] Figure 3 This is a scatter diagram of the machine learning results of an embodiment of the present invention.

[0020] Figure 4 Schematic diagram of the loss change of the MLP model in an embodiment of the present invention.

[0021] Figure 5 Graphs showing the spatial distribution of characteristic parameters before and after updating according to an embodiment of the present invention.

[0022] Figure 6 This is a schematic diagram comparing simulation results before and after optimization of machine learning parameters for the interval partitioning above Gangtuo in an embodiment of the present invention.

[0023] Figure 7 This is a schematic diagram comparing simulation results before and after machine learning parameter optimization for the Wudongde-Xiangjiaba interval partitioning according to an embodiment of the present invention.

[0024] Figure 8 Schematic diagram comparing simulation results before and after machine learning parameter optimization for the Minjiang River Basin region according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The technical solution of the present invention will be described in further detail below with reference to the accompanying drawings. It should be noted that the drawings are all in very simplified form and are only used to conveniently and clearly assist in illustrating the purpose of the embodiments of the present invention.

[0026] like Figure 1 As shown, a method for gridding hydrological parameters of a distributed hydrological model based on deep learning includes the following steps: S1. Collect characteristic data of hydrological parameters in different regions based on the VIC model and preprocess the data to form a data set; The data sources used are: (1) Geographical topographic data: Digital elevation model (DEM) with a resolution of 30 m in China was used, which was obtained from the cloud data website of the Chinese Academy of Sciences. Vegetation type data was obtained from the AVHRR Global Land Cover Classification dataset provided by the University of Maryland. Landsat 8 OLI data with a relatively high spatial resolution was used for extracting underlying surface information.

[0027] (2) Soil data: The soil parameter values ​​are set based on the 10 km × 10 km resolution global-scale soil database published by Reynolds et al.

[0028] (3) Meteorological data: The hourly maximum and minimum temperature and rainfall data of the study area from 2013 to 2024 provided by the Information Center of the Ministry of Water Resources were used; the radiation and wind speed data were generated using the mtclim model.

[0029] (4) Hydrological data: The measured hourly flow data for the upper and middle reaches of the Yangtze River Basin during the flood season from 2014 to 2024 provided by the Information Center of the Ministry of Water Resources.

[0030] As shown in Table 1, the characteristic data of hydrological parameters in step S1 include the soil characteristic data and climate characteristic data of the optimized zoning of the study area; the soil characteristic data include hydrological parameters and soil parameters; the hydrological parameters include the saturation capacity curve shape parameter B and the maximum base flow Dm of the bottom soil within 24 hours; the soil parameters include the soil saturated hydraulic conductivity variation index EXPT, the soil saturated hydraulic conductivity Ksat, the soil water diffusion coefficient PHI, the critical water content ratio WcrFT and the wilting water content ratio parameter WpFT; the climate characteristic data include the annual water surface evaporation E and the annual dryness E / P characteristic data.

[0031] Table 1

[0032] Step S1 involves matching the soil characteristic data grid with the climate characteristic values, and includes the following sub-steps: 1) Obtaining grid longitude and latitude information of soil characteristic data; 2) Calculate the geographic distance between each soil grid cell and the climate data point; 3) Extract climate characteristic data of the climate data point closest to the soil grid location and integrate them with the soil characteristic data to generate a 0.05°×0.05° high-resolution parameter dataset.

[0033] The dataset is preprocessed, including applying the MinMaxScaler function to all input features and target parameters for normalization, scaling the data to the range [0, 1] to eliminate dimensional differences.

[0034] The calculation formula of the MinMaxScaler function is: ; Where, represents the normalized value, X represents the initial input data, Indicates the minimum value in the feature data, Indicates the maximum value in the feature data.

[0035] S2. Build a multi-layer perceptron (MLP) neural network model and establish a mapping relationship between input variables and optimization parameters; like Figure 2 As shown in Figure 3, the constructed multi-layer perceptron (MLP) neural network model consists of one input layer, two hidden layers, and one output layer. The input layer receives normalized soil characteristic data and climate characteristic values, and the hidden layers are configured with 512 and 64 neurons, respectively, both using the ReLU activation function. The output layer has one neuron and uses the Sigmoid activation function.

[0036] Step S2 includes the following sub-steps: 1) Randomly shuffle the data and split it into training and validation sets in a ratio of 9:1. Set the batch size to 214 and the maximum number of training iterations to 10. Select the RMSprop optimizer and automatically adjust the learning rate. 2) For each of the six hydrological parameters to be optimized—the variable infiltration curve parameter B, the maximum baseflow velocity Dm, the Dm fraction Ds at the onset of nonlinear baseflow, the maximum soil moisture fraction Ws when nonlinear baseflow occurs, the thickness d1 of the second soil water layer, and the thickness d2 of the third soil water layer—we constructed a mapping relationship between each parameter and environmental characteristics, and trained independent MLP models for each parameter. like Figure 3 As shown in the figure, taking the capacity curve shape parameter B as an example, the historical sample values ​​of the target parameter B are extracted from the preprocessed data set, and other soil characteristic data and climate characteristic data are used as input to train the MLP model to learn the nonlinear relationship between B and environmental variables. During the training process of parameter B, the other five parameters to be optimized will not be used as input features in the training to ensure that each parameter is calibrated independently.

[0037] The mapping relationship between parameters and environmental characteristics is constructed, the historical sample values ​​of the target parameters are extracted from the preprocessed data set, and other soil characteristic data and climate characteristic data are used as input to train the MLP model to learn the nonlinear relationship between B and environmental variables.

[0038] S3. Monitor the loss changes during training and compare the fixed and calculated values ​​through scatter plots to evaluate the prediction accuracy and bias distribution of the model. Step S3 includes the following sub-steps: 1) Set the loss function and use the mean absolute error to quantify the deviation between the predicted value and the true value; by monitoring the MAE change curves of the training set and the validation set, the convergence trend of the model can be observed to ensure the stability of the training process. When both curves show a trend of steady decline and eventual convergence, it means that the model has good fitting ability. If the loss of the training set continues to decline and the loss of the validation set shows a platform or rebound, it indicates that the model may be overfitting. On the contrary, if both curves converge to a higher loss value too early, it reflects that the model is at risk of underfitting. Taking the parameter B that reflects the shape index of the saturated water storage capacity curve as an example, the loss changes as shown in the figure below. Figure 4 As shown: The loss function is calculated as: ; In the formula, i represents the counting unit, n represents the number of samples, represents the model prediction value, Represents the true value.

[0039] 2) The prediction results of the denormalized MLP neural network model on the test and on the original scale data are obtained. The fixed value and the calculated value are compared through the scatter plot. The fitting ability of the model is evaluated and the optimization parameters are analyzed through the R² score. Then, the optimized parameter set of the characteristic variables B, Ds, Dm, Ws, d1 and d2 is output, and the determination coefficient R 2 It is used to measure the degree of fit between the hydrological parameters predicted by the MLP model and the hydrological parameters determined by the zoning, reflecting the ability of the model to explain the changes in variables. The closer the R² value is to 1, the stronger the model fitting ability is, and the higher the correlation between the predicted parameters and the true parameters is. The present invention believes that R²>0.5 can indicate that the model has good predictive ability. Scatter plot as shown Figure 4 As shown, the calculation formula for the evaluation score R² is: ; Where, n represents the number of samples, y i represents the i-th true value, represents the i-th predicted value, Represents the mean of the true values.

[0040] 3) If Figure 5As shown, each hydrological parameter to be optimized corresponds to a trained MLP model. The trained model is saved in the Keras-specified format as an H5 file for immediate access. The model records the network structure, weights, and training configuration information. The trained MLP model is then called for each grid in the study area. Combined with the soil and climate data for that grid, the optimized B, Ds, Dm, Ws, d1, and d2 parameters are calculated and generated one by one. This creates a dataset of optimized soil parameters for the entire region, providing subsequent input for VIC model simulations.

[0041] S4. Update the original hydrological parameters in the model parameters to hydrological parameters automatically calibrated based on machine learning, simulate and analyze the flood process in the study area, and then verify the optimized parameters through the Nash efficiency coefficient.

[0042] Step S4 includes the following sub-steps: 1) Analyze the distribution characteristics and rationality of the optimized hydrological parameters. The rationality analysis includes the changing trend of the spatial distribution of the parameters before and after optimization, and evaluates the high-value and low-value areas of the parameters and the uniformity of the overall distribution. The spatial distribution of the parameters is then simulated in combination with the underlying surface characteristics of topography and land use. 2) The soil data set is updated through the hydrological parameters automatically calibrated by machine learning, input into the VIC model to simulate the flood process in the study area, and the hydrological process of precipitation and runoff is analyzed in combination with the optimized parameters.

[0043] In step S4, the optimization parameters are verified by the Nash efficiency coefficient, including calculating the Nash efficiency coefficient and the average relative error, evaluating the consistency between the VIC model simulation results and the measured flood process, and then quantifying the effect of the optimization parameters on the simulation accuracy.

[0044] The Nash efficiency coefficient is calculated as follows: ; Where, represents the observation value at time t, represents the simulation value at time t, It represents the average value of the observation value, that is, the average flow rate during the entire observation period, n represents the total number of moments of observation data, and t represents the counting unit.

[0045] like Figures 6 to 8 As shown in the figure, the interval division above Gangtuo, the interval division between Wudongde and Xiangjiaba, and the Minjiang River basin division are illustrated before and after the machine learning parameter optimization. Figures 6 to 8In the figure, the left figure is a schematic diagram before machine learning parameter optimization, and the right figure is a schematic diagram before machine learning parameter optimization. As shown in Table 2, the Nash efficiency coefficient and relative error of the simulation results of typical regions can reflect that the parameters optimized based on deep learning show better flood simulation capabilities in some areas, especially in flood peak identification and process trends. The parameter optimization scheme constructed by the present invention can greatly improve the efficiency of parameter calibration and reduce labor costs when facing large-scale high-resolution watershed modeling tasks. It is more reflected in the rapid acquisition of parameter sets with good simulation accuracy under limited time and computing resources, thereby significantly improving the efficiency of model construction and deployment.

[0046] Table 2

[0047] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions recited in the claims, including equivalent alternatives to the technical features of the technical solutions recited in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for gridding hydrological parameters of a distributed hydrological model based on deep learning, characterized in that: The following steps are involved: S1. Collect characteristic data of hydrological parameters in different regions based on the VIC model and preprocess the data to form a data set; S2. Build a multi-layer perceptron (MLP) neural network model and establish a mapping relationship between input variables and optimization parameters; S3. Monitor the loss changes during training and compare the fixed and calculated values ​​through scatter plots to evaluate the prediction accuracy and bias distribution of the model. S4. Update the original hydrological parameters in the model parameters to hydrological parameters automatically calibrated based on machine learning, simulate and analyze the flood process in the study area, and then verify the optimized parameters through the Nash efficiency coefficient.

2. The method for gridding hydrological parameters of a distributed hydrological model based on deep learning according to claim 1, characterized in that: The characteristic data of the hydrological parameters in step S1 include soil characteristic data and climate characteristic data of the optimized zoning of the study area; the soil characteristic data include hydrological parameters and soil parameters; the hydrological parameters include saturation capacity curve shape parameters and the maximum value of base flow of the bottom soil within 24 hours; the soil parameters include soil saturated hydraulic conductivity change index, soil saturation, soil water diffusion coefficient, critical water content ratio and wilting water content ratio parameters; the climate characteristic data include annual water surface evaporation and annual dryness characteristic data.

3. The method for gridding hydrological parameters of a distributed hydrological model based on deep learning according to claim 2 is characterized in that: The step S1 includes matching the soil characteristic data grid with the climate characteristic value, including the following sub-steps: 1) Obtaining grid longitude and latitude information of soil characteristic data; 2) Calculate the geographic distance between each soil grid cell and the climate data point; 3) Extract climate characteristic data of the climate data point closest to the soil grid location and integrate them with the soil characteristic data to generate a high-resolution parameter dataset.

4. The method for gridding hydrological parameters of a distributed hydrological model based on deep learning according to claim 3 is characterized in that: The preprocessing of the data set includes normalizing the data set and scaling the data to within an interval to eliminate dimensional differences.

5. The method for gridding hydrological parameters of a distributed hydrological model based on deep learning according to claim 1, characterized in that: The multi-layer perceptron (MLP) neural network model comprises an input layer, two hidden layers, and an output layer; the neurons in the hidden layer use a ReLU activation function; and the neurons in the output layer use a Sigmoid activation function.

6. The method for gridding hydrological parameters of a distributed hydrological model based on deep learning according to claim 1, characterized in that: The step S2 includes the following sub-steps: 1) Randomly shuffle the data, divide the dataset into training and validation sets proportionally, set the batch size and maximum number of training iterations, select the optimizer and automatically adjust the learning rate; 2) For the hydrological parameters to be optimized, the mapping relationship between the parameters and environmental characteristics is constructed one by one, and independent MLP models are trained for each parameter.

7. The method for gridding hydrological parameters of a distributed hydrological model based on deep learning according to claim 6, characterized in that: The mapping relationship between the constructed parameters and the environmental characteristics is extracted from the preprocessed data set, and the historical sample values ​​of the target parameters are used as input to train the MLP model to learn the nonlinear relationship between B and the environmental variables.

8. The method for gridding hydrological parameters of a distributed hydrological model based on deep learning according to claim 1, characterized in that: The step S3 includes the following sub-steps: 1) Set the loss function and use the mean absolute error to quantify the deviation between the predicted value and the true value; 2) Denormalize the prediction results of the MLP neural network model in the test and obtain the original scale data. Compare the fixed value and the calculated value through scatter plot to evaluate the model's fitting ability and analyze the optimization parameters, and then output the optimized parameter set of the feature variables; 3) Each hydrological parameter to be optimized corresponds to a trained MLP model. The model is saved in a specified format and can be called at any time. For each grid in the study area, the trained MLP model is called. Combined with the soil and climate characteristics data of the grid, the optimized parameters are calculated and generated one by one, forming an optimized soil parameter dataset for the entire area, which provides subsequent input for VIC model simulation.

9. The method for gridding hydrological parameters of a distributed hydrological model based on deep learning according to claim 1, characterized in that: The step S4 includes the following sub-steps: 1) Analyze the distribution characteristics and rationality of the optimized hydrological parameters. The rationality analysis includes the changing trends of the spatial distribution of the parameters before and after optimization, and evaluate the high-value and low-value areas of the parameters and the uniformity of the overall distribution. Then, the spatial distribution of the parameters is simulated in combination with the underlying surface characteristics of topography and land use. 2) The soil data set is updated through the hydrological parameters automatically calibrated by machine learning, input into the VIC model to simulate the flood process in the study area, and the hydrological process of precipitation and runoff is analyzed in combination with the optimized parameters.

10. The method for gridding hydrological parameters of a distributed hydrological model based on deep learning according to claim 1, characterized in that: Step S4, verifying the optimization parameters through the Nash efficiency coefficient, includes calculating the Nash efficiency coefficient and the average relative error, evaluating the consistency between the VIC model simulation results and the measured flood process, and then quantifying the effect of the optimization parameters on the simulation accuracy.

Images