WO2022032874A1

WO2022032874A1 - Adversarial neural network-based hydrological parameter calibration method for data region

Info

Publication number: WO2022032874A1
Application number: PCT/CN2020/123715
Authority: WO
Inventors: 李胜; 张�荣; 刘晟一; 田彪; 丁交亮; 彭江江; 刘继军
Original assignee: 贵州东方世纪科技股份有限公司
Priority date: 2020-08-14
Filing date: 2020-10-26
Publication date: 2022-02-17
Also published as: CN111914488B; CN111914488A

Abstract

An adversarial neural network-based hydrological parameter calibration method for a data region, comprising: collecting soil texture, vegetation coverage, a land utilization rate, topographic data, a runoff coefficient, total annual evaporation, and gradient and slope data; dividing a calibration region into computing units of less than 30 square kilometers; determining each underlying surface parameter of each computing unit and a meteorological related factor; and performing automatic calibration on hydrological parameters of a data watershed by using an adversarial neural network GAN to obtain an optimal hydrological parameter of each computing unit. The technical problems in the prior art such as the work is repetitive, the efficiency is low, the complexity is extremely high, and the application and popularization of a hydrologic model are not utilized are solved.

Description

A method for calibration of hydrological parameters in areas with data based on adversarial neural network

technical field

The invention belongs to the hydrological parameter calibration technology, and in particular relates to a hydrological parameter calibration method based on an adversarial neural network in an area with data.

Background technique

Hydrological models play an important role in the study of hydrological laws and in solving practical problems in production. With the rapid development of modern science and technology, information technology with computers and communications as the core is widely used in the fields of hydrology, water resources and hydraulic engineering. The research on hydrological models has developed rapidly, and is widely used in the study of basic hydrological laws, flood and drought disaster prevention, water resources evaluation and development and utilization, water environment and ecosystem protection, climate change and the analysis of the impact of human activities on water resources and water environment. and other fields. Therefore, it is of great scientific significance and application value to study how to improve the prediction accuracy of hydrological models.

Any model is accompanied by errors and uncertainties. In the model modeling work, there are a large number of error sources, and the error sources mainly include the following aspects:

(1) Errors caused by excluded factors

In the modeling process, the hydrological model needs to consider every link of the whole hydrological process of "precipitation - runoff - confluence", each link has many influencing factors, it is impossible to introduce every factor into the model arrived. Therefore, it is necessary to select these influencing factors to produce a certain prediction error.

(2) The error of the measured historical record data

The accuracy of the measured data and the size of the error are determined by the advanced and mature measurement technology, which affects the fitting degree of the model simulation, thereby affecting the prediction accuracy of the model. These data include not only traditional hydrological (flow) meteorological (rainfall) data, but also factors such as geology, vegetation, soil and land use.

(3) Parameter error

The parameters of the distributed hydrological model have clear physical meanings, and it is easy to estimate the variation range of the parameters, but it is difficult to determine the optimal value of the parameters.

(4) Model structure error

Incorrect calculation methods, inappropriate time steps, inappropriate running sequences, incomplete or biased model structures, etc. used in the model design and establishment process can all cause model prediction errors.

In order to eliminate the model prediction errors caused by the above reasons, the calibration of parameters is an important link to improve the prediction accuracy of hydrological models. Most of the watershed hydrological models, especially some parameters in small and medium-sized watersheds, cannot be directly determined through observation and experiments, and their values are not It has a certain relationship with the characteristics of the underlying surface of the watershed, but it cannot establish a relationship with the characteristics of the underlying surface of the watershed, so the calibration of parameters for the watershed hydrological model is still a difficult problem.

In the prior art for specific applications in watersheds with data, the parameter calibration of the hydrological model generally adopts the traditional trial-and-error method, that is, the parameter values of the hydrological model are continuously adjusted manually to meet the simulation accuracy requirements, but for the hydrological model without data The calibration of model parameters, using this method, has problems such as low calibration accuracy, which seriously affects the accuracy of hydrological forecasting.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is: to provide a method for calibrating hydrological parameters in areas with data based on an adversarial neural network, so as to solve the problem that the prior art adopts the traditional trial-and-error method for determining the parameters of the hydrological model of watersheds without data, that is, through The parameter values of the hydrological model are continuously adjusted manually to meet the requirements of simulation accuracy. However, this method has problems such as low calibration accuracy, which seriously affects the accuracy of hydrological forecasting.

The technical scheme of the present invention is:

An adversarial neural network-based calibration method for hydrological parameters in areas with data, which includes:

Step 1. Collect soil texture, vegetation coverage, land utilization, terrain data, runoff coefficient, total annual evaporation, gradient and slope data;

Step 2. Divide the calibration area into calculation units below 30 square kilometers;

Step 3. Determine the underlying surface and meteorological related factors of each parameter of each calculation unit according to the physical characteristics of the parameters of the hydrological model;

Step 4. The adversarial neural network GAN is used to automatically calibrate the hydrological parameters of the watershed with data. The adversarial neural network GAN takes noise as the input, and optimizes the parameters through the hydrological model to obtain the optimal hydrological parameters for each unit.

The underlying surface and weather-related factors of each parameter described in step 3 are:

The method for automatic calibration of hydrological parameters using the adversarial neural network GAN described in step 4 is:

Step 4.1. Generate samples with normally distributed noise as the input of the generator;

Step 4.2. Input the generated sample set into the hydrological model for optimization to obtain the optimal parameters;

Step 4.3: Input the optimal parameters output by the hydrological model and the samples generated by the generator to the discriminator to judge the true and false.

In step 4.2, when the generated sample set is input into the hydrological model for optimization to obtain the optimal parameters, the deterministic coefficient is used as the optimization principle.

Beneficial effects of the present invention:

The invention divides the optimized area into countless independent computing units, and then uses the confrontation neural network GAN to automatically calibrate the hydrological parameters to realize the parameter calibration of the data area, so it can effectively solve the problem of the modern hydrological model due to its strong professionalism. The resulting problem of difficulty in use can reduce a lot of tedious steps and work for professional manual parameter adjustment and calibration in practical applications. For the popularization and application of various hydrological models, it solves the problem that the existing technology adopts the traditional trial and error method for the determination of the parameters of the hydrological model in the watershed with data, that is, the parameter values of the hydrological model are continuously adjusted manually to meet the simulation accuracy requirements. Subjectivity, repetitive work, low efficiency and extremely high complexity, which are not conducive to technical problems such as the application and promotion of hydrological models.

Description of drawings:

FIG. 1 is a schematic diagram of the automatic calibration process flow of the adversarial neural network GAN of the present invention for performing hydrological parameters on similar units.

detailed description

Step 4. Use the adversarial neural network GAN to automatically calibrate the hydrological parameters of the watershed with data. The adversarial neural network GAN takes noise as the input, and optimizes the parameters through the hydrological model to obtain the optimal hydrological parameters for each unit;

Generative Adversarial Networks (GAN) is a subclass of generative models, which can estimate the potential distribution of existing data samples, build a model that can conform to the data distribution, and generate new data samples, and the model has A certain self-learning ability can be applied in semi-supervised learning.

The core idea of GAN comes from the Nash equilibrium of game theory. It sets the two parties involved as a generator and a discriminator. The purpose of the generator is to learn the real data distribution as much as possible, and the purpose of the discriminator is to discriminate as accurately as possible. Whether the input data comes from the real data or the generator; the two models need to be continuously optimized at the same time, each improving its own generating ability and discriminating ability, and the calculation is completed when the two reach a balance.

Conventional adversarial neural networks cannot directly achieve automatic parameter calibration because there are no real samples. Therefore, each time the generator outputs a generated sample, the hydrological model must be used to select the optimal parameters as the real sample input for the next discriminator iterative calculation.

It can be seen that the loss values of the discriminator and the generator are gradually close to 1, indicating that the model is converged.

The certainty coefficient of the hydrological model can be seen to be improved from 0.78 in the initial stage to 0.86. It shows that the neural network can be used for the optimization of hydrological model parameters.

It can be seen that when the certainty coefficient is better than the last time, the loss value will suddenly increase, indicating that the determiner will automatically retrain after updating the true value, and it will converge quickly. Therefore, when performing optimization to obtain the optimal parameters, the deterministic coefficient is used as the optimization principle.

The learning ability of the deep learning network is quite strong. When a real sample is given, the generated model can quickly converge to the range of the real sample. Since the real value also needs to be updated iteratively, it is extremely prone to overfitting problems, which directly lead to convergence. Slow or stuck in a local optimum. The present invention solves these problems by using dormant local neurons, weight regularization, and adjustment of neuron data.

The core problem of the present invention is to find the optimal parameters of the data basin. Therefore, an optimal search strategy can be added to optimize the generated samples to improve the performance of the entire network.

The reason why the present invention adopts the adversarial neural network to derive the optimal parameters is that each iteration is generated by random changes in the best distribution space of the last time.

Claims

An adversarial neural network-based calibration method for hydrological parameters in areas with data, which includes:

Step 1. Collect soil texture, vegetation coverage, land utilization, terrain data, runoff coefficient, total annual evaporation, gradient and slope data;

Step 2. Divide the calibration area into calculation units below 30 square kilometers;

Step 3. Determine the underlying surface and meteorological related factors of each parameter of each calculation unit according to the physical characteristics of the parameters of the hydrological model;

Step 4. The adversarial neural network GAN is used to automatically calibrate the hydrological parameters of the watershed with data. The adversarial neural network GAN takes noise as the input, and optimizes the parameters through the hydrological model to obtain the optimal hydrological parameters for each unit.
A method for calibration of hydrological parameters in areas with data based on adversarial neural network according to claim 1, characterized in that: the underlying surface of each parameter and the meteorological correlation factor described in step 3 are:
A kind of hydrological parameter calibration method based on adversarial neural network according to claim 1, it is characterized in that: the method that adopts adversarial neural network GAN to automatically calibrate hydrological parameters described in step 4 is:

Step 4.1. Generate samples with normally distributed noise as the input of the generator;

Step 4.2. Input the generated sample set into the hydrological model for optimization to obtain the optimal parameters;

Step 4.3: Input the optimal parameters output by the hydrological model and the samples generated by the generator to the discriminator to judge the true and false.
A method for calibration of hydrological parameters in areas with data based on adversarial neural network according to claim 3, characterized in that: in step 4.2, when the generated sample set is input into the hydrological model for optimization to obtain the optimal parameters, it is The deterministic coefficient is used as the preferred principle.