Machine learning based modeling method for water dynamics of river-lake composite system

Abstract

The embodiment of the application provides a river-lake composite system water dynamic force modeling method based on machine learning. Boundary conditions of a one-dimensional water dynamic force model provide input for machine learning. Output of the machine learning provides complex boundary conditions and downstream boundary conditions of a river-lake intersection for the one-dimensional water dynamic force model. The method improves the operation efficiency of the river-lake composite system water dynamic force simulation, can quickly calculate water levels and flow rates of important section positions of the river-lake composite system, reduces the requirement for topographic data, does not need to input complex lake topographic elevation survey data, and can effectively improve the water level simulation accuracy of sections near the downstream boundary.

Description

TECHNICAL FIELD

[0001] The application relates to the field of water dynamic modeling and can also be used in the field of machine learning, and particularly relates to a water dynamic modeling method for a river-lake composite system based on machine learning. BACKGROUND

[0002] Machine learning is an important tool for researching basin hydrological simulation. When machine learning is used for hydrological modeling, rainfall or upstream flow in a basin is generally taken as input, and downstream water level or flow is taken as output. The parameters of the machine learning model are trained according to the measured data, and the model that has completed parameter training can simulate and predict the water level and flow of rivers or lakes in the basin. Machine learning is a hydrological and water dynamic simulation method with high speed and high precision, but it can only simulate single stations one by one and is difficult to model hydrological and water dynamic elements of each section of a large range of river sections.

[0003] Water dynamic numerical simulation of a river-lake composite system can establish one-dimensional, two-dimensional or three-dimensional physical water dynamic models. Different model dimensions are essentially a trade-off between solving precision and calculation cost, and are also limited by data availability. For river simulation of a long river section and a long period of time, a one-dimensional model along the flow direction of the river channel helps to greatly reduce the calculation amount. In a lake area, a one-dimensional model cannot be used to generalize due to the large length and width in the horizontal direction, and a two-dimensional model or a three-dimensional model along the water depth direction is often used. However, for a large lake, high-resolution water depth measurement data or digital elevation data required for constructing a two-dimensional or three-dimensional model are difficult to obtain, and there are also problems such as low model running efficiency, repeated iteration of water level and flow at the coupled boundary and easy calculation instability. Under the premise of ensuring calculation precision, reasonable generalization of the lake model to reduce the calculation cost is a major challenge faced by simulation of hydrological and water dynamic processes of a river-lake composite system. In addition, in a changing environment without water conservancy engineering regulation, the downstream boundary condition of the river-lake composite system is unknown. The water level-flow relationship curve method commonly used is often disturbed by topography, vegetation and other factors and is not very stable, which easily affects the precision of the river-lake composite system water dynamic simulation.

[0004] Machine learning has the advantage of exploring implicit relationships in nonlinear systems. Coupling machine learning and hydrodynamic methods may provide new ideas for solving complex problems in river-lake composite system water dynamic modeling. Practical application and in-depth study show that there are some shortcomings in the current river-lake composite system modeling method of machine learning and hydrodynamics: (1) the calculation site of pure machine learning method is limited, and it is difficult to model each unit of the entire study area at the same time; (2) the pure hydrodynamic method is time-consuming and requires high terrain data, especially for the lake part; (3) the way of using hydrodynamic model to drive machine learning for alternative modeling has limited effect, and the parameters need to be updated frequently according to the changes of terrain, and the scalability is not strong, and the advantages of machine learning and hydrodynamic method are not fully combined. SUMMARY

[0005] In view of the problems in the prior art, the application provides a river-lake composite system water dynamic modeling method based on machine learning, which can effectively improve the operation efficiency of river-lake composite system water dynamic simulation.

[0006] In order to solve at least one of the above problems, the application provides the following technical scheme:

[0007] The application provides a river-lake composite system water dynamic modeling method based on machine learning, comprising:

[0008] Collecting flow data of upstream boundary, lateral boundary and river-lake intersection position station of the study area, and water level data of downstream boundary station;

[0009] According to the flow of the upstream boundary station and the flow of the lateral tributary station that may affect the water exchange between the river and the lake, a machine learning method is used to train and predict the flow at the river-lake intersection position;

[0010] According to the flow of the upstream and lateral boundary stations and the exchanged flow between the river and the lake, a machine learning method is used to train and predict the water level of the downstream boundary;

[0011] Discretize the study river section according to the river terrain data, divide a cross section every certain distance on the river, and arrange the bank edge starting point distance and elevation data on the cross section;

[0012] The initial flow of each cross section is assigned as the flow of the upstream boundary on the first day, and the initial water level is calculated by linear interpolation according to the water level of the downstream boundary on the first day and the river slope;

[0013] Discretize the Saint-Venant equation set, and input the initial and boundary conditions of each river cross section to perform one-dimensional hydrodynamic calculation of the river.

[0014] Further, the machine learning method for training and predicting the flow at the river-lake intersection position comprises:

[0015] The specific calculation paradigm adopted is:

[0016]

[0017] wherein, is the water exchange of the jth lake and river; Q in is the upstream inflow; Q B is the river tributary flow, and the water exchange of the jth lake and river is affected by i river tributaries in total; is the tributary flow of the jth lake itself, and there are k lake tributaries in total; f(g) represents a preset machine learning method.

[0018] Further, the machine learning method is used to train and predict the water level of the downstream boundary, comprising:

[0019] The specific calculation paradigm of the downstream boundary water level adopted is as follows:

[0020]

[0021] wherein, Z is the downstream boundary water level; Q in is the upstream inflow; Q B is the river tributary flow; Q L is the water exchange of the lake and river; m and n are the number of tributaries connected to the river and the number of lakes connected to the river respectively; f(g) represents a preset machine learning method.

[0022] From the above technical solution, the application provides a river-lake composite system water dynamic modeling method based on machine learning, which can improve the operation efficiency of river-lake composite system water dynamic simulation, quickly calculate the water level and flow of important cross-section positions of the river-lake composite system, reduce the requirement for topographic data, and does not need to input complex lake topographic elevation survey data, and can effectively improve the water level simulation accuracy of the cross-section near the downstream boundary. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0024] Figure 1 The flowchart of the river-lake composite system water dynamic modeling method based on machine learning in the embodiments of the application is shown in the figure.

[0025] Figure 2A comparison diagram of measured and simulated water levels and flow rates of the main site in an embodiment of the present application;

[0026] Figure 3 A comparison diagram of measured and simulated downstream boundary water levels in an embodiment of the present application. DETAILED DESCRIPTION

[0027] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0028] The acquisition, storage, use, processing and the like of data in the technical scheme of the present application all comply with relevant provisions of national laws and regulations.

[0029] In order to effectively improve the operation efficiency of the water dynamics simulation of the river-lake composite system, an embodiment of a river-lake composite system water dynamics modeling method based on machine learning is provided in the present application, which is shown in Figure 1 The river-lake composite system water dynamics modeling method based on machine learning specifically includes the following contents:

[0030] Step 1, hydrological data collection: collect the flow data of the upstream boundary, lateral boundary and river-lake intersection site of the study area, and the water level data of the downstream boundary site.

[0031] Specifically, according to the data sources such as the Hydrological Yearbook, the upstream boundary flow (Yichang Station flow), river tributary flow (Hanjiang Xiantao Station flow), lake tributary flow (Dongting Lake Shimen Station, Taoyuan Station, Taoyuan Station and Xiangtan Station flow, Poyang Lake Wai Zhou Station, Lijia Station, Meigang Station, Hushan Station, Qizhen Station, Wanjiabu Station and Dufengkeng Station flow), river-lake intersection point flow (Dongting Lake Chenglingji Station flow, Poyang Lake Hukou Station flow) and downstream boundary water level (Datong Station water level) of the middle reaches of the Yangtze River (including the main stream of the Yangtze River, Dongting Lake, Poyang Lake, Hanjiang River, etc.) from 2007 to 2016 are collected. All flow or hydrological data are daily scale hydrological data.

[0032] Step 2, calculation of river-lake exchange water volume: according to the flow of the upstream boundary site and the flow of the lateral tributary site that may affect the water exchange between the river and the lake, a machine learning method is used to train and predict the flow at the river-lake intersection position. The specific calculation paradigm is as follows:

[0033]

[0034] wherein, is the water exchange between the jth lake and the river; Q in is the upstream inflow, i.e., the flow at Yichang Station; Q B is the flow of the river tributary, and the water exchange between the jth lake and the river is affected by i river tributaries; is the flow of the tributary of the jth lake, and there are k tributaries of the lake; RNN(g) represents a machine learning method called recurrent neural network.

[0035] The flow at the river-lake junction from 2007 to 2013 is taken as the output of the machine learning, the upstream boundary flow (the flow at Yichang Station), the tributary flow of the corresponding lake (the flow at Shimen Station, Taojiang Station, Taoyuan Station, and Xiangtan Station of Dongting Lake, and the flow at Wai Zhou Station, Lijia Station, Meigang Station, Hushan Station, Qizhen Station, Wanjiabu Station, and Daofengkeng Station of Poyang Lake), and the upstream river tributary flow that can affect the river-lake junction (for example, the flow at Hukou Station of Poyang Lake and the junction of the main stream of the Yangtze River) are taken as the input of the machine learning, and the parameters of the machine learning model are trained. The RNN model with trained parameters is used to predict the flow at the river-lake junction from 2014 to 2016 (the flow at Chenglingji Station of Dongting Lake and the flow at Hukou Station of Poyang Lake).

[0036] Step 3, downstream boundary water level calculation: according to the flow at the upstream and lateral boundary stations and the water exchange between the river and the lake, a machine learning method is used to train and predict the water level at the downstream boundary. The specific calculation paradigm is as follows:

[0037]

[0038] where Z is the downstream boundary water level, i.e., the water level at Datong Station; Q in is the upstream inflow, i.e., the flow at Yichang Station; Q B is the flow of the river tributary, i.e., the flow at Xiantao Station of Hanjiang River; Q L is the water exchange between the lake and the river; m and n are the number of tributaries connected to the river and the number of lakes connected to the river, respectively, which are 1 and 2; RNN(g) represents a machine learning method called recurrent neural network.

[0039] The water level at the downstream boundary station from 2007 to 2013 (the water level at Datong Station) is taken as the target variable of the machine learning model, the upstream boundary flow (the flow at Yichang Station), the flow of all upstream river tributaries (the flow at Xiantao Station of Hanjiang River), and the flow of the lake into and out of the river (the flow at Chenglingji Station and the flow at Hukou Station) are taken as the input of the machine learning model, and the parameters of the machine learning model are trained. The RNN model with trained parameters is used to predict the water level at the downstream boundary station from 2014 to 2016 (the water level at Datong Station).

[0040] Step 4, dividing river cross-sections: Discretize the study reach according to the river topographic data, and divide a cross-section every certain distance on the river. A total of 535 cross-sections are divided, with a distance of 1-5 kilometers between cross-sections. The bank edge starting point distance and elevation data on these cross-sections are sorted.

[0041] Step 5, setting initial conditions: Assign the initial flow of the 535 cross-sections to the flow of the upstream boundary Yichang Station on the first day (January 1, 2014), and calculate the initial water level according to the water level of the downstream boundary Datong Station on the first day (January 1, 2014) and the river slope by linear interpolation;

[0042] Step 6, discretize the Saint-Venant equation group (Saint-Venant equation group, i.e. the basic equation of one-dimensional hydrodynamic model) in the following form:

[0043]

[0044] In the formula: A is the area, m; t is the time, s; Q is the flow, m 3 / s; x is the flow, m; q is the lateral flow per unit flow, positive value indicating inflow, m 2 / s; g is the acceleration of gravity, taken as 9.81 m / s 2 ; B is the water surface width, m; Z is the water level, m; n is the roughness, dimensionless; R is the wet perimeter, m.

[0045] Solve the one-dimensional hydrodynamic model Saint-Venant equation group according to the boundary conditions obtained in steps 1-3 and the initial conditions obtained in step 5. The water level and flow simulation results of the main river and lake stations are shown in Figure 2 . The simulated value and the measured value of the downstream boundary water level are shown in Figure 3 . This method better simulates the water level and flow of the river-lake complex system, and is suitable for hydrodynamic modeling of the river-lake complex system. In addition, this method can simulate the water level and flow of the study area for 1 year in 9.4 minutes, which is 38 times faster than the traditional two-dimensional coupled hydrodynamic model, which takes 6 hours to simulate the same area and time length (Xijun Lai, Jiahu Jiang, Qiuhua Liang, Qun Huang, 2013, Large-scale hydrodynamic modeling of the middle Yangtze River Basin with complex river-lake interactions, Journal of Hydrology, 492, pp. 228-243.).

[0046] From the above description, the method for modeling water dynamics of a river-lake composite system based on machine learning provided by the embodiments of the present application can improve the operation efficiency of the water dynamics simulation of the river-lake composite system through flexible modeling, quickly calculate the water level and flow of important cross-section positions of the river-lake composite system, and reduce the requirement for topographic data, without the need to input complex lake topographic elevation survey data, and can effectively improve the water level simulation accuracy of the cross-section near the downstream boundary.

[0047] The principles and implementation manners of the present application are described by using specific embodiments in the present application, and the above embodiment descriptions are only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the field, the specific implementation manners and application ranges will be changed according to the idea of the present application, and the above descriptions should not be understood as the limitation of the present application.

Claims

1.A method for modeling hydrodynamic forces in a river-lake composite system based on machine learning, characterized in that, The method comprises: Collecting flow data of upstream boundary, lateral boundary and river-lake intersection site of the study area, and water level data of downstream boundary site; According to the flow of the upstream boundary site and the flow of the lateral tributary site that may affect the water exchange between the river and the lake, a machine learning method is used to train and predict the flow at the river-lake intersection site, including: The specific calculation paradigm used is: wherein, is the water exchange between the jth lake and the river; is the upstream inflow; is the flow of the tributary alongside the river, and the water exchange of the jth lake with the river is affected by i tributaries of the river; is the flow of the tributary that the jth lake itself receives, and there are k tributaries of the lake in total; f() represents a preset machine learning method; According to the flow of the upstream and lateral boundary sites and the exchanged flow between the river and the lake, a machine learning method is used to train and predict the water level of the downstream boundary, including: The specific calculation paradigm of the downstream boundary water level is as follows: where Z is the downstream boundary water level; is the upstream inflow; is the flow of tributaries alongside the river; is the water exchange between the lake and the river; m and n are the number of tributaries and lakes connected to the river, respectively; f() represents a predetermined machine learning method; According to the river terrain data, the study river section is discretized, and a cross section is divided at a certain distance on the river, and the bank edge starting point distance and elevation data on the cross section are arranged; The initial flow of each cross section is assigned as the flow of the upstream boundary on the first day, and the initial water level is linearly interpolated according to the water level of the downstream boundary on the first day and the river slope; Discrete Saint-Venant equation set, and input initial and boundary conditions of each river cross section for one-dimensional hydrodynamic calculation of the river.

Images