Search

CN-121981019-A - River underwater topography inversion method based on physical information neural network and deviation correction network

CN121981019ACN 121981019 ACN121981019 ACN 121981019ACN-121981019-A

Abstract

The invention discloses a river underwater topography inversion method based on a physical information neural network and an offset correction network, and provides a double-stage inversion frame combining physical driving and data correction. The method comprises the steps of constructing a physical information neural network inversion model, training by utilizing water level and flow velocity data at sparse monitoring points, obtaining a preliminary topography estimation field meeting a hydrodynamic rule by minimizing a physical loss function containing a two-dimensional shallow water equation residual error, constructing a deviation correction network, training by utilizing a pre-constructed prior data set containing rich non-Gaussian topography features, learning a systematic deviation mode existing in a preliminary inversion result, inputting the preliminary topography estimation field into the trained deviation correction network, predicting a topography deviation field, carrying out superposition correction, and outputting a final high-precision river underwater topography. The invention realizes the dynamic and high-precision reconstruction of complex topographic features such as pits, sand and the like, and provides a reliable dynamic boundary condition updating method for flood forecast.

Inventors

  • LI YICHEN
  • YE QIANG
  • ZENG LINGZAO

Assignees

  • 浙江大学

Dates

Publication Date
20260505
Application Date
20260403

Claims (10)

  1. 1. A river channel underwater topography inversion method based on a physical information neural network and an offset correction network is characterized by comprising the following steps: step 1) acquiring sparse hydrodynamic observation data of a target river reach, wherein the sparse hydrodynamic observation data comprise data of water level and flow velocity change along with time at a plurality of discrete monitoring points; step 2), constructing a physical information neural network inversion model, and establishing a mapping relation between space-time coordinates and flow field state quantity and river channel topography parameters; Step 3) training the physical information neural network inversion model constructed in the step 2) by utilizing the sparse hydrodynamic observation data in the step 1), and obtaining a preliminary terrain estimation field by minimizing a loss function of physical constraint; Step 4) additionally constructing a deviation correction network, training the deviation correction network by utilizing a pre-constructed prior data set containing various known terrain estimation fields and flow field feature graphs, and learning systematic deviation distribution features of river terrain; And 5) inputting the preliminary terrain estimation field and the corresponding flow field state characteristics in the step 3) into the trained deviation correction network in the step 4) to predict the terrain deviation field, superposing the terrain deviation field and the preliminary terrain estimation field, and outputting a final river underwater terrain inversion result.
  2. 2. The method of claim 1, wherein the network architecture of the physical information neural network inversion model comprises constructing a fully connected depth neural network as the infrastructure of the physical information neural network inversion model, and wherein the inputs are space-time coordinates Outputting a topography parameter including a time-varying hydrodynamic state quantity and a time-varying hydrodynamic state quantity, wherein the hydrodynamic state quantity comprises water depth 、 Directional flow velocity And Directional flow velocity The topographic parameter is the elevation of the river bed ; During the training process, the river bed elevation Is set to be in accordance with the space coordinates only The output of the associated independent subnetwork, not over time And (3) a change.
  3. 3. The method of claim 2, wherein the loss function of the physical constraint of step 3) consists of a data fitting term and a control equation residual term; the control equation residual term is constructed based on a two-dimensional shallow water equation set and comprises a continuity equation residual and a momentum equation residual, when calculating the residual, the partial derivative of network output to input coordinates is calculated by utilizing the self-contained automatic differential technology of a neural network, and the riverbed elevation is calculated Is substituted into the bottom slope source term in the momentum equation.
  4. 4. The method of claim 1, wherein step 4) the bias correction network employs a full convolutional neural network and a U-Net architecture for establishing an end-to-end mapping from preliminary terrain estimation field and flow field feature maps to terrain bias fields; The construction process of the prior data set comprises the following steps: Generating a large number of virtual riverbed terrain samples with non-Gaussian distribution characteristics by utilizing a random generation algorithm or an image enhancement technology; Forward modeling is carried out on the virtual riverbed terrain sample by utilizing a hydrodynamic numerical model, and corresponding virtual hydrodynamic observation data are generated; Simulating a corresponding virtual preliminary terrain estimation field by using the physical information neural network inversion model based on virtual observation data; And calculating a difference value between the virtual preliminary terrain estimation field and the reference virtual riverbed terrain sample as tag data of the training deviation correction network.
  5. 5. The method of claim 1, wherein the terrain bias field of step 5) is used for compensating smoothing effect errors of the physical information neural network due to spectrum bias in the region of severe terrain variation or extreme value region, and the final river underwater terrain inversion result The calculation formula is as follows: ; Wherein, the A preliminary terrain estimation field output by the inversion model of the physical information neural network, The network predicted terrain bias field is corrected for bias.
  6. 6. A method according to claim 3, wherein the bottom friction stress term in the two-dimensional shallow water equation set is calculated by using a manning formula, and is specifically formed as follows: ; ; Wherein, the And Is that And The bottom surface of the direction is subjected to friction stress, In order to achieve a fluid density, The acceleration of the gravity is that, Is the coefficient of the roughness coefficient of Manning, 、 In order for the flow rate to be the same, Is the water depth.
  7. 7. Use of the method according to any one of claims 1-6 for river topography monitoring, flood forecasting and channel maintenance monitoring.
  8. 8. The method of claim 7, wherein the complex topography is characterized by non-Gaussian distributions.
  9. 9. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method of any of claims 1 to 6.
  10. 10. A river terrain monitoring system, comprising: 1) The data acquisition module is used for acquiring water level and flow velocity data of the river section in real time; 2) The calculation module is used for operating the inversion method according to any one of claims 1 to 6 and outputting a current river underwater topography distribution map, and when the calculation module operates the inversion method, the calculation module performs hot start optimization through a preset historical weight parameter so as to improve the convergence rate of online calculation; 3) And the updating module is used for updating boundary condition parameters in the flood forecast model according to the river underwater topography distribution map obtained by inversion.

Description

River underwater topography inversion method based on physical information neural network and deviation correction network Technical Field The invention relates to the technical field of hydraulic engineering measurement, computational hydrodynamics and artificial intelligence intersection, in particular to a river underwater topography inversion method based on a physical information neural network and an offset correction network. Background The river underwater topography is the foundation of river dynamics research, flood evolution simulation and channel planning. Accurate mastering of the distribution of the river bed topography is important to improving the forecasting accuracy of the hydrodynamic model. Traditional sonar mapping is high in cost and poor in timeliness, and rapid terrain changes caused by floods are difficult to capture. Therefore, inversion methods based on hydrodynamic observations (e.g., water level, flow rate) are becoming a research hotspot. In order to obtain unknown riverbed terrain, inversion methods based on hydrodynamic observations are typically employed. In general, the observation process of an open channel hydrodynamic system can be expressed briefly as: Wherein, the Representing observation vectors (such as water level and flow rate at sparse monitoring points),Representing model input variables (such as upstream flow),Representing the model parameters to be inverted (i.e. the riverbed terrain field),Representing the observed error. Function ofA numerical model operator representing simulated open channel flow dynamics. From a mathematical perspective, river terrain inversion is a typical Bayesian statistical inference problem. The aim being based on observed dataNumerical model operatorInferring unknown topographical parameters. The posterior probability distribution of the topographical parameters can be expressed as: Wherein, the Posterior probability distribution for topographical parameters; Is a priori distribution of the values of the distribution, Is a likelihood function that is a function of the likelihood,Is an edge likelihood or evidence item of observed data. To solve this problem, a Kalman smoother is integrated) Is one of the most widely used methods at present. At the position ofIn general, an inclusion is generatedThe set of monte carlo for each member represents a priori uncertainty of the model parameters. The core updating mechanism is shown as follows: Wherein, the AndRespectively the firstThe updated (posterior) and pre-updated (prior) topographical parameters of the individual collection members,In the form of a kalman gain matrix,Is the firstThe observation disturbance vectors corresponding to the members of the collection are used for representing the randomness of the measurement errors in the parameter updating process; For numerical model operators Predictive output of a priori parameters, representing the use of open channel hydrodynamic modeling for the firstAnd the water level or flow velocity predicted value is obtained after forward modeling is carried out on the terrain parameter samples. However, the above formula reveals an essential disadvantage of the conventional approach in that the Kalman gain matrix K is a linear operator based on covariance calculation. This linear update mechanism implies a gaussian distribution assumption, i.e. assuming that both the riverbed topography and the observed errors obey a normal distribution. However, in actual natural river courses, the topography tends to be extremely complex, including deep pits, sand, and other forms having significant non-gaussian characteristics. When faced with such non-gaussian distribution scenarios, conventional methods based on linear assumptions can result in inversion results being "flattened" and unable to recover the extreme topography features, thus producing systematic deviations. Although deep learning methods attempt to replace linear matrices with neural networksHowever, in the absence of physical constraints, the pure data driven model results are often not "true" enough, and outliers that do not fit the actual rules are easily generated and are affected by the quality and quantity of the data. Although the physical equation constraint is introduced into the physical information neural network which is rising in recent years, the physical information neural network is limited by the property that the neural network is used to learn low-frequency characteristics, and smooth solutions still tend to be generated when the high-frequency terrain change is processed, so that steep terrain change details are difficult to accurately capture. Therefore, a hybrid inversion method that combines physical mechanism constraints with data-driven error correction while focusing on high-frequency feature matching is needed to solve the high-precision reconstruction problem under complex non-gaussian terrain. Disclosure of Invention The invention provides a river underwater t