Two-dimensional hydrodynamic model data assimilation method and system and storage medium

Abstract

The invention discloses a two-dimensional hydrodynamic model data assimilation method and system and a storage medium, and relates to the technical field of data assimilation processing. The method comprises the following steps: acquiring prediction index fitting data under actually measured input parameters and model parameters in a historical flood event by using a forward modeling model, calculating a prediction error rate of the forward modeling model to prediction indexes by using the actually measured data and the fitting data, and constructing a prediction model which takes normalized values of the input parameters and the model parameters as input data, the method comprises the following steps: acquiring a prediction error rate of a neural network model taking a normalized value randomly combined in a parameter range as input data by taking a prediction error rate as output data, constructing a data mapping table, and searching input parameters which are the same as a new flood event in the data mapping table to obtain the prediction error rate of the neural network model. And taking the obtained input parameters and model parameters when the weighted values of the prediction error rates of all the prediction indexes are minimum as driving data of a forward model, and obtaining prediction index fitting data of the new flood event.

Description

Technical Field

[0001] The present invention relates to the technical field of data assimilation processing, and particularly relates to a two-dimensional hydrodynamic model data assimilation method, system and storage medium. Background Art

[0002] A two-dimensional hydrodynamic model is a mathematical model used to simulate the movement of water flow in a two-dimensional space (usually the horizontal plane), and is widely used in hydrodynamic research and engineering practices of water bodies such as rivers, lakes, estuaries, and coastal waters. It is an important tool in hydrodynamic research and engineering applications. By continuously improving the theoretical basis and numerical methods of the model and combining with actual observation data, the accuracy and applicability of the model can be improved, providing strong support for fields such as water resource management, flood control and disaster reduction, and ecological environment protection.

[0003] During the simulation process of the hydrodynamic model, it will be affected by uncertain factors such as model parameters, input data, and model structure, resulting in a deviation between the simulation result and the actual observation. Through data assimilation, the observed data and the model prediction data can be combined, and the assimilation algorithm can be used to continuously adjust the model parameters to make the model output closer to the actual observed value. In addition, data assimilation can optimize the initial field and boundary conditions of the model according to the observed data to make it more in line with the actual situation. Thus, the simulation accuracy of the model can be improved.

[0004] Currently, the existing technologies for the assimilation processing of hydrodynamic models mainly include particle filtering method, ensemble Kalman filtering method, and variational method. In the actual application process of the above methods, when the amount of variable data in the model is large, the demand for computing resources will increase significantly, not only resulting in low computing efficiency and high computing cost, but also it is difficult to meet the real-time or near-real-time hydrodynamic simulation requirements using the above methods. Especially during flood warning and emergency response, due to the high computational complexity of the existing assimilation methods, the calculation often cannot be completed in time, making the model state unable to be updated quickly, greatly reducing the timeliness of model fitting, thus limiting the application of the above methods in actual flood warning. Therefore, we propose a two-dimensional hydrodynamic model data assimilation method, system and storage medium. Summary of the Invention

[0005] The main purpose of the present invention is to provide a two-dimensional hydrodynamic model data assimilation method, system and storage medium, which can effectively solve the problems in the background art.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is, A two-dimensional hydrodynamic model data assimilation method, comprising: Collecting the measured data of the prediction indicators of historical flood events in the area to be monitored , substitute the measured data into the forward model of the flood event, and use the forward model to obtain the predicted index fitting data in the scenario where the input parameter is , and the model parameter is , where , among which, represents the data at time t in the flood event; u, v, and w are the numbers of the predicted index, input parameter, and model parameter respectively; Use the measured data and the fitting data to calculate the prediction error rate of the forward model for the u-th predicted index at time t in the flood event , where t ∈ T; T is the total fitting duration of the forward model for the flood event; Construct a neural network model with the normalized values of the input parameter and the model parameter as the input data, and the prediction error rate as the output data, and perform training. By adjusting the weight parameters, bias parameters, and hyperparameters of the neural network model until its accuracy is not less than the set threshold; Set the value range of the input parameter , and the value range of the model parameter , , and obtain the random combination of parameters within the value range , where ∈ , ; ∈ , , and obtain the prediction error rate of the neural network model when using the normalized value of the random combination as the input data, and construct a data mapping table → ; Search for the input parameter in the data mapping table that is the same as the input parameter of the new flood event at time t, and use the input parameter and the weighted value of the prediction error rate of all predicted indexes when it is the smallest for the model parameter as the driving data of the forward model, and obtain the corresponding predicted index fitting data . .

[0007] A two-dimensional hydrodynamic model data assimilation system, including: A data acquisition module for collecting the measured data of the predicted indexes of historical flood events in the area to be monitored ; A data fitting module, which is used to input the measured data of the prediction index obtained into the forward model of the flood event, and use the forward model to obtain the fitting data of the prediction index under the condition that the input parameter is and the model parameter is . Among them, , represents the data at time t in the flood event; u, v, and w are the numbers of the prediction index, input parameter, and model parameter respectively; A fitting data processing module, which is used to calculate the prediction error rate of the forward model for the u-th prediction index at time t in the flood event by using the measured data and the fitting data . Among them, t ∈ T; T is the total fitting duration of the forward model for the flood event; , A neural network construction module, which is used to construct an LSTM neural network model with the normalized values of the input parameter and the model parameter as the input data and the prediction error rate as the output data, and perform training. By adjusting the weight parameters, bias parameters, and hyperparameters of the neural network model until its accuracy is not less than the set threshold; A mapped data acquisition module, which is used to set the value range of the input parameter of the forward model , and the value range of the model parameter , , and obtain the random combination of parameters within the value range . Among them, ∈ , ; ∈ , , and obtain the prediction error rate when the neural network model uses the normalized value of the random combination as the input data, and construct a data mapping table → ; A driving data acquisition module, which is used to search for the input parameter in the data mapping table that is the same as the input parameter of the new flood event at time t, and use the input parameter and the weighted value of the prediction error rate of all prediction indexes when it is the smallest as the driving data of the forward model, and obtain the corresponding fitting data of the prediction index .

[0008] ​​A two-dimensional hydrodynamic model data assimilation storage medium stores an electronic program therein.

[0009] Furthermore, the prediction indicators include water level, flow rate, and inundation range; The input parameters include rainfall-runoff parameters and terrain parameters; The model parameters include the Manning roughness coefficient of the monitoring area, boundary condition parameters, calculation unit parameters, time step parameters, and model operation parameters.

[0010] Furthermore, the prediction error rate is calculated by the formula: | - | / × 100%.

[0011] Furthermore, the neural network model is an LSTM neural network model, and the hyperparameters of the neural network model include the hidden layer size, learning rate, and number of training epochs.

[0012] Furthermore, the process for determining the value ranges of the input parameters and the model parameters includes the following steps: Using the input parameters of the obtained historical flood events and the model parameters to construct a multi-dimensional data vector z, z = , T , and the multi-dimensional data vector z ~ N(μ, ∑), where μ is the d-dimensional mean vector representing the expected value of each data dimension; ∑ is the d×d covariance matrix representing the covariance relationship between the data dimensions; d = ; Construct the normal distribution probability density function of the multi-dimensional data vector z , where = ; is the inverse matrix of the covariance matrix; π is the pi; Sample a noise vector from the normal distribution of the constructed multi-dimensional data vector z; Construct a generative adversarial network model, where the generator G of the generative adversarial network model takes the sampled noise vector as the input and outputs a generated multi-dimensional virtual data vector; the discriminator D of the generative adversarial network model takes the normalized value of the multi-dimensional data vector z as the input and outputs the probability that the generated multi-dimensional virtual data vector is real data; ​Train the discriminator D using the real data in the multi-dimensional data vector z and the virtual data in the multi-dimensional virtual data vector generated by the generator, calculate the loss of the real data and the loss of the generated data, and update the parameters of the discriminator D; input the virtual data generated by the generator G into the discriminator D, calculate the loss of the generator G, and update the parameters of the generator G so that the generator G can generate multi-dimensional virtual data closer to the real data in the multi-dimensional data vector z. Use the trained generator G to generate multi-dimensional virtual data with a sample size of Q, and use the minimum value of the input parameters in the generated multi-dimensional virtual data as the lower limit of the input parameter value range and the maximum value of the input parameters as the upper limit of the input parameter value range and the minimum value of the model parameters as the lower limit of the model parameter value range and the maximum value of the model parameters as the upper limit of the model parameter value range .

[0013] Further, the sampling step of the noise vector is specifically as follows: Define the mean vector μ, where μ = [μ1, μ2,..., μ d T ; where μ d is the mean of the d-th dimensional data; Define the d×d dimensional covariance matrix ∑, where ∑ = , represents the covariance between the p-th dimensional data and the q-th dimensional data, and p = 1, 2,..., d; q = 1, 2,..., d; Use numerical methods to sample the noise vector from the multi-dimensional data normal distribution.

[0014] Further, the sample size Q of the multi-dimensional virtual data generated by the generator G satisfies the relation: Q ≥ C / ε 2 ; where C is the capacity of the generator G; ε is the difference between the data distribution of the virtual data generated by the generator G and the data distribution of the real data.

[0015] The present invention has the following beneficial effects ​Compared with the prior art, this technical solution uses a forward model to obtain the predicted indicator fitting data in the case of measured input parameters and model parameters in historical flood events, calculates the prediction error rate of the forward model for the predicted indicators using the measured data and the fitting data, constructs a neural network model with the normalized values of the input parameters and model parameters as input data and the prediction error rate as output data, obtains the prediction error rate when the neural network model uses randomly combined normalized values within the parameter range as input data, constructs a data mapping table, searches for the same input parameters as the new flood event in the data mapping table, and uses the input parameters obtained and the weighted value of the prediction error rates of all predicted indicators when it is minimized as the driving data of the forward model to obtain the predicted indicator fitting data of the new flood event. By using the neural network model to fit the historical data, the construction of the neural network model and the fitting prediction process of the data can be completed during the non-flood season, thereby establishing a data mapping table. When a new flood event occurs, directly searching the data mapping table can quickly determine the various setting parameters of the model, realize the cyclic driving of the forward model, and can quickly and timely complete the calculation during flood warning and emergency response, enabling the model state to be updated quickly, improving the timeliness of model fitting, and meeting the real-time or near-real-time hydrodynamic simulation requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic flowchart of a two-dimensional hydrodynamic model data assimilation method of the present invention; Figure 2 is a schematic structural diagram of a two-dimensional hydrodynamic model data assimilation system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] The following further describes the present invention in conjunction with the specific embodiments. Among them, the drawings are only for illustrative purposes and show only schematic diagrams, not physical diagrams, and should not be construed as limiting the present invention. In order to better illustrate the specific embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, and do not represent the dimensions of the actual product.

[0018] The specific implementation process of the technical solution of the present invention includes the following steps: Step 1: Collect the measured data of the predicted indicators of historical flood events in the area to be monitored , substitute the measured data into the forward model of the flood event, and use the forward model to obtain the predicted indicator fitting data in the case where the input parameter is and the model parameter is . Among them, , where represents the data at time t in the flood event; u, v, and w are the numbers of the predicted indicators, input parameters, and model parameters respectively; Among them, the predicted indicators include water level, flow rate, and inundation range; The input parameters include rainfall-runoff parameters and terrain parameters; The model parameters include the Manning roughness coefficient of the monitoring area, boundary condition parameters, calculation unit parameters, time step parameters, and model operation parameters.

[0019] In this embodiment, the forward model is taken as the LISFLOOD-FP model for illustration. In the LISFLOOD-FP model, the following parameters need to be set: Manning roughness coefficient The Manning roughness coefficient is an important parameter in hydraulics and hydrology, used to describe the frictional resistance between water flow and the riverbed, riverbank, or inner wall of a pipe.

[0020] Terrain parameters DEM data: The Digital Elevation Model (DEM) is one of the basic inputs of the model, used to define the terrain of the calculation domain. The resolution and accuracy of the DEM directly affect the accuracy of the simulation results.

[0021] Boundary condition parameters Boundary condition files: Include the upstream flow boundary (.bci file) and the downstream water level boundary (.bdy file). These files define the input and output conditions of the model and are the key to simulating flood propagation.

[0022] Point source boundary conditions: For example, the inflow of a rainwater inspection well can be calculated by the weir flow formula and input into the model as a point source boundary condition.

[0023] Calculation unit parameters Calculation unit division: The entire area to be monitored is divided into calculation units according to the terrain grid structure. Each calculation unit contains information such as elevation, roughness coefficient, infiltration parameter, percentage of impervious area, and depression depth.

[0024] Land use type division: Each calculation unit is divided into a pervious area and an impervious area according to the land use type or the type of surface cover, and runoff calculations are carried out separately.

[0025] Rainfall-runoff parameters Rainfall intensity: Rainfall intensity is an important input parameter for simulating waterlogging and affects surface runoff.

[0026] Infiltration amount: In the pervious area, the runoff volume is equal to the rainfall intensity minus the infiltration loss. Therefore, the infiltration amount is a key parameter.

[0027] Time step parameters Time step: According to the CFL condition, the time step is linearly proportional to the grid cell size to ensure the stability of the numerical simulation.

[0028] Model operation parameters Model operation file: including the setting of model operation parameters, such as total simulation time, time step, etc.

[0029] After setting the LISFLOOD-FP model parameters, prepare the input files: Prepare input files, including DEM files, boundary condition files (.bci and.bdy), channel information files (.river), etc. These files need to be organized and named according to the format requirements of the LISFLOOD-FP model.

[0030] Model execution: Run the simulation using the execution file of the LISFLOOD-FP model. The model will simulate the evolution of floods in time and space based on the input DEM, boundary conditions, and channel information.

[0031] Result analysis Interpretation of output files: The model output includes parameters such as water depth and flow velocity in raster format, as well as predicted water level and flow hydrographs. The output files are usually provided at an hourly resolution to help analyze the spatio-temporal evolution of floods.

[0032] Step 2: Use measured data and fitted data Calculate the prediction error rate of the forward model for the u-th prediction index at time t during a flood event , and the calculation formula is: | | / × 100%; where t ∈ T; T is the total fitting duration of the forward model for the flood event.

[0033] Step 3: Construct an LSTM neural network model with the normalized values of input parameters , model parameters as input data and the prediction error rate as output data, and train it. By adjusting the weight parameters, bias parameters, and hyperparameters of the neural network model until its accuracy is not less than the set threshold, where the hyperparameters include hidden layer size, learning rate, and number of training epochs.

[0034] Step 4: Set the value range of the input parameters , and the value range of the model parameters , .

[0035] Specifically, the process for determining the value range is as follows: Step 41: Use the input parameters of the obtained historical flood events and model parameters Construct a multi-dimensional data vector z, where z = , T , and the multi-dimensional data vector z ~ N(μ, Σ), where μ is a d-dimensional mean vector representing the expected value of each data dimension; Σ is a d×d covariance matrix representing the covariance relationship between data dimensions; d = ; Step 42: Construct the probability density function of the normal distribution of the multi-dimensional data vector z , where = ; is the inverse matrix of the covariance matrix; π is the pi; Step 43: Sample a noise vector from the normal distribution of the constructed multi-dimensional data vector z. The specific process is as follows: I. Define the mean vector μ, where μ = [μ1, μ2,..., μ d T ; where μ d is the mean of the d-th dimensional data; II. Define a d×d dimensional covariance matrix Σ, where Σ = , represents the covariance between the p-th dimensional data and the q-th dimensional data, and p = 1, 2,..., d; q = 1, 2,..., d; III. Use numerical methods to sample the noise vector from the multi-dimensional data normal distribution; Step 44: Construct a generative adversarial network model. Among them, the generator G of the generative adversarial network model takes the sampled noise vector as input and outputs a generated multi-dimensional virtual data vector; the discriminator D of the generative adversarial network model takes the normalized value of the multi-dimensional data vector z as input and outputs the probability that the generated multi-dimensional virtual data vector is real data; Step 45: Use the real data in the multi-dimensional data vector z and the virtual data in the multi-dimensional virtual data vector generated by the generator to train the discriminator D, calculate the loss of the real data and the loss of the generated data, and update the parameters of the discriminator D; input the virtual data generated by the generator G into the discriminator D, calculate the loss of the generator G, and update the parameters of the generator G so that the generator G can generate multi-dimensional virtual data closer to the real data in the multi-dimensional data vector z; Step 46: Use the trained generator G to generate multi-dimensional virtual data with a sample size of Q, and use the minimum value of the input parameters in the generated multi-dimensional virtual data as the lower limit of the input parameter value range and the maximum value of the input parameters as the upper limit of the input parameter value range and the minimum value of the model parameters as the lower limit of the model parameter value range ​​, the maximum value of the model parameters is used as the upper limit of the range of model parameter values .

[0036] Step 5: Obtain a random combination of parameters within the value range , where ∈ , ; ∈ , , obtain the prediction error rate when the neural network model takes the normalized value of the random combination as the input data, and construct a data mapping table , → .

[0037] Step 6: Search for the input parameter in the data mapping table that is the same as the input parameter of the new flood event at time t , and use the input parameter and the model parameter when the weighted value of the prediction error rate of all prediction indicators is minimized as the driving data of the forward model, where the weighted value of the prediction error rate = , represents the weight of the u-th prediction indicator. In this embodiment, the weight coefficients of water level, flow rate, and inundation area can be taken as equal values of 1 / 3, then the weighted value of the prediction error rate = ; At this time, the input parameter is the collected data of the new flood event at time t. When the weighted value of the prediction error rate is minimized, by searching the data mapping table → , the best model parameter of the LISFLOOD-FP model at this time can be determined. Take the input parameter of the new flood event collected at time t and the best model parameter as the driving data of the LISFLOOD-FP model at this time, and the fitting data of each prediction indicator of the LISFLOOD-FP model can be obtained. Repeat this process to achieve the data assimilation process.

[0038] The technical solution of the present invention uses a neural network model to fit historical data, and can complete the construction of the neural network model and the fitting prediction process of the data during the non-flood season, thereby establishing a data mapping table. When a new flood event occurs, the data mapping table can be directly searched to quickly determine the various setting parameters of the model, realizing the cyclic drive of the forward model. When carrying out flood warning and emergency response, the calculation can be quickly and timely completed, enabling the model state to be updated quickly, improving the timeliness of model fitting, and meeting the real-time or near-real-time hydrodynamic simulation requirements.

[0039] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification only illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A two-dimensional hydrodynamic model data assimilation method, characterized in that, Including: Collect the measured data of the prediction indicators of historical flood events in the area to be monitored ; Obtain the predicted index fitting data in the scenario where the input parameter is and the model parameter is ; ; Using measured data and fitted data Calculate the prediction error rate of the forward model for the u-th prediction index at time t during the flood event ; Construct a neural network model with the normalized values of the input parameters , model parameters as the input data and the prediction error rate as the output data; Set the value range of the input parameters of the forward model , and the value range of the model parameters , , and obtain random combinations of parameters within the value range ; Construct a data mapping table, where the data mapping table is from a random combination mapped to the prediction error rate when the normalized value of a random combination is used as the input data to the neural network model ; Search for the input parameters of the new flood event at time t in the data mapping table with the same input parameters , and use the input parameters and the model parameters when the weighted value of the prediction error rates of all prediction indicators is minimized as the driving data of the forward model to obtain the corresponding predicted indicator fitting data .

2. A two-dimensional hydrodynamic model data assimilation method according to claim 1, characterized in that the prediction indicators include water level, flow rate, and inundation range; the input parameters include rainfall runoff parameters and terrain parameters; the model parameters include the Manning roughness coefficient of the monitoring area, boundary condition parameters, calculation unit parameters, time step parameters, and model operation parameters.

3. A two-dimensional hydrodynamic model data assimilation method according to claim 1, characterized in that Prediction error rate The calculation formula is: | - | / × 100%.

4. A two-dimensional hydrodynamic model data assimilation method according to claim 1, characterized in that The neural network model is an LSTM neural network model, and the hyperparameters of the neural network model include the hidden layer size, learning rate, and number of training epochs.

5. A two-dimensional hydrodynamic model data assimilation method according to claim 1, characterized in that The process for determining the value ranges of the input parameters and the model parameters includes the following steps: Using the input parameters of historical flood events obtained and model parameters Construct a multi-dimensional data vector z, z = , T , and the multi-dimensional data vector z ~ N(μ, ∑), where μ is the mean vector of dimension d, representing the expected value of each data dimension; ∑ is the covariance matrix of d×d, representing the covariance relationship between data dimensions; d = ;​ Construct the probability density function of the normal distribution of the multi-dimensional data vector z , where = ; is the inverse matrix of the covariance matrix; π is the ratio of a circle's circumference to its diameter Sampling a noise vector from the normal distribution of the constructed multi-dimensional data vector z; Constructing a generative adversarial network model, wherein the generator G of the generative adversarial network model takes the sampled noise vector as input and outputs a generated multi-dimensional virtual data vector; the discriminator D of the generative adversarial network model takes the normalized value of the multi-dimensional data vector z as input and outputs the probability that the generated multi-dimensional virtual data vector is real data; Training the discriminator D using the real data in the multi-dimensional data vector z and the virtual data in the multi-dimensional virtual data vector generated by the generator, calculating the loss of the real data and the loss of the generated data, and updating the parameters of the discriminator D; inputting the virtual data generated by the generator G into the discriminator D, calculating the loss of the generator G, and updating the parameters of the generator G so that the generator G can generate multi-dimensional virtual data closer to the real data in the multi-dimensional data vector z; Use the trained generator G to generate multi-dimensional virtual data with a sample size of Q, and use the minimum value of the input parameters in the generated multi-dimensional virtual data as the lower limit of the input parameter value range and the maximum value of the input parameters as the upper limit of the input parameter value range and the minimum value of the model parameters as the lower limit of the model parameter value range and the maximum value of the model parameters as the upper limit of the model parameter value range .

6. A two-dimensional hydrodynamic model data assimilation method according to claim 5, characterized in that The specific steps for sampling the noise vector are as follows: Define the mean vector μ, where μ = [μ1, μ2,..., μ d T ; where μ d is the mean of the d-th dimensional data;​ Define a d×d dimensional covariance matrix ∑, where, ∑ = , denotes the covariance between the p-th dimensional data and the q-th dimensional data, and p = 1, 2, ..., d; q = 1, 2, ..., d; Sampling the noise vector from the multi-dimensional data normal distribution using a numerical method.

7. A two-dimensional hydrodynamic model data assimilation method according to claim 5, characterized in that, The sample size Q of the multi-dimensional virtual data generated by the generator G satisfies the relation: Q ≥ C / ε 2 ; where C is the capacity of the generator G; and ε is the difference between the data distribution of the virtual data generated by the generator G and the data distribution of the real data.

8. A two-dimensional hydrodynamic model data assimilation system, characterized in that, Including: A data acquisition module for collecting measured data of prediction indicators of historical flood events in the area to be monitored ; A data fitting module, which is used to obtain the predicted index fitting data under the condition that the input parameter is and the model parameter is ; The fitting data processing module is used to utilize the measured data and the fitting data to calculate the prediction error rate of the forward model for the u-th prediction index at time t during the flood event ; A neural network construction module for constructing an LSTM neural network model with the normalized values of the input parameters , model parameters as input data and the prediction error rate as output data; A mapping data acquisition module, used to construct a mapping from a random combination to the prediction error rate when the normalized value of the neural network model with a random combination is used as input data data mapping table; A driving data acquisition module for searching for input parameters of a new flood event at time t in a data mapping table with the same input parameters , and taking the model parameters when the weighted value of the prediction error rate of the input parameters and all prediction indicators is minimized as the driving data of the forward model, and obtaining the corresponding prediction indicator fitting data .

9. A two-dimensional hydrodynamic model data assimilation storage medium, characterized in that, An electronic program is stored in the storage medium, wherein when the electronic program is run by a processor, it can implement the steps of a two-dimensional hydrodynamic model data assimilation method according to any one of claims 1-7.

Images