A model-coupled lake area water level prediction method, medium and computer device

Abstract

The application relates to the technical field of flood disaster prediction, and discloses a model-coupled lake area water level prediction method, a medium and computer equipment. The prediction method comprises the following steps: establishing a two-dimensional mathematical water dynamic model to obtain an optimal water level prediction neural network model; obtaining actual measurement values of upstream water levels; taking predicted downstream water level prediction values as downstream boundary values of the two-dimensional mathematical water dynamic model, and taking the actual measurement values of the upstream water levels as upstream boundary values; running the two-dimensional mathematical water dynamic model to output water level data conforming to physical laws; respectively establishing neural network models for prediction sites in the lake area to realize point water level prediction based on physically consistent labels; and dynamically updating and outputting water level prediction values. Effects are as follows: physical consistency is maintained, and the physical credibility of prediction results is improved; downstream boundary conditions are dynamically updated in response to real-time hydrology, and the method is suitable for lake area water level prediction scenarios; high-dimensional field simulation and low-dimensional key point prediction are decoupled, and prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of flood disaster forecasting technology, and in particular to a model-coupled method, medium, and computer equipment for predicting water levels in lake areas. Background Technology

[0002] Floods are a type of natural disaster that has a severe impact on human society. Climate change and human activities are increasing the frequency of floods and the associated losses.

[0003] With the rise of big data technology, data-driven artificial intelligence models have gained widespread attention in the field of water level prediction. These models reveal potential knowledge and patterns by analyzing and mining data, reducing errors caused by insufficient human understanding of physical laws and model simplification. Artificial intelligence has demonstrated powerful capabilities in handling complex nonlinear dynamic systems and has been widely used in areas such as rainfall-runoff, flood, and sediment forecasting since the 1990s.

[0004] Invention application CN202511181867.6 provides a rapid flood forecasting method for rivers based on a physical information neural network. Specifically, it proposes a physical information neural network for rapid flood forecasting, introducing boundary condition parameters as input variables to learn the flood wave propagation patterns under different boundary conditions, thereby rapidly predicting the water depth changes at various river cross-sections under arbitrary boundary conditions. Furthermore, to address the computational accuracy issue, this invention proposes an integrated hydrodynamic method-based physical information neural network flood rapid forecasting framework, RFF-PINN. This framework uses piecewise polynomial interpolation to reconstruct the grid-based discrete numerical solution into a continuous function in the spatiotemporal domain, constructing a residual loss between the network output and the hydrodynamic model simulation values. This residual loss, along with the PDE loss, co-optimizes the network parameters, avoiding the reduced optimization effect caused by the difficulty in converging the PDE loss in complex flow regions.

[0005] Invention patent ZL202410487324.6 provides a rapid flood forecasting method and system based on machine learning. It constructs a hydro-hydrodynamic coupling model to simulate flood inundation under different rainfall return periods and patterns, using a preset water depth threshold to determine inundated and non-inundated areas. Furthermore, this invention constructs an optimal flood forecasting model and an optimal water depth prediction model to quickly calculate the inundation range, inundation depth, and inundation evolution for each grid. Compared to previous studies that used neural network models to output single variables and required building numerous models to predict multiple spatial points, the CNN model constructed in this invention considers the spatial correlation of multiple grid points, resulting in higher prediction accuracy and efficiency. It also overcomes the time-consuming nature of hydro-hydrodynamic models, enabling rapid output of forecast results.

[0006] Invention patent ZL202411631896.3 discloses an urban flood prediction method based on a Bayesian convolutional neural network, comprising: collecting historical rainfall events in the study area; simulating different rainfall events using hydrodynamic simulation software to obtain the maximum inundation depth; selecting characteristic variables affecting the inundation depth, including 10 rainfall characteristic variables and 10 spatial characteristic variables; constructing a Bayesian convolutional neural network prediction model considering uncertainty; training the model based on the loss function MAE; inputting validation set data into the trained model, and calculating the mean of the approximate posterior distribution of the output as the predicted value of the inundation depth. This method, which calculates and analyzes the relationship between characteristic variables and target variables, effectively integrates different data, and considers potential sources of uncertainty, not only predicts the magnitude and range of inundation depth in the study area but also quantifies the uncertainty of the prediction, providing support for real-time flood disaster management.

[0007] Although existing patent documents disclose technical solutions for combining neural networks with hydrodynamic models for flood simulation, such as the pure data-driven models proposed in ZL202410487324.6 and ZL202411631896.3, and the Physical Information Neural Network (PINN) proposed in CN202511181867.6, their coupling methods are generally offline, unidirectional, and static. That is, the hydrodynamic model is only used to generate training samples, and the neural network is only used as a substitute model for prediction. The hydrodynamic model runs all historical floods at once, outputs grid water depth or cross-sectional flow, and then completely exits the system. The neural network only uses these static results as labels for supervised learning, and no longer interacts with the hydrodynamic model during the prediction stage. For the PINN, although it attempts to make the neural network fit the observed data and the physical law grid simultaneously during training by using the residuals of the control equations as part of the loss function, it is extremely unstable and difficult to converge during training. Secondly, due to the constraints of physical laws, it is often limited to the boundary conditions used for training, and cannot strictly follow physical laws when the boundary is exceeded. In essence, it is also a static prediction tool. It has the following insurmountable defects: 1) Downstream boundary conditions cannot be dynamically updated: Existing technologies mostly use fixed tide levels or empirical relationships as downstream boundaries, which cannot be adaptively adjusted according to real-time hydrological responses. This leads to rapid accumulation of simulation errors in areas where tidal influence is significant or downstream water level changes are drastic. 2) Lack of a real-world data feedback mechanism: Because the hydrodynamic model is discarded after training, the prediction phase relies entirely on a black-box network, and flood evolution no longer satisfies the conservation laws of the Saint-Venant or shallow water equations. When encountering extreme rainfall, levee breaches, gate scheduling, or other abrupt boundary changes, the network output often exhibits serious flaws such as "unreasonable water level jumps, mismatched peak times, and non-conservation of water volume," and cannot be corrected through real-time observation. Once the neural network is trained, its output is no longer compared or corrected with real-world data, making it unable to cope with extreme events or abrupt boundary conditions. 3) Redundant output dimensions, making it difficult to consider low-dimensional, highly sensitive representative points: Existing technologies mostly focus on river or urban flooding scenarios, and their hydrodynamic-neural network coupled models generally adopt a "full-grid output" approach, that is, predicting the water depth of tens of thousands to millions of nodes at once. For slowly changing water bodies such as lakes with wide water surfaces, minimal water level gradients, and few representative points, this mode not only causes huge storage and computational redundancy, but is also prone to overfitting due to the curse of dimensionality, resulting in the worst accuracy in predicting water levels for "low-dimensional but highly sensitive" representative points in lake areas; 4) Rigid representative point configuration makes it difficult to flexibly add or remove target points as needed: Existing hydrodynamic-neural network coupled models typically pre-fix a set of spatial output locations (such as full-grid nodes, regular sampling nodes, or several predetermined monitoring sections) during the modeling phase, and determine the output dimension of the neural network accordingly. The network structure is strongly bound to this pre-set set of points. The prediction phase often only provides results for this set of predetermined locations. If, during actual flood control scheduling or monitoring deployment, it is necessary to temporarily add water level predictions for a specific lake station, sluice gate section, or a local area of ​​interest, this can usually only be achieved by increasing the output dimension, adjusting the network structure, and retraining the overall model. It is difficult to rapidly incrementally model for any specified point based on the existing model. This rigid output design cannot meet the flexible requirements of lake area operations, which require "freely specifying a small number of key points within a given research area and quickly establishing a prediction model."

[0008] Therefore, there is an urgent need for a lake level prediction method that is physically consistent and can be corrected online to solve the problems in the existing technology. Summary of the Invention

[0009] The main objective of this invention is to provide a model-coupled lake level prediction method to address the problem that existing technologies are difficult to apply to lake level prediction due to inherent defects in the coupling mode. These defects include: rapid accumulation of simulation errors caused by the inability to dynamically update downstream boundary conditions; inability to cope with extreme events or sudden changes in boundary conditions due to the lack of a measured data feedback mechanism; and insufficient prediction accuracy and flexibility due to redundant output dimensions and rigid representative point configuration. The specific technical solution is as follows: A model-coupled method for predicting lake water levels includes the following steps: Step 1: Determine the scope of the river network and lake area to be studied, and obtain basic data; the basic data includes forecasting stations within the lake area, flow data from upstream stations, rainfall data from the area and water level data from downstream stations, as well as model topographic data; Step 2: Establish a two-dimensional mathematical hydrodynamic model for the river network and lake area under study. Specifically, this involves: using triangular mesh division and refining the mesh for narrow and winding areas; and selecting the Manning coefficient. A water level forecasting neural network model for river networks and lake areas was established in the PyTorch framework. The hyperparameters were determined by the control variable method, and the optimal water level forecasting neural network model was built. Step 3: Obtain the actual measured value of the upstream water level; Step 3: Use the downstream water level prediction value obtained from the optimal water level prediction neural network model as the downstream boundary value of the two-dimensional mathematical hydrodynamic model, and use the actual measured value of the upstream water level as the upstream boundary value; run the two-dimensional mathematical hydrodynamic model and output water level data for the research river network and lake area. Step 4: Based on the water level data of the river network and lake area obtained in Step 3, establish neural network models for the forecast stations in the lake area to realize point water level prediction based on physical consistency labels. Step 5: Dynamically update and output the water level forecast values ​​of the forecast stations in the lake area based on changes in upstream water levels.

[0010] Preferably, the water level forecast values ​​of forecast stations within the lake area are dynamically updated and output based on changes in upstream water levels, including: Step 5.1: Using the latest measured value of the upstream water level as the inflow boundary, input the neural network models whose weights remain unchanged in Step 4; Step 5.2: Use a neural network model to predict the downstream boundary water level in real time. The downstream boundary of the two-dimensional mathematical hydrodynamic model is used to drive the two-dimensional mathematical hydrodynamic model to roll forward and calculate. The input includes the measured data from the upstream to predict the downstream water level of the lake area in advance. The two-dimensional mathematical hydrodynamic model simulates the water level field of the entire lake area between the current time and the prediction time span. Step 5.3: Extract the water level sequence of any forecast station within the lake area from the water level field of the entire lake area in Step 5.2, input it into the corresponding neural network model, and output the water level forecast value of the forecast station within the lake area.

[0011] Preferably, in step two: the Manning coefficient n is divided into substrate-vegetation zones: bare sand 0.013, sparse aquatic plants 0.021, dense vegetation 0.045, with a fine-tuning step size of 0.001, and calibrated to NSE≥0.90.

[0012] Preferably, a water level forecasting neural network model for river networks and lake areas is established within the PyTorch framework. The hyperparameters are determined using the controlled variable method, and the optimal water level forecasting neural network model is constructed. The specific steps are as follows: Using RMSE as the objective function, a grid search method is employed to iterate through three sets of hyperparameters: the number of hidden layers, neurons in each layer, and the learning rate. The combination corresponding to the minimum RMSE in the validation set is selected as the optimal hyperparameter, thus completing the construction of the optimal water level forecast neural network model. The model is trained using partial measured data and validated using partial data to output the target water level. The collected measured data is divided into training, validation, and test sets according to time sequence. Input variables include upstream measured water level, downstream measured flow, interval rainfall, and gate opening. The output variable is the measured water level at the target representative point. After training, the corresponding water level prediction value and two accuracy indicators, Nash efficiency coefficient (NSE) and RMSE, are output. The optimal water level forecast neural network model ultimately outputs the downstream water level.

[0013] Preferably, in step four: the neural network models of each forecasting station in the lake area share the same network structure, but each forecasting station trains its own sub-model independently, that is, the weights and biases are saved independently; The network structure includes: 1 hidden layer with 64 neurons, ReLU activation function; 1 output neuron with linear activation, corresponding to the water level; loss function: MSE; optimizer: Adam; learning rate: 0.001.

[0014] Preferably, the training method is as follows: for each important forecasting station, representative forecasting station, or forecasting station of special concern, the simulated sequence of the forecasting station and the measured water level of the same period are used as physical consistency labels, and training, validation and test sets are divided, and a sub-model weight file is trained and saved independently. Each sub-model outputs the predicted water level for the corresponding forecast station over the next 1-72 hours, as well as the NSE and RMSE accuracy indices.

[0015] Preferably, when adding a new forecast site, only the new sub-model needs to be trained, without changing the network framework.

[0016] Preferably, the training, validation, and test sets are divided in a ratio of 7:1.5:1.5.

[0017] This invention discloses a model-coupled method for predicting lake levels, comprising: determining the scope of the study river network and lake area and acquiring basic data; establishing a two-dimensional mathematical hydrodynamic model for the study river network and lake area to obtain an optimal water level prediction neural network model; acquiring actual upstream water level measurements; using the downstream water level prediction value obtained from the optimal water level prediction neural network model as the downstream boundary value of the two-dimensional mathematical hydrodynamic model, while the upstream boundary value uses the actual upstream water level measurements; running the two-dimensional mathematical hydrodynamic model to output water level data for the study river network and lake area; establishing neural network models for prediction stations within the lake area based on the water level data obtained in step three, achieving point water level prediction based on physically consistent labels; and dynamically updating the output water level prediction values ​​for prediction stations within the lake area based on upstream water level changes. This method has the following technical effects: 1. This invention addresses the problem of the inability to dynamically update downstream boundary conditions. Specifically, it establishes a water level forecasting neural network model within the study area. Using upstream measured flow and water level data, as well as interval rainfall data, as input, it outputs a predicted downstream water level. This predicted value is directly used as the downstream boundary condition for the two-dimensional hydrodynamic model, while the upstream boundary uses real-time measured water levels. With the arrival of new upstream water level and rainfall information for a new time period, the neural network re-provides the downstream prediction value, and the hydrodynamic model recalculates the water level field across the entire lake area. This achieves automatic updating of downstream boundary conditions with real-time hydrological response, avoiding the rapid accumulation of errors caused by fixed tidal levels or empirical relationships.

[0018] 2. To address the issues of insufficient feedback from measured data and the disconnect between prediction and physical constraints, this invention maintains the online coupling between the hydrodynamic model and the neural network during the prediction phase. On one hand, the downstream boundary is provided in real time by the neural network, while the hydrodynamic model performs numerical simulations according to shallow water equations, ensuring that the entire flood evolution process satisfies conservation laws and basic hydrodynamic constraints. On the other hand, based on the water level field of the entire lake area output by the hydrodynamic model and the latest measured water level, the neural network model for important stations is continuously updated and trained, allowing its parameters to be corrected under new observation information. Through this closed-loop feedback mechanism of "neural network prediction—hydrodynamic constraints—observation correction," this invention can suppress problems such as unreasonable water level jumps, mismatched peak times, and non-conservation of water volume under extreme rainfall, dike breach, and gate scheduling scenarios with sudden boundary changes, achieving continuous comparison and correction of prediction results against measured data.

[0019] 3. To address the issues of redundant output dimensions and difficulty in simultaneously considering low-dimensional, highly sensitive representative points, this invention decouples the high-dimensional spatial distribution task from the low-dimensional key point prediction task. Specifically, a two-dimensional hydrodynamic model is responsible for simulating the water level field of the entire lake area in continuous space. The neural network no longer undertakes high-dimensional prediction with "full-grid output," but instead establishes water level prediction models only for a small number of representative points, important monitoring stations, or points of special interest in the lake area. The training samples for the representative point neural network mainly come from the water level sequence of that point obtained from hydrodynamic simulation, supplemented by measured data for verification and correction. Through this "small-scale, low-dimensional output" point model design, this invention significantly reduces the curse of dimensionality and the risk of overfitting without increasing storage and computational burden, and improves the prediction accuracy of water levels for "low-dimensional but highly sensitive" representative points in the lake area.

[0020] 4. This invention addresses the problems of rigid representative point configuration, difficulty in flexibly adding or removing target points as needed, and difficulty in obtaining samples. Specifically, it pre-constructs a unified two-dimensional hydrodynamic model within the study area. Once parameter calibration is completed, it can be repeatedly run under various inflow, rainfall, and scheduling conditions to generate water level simulation results for any location and time period across the entire lake area. For existing or newly added lake stations, sluice gate sections, and points of local interest, this invention directly extracts the water level time series of the corresponding points from existing hydrodynamic simulation results. Combined with a small amount of measured data for that point, a corresponding lightweight neural network prediction model can be quickly trained. This eliminates the need to modify the original large network structure or to accumulate and organize measured samples for each new point over a long period. Thus, it enables the free designation of a small number of key points within a given study area and rapid numerical modeling based on the hydrodynamic model, avoiding the problems of traditional schemes where the representative point output set is fixed in the modeling stage and difficult to add, remove, or expand later.

[0021] The present invention also discloses a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the model-coupled lake level prediction method described above.

[0022] The present invention also provides a computer device, comprising: a processor and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the model-coupled lake level prediction method described above. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0024] Figure 1 This is a flowchart of the lake water level prediction method based on model coupling of the present invention; Figure 2 The simulated and measured values ​​of farmland in the Dongting Lake area during the period of 1998 are displayed; Figure 3 The simulated and measured values ​​of farmland in the Dongting Lake area for the period of 2020 are displayed; Figure 4 The simulated and measured values ​​of deer antlers in the Dongting Lake area during the period of 1998 are shown; Figure 5 The simulated and measured values ​​of deer antlers in the Dongting Lake area for the period of 2020 are displayed; Figure 6 The simulated and measured values ​​of Chenglingji in the Dongting Lake area during the period of 1998 are shown; Figure 7 The simulated and measured values ​​of Chenglingji in the Dongting Lake area for the period of 2020 are displayed; Figure 8 The simulated and measured values ​​for Nanzui in the Dongting Lake area during the period of 1998 are displayed; Figure 9 The simulated and measured values ​​for Nanzui in the Dongting Lake area during the period of 2020 are displayed; Figure 10 The simulated and measured values ​​of Xiaozui in the Dongting Lake area during the period of 1998 are shown; Figure 11 The simulated and measured values ​​of Xiaozui in the Dongting Lake area during the period of 2020 are displayed; Figure 12 The simulated and measured values ​​of Yangliutan in the Dongting Lake area during 1998 are displayed. Figure 13 The simulated and measured values ​​of Yangliutan in the Dongting Lake area during the period of 2020 are displayed; Figure 14 The simulated and measured values ​​for Xiangyin County within the Dongting Lake area during the period of 1998 are displayed. Figure 15 The simulated and measured values ​​for Xiangyin County within the Dongting Lake area during the period of 2020 are displayed. Figure 16 These are the results of the neural network training and testing for the Luoshan Station prediction period, conducted over 1 day. Figure 17 These are the 2-day training and testing results of the neural network for the Luoshan Station prediction period; Figure 18 These are the 3D training and testing results of the neural network for the Luoshan Station prediction period; Figure 19 This is the result of a 1-day simulation verification of the prediction period at Luoshan Station using a neural network in 1998; Figure 20 This is the result of a 2-day simulation verification of the prediction period at Luoshan Station using a neural network in 1998; Figure 21 This is the result of a 3D simulation verification of the prediction period of the Luoshan Station using neural networks in 1998; Figure 22 This is the 1-day simulation verification result of the neural network for the prediction period of Luoshan Station from 2017 to 2022; Figure 23 This is the verification result of a 2-day simulation of the prediction period for Luoshan Station using a neural network from 2017 to 2022. Figure 24 This is the verification result of a 3D simulation of the prediction period of Luoshan Station from 2017 to 2022 using a neural network; Figure 25 This is a schematic diagram of the results of the 1998 neural network-hydrodynamic simulation of Nanzui. Figure 26 This is a schematic diagram of the results of the 1998 neural network-hydrodynamic simulation of Xiaozui. Figure 27 This is a schematic diagram of the results of the 1998 neural network-hydrodynamic simulation in Xiangyin. Figure 28 This is a schematic diagram of the results of the 1998 neural network-hydraulic simulation of farmland. Figure 29 This is a schematic diagram of the results of a neural network-hydrodynamic simulation of antlers in 1998; Figure 30 This is a schematic diagram of the results of the 1998 neural network-hydrodynamic simulation of Chenglingji. Figure 31 This is a schematic diagram of the results of the 2017 neural network-hydrodynamic simulation of Nanzui. Figure 32 This is a schematic diagram of the results of the 2017 neural network-hydrodynamic simulation of Xiaozui. Figure 33 This is a schematic diagram of the results of the 2017 neural network-hydrodynamic simulation in Xiangyin. Figure 34 This is a schematic diagram of the results of the 2017 neural network-hydraulic simulation of farmland. Figure 35 This is a schematic diagram of the results of neural network-hydrodynamic simulation of antlers in 2017; Figure 36This is a schematic diagram of the results of the 2017 neural network-hydrodynamic simulation of Chenglingji. Figure 37 This is a schematic diagram of the results of the 2020 neural network-hydrodynamic simulation of Nanzui. Figure 38 This is a schematic diagram of the results of the 2020 neural network-hydrodynamic simulation of Xiaozui. Figure 39 This is a schematic diagram of the results of the 2020 neural network-hydrodynamic simulation in Xiangyin. Figure 40 This is a schematic diagram of the results of the 2020 neural network-hydraulic simulation of farmland. Figure 41 This is a schematic diagram of the results of the 2020 neural network-hydrodynamic simulation of deer antlers; Figure 42 This is a schematic diagram of the results of the 2020 neural network-hydrodynamic simulation of Chenglingji.

[0025] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0027] Example: A model-coupled method for predicting lake levels, specifically a method for predicting lake levels by coupling an artificial intelligence model with a hydrodynamic model, see [link to relevant documentation]. Figure 1 This includes the following steps: Step 1: Determine the scope of the river network and lake area to be studied, and obtain basic data; the basic data includes flow data from upstream stations, rainfall in the area, water level data from downstream stations, and model topographic data; Step 2: Establish a two-dimensional mathematical hydrodynamic model for the river network and lake area under study. Specifically, this involves: using triangular mesh division and refining the mesh for narrow and winding areas; and selecting the Manning coefficient. A water level forecasting neural network model for river networks and lake areas was established in the PyTorch framework. The hyperparameters were determined by the control variable method, and the optimal water level forecasting neural network model was built. Step 3: Obtain the actual measured value of the upstream water level; Step 3: Use the downstream water level prediction value obtained from the optimal water level prediction neural network model as the downstream boundary value of the two-dimensional mathematical hydrodynamic model, and use the actual measured value of the upstream water level as the upstream boundary value; run the two-dimensional mathematical hydrodynamic model and output water level data for the research river network and lake area. Step 4: Based on the water level data of the river network and lake area obtained in Step 3, establish neural network models for the forecast stations in the lake area to realize point water level prediction based on physical consistency labels. Step 5: Dynamically update and output the water level forecast values ​​of the forecast stations within the lake area based on changes in upstream water levels, including: Step 5.1: Using the latest measured value of the upstream water level as the inflow boundary, input the neural network models whose weights remain unchanged in Step 4; Step 5.2: A neural network model is used to predict the downstream boundary water level in real time. This downstream boundary serves as the driving force for the two-dimensional mathematical hydrodynamic model to perform forward calculations. Input includes measured data from upstream to predict the downstream water level in the lake area. The two-dimensional mathematical hydrodynamic model simulates the water level field of the entire lake area between the current moment and the predicted time span. The two-dimensional hydrodynamic model calculation uses the central finite volume method to solve the two-dimensional shallow water equations, adhering to the conservation of mass and momentum. The main equations include the continuity equation and the momentum equation, i.e. Continuity equation: ; Momentum equation: ; ; In the formula, For time; , The coordinates are in a rectangular coordinate system. , respectively along , Flow velocity component in the direction; For water depth; Water level; Coriolis parameters; It is the acceleration due to gravity; The density of water; The coefficient of turbulent viscosity; , They are respectively , Surface wind stress; , They are respectively , Downward shear stress.

[0028] Step 5.3: Extract the water level sequence of any forecast station within the lake area from the water level field of the entire lake area in Step 5.2, input it into the corresponding neural network model, and output the water level forecast value of the forecast station within the lake area.

[0029] In this preferred embodiment, in step two: the Manning coefficient n is divided into substrate-vegetation zones: bare sand 0.013, sparse aquatic plants 0.021, dense vegetation 0.045, with a fine-tuning step size of 0.001, and calibrated to NSE≥0.90.

[0030] In this preferred embodiment, a water level forecasting neural network model for river networks and lake areas is established within the PyTorch framework. The hyperparameters are determined using the controlled variable method, and the optimal water level forecasting neural network model is constructed. The specific operations are as follows: Using RMSE as the objective function, a grid search method is employed to iterate through three sets of hyperparameters: the number of hidden layers, neurons in each layer, and the learning rate. The combination corresponding to the minimum RMSE in the validation set is selected as the optimal hyperparameter, thus completing the construction of the optimal water level forecast neural network model. The model is trained using partial measured data and validated using partial data to output the target water level. The collected measured data is divided into training, validation, and test sets according to time sequence. Input variables include upstream measured water level, downstream measured flow, interval rainfall, and gate opening. The output variable is the measured water level at the target representative point. After training, the corresponding water level prediction value and two accuracy indicators, Nash efficiency coefficient (NSE) and RMSE, are output. The optimal water level forecast neural network model ultimately outputs the downstream water level.

[0031] In this preferred embodiment, in step four: the neural network models of forecast stations in each lake area share the same network structure, but each station trains its own sub-model independently, that is, the weights and biases are saved independently. The network structure includes: 1 hidden layer with 64 neurons, ReLU activation function; 1 output layer with linear activation, corresponding to water level; loss function: MSE; optimizer: Adam; learning rate: 0.001. The specific training method is as follows: for each important station, representative point, or special point of interest, the simulated sequence of the station and the measured water level of the same period are used as physical consistency labels, and training, validation and test sets are divided. A sub-model weight file is trained independently and saved. Each sub-model outputs the predicted water level for the corresponding site over the next 1-72 hours, as well as NSE and RMSE accuracy indicators.

[0032] In this preferred embodiment, when adding a new site, only the new sub-model needs to be trained, without modifying the network framework.

[0033] In this embodiment, the training, validation, and test sets are preferably divided in a ratio of 7:1.5:1.5.

[0034] Taking Dongting Lake as an example, the solution in this embodiment is implemented as follows: First, a two-dimensional hydrodynamic model of Dongting Lake was created. The model covers the entire Jingjiang River section and the Dongting Lake area, with the upstream boundary extending to Zhicheng and the downstream boundary reaching Luoshan, covering the four main rivers flowing into Dongting Lake.

[0035] In establishing river and lake hydrodynamic models, the Manning roughness coefficient (n) is a key parameter affecting the model's accuracy. This invention calculates the Manning coefficient using the following method: measured hydrological data from 1998 were input into the Dongting Lake hydrodynamic model to calibrate the Manning roughness coefficient (n). See details below. Figures 2-15 (Seven characteristic stations within the lake area: Yingtian, Lujiao, Chenglingji, Nanzui, Xiaohezui, Yangliutan, and Xiangyin). Combined with... Figures 2-15 It can be seen that the simulated values ​​of the model agree well with the measured values. The error analysis of the determination of the water level ratio of Dongting Lake is detailed in Table 1.

[0036] Table 1. Error Analysis for Determining the Water Level of Dongting Lake

[0037] Subsequently, a neural network model for the Luoshan station in the lower reaches of the Dongting Lake area was constructed using the machine learning framework PyTorch and the Python programming language. The dataset for model training and testing consisted of measured hydrological flow data from January 1, 2003 to December 31, 2016 (a total of 5114 days of valid data). The recurrent neural network model was trained using the first 80% of the data, and tested using the last 20% of the data.

[0038] In machine learning, data preprocessing is essential for a smoother training process; data normalization is commonly used. The following formula was chosen to scale all data to the range of 0 to 1.

[0039] ; in: This is the normalized value, i.e., the output result, ranging from 0 to 1; This refers to the original data value, i.e., the input value to be processed; It is the minimum value in this set of feature data; It is the maximum value among the features in this set.

[0040] Standardized data can lead to faster model convergence and higher accuracy during model development.

[0041] Table 2 shows the evaluation metrics for the simulation performance of neural network models with different training iterations (10–70 times), as follows: Table 2 Evaluation Indicators of Neural Network Model Simulation Performance for Different Training Runs (10–70 times)

[0042] As shown in Table 2, with the increase of training iterations, the Nash coefficient (NSE) and root mean square error (RMSE) exhibit trends of first increasing and then decreasing, and first decreasing and then increasing, respectively. In this case, 50 iterations were chosen as the training iterations for the neural network model to predict water levels. During model training, adjustments need to be made to the number of hidden layers and the number of neurons in each hidden layer, as detailed in Tables 3 and 4.

[0043] Table 3. Effect of the number of hidden layer neurons on the model.

[0044] Table 4 Calculation Results of Evaluation Indicators for the Number of Hidden Layers

[0045] Regarding parameter selection, this case study uses a controlled variable approach to determine the optimal parameters. The LSTM neural network model has the following parameters: 50 training iterations, 1 hidden layer, and 80 neurons. The ADAM adaptive learning rate optimization algorithm, designed to combat gradient explosion, is used as the optimizer, with an initial learning rate of 0.001. Dropout regularization is used to prevent overfitting, with a Dropout regularization parameter of 0.2. The tanh function is chosen as the activation function, while other different models are also explored.

[0046] The model was trained using measured upstream flow data and measured downstream water level data from the Luoshan hydrological station during typical flood years in Dongting Lake from January 1, 2003 to December 31, 2016. The sliding window size was set to 10 times the target duration to predict the water level at the Luoshan station at the target time. The trained model was then used to make lead-time forecasts for flood events at the Luoshan station in 1998 and from 2017 to 2023, thus completing the forecast of water level changes at the Luoshan hydrological station.

[0047] The water level prediction results of model training and validation are as follows: Figures 16-24 As shown, the predicted water level at Luoshan Station is basically consistent with the measured water level, with a relative error of no more than 1.0%. The overall prediction results of the model have strong reliability. Tables 5 and 6 show the various indicators of the LSTM and other model validation phases. It can be seen that the predicted water level values ​​obtained through training have a high degree of agreement with the measured water level values ​​at Luoshan Station.

[0048] Table 5. Calculation results of evaluation indicators during the testing phase of BP and RNN models.

[0049] Table 6. Calculation Results of Evaluation Indicators for the LSTM and GRU Model Testing Phase

[0050] Finally, a neural network-hydraulic coupling model was established based on the neural network model and the hydrodynamic model. The LSTM model showed better prediction results for the water level at Luoshan station in the lower reaches of Dongting Lake, so this result was used as the input for the downstream Luoshan station in the hydrodynamic model, while measured data were used as the input for the upstream station. The corresponding data were then input into the hydrodynamic model to simulate the flood events of 1998, 2017, and 2020. See details... Figures 25-30 , Figures 31-36 as well as Figures 37-42 The simulation results of water levels at six characteristic stations in Dongting Lake for different forecast periods in 1998, 2017, and 2020 are shown respectively.

[0051] Based on this model, data changes at any point within the lake area can be obtained, and corresponding neural network prediction models can be established for different points. The prediction results from the neural network prediction model of downstream stations in the lake area are updated and input into the hydrodynamic model to update the data and further predict target stations. For values ​​of stations to be set in the future, prediction models can be established as soon as possible, using a coupled-distributed approach to establish prediction models for forecast stations within the lake area. For the water levels of newly constructed flood diversion dikes in the Dongting Lake area, it is difficult to establish a direct prediction model due to the lack of historical measured data. Based on the above approach, accurate water level prediction can be achieved, as shown in Table 7.

[0052] Table 7. Opening and Closing Times of the Three Major Sluice Gates during the Dongting Lake Flood Processes in 1996, 1998, and 2017

[0053] The effects of applying the technical solution of this embodiment are as follows: 1. Achieving bidirectional online coupling between neural networks and hydrodynamic models while maintaining physical consistency: Specifically, this invention does not "abandon" the hydrodynamic model during the prediction phase. Instead, it constructs a closed-loop link of "neural network prediction—hydrodynamic calculation—result feedback" by generating downstream boundaries in real time through neural networks and feeding back the water level field of the entire lake area from the hydrodynamic model. The entire flood evolution process is always constrained by physical control equations such as the shallow water equation, avoiding problems such as non-conservation and unreasonable water level jumps in the output of a pure black-box network, significantly improving the physical reliability of the forecast results.

[0054] 2. The downstream boundary conditions are dynamically updated with real-time hydrological response, adapting to tidal and drastic fluctuation scenarios. Specifically, by directly using the downstream water level prediction value from the neural network as the downstream boundary of the two-dimensional hydrodynamic model, combined with the real-time measured water level input from the upstream, this invention can update the boundary conditions in each forecast period, achieving adaptive adjustment of the downstream boundary. This mechanism effectively suppresses the accumulation of errors caused by using fixed tidal levels or empirical relationships in areas with significant tidal influence or drastic downstream water level changes, improving the stability of long-duration flood process simulation and forecasting.

[0055] 3. Decoupling high-dimensional field simulation from low-dimensional key point prediction reduces the risk of dimensionality curse and overfitting. Specifically, this invention uses a two-dimensional hydrodynamic model to handle the high-dimensional calculation of the water level field across the entire lake area, while the neural network only performs low-dimensional output predictions for a small number of representative points, important monitoring stations, or points of special interest. It no longer uses a high-dimensional replacement network with "full-grid output." This collaborative mode of "full hydrodynamic field + representative neural network points" significantly reduces the dimensionality of the neural network output without increasing the overall computational burden, mitigating the dimensionality curse and overfitting problems, and significantly improving the accuracy of low-dimensional but highly sensitive water level predictions at key points in the lake area.

[0056] 4. This invention possesses the capability to flexibly add or remove target points as needed and to rapidly model point-to-point. Specifically, it pre-constructs a unified two-dimensional hydrodynamic model within the study area. After calibration, this model can be repeatedly run under various inflow, rainfall, and scheduling conditions to generate water level simulation results for any location within the entire lake area. For existing or newly added lake stations, sluice gate sections, and points of local interest, the water level time series for those points can be directly extracted from the hydrodynamic simulation results. Combined with a small amount of measured data, a corresponding lightweight neural network model can be quickly trained without modifying the original large network structure or accumulating historical sample data over a long period. This enables the ability to freely specify a small number of key points within a fixed study area and perform point-to-point distributed predictions, significantly improving the flexibility of model expansion and maintenance.

[0057] 5. It possesses online update and scenario migration capabilities, adapting to extreme events and new operating conditions. Specifically, by repeating neural network prediction and hydrodynamic simulation after updating the upstream water level, this invention can continuously update the neural network model of important stations after receiving a new round of measured data, allowing the model parameters to continuously absorb new information and correct old deviations during operation. This mechanism not only enhances the adaptability to sudden operating conditions such as extreme rainfall, dike breaches, and gate scheduling, but also facilitates its application in other lake areas with similar hydrological and climatic conditions, achieving a certain degree of scenario migration.

[0058] 6. Facilitates engineering integration and practical application deployment. Specifically, the method of this invention adopts a mature two-dimensional mathematical hydrodynamic model in the hydrodynamic part and is implemented based on the mainstream deep learning framework in the neural network part, which can be smoothly integrated with existing flood control scheduling and hydrological forecasting platforms. Through the clearly defined "hydrodynamic calculation + representative point intelligent forecasting" architecture, the computational complexity is controlled while ensuring forecast accuracy, making it suitable for long-term stable operation in engineering scenarios such as flood control duty and real-time rolling forecasting.

[0059] This embodiment also provides a computer device, including: a processor and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the model-coupled lake level prediction method described above.

[0060] In addition, this embodiment also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the model-coupled lake level prediction method described above.

[0061] It should be noted that the division of the various modules in the above system is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. These modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. Each module can be a separate processing element, or it can be integrated into a chip in the aforementioned device. Alternatively, it can be stored as program code in the memory of the aforementioned device, and its functions can be called and executed by a processing element of the device. Furthermore, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through the integrated logic circuits in the hardware of the processor element or through software instructions.

[0062] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0063] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0064] When an integrated unit / module is implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, etc. Unless otherwise specified, the processor can be any suitable hardware processor, such as a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic component. Unless otherwise specified, memory can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as a USB flash drive, Random-Access Memory (RAM), Static Random-Access Memory (SRAM), Dynamic Random-Access Memory (DRAM), Enhanced Dynamic Random-Access Memory (EDRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), Resistive Random Access Memory (RRAM), High-Bandwidth Memory (HBM), and Hybrid Memory Cube (HMC). Cube, magnetic storage, flash memory, disk, optical disk, portable hard drive or magnetic disk, and other media that can store program code.

[0065] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application.

[0066] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A model-coupled method for predicting lake water levels, characterized in that, Includes the following steps: Step 1: Determine the scope of the river network and lake area to be studied, and obtain basic data; the basic data includes forecasting stations within the lake area, flow data from upstream stations, rainfall data from the area and water level data from downstream stations, as well as model topographic data; Step 2: Establish a two-dimensional mathematical hydrodynamic model for the study area of ​​river network and lake area. Specifically, triangular mesh division is adopted, and the mesh is refined for narrow and curved areas. Select the Manning coefficient; A water level forecasting neural network model for river networks and lake areas was established in the PyTorch framework. The hyperparameters were determined by the control variable method, and the optimal water level forecasting neural network model was built. Step 3: Obtain the actual measured value of the upstream water level; Step 3: Use the downstream water level prediction value obtained from the optimal water level prediction neural network model as the downstream boundary value of the two-dimensional mathematical hydrodynamic model, and use the actual measured value of the upstream water level as the upstream boundary value; run the two-dimensional mathematical hydrodynamic model and output water level data for the research river network and lake area. Step 4: Based on the water level data of the river network and lake area obtained in Step 3, establish neural network models for the forecast stations in the lake area to realize point water level prediction based on physical consistency labels. Step 5: Dynamically update and output the water level forecast values ​​of the forecast stations in the lake area based on changes in upstream water levels.

2. The model-coupled lake level prediction method as described in claim 1, characterized in that, The water level forecast values ​​for forecast stations within the lake area are dynamically updated and output based on changes in upstream water levels, including: Step 5.1: Using the latest measured value of the upstream water level as the inflow boundary, input the neural network models whose weights remain unchanged in Step 4; Step 5.2: Use a neural network model to predict the downstream boundary water level in real time. The downstream boundary of the two-dimensional mathematical hydrodynamic model is used to drive the two-dimensional mathematical hydrodynamic model to roll forward and calculate. The input includes the measured data from the upstream to predict the downstream water level of the lake area in advance. The two-dimensional mathematical hydrodynamic model simulates the water level field of the entire lake area between the current time and the prediction time span. Step 5.3: Extract the water level sequence of any forecast station within the lake area from the water level field of the entire lake area in Step 5.2, input it into the corresponding neural network model, and output the water level forecast value of the forecast station within the lake area.

3. The model-coupled lake level prediction method as described in claim 1, characterized in that, In step two: the Manning coefficient n is divided into substrate-vegetation zones: bare sand 0.013, sparse aquatic plants 0.021, dense vegetation 0.045, with a fine-tuning step size of 0.001, and calibrated to NSE≥0.

90.

4. The model-coupled lake level prediction method as described in claim 1, characterized in that, A neural network model for water level forecasting in river networks and lake areas was established using the PyTorch framework. Hyperparameters were determined using the controlled variable method, and the optimal neural network model for water level forecasting was constructed. The specific steps are as follows: Using RMSE as the objective function, a grid search method is employed to iterate through three sets of hyperparameters: the number of hidden layers, neurons in each layer, and the learning rate. The combination corresponding to the minimum RMSE in the validation set is selected as the optimal hyperparameter, thus completing the construction of the optimal water level forecast neural network model. The model is trained using partial measured data and validated using partial data to output the target water level. The collected measured data is divided into training, validation, and test sets according to time sequence. Input variables include upstream measured water level, downstream measured flow, interval rainfall, and gate opening. The output variable is the measured water level at the target representative point. After training, the corresponding water level prediction value and two accuracy indicators, Nash efficiency coefficient (NSE) and RMSE, are output. The optimal water level forecast neural network model ultimately outputs the downstream water level.

5. The model-coupled lake level prediction method as described in claim 1, characterized in that, In step four: the neural network models of each forecasting station in the lake area share the same network structure, but each forecasting station trains its own sub-model independently, that is, the weights and biases are saved independently. The network structure includes: 1 hidden layer with 64 neurons, ReLU activation function; 1 output neuron with linear activation, corresponding to the water level; loss function: MSE; optimizer: Adam; learning rate: 0.

001.

6. The model-coupled lake level prediction method as described in claim 5, characterized in that, The specific training method is as follows: for each important forecasting station, representative forecasting station, or forecasting station of special concern, the simulated sequence of the forecasting station and the measured water level of the same period are used as physical consistency labels, and training, validation and test sets are divided. A sub-model weight file is trained independently and saved. Each sub-model outputs the predicted water level for the corresponding forecast station over the next 1-72 hours, as well as the NSE and RMSE accuracy indices.

7. The model-coupled lake level prediction method as described in claim 5, characterized in that, When adding new forecast sites, only the new sub-model needs to be trained; there is no need to modify the network framework.

8. The model-coupled lake level prediction method as described in claim 5, characterized in that, The training, validation, and test sets were divided in a ratio of 7:1.5:1.

5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the model-coupled lake level prediction method as described in any one of claims 1-8.

10. A computer device, characterized in that, include: A processor and a memory communicatively connected to the processor; The memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the model-coupled lake level prediction method as described in any one of claims 1-8.

Images