Salt tide probability forecasting method and device based on physical scene adaptive driving

Abstract

The invention relates to the technical field of ocean hydrology prediction, in particular to a salt tide probability forecasting method and device based on physical scene adaptive driving, and the method comprises the steps: obtaining a salinity actual measurement result, physical scene condition data, and a salinity prediction result of a numerical model and a deep learning model in the same time period; calculating the physical scene condition data by adopting a hidden Markov method to obtain a physical scene of each time step length; evaluating the performance of each model in each physical scene according to the data; inputting the data into a deep neural network coupled based on a time sequence two-dimensional change modeling network and a gating fusion network so as to calculate the optimal time-varying weight of each model; and inputting the data into a Bayesian model averaging method based on a Copula function so as to calculate a set probability forecast result based on coupling of a plurality of models. Therefore, the problems that in the prior art, salinity prediction under different scenes is poor in robustness, evaluation is uncertain and the like are solved.

Description

Technical Field

[0001] This invention relates to the field of marine hydrological prediction technology, and in particular to a method and apparatus for predicting saltwater intrusion probability based on physical scene adaptive driving. Background Technology

[0002] Saltwater intrusion is driven by multiple factors, including upstream runoff conditions, tidal dynamics, wind fields, temperature variations, estuarine topography and hydrodynamic patterns, and sea level elevation. Its evolution is typically characterized by large salinity amplitudes, significant temporal correlations, highly nonlinear and coupled influencing mechanisms, and substantial uncertainty. Therefore, developing high-precision, interpretable saltwater intrusion forecasts with the ability to characterize uncertainties is crucial for coastal city water intake scheduling and emergency response, optimization of agricultural irrigation plans, and management of estuarine ecological risks.

[0003] Currently, saltwater intrusion forecasting mainly relies on two types of methods: one is the hydrodynamic-salinity transport numerical model, which can reflect the physical process, but it is highly sensitive to boundary conditions and parameter calibration, has a high computational cost, and its error is easily amplified under the superposition of multiple uncertainties or extreme events; the other is a single machine learning model, which is computationally efficient and has a strong ability to fit nonlinearity, but it is heavily dependent on the coverage of training samples, lacks robustness in the face of extreme low water levels, strong storm surges, and sea level anomalies, and is insufficient in characterizing multi-factor coupling, time-varying dynamic mechanisms and error structures. It can usually only output a single predicted value, which is difficult to provide reliable uncertainty information for risk decision-making. Summary of the Invention

[0004] This invention provides a method and apparatus for predicting salinity probability based on physical scene adaptive driving, which solves the problems of poor robustness and uncertainty in the evaluation of salinity prediction under different scenarios in the existing technology. It can dynamically adjust the model weight structure between different models and different scenarios, realize targeted learning of multi-factor coupling and time-varying mechanisms, and improve the accuracy and generalization ability of salinity prediction.

[0005] A first aspect of the present invention provides a method for predicting the probability of saltwater intrusion based on physical scene adaptive driving, comprising the following steps: Acquire the actual salinity measurement results, physical scene condition data, the first salinity prediction result of the pre-built numerical model, and the second salinity prediction result of the pre-built deep learning model under the same time period; The hidden Markov method is used to calculate the physical scene condition data to obtain multiple physical scenes, and each physical scene is assigned to a corresponding time step to obtain the physical scene for each time step. Based on the physical scenario at each time step, the performance of each model in each physical scenario is evaluated according to the measured salinity results, the first salinity prediction results, and the second salinity prediction results. The measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, and the performance of each model in each physical scene are integrated into the first input data. The first input data is then input into a pre-constructed deep neural network based on a temporal two-dimensional change modeling network and a gated fusion network to calculate the optimal time-varying weights for each model. The measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, the physical scene at each time step, and the optimal time-varying weights of each model are integrated into the second input data. The second input data is then input into a pre-built Bayesian model averaging method based on the Copula function to calculate the ensemble probability prediction results based on the coupling of multiple models.

[0006] Optionally, the physical scene condition data includes tide level, upstream flow rate, and wind field.

[0007] Optionally, the numerical model is a finite-volume ocean community model, and the deep learning model includes a temporal convolutional network model, a deep recurrent autoregressive coupled self-attention mechanism-temporal convolutional network model, and a Fourier neural operator model.

[0008] Optionally, the measured salinity results, the physical scene condition data, the first salinity prediction result, and the second salinity prediction result are all time series data on an hourly scale.

[0009] Optionally, the evaluation of the performance of each model in each physical scenario based on the measured salinity results, the first salinity prediction result, and the second salinity prediction result for each time step includes: Based on the physical scenario at each time step, the measured salinity results, the first salinity prediction results, and the second salinity prediction results are input into preset evaluation metrics to obtain the performance of each model in each physical scenario. The preset evaluation metrics include the Kling-Gupta efficiency coefficient, the Nash-Satcliffe efficiency coefficient, the root mean square error, the mean absolute error, and the coefficient of determination.

[0010] Optionally, the step of integrating the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, and the performance of each model in each physical scene into first input data, and inputting the first input data into a pre-constructed deep neural network coupled with a temporal two-dimensional variation modeling network and a gated fusion network to calculate the optimal time-varying weights for each model, includes: Based on a time-series two-dimensional variation modeling network, an input sequence is constructed that includes the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, and the performance of each model in each physical scene. The input sequence is then linearly projected to obtain the corresponding aggregated sequence. Perform a discrete Fourier transform on the aggregated sequence to obtain the corresponding amplitude and period; The input sequence is sliced ​​using the period to obtain one-dimensional time series features; Multiple amplitudes are weighted and fused according to the significance of the period to obtain the period weight; Calculate the scenario state representation based on the one-dimensional time series features and the periodic weights; Based on the gated fusion network, the optimal time-varying weights of each model are calculated according to the scenario state representation.

[0011] Optionally, the step of integrating the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, the physical scene at each time step, and the optimal time-varying weights of each model into second input data, and inputting the second input data into a pre-built Bayesian model averaging method based on the Copula function to calculate the ensemble probability prediction result based on the coupling of multiple models, includes: The residuals of each model are defined based on the measured salinity results, the first salinity prediction result, and the second salinity prediction result; Based on the physical scene condition data and the physical scene at each time step, the residuals of each model are constructed to follow the scene condition distribution. Construct edge density and cumulative distribution functions based on the residuals following the scene condition distribution; The Copula joint distribution function is constructed based on the optimal time-varying weights of each model; A joint density function is constructed based on the marginal density and cumulative distribution function and the Copula joint distribution function, and the joint density function is weighted and fused using Bayesian methods to obtain the ensemble probability prediction result based on the coupling of multiple models.

[0012] A second aspect of the present invention provides a saltwater intrusion probability forecasting device based on physical scene adaptive driving, comprising: The acquisition module is used to acquire the measured salinity results, physical scene condition data, the first salinity prediction results of the pre-built numerical model, and the second salinity prediction results of the pre-built deep learning model under the same time period. The allocation module is used to calculate the physical scene condition data using the Hidden Markov Method to obtain multiple physical scenes, and to allocate each physical scene to the corresponding time step to obtain the physical scene for each time step. The evaluation module is used to evaluate the performance of each model in each physical scenario based on the physical scenario at each time step, according to the measured salinity results, the first salinity prediction results, and the second salinity prediction results. The weight calculation module is used to integrate the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, and the performance of each model in each physical scene into the first input data, and input the first input data into a pre-constructed deep neural network based on the coupling of a time-series two-dimensional change modeling network and a gated fusion network to calculate the optimal time-varying weights of each model. The probability prediction module is used to integrate the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, the physical scene at each time step, and the optimal time-varying weights of each model into second input data, and input the second input data into a pre-built Bayesian model averaging method based on the Copula function to calculate the ensemble probability prediction results based on the coupling of multiple models.

[0013] A third aspect of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the physical scene adaptive driving method for saltwater intrusion probability forecasting as described in the above embodiments.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting saline tide probability based on physical scene adaptive driving.

[0015] The proposed method and apparatus for predicting saltwater intrusion based on physical scene adaptive driving in this invention constructs a scenario representation that reflects the dominant mechanism of saltwater intrusion. It dynamically adjusts the model weight structure using a neural network method across different models and scenarios to achieve targeted learning of multi-factor coupling and time-varying mechanisms. Simultaneously, it evaluates the dependency distribution of different models under various physical scenarios and outputs the prediction distribution and interval uncertainty in probabilistic form. This not only improves accuracy and generalization ability but also directly serves water intake safety early warning, scheduling optimization, and ecological risk management.

[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0017] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart of a saltwater intrusion probability prediction method based on physical scene adaptive driving according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the specific execution of a physical scene adaptive-driven saltwater intrusion probability prediction method according to an embodiment of the present invention. Figure 3 This is a schematic diagram of a physical scene partitioning result provided by an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the error results of various prediction models under different physical scenarios according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a final set probability prediction output result provided by an embodiment of the present invention; Figure 6 This is a block diagram of a saltwater intrusion probability forecasting device based on physical scene adaptive driving according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention.

[0018] Explanation of reference numerals in the attached figures: 60-A probability forecasting device for saltwater intrusion based on physical scene adaptive drive; 601-Acquisition module; 602-Allocation module; 603-Evaluation module; 604-Weight calculation module; 605-Probability forecasting module; 701-Memory; 702-Processor; 703-Communication interface. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0020] The following describes, with reference to the accompanying drawings, a method and apparatus for predicting saltwater intrusion probability based on physical scene adaptive driving according to embodiments of the present invention.

[0021] Figure 1 This is a flowchart illustrating a physical scene adaptive-driven method for predicting the probability of saltwater intrusion, as provided in an embodiment of the present invention.

[0022] like Figure 1 As shown, the physical scene-adaptive driven method for predicting saltwater intrusion probability includes the following steps: In step S101, the measured salinity results, physical scene condition data, the first salinity prediction result of the pre-built numerical model, and the second salinity prediction result of the pre-built deep learning model are obtained under the same time period.

[0023] In some embodiments, physical scene condition data includes tide level, upstream flow rate, and wind field.

[0024] In some embodiments, the numerical model is a finite-volume ocean community model, and the deep learning model includes a temporal convolutional network model, a deep recurrent autoregressive coupled self-attention mechanism-temporal convolutional network model, and a Fourier neural operator model.

[0025] In some embodiments, the measured salinity results, physical scene condition data, first salinity prediction results, and second salinity prediction results are all time series data on an hourly scale.

[0026] like Figure 2 As shown, in the actual execution process, scenario condition data including tide level, upstream flow, and wind field, as well as measured salinity results, are obtained for the same time period. A pre-built numerical model is used to solve for the first salinity prediction result for the same time period, and a pre-built deep learning model is used to solve for the second salinity prediction result for the same time period. The numerical model includes, but is not limited to, the Finite Volume Ocean Community (FVCOM) model, and the deep learning model includes, but is not limited to, the Temporal Convolutional Network (TCN) model, the SA-TCN model based on a deep recurrent autoregressive coupled self-attention mechanism, and the Fourier Neural Operator (FNO) model.

[0027] It should be noted that the method for constructing the numerical model is as follows: Topographical data, upstream inflow, downstream open-boundary tides, surface wind field, and temperature data are obtained as initial and boundary conditions. A computational grid for the study watershed is constructed, and the aforementioned initial and boundary condition data are input into the computational grid for calculation. During this process, the model is debugged by adjusting the bottom friction coefficient to minimize the error between the model results and the measured data, thus obtaining a well-calibrated numerical model. Then, the initial and boundary conditions for the required calculation period are input into the model to obtain the first salinity prediction result for the required period.

[0028] The method for constructing a deep learning model is as follows: Real-world test data is divided into training, validation, and testing phases and input into the model. The Bayesian optimization algorithm is used to train the deep learning model, updating its parameters. The validation and testing phases are then used to validate and test the deep learning model and its results, resulting in a well-constructed deep learning model. Subsequently, historical test data is input into the model to obtain the second salinity prediction result from the deep learning model.

[0029] In step S102, the hidden Markov method is used to calculate the physical scene condition data to obtain multiple physical scenes, and each physical scene is assigned to the corresponding time step to obtain the physical scene for each time step.

[0030] like Figure 2 As shown, in actual execution, this embodiment of the invention uses the Hidden Markov Model to calculate the acquired physical scene condition data, obtain the frequency of each physical scene, and set the number of physical scene outputs to n. Then, the n physical scenes with the highest output frequency are numbered starting from 0, and the physical scene numbers are assigned to each time step to obtain the physical scene for each time step.

[0031] For example, such as Figure 3 As shown, based on the data, the number of independent scene outputs can be set to 6, numbered from 0 to 5. Each 20% of the data is used as a threshold for scene division. Each number corresponds to a specific physical scene as follows: Cluster 0 represents westerly winds, low southerly winds, medium water level, and low flow; Cluster 1 represents westerly winds, low southerly winds, medium water level, and high flow; Cluster 2 represents high westerly winds, high southerly winds, high water level, and low flow; Cluster 3 represents westerly winds, low southerly winds, medium water level, and low flow; Cluster 4 represents extremely low westerly winds, low southerly winds, medium water level, and medium flow; Cluster 5 represents high easterly winds, high southerly winds, medium water level, and high flow.

[0032] In step S103, based on the physical scenario at each time step, the performance of each model in each physical scenario is evaluated according to the measured salinity results, the first salinity prediction results, and the second salinity prediction results.

[0033] In some embodiments, based on the physical scenario at each time step, the performance of each model in each physical scenario is evaluated according to the measured salinity results, the first salinity prediction result, and the second salinity prediction result, including: Based on the physical scenario at each time step, the measured salinity results, the first salinity prediction results, and the second salinity prediction results are input into the preset evaluation metrics to obtain the performance of each model in each physical scenario. The preset evaluation metrics include the Kling-Gupta efficiency coefficient, the Nash-Satcliffe efficiency coefficient, the root mean square error, the mean absolute error, and the coefficient of determination.

[0034] like Figure 2 and 4 As shown, in the actual execution process, based on the physical scenario at each time step, the performance of each model in each physical scenario is evaluated by combining the first salinity prediction result of the numerical model, the second salinity prediction result of the deep learning model, and the salinity measurement results, so as to determine the error results of each prediction model in different physical scenarios.

[0035] The evaluation metrics include the Kling-Gupta efficiency coefficient (KGE), Nash-Sutcliffe efficiency coefficient (NSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The closer the values ​​of KGE, NSE, and R² are to 1, the better the model performance; the closer the values ​​of RMSE and MAE are to 0, the better the model performance.

[0036] The specific formula for calculating the evaluation indicators is as follows:

[0037]

[0038]

[0039]

[0040]

[0041] in, The Pearson correlation coefficient is the ratio between the measured and predicted values. It is the ratio of the standard deviations between the measured value and the predicted value. This is the ratio of the mean between the measured value and the predicted value. For the entire cycle, For each One cycle, and These are the measured value and the predicted value, respectively. These are the simulated values ​​of the variables in the regression model. This represents the average of the measured values.

[0042] In step S104, the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, and the performance of each model in each physical scene are integrated into the first input data. The first input data is then input into a pre-constructed deep neural network based on the coupling of a time-series two-dimensional change modeling network and a gated fusion network to calculate the optimal time-varying weights for each model.

[0043] In some embodiments, the measured salinity results, the first salinity prediction results, the second salinity prediction results, physical scene condition data, and the performance of each model in each physical scene are integrated into the first input data. This first input data is then fed into a pre-constructed deep neural network coupled with a temporal two-dimensional variation modeling network and a gated fusion network to calculate the optimal time-varying weights for each model, including: Based on the temporal two-dimensional variation modeling network, an input sequence is constructed that includes measured salinity results, first salinity prediction results, second salinity prediction results, physical scene condition data, and the performance of each model in each physical scene. The input sequence is then linearly projected to obtain the corresponding aggregate sequence. Perform a discrete Fourier transform on the aggregated sequence to obtain the corresponding amplitude and period; The input sequence is sliced ​​using the period to obtain one-dimensional time series features; Multiple amplitudes are weighted and fused according to the significance of the period to obtain the period weight; Scenario state representation is calculated based on one-dimensional time series characteristics and periodic weights; Based on the gated fusion network, the optimal time-varying weights of each model are calculated according to the scenario state representation.

[0044] like Figure 2 As shown, in the actual implementation process, let the time series of the salinity measurement results be... , For time steps, there exists Results of numerical models and deep learning models For the driving feature vector that includes physical scene conditions, the first... The prediction given by the basic model is First, a Fast Fourier Transform (FFT) operation is performed to find the dominant period: using a period of length... Sliding window construction of input sequence And obtained through linear projection Corresponding aggregation sequence The amplitude is obtained by performing a Discrete Fourier Transform (DFT) on the aggregated sequence. :

[0045] Furthermore, select the one with the largest amplitude. frequency And convert it into a period:

[0046] in, The most prominent one in the corresponding window The periodic scale of an individual.

[0047] and the initial value of each cycle For the input sequence Slicing to obtain This forms a two-dimensional image structure of "number of blocks × periodic phase," which facilitates convolution and multi-scale extraction. Then, convolution kernels of different lengths are used... right By performing splicing and normalization, short-scale kernels focus on abrupt changes, while long-scale kernels capture slow-changing trends. The combination can simultaneously describe both periodic structure and the speed of change. Then, the dimensionality is reduced back to one dimension to obtain one-dimensional time series features. .

[0048] Furthermore, the Softmax operation is used to fuse multiple periodic branches by periodic significance weighting. Each period, its period weight for:

[0049] Among them, parameters Control the degree of strong cyclical preferences.

[0050] After pooling, the TimeNet network for modeling two-dimensional temporal variations finally outputs a... Contextual state representation at any moment :

[0051] Furthermore, based on the gated fusion network GFN, according to the scenario state representation, it performs... Time-based dynamic weights The calculation is performed, and the output is the optimal time-varying weight for each prediction model at each time step. The specific formula is as follows:

[0052] in, The scenario state output by TimesNet. and These are the hidden layer mapping parameters. It is a ReLU nonlinear activation function. and For output to Scoring of each model It was ensured that the weights assigned to each model were all greater than 0 and The sum is 1.

[0053] In this embodiment, the training results of the above process are: KGE value 0.874, NSE value 0.836, R² value 0.836, MAE value 0.936, and RMSE value 1.212. The validation results are: KGE value 0.855, NSE value 0.715, R² value 0.736, MAE value 0.936, and RMSE value 1.212. It can be seen that the error indicators are relatively good.

[0054] In step S105, the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, the physical scene at each time step, and the optimal time-varying weights of each model are integrated into the second input data. The second input data is then input into the pre-built Bayesian model averaging method based on the Copula function to calculate the ensemble probability prediction results based on the coupling of multiple models.

[0055] In some embodiments, measured salinity results, first salinity prediction results, second salinity prediction results, physical scene condition data, physical scenes at each time step, and the optimal time-varying weights of each model are integrated into second input data. This second input data is then fed into a pre-built Bayesian model averaging method based on a Copula function to calculate ensemble probability prediction results based on the coupling of multiple models, including: The residuals of each model are defined based on the measured salinity results, the first salinity prediction result, and the second salinity prediction result; Based on the physical scene condition data and the physical scene at each time step, the residuals of each model are constructed to follow the scene condition distribution. Construct edge density and cumulative distribution functions based on the residuals following the scene condition distribution; Construct the Copula joint distribution function based on the optimal time-varying weights for each model; A joint density function is constructed based on the marginal density, cumulative distribution function, and Copula joint distribution function. The joint density function is then weighted and fused using Bayesian methods to obtain ensemble probability prediction results based on the coupling of multiple models.

[0056] like Figure 2As shown, in the actual execution process, the input data is assumed to be: the time series of salinity measurement results. , A prediction model Provide its predicted value Each time step Corresponding to a scene category tag Parameters containing physical scene condition data The weights of each prediction model at each time step, calculated by the TimeNet two-dimensional variation modeling network and the gated fusion network GFN. .

[0057] Conditional error modeling is performed on the edge layer of each prediction model. Specifically, the residual of each model is defined based on the measured salinity results, the first salinity prediction result, and the second salinity prediction result:

[0058] Furthermore, based on the physical scene condition data and the physical scene at each time step, the residuals of each model are constructed to follow the scene condition distribution:

[0059] in, For specific scene numbers, For the model In the scene The scene distribution function under, The models are respectively In the scene The mean and standard deviation are given. Therefore, the marginal density and cumulative distribution function are:

[0060]

[0061] in, Indicates the scene c Next, the k The probability density of the model residuals. Let be the probability density function of the standard normal distribution. For the corresponding cumulative distribution, Map the residual to [0 , [1] This space serves as the input variable for subsequent Copula layers. Different scenarios... and These differences can reflect changes in the model's error characteristics.

[0062] Furthermore, conditional dependency structure modeling is performed at the Copula layer, and the joint dependencies between models are represented by scene-conditional Vine-Copula:

[0063] Parameters change dynamically:

[0064] In the formula, For the scene c The Copula joint distribution function under the following conditions Copula is a binary condition for Vine-Copula. For a given condition variable In the case of the first The conditional distribution function values ​​of the variables. For a given condition variable In the case of the first The conditional distribution function values ​​of the variables. and These are all corresponding dynamic parameters that control dependency strength or tail relationships. The corresponding mean, The time autocorrelation coefficient, This is a random perturbation term. This layer is used to characterize the dependency structure between model predictions under different scenarios.

[0065] Constructing a complete joint density function by combining the marginal density, cumulative distribution function, and Copula joint distribution function. :

[0066] In the formula, Representing a scene Joint dependency density of the model residuals The edge density of each model.

[0067] like Figure 5 As shown, the joint density function is weighted and fused using Bayesian weighting to obtain the final conditional distribution, which is the ensemble probability prediction result based on the coupling of multiple models:

[0068] In the formula, This is an ensemble probabilistic forecast result based on the coupling of multiple models. For the scene Below, model The weight value.

[0069] In summary, the saltwater intrusion probability forecasting method based on physical scenario adaptive driving proposed in this embodiment of the invention constructs a scenario representation that reflects the dominant mechanism of saltwater intrusion, and dynamically adjusts the model weight structure using a neural network method among different models and scenarios to achieve targeted learning of multi-factor coupling and time-varying mechanisms. At the same time, it evaluates the dependency distribution relationship of different models under various physical scenarios and outputs the prediction distribution and interval uncertainty in probabilistic form. This method can improve accuracy and generalization ability, and directly serve water intake safety early warning, scheduling optimization and ecological risk management.

[0070] Next, referring to the accompanying drawings, a saltwater intrusion probability forecasting device based on physical scene adaptive driving proposed according to an embodiment of the present invention is described.

[0071] Figure 6 This is a block diagram of a saltwater intrusion probability forecasting device based on physical scene adaptive driving provided in an embodiment of the present invention.

[0072] like Figure 6 As shown, the saltwater intrusion probability forecasting device 60 based on physical scene adaptive driving includes: an acquisition module 601, an allocation module 602, an evaluation module 603, a weight calculation module 604, and a probability forecasting module 605.

[0073] The system comprises the following modules: Acquisition module 601 acquires measured salinity results, physical scene condition data, a first salinity prediction result from a pre-built numerical model, and a second salinity prediction result from a pre-built deep learning model within the same time period. Allocation module 602 uses a Hidden Markov Model to calculate the physical scene condition data to obtain multiple physical scenes, and allocates each physical scene to a corresponding time step to obtain physical scenes for each time step. Evaluation module 603 evaluates the performance of each model in each physical scene based on the measured salinity results, the first salinity prediction result, and the second salinity prediction result, according to the physical scenes at each time step. Weight calculation module 604 integrates the measured salinity results, the first salinity prediction result, the second salinity prediction result, the physical scene condition data, and the performance of each model in each physical scene into first input data, and inputs this first input data into a pre-built deep neural network coupled with a temporal two-dimensional variation modeling network and a gated fusion network to calculate the optimal time-varying weights for each model. The probability prediction module 605 is used to integrate the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, the physical scene at each time step, and the optimal time-varying weights of each model into the second input data, and input the second input data into the pre-built Bayesian model averaging method based on the Copula function to calculate the ensemble probability prediction results based on the coupling of multiple models.

[0074] In some embodiments, physical scene condition data includes tide level, upstream flow rate, and wind field.

[0075] In some embodiments, the numerical model is a finite-volume ocean community model, and the deep learning model includes a temporal convolutional network model, a deep recurrent autoregressive coupled self-attention mechanism-temporal convolutional network model, and a Fourier neural operator model.

[0076] In some embodiments, the measured salinity results, physical scene condition data, first salinity prediction results, and second salinity prediction results are all time series data on an hourly scale.

[0077] In some embodiments, the evaluation module 603 includes: Based on the physical scenario at each time step, the measured salinity results, the first salinity prediction results, and the second salinity prediction results are input into the preset evaluation metrics to obtain the performance of each model in each physical scenario. The preset evaluation metrics include the Kling-Gupta efficiency coefficient, the Nash-Satcliffe efficiency coefficient, the root mean square error, the mean absolute error, and the coefficient of determination.

[0078] In some embodiments, the weight calculation module 604 includes: The projection unit is used to construct an input sequence for the time-series two-dimensional variation modeling network, which includes measured salinity results, first salinity prediction results, second salinity prediction results, physical scene condition data, and the performance of each model in each physical scene. The input sequence is then linearly projected to obtain the corresponding aggregated sequence. The transformation unit is used to perform a discrete Fourier transform on the aggregated sequence to obtain the corresponding amplitude and period; A slicing unit is used to slice the input sequence using a period to obtain one-dimensional time series features; The first weighted fusion unit is used to weight and fuse multiple amplitudes according to the significance of the period to obtain the period weight; The first computing unit is used to calculate the scenario state representation based on one-dimensional time series features and periodic weights. The second computational unit is used to calculate the optimal time-varying weights of each model based on the scenario state representation using a gated fusion network.

[0079] In some embodiments, the probability prediction module 605 includes: Define the unit to define the residuals of each model based on the measured salinity results, the first salinity prediction result, and the second salinity prediction result; The first building unit is used to construct residuals that conform to the scene condition distribution based on physical scene condition data and the physical scene at each time step, according to the residuals of each model. The second building unit is used to construct the edge density and cumulative distribution function based on the residual following the scene condition distribution. The third building block is used to construct the Copula joint distribution function based on the optimal time-varying weights of each model; The second weighted fusion unit is used to construct a joint density function based on the marginal density, cumulative distribution function and Copula joint distribution function, and to perform weighted fusion of the joint density function through Bayesian methods to obtain ensemble probability prediction results based on the coupling of multiple models.

[0080] It should be noted that the foregoing explanation of the embodiment of the saltwater intrusion probability forecasting method based on physical scene adaptive driving also applies to the saltwater intrusion probability forecasting device based on physical scene adaptive driving in this embodiment, and will not be repeated here.

[0081] The saltwater intrusion probability forecasting device based on physical scene adaptive driving proposed in this embodiment of the invention constructs a scenario representation that reflects the dominant mechanism of saltwater intrusion, and dynamically adjusts the model weight structure using a neural network method among different models and scenarios to achieve targeted learning of multi-factor coupling and time-varying mechanisms; at the same time, it evaluates the dependency distribution relationship of different models under various physical scenarios, and outputs the prediction distribution and interval uncertainty in probabilistic form, which can improve accuracy and generalization ability, and directly serve water intake safety early warning, scheduling optimization and ecological risk management.

[0082] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0083] The electronic device may include: a memory 701, a processor 702, and a computer program stored on the memory 701 and capable of running on the processor 702.

[0084] When the processor 702 executes the program, it implements the saltwater intrusion probability prediction method based on physical scene adaptive driving provided in the above embodiments.

[0085] Furthermore, electronic devices also include: Communication interface 703 is used for communication between memory 701 and processor 702.

[0086] The memory 701 is used to store computer programs that can run on the processor 702.

[0087] The memory 701 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0088] If the memory 701, processor 702, and communication interface 703 are implemented independently, then the communication interface 703, memory 701, and processor 702 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 7 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0089] Optionally, in a specific implementation, if the memory 701, processor 702, and communication interface 703 are integrated on a single chip, then the memory 701, processor 702, and communication interface 703 can communicate with each other through an internal interface.

[0090] The processor 702 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0091] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting saline tide probability based on physical scene adaptive driving.

[0092] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0093] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0094] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0095] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0096] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0097] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments.

[0098] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0099] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting the probability of saltwater intrusion based on physical scene adaptive driving, characterized in that, Includes the following steps: Acquire the actual salinity measurement results, physical scene condition data, the first salinity prediction results of the pre-built numerical model, and the second salinity prediction results of the pre-built deep learning model under the same time period; The hidden Markov method is used to calculate the physical scene condition data to obtain multiple physical scenes, and each physical scene is assigned to a corresponding time step to obtain the physical scene for each time step. Based on the physical scenario at each time step, the performance of each model in each physical scenario is evaluated according to the measured salinity results, the first salinity prediction results, and the second salinity prediction results. The measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, and the performance of each model in each physical scene are integrated into the first input data. The first input data is then input into a pre-constructed deep neural network based on a temporal two-dimensional change modeling network and a gated fusion network to calculate the optimal time-varying weights for each model. The measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, the physical scene at each time step, and the optimal time-varying weights of each model are integrated into the second input data. The second input data is then input into a pre-built Bayesian model averaging method based on the Copula function to calculate the ensemble probability prediction results based on the coupling of multiple models.

2. The method for predicting saline tide probability based on physical scene adaptive driving according to claim 1, characterized in that, The physical scene condition data includes tide level, upstream flow rate, and wind field.

3. The method for predicting saline tide probability based on physical scene adaptive driving according to claim 1, characterized in that, The numerical model is a finite-volume ocean community model, and the deep learning model includes a temporal convolutional network model, a deep recurrent autoregressive coupled self-attention mechanism-temporal convolutional network model, and a Fourier neural operator model.

4. The method for predicting saline tide probability based on physical scene adaptive driving according to claim 1, characterized in that, The measured salinity results, the physical scene condition data, the first salinity prediction result, and the second salinity prediction result are all time series data on an hourly scale.

5. The method for predicting saline tide probability based on physical scene adaptive driving according to claim 1, characterized in that, The physical scenario based on each time step, evaluating the performance of each model in each physical scenario according to the measured salinity results, the first salinity prediction result, and the second salinity prediction result, includes: Based on the physical scenario at each time step, the measured salinity results, the first salinity prediction results, and the second salinity prediction results are input into preset evaluation metrics to obtain the performance of each model in each physical scenario. The preset evaluation metrics include the Kling-Gupta efficiency coefficient, the Nash-Satcliffe efficiency coefficient, the root mean square error, the mean absolute error, and the coefficient of determination.

6. The method for predicting saline tide probability based on physical scene adaptive driving according to claim 1, characterized in that, The process of integrating the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, and the performance of each model in each physical scene into first input data, and inputting the first input data into a pre-constructed deep neural network based on a temporal two-dimensional variation modeling network and a gated fusion network to calculate the optimal time-varying weights for each model, includes: Based on a time-series two-dimensional variation modeling network, an input sequence is constructed that includes the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, and the performance of each model in each physical scene. The input sequence is then linearly projected to obtain the corresponding aggregated sequence. Perform a discrete Fourier transform on the aggregated sequence to obtain the corresponding amplitude and period; The input sequence is sliced ​​using the period to obtain one-dimensional time series features; Multiple amplitudes are weighted and fused according to the significance of the period to obtain the period weight; Calculate the scenario state representation based on the one-dimensional time series features and the periodic weights; Based on the gated fusion network, the optimal time-varying weights of each model are calculated according to the scenario state representation.

7. The method for predicting saline tide probability based on physical scene adaptive driving according to claim 1, characterized in that, The process of integrating the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, the physical scene at each time step, and the optimal time-varying weights of each model into second input data, and inputting the second input data into a pre-constructed Bayesian model averaging method based on the Copula function to calculate the ensemble probability prediction result based on the coupling of multiple models, includes: The residuals of each model are defined based on the measured salinity results, the first salinity prediction result, and the second salinity prediction result; Based on the physical scene condition data and the physical scene at each time step, the residuals of each model are constructed to follow the scene condition distribution. Construct edge density and cumulative distribution functions based on the residuals following the scene condition distribution; The Copula joint distribution function is constructed based on the optimal time-varying weights of each model; A joint density function is constructed based on the marginal density and cumulative distribution function and the Copula joint distribution function, and the joint density function is weighted and fused using Bayesian methods to obtain the ensemble probability prediction result based on the coupling of multiple models.

8. A saltwater intrusion probability forecasting device based on physical scene adaptive driving, characterized in that, include: The acquisition module is used to acquire the measured salinity results, physical scene condition data, the first salinity prediction results of the pre-built numerical model, and the second salinity prediction results of the pre-built deep learning model under the same time period. The allocation module is used to calculate the physical scene condition data using the Hidden Markov Method to obtain multiple physical scenes, and to allocate each physical scene to the corresponding time step to obtain the physical scene for each time step. The evaluation module is used to evaluate the performance of each model in each physical scenario based on the physical scenario at each time step, according to the measured salinity results, the first salinity prediction results, and the second salinity prediction results. The weight calculation module is used to integrate the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, and the performance of each model in each physical scene into the first input data, and input the first input data into a pre-constructed deep neural network based on the coupling of a time-series two-dimensional change modeling network and a gated fusion network to calculate the optimal time-varying weights of each model. The probability prediction module is used to integrate the measured salinity results, the first salinity prediction results, the second salinity prediction results, the physical scene condition data, the physical scene at each time step, and the optimal time-varying weights of each model into second input data, and input the second input data into a pre-built Bayesian model averaging method based on the Copula function to calculate the ensemble probability prediction results based on the coupling of multiple models.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the physical scene adaptive-driven saltwater tide probability forecasting method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the physical scene adaptive-driven saltwater intrusion probability forecasting method as described in any one of claims 1-7.

Images