Land subsidence risk prediction method, device and system based on recurrent neural network

Abstract

This invention provides a method, apparatus, and system for predicting land subsidence hazard based on a recurrent neural network. The method includes: acquiring spatiotemporal distribution information of land subsidence and water level; obtaining geological control parameters and regression parameters characterizing the spatial relationship between the input variables of the recurrent neural network and land subsidence based on the spatiotemporal distribution information of land subsidence and water level; constructing a spatiotemporal model of the recurrent neural network for predicting land subsidence hazard; inputting the geological control parameters and subsidence inducing factors as input variables into the input layer of the spatiotemporal model of the recurrent neural network, and inputting the regression parameters into the spatial relationship layer of the spatiotemporal model of the recurrent neural network; and predicting the land subsidence value using the spatiotemporal model of the recurrent neural network. This invention can improve prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of land subsidence risk assessment caused by groundwater extraction, and in particular to a method, apparatus and system for predicting land subsidence risk based on recurrent neural networks. Background Technology

[0002] Land subsidence caused by excessive groundwater extraction is a geological hazard that hinders sustainable urban development. Land subsidence simulation and prediction are crucial for the rational extraction of groundwater resources and the prevention and control of land subsidence. Constructing constitutive-based land subsidence simulation models requires numerous hydrogeological and soil mechanics parameters, as well as geological structures that accurately reflect the lithological distribution of strata. However, the spatial distribution of these parameters is sparse, often making it difficult to meet the requirements for regional-scale models. Data-driven temporal simulation models of land subsidence (e.g., Grey Model (GM), regression analysis models, Artificial Intelligence (AI) models) start from the data itself, exploring the temporal development relationship between land subsidence and influencing factors. These models offer high computational efficiency, require fewer parameters, can perform short-term predictions, and demonstrate good simulation results.

[0003] For stratigraphic structures with strong spatial heterogeneity, there is a temporal nonlinear relationship between land subsidence and related influencing factors such as hydrogeological parameters. This relationship also varies spatially, and the degree of land subsidence development is closely related to soil physical properties. Existing data-driven models do not simultaneously consider the spatiotemporal relationship between hydrogeological parameters and land subsidence, nor the controlling effect of soil physical properties (such as soil compressibility) on land subsidence. Persistent Scattered Interferometric Synthetic Aperture Radar (PSI) technology can extract temporal deformation of permanent scatterers (PS) to obtain large-scale, long-term, high-precision, and high-density land subsidence information. Combined with the rapidly developing deep learning technology in recent years, it provides data support and technical backing for data-driven land subsidence simulation.

[0004] Traditional hydrogeological models are difficult to apply due to the large number of parameters and the difficulty in obtaining these parameters. Data-driven models also fail to consider the spatiotemporal relationship and physical mechanisms of influencing factors and land subsidence simultaneously, resulting in low prediction accuracy. Summary of the Invention

[0005] In view of this, the present invention provides a method, apparatus and system for predicting ground subsidence risk based on recurrent neural networks, so as to improve prediction accuracy.

[0006] On the one hand, the present invention provides a method for predicting land subsidence risk based on recurrent neural networks, including: acquiring spatiotemporal distribution information of land subsidence and spatiotemporal distribution information of water level;

[0007] Based on the spatiotemporal distribution information of ground subsidence and the spatiotemporal distribution information of water level, geological control parameters and regression parameters used to characterize the spatial relationship between the input variables of the recurrent neural network and ground subsidence are obtained.

[0008] Construct a spatiotemporal model of a recurrent neural network for predicting land subsidence risk;

[0009] The geological control parameters and subsidence inducing factors are input as input variables into the input layer of the recurrent neural network spatiotemporal model, and the regression parameters are input into the spatial relationship layer of the recurrent neural network spatiotemporal model. The ground subsidence value is then predicted using the recurrent neural network spatiotemporal model.

[0010] Furthermore, the steps of obtaining the spatiotemporal distribution information of ground subsidence and the spatiotemporal distribution information of water level include:

[0011] Extracting spatiotemporal distribution information of ground subsidence based on PSI technology;

[0012] Information on the spatiotemporal distribution of water level is obtained using universal kriging interpolation based on error correction.

[0013] Furthermore, the geological control parameters include geological, hydrogeological, and soil mechanics parameters such as the inelastic release coefficient, elastic release coefficient, compressibility coefficient, and compressible layer thickness of the aquifer system;

[0014] The steps of obtaining geological control parameters and regression parameters for characterizing the spatial relationship between the input variables of the recurrent neural network and land subsidence based on the spatiotemporal distribution information of land subsidence and water level include:

[0015] The inelastic water release coefficient of the aquifer system is obtained based on the spatiotemporal distribution information of ground subsidence and the spatiotemporal distribution information of water level.

[0016] A geographic weighted regression model was used to analyze the spatial relationship between the input variables of the recurrent neural network and ground subsidence at each time point in the time series, and the regression parameters at different locations and times were obtained.

[0017] Furthermore, the step of constructing a recurrent neural network spatiotemporal model for predicting land subsidence risk includes:

[0018] A recurrent neural network spatiotemporal model for predicting land subsidence risk is constructed, consisting of an input layer, a spatial relationship layer, a hidden layer, an output layer, and an optimization layer.

[0019] The input layer is used to receive the geological control parameters and subsidence inducing factors;

[0020] The spatial relationship layer is used to characterize the spatial correlation between the input variables of the input layer and the PSI-observed settlement;

[0021] The hidden layer includes a multi-layer LSTM model. In each layer of the multi-layer LSTM model, the unit state value and output value of the previous time step are input into the memory cell of the next time step. Each memory cell contains multiple neurons. The dimension transformation is performed through a fully connected layer to convert the output dimension into the dimension used for optimizing the spatiotemporal model of the recurrent neural network.

[0022] The output layer is used to output the ground settlement value at each moment;

[0023] The optimization layer is used to jointly calculate the loss based on the ground subsidence value and PSI observation value at each time moment, and to use the gradient descent method to solve for the recurrent neural network spatiotemporal model parameters when the loss is minimized, thereby optimizing the recurrent neural network spatiotemporal model.

[0024] Furthermore, after the step of predicting the ground subsidence value using a recurrent neural network spatiotemporal model, the method further includes:

[0025] Spatial interpolation is performed on the predicted ground settlement values ​​at PS points to create a regional settlement distribution map.

[0026] Furthermore, the settlement-inducing factors include the water level parameters of the confined aquifer.

[0027] The present invention also provides a ground subsidence hazard prediction device based on a recurrent neural network, comprising:

[0028] The data acquisition module is used to acquire spatiotemporal distribution information of ground subsidence and spatiotemporal distribution information of water level;

[0029] The data processing module is used to obtain geological control parameters and regression parameters that characterize the spatial relationship between the input variables of the recurrent neural network and the ground subsidence, based on the spatiotemporal distribution information of ground subsidence and the spatiotemporal distribution information of water level.

[0030] The model building module is used to build a recurrent neural network spatiotemporal model for predicting land subsidence risk.

[0031] The settlement hazard prediction module is used to input the geological control parameters and settlement inducing factors as input variables into the input layer of the recurrent neural network spatiotemporal model, and input the regression parameters into the spatial relationship layer of the recurrent neural network spatiotemporal model, and use the recurrent neural network spatiotemporal model to predict the ground settlement value.

[0032] Furthermore, the model building module is specifically used to construct the input layer, spatial relationship layer, hidden layer, output layer, and optimization layer of a recurrent neural network spatiotemporal model for predicting land subsidence risk, wherein;

[0033] The input layer is used to receive the geological control parameters and subsidence inducing factors;

[0034] The spatial relationship layer is used to characterize the spatial correlation between the input variables of the input layer and the PSI-observed settlement;

[0035] The hidden layer includes a multi-layer LSTM model. In each layer of the multi-layer LSTM model, the unit state value and output value of the previous time step are input into the memory cell of the next time step. Each memory cell contains multiple neurons. The dimension transformation is performed through a fully connected layer to convert the output dimension into the dimension used for optimizing the spatiotemporal model of the recurrent neural network.

[0036] The output layer is used to output the ground settlement value at each moment;

[0037] The optimization layer is used to jointly calculate the loss based on the ground subsidence value and PSI observation value at each time moment, and to use the gradient descent method to solve for the recurrent neural network spatiotemporal model parameters when the loss is minimized, thereby optimizing the recurrent neural network spatiotemporal model.

[0038] The geological control parameters include geological, hydrogeological, and soil mechanics parameters such as the inelastic water release coefficient, elastic water release coefficient, compressibility coefficient, and compressible layer thickness of the aquifer system.

[0039] The subsidence-inducing factors include the water level parameters of the confined aquifer.

[0040] The present invention also provides a computer-readable storage medium storing a program, the program including instructions for performing the above-described methods.

[0041] The present invention also provides a ground subsidence hazard prediction system based on a recurrent neural network, comprising a readable medium storing a computer program, the program including instructions for performing the above-described method.

[0042] This invention relates to a method, device, and system for predicting land subsidence risk based on recurrent neural networks. When establishing a data-driven model, it considers the nonlinear temporal changes and spatial heterogeneity of geological, hydrogeological, and soil mechanics parameters in strongly heterogeneous strata, leading to the nonlinear evolution of land subsidence and spatial differences. It establishes a spatiotemporal simulation method for land subsidence that considers the different compressibility properties of soil at different locations, for regional-scale land subsidence prediction, thereby providing support for the scientific prevention and control of land subsidence. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments 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 these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a ground subsidence risk prediction method based on a recurrent neural network according to an exemplary first embodiment of the present invention.

[0045] Figure 2 This is a flowchart of a ground subsidence risk prediction method based on a recurrent neural network according to an exemplary second embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the spatiotemporal recurrent neural network model (Geo-GWLSTM) for land subsidence risk prediction method based on recurrent neural networks according to an exemplary third embodiment of the present invention.

[0047] Figure 4 This is a schematic diagram of the spatiotemporal recurrent neural network model (Geo-GWLSTM) for land subsidence risk prediction method based on recurrent neural networks in an exemplary third embodiment of the present invention.

[0048] Figure 5 This is a schematic diagram of fixed-duration data preprocessing during the prediction period of the Geo-GWLSTM spatiotemporal recurrent neural network model for land subsidence risk prediction method based on a recurrent neural network according to an exemplary third embodiment of the present invention.

[0049] Figure 6 This document describes the optimization training process of the spatiotemporal recurrent neural network model (Geo-GWLSTM) for land subsidence hazard prediction in the exemplary third embodiment of the present invention.

[0050] Figure 7 This is a structural block diagram of a ground subsidence hazard prediction device based on a recurrent neural network according to an exemplary fourth embodiment of the present invention. Detailed Implementation

[0051] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0052] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0053] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0054] like Figure 1 As shown, the present invention provides a method for predicting ground subsidence risk based on a recurrent neural network, comprising:

[0055] Step 101: Obtain the spatiotemporal distribution information of ground subsidence and water level.

[0056] Step 102: Based on the spatiotemporal distribution information of ground subsidence and the spatiotemporal distribution information of water level, obtain geological control parameters and regression parameters used to characterize the spatial relationship between the input variables of the recurrent neural network and ground subsidence.

[0057] Step 103: Construct a recurrent neural network spatiotemporal model for predicting land subsidence risk.

[0058] Step 104: Input the geological control parameters and settlement inducing factors as input variables into the input layer of the recurrent neural network spatiotemporal model, and input the regression parameters into the spatial relationship layer of the recurrent neural network spatiotemporal model. Use the recurrent neural network spatiotemporal model to predict the ground settlement value.

[0059] This embodiment establishes a data-driven spatiotemporal simulation method for land subsidence that simultaneously considers spatiotemporal and physical characteristics, for regional-scale land subsidence simulation and prediction, thereby providing support for the scientific prevention and control of land subsidence.

[0060] like Figure 2 As shown, another method for predicting ground subsidence risk based on recurrent neural networks is presented in this invention. Figure 1The preferred embodiment of the method described herein, the land subsidence recurrent neural network spatiotemporal prediction method considering physical mechanisms, optimizes and integrates PSI, a geographically weighted recurrent neural network model, and a land subsidence physical model. Considering the physical mechanisms of land subsidence, and addressing the land subsidence problem induced by groundwater extraction in heterogeneous geological environments, it constructs a data-driven land subsidence spatiotemporal simulation model (Geo-GWLSTM) that reflects physical mechanisms, providing a land subsidence spatiotemporal simulation framework based on a physical model, PSI technology, and deep learning. Specifically, it includes three parts: spatiotemporal dataset acquisition, construction of a land subsidence recurrent neural network spatiotemporal model considering physical mechanisms, and land subsidence spatiotemporal prediction. Wherein:

[0061] Part 1: Spatiotemporal Dataset Acquisition. Includes data acquisition and data analysis modules.

[0062] (1) Data acquisition module. This module includes the extraction of spatiotemporal distribution information of land subsidence based on PSI technology and the acquisition of spatiotemporal distribution information of water level based on error-corrected universal kriging interpolation.

[0063] ① Extracting spatiotemporal distribution information of land subsidence based on PSI technology. The PSI technology is used to extract land subsidence information B over time series t at the regional scale, denoted as... k, i≥1, where This represents the cumulative settlement at the m-th PS point at the k-th time, expressed in millimeters (mm).

[0064] ② Obtaining the spatiotemporal distribution information of water level using the universal Kriging interpolation method based on error correction. Combining water level contour data, the spatial error of the interpolation results is analyzed, and the spatial interpolation results of the water level observation points are corrected to obtain the temporal groundwater level information H for each PS point, denoted as... in This represents the groundwater level at the m-th PS point at the k-th time, in meters (m).

[0065] The specific method for correction is as follows:

[0066] The first step involves spatial interpolating the groundwater level observation data using the universal kriging method to obtain a preliminary spatial distribution GL of the groundwater level. int The groundwater level contour data is discretized, and then the discretized point data is interpolated to obtain the true spatial distribution of groundwater levels.

[0067] The second step is to compare and analyze the spatial distribution characteristics of the flow field data obtained from the observation points and contour lines respectively. Taking into account the wet season, dry season and monthly precipitation, the spatial distribution of the error ε of the two flow field data is calculated according to the following formula (1). For the wet season, ε is calculated as the contour line water level GL during the wet season. isoJand interpolated groundwater level GL intJ Subtraction: For the dry season, ε is calculated using the contour water level GL during the dry season. isoD and interpolated groundwater level GL intD Subtract. Remove the interpolation error from the original interpolation result according to the following formula (2), thus obtaining the corrected spatial distribution of groundwater level GL. new .

[0068]

[0069] GL new =GL int +ε (2)

[0070] (2) Data Analysis Module. Based on the acquired water level and settlement datasets, hydrogeological information is inverted and spatiotemporal relationships are analyzed. It includes two sub-modules: hydrogeological information inversion and spatiotemporal relationship analysis.

[0071] ① Hydrogeological Information Inversion Submodule. Inelastic water release coefficient S of aquifer system. kv Reflecting the compressibility properties of the soil, it is a crucial physical parameter controlling the intensity of ground settlement. The inelastic release coefficient of the aquifer system is retrieved by combining settlement and water level information. The preconsolidation head h is selected based on historical water level information and groundwater development and utilization. k By combining the pre-consolidation head, the groundwater level is decomposed into a water level that exhibits a long-term downward trend (i.e., the water level Δh acting on inelastic deformation). v It consists of two parts: water level and seasonal fluctuations.

[0072] When the compressibility of the liquid is negligible and elastic compression is ignored, the ground settlement obtained in step (1) is all due to the inelastic deformation of the soil. Therefore, the change in water level Δh below the preconsolidation head is used to determine the settlement. v and the inelastic deformation Δb caused by the change in water level. v The inelastic release coefficient of the aquifer system was calculated, as shown in formula (3) below. The inelastic release coefficient of the aquifer system at different spatial locations (i.e., the m-th PS point) This refers to the water level drop in the same month of different years compared to the pre-consolidation head. and contemporaneous deformation The ratio of .

[0073]

[0074] ② Spatiotemporal Relationship Analysis Submodule. This module uses a geographically weighted regression (GWR) model to analyze the spatial relationship between the model input variables and ground subsidence at each time point in the time series, obtaining regression parameters at different locations and times, which are then used as spatial weights.

[0075] Part 2: Construction of a spatiotemporal model of a recurrent neural network for land subsidence considering physical mechanisms.

[0076] To address the temporal nonlinearity and spatial variability between land settlement and groundwater level in heterogeneous strata and lithological environments, as well as the control effect of soil compressibility on land settlement, a spatiotemporal simulation model of land settlement recurrent neural network considering physical mechanisms is constructed, and the settlement simulation model is trained and optimized.

[0077] The first step is model architecture construction. Inherent hydrogeological and geological conditions are used as physical control variables to construct a spatiotemporal recurrent neural network model of land subsidence (Geology and Geographically Weighted Long Short-Term Memory, Geo-GWLSTM), which considers physical mechanisms. Figure 3 As shown, a Long Short-Term Memory (LSTM) model is used in the temporal sequence to simulate the nonlinear evolution of the relationship between groundwater level changes and land settlement. Spatially, the land settlement response under water level changes is controlled by soil physical properties, and spatial weights are used to represent the spatial differences between input variables and land settlement (different correlations at different locations). The temporal input feature X... t It includes precipitating factors and intrinsic geological controlling factors, where precipitating factors are represented as follows: j represents the factors influencing land subsidence (including deformation-inducing factors such as water level), and G represents the intrinsic geological control factors. (m) The spatial weights corresponding to the input features are represented as follows: That is, the spatial coordinates are (u x v x The regression parameters between the m-th PS point and the j-th input feature at time t and the ground subsidence. Regression parameters between soil physical properties and ground settlement The output feature is the ground subsidence value. The recurrent computation unit of the neural network is LSTM, which learns the temporal evolution relationship between input and output variables over a long period of time through three gating mechanisms: input gate, forget gate, and output gate.

[0078] The second step is model network structure design. Based on the model building method, a control variable layer and a spatial relationship layer are added to the input-output structure of the recurrent neural network. Combined with the input sedimentation inducing factors, the Geo-GWLSTM model network structure is designed as follows: Figure 4 As shown.

[0079] ① Input layer: includes anthropogenic factors that induce land subsidence and intrinsic geological conditions that control land subsidence.

[0080] ② Spatial Relationship Layer: This layer identifies the spatial correlation between input variables and PSI-observed settlement, enhancing the correlation between input variables and settlement at each sample point. For each time step in the time series, the spatial weights at each sample point (PS point) in the input time series are used, and this matrix is ​​integrated into the corresponding PS point to obtain the geographically weighted PS points. For large datasets, to improve model training efficiency, a mini-batch method is used for model training, meaning the input sample size for each training iteration is the set batch size.

[0081] ③ Hidden Layers: To learn from samples more effectively, a multi-layer LSTM model is used, with the number of recurrent layers designed based on the model training and optimization results. In each layer, the state and output values ​​of the unit at the previous time step are input into the memory cell at the next time step. Each memory cell contains multiple neurons, and a fully connected layer is used for dimensionality transformation, converting the output dimension into the dimension used for model optimization.

[0082] ④ Output layer: Employs a multi-output approach, outputting the model's predicted values ​​(i.e., ground settlement values) at each time step. This improves model accuracy.

[0083] ⑤ Optimization layer: The model output value and PSI observation value y at each time step t Joint calculation of loss L t The gradient descent method is used to solve for the model parameters that minimize the loss, and the model is optimized.

[0084] The third part is the spatiotemporal prediction of land subsidence. Water level information for n time periods is added to the model training time series t as the prediction period, and the subsidence at these n times is predicted. This mainly involves two steps: data preprocessing and subsidence prediction.

[0085] The first step is data preprocessing. Input features and spatial weights are set for n time points in the prediction period. For water level data at a single PS point, preprocessing is required. This involves extracting and arranging the data sequentially with a fixed time series length t-1, starting from time 2, to obtain the model input data for that single point, such as... Figure 5 As shown. Then, batch process all PS points in the same way to obtain the model input data for the region.

[0086] The second step is settlement prediction. First, the trained model is invoked to predict the settlement change over time series n. Then, starting from the prediction period, the settlement change value output by the model at the last moment of each time series for each PS point is extracted; this value is the predicted settlement value for that PS point. Finally, the annual settlement during the prediction period and the cumulative settlement over the entire period are calculated using the predicted settlement change values.

[0087] The third step is regional settlement mapping. Spatial interpolation is performed on the predicted settlement values ​​at PS points to create a regional settlement distribution map and analyze the settlement risk.

[0088] This embodiment addresses the challenges of traditional hydrogeological models, which suffer from numerous and difficult-to-obtain parameters, and the problem that data-driven models fail to simultaneously consider the spatiotemporal relationship and physical mechanisms of influencing factors and land subsidence. It employs the spatiotemporal prediction method for land subsidence using a recurrent neural network that considers physical mechanisms, as described in this invention. This method utilizes readily available and abundant time-series SAR remote sensing imagery as data support and leverages deep learning methods for land subsidence simulation and prediction. This invention comprehensively considers the combined impact of anthropogenic factors and geological control factors on land subsidence, as well as the spatiotemporal variability of this impact, thereby improving the reliability of data-driven land subsidence models and providing a reliable method for effective land subsidence simulation and prediction.

[0089] The following detailed explanation of the spatiotemporal prediction method for land subsidence of this invention is illustrated with a specific example, comprising three parts: spatiotemporal dataset acquisition, land subsidence simulation model training, and land subsidence prediction. This invention integrates PSI technology, spatial analysis methods, and recurrent neural network models to perform spatiotemporal simulation and prediction of regional-scale land subsidence under strongly heterogeneous geological structures. In this embodiment, the input inducing factor is the water level change of the confined aquifer, and the geological controlling factor is the inelastic release coefficient of the aquifer system.

[0090] The first step is to acquire time-series ground subsidence information based on multiple time-series SAR remote sensing images using PSI technology.

[0091] The second step involves combining long-term water level observation station data with universal kriging interpolation to obtain water level information in the same time series as the land subsidence information extracted by PSI. The spatial error between the interpolation results from the June (high-water season) water level observation points and the contour data is used to correct the interpolation results from April to September of the same year. Similarly, the spatial error between the interpolation results from the December (low-water season) water level observation points and the contour data is used to correct the interpolation results from October to December of the same year and January to March of the following year. This yields the time-series water level information for each PS point.

[0092] The third step involves combining the settlement and water level information obtained in the first and second steps, selecting the water level in June of the first year of the time-series settlement as the pre-consolidation head, and calculating the inelastic release coefficient of the aquifer system at each PS point by the water level drop relative to the pre-consolidation head in June of the last year during the water level decline period and the corresponding settlement.

[0093] The fourth step involves performing differential calculations on the temporal settlement and water level information obtained in the first and second steps to obtain the water level change at each PS point at each time step. (i.e., the water level change at the m-th PS point at time k) and settlement (That is, the settlement of the m-th PS point at time k). Then the input sequence for water level information is:

[0094] The fifth step is to use the GWR model to calculate the position of each PS point at each time step. and corresponding settlement change The regression parameters between them are expressed as follows: and

[0095] Step 6: Input model data, set hyperparameters, and train and optimize the model. The model input variables are... The corresponding output sedimentation sequence is

[0096] Step 7: Select the optimal hyperparameters based on the principles of minimizing the loss function value and maximizing accuracy to obtain the optimal trained model. The model training and optimization methods and steps are as follows: Figure 6 As shown.

[0097] ① Dataset partitioning: Using the thinning tool in ArcGIS software, all PS points are randomly divided into training and test sets, and the spatial distribution of the training and test sets is consistent after partitioning.

[0098] ② Model hyperparameter initialization: Set the model hyperparameters, including the number of neurons, the number of hidden layers, the batch size, the learning rate, the loss function, and the optimization algorithm.

[0099] ③ Model training: Input the feature matrix and spatial weight matrix, input batch size samples each time, and after completing one epoch (i.e. all samples have completed one training), use gradient descent to calculate the loss between the predicted settlement value and the PSI observed settlement value at each time step. The model is iteratively trained (to complete epochs of training) to obtain the minimum loss value.

[0100] ④ Model overfitting assessment: The model converges when the loss value stops decreasing. This value is used to assess model overfitting and select the optimal number of iterations.

[0101] ⑤ Optimize model training: Based on the trained model, evaluate the model's generalization ability and prediction effect on the test set data, analyze the model accuracy, and select the optimal hyperparameter combination by combining the loss function value, loss function curve, and test set accuracy to obtain the optimal model.

[0102] Step 8: For the PS point on the test set, the inelastic release coefficient of the aquifer system at the n time points of the prediction period is fixed (i.e., the same value as in the training period), and the spatial weight is the average of the spatial weights during the training period (i.e., the average of the spatial weights calculated at all time points in the time series during the training period); the water level sequence is differentially calculated to obtain the water level change at each time point in the prediction period, expressed as... Preprocessing is performed according to a fixed time series length t-1. The model trained in step 7 is called to predict the settlement at point PS in the test set at n time points, calculate the settlement during the prediction period and the cumulative settlement from time 1, and perform spatial interpolation to obtain the regional ground settlement distribution map.

[0103] like Figure 7 As shown, this invention provides a ground subsidence risk prediction device based on a recurrent neural network. Figure 1-6 The explanations and descriptions of the method embodiments shown apply to this embodiment, such as... Figure 7 As shown, it includes:

[0104] Data acquisition module 701 is used to acquire spatiotemporal distribution information of ground subsidence and spatiotemporal distribution information of water level;

[0105] Data processing module 702 is used to obtain geological control parameters and regression parameters for characterizing the spatial relationship between the input variables of the recurrent neural network and ground subsidence based on the spatiotemporal distribution information of ground subsidence and the spatiotemporal distribution information of water level.

[0106] Model building module 703 is used to build a recurrent neural network spatiotemporal model for predicting land subsidence risk;

[0107] The settlement hazard prediction module 704 is used to input the geological control parameters and settlement inducing factors as input variables into the input layer of the recurrent neural network spatiotemporal model, and input the regression parameters into the spatial relationship layer of the recurrent neural network spatiotemporal model, and use the recurrent neural network spatiotemporal model to predict the ground settlement value.

[0108] Furthermore, the model building module is specifically used to construct the input layer, spatial relationship layer, hidden layer, output layer, and optimization layer of a recurrent neural network spatiotemporal model for predicting land subsidence risk, wherein;

[0109] The input layer is used to receive the geological control parameters and subsidence inducing factors;

[0110] The spatial relationship layer is used to characterize the spatial correlation between the input variables of the input layer and the PSI-observed settlement;

[0111] The hidden layer includes a multi-layer LSTM model. In each layer of the multi-layer LSTM model, the unit state value and output value of the previous time step are input into the memory cell of the next time step. Each memory cell contains multiple neurons. The dimension transformation is performed through a fully connected layer to convert the output dimension into the dimension used for optimizing the spatiotemporal model of the recurrent neural network.

[0112] The output layer is used to output the ground settlement value at each moment;

[0113] The optimization layer is used to jointly calculate the loss based on the ground subsidence value and PSI observation value at each time moment, and to use the gradient descent method to solve for the recurrent neural network spatiotemporal model parameters when the loss is minimized, thereby optimizing the recurrent neural network spatiotemporal model.

[0114] The geological control parameters include geological, hydrogeological, and soil mechanics parameters such as the inelastic water release coefficient, elastic water release coefficient, compressibility coefficient, and compressible layer thickness of the aquifer system.

[0115] The subsidence-inducing factors include the water level parameters of the confined aquifer.

[0116] The present invention also provides a computer-readable storage medium storing a program, the program including instructions for performing the above-described methods. Furthermore, the present invention provides a land subsidence hazard prediction system based on a recurrent neural network, comprising a readable medium storing a computer program, the program including instructions for performing the above-described methods.

[0117] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting ground subsidence risk based on recurrent neural networks, characterized in that, include: Acquire information on the spatiotemporal distribution of ground subsidence and water level; Based on the spatiotemporal distribution information of ground subsidence and the spatiotemporal distribution information of water level, geological control parameters and regression parameters used to characterize the spatial relationship between the input variables of the recurrent neural network and ground subsidence are obtained. Construct a spatiotemporal model of a recurrent neural network for predicting land subsidence risk; The geological control parameters and subsidence inducing factors are input as input variables into the input layer of the recurrent neural network spatiotemporal model, and the regression parameters are input into the spatial relationship layer of the recurrent neural network spatiotemporal model. The ground subsidence value is then predicted using the recurrent neural network spatiotemporal model. The steps for constructing a recurrent neural network spatiotemporal model for predicting land subsidence risk include: A recurrent neural network spatiotemporal model for predicting land subsidence risk is constructed, consisting of an input layer, a spatial relationship layer, a hidden layer, an output layer, and an optimization layer. The input layer is used to receive the geological control parameters and subsidence inducing factors; The spatial relationship layer is used to characterize the spatial correlation between the input variables of the input layer and the PSI observed settlement; for each moment in the time series, the spatial weight of each PS point in the time series is input to obtain the geographically weighted PS point; The hidden layer includes a multi-layer LSTM model. In each layer of the multi-layer LSTM model, the unit state value and output value of the previous time step are input into the memory cell of the next time step. Each memory cell contains multiple neurons. The dimension transformation is performed through a fully connected layer to convert the output dimension into the dimension used for optimizing the spatiotemporal model of the recurrent neural network. The output layer is used to output the ground settlement value at each moment; The optimization layer is used to jointly calculate the loss based on the ground subsidence value and PSI observation value at each time moment, and to use the gradient descent method to solve for the recurrent neural network spatiotemporal model parameters when the loss is minimized, thereby optimizing the recurrent neural network spatiotemporal model. The method of predicting ground subsidence values ​​using a recurrent neural network spatiotemporal model includes: Set the input features and spatial weights at n time points during the prediction period. Starting from the second time point, extract and arrange them sequentially with a fixed time series length t-1 to obtain the model input data at a single point. Batch process all PS points in the same way to obtain the model input data over the region. Call the trained model to predict the settlement change over time series n; extract the settlement change value output by the model at the last moment of each time series for each PS point from the beginning of the prediction period to obtain the predicted settlement value for that PS point. The annual settlement during the forecast period and the cumulative settlement over the entire period are calculated based on the predicted settlement change values.

2. The method for predicting ground subsidence risk based on recurrent neural networks according to claim 1, characterized in that, The steps for obtaining the spatiotemporal distribution information of ground subsidence and the spatiotemporal distribution information of water level include: Extracting spatiotemporal distribution information of ground subsidence based on PSI technology; Information on the spatiotemporal distribution of water level is obtained using universal kriging interpolation based on error correction.

3. The method for predicting ground subsidence risk based on recurrent neural networks according to claim 1, characterized in that, The geological control parameters include geological, hydrogeological, and soil mechanics parameters such as the inelastic water release coefficient, elastic water release coefficient, compressibility coefficient, and compressible layer thickness of the aquifer system. The steps of obtaining geological control parameters and regression parameters for characterizing the spatial relationship between the input variables of the recurrent neural network and land subsidence based on the spatiotemporal distribution information of land subsidence and water level include: The inelastic water release coefficient of the aquifer system is obtained based on the spatiotemporal distribution information of ground subsidence and the spatiotemporal distribution information of water level. A geographic weighted regression model was used to analyze the spatial relationship between the input variables of the recurrent neural network and ground subsidence at each time point in the time series, and the regression parameters at different locations and times were obtained.

4. The method for predicting ground subsidence risk based on recurrent neural networks according to claim 3, characterized in that, Following the step of predicting ground subsidence values ​​using a recurrent neural network spatiotemporal model, the method further includes: Spatial interpolation is performed on the predicted ground settlement values ​​at PS points to create a regional settlement distribution map.

5. The method for predicting ground subsidence risk based on recurrent neural networks according to claim 4, characterized in that, The subsidence-inducing factors include the water level parameters of the confined aquifer.

6. A ground subsidence hazard prediction device based on a recurrent neural network according to any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire spatiotemporal distribution information of ground subsidence and spatiotemporal distribution information of water level; The data processing module is used to obtain geological control parameters and regression parameters that characterize the spatial relationship between the input variables of the recurrent neural network and the ground subsidence based on the spatiotemporal distribution information of ground subsidence and the spatiotemporal distribution information of water level. The model building module is used to build a recurrent neural network spatiotemporal model for predicting land subsidence risk. The settlement hazard prediction module is used to input the geological control parameters and settlement inducing factors as input variables into the input layer of the recurrent neural network spatiotemporal model, and input the regression parameters into the spatial relationship layer of the recurrent neural network spatiotemporal model, and use the recurrent neural network spatiotemporal model to predict the ground settlement value.

7. The ground subsidence hazard prediction device based on recurrent neural network according to claim 6, characterized in that, The model building module is specifically used to construct the input layer, spatial relationship layer, hidden layer, output layer, and optimization layer of a recurrent neural network spatiotemporal model for predicting land subsidence risk. The input layer is used to receive the geological control parameters and subsidence inducing factors; The spatial relationship layer is used to characterize the spatial correlation between the input variables of the input layer and the PSI-observed settlement; The hidden layer includes a multi-layer LSTM model. In each layer of the multi-layer LSTM model, the unit state value and output value of the previous time step are input into the memory cell of the next time step. Each memory cell contains multiple neurons. The dimension transformation is performed through a fully connected layer to convert the output dimension into the dimension used for optimizing the spatiotemporal model of the recurrent neural network. The output layer is used to output the ground settlement value at each moment; The optimization layer is used to jointly calculate the loss based on the ground subsidence value and PSI observation value at each time moment, and to use the gradient descent method to solve for the recurrent neural network spatiotemporal model parameters when the loss is minimized, thereby optimizing the recurrent neural network spatiotemporal model. The geological control parameters include geological, hydrogeological, and soil mechanics parameters such as the inelastic water release coefficient, elastic water release coefficient, compressibility coefficient, and compressible layer thickness of the aquifer system. The subsidence-inducing factors include the water level parameters of the confined aquifer.

8. A computer-readable storage medium storing a program, characterized in that, The program includes instructions for performing the method as described in any one of claims 1-5.

9. A ground subsidence hazard prediction system based on recurrent neural networks, comprising a readable medium storing a computer program, characterized in that, The program includes instructions for performing the method as described in any one of claims 1-5.

Images