CN-122022041-A - Soil heavy metal pollution space-time dynamic prediction method based on LSTM-GWR coupling

CN122022041ACN 122022041 ACN122022041 ACN 122022041ACN-122022041-A

Abstract

The invention relates to the technical field of environmental science, in particular to a soil heavy metal pollution space-time dynamic prediction method based on LSTM-GWR coupling. The method comprises the steps of carrying out Bayesian correction on multi-source data by using a monitoring terminal, inverting the acid exchangeable thallium content through support vector regression, uploading the acid exchangeable thallium content to a cloud end, constructing a core feature matrix by the cloud end based on game theory combination weighting, extracting a time reference predicted value by using an LSTM, constructing a GWR model output space residual error correction value aiming at a time sequence residual error, finally superposing the two predicted results to obtain a predicted result, and generating a dynamic threshold value and a management and control instruction by combining a natural breakpoint method and pushing the dynamic threshold value and the management and control instruction to the terminal. According to the invention, through a terminal cloud cooperation and space-time residual error coupling mechanism, the problem of data uncertainty and space-time characteristic fracture is effectively solved, and accurate dynamic early warning and intelligent management and control of high-toxicity thallium form of soil are realized.

Inventors

ZHANG HUI
HE HANG
ZHOU FURONG
CHEN JUAN
WEN PAN

Assignees

四川省自然资源实验测试研究中心(四川省核应急技术支持中心)

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (8)

1. The soil heavy metal pollution space-time dynamic prediction method based on LSTM-GWR coupling is characterized by comprising the following steps of: the monitoring terminal acquires multi-source heterogeneous data of the area to be predicted, corrects the monitoring data by using a Bayesian probability model, inputs the corrected monitoring data into a pre-trained support vector regression model, inverts to obtain acid exchangeable thallium content, and uploads the acid exchangeable thallium content to the cloud; The cloud performs objective weighting and subjective weighting on the environmental impact factors, and introduces a game theory combined weighting algorithm to calculate optimal coefficients to generate a global comprehensive weight vector; The cloud end utilizes an LSTM network to extract the time evolution characteristic of a core characteristic matrix and output a time reference predicted value, calculates a time sequence residual error between the predicted value and a true value in the training stage of the LSTM network, constructs a GWR model by taking the time sequence residual error as a dependent variable and taking a space covariate as an independent variable, calculates a local regression coefficient of each grid point of a region to be predicted by utilizing the GWR model, and outputs a space residual error correction value; And calculating a dynamic early warning threshold value and a hierarchical control instruction based on a natural breakpoint method, and pushing the dynamic early warning threshold value and the hierarchical control instruction to a monitoring terminal.
2. The LSTM-GWR coupling-based soil heavy metal pollution space-time dynamic prediction method according to claim 1 is characterized in that the multi-source heterogeneous data comprise meteorological data, soil physicochemical property data, monitoring data and pollution source data, the specific steps of acquiring the multi-source heterogeneous data of a region to be predicted comprise the steps of acquiring real-time soil physicochemical property data comprising soil pH value, organic matter content and oxidation-reduction potential through a sensor array, connecting rainfall and wind speed data and a pollution source emission list through a communication interface, and taking a laboratory accurate detection value as a calibration standard, wherein the monitoring data comprise the steps of acquiring the rapid detection value of the soil heavy metal content and the laboratory accurate detection value through an integrated portable X-ray fluorescence spectrometer, performing space-time gridding resampling processing, utilizing a bilinear interpolation algorithm to map data of different spatial resolutions into a preset standard grid, and converging data of different time frequencies into daily scale data to form a space-time aligned basic data set.
3. The LSTM-GWR coupling-based soil heavy metal pollution space-time dynamic prediction method of claim 1 is characterized by comprising the specific steps of constructing a Bayesian inference network, taking statistical distribution characteristics of historical monitoring data of a region to be predicted as prior probability distribution, obtaining the monitoring data, constructing likelihood functions of observation data, sampling by a Markov chain Monte Carlo method based on the prior probability distribution and the likelihood functions, calculating posterior probability distribution of the monitoring data, extracting mathematical expected values of the posterior probability distribution as corrected monitoring data true values, calculating confidence interval width of the posterior probability, directly outputting correction data when the confidence interval width is smaller than a preset precision threshold, supplementing the current moment data by a linear interpolation method when the confidence interval width is not smaller than the preset precision threshold, and uploading a basic data set, the corrected monitoring data, collected environment factor data and acid exchangeable thallium content obtained by inversion to a cloud as a fusion data set.
4. The LSTM-GWR coupling-based soil heavy metal pollution space-time dynamic prediction method of claim 1 is characterized in that a pre-training process of a support vector regression model comprises the steps of constructing a historical sample set containing total thallium content of soil, physical and chemical property parameters of soil and corresponding acid exchangeable thallium content by cloud, selecting a radial basis function as a kernel function, mapping low-dimensional input features to a high-dimensional feature space, introducing a relaxation variable and a penalty factor to construct an optimized objective function, solving a dual problem by utilizing a Lagrangian multiplier method to obtain a support vector and a corresponding bias item, and performing cross-validation optimization on the penalty factor and the kernel function parameters by adopting a grid search method until inversion root mean square error of the acid exchangeable thallium content is smaller than a preset standard value.
5. The LSTM-GWR coupling-based soil heavy metal pollution space-time dynamic prediction method according to claim 1, wherein the specific steps of objectively and subjectively weighting the environmental impact factors comprise calculating objective weight vectors of the environmental impact factors by adopting an information entropy algorithm; The specific steps of introducing a game theory combined weighting algorithm to calculate optimal coefficients comprise the steps of constructing a game theory aggregation model, regarding objective weight vectors and subjective weight vectors as game opponents, defining an objective function, aiming at minimizing the sum of the discrete degree between the optimal linear combined weight and each basic weight vector, calculating first-order partial derivatives of the objective function on each combined coefficient, constructing a linear equation set with zero first-order partial derivatives, solving the linear equation set to obtain the optimal linear combined coefficient enabling the objective function to reach the minimum value, carrying out weighted summation on the objective weight vectors and the subjective weight vectors by utilizing the optimal linear combined coefficient, and generating a global comprehensive weight vector after normalization.
6. The method for predicting the time evolution characteristics of the soil heavy metal pollution based on LSTM-GWR coupling is characterized by comprising the specific steps of reconstructing a screened core characteristic matrix into a three-dimensional time sequence tensor according to a time step and inputting the three-dimensional time sequence tensor into the LSTM, reading a forgetting proportion of a cell state at the last moment from zero to one through a forgetting gate, determining the retention degree of historical information, generating a new candidate value vector at the current moment through a hyperbolic tangent activation function through an input gate, updating the cell state through combining with an input threshold value, calculating a hidden layer output value at the current moment based on the updated cell state through the output gate, wherein the hidden layer output value is a time reference predicted value, and representing the time evolution trend without considering spatial heterogeneous interference.
7. The LSTM-GWR coupling-based soil heavy metal pollution space-time dynamic prediction method is characterized by comprising the specific steps of calculating a water flow direction matrix of a region to be predicted by utilizing digital elevation model data, constructing a hydrographic distance matrix based on a hydrographic path, constructing an anisotropic space weight function, determining space weight based on the hydrographic distance, carrying out self-adaptive optimization on the bandwidth of the anisotropic space weight function by correcting a red pool information quantity criterion, determining an optimal bandwidth value, constructing a local weighted regression equation for each grid point to be predicted based on the optimal bandwidth value and the anisotropic space weight function, solving by utilizing a least square method to obtain a constant term estimated value corresponding to each grid point and a local regression coefficient of each space covariate, and calculating the space correction value of each grid point by utilizing a linear regression equation according to the local regression coefficient, the constant term estimated value and the value of each space covariate.
8. The LSTM-GWR coupling-based soil heavy metal pollution space-time dynamic prediction method is characterized by comprising the specific steps of collecting acid exchangeable thallium concentration prediction results of all grid points in a region to be predicted, constructing a prediction value set, carrying out multi-round iterative computation on the prediction value set, searching a group of classification breakpoints in each round of iteration to enable the sum of data variances in the same classified type to be minimum and the difference of data average values among different types to be maximum, carrying out weighted fusion on the classification breakpoint values obtained through computation and the soil environment quality standard limit value to determine a dynamic early warning threshold value suitable for the current space-time background, dividing the region into a safety region, a warning region and a management region according to the dynamic early warning threshold value, and matching corresponding classification management instructions from preset knowledge maps according to environment influence factors with maximum regression coefficient absolute values identified by GWR models in the grid points.

Description

Soil heavy metal pollution space-time dynamic prediction method based on LSTM-GWR coupling Technical Field The invention relates to the technical field of environmental science, in particular to a soil heavy metal pollution space-time dynamic prediction method based on LSTM-GWR coupling. Background With the acceleration of industrialization progress, soil heavy metal pollution has become a global environmental problem, wherein thallium is a highly toxic heavy metal element with high mobility, and the concealment and sudden environmental risks are particularly prominent. In the existing soil heavy metal pollution prediction technology, two main types of main stream methods mainly exist, including a time sequence prediction method and a space analysis method. The prior art attempts to predict temporal and spatial combinations, but the spatial model of their combination generally assumes that the spatial relationship is smooth and uniform, ignoring the spatial non-stationarity of significant soil heavy metal distribution. On the other hand, some prior art solves the problem of spatial heterogeneity through the geographic weighting idea, but basically still belongs to the category of linear regression or statistical interpolation, and when processing long-sequence and complex nonlinear time dynamic characteristics, the generalization capability and the prediction precision of the method are far less than those of a deep learning model. Therefore, the prior art generally has the problem of cleavage or limitation of a space-time coupling mechanism, and is difficult to simultaneously consider the nonlinear characteristics of the time dimension and the non-stationarity of the space dimension, so that the space-time precision of a prediction result is limited. In addition, the risk of heavy metal pollution of soil is influenced by various isomerism factors such as soil properties, meteorological conditions, pollution source emission and the like. When determining the weight of the influencing factors, the prior method usually depends on the objective method driven by data or the subjective method which depends on expert experience only, and lacks a comprehensive weighting mechanism combining subjective and objective, thus easily causing the inundation or misjudgment of key risk factors. Finally, aiming at thallium which is a specific pollutant, the prior early warning technology mostly uses the general heavy metal evaluation thought, and mainly predicts and risk divides based on the total amount of elements. However, thallium biotoxicity and environmental risk are mainly dependent on its chemical morphology, in particular the highly bioavailable acid exchangeable state, and predictions based on total amounts alone often do not truly reflect its actual harm to the ecological environment and human body. In conclusion, how to realize space-time dynamic prediction of pollution risk of heavy metal thallium in soil based on multi-source heterogeneous data is a technical problem to be solved in the field. Therefore, a soil heavy metal pollution space-time dynamic prediction method based on LSTM-GWR coupling is provided. Disclosure of Invention The invention aims to provide a soil heavy metal pollution space-time dynamic prediction method based on LSTM-GWR coupling. The method comprises the steps of carrying out Bayesian correction on multi-source data by using a monitoring terminal, inverting the acid exchangeable thallium content through support vector regression, uploading the acid exchangeable thallium content to a cloud end, constructing a core feature matrix by the cloud end based on game theory combination weighting, extracting a time reference predicted value by using an LSTM, constructing a GWR model output space residual error correction value aiming at a time sequence residual error, finally superposing the two predicted results to obtain a predicted result, and generating a dynamic threshold value and a management and control instruction by combining a natural breakpoint method and pushing the dynamic threshold value and the management and control instruction to the terminal. According to the invention, through a terminal cloud cooperation and space-time residual error coupling mechanism, the problem of data uncertainty and space-time characteristic fracture is effectively solved, and accurate dynamic early warning and intelligent management and control of high-toxicity thallium form of soil are realized. In order to achieve the above purpose, the present invention provides the following technical solutions: A soil heavy metal pollution space-time dynamic prediction method based on LSTM-GWR coupling comprises the following steps: the monitoring terminal acquires multi-source heterogeneous data of the area to be predicted, corrects the monitoring data by using a Bayesian probability model, inputs the corrected data into a pre-trained support vector regression model, inverts to obtain acid exchangeable thallium co