CN-121834626-B - Tunnel or mine water burst space-time prediction method coupled with hydrodynamic numerical model

CN121834626BCN 121834626 BCN121834626 BCN 121834626BCN-121834626-B

Abstract

The invention discloses a tunnel or mine water burst space-time prediction method and a system coupled with a hydrodynamic numerical model, wherein the method comprises the following steps: and outputting multi-source data based on the identified and verified groundwater numerical model, complementing the loss of the measured data, quantifying the difference between the permeability characteristics of the fault and the normal stratum, coupling the fault and the normal stratum to a data system, and merging the fault and the normal stratum into tunnel or mine excavation space data. And constructing an LSTM-isolated forest-K nearest neighbor regression coupling model, and configuring a multifunctional module to realize the common training of the multi-scene data. Dividing the preprocessed multi-element time sequence data into a training set and a testing set, extracting hidden features through a coupling model, fusing abnormal detection results, synchronously training a residual error correction model, and adaptively optimizing super parameters and weights according to multi-engineering prediction error feedback. Based on the trained coupling model, a window rolling strategy is adopted to conduct synchronous water inrush space-time prediction, and prediction data meeting engineering accuracy is output by combining residual correction. And a reliable technical support is provided for engineering construction safety prevention and control.

Inventors

CHENG JIANMEI
HU MINGHANG
LUO YIMING
SHI TINGTING
WANG NINGTAO
WANG BINGGUO
ZHENG KUN

Assignees

中国地质大学（武汉）
中国地质调查局武汉地质调查中心(中南地质科技创新中心)

Dates

Publication Date: 20260512
Application Date: 20260311

Claims (9)

1. A tunnel or mine water burst space-time prediction method coupled with a hydrodynamic numerical model is characterized by comprising the following steps: S1, outputting tunnel or mine water burst related prediction data, groundwater environment data and geologic body permeability characteristic data based on an identified and verified groundwater dynamic numerical model, complementing the missing or insufficient part of actual monitoring data by adopting the numerical model prediction data, quantifying the permeability characteristic difference between faults and normal stratum by a difference analysis method, and coupling the fault and normal stratum permeability characteristic difference to a data system; s2, constructing a coupling model by adopting a method comprising long-term and short-term time memory neural network LSTM, isolated forest iForest and K nearest neighbor regression, and simultaneously configuring feature extraction, rolling prediction, data expansion, label statistics, outlier detection and residual error correction modules, wherein each module completes data interaction and training through a preset interface; S3, dividing the preprocessed multi-element time sequence training data into a training set and a testing set according to time sequence, inputting the training set and the testing set into a constructed coupling machine learning model, training an LSTM model and extracting hidden time sequence characteristics, respectively training three abnormal detection models, integrating learning results through a weighted fusion strategy, and synchronously training a residual correction model; s4, synchronously and dynamically predicting water inflow time and space of a plurality of tunnels or mines by utilizing the optimized coupled machine learning model through a window rolling strategy of the rolling prediction module, outputting water inflow prediction results of different time steps and different tunnel or mine sections; the self-adaptive adjustment of the super parameters of each module in the step S3 comprises the following steps: Firstly, constructing a multi-objective error evaluation function, and calculating root mean square error based on prediction results and measured data of a plurality of tunnels or mines Average absolute percentage error Constructing a comprehensive error index Wherein Represent the first A strip tunnel or a mine shaft, And Is the inherent weight of the error index, is naturally determined by the data distribution characteristics and engineering precision requirements, , , , Is the first Predicted number of time steps for a strip tunnel or mine, Is the first Strip tunnel or mine The measured water inflow of the time step, Predicting water inflow for the model of the corresponding time step; Defining a model hyper-parameter set Wherein Super-parametric vectors for LSTM model, comprising hidden layer neuron number Number of iterations Rate of learning A descent rate, a regularization parameter; Super-parametric vectors for isolated forest models, including the number of decision trees Sample amount Coefficient of variation; the super-parameter vector of the K neighbor regression model comprises a neighbor sample number K and a Euclidean distance measurement mode; Defining residual correction weight vectors , For the total number of tunnels or mines, Is the first Residual error correction weight corresponding to strip tunnel or mine meets the following requirements ; With the average value of the integrated errors of all tunnels or mines minimized as an objective function, i.e And the grid search method is used for performing optimization on the super parameters of the model, and through exhaustive traversal of all parameter combinations in a preset super parameter space, verifying each group of parameter combinations one by one and calculating a corresponding comprehensive error mean value, and finally selecting an optimal super parameter combination which enables the comprehensive error mean value to reach a minimum value, the global optimization of the super parameters of the model is completed.
2. The tunnel or mine water burst space-time prediction method coupled with the hydrodynamic force numerical model according to claim 1, wherein the difference analysis method in S1 is specifically as follows: Firstly, the core difference of the permeability characteristics of a fault zone and a normal stratum is clarified, firstly, the average difference of the overall permeability of the stratum is the average difference, secondly, the non-uniformity difference of the distribution of the internal permeability is the difference, and in order to comprehensively characterize the two differences, the sequential sequence of the permeability coefficient of the fault zone is extracted firstly based on continuous time step data output by a groundwater numerical model , wherein, For numerical simulation of time steps, each data point corresponds to a global permeability coefficient value of a fault zone of a specific time step, and is derived from a physical simulation result of a model on the fault zone lithology and the fracture development state, and a synchronous normal stratum permeability coefficient time sequence The data of the fault zone and the fault zone are based on the same numerical simulation system, so that the consistency of hydrogeologic environment is ensured, and the simulated value of the global permeability coefficient of the normal stratum is corresponding; Secondly, calculating the cumulative average value of the permeability coefficients of the fault zone and the normal stratum, and calculating the cumulative average value of the permeability coefficients of the fault zone Normal stratum permeability coefficient cumulative average As the average water permeability reference value of normal surrounding rock around fault, and Forming a differential contrast of mean dimensions, i.e The difference directly reflects the strength difference of the overall water permeability of the two types of stratum; thirdly, considering the influence of the heterogeneity of the permeability in the stratum on the water permeability, the degree of the dispersion of the permeability coefficient of the fault zone and the normal stratum needs to be quantified, and the variance of the permeability coefficient of the fault zone is calculated Variance of normal stratum permeability coefficient Representing the permeability discrete characteristic formed by the weathering degree and rock mass structural difference in the normal surrounding rock, and constructing a comprehensive dispersion index in order to unify the quantization dimensionality of the mean value difference and the variance difference The index integrates the respective heterogeneity characteristics of two types of stratum, and provides a standardized comparison standard for mean value difference; finally, deriving a differential quantization model expression based on the ratio relation of the mean value difference and the comprehensive dispersion: Wherein, the method comprises the steps of, And (3) performing coupling calculation on the average value difference and the comprehensive dispersion for the comprehensive difference value of the fault zone and the normal stratum permeability coefficient, and completing objective and dimensionless quantification on the two types of stratum permeability characteristic differences.
3. The method for predicting tunnel or mine water burst time and space coupled with hydrodynamic force numerical model as set forth in claim 1, wherein S1 further comprises performing abnormal signal identification on the measured data, and obtaining a construction section comprehensive permeability coefficient by weighting calculation in combination with surrounding rock classification, weathering degree and lithology structure And preprocessing all the coupled data by adopting a normalization method adapted to a corresponding module to form multi-element time sequence training data.
4. A method of spatio-temporal prediction of water gushing in tunnels or mines coupled to a hydrodynamic numerical model according to claim 3, characterized in that said complex permeability coefficients are The calculation method of (1) is as follows: First, the formation background permeability coefficient is defined The parameters are taken from an identified and verified groundwater numerical model, the corresponding construction section corresponds to a permeability reference value of a rock stratum under the condition that the rock stratum is not disturbed by construction and the natural geological state is maintained, and the data are cross-checked to be consistent with geological investigation reports and rock sample test data of a research area; secondly, selecting a core geological influence factor of the construction section, and setting surrounding rock classification quantization factors as The rock mass weathering degree quantization factor is The lithology structure breaking degree quantization factor is Each factor realizes quantization through interval mapping, corresponding dimensionless quantization values are respectively endowed for different surrounding rock types, weathering grades and lithology structure types, and the mapping rule that the more obvious the enhancement effect of geological features on permeability is met, the larger the corresponding factor value is; Introducing dynamic correction terms of the geological influence factors, wherein the expression is in the form of the corresponding exponential power product of each geological influence factor, and obtaining the generalized calculation expression of the comprehensive permeability coefficient as follows Wherein In order to participate in the number of geological impact factors calculated, Not less than 3; Natural weight index for each factor; the comprehensive permeability coefficient of the construction section is used for representing the real permeability of the construction section under the actual geological working condition; The physical simulation result is derived from the underground water numerical model and is the stratum background permeability coefficient; is the first And the quantized value of each geological influence factor is obtained based on objective geological feature quantization, so that the normalization of the geological feature characterization is ensured.
5. The tunnel or mine water burst space-time prediction method coupled with the hydrodynamic force numerical model according to claim 1, wherein the three anomaly detection models in the step S3 are respectively a long-short-period time memory neural network model, an isolated forest model and a K nearest neighbor regression model, and the three anomaly detection models are respectively responsible for different anomaly identification tasks based on design logic of functional complementation and cooperative coupling: The long-term and short-term time memory neural network model relies on the long-term dependency capturing capability of the neural network model on time sequence data, starts from a time sequence evolution rule of tunnel or mine water inflow data, identifies trend abnormality caused by dynamic change of hydrogeological conditions and continuous disturbance of construction process, and locates continuous abnormal change of tunnel or mine water inflow generated along with time by implicit association in mining data time sequence characteristics; The isolated forest model focuses on global discrete features in water inflow data of tunnels or mines based on a random division principle, specifically identifies global isolated abnormal data points caused by factors including hydrologic environmental condition changes of excavation surfaces and abrupt changes of geological structures, and captures abnormal information features which deviate remarkably from peripheral data distribution rules by randomly cutting and dividing data space and performing discrete evaluation on the data; and the K neighbor regression model identifies the local abnormal water surge event caused by factors including tunnel or mine construction plan change and equipment transient fault adjustment from the statistical rule level of local data, and screens out abnormal information in the local data by delineating the local data.
6. The tunnel or mine water burst space-time prediction method coupled with the hydrodynamic force numerical model according to claim 1, wherein the weighting fusion strategy in the step S3 is specifically a linear weighting strategy: Defining the weight coefficients of the detection results of the three types of models as respectively The weight coefficient is comprehensively determined based on the abnormal detection accuracy and recall rate of various models on the verification set, and higher weight is allocated to the model with better detection performance.
7. The method for predicting the water burst time and space of a tunnel or a mine coupled with a hydrodynamic numerical model according to claim 1, wherein the correction process of the residual correction weight in the step S3 is as follows: Dynamic allocation of residual correction weights is completed based on the prediction precision of each tunnel or mine, the weight allocation is realized by adopting a conditional mapping function, and the prediction precision quality scores of the single tunnels or mines are calculated first The calculation formula is as follows: In the formula, For determining the coefficient, the value of the coefficient is derived from fitting analysis of the predicted value of a single tunnel or mine and measured water inflow data; for the correlation coefficient of the predicted value and the measured value, Quality score for MAE improvement rate The higher the value, the better the prediction precision after residual error correction of the corresponding tunnel or mine, and the quality score calculated for the short tunnel with insufficient data sample size Performing deduction correction, and grading the quality after correction According to the corrected quality score Five-level prediction precision evaluation standards are defined, and corresponding residual error correction weights are directly matched through a conditional mapping function The mapping relation is as follows: The corresponding evaluation grades are excellent, good, general, poor and very poor in sequence.
8. A tunnel or mine water burst space-time prediction system coupled with a hydrodynamic numerical model, comprising: The multi-source data coupling preprocessing unit is used for outputting tunnel or mine water burst related prediction data, groundwater environment data and geologic body permeability characteristic data based on the recognized and verified groundwater dynamic numerical model, supplementing the missing or insufficient part of actual monitoring data by adopting the numerical model prediction data, quantifying the permeability characteristic difference between faults and normal stratum by a difference analysis method and coupling the fault and normal stratum to a data system; The coupling model construction configuration unit is used for constructing a coupling model by adopting a method comprising long-short-term time memory neural network LSTM, isolated forest iForest and K neighbor regression, and simultaneously configuring feature extraction, rolling prediction, data expansion, label statistics, outlier detection and residual error correction modules, wherein each module completes data interaction and training through a preset interface; The model training dynamic optimization unit is used for dividing the preprocessed multi-element time sequence training data into a training set and a testing set according to time sequence, inputting the training set and the testing set into a constructed coupling machine learning model, firstly training an LSTM model, extracting hidden time sequence characteristics, respectively training three abnormal detection models, integrating learning results through a weighted fusion strategy, and synchronously training a residual correction model; The tunnel or mine water inflow prediction result correction unit is used for carrying out synchronous water inflow space-time dynamic prediction on a plurality of tunnels or mines by utilizing the optimized coupled machine learning model through a window rolling strategy of the rolling prediction module and outputting water inflow prediction results of different time steps and different tunnel or mine sections; abnormal data in the prediction process is identified by combining with the outlier detection module, the prediction result is corrected in real time by the residual error correction module, and finally the tunnel or mine water inflow space-time dynamic prediction data meeting the engineering precision requirement is output The process of self-adaptively adjusting the super parameters of each module in the model training dynamic optimization unit is as follows: Firstly, constructing a multi-objective error evaluation function, and calculating root mean square error based on prediction results and measured data of a plurality of tunnels or mines Average absolute percentage error Constructing a comprehensive error index Wherein Represent the first A strip tunnel or a mine shaft, And Is the inherent weight of the error index, is naturally determined by the data distribution characteristics and engineering precision requirements, , , , Is the first Predicted number of time steps for a strip tunnel or mine, Is the first Strip tunnel or mine The measured water inflow of the time step, Predicting water inflow for the model of the corresponding time step; Defining a model hyper-parameter set Wherein Super-parametric vectors for LSTM model, comprising hidden layer neuron number Number of iterations Rate of learning A descent rate, a regularization parameter; Super-parametric vectors for isolated forest models, including the number of decision trees Sample amount Coefficient of variation; the super-parameter vector of the K neighbor regression model comprises a neighbor sample number K and a Euclidean distance measurement mode; Defining residual correction weight vectors , For the total number of tunnels or mines, Is the first Residual error correction weight corresponding to strip tunnel or mine meets the following requirements ; With the average value of the integrated errors of all tunnels or mines minimized as an objective function, i.e And the grid search method is used for performing optimization on the super parameters of the model, and through exhaustive traversal of all parameter combinations in a preset super parameter space, verifying each group of parameter combinations one by one and calculating a corresponding comprehensive error mean value, and finally selecting an optimal super parameter combination which enables the comprehensive error mean value to reach a minimum value, the global optimization of the super parameters of the model is completed.
9. A computer readable storage medium having stored thereon a computer program, the computer program being executable by a processor to perform a method of spatiotemporal prediction of water gushing in a tunnel or mine coupled with a hydrodynamic numerical model according to any of claims 1 to 7.

Description

Tunnel or mine water burst space-time prediction method coupled with hydrodynamic numerical model Technical Field The invention belongs to the technical field of prediction of hydrogeological disasters of tunnels and mine engineering, and particularly relates to a method and a system for predicting water burst time of tunnels or mines by coupling a hydrodynamic numerical model. Background In the tunnel and mine engineering construction and operation processes, water gushing disasters are one of core risk factors which threaten construction safety and influence engineering progress. As engineering construction advances to underground deep complex geological areas, hydrogeological conditions become more and more complex, and the characteristics of fault fracture zone development, remarkable rock stratum permeability difference and the like cause remarkable improvement of burst and uncertainty of tunnel or mine water gushing disasters, so that higher requirements are put on the accuracy and timeliness of water gushing prediction under the complex geological environment. The current tunnel or mine water burst prediction method is mainly divided into two types, namely a traditional geological investigation method and a conventional hydrodynamic numerical simulation method. Traditional geological survey methods rely on-site drilling sampling and geological mapping, are limited by survey ranges and precision, are difficult to comprehensively reflect dynamic changes of global geological hydrologic environmental conditions of engineering areas, are high in subjectivity of prediction results, and are insufficient in prejudgment capability on hidden water inrush risks. Although the conventional hydrodynamic force numerical simulation method can construct a model based on geological parameters, a single model structure is adopted, water burst or mine time sequence monitoring data and geological characteristic difference information cannot be effectively coupled, and the adaptability to dynamic change of rock stratum permeability in the construction process is poor, so that a prediction result and an engineering actual working condition deviate. Meanwhile, the existing water burst or mine prediction technology generally has the problems of insufficient data integration, weak abnormality recognition capability, missing space-time dynamic prediction and the like. In actual engineering, the monitoring data is easy to be lost or abnormal due to equipment faults and environmental interference, the conventional method is difficult to complement and discriminate effectively, and aiming at synchronous prediction requirements of a plurality of tunnels or mines, the prior art lacks an efficient model adaptation and dynamic optimization mechanism, and is difficult to meet the safety prevention and control requirements of large-scale engineering. Therefore, the tunnel and mine water burst prediction method capable of integrating multi-source geological hydrologic data, accurately quantifying geological feature differences and realizing space-time dynamic prediction is developed, and has important practical significance for improving the water burst disaster prevention and control capability of engineering, guaranteeing construction safety and reducing engineering loss. Disclosure of Invention The method aims to solve the problems of insufficient data integration, inaccurate geological feature quantification and weak space-time dynamic prediction capability in the water burst prediction of tunnels and mine engineering, and realizes the effective fusion of multi-source data, dynamic model optimization and water burst space-time accurate prediction by coupling a hydrodynamic numerical model and a machine learning algorithm, thereby providing reliable technical support for engineering construction safety prevention and control. In view of the above-mentioned drawbacks or improvements of the prior art, as a first aspect of the present invention, the present invention provides a method for spatiotemporal prediction of water gushing in tunnels or mines coupled with a hydrodynamic numerical model, comprising: S1, outputting tunnel or mine water burst related prediction data, groundwater environment data and geologic body permeability characteristic data based on an identified and verified groundwater dynamic numerical model, complementing the missing or insufficient part of actual monitoring data by adopting the numerical model prediction data, quantifying the permeability characteristic difference between faults and normal stratum by a difference analysis method, and coupling the fault and normal stratum permeability characteristic difference to a data system; s2, constructing a coupling model by adopting a method comprising long-term and short-term time memory neural network LSTM, isolated forest iForest and K nearest neighbor regression, and simultaneously configuring feature extraction, rolling prediction, data expansion, label statist