A tunnel or mine gushing water space-time prediction method coupled with a water power numerical model

Abstract

The application discloses a tunnel or mine gushing water space-time prediction method and system coupled with a water power numerical model, and the method comprises the following steps: based on the identified and verified underground water numerical model, outputting multi-source data, complementing the missing measured data, quantifying the difference between fault and normal stratum permeability characteristics and coupling to the data system, and incorporating the tunnel or mine excavation space data. An LSTM-isolation forest-K nearest neighbor regression coupled model is constructed, and a multifunctional module is configured to realize common training of multi-scene data. The pretreated multivariate time series data is divided into a training set and a test set, hidden features are extracted through the coupled model, abnormal detection results are fused, a residual correction model is trained synchronously, and the hyperparameters and weights are adaptively optimized according to the multi-project prediction error feedback. Based on the trained coupled model, a window rolling strategy is adopted to carry out synchronous gushing water space-time prediction, residual correction is combined, and prediction data meeting the engineering precision is output. Reliable technical support is provided for engineering construction safety control.

Description

Technical Field

[0001] This invention belongs to the field of hydrogeological disaster prediction technology for tunnel and mine engineering, and more specifically, relates to a method and system for spatiotemporal prediction of water inrush in tunnels or mines coupled with a hydrodynamic numerical model. Background Technology

[0002] During the construction and operation of tunnels and mines, water inrush is one of the core risk factors threatening construction safety and affecting project progress. As engineering construction advances into deep underground complex geological areas, hydrogeological conditions become increasingly complex. Geological features such as the development of fault fracture zones and significant differences in rock strata permeability lead to a significant increase in the suddenness and uncertainty of water inrush disasters in tunnels or mines, placing higher demands on the accuracy and timeliness of water inrush prediction in such complex geological environments.

[0003] Current methods for predicting water inrush in tunnels or mines are mainly divided into two categories: traditional geological exploration methods and conventional hydrodynamic numerical simulation methods. Traditional geological exploration methods rely on on-site drilling and geological mapping. Due to limitations in the scope and accuracy of the exploration, they are unable to comprehensively reflect the dynamic changes in the geological and hydrological environmental conditions of the entire project area. The prediction results are highly subjective and lack the ability to predict hidden water inrush risks. Although conventional hydrodynamic numerical simulation methods can build models based on geological parameters, they often use a single model structure, which cannot effectively couple the time-series monitoring data of water inrush or mines with the differences in geological characteristics. Furthermore, they have poor adaptability to the dynamic changes in rock permeability during construction, resulting in deviations between the prediction results and the actual engineering conditions.

[0004] Meanwhile, existing water inrush or mine inrush prediction technologies generally suffer from problems such as insufficient data integration, weak anomaly identification capabilities, and lack of spatiotemporal dynamic prediction. In actual engineering projects, monitoring data is prone to loss or anomalies due to equipment failure and environmental interference, and conventional methods are insufficient to effectively supplement and identify these data. Furthermore, for the simultaneous prediction needs of multiple tunnels or mines, existing technologies lack efficient model adaptation and dynamic optimization mechanisms, making it difficult to meet the safety and control requirements of large-scale projects. Therefore, developing a method for predicting tunnel and mine water inrush that can integrate multi-source geological and hydrological data, accurately quantify differences in geological characteristics, and achieve spatiotemporal dynamic prediction is of significant practical importance for improving the ability to prevent and control water inrush disasters in engineering projects, ensuring construction safety, and reducing engineering losses. Summary of the Invention

[0005] This invention aims to solve the problems of insufficient data integration, inaccurate quantification of geological features, and weak spatiotemporal dynamic prediction capabilities in the prediction of water inrush in tunnel and mine engineering. By coupling a hydrodynamic numerical model with a machine learning algorithm, it achieves effective fusion of multi-source data, dynamic optimization of the model, and accurate spatiotemporal prediction of water inrush, providing reliable technical support for the safety control of engineering construction.

[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides a method for spatiotemporal prediction of water inrush in tunnels or mines coupled with a hydrodynamic numerical model, comprising:

[0007] S1. Based on the identified and verified groundwater dynamic numerical model, output the predicted data related to water inrush in tunnels or mines, groundwater environment data and geological body permeability data. Use the numerical model prediction data to supplement the missing or insufficient parts of the actual monitoring data. Use the difference analysis method to quantify the difference in permeability characteristics between faults and normal strata and couple it to the data system.

[0008] S2. A coupled model is constructed using methods including Long Short-Term Memory Neural Network (LSTM), Isolation Forest (iForest), and K-Nearest Neighbor Regression. At the same time, modules for feature extraction, rolling prediction, data expansion, label statistics, outlier detection, and residual correction are configured. Each module completes data interaction and training through a preset interface.

[0009] S3. The preprocessed multivariate time-series training data is divided into training and test sets according to time order and input into the constructed coupled machine learning model. First, the LSTM model is trained and hidden time-series features are extracted. Then, three anomaly detection models are trained separately and the learning results are integrated through a weighted fusion strategy. The residual correction model is trained simultaneously. Based on the feedback of water inflow prediction errors from multiple tunnels or mines, the hyperparameters and residual correction weights of each module are adaptively adjusted to complete the dynamic optimization of the model training process.

[0010] S4. Using the window scrolling strategy of the scrolling prediction module, the optimized coupled machine learning model is used to perform synchronous spatiotemporal dynamic prediction of water inflow for multiple tunnels or mines, and output the water inflow prediction results for different time steps and different tunnel or mine cross sections; combined with the outlier detection module to identify abnormal data in the prediction process, the prediction results are corrected in real time through the residual correction module, and finally the spatiotemporal dynamic prediction data of tunnel or mine water inflow that meets the engineering accuracy requirements is output.

[0011] Furthermore, the difference analysis method in S1 is specifically as follows:

[0012] First, it is clear that the core differences between the permeability characteristics of fault zones and normal strata lie in two aspects: the average difference in the overall permeability of the strata, and the difference in the heterogeneity of the internal permeability distribution. To comprehensively characterize these two types of differences, a time series of permeability coefficients of fault zones is extracted based on continuous time step data output from the groundwater numerical model. ,in, The time steps in the numerical simulation represent the global permeability coefficient value of the fault zone at a specific time step for each data point. This value originates from the physical simulation results of the fault zone's lithology and fracture development, as well as the time series of permeability coefficients of normal strata during the same period. Based on the same numerical simulation system as the fault zone data, the consistency of the hydrogeological environment is ensured, and the simulated value of the permeability coefficient of the normal strata is obtained for the whole area.

[0013] Secondly, calculate the cumulative mean of the permeability coefficients of the fault zone and the normal formation, and the cumulative mean of the permeability coefficient of the fault zone. Cumulative mean of normal formation permeability As a benchmark value for the average permeability of the normal surrounding rock around the fault, and To form a comparison of differences along the mean dimension, i.e. This difference directly reflects the difference in the overall permeability of the two types of strata;

[0014] Furthermore, considering the impact of the heterogeneity of permeability within the formation on water permeability, it is necessary to quantify the dispersion of permeability coefficients between the fault zone and normal formations, and calculate the variance of the permeability coefficient in the fault zone. Variance of normal formation permeability coefficient To characterize the discrete characteristics of permeability within normal surrounding rock due to differences in weathering degree and rock mass structure, a comprehensive dispersion index is constructed to unify the quantitative dimensions of mean difference and variance difference. This indicator integrates the heterogeneous characteristics of the two types of strata, providing a standardized comparison benchmark for mean differences;

[0015] Finally, based on the ratio of mean difference to overall dispersion, the expression for the difference quantification model is derived: ;in, The comprehensive difference in permeability coefficients between fault zones and normal formations is calculated by coupling the mean difference with the comprehensive dispersion, thus achieving an objective and dimensionless quantification of the differences in permeability characteristics between the two types of formations.

[0016] Furthermore, S1 also includes: identifying abnormal signals from the measured data, and obtaining the comprehensive permeability coefficient of the construction section through weighted calculation based on the surrounding rock classification, weathering degree, and lithological structure. All coupled data are preprocessed using a normalization method adapted to the corresponding module to form multivariate time-series training data.

[0017] Furthermore, the overall permeability coefficient The calculation method is as follows:

[0018] First, define the background permeability coefficient of the rock formation. The parameter is taken from the identified and verified groundwater numerical model, and corresponds to the benchmark value of the permeability of the rock strata at the construction section under the natural geological state without construction disturbance. Its data is consistent with the geological survey report and rock sample test data of the study area.

[0019] Secondly, the core geological influencing factors of the construction section are selected, and the surrounding rock classification quantification factor is set as follows: The quantification factor for the degree of weathering of rock mass is: The quantification factor for the degree of lithological structure fragmentation is: Each factor is quantified through interval mapping. Different dimensionless quantification values ​​are assigned to different surrounding rock types, weathering grades, and lithological structures. The mapping rule is that the more significant the enhancement effect of geological features on permeability, the larger the corresponding factor value.

[0020] Furthermore, a dynamic correction term for geological influence factors is introduced, expressed as a product of the corresponding exponents of each geological influence factor, resulting in a generalized calculation expression for the comprehensive permeability coefficient. ;in The number of geological influence factors involved in the calculation. Not less than 3; This represents the natural weight index of each factor; The comprehensive permeability coefficient of the construction section is used to characterize the actual permeability of the construction section under actual geological conditions. The background permeability coefficient of the rock strata is derived from the physical simulation results of the groundwater numerical model; For the first The quantitative values ​​of each geological influence factor are obtained based on the quantification of objective geological characteristics, ensuring the standardization of geological characteristic representation.

[0021] Furthermore, the three anomaly detection models in S3 are the Long Short-Term Memory Neural Network model, the Isolation Forest model, and the K-Nearest Neighbors Regression model. Based on the design logic of functional complementarity and synergistic coupling, they each undertake different anomaly identification tasks:

[0022] The Long Short-Term Memory Neural Network Model relies on its ability to capture long-term dependencies in time-series data. Starting from the time-series evolution of tunnel or mine water inflow data, it identifies trend anomalies caused by dynamic changes in hydrogeological conditions and continuous disturbances during construction. By mining the implicit correlations in the time-series characteristics of the data, it locates the continuous abnormal changes in tunnel or mine water inflow over time.

[0023] The isolated forest model is based on the principle of random partitioning. It focuses on the global discrete features in the water inflow data of tunnels or mines. It specifically identifies globally isolated anomalous data points caused by factors such as changes in hydrological environmental conditions at the excavation face and sudden changes in geological structure. By randomly cutting and partitioning the data space and then evaluating the discreteness of the data, it captures anomalous information features that deviate significantly from the distribution patterns of surrounding data.

[0024] The K-nearest neighbor regression model identifies localized abnormal water inrush events caused by factors such as changes in tunnel or mine construction plans and adjustments due to momentary equipment failures, based on the statistical regularity of local data. By defining local data, it filters out abnormal information in this part of the data.

[0025] Furthermore, the weighted fusion strategy in S3 is specifically a linear weighted strategy:

[0026] The weighting coefficients for the detection results of the three types of models are defined as follows: The weighting coefficients are determined based on the anomaly detection accuracy and recall of various models on the validation set, with higher weights assigned to models with better detection performance.

[0027] Furthermore, the process of adaptively adjusting the hyperparameters of each module in S3 is as follows:

[0028] First, a multi-objective error evaluation function is constructed. Based on the prediction results and measured data of multiple tunnels or mines, the root mean square error is calculated. and mean absolute percentage error Construct a comprehensive error index ,in Indicates the first A tunnel or mine, and The inherent weights of the error index are naturally determined by the data distribution characteristics and engineering accuracy requirements. , , , For the first The predicted time steps for a tunnel or mine. For the first The first tunnel or mine The measured inflow rate at the time step Predict the inflow rate for the model at the corresponding time step;

[0029] Define the set of model hyperparameters ,in This is the hyperparameter vector of the LSTM model, containing the number of hidden layer neurons. Number of iterations Learning rate descent rate, regularization parameter; This is the hyperparameter vector for the Isolation Forest model, containing the number of decision trees. Sample size Coefficient of variation; This is the hyperparameter vector of the K-nearest neighbor regression model, which includes the number of nearest neighbor samples K and the Euclidean distance metric.

[0030] Define residual correction weight vector , This represents the total number of tunnels or mines. For the first The residual correction weights corresponding to each tunnel or mine shaft satisfy the following conditions: ;

[0031] The objective function is to minimize the mean of the overall error of all tunnels or mines. The model's hyperparameters are optimized using a grid search method. This grid search method exhaustively traverses all parameter combinations within a preset hyperparameter space, verifies each parameter combination, calculates the corresponding mean comprehensive error, and finally selects the optimal hyperparameter combination that minimizes the mean comprehensive error, thus completing the global optimization of the model's hyperparameters.

[0032] Furthermore, the correction process for the residual correction weights in S3 is as follows:

[0033] The residual correction weights are dynamically allocated based on the prediction accuracy of each tunnel or mine. This weight allocation is implemented using a conditional mapping function, which first calculates the prediction accuracy quality score for each tunnel or mine. The calculation formula is:

[0034]

[0035] In the formula, The coefficient of determination is derived from the fitting analysis of the predicted value and the measured water inflow data of a single tunnel or mine. The correlation coefficient between the predicted and measured values. MAE improvement rate, quality score A higher value indicates better prediction accuracy after residual correction for the corresponding tunnel or mine; for short tunnels with insufficient data samples, the calculated quality score is... Deductions and corrections are made, resulting in a revised quality score. Based on the revised quality score Five levels of prediction accuracy evaluation criteria were defined, and the corresponding residual correction weights were directly matched through a conditional mapping function. The mapping relationship is as follows:

[0036]

[0037] The corresponding evaluation levels are, in order: excellent, good, average, poor, and very poor.

[0038] As a second aspect of the present invention, a spatiotemporal prediction system for tunnel or mine water inrush coupled with a hydrodynamic numerical model is also provided, comprising:

[0039] The multi-source data coupling preprocessing unit is used to output predicted data related to water inrush in tunnels or mines, groundwater environment data, and geological body permeability data based on the identified and verified groundwater dynamic numerical model. It uses numerical model prediction data to supplement the missing or insufficient parts of the actual monitoring data, and uses difference analysis method to quantify the difference in permeability characteristics between faults and normal strata and couple it to the data system.

[0040] The coupled model construction configuration unit is used to construct coupled models using methods including Long Short-Term Memory Neural Network (LSTM), Isolation Forest (iForest), and K-Nearest Neighbor Regression. It also configures feature extraction, rolling prediction, data expansion, label statistics, outlier detection, and residual correction modules. Each module completes data interaction and training through a preset interface.

[0041] The model training dynamic optimization unit is used to divide the preprocessed multivariate time-series training data into training and test sets according to time order, and input them into the constructed coupled machine learning model. First, the LSTM model is trained and hidden time-series features are extracted. Then, three anomaly detection models are trained separately and the learning results are integrated through a weighted fusion strategy. The residual correction model is trained simultaneously. Based on the feedback of water inflow prediction errors from multiple tunnels or mines, the hyperparameters and residual correction weights of each module are adaptively adjusted to complete the dynamic optimization of the model training process.

[0042] The tunnel or mine water inflow prediction result correction unit is used to perform synchronous spatiotemporal dynamic prediction of water inflow for multiple tunnels or mines using the window scrolling strategy of the rolling prediction module and the optimized coupled machine learning model. It outputs the water inflow prediction results for different time steps and different tunnel or mine cross sections. Combined with the outlier detection module to identify abnormal data in the prediction process, the prediction results are corrected in real time by the residual correction module. Finally, the spatiotemporal dynamic prediction data of tunnel or mine water inflow that meets the engineering accuracy requirements is output.

[0043] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor as described in any one of the claims, a method for spatiotemporal prediction of water inrush in tunnels or mines coupled with a hydrodynamic numerical model.

[0044] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0045] 1. The spatiotemporal prediction method for tunnel or mine water inrush using a coupled hydrodynamic numerical model of the present invention outputs multi-source geological and hydrological data based on an identified and verified groundwater numerical model. It uses numerical model prediction data to supplement missing parts of actual monitoring data and simultaneously constructs a quantitative model of geological body permeability differences to quantify the differences in permeability characteristics between faults and normal strata and couples this model to the data system, achieving effective integration and preprocessing of multi-source data. This method constructs a multi-factor coupled comprehensive permeability coefficient generalization calculation model targeting core geological influencing factors such as surrounding rock classification, weathering degree, and lithological structure of the construction section. This accurately characterizes the actual permeability capacity of the construction section, solving the problems of incomplete measured data and inaccurate quantification of geological characteristics in traditional methods. It provides high-quality multivariate time-series training data with both physical meaning and operational condition adaptability for subsequent model training, ensuring the reliability and consistency of data input and laying a solid data foundation for water inrush prediction.

[0046] 2. The spatiotemporal prediction method for tunnel or mine water inrush using a coupled hydrodynamic numerical model of the present invention constructs a weighted coupled model of long short-term memory neural network-isolation forest-K-nearest neighbor regression, configuring functional modules such as feature extraction, rolling prediction, outlier detection, and residual correction. The model is trained by dividing the training and testing sets according to time sequence. During training, the long short-term memory neural network is first used to extract hidden features from the data. Then, three anomaly detection models are trained, and the results are integrated using a weighted fusion strategy. Simultaneously, the residual correction model is trained. The hyperparameters of each module and the residual correction weights are adaptively adjusted based on the prediction error feedback from multiple tunnels or mines, completing the dynamic optimization of model training. This process enables the model to learn the correlation between the temporal variation of water inrush and geological features, improving the model's ability to identify trend-based, locally isolated, and global anomalies, and ensuring the model's generalization ability and prediction accuracy.

[0047] 3. The spatiotemporal prediction method for tunnel or mine water inflow using a coupled hydrodynamic numerical model of the present invention employs a window scrolling strategy of a scrolling prediction module to perform synchronous spatiotemporal dynamic prediction of water inflow for multiple tunnels or mines using an optimized coupled machine learning model, outputting water inflow prediction results at different time steps and cross-sections. During the prediction process, an outlier detection module identifies abnormal prediction data, and a residual correction module corrects the prediction results in real time, ultimately outputting spatiotemporal dynamic prediction data of tunnel or mine water inflow that meets engineering accuracy requirements. This strategy, by setting reasonable prediction windows, step sizes, and prediction step lengths, achieves continuous scrolling of the prediction window, ensuring the continuity and timeliness of prediction results, and solves the problem that traditional prediction methods struggle to achieve spatiotemporal dynamic prediction, providing continuous and accurate water inflow prediction support for safety control during tunnel or mine construction. Attached Figure Description

[0048] Figure 1This is a flowchart of spatiotemporal prediction of water inrush in tunnels or mines using a coupled hydrodynamic numerical model, according to an embodiment of the present invention.

[0049] Figure 2 This is a virtual borehole distribution map for groundwater level monitoring in a tunnel numerical model according to an embodiment of the present invention.

[0050] Figure 3 This is a schematic diagram of the verification results of numerical simulation prediction of tunnel water inflow in an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram of the simulation results of the periodic change of groundwater level in the tunnel numerical model according to an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram of a groundwater drawdown cone generated under simulated plugging conditions according to an embodiment of the present invention.

[0053] Figure 6 This is a schematic diagram of a groundwater drawdown funnel generated under simulated heavy rain conditions according to an embodiment of the present invention.

[0054] Figure 7 This is a schematic diagram of a wavelet decomposition anomaly signal scheme according to an embodiment of the present invention;

[0055] Figure 8 This is a diagram showing the simultaneous prediction results of different tunnel water inflow rates based on the method in an embodiment of the present invention.

[0056] Figure 9 This is a diagram illustrating the improvement in prediction accuracy achieved by adaptive residual correction in an embodiment of the present invention.

[0057] Figure 10 This is a comparison diagram of machine learning and pure numerical simulation prediction based on coupled numerical simulation of the No. 2 branch tunnel section in an embodiment of the present invention.

[0058] Figure 11 This is a comparison chart of machine learning and pure numerical simulation prediction based on coupled numerical simulation of the inlet section in an embodiment of the present invention.

[0059] Figure 12 This is a system unit diagram of an embodiment of the present invention;

[0060] In all the attached figures, d represents the day. It indicates the daily water inflow, specifically referring to the cubic meter value corresponding to the daily water inflow. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0062] Example 1

[0063] Please refer to Figure 1 This embodiment 1 provides a method for spatiotemporal prediction of water inrush in tunnels or mines using a coupled hydrodynamic numerical model, including:

[0064] S1. Based on the identified and verified groundwater dynamic numerical model, output the predicted data related to water inrush in tunnels or mines, groundwater environment data and geological body permeability data. Use the numerical model prediction data to supplement the missing or insufficient parts of the actual monitoring data. Use the difference analysis method to quantify the difference in permeability characteristics between faults and normal strata and couple it to the data system.

[0065] S2. A coupled model is constructed using methods including Long Short-Term Memory Neural Network (LSTM), Isolation Forest (iForest), and K-Nearest Neighbor Regression. At the same time, modules for feature extraction, rolling prediction, data expansion, label statistics, outlier detection, and residual correction are configured. Each module completes data interaction and training through a preset interface.

[0066] S3. The preprocessed multivariate time-series training data is divided into training and test sets according to time order and input into the constructed coupled machine learning model. First, the LSTM model is trained and hidden time-series features are extracted. Then, three anomaly detection models are trained separately and the learning results are integrated through a weighted fusion strategy. The residual correction model is trained simultaneously. Based on the feedback of water inflow prediction errors from multiple tunnels or mines, the hyperparameters and residual correction weights of each module are adaptively adjusted to complete the dynamic optimization of the model training process.

[0067] S4. Using the window scrolling strategy of the scrolling prediction module, the optimized coupled machine learning model is used to perform synchronous spatiotemporal dynamic prediction of water inflow for multiple tunnels or mines, and output the water inflow prediction results for different time steps and different tunnel or mine cross sections; combined with the outlier detection module to identify abnormal data in the prediction process, the prediction results are corrected in real time through the residual correction module, and finally the spatiotemporal dynamic prediction data of tunnel or mine water inflow that meets the engineering accuracy requirements is output.

[0068] The “numerical model based on identified and verified groundwater” used in this embodiment is a three-dimensional unsteady groundwater flow numerical model constructed based on actual engineering survey and monitoring data and verified by multi-dimensional measured data.

[0069] The model is constructed based on fundamental hydrogeological data obtained from the preliminary geological survey of the engineering area, including topography, geological structure, stratigraphy, fault development, and source-sink parameters. Parameters, including permeability coefficient and porosity data, were obtained through literature review and groundwater flow field fitting. Figure 2 As shown, the model also incorporates groundwater levels obtained through actual measurements and virtual observation holes (such as XNZK001, XNZK002, and XNZK003) during construction, and compiles monitoring data such as daily tunnel water inflow detection data provided by the construction party, providing a real and reliable physical basis for the model.

[0070] Guided by groundwater dynamics and numerical simulation theory, the model sets the hydrogeological unit where the tunnel or mine is located as the three-dimensional simulation calculation domain. Combined with source and sink terms, the corresponding boundary conditions and initial conditions are clarified. After the irregular grid is discretized by groundwater simulation software, a three-dimensional steady flow numerical model that can simulate groundwater movement and the interaction between the tunnel and groundwater is formed. On the basis of the steady flow model, the tunnel or mine excavation time series is added to simulate the unsteady flow model.

[0071] After the unsteady flow model is simulated, its reliability is verified through multi-dimensional measured data: such as... Figure 3 As shown, the simulated water inflow results for key areas such as the inlet tunnel section and the No. 2 branch tunnel control section are compared with the measured data, and the error is within a controllable range; Figure 4 As shown, the simulated groundwater level conforms to the seasonal cycle and falls within the range of measured water level fluctuations, accurately depicting the changes in the groundwater level flow field at different time points; for example... Figure 5 as well as Figure 6 As shown, for both blockage and heavy rain conditions, the model can predict the range of the groundwater drawdown cone and the groundwater level drop caused by tunnel excavation, while accurately reflecting the actual characteristic that heavy rain has a negligible impact on the tunnel's water inflow in the area.

[0072] The groundwater numerical model, verified by measured groundwater level and tunnel inflow data, exhibits a close "generation-dependency" relationship with the data source of the method in this embodiment. This model systematically outputs inflow-related prediction data, groundwater environmental data, and geological body permeability data by simulating physical processes such as groundwater movement, geological body permeability characteristics, and the impact of tunnel excavation. These data derived from numerical simulations, based on physical processes, can form multi-level coupling with the measured data, filling in missing data in the measured data and assisting machine learning algorithms in identifying complex hydrogeological structures in the study area.

[0073] Next, this embodiment 1 will further elaborate on the specific steps of the method used in this embodiment.

[0074] (1) Multi-source data coupling preprocessing

[0075] In tunnel and mine water inrush prediction, accurate quantification of geological permeability characteristics and effective integration of measured data are crucial prerequisites for ensuring prediction accuracy. Due to factors such as the engineering site environment, equipment operating status, and adjustments to construction procedures, actual monitoring data on tunnel or mine water inrush volume and groundwater levels often suffer from gaps or insufficient data, directly impacting the training effectiveness of subsequent prediction models. Furthermore, the permeability characteristics of faults and normal strata differ significantly; failure to accurately quantify these differences will prevent the model from accurately identifying key control factors for water inrush risk.

[0076] Based on the identified and verified groundwater numerical model, it can output prediction data related to water inflow, groundwater environmental data, and geological body permeability characteristics data. For missing portions of measured tunnel water inflow data due to factors such as holiday shutdowns and equipment failures, the corresponding time period's predicted tunnel water inflow data output by the numerical model is directly used to fill in the gaps. Similarly, for missing or insufficient groundwater level monitoring data, the water level data output by the numerical model is used to supplement the data, thereby ensuring the integrity and continuity of the data.

[0077] When quantifying the differences in permeability characteristics between fault zones and normal formations, the core differences between these two types of formations are firstly identified in two aspects: the average difference in overall permeability and the heterogeneity of internal permeability distribution. To comprehensively characterize these two types of differences, a time series sequence of permeability coefficients of the fault zone is extracted based on continuous time step data output from the groundwater numerical model. ,in, The time steps in the numerical simulation represent the global permeability coefficient value of the fault zone at a specific time step for each data point. This value originates from the physical simulation results of the fault zone's lithology and fracture development, as well as the time series of permeability coefficients of normal strata during the same period. Based on the same numerical simulation system as the fault zone data, the consistency of the hydrogeological environment is ensured, and the simulated value of the permeability coefficient of the normal strata is obtained for the whole area.

[0078] Secondly, calculate the cumulative mean of the permeability coefficients of the fault zone and the normal formation, and the cumulative mean of the permeability coefficient of the fault zone. Cumulative mean of normal formation permeability As a benchmark value for the average permeability of the normal surrounding rock around the fault, and To form a comparison of differences along the mean dimension, i.e. This difference directly reflects the difference in the overall permeability of the two types of strata;

[0079] Furthermore, considering the impact of the heterogeneity of permeability within the formation on water permeability, it is necessary to quantify the dispersion of permeability coefficients between the fault zone and normal formations, and calculate the variance of the permeability coefficient in the fault zone. Variance of normal formation permeability coefficient To characterize the discrete characteristics of permeability within normal surrounding rock due to differences in weathering degree and rock mass structure, a comprehensive dispersion index is constructed to unify the quantitative dimensions of mean difference and variance difference. This indicator integrates the heterogeneous characteristics of the two types of strata, providing a standardized comparison benchmark for mean differences;

[0080] Finally, based on the ratio of mean difference to overall dispersion, the expression for the difference quantification model is derived: ;in, This method calculates the comprehensive difference in permeability coefficients between fault zones and normal formations by coupling the mean difference with the comprehensive dispersion, thus achieving an objective and dimensionless quantification of the differences in permeability characteristics between the two types of formations. This difference value is then coupled into the data system.

[0081] After completing data imputation and differential quantization coupling, preprocessing of the measured data is necessary. First, using MATLAB's wavelet signal analysis toolbox, signal decomposition is performed on the water inflow-time data of different tunnels or mine sections, and abnormal tunnel water inflow signals (i.e., drastic short-term changes in tunnel water inflow) are filtered out. The relevant time of the abnormal water inflow event is selected based on the location of these signals. Then, combined with actual engineering logging, the causes of abnormal tunnel water inflow changes are identified, ensuring accurate identification of the causes of special events. An example of wavelet decomposition of abnormal signals can be seen... Figure 7 ;

[0082] The specific identification principle is to process the tunnel water inflow monitoring signal by frequency division (highest frequency, second highest frequency, mid-high frequency, and low frequency) using the db4 wavelet basis, and extract the abrupt change and fluctuation characteristics of the high-frequency or mid-high frequency sub-signals. The correspondence between the wavelet decomposition signal and the abnormal phenomenon is as follows: the instantaneous abrupt change point in the high-frequency sub-signal corresponds to a single-step depression or bulge, while its repeated fluctuation corresponds to an oscillation anomaly; the continuous fluctuation in the second highest frequency and mid-high frequency sub-signals corresponds to a continuous depression or bulge.

[0083] The correspondence between the five types of data anomalies identified above and the actual phenomena is as follows: 1. Oscillation anomaly: indicates repeated fluctuations in tunnel water inflow within a short period; 2. Single-step depression: indicates a significant decrease in tunnel water inflow on a particular day, while the water inflow on subsequent days remains stable; 3. Single-step bulge: indicates a significant increase in tunnel water inflow on a particular day, while the water inflow on subsequent days remains stable; 4. Continuous depression: indicates a continuous decrease in tunnel water inflow; 5. Continuous bulge: indicates a continuous increase in tunnel water inflow. (Comparison) Figure 7Different frequency signals from the inlet construction section and the upstream construction section of the adit tunnel reveal that the abnormal water inflow in the upstream construction section of the adit tunnel is significantly more frequent than that in the inlet construction section.

[0084] Subsequently, the permeability coefficient of the construction section of tunnels or mines was differentiated. To enable the prediction model to effectively identify different cross-sectional structures, the surrounding rock classification, weathering degree, and lithological structure were selected as the core influencing factors of the permeability coefficient of the construction section. First, the background permeability coefficient of the rock strata was defined. The parameter is taken from the identified and verified groundwater numerical model, and corresponds to the benchmark value of the permeability of the rock strata at the construction section under the natural geological state without construction disturbance. Its data is consistent with the geological survey report and rock sample test data of the study area.

[0085] Secondly, the core geological influencing factors of the construction section are selected, and the surrounding rock classification quantification factor is set as follows: The quantification factor for the degree of weathering of rock mass is: The quantification factor for the degree of lithological structure fragmentation is: Each factor is quantified through interval mapping. Different dimensionless quantification values ​​are assigned to different surrounding rock types, weathering grades, and lithological structures. The mapping rule is that the more significant the enhancement effect of geological features on permeability, the larger the corresponding factor value.

[0086] Furthermore, a dynamic correction term for geological influence factors is introduced, expressed as a product of the corresponding exponents of each geological influence factor, resulting in a generalized calculation expression for the comprehensive permeability coefficient. ;in The number of geological influence factors involved in the calculation. Not less than 3, and factors such as rock strata water content and fracture development density can be added according to the complexity of engineering geology; The natural weight index for each factor is determined by correlation analysis between geological influence factors and permeability; the stronger the correlation, the lower the weight index. The larger the absolute value, and When the value is greater than 0, the characterization factor enhances the permeability. A value less than 0 indicates a weakening effect; The comprehensive permeability coefficient of the construction section is used to characterize the actual permeability of the construction section under actual geological conditions. The base permeability coefficient of the rock strata is derived from the physical simulation results of the groundwater numerical model, ensuring the reliability of the physical basis of the calculation. For the first The quantitative values ​​of each geological influence factor are obtained based on the quantification of objective geological characteristics, ensuring the standardization of geological characteristic representation.

[0087] Finally, for all the full-dimensional data after coupling integration, anomaly identification, and feature quantization, according to the adaptation characteristics and operation logic of each functional algorithm module in the coupled machine learning system, corresponding dedicated normalization preprocessing methods are used to complete the data regularization process. Each normalization method is selected specifically to fit the algorithm operation conditions and data processing needs of the corresponding module, ensuring the operation adaptability and data correlation of each module.

[0088] For core feature data, maximum-minimum normalization is employed to scale all feature indicators to a fixed numerical range, completely eliminating dimensional differences between different geological and hydrological indicators and ensuring the comparability of feature data. For feature data in the rolling prediction stage, Z-score standardization is used, relying entirely on the mean and standard deviation statistics of the training set to standardize new input data, effectively maintaining data distribution consistency between training and prediction stages and avoiding prediction bias caused by data shifts. For tunnel or mine-related data with insufficient data, trend extrapolation is used for extended normalization, supplementing data and scaling based on historical data evolution trends, balancing data completeness and trend consistency.

[0089] In the label data processing stage, a local sliding statistical normalization method is adopted. This method extracts the local temporal mean and standard deviation features of the label data through a sliding window, effectively enhancing the temporal correlation information of the data and meeting the core requirements of time series prediction. For the outlier detection module, a... The algorithm performs data normalization by defining reasonable data boundaries based on the overall mean and standard deviation of the data, identifying outlier data points that exceed the normal distribution range, and providing quantitative evidence for outlier identification. The residual correction stage employs a weighted moving average normalization method, combining the weights of predicted and historical true values ​​to correct the residuals and calibrate the current prediction results. For residual feature data, implicit normalization is used, with the feature input layer of the neural network autonomously performing the normalization calculation, adapting to the internal computational logic of the neural network.

[0090] After adaptation and normalization preprocessing of each module, all data retain the original regularity and temporal evolution attributes of geological and hydrological characteristics while eliminating dimensional interference and avoiding distribution shifts. The data interaction between modules is smooth and the feature representation is accurate, ultimately forming standardized multivariate time-series training data with completeness, consistency and accuracy.

[0091] (2) Configuration of Coupled Model Construction

[0092] In practical applications of spatiotemporal prediction of water inrush in tunnels and mines, a single machine learning algorithm is insufficient to meet the full-dimensional requirements of feature mining, local anomaly identification, and global pattern capture of time-series data. Each algorithm has its own adaptability and application limitations, and cannot independently meet the engineering requirements of water inrush prediction in multiple tunnels and multiple cross-sections. It is also difficult to cope with the complex characteristics of water inrush data, which have both temporal continuity, local suddenness, and global correlation. Therefore, it is necessary to build a more adaptable prediction system by coupling multiple algorithms with dedicated functional modules to achieve complementary advantages and functional synergy among the algorithms.

[0093] Based on this, a weighted coupled model of Long Short-Term Memory Neural Network (LSTM), Isolation Forest (iForest), and K Nearest Neighbor Regression (KNN) is built. This model integrates three core algorithms: LSTM, Isolation Forest (iForest), and K Nearest Neighbor Regression (KNN). It also includes targeted configurations for feature extraction, rolling prediction, short tunnel and mine data expansion, label statistics, outlier detection, and residual correction modules. Each functional module has a preset interaction interface, and various types of data and calculation results can be efficiently transmitted and connected through the interface. Each module and the core algorithm cooperate and work together to form a complete coupled prediction system, realizing a closed-loop operation from data feature processing to final result output.

[0094] In this coupled model, various core algorithms play differentiated and irreplaceable roles, with their functional positioning tailored to the characteristics and predictive needs of water inflow data. The core function of the Long Short-Term Memory (LSTM) neural network module is to extract the temporal features of the training data. This algorithm can deeply mine the continuous evolution of water inflow data in tunnels or mines over time, capture the trend characteristics of water inflow caused by changes in hydrogeological conditions and construction conditions, accurately identify the long-term temporal correlation information hidden in the data, and provide core temporal pattern support for overall prediction.

[0095] The K-Nearest Neighbors (KNN) module is specifically designed for learning local abnormal water inrush events. This algorithm can focus on local discrete features in tunnel or mine water inrush data, accurately identify local isolated abnormal water inrush data caused by factors such as sudden geological events and instantaneous equipment failures during construction, capture local abnormal information that deviates significantly from the distribution pattern of surrounding normal data, and effectively identify single-point, sudden local abnormal water inrush events, thus filling the gap in the LSTM time series algorithm's ability to identify local anomalies.

[0096] The Isolation Forest (iForest) module is mainly used to capture global abnormal water inrush characteristics. Relying on its own advantage in fitting global data distribution characteristics, it starts from the statistical regularity of the overall data to identify global water inrush anomalies caused by changes in the groundwater migration environment in tunnels or mines due to construction excavation and geological and climatic changes. It can accurately capture abnormal fluctuation patterns within the overall data distribution range and effectively identify global and trend-based large-scale water inrush anomalies.

[0097] Various functional modules provide adaptive support for three core algorithms: the feature extraction module extracts effective data features for algorithm operation; the rolling prediction module supports continuous dynamic prediction needs across multiple time steps; the data expansion module addresses the data volume shortage in short tunnels and mines; the label statistics module strengthens the temporal correlation attributes of data; the outlier detection module accurately identifies abnormal data during the prediction process; and the residual correction module accurately calibrates the prediction results.

[0098] Each algorithm module performs its own function and works in concert, relying on preset interfaces to complete data interaction and computation connection. This allows the coupled model to not only deeply learn the temporal patterns of water inrush data, but also accurately identify various abnormal water inrush characteristics in both local and global contexts, thus adapting to the actual engineering needs of spatiotemporal prediction of water inrush in multiple tunnels and mines.

[0099] (3) Dynamic optimization of model training

[0100] After completing the entire data preprocessing process, the resulting multivariate time-series training data needs to be imported into the constructed Long Short-Term Memory Neural Network (LSTM)-Isolation Forest (iForest)-K Nearest Neighbor Regression (KNN) weighted coupling model for system training. Since multiple tunnels or mines exist in actual engineering projects, the length of monitoring data from different tunnels or mines may vary, and it is necessary to ensure accurate correspondence of the time dimension during prediction. Therefore, before training begins, data length standardization must be completed. Data expansion methods are used to align short tunnel data with long tunnel data, ensuring that all data are strictly matched one-to-one in the time series, laying the foundation for subsequent synchronous training and prediction across multiple projects.

[0101] After data normalization, the feature and label data are normalized according to the adaptation requirements of each algorithm module to eliminate the differences in the units of measurement of data in different dimensions and accelerate the convergence speed of the model. The data is divided into training and testing sets, and the temporal order of the data is strictly maintained to avoid the model learning false temporal correlations due to disordered time order. This ensures that the training set can truly reflect the historical evolution of the water inrush data, and the testing set can effectively verify the model's predictive ability for future data.

[0102] The training process proceeds step by step with multi-algorithm collaborative learning and dynamic optimization as its core logic. First, the LSTM model is trained, relying on its ability to capture long-term dependencies in time-series data to deeply mine the evolutionary patterns of water inflow contained in the training data. Simultaneously, hidden features generated during model training are extracted. These hidden features can accurately characterize the trend and correlation information of tunnel or mine water inflow hidden in geological engineering data, providing core feature support for subsequent anomaly detection.

[0103] Next, we trained three anomaly detection models: LSTM, iForest, and KNN. Based on the design logic of functional complementarity and synergistic coupling, these three models undertook different anomaly identification tasks.

[0104] Among them, the Long Short-Term Memory Neural Network (LSTM) model relies on its ability to capture long-term dependencies of time-series data. Starting from the time-series evolution of tunnel or mine water inflow data, it identifies trend anomalies caused by dynamic changes in hydrogeological conditions and continuous disturbances during construction. By mining the implicit correlations in the time-series characteristics of the data, it locates the continuous trend of tunnel or mine water inflow over time.

[0105] The Isolation Forest (iForest) model is based on the principle of random partitioning. It focuses on the global discrete features in tunnel or mine water inflow data. It specifically identifies globally isolated anomalous data points caused by factors such as changes in the hydrological environment of the excavation face and sudden changes in geological structure. By randomly cutting the data space and then evaluating the discreteness of the data, it captures anomalous information that deviates significantly from the distribution pattern of the surrounding data.

[0106] The K-Nearest Neighbors (KNN) model identifies localized abnormal water inrush events caused by factors such as changes in construction plans and momentary equipment failures from the perspective of statistical regularities in local data. By defining local data, it filters out abnormal information in this part of the data.

[0107] After the three models are trained, a weighted fusion strategy is used to integrate their respective anomaly detection results, fully leveraging the advantages of each model to achieve comprehensive and accurate identification of different types of abnormal water inrush events. The weighted fusion strategy used in this embodiment is a linear weighting strategy: the weight coefficients of the detection results of the three models are defined as follows: The weighting coefficients are determined based on the anomaly detection accuracy and recall of various models on the validation set, with higher weights assigned to models with better detection performance.

[0108] Simultaneously with the training of the aforementioned model, a residual correction model is trained. This model aims to learn the residual relationship between the predicted results and the actual water inflow data of tunnels or mines, providing support for the accurate calibration of subsequent prediction results. Model training is not completed in a single round, but rather relies on an error feedback mechanism for dynamic optimization. By adaptively adjusting the hyperparameters of each module and the residual correction weights, the model's prediction accuracy and generalization ability are continuously improved.

[0109] The dynamic optimization process is driven by the core objective of minimizing the overall error. First, a multi-objective error evaluation function is constructed, and the root mean square error is calculated based on the prediction results and measured data from multiple tunnels or mines. and mean absolute percentage error Construct a comprehensive error index ,in Indicates the first A tunnel or mine, and The inherent weights of the error index are naturally determined by the data distribution characteristics and engineering accuracy requirements. , , For the first The predicted time steps for a tunnel or mine. For the first The first tunnel or mine The measured inflow rate at the time step Predict the inflow rate for the model at the corresponding time step;

[0110] Define the set of model hyperparameters ,in This is the hyperparameter vector of the LSTM model, containing the number of hidden layer neurons. Number of iterations Learning rate descent rate, regularization parameter; This is the hyperparameter vector for the Isolation Forest model, containing the number of decision trees. Sample size Coefficient of variation; This is the hyperparameter vector of the K-nearest neighbor regression model, which includes the number of nearest neighbor samples K and the Euclidean distance metric.

[0111] Define residual correction weight vector , This represents the total number of tunnels or mines. For the first The residual correction weights corresponding to each tunnel or mine shaft satisfy the following conditions: ;

[0112] The objective function is to minimize the mean of the overall error of all tunnels or mines. The model's hyperparameters are optimized using a grid search method. This grid search method exhaustively traverses all parameter combinations within the preset hyperparameter space, verifies each parameter combination one by one, calculates the corresponding mean comprehensive error, and finally selects the optimal hyperparameter combination that minimizes the mean comprehensive error, thus completing the global optimization of the model's hyperparameters.

[0113] Simultaneously, the residual correction weights are dynamically allocated based on the prediction accuracy of each tunnel or mine. This weight allocation is implemented using a conditional mapping function, first calculating the prediction accuracy quality score for each tunnel or mine. The calculation formula is:

[0114]

[0115] In the formula, As the coefficient of determination, in this quality score calculation, The value is derived from the fitting analysis of the predicted value and the measured water inflow data of a single tunnel or mine, and its calculation formula is:

[0116]

[0117] In the formula, For the first The first tunnel or mine The measured inflow rate at the time step To predict the inflow rate for the corresponding time step of the model, For the first The average measured water inflow at all time steps in a tunnel or mine. For the first The predicted time steps for a tunnel or mine; The range of values ​​is The closer the value is to 1, the better the fit between the predicted value and the measured value.

[0118] The correlation coefficient between the predicted and measured values. MAE improvement rate, quality score The higher the value, the better the prediction accuracy after residual correction for the corresponding tunnel or mine;

[0119] For short tunnels with insufficient data sample size, the calculated quality score is... Deductions and corrections are made, resulting in a revised quality score. Based on the revised quality score Five levels of prediction accuracy evaluation criteria were defined, and the corresponding residual correction weights were directly matched through a conditional mapping function. The mapping relationship is as follows:

[0120]

[0121] The corresponding evaluation levels are, in order: excellent, good, average, poor, and very poor.

[0122] (4) Correction of prediction results for water inrush in tunnels or mines

[0123] After completing model training and dynamic optimization, the spatiotemporal dynamic prediction stage of water inflow can begin. Because water inflow data in tunnels and mines exhibit significant temporal continuity, and hydrogeological conditions and construction conditions are constantly changing, predictions at a single point in time are insufficient to meet the full-cycle requirements of engineering safety control. Therefore, it is necessary to rely on the window rolling strategy of the rolling prediction module to achieve synchronous and continuous prediction of multiple tunnels or mines, ensuring that the predicted water inflow results can correspond in real time to changes in working conditions and geological conditions.

[0124] The implementation of the window rolling strategy is divided into three stages. The initial prediction stage is to use the last subsequence of the preprocessed multivariate time series data as the initial input window. For example, the initial window contains feature data from row 48 to row 50. The window data is input into the trained LSTM model and the forward propagation operation is performed to obtain the predicted value of tunnel or mine water inflow in row 51.

[0125] Subsequently, three types of anomaly detection (LSTM anomaly detection, isolated forest anomaly detection, and KNN anomaly detection) are performed and the results are fused. The predicted values ​​are then calibrated by the residual correction module (a constraint coefficient of 0.3 is added to avoid over-correction). If an outlier is detected, the previous real value is partially introduced for adjustment. Finally, the inverse normalized prediction result is output, and the input window is updated to include the feature data from rows 49 to 51.

[0126] Iterative rolling phase: Based on the updated window (e.g., containing feature data from rows 49 to 51), input the LSTM model to perform forward propagation and predict the tunnel or mine water inflow in row 52; repeat the process of anomaly detection fusion, residual correction, outlier adjustment (introducing prior true values), and denormalization output, and then update the window to contain feature data from rows 50 to 52.

[0127] Execution in a loop: The above process of "window input → LSTM forward prediction → anomaly detection fusion → residual correction → anomaly adjustment → denormalization output → window update" is repeated continuously. In each iteration, the window is scrolled forward one line, and the prediction target is the next line after the end of the current window (e.g., if the window contains lines 50 to 52, predict line 53), until the prediction of all time steps is completed.

[0128] Among them, anomaly detection and residual correction are performed simultaneously. Anomaly detection provides support for risk warning, while residual correction provides numerical calibration. The two are in parallel.

[0129] Then repeat the entire process of window scrolling, model prediction, anomaly identification and result correction until the tunnel or mine water inflow prediction for all time steps within the preset prediction period is completed.

[0130] Finally, by stitching together the effective results of each rolling prediction, a continuous spatiotemporal dynamic prediction sequence of tunnel or mine water inflow is formed, covering different time steps and different tunnel or mine cross-sections. This sequence can fully present the water inflow variation trend of each tunnel and each cross-section as the cross-section position changes within the prediction period, achieving comprehensive coverage of the spatiotemporal dimensions, and ultimately outputting spatiotemporal dynamic prediction data of tunnel or mine water inflow that meets engineering accuracy requirements.

[0131] After each round of tunnel or mine water inflow prediction results is generated, anomaly identification and real-time correction processes must be initiated simultaneously to ensure the reliability and accuracy of the prediction data. First, the outlier detection module comprehensively screens the prediction data output for the current round. This module integrates the collaborative identification logic of three models: Long Short-Term Memory Neural Network (LSTM), Isolation Forest (iForest), and K-Nearest Neighbor Regression (KNN), enabling targeted identification of different types of anomalous data. Specifically, the LSTM submodule, based on temporal evolution patterns, captures trend-based anomalous predictions caused by continuous changes in hydrogeological conditions; the Isolation Forest submodule focuses on global data distribution, accurately locating global anomalies caused by overall changes in the regional hydrological environment; and the K-Nearest Neighbor Regression submodule identifies local anomalous data caused by sudden geological events and instantaneous equipment fluctuations from the perspective of local data statistical patterns. Through the comprehensive judgment of the identification results from these three models, all kinds of anomalies that may occur during the prediction process can be comprehensively covered, reducing the limitations of single-model learning and improving the accuracy of anomalous water inflow predictions.

[0132] After accurately identifying and labeling abnormal data, the residual correction module is immediately activated for real-time correction. This module, based on the residual correction model optimized during the training phase and adaptive configuration, implements differentiated correction strategies for different types of abnormal data and normal prediction data. For labeled abnormal prediction values, the module utilizes the residual patterns learned during training, combined with historical measured data from the corresponding tunnel or mine, concurrent hydrogeological monitoring information, and error characteristics from previous predictions, to calculate accurate correction coefficients. Simultaneously, based on dynamically optimized residual correction weights, it integrates effective information from historical true values ​​and current prediction values ​​to perform targeted calibration of abnormal values, eliminating abnormal biases. For unlabeled normal prediction data, the module also performs minor adjustments based on the compensation logic for small residuals to further improve prediction accuracy.

[0133] During the correction process, the residual correction module provides real-time feedback on the correction effect and compares it with the preset engineering accuracy threshold. If the corrected prediction data still does not meet the accuracy requirements, the outlier detection module is retrieved for a second screening to check for any missed anomalies. Simultaneously, the correction parameters are adjusted and recalibrated until the prediction data meets the engineering accuracy standards. This closed-loop process of "anomaly identification - accurate correction - accuracy verification" effectively eliminates various deviations in the prediction process. The final output of spatiotemporal dynamic prediction data for tunnel or mine water inflow not only fully presents the water inflow change trends at different time steps and cross-sections but also ensures that the data accuracy fully meets the engineering requirements for construction safety control, providing a reliable technical basis for on-site construction decisions.

[0134] In addition, this example also uses the training case results of eight tunnel sections in a certain tunnel research area as a demonstration, which can intuitively show the application effect of this method:

[0135] In the data preprocessing stage, through Figure 7 As can be seen from the wavelet decomposition anomaly signal case, only a small number of local anomaly signals appeared in the inlet construction section, while the high-frequency and sub-high-frequency sub-signal fluctuations were significantly more numerous in the upstream construction section of the branch tunnel. This result matches the distribution characteristics of the actual water inflow anomaly on site and verifies the effectiveness of the anomaly identification module.

[0136] After entering the prediction phase, Figure 8 The results of water inflow prediction for different tunnel sections are presented. The mean absolute percentage error (MAPE) of the prediction for the inlet section is only 8.7%. The dynamic prediction curve is in good agreement with the actual water inflow curve. Furthermore, the results of synchronous prediction for multiple tunnels can fully cover the water inflow fluctuation trend of each section, proving that the method can achieve synchronous prediction in multiple engineering scenarios.

[0137] In terms of accuracy improvement, Figure 9 The residual correction analysis shows that after adaptive residual correction, the MAPE values ​​of each tunnel section are reduced to varying degrees, with some tunnel sections showing an improvement of up to 4.17%. At the same time, the residual correction weights are dynamically allocated according to the error characteristics of each tunnel, further ensuring the reliability of the corrected data.

[0138] Compared to pure numerical simulation methods, Figure 10 , Figure 11 The prediction results for the branch tunnel section and the inlet section are shown separately: the pure numerical simulation curve deviates significantly from the measured water inflow of the tunnel or mine, while the prediction curve of the coupled model of this method has a significantly improved fit with the measured data. Especially at the time node of sudden change in water inflow in the tunnel or mine, the coupled model can more accurately capture the fluctuation characteristics, which reflects the advantages of this method after combining numerical model and machine learning.

[0139] In summary, this method has achieved the identification of water inrush anomalies and the synchronous prediction of water inrush in multiple tunnels and multiple segments in a certain tunnel project. It has also effectively improved the accuracy of water inrush prediction. Furthermore, it can complement pure numerical simulation methods, fill the gap in the high demand for high-precision data in pure numerical simulation, learn the data of the physical processes based on numerical simulation, and has higher physical interpretability compared to pure machine learning.

[0140] The spatiotemporal prediction method for water inrush proposed in this embodiment has broad application prospects in engineering fields such as deep-buried tunnels and mines with complex geological conditions. As engineering construction progresses towards high-risk and complex geological areas, the suddenness and uncertainty of water inrush disasters significantly increase, making traditional prediction methods insufficient to meet prevention and control needs. This method, through core designs such as multi-source geological and hydrological data coupling, differentiated anomaly identification, and dynamic residual correction, can capture the abnormal characteristics of water inrush in tunnels or mines at different scales, achieving simultaneous spatiotemporal dynamic prediction for multiple projects. It can be applied to construction safety monitoring and water inrush disaster early warning for various underground engineering projects such as traffic tunnels, water conservancy tunnels, and mining operations, providing data support for risk prediction and emergency response at engineering sites, and effectively reducing casualties and economic losses caused by water inrush disasters.

[0141] Meanwhile, the modular design and flexible adaptability of this method give it significant potential for widespread application and expansion. On one hand, its data preprocessing and model optimization logic can be adjusted according to the geological conditions and monitoring data characteristics of different projects, adapting to tunnels and mines of varying scales and types. On the other hand, with the deep application of IoT and big data technologies in the engineering field, this method can further integrate real-time monitoring sensor network data to construct a fully intelligent prevention and control system encompassing "data acquisition - intelligent analysis - accurate prediction - early warning push." ​​This provides important reference for the technological upgrading of the underground engineering and geological disaster prevention and control industry, helping to improve the overall safety and prevention level of the industry.

[0142] Example 2

[0143] Please refer to Figure 12 This embodiment 2 provides a spatiotemporal prediction system for tunnel or mine water inrush coupled with a hydrodynamic numerical model, including:

[0144] The multi-source data coupling preprocessing unit is used to output predicted data related to water inrush in tunnels or mines, groundwater environment data, and geological body permeability data based on the identified and verified groundwater dynamic numerical model. It uses numerical model prediction data to supplement the missing or insufficient parts of the actual monitoring data, and uses difference analysis method to quantify the difference in permeability characteristics between faults and normal strata and couple it to the data system.

[0145] The coupled model construction configuration unit is used to construct coupled models using methods including Long Short-Term Memory Neural Network (LSTM), Isolation Forest (iForest), and K-Nearest Neighbor Regression. It also configures feature extraction, rolling prediction, data expansion, label statistics, outlier detection, and residual correction modules. Each module completes data interaction and training through a preset interface.

[0146] The model training dynamic optimization unit is used to divide the preprocessed multivariate time-series training data into training and test sets according to time order, and input them into the constructed coupled machine learning model. First, the LSTM model is trained and hidden time-series features are extracted. Then, three anomaly detection models are trained separately and the learning results are integrated through a weighted fusion strategy. The residual correction model is trained simultaneously. Based on the feedback of water inflow prediction errors from multiple tunnels or mines, the hyperparameters and residual correction weights of each module are adaptively adjusted to complete the dynamic optimization of the model training process.

[0147] The tunnel or mine water inflow prediction result correction unit is used to perform synchronous spatiotemporal dynamic prediction of water inflow for multiple tunnels or mines using the window scrolling strategy of the rolling prediction module and the optimized coupled machine learning model. It outputs the water inflow prediction results for different time steps and different tunnel or mine cross sections. Combined with the outlier detection module to identify abnormal data in the prediction process, the prediction results are corrected in real time by the residual correction module. Finally, the spatiotemporal dynamic prediction data of tunnel or mine water inflow that meets the engineering accuracy requirements is output.

[0148] Example 3

[0149] This embodiment 3 also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement any step of a method for spatiotemporal prediction of water inrush in tunnels or mines coupled with a hydrodynamic numerical model.

[0150] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0151] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0152] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for spatiotemporal prediction of water inrush in tunnels or mines using a coupled hydrodynamic numerical model, characterized in that, include: S1. Based on the identified and verified groundwater dynamic numerical model, output the predicted data related to water inrush in tunnels or mines, groundwater environment data and geological body permeability data. Use the numerical model prediction data to supplement the missing or insufficient parts of the actual monitoring data. Use the difference analysis method to quantify the difference in permeability characteristics between faults and normal strata and couple it to the data system. S2. A coupled model is constructed using methods including Long Short-Term Memory Neural Network (LSTM), Isolation Forest (iForest), and K-Nearest Neighbor Regression. At the same time, modules for feature extraction, rolling prediction, data expansion, label statistics, outlier detection, and residual correction are configured. Each module completes data interaction and training through a preset interface. S3. The preprocessed multivariate time-series training data is divided into training and test sets according to time order and input into the constructed coupled machine learning model. First, the LSTM model is trained and hidden time-series features are extracted. Then, three anomaly detection models are trained separately and the learning results are integrated through a weighted fusion strategy. The residual correction model is trained simultaneously. Based on the feedback of water inflow prediction errors from multiple tunnels or mines, the hyperparameters and residual correction weights of each module are adaptively adjusted to complete the dynamic optimization of the model training process. S4. Using the window scrolling strategy of the scrolling prediction module, the optimized coupled machine learning model is used to perform synchronous spatiotemporal dynamic prediction of water inflow for multiple tunnels or mines, and output the water inflow prediction results for different time steps and different tunnel or mine cross sections; combined with the outlier detection module to identify abnormal data in the prediction process, the prediction results are corrected in real time through the residual correction module, and finally the spatiotemporal dynamic prediction data of tunnel or mine water inflow that meets the engineering accuracy requirements is output. The process of adaptively adjusting the hyperparameters of each module in S3 is as follows: First, a multi-objective error evaluation function is constructed. Based on the prediction results and measured data of multiple tunnels or mines, the root mean square error is calculated. and mean absolute percentage error Construct a comprehensive error index ,in Indicates the first A tunnel or mine, and The inherent weights of the error index are naturally determined by the data distribution characteristics and engineering accuracy requirements. , , , For the first The predicted time steps for a tunnel or mine. For the first The first tunnel or mine The measured inflow rate at the time step Predict the inflow rate for the model at the corresponding time step; Define the set of model hyperparameters ,in This is the hyperparameter vector of the LSTM model, containing the number of hidden layer neurons. Number of iterations Learning rate descent rate, regularization parameter; This is the hyperparameter vector for the Isolation Forest model, containing the number of decision trees. Sample size Coefficient of variation; This is the hyperparameter vector of the K-nearest neighbor regression model, which includes the number of nearest neighbor samples K and the Euclidean distance metric. Define residual correction weight vector , This represents the total number of tunnels or mines. For the first The residual correction weights corresponding to each tunnel or mine shaft satisfy the following conditions: ; The objective function is to minimize the mean of the overall error of all tunnels or mines. The model's hyperparameters are optimized using a grid search method. This grid search method exhaustively traverses all parameter combinations within a preset hyperparameter space, verifies each parameter combination, calculates the corresponding mean comprehensive error, and finally selects the optimal hyperparameter combination that minimizes the mean comprehensive error, thus completing the global optimization of the model's hyperparameters.

2. The method for spatiotemporal prediction of water inrush in tunnels or mines using a coupled hydrodynamic numerical model according to claim 1, characterized in that, The specific method for difference analysis in S1 is as follows: First, it is clear that the core differences between the permeability characteristics of fault zones and normal strata lie in two aspects: the average difference in the overall permeability of the strata, and the difference in the heterogeneity of the internal permeability distribution. To comprehensively characterize these two types of differences, a time series of permeability coefficients of fault zones is extracted based on continuous time step data output from the groundwater numerical model. ,in, The time steps in the numerical simulation represent the global permeability coefficient value of the fault zone at a specific time step for each data point. This value originates from the physical simulation results of the fault zone's lithology and fracture development, as well as the time series of permeability coefficients of normal strata during the same period. Based on the same numerical simulation system as the fault zone data, the consistency of the hydrogeological environment is ensured, and the simulated value of the permeability coefficient of the normal strata is obtained for the whole area. Secondly, calculate the cumulative mean of the permeability coefficients of the fault zone and the normal formation, and the cumulative mean of the permeability coefficient of the fault zone. Cumulative mean of normal formation permeability As a benchmark value for the average permeability of the normal surrounding rock around the fault, and To form a comparison of differences along the mean dimension, i.e. This difference directly reflects the difference in the overall permeability of the two types of strata; Furthermore, considering the impact of the heterogeneity of permeability within the formation on water permeability, it is necessary to quantify the dispersion of permeability coefficients between the fault zone and normal formations, and calculate the variance of the permeability coefficient in the fault zone. Variance of normal formation permeability coefficient To characterize the discrete characteristics of permeability within normal surrounding rock due to differences in weathering degree and rock mass structure, a comprehensive dispersion index is constructed to unify the quantitative dimensions of mean difference and variance difference. This indicator integrates the heterogeneous characteristics of the two types of strata, providing a standardized comparison benchmark for mean differences; Finally, based on the ratio of mean difference to overall dispersion, the expression for the difference quantification model is derived: ;in, The comprehensive difference in permeability coefficients between fault zones and normal formations is calculated by coupling the mean difference with the comprehensive dispersion, thus achieving an objective and dimensionless quantification of the differences in permeability characteristics between the two types of formations.

3. The method for spatiotemporal prediction of water inrush in tunnels or mines using a coupled hydrodynamic numerical model according to claim 1, characterized in that, S1 further includes: identifying abnormal signals from the measured data, and obtaining the comprehensive permeability coefficient of the construction section through weighted calculation based on the surrounding rock classification, weathering degree, and lithological structure. All coupled data are preprocessed using a normalization method adapted to the corresponding module to form multivariate time-series training data.

4. The method for spatiotemporal prediction of water inrush in tunnels or mines using a coupled hydrodynamic numerical model according to claim 3, characterized in that, The comprehensive permeability coefficient The calculation method is as follows: First, define the background permeability coefficient of the rock formation. The parameter is taken from the identified and verified groundwater numerical model, and corresponds to the benchmark value of the permeability of the rock strata at the construction section under the natural geological state without construction disturbance. Its data is consistent with the geological survey report and rock sample test data of the study area. Secondly, the core geological influencing factors of the construction section are selected, and the surrounding rock classification quantification factor is set as follows: The quantification factor for the degree of weathering of rock mass is: The quantification factor for the degree of lithological structure fragmentation is: Each factor is quantified through interval mapping. Different dimensionless quantification values ​​are assigned to different surrounding rock types, weathering grades, and lithological structures. The mapping rule is that the more significant the enhancement effect of geological features on permeability, the larger the corresponding factor value. Introducing a dynamic correction term for geological influence factors, expressed as a product of the corresponding exponents of each geological influence factor, we obtain the generalized calculation expression for the comprehensive permeability coefficient: ;in The number of geological influence factors involved in the calculation. Not less than 3; This represents the natural weight index of each factor; The comprehensive permeability coefficient of the construction section is used to characterize the actual permeability of the construction section under actual geological conditions. The background permeability coefficient of the rock strata is derived from the physical simulation results of the groundwater numerical model; For the first The quantitative values ​​of each geological influence factor are obtained based on the quantification of objective geological characteristics, ensuring the standardization of geological characteristic representation.

5. The method for spatiotemporal prediction of water inrush in tunnels or mines using a coupled hydrodynamic numerical model according to claim 1, characterized in that, The three anomaly detection models in S3 are the Long Short-Term Memory Neural Network Model, the Isolation Forest Model, and the K-Nearest Neighbor Regression Model. Based on the design logic of functional complementarity and synergistic coupling, they each undertake different anomaly identification tasks: The Long Short-Term Memory Neural Network Model relies on its ability to capture long-term dependencies in time-series data. Starting from the time-series evolution of tunnel or mine water inflow data, it identifies trend anomalies caused by dynamic changes in hydrogeological conditions and continuous disturbances during construction. By mining the implicit correlations in the time-series characteristics of the data, it locates the continuous abnormal changes in tunnel or mine water inflow over time. The isolated forest model is based on the principle of random partitioning. It focuses on the global discrete features in the water inflow data of tunnels or mines. It specifically identifies globally isolated anomalous data points caused by factors such as changes in hydrological environmental conditions at the excavation face and sudden changes in geological structure. By randomly cutting and partitioning the data space and then evaluating the discreteness of the data, it captures anomalous information features that deviate significantly from the distribution patterns of surrounding data. The K-nearest neighbor regression model identifies localized abnormal water inrush events caused by factors such as changes in tunnel or mine construction plans and adjustments due to momentary equipment failures, based on the statistical regularity of local data. By defining local data, it filters out abnormal information in this part of the data.

6. The method for spatiotemporal prediction of water inrush in tunnels or mines using a coupled hydrodynamic numerical model according to claim 1, characterized in that, The weighted fusion strategy in S3 is specifically a linear weighted strategy: The weighting coefficients for the detection results of the three types of models are defined as follows: The weighting coefficients are determined based on the anomaly detection accuracy and recall of various models on the validation set, with higher weights assigned to models with better detection performance.

7. The method for spatiotemporal prediction of water inrush in tunnels or mines using a coupled hydrodynamic numerical model according to claim 1, characterized in that, The correction process for the residual correction weights in S3 is as follows: The residual correction weights are dynamically allocated based on the prediction accuracy of each tunnel or mine. This weight allocation is implemented using a conditional mapping function, which first calculates the prediction accuracy quality score for each tunnel or mine. The calculation formula is: In the formula, The coefficient of determination is derived from the fitting analysis of the predicted value and the measured water inflow data of a single tunnel or mine. The correlation coefficient between the predicted and measured values. MAE improvement rate, quality score A higher value indicates better prediction accuracy after residual correction for the corresponding tunnel or mine; for short tunnels with insufficient data samples, the calculated quality score is... Deductions and corrections are made, resulting in a revised quality score. Based on the revised quality score Five levels of prediction accuracy evaluation criteria were defined, and the corresponding residual correction weights were directly matched through a conditional mapping function. The mapping relationship is as follows: The corresponding evaluation levels are, in order: excellent, good, average, poor, and very poor.

8. A spatiotemporal prediction system for water inrush in tunnels or mines coupled with a hydrodynamic numerical model, characterized in that, include: The multi-source data coupling preprocessing unit is used to output predicted data related to water inrush in tunnels or mines, groundwater environment data, and geological body permeability data based on the identified and verified groundwater dynamic numerical model. It uses numerical model prediction data to supplement the missing or insufficient parts of the actual monitoring data, and uses difference analysis method to quantify the difference in permeability characteristics between faults and normal strata and couple it to the data system. The coupled model construction configuration unit is used to construct coupled models using methods including Long Short-Term Memory Neural Network (LSTM), Isolation Forest (iForest), and K-Nearest Neighbor Regression. It also configures feature extraction, rolling prediction, data expansion, label statistics, outlier detection, and residual correction modules. Each module completes data interaction and training through a preset interface. The model training dynamic optimization unit is used to divide the preprocessed multivariate time-series training data into training and test sets according to time order, and input them into the constructed coupled machine learning model. First, the LSTM model is trained and hidden time-series features are extracted. Then, three anomaly detection models are trained separately and the learning results are integrated through a weighted fusion strategy. The residual correction model is trained simultaneously. Based on the feedback of water inflow prediction errors from multiple tunnels or mines, the hyperparameters and residual correction weights of each module are adaptively adjusted to complete the dynamic optimization of the model training process. The tunnel or mine water inflow prediction result correction unit is used to perform synchronous spatiotemporal dynamic prediction of water inflow for multiple tunnels or mines using an optimized coupled machine learning model through the window scrolling strategy of the rolling prediction module. It outputs water inflow prediction results for different time steps and different tunnel or mine cross-sections. Combined with an outlier detection module to identify abnormal data during the prediction process, the unit uses a residual correction module to correct the prediction results in real time, ultimately outputting spatiotemporal dynamic prediction data of tunnel or mine water inflow that meets engineering accuracy requirements. The process of adaptively adjusting the hyperparameters of each module in the model training dynamic optimization unit is as follows: First, a multi-objective error evaluation function is constructed. Based on the prediction results and measured data of multiple tunnels or mines, the root mean square error is calculated. and mean absolute percentage error Construct a comprehensive error index ,in Indicates the first A tunnel or mine, and The inherent weights of the error index are naturally determined by the data distribution characteristics and engineering accuracy requirements. , , , For the first The predicted time steps for a tunnel or mine. For the first The first tunnel or mine The measured inflow rate at the time step Predict the inflow rate for the model at the corresponding time step; Define the set of model hyperparameters ,in This is the hyperparameter vector of the LSTM model, containing the number of hidden layer neurons. Number of iterations Learning rate descent rate, regularization parameter; This is the hyperparameter vector for the Isolation Forest model, containing the number of decision trees. Sample size Coefficient of variation; This is the hyperparameter vector of the K-nearest neighbor regression model, which includes the number of nearest neighbor samples K and the Euclidean distance metric. Define residual correction weight vector , This represents the total number of tunnels or mines. For the first The residual correction weights corresponding to each tunnel or mine shaft satisfy the following conditions: ; The objective function is to minimize the mean of the overall error of all tunnels or mines. The model's hyperparameters are optimized using a grid search method. This grid search method exhaustively traverses all parameter combinations within a preset hyperparameter space, verifies each parameter combination, calculates the corresponding mean comprehensive error, and finally selects the optimal hyperparameter combination that minimizes the mean comprehensive error, thus completing the global optimization of the model's hyperparameters.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as described in any one of claims 1-7: a method for spatiotemporal prediction of water inrush in tunnels or mines using a coupled hydrodynamic numerical model.

Images