A method for dynamic prediction of tunnel water inflow

Abstract

This invention discloses a dynamic prediction method for tunnel water inflow, belonging to the fields of tunnel engineering and hydrogeology. The method first constructs an initial hydrogeological conceptual model; then establishes a dynamic database integrating geological, geophysical, monitoring, and meteorological data; furthermore, it establishes a dynamically coupled prediction model that combines a physical numerical model and a machine learning proxy model; utilizing newly revealed data during construction, it dynamically inverts and updates model parameters through a data assimilation algorithm; the updated model is then used for probabilistic water inflow prediction and risk level classification; finally, the results are visualized and a feedback loop is formed. This invention, through a dynamic update mechanism, enables the prediction model to continuously approximate real geological conditions, solving the problems of low accuracy and poor adaptability of traditional static prediction methods, and significantly improving the ability of tunnel construction to cope with water inflow risks.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering and hydrogeology, and specifically relates to a method for dynamic prediction of tunnel water inflow. Background Technology

[0002] Water inrush in tunnels is one of the most significant geological hazards during tunnel construction. Accurately predicting the water inrush volume is crucial for ensuring construction safety, optimizing support design, and reducing engineering risks.

[0003] Existing methods for predicting tunnel water inflow can be mainly categorized as follows: Analytical methods: such as the Oshima Hiroshi method and the Goodman empirical formula. These methods are based on simplified hydrogeological models, treating the aquifer as a homogeneous, isotropic, infinitely extending body. While the calculation formulas are simple, they neglect the complexity and spatial variability of actual geological conditions, resulting in generally low prediction accuracy and difficulty in reflecting the hydrogeological characteristics of unknown sections ahead of the tunnel face. Numerical simulation methods: such as the finite element method and the finite difference method. These methods can handle complex geological boundaries and aquifer structures, theoretically offering high accuracy. However, their modeling process is complex and heavily reliant on prior geological survey data, which is often limited and inaccurate. Furthermore, the determination of numerical model parameters (such as permeability coefficient and specific yield) is subject to significant human error, and once established, the model is relatively fixed, making rapid and dynamic updates and corrections based on new geological information revealed during construction difficult. This leads to poor applicability of the prediction results in long tunnel construction. Analogy methods: predicting the water inflow of new tunnels based on existing tunnel water inflow data. This method is highly empirical, greatly limited by geographical location and engineering geological conditions, and lacks universality.

[0004] In summary, the existing technologies generally have the following defects: (1) Static nature: the prediction model cannot be dynamically updated with tunnel excavation, and the reliability of prediction for unknown sections ahead decreases over time; (2) Insufficient data utilization: it fails to effectively integrate multi-source heterogeneous data such as geological exploration, geophysical exploration, and real-time monitoring; (3) Poor model adaptability: it is difficult to accurately describe the heterogeneous and anisotropic characteristics of water-conducting channels such as rock fractures and fault fracture zones.

[0005] Therefore, there is an urgent need for a method for predicting water inflow that can dynamically adapt, learn on its own, and whose accuracy continuously improves as the project progresses. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a method for dynamic prediction of tunnel water inflow. This method can integrate multi-source data and dynamically update the prediction model as the tunnel is excavated, thereby achieving high-precision and dynamic prediction of water inflow in the unexcavated section in front of the tunnel face.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a method for dynamically predicting tunnel water inflow, comprising the following steps: S1. Constructing an initial hydrogeological conceptual model: Based on geological survey data, borehole data, and geophysical exploration data of the tunnel route, establish a three-dimensional hydrogeological conceptual model that includes stratigraphic lithology, geological structure, and initial groundwater flow field. S2. Construct a multi-source data fusion and dynamic database: Create a dynamic database to store and manage multi-source heterogeneous data from geological sketches, advanced geological forecasts, monitoring points at the working face and inside the tunnel, as well as meteorological and hydrological stations; S3. Establish a dynamic coupled prediction model: Construct a hybrid model architecture that integrates physical mechanisms and data-driven algorithms. This architecture includes a distributed hydrogeological numerical model based on the partial differential equation of groundwater flow, and a machine learning proxy model for real-time correction and rapid prediction. S4. Dynamic inversion and update of model parameters: During tunnel excavation, the latest geological information of the tunnel face and water inrush monitoring data are used as observation constraints. The particle swarm optimization algorithm is used to invert and update the key hydrogeological parameter fields of the hydrogeological numerical model. S5. Conduct dynamic prediction and risk assessment of water inflow: Use the inverted and updated dynamic coupled prediction model to simulate and predict the water inflow in the unexcavated section in front of the tunnel face, and classify the water inflow risk level based on the probability distribution of the prediction results. S6. Visualization and Feedback Optimization of Prediction Results: The risk assessment results are rendered and displayed on a 3D visualization platform, and the subsequent actual water inrush data are fed back to the dynamic database as new samples to initiate a new round of model optimization.

[0008] Furthermore, in step S2, the dynamic database is implemented as follows: A dynamic database is constructed using a relational database management system or a time-series database system, and a unified data structure, spatial coordinate index, and timestamp are defined for all data entering the database. The geological sketch data is structured and coded to convert qualitative descriptions of lithology identification, joint occurrence, fracture development density, and karst infill characteristics into quantitative or semi-quantitative index data that characterize the permeability and water-bearing capacity of the rock mass. For advanced geological prediction data, the original geophysical signal data and the spatial geometric information of the interface of the unfavorable geological body and the conclusions on its water-bearing properties, which are obtained after professional interpretation, are stored at the same time. An automatic data acquisition and wireless transmission network was established for monitoring points at the working face and inside the tunnel, as well as for meteorological and hydrological stations, to achieve high-frequency and automated data collection and exchange. A data quality control module was also set up to remove transmission noise and abnormal equipment values. The dynamic database provides a standardized application programming interface (API) that allows dynamically coupled predictive models and data assimilation algorithms to call historical data sequences and real-time data streams on demand.

[0009] Furthermore, the coupling mechanism between the machine learning surrogate model and the hydrogeological numerical model is as follows: The machine learning proxy model adopts a long short-term memory network architecture. Its input feature vector is constructed through feature engineering and includes at least: the current three-dimensional spatial coordinates of the working face, the distance of the interface of the adverse geological body in front of the working face and the intensity of the abnormal physical property parameters based on the interpretation of advanced geological forecast data, the probability distribution of lithology in front based on the initial hydrogeological conceptual model, the statistical characteristic value of the permeability field updated by inversion in the predicted influence domain, and the effective rainfall sequence of the previous period considering the lag effect. The operation process of the coupling mechanism is as follows: In each prediction period, a forward simulation is first performed using the hydrogeological numerical model under the current parameter state to obtain the baseline water inflow prediction value; then, the input feature vector is input into the pre-trained long short-term memory network proxy model, and the long short-term memory network proxy model outputs the nonlinear correction amount for the baseline prediction value, or directly outputs the final water inflow prediction value after fusion.

[0010] Furthermore, in step S4, the data assimilation algorithm is an ensemble Kalman filter algorithm, and its specific execution process includes: The key parameter field of the hydrogeological numerical model is defined as the state vector of the ensemble Kalman filter algorithm. The key parameter field includes at least the permeability tensor and specific yield of each rock layer, and the parameter set is initialized to characterize its prior uncertainty. At each assimilation moment, the time series of water inflow at the working face collected within the predetermined time window and the water level change sequence of the monitoring well are combined to form an observation vector; The prediction step of the Kalman filter algorithm is executed, and the evolution of all parameter set members in the state vector is advanced through the hydrogeological numerical model. The update step of the Kalman filter algorithm is executed, and the residual between the actual measured value and the model prediction value of the observed vector is used to update the state vector, thereby obtaining an updated set of parameter fields with reduced uncertainty. The mean of this set is used as the optimal parameter estimate for the next prediction step.

[0011] Furthermore, in step S5, the probability distribution is generated as follows: The dynamic coupling prediction model achieves probabilistic prediction through the Monte Carlo simulation framework. Specifically, it is based on the parameter field set updated by the Kalman filter algorithm. Each member of the parameter set drives the hydrogeological numerical model or machine learning proxy model to perform a water inflow prediction to obtain a set of predicted values. Kernel density estimation is performed on the set of predicted values ​​to fit the water inflow rate in a specific section ahead of the tunnel face. probability density function : In the formula, Let be the number of members in the predicted value set. For the first The predicted value of each member, For kernel function, For bandwidth parameters; The cumulative distribution function is calculated based on the probability density function to determine the threshold range of water inflow corresponding to different risk levels.

[0012] Furthermore, the training and updating of the Long Short-Term Memory (LSTM) network agent model adopts a two-stage strategy: Before the tunnel project starts, offline pre-training is carried out: using the initial hydrogeological conceptual model, combined with parameter sensitivity analysis and Latin hypercube sampling technology, a large-scale training sample set covering different hydrogeological scenarios is generated in the possible value space of key parameters. The sample set contains input feature vectors and corresponding water inflow labels. The sample set is used to initially train the long short-term memory network proxy model to establish the initial nonlinear mapping relationship between input and output. During tunnel construction, online fine-tuning is carried out: when new monitoring data accumulates to a preset scale, it is combined with the corresponding real-time input feature vector to form an incremental training sample set. The weight parameters of the long short-term memory network proxy model are fine-tuned with a learning rate lower than that in the pre-training stage to adapt it to the local hydrogeological conditions actually exposed by the tunnel.

[0013] Furthermore, in step S4, the construction and optimization process of the particle swarm optimization algorithm is as follows: Define the objective function The error norm between the simulation output of the hydrogeological numerical model and the actual monitoring data for the corresponding time period: In the formula, The vector of hydrogeological parameters to be inverted. The length of the data time window used for inversion. and They are time points Simulated and observed water inflow rates and They are time points Simulated water level and observed water level and These are the weighting coefficients for the water inflow and water level data, respectively. Initialize a group of particles, where the position of each particle represents a vector of potential hydrogeological parameters. By iteratively updating the velocity and position of the particles, the particle swarm is made to move towards the objective function. Minimize the region movement to ultimately obtain the optimal parameter estimate. .

[0014] Furthermore, in step S5, the risk level of water inrush is determined. The division rules are as follows: In the formula, The tunnel is designed with drainage capacity. Low risk Medium risk. High risk, and To be a threshold constant preset according to engineering safety criteria, As a risk index, To predict the inflow The probability density function.

[0015] Furthermore, the specific implementation method of step S6 is as follows: By integrating the 3D visualization platform with the tunnel construction information model or digital twin system, the risk level R is rendered in the form of 3D isosurfaces with different colors and transparency in the 3D spatial model of the unexcavated section in front of the tunnel face. Automatically generate and output the water inflow prediction curve and risk level profile along the tunnel axis, with the start and end mileage of high-risk sections marked on the map. The feedback optimization process is as follows: when the tunnel excavation exposes the predicted section, the actual water inflow data recorded by the monitoring equipment deployed in that section is automatically captured by the system and correlated and compared with the historical prediction records stored in the dynamic database. Based on the comparison results, a model accuracy evaluation report is generated, and the actual water inflow data is used as a new training sample with real labels and stored in the dynamic database. This triggers the re-execution of the dynamic inversion and update steps of model parameters, as well as the online fine-tuning of the machine learning agent model.

[0016] The beneficial effects of this invention are as follows: 1. Dynamism and Adaptability: By treating the construction process as a "dynamic exploration" process, the model parameters are continuously inverted and updated using real-time monitoring data, enabling the prediction model to continuously approximate the actual hydrogeological conditions, which significantly improves the accuracy and reliability of the prediction of the unexcavated section ahead.

[0017] 2. Deep integration of multi-source data: It breaks down the barriers between geological, geophysical and monitoring data, and manages and integrates them through a unified data platform, making the most of all available information and reducing the uncertainty of a single data source.

[0018] 3. Advantages of Model Coupling: It combines the advantages of a well-defined physical model and the computational efficiency and ability to handle nonlinear relationships of a machine learning model. The physical model provides a reliable physical foundation and extrapolation capabilities, while the machine learning surrogate model is responsible for rapid correction and deviation compensation. The two work together to improve computational efficiency while ensuring accuracy, meeting the real-time requirements of construction.

[0019] 4. High practicality: The final prediction results are presented in intuitive visual graphics and clearly defined risk levels, making them easy for on-site engineers to understand and apply. This provides direct and scientific decision support for advanced support design and drainage scheme development, effectively ensuring tunnel construction safety.

[0020] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0021] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0022] like Figure 1 As shown, the present invention provides a method for dynamic prediction of tunnel water inflow, comprising the following steps: S1. Constructing an initial hydrogeological conceptual model: Based on geological survey data, borehole data, and geophysical exploration data of the tunnel route, establish a three-dimensional hydrogeological conceptual model that includes stratigraphic lithology, geological structure, and initial groundwater flow field. S2. Construct a multi-source data fusion and dynamic database: Create a dynamic database to store and manage multi-source heterogeneous data from geological sketches, advanced geological forecasts, monitoring points at the working face and inside tunnels, and meteorological and hydrological stations; S3. Establish a dynamic coupled prediction model: Construct a hybrid model architecture that integrates physical mechanisms and data-driven algorithms. This architecture includes a distributed hydrogeological numerical model based on the partial differential equation of groundwater flow, and a machine learning proxy model for real-time correction and rapid prediction. S4. Dynamic inversion and update of model parameters: During tunnel excavation, the latest geological information of the tunnel face and water inrush monitoring data are used as observation constraints. The particle swarm optimization algorithm is used to invert and update the key hydrogeological parameter fields of the hydrogeological numerical model. S5. Conduct dynamic prediction and risk assessment of water inflow: Use the inverted and updated dynamic coupled prediction model to simulate and predict the water inflow in the unexcavated section in front of the tunnel face, and classify the water inflow risk level based on the probability distribution of the prediction results. S6. Visualization and Feedback Optimization of Prediction Results: The risk assessment results are rendered and displayed on a 3D visualization platform, and the subsequent actual water inrush data are fed back to the dynamic database as new samples to initiate a new round of model optimization.

[0023] In this approach, the parameters of the physical model are reverse-calibrated using the latest revealed data (such as actual water inflow) through data assimilation or optimization algorithms to make them closer to real geological conditions. Then, the calibrated model is used to predict the water inflow ahead and provide a risk probability rather than a single, fixed value. Finally, all prediction results are visualized in three dimensions, and subsequent actual data is fed back into the system to initiate the next round of model optimization, forming a closed loop of "prediction-validation-learning-re-prediction".

[0024] This solution overcomes the drawbacks of traditional static models that offer only a "one-and-done" solution, enabling the prediction model to evolve as tunnel excavation progresses, with accuracy continuously improving over time. It combines the clear mechanisms of physical models with the computational efficiency and ability to handle complex patterns of AI models (machine learning proxy models), overcoming the shortcomings of inaccurate parameters in pure physical models and poor extrapolation in pure AI models. It provides probabilistic predictions and risk assessments, allowing decision-makers to understand the uncertainties in the predictions and thus formulate more scientific and prepared responses. Through continuous feedback, the system possesses the ability to learn and continuously improve, making the prediction system increasingly "intelligent" with use, which is particularly beneficial for long tunnel projects.

[0025] In one embodiment of the present invention, in step S2, the dynamic database is implemented as follows: A dynamic database is constructed using a relational database management system or a time-series database system, and a unified data structure, spatial coordinate index, and timestamp are defined for all data entering the database. The geological sketch data is structured and coded to convert qualitative descriptions of lithology identification, joint occurrence, fracture development density, and karst infill characteristics into quantitative or semi-quantitative index data that characterize the permeability and water-bearing capacity of the rock mass. For advanced geological prediction data, the original geophysical signal data and the spatial geometric information of the interface of the unfavorable geological body and the conclusions on its water-bearing properties, which are obtained after professional interpretation, are stored at the same time. An automatic data acquisition and wireless transmission network was established for monitoring data of water inflow points at the tunnel face and data from long-term monitoring boreholes inside the tunnel, so as to achieve high-frequency and automated data exchange. A data quality control module was also set up to remove transmission noise and abnormal equipment values. The dynamic database provides a standardized application programming interface (API) that allows dynamically coupled predictive models and data assimilation algorithms to call historical data sequences and real-time data streams on demand.

[0026] In this solution, data from different sources and in different formats, such as geology, geophysics, monitoring, and meteorology, are unified on a single platform. Through the quantitative processing of qualitative data, expert experience that is difficult to use directly is transformed into feature values ​​that the model can recognize, which greatly enriches the model input information. The automatic acquisition and quality control modules reduce human error and delays, ensuring that the data used for model updates and predictions are accurate and timely. The standardized interface decouples the data layer from the model calculation layer, facilitating the independent development and maintenance of each module of the system and improving the engineering application capabilities of the entire system.

[0027] In one embodiment of the present invention, the coupling mechanism between the machine learning proxy model and the hydrogeological numerical model is specifically as follows: The machine learning proxy model adopts a long short-term memory network (LSTM) architecture. Its input feature vector is constructed through feature engineering and includes at least: the current three-dimensional spatial coordinates of the face, the distance of the interface of the adverse geological body in front of the face and the intensity of the abnormal physical property parameters based on the interpretation of advanced geological forecast data, the probability distribution of lithology in front based on the initial hydrogeological conceptual model, the statistical characteristic value of the permeability field updated by inversion in the predicted influence domain, and the effective rainfall sequence of the previous period considering the lag effect. The operation process of the coupling mechanism is as follows: In each prediction period, a forward simulation is first performed using the hydrogeological numerical model under the current parameter state to obtain the baseline water inflow prediction value; then, the input feature vector is input into the pre-trained long short-term memory network proxy model, which outputs a nonlinear correction amount for the baseline prediction value, or directly outputs the fused final water inflow prediction value.

[0028] In this scheme, a hydrogeological numerical model (such as a MODFLOW-based numerical model) is responsible for simulating the basic physical laws of groundwater flow, but it is slow to compute and its parameters may be biased. A machine learning surrogate model acts as an "intelligent corrector." Its input consists of carefully designed features reflecting various geological and hydrological conditions (such as location, geophysical anomalies, and inverted parameter field characteristics). During prediction, the physical model first provides a baseline prediction. Then, these features are fed into a pre-trained LSTM model, which will output a correction value. Or a final fusion prediction value .Right now Or directly output from LSTM This coupling is equivalent to having a "physical expert" based on physical laws work together with an "AI assistant" adept at finding patterns in data. The LSTM model can learn complex local nonlinear relationships (such as the influence of micro-fracture networks) that hydrogeological numerical models fail to accurately describe, and compensate for the systematic biases of the physical model. Once the LSTM model is trained, its forward prediction speed is extremely fast, meeting the needs of real-time prediction in construction, without having to perform a full simulation of the computationally expensive physical model every time. The AI ​​model, which combines physical priors, has stronger reasoning and generalization capabilities than a purely data-driven AI model when faced with unseen geological conditions.

[0029] In one embodiment of the present invention, in step S4, the data assimilation algorithm is the ensemble Kalman filter (EnKF) algorithm, and its specific execution process includes: The key parameter field of the hydrogeological numerical model is defined as the state vector of the ensemble Kalman filter algorithm. The key parameter field includes at least the permeability tensor and specific yield of each rock layer, and the parameter set is initialized to characterize its prior uncertainty. At each assimilation moment, the time series of water inflow at the working face collected within the predetermined time window and the water level change sequence of the monitoring well are combined to form an observation vector; The prediction step of the Kalman filter algorithm is executed, and the evolution of all parameter set members in the state vector is advanced through the hydrogeological numerical model. The update step of the Kalman filter algorithm is executed, and the residual between the actual measured value and the model prediction value of the observed vector is used to update the state vector, thereby obtaining an updated set of parameter fields with reduced uncertainty. The mean of this set is used as the optimal parameter estimate for the next prediction step.

[0030] In this scheme, the ensemble Kalman filter algorithm characterizes our uncertainty about the actual geological parameters by maintaining a "parameter set." Each member of the set is a set of possible hydrogeological parameters. During tunnel excavation, each step is divided into a prediction step and an update step; among them, Prediction step: Drive the hydrogeological model with all current parameter set members to simulate and obtain a set of predicted inflow and water level.

[0031] Update step: Compare this set of predicted values ​​with the actual monitored values. The EnKF algorithm intelligently adjusts the values ​​of each member of the parameter set based on the comparison results, giving greater weight to members whose predicted results are closer to the actual values. Ultimately, the mean of the updated parameter set is the best estimate of the true parameters.

[0032] This solution requires no manual intervention. The system can automatically utilize a continuous stream of new data to optimize the internal state of the model, ensuring that the model always maintains optimal performance. The distribution of the parameter set directly reflects the uncertainty of the parameters after inversion, which provides a source for subsequent probabilistic predictions. Compared with traditional variational assimilation methods, the EnKF algorithm has better robustness and practicality for nonlinear problems such as groundwater simulation.

[0033] In one embodiment of the present invention, the method for generating the probability distribution in step S5 is as follows: The dynamic coupling prediction model achieves probabilistic prediction through the Monte Carlo simulation framework. Specifically, it is based on the parameter field set updated by the Kalman filter algorithm. Each member of the parameter set drives the hydrogeological numerical model or machine learning proxy model to perform a water inflow prediction to obtain a set of predicted values. Kernel density estimation is performed on the set of predicted values ​​to fit the water inflow rate in a specific section ahead of the tunnel face. probability density function Its mathematical expression is: In the formula, Let be the number of members in the predicted value set. For the first The predicted value of each member, For kernel function, For bandwidth parameters; The cumulative distribution function is calculated based on the probability density function, and the threshold range of water inflow corresponding to different risk levels is determined accordingly.

[0034] In this scheme, the prediction is driven by a "parameter set" (where each member represents a possible geological condition) updated using the EnKF algorithm. Specifically, by running the coupled prediction model with each of these dozens or hundreds of parameter set members, dozens or hundreds of different predicted inflow values ​​are obtained, forming a "predicted value set." This set is essentially a sample of possible future inflow values. Then, kernel density estimation (KDE) is used to smooth these samples and fit a complete probability density curve (PDF). This curve can be used to calculate the probability that the inflow will exceed a certain danger level.

[0035] This design allows decision-makers to know not only the "most likely" inflow volume, but also the "maximum" likely inflow volume and the probability of exceeding drainage capacity. This is something that a single deterministic prediction cannot provide, enabling the formulation of construction plans (such as drainage equipment configuration and support scheme selection) to shift from "experience-driven" to "data-driven, risk-quantified" scientific decision-making. The probability distribution format is easier for engineers to understand and accept than complex model outputs.

[0036] In one embodiment of the present invention, the training and updating of the Long Short-Term Memory (LSTM) network agent model adopts a two-stage strategy: Before or in the early stages of tunnel construction, offline pre-training is conducted: using an initial hydrogeological conceptual model, combined with parameter sensitivity analysis and Latin hypercube sampling technology, a large-scale training sample set covering different hydrogeological scenarios is generated in the possible value space of key parameters. The sample set contains input feature vectors and corresponding water inflow labels. The sample set is used to initially train the long short-term memory network proxy model to establish an initial nonlinear mapping relationship between input and output. During tunnel construction, online fine-tuning is carried out: when new monitoring data accumulates to a preset scale, it is combined with the corresponding real-time input feature vector to form an incremental training sample set. The weight parameters of the long short-term memory network proxy model are fine-tuned with a learning rate lower than that in the pre-training stage to adapt it to the local hydrogeological conditions actually exposed by the tunnel.

[0037] In this approach, offline pre-training is performed: Before the project begins, existing geological knowledge and parameter ranges are used to generate thousands of "virtual" geological scenarios and corresponding water inflows in batches via computer. This large-scale, wide-coverage dataset is then used to train the LSTM for the first time, allowing it to learn the basic response patterns of water inflows.

[0038] Online fine-tuning: During construction, once sufficient real-world data has been accumulated, this data is used to refine the already well-established LSTM model. Because the learning rate is set very low, this fine-tuning does not destroy previously learned general knowledge; it simply adjusts the model's parameters to better adapt to the specific tunnel. In the early stages of the project, when real-world data is scarce, this approach provides an initially usable and relatively accurate surrogate model through pre-training. The fine-tuning strategy allows the model to adapt to the specificities of the actual work site while avoiding overfitting or performance degradation due to insufficient data. Compared to training from scratch, fine-tuning based on a pre-trained model converges faster and requires less real-world data.

[0039] In one embodiment of the present invention, in step S4, the construction and optimization process of the particle swarm optimization algorithm (PSO) is as follows: Define the objective function The error norm between the simulation output of the hydrogeological numerical model and the actual monitoring data for the corresponding time period: In the formula, The vector of hydrogeological parameters to be inverted. The length of the data time window used for inversion. and They are time points Simulated and observed water inflow rates and They are time points Simulated water level and observed water level and These are the weighting coefficients for the water inflow and water level data, respectively. Initialize a group of particles, where the position of each particle represents a vector of potential hydrogeological parameters. By iteratively updating the velocity and position of the particles, the particle swarm is made to move towards the objective function. Minimize the region movement to ultimately obtain the optimal parameter estimate. .

[0040] In this scheme, the particle swarm optimization algorithm transforms the parameter inversion problem into an optimization problem: finding a set of parameters such that the error between the model's simulated output and the measured data (i.e., the objective function) is equal to the error between the simulated output and the measured data. The objective function (PSO) simulates the foraging behavior of flocks of birds, where each "particle" represents a set of possible parameters. The particles fly within the parameter space, constantly adjusting their flight direction (i.e., the direction of parameter change) through communication and self-memory. Ultimately, the entire flock gathers in the region where the objective function value is minimized, thus finding the optimal parameters.

[0041] This scheme employs the PSO algorithm, which is less prone to getting trapped in local optima and more likely to find the globally optimal parameter combination, especially when the parameter space is very complex. It is suitable for commercial or legacy numerical simulation software that cannot be directly embedded with sequential assimilation algorithms such as EnKF; it can be called simply as a "black box." Compared to some traditional optimization algorithms (such as genetic algorithms), PSO often converges to a satisfactory solution with fewer iterations.

[0042] In one embodiment of the present invention, in step S5, the risk level of water inrush is determined. The division rules are as follows: In the formula, The tunnel is designed with drainage capacity. Low risk Medium risk. High risk, and To be a threshold constant preset according to engineering safety criteria, As a risk index, To predict the inflow The probability density function.

[0043] In this solution, the prediction system is deeply integrated with the tunnel's digital twin model. Predicted risk zones are no longer abstract numbers on blueprints, but are visually represented in the 3D model as bright red "danger zones." The system automatically generates cross-sectional diagrams to guide construction. Most importantly, when the tunneling machine reaches a previously predicted area, the system automatically compares the predicted value with the actual value and immediately stores the actual value as the "standard answer" in the database, triggering a new round of learning and updating of the model.

[0044] This plan mathematically quantifies the risk level classification, considering not only whether the predicted inflow exceeds the design value, but also the "probability of exceeding" and the "severity of exceeding." Risk Index ( In mathematical terms, this equals the probability density curve, exceeding the designed drainage capacity. The product of the area of ​​that "tail" and the excess amount. Simply put, = Probability of exceeding the limit × Average amount exceeding the limit. For example, a scenario with a "high probability of slightly exceeding the limit" and a scenario with a "low probability of significantly exceeding the limit" might yield similar results. The value can be determined, but traditional methods may not be effective in distinguishing them. By setting a preset threshold... and It can make continuous The values ​​are mapped to three discrete risk levels: "low", "medium", and "high".

[0045] This plan unifies the "probability" and "consequences" of risk into a single indicator, avoiding the one-sidedness of evaluation based on a single dimension. This enables construction companies to precisely allocate different levels of response resources and measures according to different risk levels, achieving a balance between safety and economy.

[0046] In one embodiment of the present invention, step S6 is specifically implemented as follows: The three-dimensional visualization platform is integrated with the tunnel construction information model or digital twin system, and the risk level R is rendered in the form of three-dimensional isosurfaces with different colors and transparency in the three-dimensional spatial model of the unexcavated section in front of the tunnel face. Automatically generate and output the water inflow prediction curve and risk level profile along the tunnel axis, with the start and end mileage of high-risk sections marked on the map. The feedback optimization process is as follows: when the tunnel excavation exposes the predicted section, the actual water inflow data recorded by the monitoring equipment deployed in that section is automatically captured by the system and correlated and compared with the historical prediction records stored in the dynamic database. The comparison results generate a model accuracy evaluation report on the one hand, and store the actual water inflow data as new training samples with real labels in the dynamic database on the other hand, triggering the re-execution of the dynamic inversion and update steps of model parameters, as well as the online fine-tuning of the LSTM proxy model.

[0047] This solution presents complex professional prediction results in the most intuitive graphical way, greatly improving the efficiency of information transmission and lowering the barrier to entry. It achieves a fully automated closed loop: from data collection, model updates, and prediction to result verification and feedback, the entire process is automated as much as possible, reducing human intervention and improving the system's intelligence and response speed. It constructs a continuously evolving digital twin: making the tunnel's digital model no longer a static blueprint, but a "living" twin that can evolve synchronously with the physical tunnel and continuously learn, providing core support for intelligent construction and operation of tunnels throughout their entire lifecycle.

[0048] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A method for dynamic prediction of tunnel water inflow, characterized in that, Includes the following steps: S1. Constructing an initial hydrogeological conceptual model: Based on geological survey data, borehole data, and geophysical exploration data of the tunnel route, establish a three-dimensional hydrogeological conceptual model that includes stratigraphic lithology, geological structure, and initial groundwater flow field. S2. Construct a multi-source data fusion and dynamic database: Create a dynamic database to store and manage multi-source heterogeneous data from geological sketches, advanced geological forecasts, monitoring points at the working face and inside the tunnel, as well as meteorological and hydrological stations; S3. Establish a dynamic coupled prediction model: Construct a hybrid model architecture that integrates physical mechanisms and data-driven algorithms. This architecture includes a distributed hydrogeological numerical model based on the partial differential equation of groundwater flow, and a machine learning proxy model for real-time correction and rapid prediction. S4. Dynamic inversion and update of model parameters: During tunnel excavation, the latest geological information of the tunnel face and water inrush monitoring data are used as observation constraints. The particle swarm optimization algorithm is used to invert and update the key hydrogeological parameter fields of the hydrogeological numerical model. S5. Conduct dynamic prediction and risk assessment of water inflow: Use the inverted and updated dynamic coupled prediction model to simulate and predict the water inflow in the unexcavated section in front of the tunnel face, and classify the water inflow risk level based on the probability distribution of the prediction results. S6. Visualization and Feedback Optimization of Prediction Results: The risk assessment results are rendered and displayed on a 3D visualization platform, and the subsequent actual water inrush data are fed back to the dynamic database as new samples to initiate a new round of model optimization.

2. The method for dynamic prediction of tunnel water inflow according to claim 1, characterized in that: In step S2, the dynamic database is implemented as follows: A dynamic database is constructed using a relational database management system or a time-series database system, and a unified data structure, spatial coordinate index, and timestamp are defined for all data entering the database. The geological sketch data is structured and coded to convert qualitative descriptions of lithology identification, joint occurrence, fracture development density, and karst infill characteristics into quantitative or semi-quantitative index data that characterize the permeability and water-bearing capacity of the rock mass. For advanced geological prediction data, the original geophysical signal data and the spatial geometric information of the interface of the adverse geological body and the conclusions of its water-bearing inference are stored simultaneously. An automatic data acquisition and wireless transmission network was established for monitoring points at the working face and inside the tunnel, as well as for meteorological and hydrological stations, to achieve high-frequency and automated data collection and exchange. A data quality control module was also set up to remove transmission noise and abnormal equipment values. The dynamic database provides a standardized application programming interface (API) that allows dynamically coupled predictive models and data assimilation algorithms to call historical data sequences and real-time data streams on demand.

3. The method for dynamic prediction of tunnel water inflow according to claim 2, characterized in that: The coupling mechanism between the machine learning surrogate model and the hydrogeological numerical model is as follows: The machine learning proxy model adopts a long short-term memory network architecture. Its input feature vector is constructed through feature engineering and includes at least: the current three-dimensional spatial coordinates of the working face, the distance of the interface of the adverse geological body in front of the working face and the intensity of the abnormal physical property parameters based on the interpretation of advanced geological forecast data, the probability distribution of lithology in front based on the initial hydrogeological conceptual model, the statistical characteristic value of the permeability field updated by inversion in the predicted influence domain, and the effective rainfall sequence of the previous period considering the lag effect. The operation process of the coupling mechanism is as follows: In each prediction period, a forward simulation is first performed using the hydrogeological numerical model under the current parameter state to obtain the baseline water inflow prediction value; then, the input feature vector is input into the pre-trained long short-term memory network proxy model, and the long short-term memory network proxy model outputs the nonlinear correction amount for the baseline prediction value, or directly outputs the final water inflow prediction value after fusion.

4. The method for dynamic prediction of tunnel water inflow according to claim 3, characterized in that: In step S4, the data assimilation algorithm is the ensemble Kalman filter algorithm, and its specific execution process includes: The key parameter field of the hydrogeological numerical model is defined as the state vector of the ensemble Kalman filter algorithm. The key parameter field includes at least the permeability tensor and specific yield of each rock layer, and the parameter set is initialized to characterize its prior uncertainty. At each assimilation moment, the time series of water inflow at the working face collected within the predetermined time window and the water level change sequence of the monitoring well are combined to form an observation vector; The prediction step of the Kalman filter algorithm is executed, and the evolution of all parameter set members in the state vector is advanced through the hydrogeological numerical model. The update step of the Kalman filter algorithm is executed, and the residual between the actual measured value and the model prediction value of the observed vector is used to update the state vector, thereby obtaining an updated set of parameter fields with reduced uncertainty. The mean of this set is used as the optimal parameter estimate for the next prediction step.

5. The method for dynamic prediction of tunnel water inflow according to claim 4, characterized in that: In step S5, the probability distribution is generated as follows: The dynamic coupling prediction model achieves probabilistic prediction through the Monte Carlo simulation framework. Specifically, it is based on the parameter field set updated by the Kalman filter algorithm. Each member of the parameter set drives the hydrogeological numerical model or machine learning proxy model to perform a water inflow prediction to obtain a set of predicted values. Kernel density estimation is performed on the set of predicted values ​​to fit the water inflow rate in a specific section ahead of the tunnel face. probability density function : In the formula, Let be the number of members in the predicted value set. For the first The predicted value of each member, For kernel function, For bandwidth parameters; The cumulative distribution function is calculated based on the probability density function to determine the threshold range of water inflow corresponding to different risk levels.

6. The method for dynamic prediction of tunnel water inflow according to claim 5, characterized in that: The training and updating of the Long Short-Term Memory (LSTM) network agent model employs a two-stage strategy: Before the tunnel project starts, offline pre-training is carried out: using the initial hydrogeological conceptual model, combined with parameter sensitivity analysis and Latin hypercube sampling technology, a large-scale training sample set covering different hydrogeological scenarios is generated in the possible value space of key parameters. The sample set contains input feature vectors and corresponding water inflow labels. The sample set is used to initially train the long short-term memory network proxy model to establish the initial nonlinear mapping relationship between input and output. During tunnel construction, online fine-tuning is carried out: when new monitoring data accumulates to a preset scale, it is combined with the corresponding real-time input feature vector to form an incremental training sample set. The weight parameters of the long short-term memory network proxy model are fine-tuned with a learning rate lower than that in the pre-training stage to adapt it to the local hydrogeological conditions actually exposed by the tunnel.

7. The method for dynamic prediction of tunnel water inflow according to claim 6, characterized in that: In step S4, the construction and optimization process of the particle swarm optimization algorithm is as follows: Define the objective function The error norm between the simulation output of the hydrogeological numerical model and the actual monitoring data for the corresponding time period: In the formula, The vector of hydrogeological parameters to be inverted. The length of the data time window used for inversion. and They are time points Simulated and observed water inflow rates and They are time points Simulated water level and observed water level and These are the weighting coefficients for the water inflow and water level data, respectively. Initialize a group of particles, where the position of each particle represents a vector of potential hydrogeological parameters. By iteratively updating the velocity and position of the particles, the particle swarm is made to move towards the objective function. Minimize the region movement to ultimately obtain the optimal parameter estimate. .

8. The method for dynamic prediction of tunnel water inflow according to claim 7, characterized in that: In step S5, the risk level of water inrush is determined. The division rules are as follows: In the formula, The tunnel is designed with drainage capacity. Low risk Medium risk High risk, and To be a threshold constant preset according to engineering safety criteria, As a risk index, To predict the inflow The probability density function.

9. The method for dynamic prediction of tunnel water inflow according to claim 8, characterized in that: The specific implementation method of step S6 is as follows: By integrating the 3D visualization platform with the tunnel construction information model or digital twin system, the risk level R is rendered in the form of 3D isosurfaces with different colors and transparency in the 3D spatial model of the unexcavated section in front of the tunnel face. Automatically generate and output the water inflow prediction curve and risk level profile along the tunnel axis, with the start and end mileage of high-risk sections marked on the map. The feedback optimization process is as follows: when the tunnel excavation exposes the predicted section, the actual water inflow data recorded by the monitoring equipment deployed in that section is automatically captured by the system and correlated and compared with the historical prediction records stored in the dynamic database. Based on the comparison results, a model accuracy evaluation report is generated, and the actual water inflow data is used as a new training sample with real labels and stored in the dynamic database. This triggers the re-execution of the dynamic inversion and update steps of model parameters, as well as the online fine-tuning of the machine learning agent model.

Images