Method for evaluating applicability of TBM (Tunnel Boring Machine) of underground cavern of pumped storage power station
By constructing a non-AHP-driven dynamic multi-source data fusion machine learning evaluation framework, combined with a real-time monitoring system and high real-time sensors, the problems of the specificity and single index system of existing TBM applicability evaluation methods are solved, and efficient and safe construction of underground cavern TBMs for pumped storage power stations is realized.
Patent Information
- Application Number
- CN202511865146.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-17
AI Technical Summary
Existing TBM applicability evaluation methods are insufficient to meet the special needs of underground caverns in pumped storage power stations. They lack specificity and have a single indicator system, resulting in evaluation results that are out of touch with actual needs.
We adopt a non-AHP-driven, dynamic multi-source data fusion, and machine learning-enabled evaluation framework to construct static and dynamic indicator systems. By combining real-time monitoring systems and high real-time sensors, we automatically learn indicator weights through machine learning models, dynamically update applicability scores, and use SHAP values to explain the model decision-making process.
It has achieved a highly targeted and comprehensive TBM applicability evaluation, which has improved construction efficiency and safety, covered non-traditional indicators such as ecological environment and life cycle cost, and ensured the timeliness and understandability of the evaluation results.
Smart Images
Figure CN121684306A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of TBM equipment technology, specifically to a method for evaluating the suitability of underground cavern TBMs in pumped storage power stations. Background Technology
[0002] With the large-scale grid connection of new energy sources such as wind power and photovoltaics, pumped storage power stations, as the "regulation hub of the new power system," have become key equipment to support the consumption of new energy and ensure the safety of the power grid due to their multiple functions such as peak shaving and valley filling, frequency regulation and phase regulation, and energy storage backup.
[0003] Underground caverns (including water conveyance systems, underground powerhouses, drainage corridors, and ventilation / safety tunnels) are the core components of pumped storage power stations, and their construction quality directly affects the station's operational efficiency and lifespan. These caverns are characterized by diverse types (horizontal tunnels, inclined shafts, caverns with small turning radii, etc.), significant functional differences (e.g., drainage corridors must meet low permeability requirements, and water intake inclined shafts must withstand high water pressure), and complex geological conditions (hard rock, faults, groundwater, etc.). Traditional drill-and-blast methods suffer from drawbacks such as low tunneling efficiency, poor working environment, and significant disturbance to the surrounding rock, making it difficult to meet the demands of large-scale, high-quality construction.
[0004] Hard rock tunnel boring machines (TBMs), as efficient, safe, and environmentally friendly tunnel construction equipment, have been widely used in the construction of underground caverns for pumped storage power stations in recent years. Compared with the drill-and-blast method, TBMs have advantages such as fast tunneling speed, high construction accuracy, and minimal disturbance to the surrounding rock. They are particularly suitable for the construction of underground caverns in pumped storage power stations with medium to long distances, large cross sections, or complex geological conditions, and have become key equipment for promoting the "green and intelligent construction" of pumped storage power stations. However, existing TBM applicability evaluation methods are difficult to meet the special needs of pumped storage power stations: lack of specificity: existing evaluations are mostly based on experience from railway and subway projects, without considering the characteristics of "diverse types and large functional differences" of pumped storage caverns, resulting in evaluation results that are out of touch with actual needs; and a single indicator system. Summary of the Invention
[0005] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method for evaluating the applicability of underground tunnel boring machines (TBMs) in pumped storage power stations. This method has advantages such as strong targeting and a comprehensive indicator system, thus solving the aforementioned problems.
[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a method for evaluating the suitability of a TBM (Toyota Burglar Muller) for underground caverns in pumped storage power stations, comprising the following steps: S1. First, we will review the existing framework for the applicability evaluation of underground tunnel boring machines (TBMs) in pumped storage power stations, clarify its core technical elements and protection scope, and propose an evaluation framework that is non-AHP driven, dynamic multi-source data fusion, and machine learning empowered. S2. Construct an indicator system: divided into static indicators (design stage: tunnel depth, surrounding rock type, turning radius) and dynamic indicators (construction stage: TBM posture error, tool wear rate, surrounding rock deformation rate, ventilation efficiency). S3. Collect historical data on TBM construction cases for pumped storage power stations (including index values and TBM applicability results: such as tunneling efficiency and accident rate). S4. Integrate the real-time monitoring system to input data such as TBM attitude, tool wear, and surrounding rock deformation into the model in real time and dynamically update the applicability score. For example, when the "tool wear rate" exceeds the threshold (e.g., 15%), the model automatically lowers the "TBM applicability score" and gives a suggestion to "replace the tool". S5. Use SHAP values to explain the decision-making process of the machine learning model (such as the percentage of the impact of "surrounding rock deformation rate" on the score) to ensure that the results can be understood by engineers.
[0007] Preferably, the existing framework in S1 is mostly based on the Analytic Hierarchy Process (AHP) to construct a static evaluation model of "target layer - criterion layer - sub-criterion layer". It does not fully integrate dynamic data of the construction process (such as TBM real-time attitude, tool wear, and surrounding rock deformation), the evaluation results are lagging, rely on expert subjective judgment (AHP judgment matrix assignment), have insufficient repeatability, and do not cover non-traditional indicators such as ecological and environmental impact (such as the disturbance of groundwater and surface vegetation by tunnel construction) and life cycle cost (such as equipment maintenance and slag disposal).
[0008] Preferably, S1 replaces the AHP judgment matrix assignment with a machine learning model (such as random forest or XGBoost), trains the model with historical project data (such as TBM tunneling speed, surrounding rock type, and construction cost), and automatically learns the indicator weights; integrates dynamic real-time data: adds dynamic indicators of the construction stage (such as TBM real-time attitude error, tool wear rate, surrounding rock deformation rate, and ventilation smoke dispersion efficiency) to construct a dynamic evaluation of the entire "design-construction" process; and adds dimensions such as ecological and environmental impact (such as the degree of disturbance to groundwater runoff during tunnel construction), life cycle cost (such as equipment maintenance cost and comprehensive utilization benefits of slag), and ergonomic adaptability (such as the correlation between the tilt angle of the TBM operating platform and worker fatigue).
[0009] Preferably, the static indicators in S2 are determined during the project design phase, reflecting geological and engineering planning characteristics, and are mainly used for site selection assessment, scheme comparison, and risk prediction, possessing early certainty: Cavern depth: affects ground stress level and support design; Surrounding rock type: determines support strength and tunneling parameters; Turning radius: affects TBM selection and tunneling path feasibility. Once determined, these indicators typically do not change with construction, but data accuracy and representativeness must be ensured during the design phase. Dynamic indicators are used for real-time monitoring, early warning, and optimization decision-making during construction, possessing time-varying characteristics and operability: TBM attitude error: reflects tunneling progress... Directional control accuracy requires real-time correction; tool wear rate affects tunneling efficiency and cost, and determines the tool replacement cycle; surrounding rock deformation rate reflects geological stability and is key to safety early warning; ventilation efficiency relates to the safety of the working environment and the health of personnel. For these indicators, a threshold alarm mechanism needs to be established and linked with process optimization. A "expert interview + machine learning" approach is adopted to jointly screen indicators: first, core indicators are identified by experts, and then redundant indicators are eliminated using machine learning (such as correlation analysis) to ensure the simplicity and effectiveness of the indicator system. This phased indicator system helps to achieve closed-loop management of "pre-design control + in-process monitoring".
[0010] Preferably, S3 uses a random forest model to train the mapping relationship between "indicators and applicability" and automatically outputs the weight of each indicator (e.g., the weight of "surrounding rock deformation rate" is higher than that of "cavity burial depth").
[0011] Preferably, in step S4, highly real-time and highly reliable sensors need to be deployed for the core aspects of TBM construction (attitude, cutting tools, surrounding rock) to ensure data coverage of the entire construction process. Industrial Ethernet (Ethernet / IP) or 5G / 4G wireless communication is used to meet the high bandwidth and low latency requirements of the TBM construction environment (such as underground tunnels). Edge servers are deployed locally on the TBM to perform preliminary data processing (such as noise filtering and feature extraction) to reduce cloud transmission pressure and improve real-time performance. Digital filtering algorithms (such as low-pass filtering and wavelet transform) are used to remove high-frequency interference (such as motor noise and mechanical vibration) in vibration and force signals. Time stamp alignment (such as GPS time synchronization) is used to unify the time reference of multi-source data. Features with clear engineering significance are extracted from the original signals to reduce model complexity: Vibration signals: extract RMS (root mean square), peak value, kurtosis, and frequency spectrum (main frequency after FFT transformation); Force signals: extract mean, variance, and maximum cutting force; AE signals: extract RMS, event count, and energy. The core of dynamic applicability scoring is to map real-time data into a quantitative assessment of the construction status, reflecting whether the current TBM is suitable for continued tunneling.
[0012] Preferably, in S5, SHAP is a model interpretation method based on Shapley values in game theory. It treats the decision of the machine learning model as the result of feature "cooperation" (such as "surrounding rock deformation rate", "burial depth", "lithology" and other features jointly determining the "stability score"). By quantifying the marginal contribution of each feature to the decision, it provides a fair and consistent interpretation. The SHAP contribution of "surrounding rock deformation rate" to the model decision (such as "surrounding rock stability score") needs to follow the process of "benchmark value setting → feature subset generation → marginal contribution calculation → weighted average". The benchmark value is usually the mean / median of the features in the training data (such as the mean of "typical surrounding rock deformation rate" in engineering, which represents the "average decision value when there is no feature information"). Feature subset generation is to generate feature combinations that include / do not include "surrounding rock deformation rate" through random sampling. Marginal contribution is to calculate the difference in decision when "this feature is present or absent". Weighted average is to weight the marginal contribution according to the probability of the feature combination to obtain the final SHAP value.
[0013] Compared with the prior art, the present invention provides a method for evaluating the suitability of underground tunnel boring machines (TBMs) for pumped storage power stations, which has the following beneficial effects: 1. This invention proposes an evaluation framework that is non-AHP driven, dynamically multi-source data fusion, and machine learning-enabled, and constructs an indicator system: divided into static indicators (design stage: tunnel depth, surrounding rock type, turning radius) and dynamic indicators (construction stage: TBM attitude error, tool wear rate, surrounding rock deformation rate, ventilation efficiency). It integrates a real-time monitoring system, inputting data such as TBM attitude, tool wear, and surrounding rock deformation into the model in real time, and dynamically updating the applicability score. Thus, this method has the advantages of strong targeting and comprehensive indicator system.
[0014] 2. This invention replaces the judgment matrix assignment of AHP with machine learning models (such as random forest and XGBoost), trains the model with historical project data (such as TBM tunneling speed, surrounding rock type, and construction cost), and automatically learns the index weights; it integrates dynamic real-time data: adds dynamic indicators of the construction stage (such as TBM real-time attitude error, tool wear rate, surrounding rock deformation rate, and ventilation smoke dispersion efficiency) to construct a dynamic evaluation of the entire "design-construction" process.
[0015] 3. This invention deploys highly real-time and highly reliable sensors for the core aspects of TBM construction (attitude, cutting tools, and surrounding rock) to ensure data coverage of the entire construction process. It adopts industrial Ethernet (Ethernet / IP) or 5G / 4G wireless communication to meet the high bandwidth and low latency requirements of TBM construction environments (such as underground tunnels). Edge servers are deployed locally on the TBM to perform preliminary data processing (such as noise filtering and feature extraction), reducing the pressure on cloud transmission and improving real-time performance. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] Please see Figure 1 A method for evaluating the suitability of underground tunnel boring machines (TBMs) for pumped storage power stations includes the following steps: S1. First, we will review the existing framework for the applicability evaluation of underground tunnel boring machines (TBMs) in pumped storage power stations, clarify its core technical elements and protection scope, and propose an evaluation framework that is non-AHP driven, dynamic multi-source data fusion, and machine learning empowered. S2. Construct an indicator system: divided into static indicators (design stage: tunnel depth, surrounding rock type, turning radius) and dynamic indicators (construction stage: TBM posture error, tool wear rate, surrounding rock deformation rate, ventilation efficiency). S3. Collect historical data on TBM construction cases for pumped storage power stations (including index values and TBM applicability results: such as tunneling efficiency and accident rate). S4. Integrate the real-time monitoring system to input data such as TBM attitude, tool wear, and surrounding rock deformation into the model in real time and dynamically update the applicability score. For example, when the "tool wear rate" exceeds the threshold (e.g., 15%), the model automatically lowers the "TBM applicability score" and gives a suggestion to "replace the tool". S5. Use SHAP values to explain the decision-making process of the machine learning model (such as the percentage of the impact of "surrounding rock deformation rate" on the score) to ensure that the results can be understood by engineers.
[0019] The existing framework in S1 is mostly based on the Analytic Hierarchy Process (AHP) to construct a static evaluation model of "target layer - criterion layer - sub-criterion layer". It does not fully integrate dynamic data of the construction process (such as TBM real-time attitude, tool wear, and surrounding rock deformation), the evaluation results are lagging, rely on expert subjective judgment (AHP judgment matrix assignment), have insufficient repeatability, and do not cover non-traditional indicators such as ecological and environmental impact (such as the disturbance of groundwater and surface vegetation by tunnel construction) and life cycle cost (such as equipment maintenance and slag disposal).
[0020] S1 replaces the judgment matrix assignment in AHP with machine learning models (such as Random Forest and XGBoost). It trains the model using historical project data (such as TBM tunneling speed, surrounding rock type, and construction cost) to automatically learn indicator weights. It integrates dynamic real-time data, incorporating dynamic indicators from the construction phase (such as TBM real-time attitude error, tool wear rate, surrounding rock deformation rate, and ventilation smoke dissipation efficiency) to construct a dynamic evaluation of the entire "design-construction" process. It adds dimensions such as ecological and environmental impact (e.g., the degree of disturbance to groundwater runoff during tunnel construction), life-cycle cost (e.g., equipment maintenance costs and the benefits of comprehensive utilization of slag), and ergonomic adaptability (e.g., the correlation between the tilt angle of the TBM operating platform and worker fatigue). The core formula, Random Forest, focuses on feature importance. , ( (This refers to the reduction in Gini impurity during node splitting).
[0021] Static indicators in S2: Determined during the project design phase, reflecting geological and engineering planning characteristics, primarily used for site selection assessment, scheme comparison, and risk prediction, possessing early-stage certainty: Cavern depth: Affects ground stress level and support design; Surrounding rock type: Determines support strength and tunneling parameters; Turning radius: Affects TBM selection and tunneling path feasibility. Once determined, these indicators typically do not change with construction, but data accuracy and representativeness must be ensured during the design phase. Dynamic indicators: Used for real-time monitoring, early warning, and optimization decision-making during construction, possessing time-varying characteristics and operability: TBM attitude error: Reflects tunneling direction control... Precision control requires real-time correction; tool wear rate affects tunneling efficiency and cost, and determines the tool replacement cycle; surrounding rock deformation rate reflects geological stability and is key to safety early warning; ventilation efficiency relates to the safety of the working environment and the health of personnel. For these indicators, a threshold alarm mechanism needs to be established and linked with process optimization. A "expert interview + machine learning" approach is adopted to jointly screen indicators: first, core indicators are identified by experts, and then redundant indicators are eliminated using machine learning (such as correlation analysis) to ensure the conciseness and effectiveness of the indicator system. This phased indicator system helps to achieve closed-loop management of "pre-design control + in-process monitoring".
[0022] S3 uses a random forest model to train the mapping relationship between "indicators and applicability" and automatically outputs the weight of each indicator (e.g., the weight of "surrounding rock deformation rate" is higher than that of "cavity burial depth").
[0023] In S4, highly real-time and reliable sensors need to be deployed for core aspects of TBM construction (attitude, cutting tools, surrounding rock) to ensure data coverage of the entire construction process. Industrial Ethernet (Ethernet / IP) or 5G / 4G wireless communication is used to meet the high bandwidth and low latency requirements of TBM construction environments (such as underground tunnels). Edge servers are deployed locally on the TBM to perform preliminary data processing (such as noise filtering and feature extraction) to reduce cloud transmission pressure and improve real-time performance. Digital filtering algorithms (such as low-pass filtering and wavelet transform) are used to remove high-frequency interference (such as motor noise and mechanical vibration) from vibration and force signals. Time stamp alignment (such as GPS time synchronization) is used to unify the time reference of multi-source data. Features with clear engineering significance are extracted from the original signals to reduce model complexity: Vibration signals: extract RMS (root mean square), peak value, kurtosis, and frequency spectrum (main frequency after FFT transformation); Force signals: extract mean, variance, and maximum cutting force; AE signals: extract RMS, event count, and energy. The core of dynamic suitability scoring is to map real-time data into a quantitative assessment of the construction status, reflecting whether the current TBM is suitable for continued tunneling.
[0024] In S5, SHAP is a model interpretation method based on Shapley values in game theory. It treats the decision of the machine learning model as the result of feature "cooperation" (such as features like "surrounding rock deformation rate," "burial depth," and "lithology" jointly determining the "stability score"). By quantifying the marginal contribution of each feature to the decision, it provides a fair and consistent interpretation. Calculating the SHAP contribution of "surrounding rock deformation rate" to the model decision (such as "surrounding rock stability score") requires following the process of "benchmark value setting → feature subset generation → marginal contribution calculation → weighted average." The benchmark value is usually the mean / median of features in the training data (e.g., in engineering, the mean of "typical surrounding rock deformation rate" is used, representing the "average decision value without feature information"). Feature subset generation involves generating feature combinations that include / do not include "surrounding rock deformation rate" through random sampling. Marginal contribution calculates the difference in decision depending on whether the feature is present or absent. Weighted average weights the marginal contribution based on the probability of the feature combination to obtain the final SHAP value.
[0025] When using this technology, we first review the existing framework for the suitability evaluation of underground tunnel boring machines (TBMs) in pumped storage power stations, clarifying their core technical elements and protection scope. We then propose an evaluation framework that is not driven by AHP (Automatic Hierarchical Construction), uses dynamic multi-source data fusion, and is empowered by machine learning. We construct an indicator system, which is divided into static indicators (design stage: tunnel depth, surrounding rock type, turning radius) and dynamic indicators (construction stage: TBM attitude error, tool wear rate, surrounding rock deformation rate, ventilation efficiency). We collect historical data from TBM construction cases in pumped storage power stations, integrate real-time monitoring systems, and input data such as TBM attitude, tool wear, and surrounding rock deformation into the model in real time to dynamically update the suitability score. We use SHAP (Shape-Up Point) values to explain the decision-making process of the machine learning model (such as the percentage of influence of "surrounding rock deformation rate" on the score) to ensure that the results can be understood by engineers.
[0026] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for evaluating TBM applicability of underground caverns of pumped storage power stations, characterized in that: Comprise the following steps: S1, First, sort out the existing framework of TBM applicability evaluation of underground cavern of pumped storage power station, clear its core technical elements and protection scope, put forward the evaluation framework of non-AHP driving, dynamic multi-source data fusion and machine learning empowerment; S2, Build index system: divided into static index (design stage: cavern burial depth, surrounding rock type, turning radius) and dynamic index (construction stage: TBM attitude error, tool wear rate, surrounding rock deformation rate, ventilation efficiency); S3, Collect historical data of TBM construction case of pumped storage power station (including index value and TBM applicability result: such as tunneling efficiency, safety accident rate); S4, Integrate real-time monitoring system, input TBM attitude, tool wear, surrounding rock deformation and other data into the model in real time, and dynamically update the applicability score; S5, Use SHAP value to explain the decision-making process of machine learning model (such as the proportion of "surrounding rock deformation rate" to the score), ensure that the result can be understood by engineering personnel.
2. The method for evaluating the TBM applicability of underground caverns of pumped storage power stations according to claim 1, characterized in that: The existing framework in S1 is mostly based on analytic hierarchy process (AHP) to build a static evaluation model of "target layer-criteria layer-sub-criteria layer", which does not fully integrate dynamic data in the construction process (such as TBM real-time attitude, tool wear, surrounding rock deformation), the evaluation result is lagging behind, the repeatability is insufficient, and it does not cover non-traditional indicators such as ecological environmental impact (such as the disturbance of cavern construction to underground water and surface vegetation) and whole life cycle cost (such as equipment maintenance and slag treatment).
3. The method for evaluating the TBM applicability of underground caverns of pumped storage power stations according to claim 1, characterized in that: The S1 uses machine learning model (such as random forest, XGBoost) to replace the judgment matrix assignment of AHP, trains the model through historical project data (such as TBM tunneling speed, surrounding rock type, construction cost), automatically learns the index weight; Integrate dynamic real-time data: add dynamic indexes in construction stage (such as TBM real-time attitude error, tool wear rate, surrounding rock deformation rate, ventilation and smoke efficiency), build "design-construction" whole process dynamic evaluation; Increase the dimensions of ecological environmental impact (such as the disturbance degree of cavern construction to groundwater runoff), whole life cycle cost (such as equipment maintenance cost, slag comprehensive utilization benefit), man-machine engineering adaptability (such as the correlation between the inclination angle of TBM operation platform and worker fatigue).
4. The method for evaluating the TBM applicability of underground caverns of pumped storage power stations according to claim 1, characterized in that: The static indicators in S2 are determined in the project design phase, reflecting geological and engineering planning characteristics, and are mainly used for site selection evaluation, scheme comparison and risk prediction, with early determination: tunnel depth: affecting ground stress level and support design; surrounding rock type: determining support strength and tunneling parameters; turning radius: affecting TBM selection and tunneling path feasibility; once these indicators are determined, they usually do not change with construction, but the accuracy and representativeness of the data need to be ensured during the design phase. Dynamic indicators: used for real-time monitoring, early warning and optimization decision-making during construction, with time-varying and intervenable characteristics: TBM attitude error: reflecting the control accuracy of the tunneling direction, which needs to be corrected in real time; cutter wear rate: affecting tunneling efficiency and cost, determining the cutter replacement period; surrounding rock deformation rate: reflecting geological stability, which is a key safety warning; ventilation efficiency: related to the safety of the working environment and the health of personnel, such threshold alarm mechanisms need to be established for such indicators, and linked with process optimization, using "expert interview + machine learning" to jointly select indicators: first identify core indicators through expert identification, then use machine learning (such as correlation analysis) to eliminate redundant indicators, ensuring the simplicity and effectiveness of the indicator system.
5. The method for evaluating the TBM applicability of underground caverns of pumped storage power stations according to claim 1, characterized in that: S3 uses a random forest model to train the mapping relationship of "indicator-applicability", automatically outputting the weight of each indicator (such as the weight of "surrounding rock deformation rate" being higher than "tunnel depth").
6. The method for evaluating the TBM applicability of underground caverns of pumped storage power stations according to claim 1, characterized in that: In S4, high real-time and reliable sensors need to be deployed for the core aspects of TBM construction (attitude, cutter, surrounding rock) to ensure data coverage throughout the construction process, using industrial Ethernet (Ethernet / IP) or 5G / 4G wireless communication to meet the high bandwidth and low latency requirements of TBM construction environments (such as underground tunnels), deploying edge servers locally on the TBM for preliminary processing of data (such as noise filtering and feature extraction), reducing cloud transmission pressure and improving real-time performance, using digital filtering algorithms (such as low-pass filtering and wavelet transform) to remove high-frequency interference (such as motor noise and mechanical vibration) in vibration and force signals, aligning the time base of multi-source data through timestamp alignment (such as GPS time), extracting engineering-specific features from raw signals to reduce model complexity: vibration signals: extracting RMS (root mean square), peak value, kurtosis, and frequency spectrum (main frequency after FFT transformation); force signals: extracting mean, variance, and maximum cutting force; AE signals: extracting RMS, event count, and energy; the core of dynamic applicability scoring is to map real-time data into a quantitative assessment of the construction state, reflecting whether the current TBM is suitable for continued tunneling.
7. The method for evaluating the TBM applicability of underground caverns of pumped storage power stations according to claim 1, characterized in that: The SHAP in S5 is a model explanation method based on the Shapley value in game theory, which regards the decision of the machine learning model as the result of the "cooperation" of the features (such as "rock deformation rate", "buried depth", "lithology", etc. Features such as "stability score" determine the "stability score"), by quantifying the marginal contribution of each feature to the decision, a fair and consistent explanation is provided. The SHAP contribution of "rock deformation rate" to the model decision (such as "rock stability score") needs to follow the process of "baseline setting → feature subset generation → marginal contribution calculation → weighted average", where the baseline value: usually take the mean / median of the feature in the training data (such as taking the mean of "typical rock deformation rate" in engineering, representing "average decision value when there is no feature information"); Feature subset generation: generate feature combinations containing / containing "rock deformation rate" by random sampling; Marginal contribution: calculate the decision difference when "with / without the feature"; Weighted average: according to the probability of the feature combination, the marginal contribution is weighted to get the final SHAP value.