Training method and device of cavern deformation prediction model based on geological space collaborative clustering

Abstract

This disclosure provides a training method and apparatus for a cavern deformation prediction model based on geological spatial collaborative clustering. The method includes: acquiring multi-source feature data of each monitoring point in the target underground cavern; performing spatial clustering on each monitoring point based on its spatial coordinate data and geological feature data to obtain multiple clusters and category labels for each cluster; determining the geological spatial collaborative weights corresponding to each cluster based on the temporal deformation data of the monitoring points within each cluster; constructing an input feature set based on the multi-source feature data and corresponding category labels of each monitoring point; determining the weighted loss function in a pre-constructed initial gradient boosting tree model based on the geological spatial collaborative weights of each cluster; and iteratively training the initial gradient boosting tree model using the input feature set and the weighted loss function to obtain the cavern deformation prediction model. This disclosure can improve the geological interpretability of the prediction model in practical engineering applications.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to a training method and apparatus for a cavern deformation prediction model based on geological spatial collaborative clustering. Background Technology

[0002] The geological conditions of underground caverns in water conservancy and hydropower projects are extremely complex, and scientifically predicting the deformation trend of the surrounding rock is crucial to ensuring construction and operational safety. However, most deformation prediction models in related technologies are typical black-box models, whose internal working mechanisms rely solely on mathematical statistics and data mapping, detached from the actual physical and mechanical mechanisms and geological laws of geotechnical engineering. During training, they generally adopt globally fixed parameter tuning strategies and standard error loss functions, ignoring the geological spatial heterogeneity of underground caverns.

[0003] Therefore, when dealing with underground caverns with large spatial spans and significant variations in geological conditions, the training mechanisms in related technologies that are detached from geological constraints result in prediction models that lack engineering logic support, leading to poor interpretability of the prediction models in practical applications. Summary of the Invention

[0004] To overcome the problems existing in related technologies, this disclosure provides a training method and device for a cavern deformation prediction model based on geological spatial collaborative clustering, which can improve the interpretability of the prediction model in practical engineering applications.

[0005] According to a first aspect of the present disclosure, a method for training a cavern deformation prediction model based on geological spatial collaborative clustering is provided, the method comprising: Acquire multi-source feature data of each monitoring point in the target underground cavern. The multi-source feature data includes at least the spatial coordinate data, geological feature data, and temporal deformation data of each monitoring point. Based on the spatial coordinate data of each monitoring point and the geological feature data, the collaborative metric distance between any two monitoring points is calculated, and each monitoring point is spatially clustered according to the collaborative metric distance to obtain multiple clusters and category labels for each cluster. Based on the temporal deformation data of the monitoring points in each cluster, the geological spatial collaborative weights corresponding to each cluster are determined. Based on the multi-source feature data and corresponding category labels of each monitoring point, an input feature set is constructed; The weighted loss function in the pre-constructed initial gradient boosting tree model is determined based on the geological spatial collaborative weights of each of the aforementioned clusters. The initial gradient boosting tree model is then iteratively trained using the input feature set and the weighted loss function to obtain the cavern deformation prediction model.

[0006] In one exemplary embodiment of this disclosure, the multi-source feature data further includes construction feature data of each of the monitoring points; The spatial coordinate data is used to indicate the three-dimensional spatial position of each monitoring point in the target underground cavern; the geological feature data includes at least the maximum principal stress and uniaxial compressive strength of the surrounding rock in the area where each monitoring point is located; the time-series deformation data includes the cumulative deformation of each monitoring point in the historical observation period, and / or the deformation rate; the construction feature data includes the excavation disturbance parameters and support compensation parameters for the area where each monitoring point is located.

[0007] In one exemplary embodiment of this disclosure, calculating the collaborative metric distance between any two monitoring points based on the spatial coordinate data of each monitoring point and the geological feature data includes: The distance between any two monitoring points is determined based on the spatial coordinate data of each monitoring point. Based on the geological characteristic data of each monitoring point, the ratio of the maximum principal stress of the surrounding rock to the uniaxial compressive strength of the surrounding rock at each monitoring point is determined to obtain the strength-stress ratio of the surrounding rock, and a geological difference penalty term is determined based on the difference between the strength-stress ratios of the surrounding rock at any two monitoring points. The collaborative measurement distance between any two monitoring points is determined based on the product of the distance and the geological difference penalty term.

[0008] In an exemplary embodiment of this disclosure, determining the geological spatial collaborative weights corresponding to each cluster based on the temporal deformation data of monitoring points in each cluster includes: Based on the time-series deformation data of the monitoring points in each of the aforementioned clusters, the intra-cluster deformation standard deviation of each of the aforementioned clusters is calculated respectively; The maximum value among the intra-cluster deformation standard deviations of each of the aforementioned clusters is determined as the global deformation standard deviation; The ratio of the intra-cluster deformation standard deviation to the global deformation standard deviation of each cluster is determined as the geological spatial collaborative weight corresponding to each cluster.

[0009] In an exemplary embodiment of this disclosure, after spatially clustering each monitoring point according to the collaborative metric distance to obtain multiple clusters and category labels for each cluster, the method further includes: Calculate the probability density distribution curves for each of the aforementioned clusters; If the probability density distribution curve of each cluster exhibits a single-peak distribution and the peak value is greater than or equal to a preset threshold, then the category label of each cluster is determined.

[0010] In one exemplary embodiment of this disclosure, an input feature set is constructed based on the multi-source feature data of each monitoring point and its corresponding category label, including: The multi-source feature data of each monitoring point and its corresponding category label are sequentially concatenated to obtain the input feature vector of each monitoring point. The input feature set is obtained based on the input feature vectors of each monitoring point.

[0011] In one exemplary embodiment of this disclosure, determining the weighted loss function in the pre-constructed initial gradient boosting tree model based on the geological spatial collaborative weights of each of the said clusters includes: Obtain the predicted and actual values ​​of the monitoring points within each cluster using the initial gradient boosting tree model; Based on the predicted and actual values, the mean square error of the initial gradient boosting tree model in each of the clusters is calculated. The weighted loss function is obtained by weighting and summing the prediction mean square error of each cluster with the corresponding geological spatial collaborative weight.

[0012] According to a second aspect of the present disclosure, a method for predicting cavern deformation based on geological spatial collaborative clustering is provided, the method comprising: Acquire multi-source feature data of the node to be predicted in the target underground cavern, wherein the multi-source feature data includes at least the spatial coordinate data and geological feature data of the node to be predicted; Based on the multi-source feature data and corresponding category labels of the nodes to be predicted, construct the input vector to be predicted; The input vector to be predicted is input into the pre-trained cavern deformation prediction model to obtain the deformation prediction value of the node at the location to be predicted. The cavern deformation prediction model is trained based on the training method described above.

[0013] According to a third aspect of the present disclosure, a training device for a cavern deformation prediction model based on geological spatial collaborative clustering is provided, comprising: The first acquisition module is used to acquire multi-source feature data of each monitoring point in the target underground cavern. The multi-source feature data includes at least the spatial coordinate data, geological feature data and temporal deformation data of each monitoring point. The determination module is used to calculate the collaborative metric distance between any two monitoring points based on the spatial coordinate data of each monitoring point and the geological feature data, and to perform spatial clustering of each monitoring point according to the collaborative metric distance to obtain multiple clusters and category labels of each cluster. Based on the temporal deformation data of the monitoring points in each cluster, the geological spatial collaborative weights corresponding to each cluster are determined. The construction module is used to construct an input feature set based on the multi-source feature data of each monitoring point and the corresponding category label; The training module is used to determine the weighted loss function in the pre-constructed initial gradient boosting tree model based on the geological spatial collaborative weights of each of the aforementioned clusters, and to iteratively train the initial gradient boosting tree model using the input feature set and the weighted loss function to obtain the cavern deformation prediction model.

[0014] According to a fourth aspect of the present disclosure, a cavern deformation prediction device based on geological spatial collaborative clustering is provided, comprising: The second acquisition module is used to acquire multi-source feature data of the node to be predicted in the target underground cavern. The multi-source feature data includes at least the spatial coordinate data and geological feature data of the node to be predicted. The input module is used to construct the input vector to be predicted based on the multi-source feature data of the node to be predicted and the corresponding category label; The prediction module is used to input the input vector to be predicted into the pre-trained cavern deformation prediction model to obtain the deformation prediction value of the node to be predicted. The cavern deformation prediction model is trained based on the cavern deformation prediction model based on the above-mentioned geological spatial collaborative clustering training method.

[0015] This disclosure provides a training method for a cavern deformation prediction model based on geological spatial collaborative clustering. By calculating the collaborative metric distance based on the spatial coordinate data and geological feature data of each monitoring point and performing spatial clustering, multiple clusters and category labels that integrate physical spatial location and prior geological attributes can be directly obtained without relying on complex black-box clustering algorithms. Therefore, compared with the traditional global modeling method that ignores the heterogeneity of geological space, this method can explicitly transform the geomechanical properties of geotechnical engineering into the spatial partitioning boundary of the data, reducing the statistical bias caused by a single feature dimension. At the same time, by determining the geological spatial collaborative weights corresponding to each cluster based on the temporal deformation data of monitoring points in each cluster, the difference in dynamic deformation degree between different spatial clusters can be quantified, providing a data foundation for the differentiated training of the subsequent model.

[0016] Furthermore, the multi-source feature data of each monitoring point and the corresponding category labels are used to construct an input feature set, and the geological spatial collaborative weight configuration of each cluster is introduced as a penalty coefficient into the loss function of the initial gradient boosting tree model to construct a weighted loss function. Thus, in the iterative training stage of the prediction model, differentiated penalty processing of prediction errors of different geological clusters can be achieved, so that when facing the overall cavern monitoring data, the error penalty intensity of high-weight clusters is amplified, while maintaining the normal convergence of low-weight stable sections.

[0017] Furthermore, by iteratively training the initial gradient boosting tree model using the input feature set and a weighted loss function, the dilution and masking of local abnormal deformation features by the globally uniform loss function can be reduced under complex working conditions with significant geological spatial heterogeneity. This allows the prediction data of the final generated cavern deformation prediction model to accurately focus on and capture the abnormal deformation trends of high-risk geological sections. Through this method, this disclosure establishes a logical correspondence between the internal error backpropagation mechanism of the prediction model and the zoning characteristics of the actual geological space, thereby improving the geological interpretability of the prediction model in engineering applications while ensuring the control accuracy of local high-risk sections.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this disclosure, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0020] Figure 1 This is a flowchart illustrating a training method for a cavern deformation prediction model based on geological spatial collaborative clustering, according to an exemplary embodiment of this disclosure.

[0021] Figure 2 This is a flowchart illustrating a cavern deformation prediction method based on geological spatial collaborative clustering according to an exemplary embodiment of the present disclosure.

[0022] Figure 3 This is a schematic diagram of the structure of a training device for a cavern deformation prediction model based on geological spatial collaborative clustering, according to an exemplary embodiment of this disclosure.

[0023] Figure 4 This is a schematic diagram of a cavern deformation prediction device based on geological spatial collaborative clustering, according to an exemplary embodiment of the present disclosure.

[0024] Figure 5 This is a hardware structure diagram of a computer device shown in an embodiment of this disclosure. Detailed Implementation

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0026] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0027] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0028] The embodiments of this disclosure will now be described in detail.

[0029] like Figure 1 As shown, Figure 1 This disclosure is a flowchart illustrating a training method for a cavern deformation prediction model based on geological spatial collaborative clustering according to an exemplary embodiment, comprising the following steps: Step 101: Obtain multi-source feature data of each monitoring point in the target underground cavern.

[0030] Step 102: Based on the spatial coordinate data and geological feature data of each monitoring point, calculate the collaborative metric distance between any two monitoring points, and perform spatial clustering of each monitoring point according to the collaborative metric distance to obtain multiple clusters and category labels of each cluster. Based on the temporal deformation data of the monitoring points in each cluster, determine the geological spatial collaborative weights corresponding to each cluster.

[0031] Step 103: Construct an input feature set based on the multi-source feature data and corresponding category labels of each monitoring point.

[0032] Step 104: Determine the weighted loss function in the pre-constructed initial gradient boosting tree model based on the geological spatial collaborative weights of various clusters, and use the input feature set and the weighted loss function to iteratively train the initial gradient boosting tree model to obtain the cavern deformation prediction model.

[0033] The training method for a cavern deformation prediction model based on geological spatial collaborative clustering provided in this exemplary embodiment acquires multi-source feature data from each monitoring point, calculates collaborative metric distance based on spatial coordinate data and geological feature data, and performs spatial clustering. This enables the data preprocessing stage to cluster the originally spatially discrete monitoring point data with varying geological attributes into multiple independent clusters with spatial proximity and geological similarity. Furthermore, by extracting temporal deformation data from monitoring points within each cluster, geological spatial collaborative weights are determined, quantifying the actual deformation fluctuations in different local areas of the underground cavern into specific weight values ​​independent of the feature set.

[0034] Furthermore, by constructing an input feature set from multi-source feature data and category labels, and introducing geological spatial collaborative weights into the initial gradient boosting tree model, a weighted loss function for different clusters is determined, and finally iterative training is performed. Thus, compared to the traditional method of using a globally uniform loss function to calculate errors indiscriminately for all monitoring points, this method maps the deformation weights of spatial clusters to error penalty coefficients, enabling the model to apply differentiated parameter update intensity based on the weights of different clusters during gradient descent optimization. This reduces the dilution and masking effect of a large amount of globally stable data on the abnormal features of locally high-fluctuation sections, avoiding underfitting of the prediction model in high-risk geological sections. Through the differentiated iterative training driven by geological spatial collaborative weights, the originally purely data-mapped prediction algorithm incorporates constraint logic on the heterogeneity of underground space geomechanics during its internal network parameter optimization process. This gives the final generated cavern deformation prediction model geological spatial perception capabilities, improving the adaptive prediction accuracy of the model for abnormal deformation sections while ensuring global convergence, and enhancing the interpretability of the prediction model in engineering decision-making.

[0035] The training method of the cavern deformation prediction model based on geological spatial collaborative clustering in this example embodiment will be described in detail below.

[0036] In step 101, multi-source feature data of each monitoring point in the target underground cavern are acquired.

[0037] In the exemplary embodiments disclosed herein, the geological environment of the target underground cavern is extremely complex, and single-dimensional data often cannot accurately reflect the true deformation trend of the surrounding rock. Therefore, obtaining multi-dimensional feature data is the foundation for constructing a high-precision prediction model with geological interpretability.

[0038] Specifically, multi-source feature data can include at least spatial coordinate data, geological feature data, and temporal deformation data. Spatial coordinate data indicates the three-dimensional spatial location of each monitoring point within the target underground cavern, such as the three-dimensional coordinates of each monitoring point. Geological feature data indicates the geological conditions of the area where each monitoring point is located. These geological conditions can include the physical and mechanical properties of the rock mass and information on the geological environment. This geological feature data includes at least the maximum principal stress and uniaxial compressive strength of the surrounding rock in the area where each monitoring point is located. Optionally, this geological feature data can also include the type of surrounding rock and external water pressure. The maximum principal stress and uniaxial compressive strength of the surrounding rock directly reflect the stress state of the rock mass and its ability to resist failure, and are core indicators for subsequent quantification of geological differences and correction of physical spatial distances. Temporal deformation data indicates the dynamic characteristics of the surrounding rock deformation evolution at each monitoring point within the historical observation period. This temporal deformation data includes the cumulative deformation amount and / or deformation rate of each monitoring point within the historical observation period. This temporal deformation data reflects the current degree of deformation intensity of the surrounding rock, used for subsequent extraction of fluctuation features to generate risk weights. It can also serve as the true label to be fitted by the prediction model, i.e., the prediction target.

[0039] The multi-source feature data may also include construction feature data. This construction feature data is used to indicate the engineering disturbance state caused by tunnel excavation and support operations to the area where each monitoring point is located. Specifically, the construction feature data includes excavation disturbance parameters and support compensation parameters. Among them, the excavation disturbance parameters include at least one of the following: excavation method, excavation advance, distance from the tunnel face, and blasting vibration intensity, used to quantify the spatial unloading effect generated by the excavation operation; the support compensation parameters include at least one of the following: support type and support lag time, used to characterize the inhibitory effect of the support measures on the deformation of the surrounding rock.

[0040] For example, in practical engineering, the raw dataset of multi-source feature data from all monitoring points can be collected using various sensors and surveying equipment deployed in the underground caverns of the target hydropower station. Since different physical dimensions and orders of magnitude are involved, after obtaining the raw dataset and before applying it to calculate the collaborative distance metric or construct the input feature set, the Z-score normalization method can be used to perform dimensionless processing on the raw dataset, resulting in processed multi-source feature data. This eliminates the dimensional differences and numerical orders of magnitude between the heterogeneous multi-source data, ensuring the accuracy of each feature in subsequent spatial distance calculations and gradient boosting tree model iterative training.

[0041] In step 102, based on the spatial coordinate data and geological feature data of each monitoring point, the collaborative metric distance between any two monitoring points is calculated, and spatial clustering is performed on each monitoring point according to the collaborative metric distance to obtain multiple clusters and category labels of each cluster. Based on the temporal deformation data of the monitoring points in each cluster, the geological spatial collaborative weights corresponding to each cluster are determined.

[0042] In the exemplary embodiment of this disclosure, the collaborative measurement distance between any two monitoring points is calculated based on the spatial coordinate data and geological feature data of each monitoring point, specifically including the following steps 1021-1023: Step 1021: Determine the distance between any two monitoring points based on the spatial coordinate data of each monitoring point.

[0043] This distance characterizes the geometric proximity of any two monitoring points in a three-dimensional physical space within an underground cavern. Specifically, the spatial coordinate data includes the three-dimensional spatial coordinates of the monitoring points after eliminating the influence of dimensions. Based on the above three-dimensional spatial coordinates, the physical distance between the two monitoring points can be calculated using the following formula: in, Represents any two monitoring points i and monitoring points j Physical distance between them, monitoring points i The corresponding three-dimensional spatial coordinates are Monitoring points j The corresponding three-dimensional spatial coordinates are .

[0044] Step 1022: Based on the geological characteristic data of each monitoring point, determine the ratio of the maximum principal stress of the surrounding rock to the uniaxial compressive strength of the surrounding rock at each monitoring point, and obtain the strength-stress ratio of the surrounding rock.

[0045] The maximum principal stress of the surrounding rock in geological characteristic data represents the magnitude of the external tectonic stress on the rock mass, while the uniaxial compressive strength of the surrounding rock represents the inherent ability of the rock mass to resist failure. The maximum principal stress and the uniaxial compressive strength of the surrounding rock are two core physical quantities that determine the deformation and failure characteristics of the surrounding rock.

[0046] For example, monitoring points are extracted separately. i With monitoring points j Maximum principal stress of surrounding rock after eliminating dimensional influence and and the uniaxial compressive strength of the surrounding rock and Furthermore, the ratio of principal stress to compressive strength at each monitoring point is calculated, i.e., the surrounding rock strength-stress ratio. This ratio is a quantitative basis for determining whether a rock mass is prone to large deformation or rockburst. It directly quantifies the antagonistic balance between the rock mass's own bearing capacity and external geostress, thus determining the risk of large deformation in the surrounding rock of the area. Calculating this ratio transforms static geological survey data into a dimensionless mechanical index that directly reflects the dynamic deformation potential of the rock mass, providing a physical basis for subsequently quantifying the differences in geological characteristics between monitoring points.

[0047] Step 1023: Determine the geological difference penalty term based on the difference between the strength stress ratio of the surrounding rock at any two monitoring points.

[0048] The core reason for the deformation of the surrounding rock in underground caverns is the imbalance between the release of in-situ stress and the bearing capacity of the rock mass. Two spatial sections with similar strength-stress ratios of the surrounding rock often exhibit a high degree of similarity and spatial correlation in their stress state, deformation degree, and deformation evolution trend.

[0049] Specifically, monitoring points can be calculated. i The ratio of surrounding rock strength to stress and monitoring points j The difference between the strength and stress ratio of the surrounding rock is used to determine the geological difference penalty. For example, this geological difference penalty can be calculated using the following formula: in, Indicates a penalty for geological differences. Indicates monitoring point i The corresponding maximum principal stress of the surrounding rock at that location, This represents the maximum principal stress in the surrounding rock at monitoring point j. Indicates monitoring point i The corresponding uniaxial compressive strength of the surrounding rock at that location. Indicates monitoring point j The corresponding uniaxial compressive strength of the surrounding rock at that location. This indicates the introduced correction factor, which is used to avoid the geological difference penalty term being 0. This correction factor can be adjusted according to the actual situation, and this disclosure does not impose any special restrictions.

[0050] This geological difference penalty term is used to quantitatively assess the degree of heterogeneity between two monitoring points in terms of their inherent geological environment and subsequent stress-induced damage. The closer the stress-to-strength ratio of the surrounding rock at the two monitoring points is, the smaller the value of the penalty term, indicating that the stress state and deformation evolution trend of the surrounding rock are more similar. Conversely, when one point is located in a high-stress, low-strength soft rock zone with large deformation, while the other point is located in a low-stress, high-strength hard rock zone with stable conditions, the stress ratio difference between the two is large, and the value of this penalty term will increase synchronously.

[0051] Step 1024: Determine the collaborative metric distance based on the product of distance and geological difference penalty term.

[0052] Since the distance calculated above only reflects the geometric topological relationship between monitoring points, in actual engineering, two points that are close together may be located in a high-stress hard rock brittle fracture zone and a low-stress intact and stable zone, respectively. If this is used as the basis for clustering, subsequent prediction models will be unable to detect abrupt changes in the mechanical properties of the surrounding rock. Therefore, this distance can be used as a benchmark distance, and the collaborative metric distance can be determined based on the product of this distance and the geological difference penalty term. This can be determined as follows: in, Represents any two monitoring points i and monitoring points j The collaborative metric distance between them Represents any two monitoring points i and monitoring points j The physical distance between them This indicates a penalty for geological differences.

[0053] By calculating the product of distance and geological difference penalty, an explicit geological correction is applied to the originally singular distance. If two monitoring points are physically close but have vastly different geological conditions, the resulting co-metric distance will be significantly stretched, preventing subsequent clustering algorithms from incorrectly classifying them into the same cluster. Conversely, if two points are physically close and have extremely similar geomechanical states, the co-metric distance will remain small, allowing subsequent clustering algorithms to cluster them together. This ensures that the final co-metric distance accurately reflects the actual stress and deformation evolution of the surrounding rock in underground caverns.

[0054] In the exemplary implementation of this disclosure, after obtaining the collaborative metric distance between each monitoring point, spatial clustering can be performed on each monitoring point based on this distance. Exemplarily, the following minimization objective function can be constructed. J Perform clustering optimization: in, C The total number of clusters. For the first c Within the cluster, the th i The collaborative metric distance between each monitoring point and the centroid of the cluster. Let c be the total number of monitoring points within the c-th cluster. For the first c Geological spatial synergy weights of each cluster.

[0055] In the exemplary implementation of this disclosure, the geological spatial collaborative weights corresponding to each type of cluster are determined based on the temporal deformation data of monitoring points in each type of cluster, specifically including steps 1025-1027: Step 1025: Based on the time-series deformation data of monitoring points in various clusters, calculate the intra-cluster deformation standard deviation for each type of cluster.

[0056] For example, the first c The standard deviation of intra-cluster deformation for a given cluster can be calculated as follows: in, Indicates the first c The standard deviation of intra-cluster deformation for each cluster. For the first c The total number of monitoring points within each cluster; For the first c Within the cluster, the th i The measured values ​​of time-series deformation data corresponding to each monitoring point; For the first c The arithmetic mean of the time-series deformation data of all monitoring points within a cluster.

[0057] Step 1026: The maximum value among the intra-cluster deformation standard deviations of various clusters is determined as the global deformation standard deviation.

[0058] After calculating the intra-cluster deformation standard deviation for each cluster, these standard deviation values ​​can be iterated through, and the maximum value can be extracted and defined as the global deformation standard deviation. This global deformation standard deviation characterizes the fluctuation level of the most volatile and geologically challenging cluster within the current overall spatial range of the target underground cavern.

[0059] Step 1027: The ratio of the intra-cluster deformation standard deviation to the global deformation standard deviation of each type of cluster is determined as the geological spatial collaborative weight corresponding to each type of cluster.

[0060] Based on the local fluctuation characteristics of each cluster and the global extreme fluctuation characteristics, the final determination can be made. c The geological spatial synergy weight of each cluster can be determined by the following formula: in, Indicates the first c Geological spatial synergy weights of individual clusters, Indicates the first c The standard deviation of intra-cluster deformation of each cluster. This represents the global deformation standard deviation.

[0061] Through the above method, this disclosure transforms the degree of deformation in different sections of the underground cavern into specific, calculable weight values. Mathematically, this ratio naturally and smoothly normalizes all weight values, placing them within the (0,1] interval, thus avoiding gradient explosion caused by weight imbalance during subsequent model training. Secondly, from an engineering geological perspective, the ratio... The closer the value is to 1, the closer the deformation characteristics of the cluster are to the most dangerous state globally. Outputting this collaborative weight, which reflects the risk of geological fluctuations, into the subsequent prediction model can explicitly give high-risk monitoring points higher attention, thus laying a solid quantitative foundation for improving the geological interpretability of the prediction model and the prediction accuracy of local high-risk sections.

[0062] In the exemplary embodiments of this disclosure, due to the complexity of the geological conditions of underground caverns, the clustering algorithm may, during the initial iteration, forcibly group monitoring points with significantly different geological attributes into the same cluster due to parameter settings or local extrema issues. To ensure the physical rationality of the clustering, after clustering is completed, non-parametric verification methods such as kernel density estimation can be used to perform probability density fitting on the characteristic data of all monitoring points within the cluster, such as the collaborative metric distance distribution or deformation fluctuation distribution, to calculate the probability density distribution curves of each cluster. Furthermore, the morphological characteristics of the calculated probability density distribution curves are quantitatively evaluated to determine whether the probability density distribution curves of each cluster exhibit a unimodal distribution, and to assess whether their peak value is greater than or equal to a preset threshold. For example, the preset threshold can be set to 0.8 based on engineering experience.

[0063] If the probability density distribution curve shows a concentrated single-peak distribution with a high peak value, it indicates that the monitoring points contained in this cluster are highly concentrated in terms of spatial location and geomechanical characteristics, and have strong internal consistency. If the curve shows a bimodal / multimodal distribution or a low peak value, it indicates that the cluster actually contains geological structures with significantly different surrounding rock stress states and failure tendencies.

[0064] Based on the above evaluation results, if the probability density distribution curves of all clusters exhibit a unimodal distribution and the peak value is greater than or equal to a preset threshold, the current spatial clustering result is deemed valid, and the category labels of each cluster are ultimately determined. These labels serve as discrete input features characterizing spatial geological heterogeneity and are output to the next feature construction stage. Conversely, if there are clusters that do not meet the above conditions, the current clustering is deemed invalid, triggering parameter correction. The optimal number of clusters is recalculated, or the cluster centers are adjusted, and the spatial clustering step is re-executed. The category label uniquely identifies the geological risk zone corresponding to each monitoring point. Specifically, since the category label is generated by clustering using a co-metric distance that integrates distance and geological difference penalty terms, it represents not only the geometric proximity of each monitoring point in physical space but also the homogeneity of each monitoring point in terms of surrounding rock stress state and deformation evolution trend. For example, monitoring points with the same category label can be considered to belong to a specific geological risk zone. Such geological risk zones may include high-risk zones, low-risk zones, or medium-risk zones, or they may be high-stress soft rock deformation zones, low-stress hard rock stability zones, high-stress hard rock brittle fracture zones, medium-stress transition zones, etc. The specifics can be determined based on the actual situation, and this disclosure does not impose any special limitations.

[0065] This category label transforms continuous and complex underlying geological and spatial parameters into discrete features, and inputs them as explicit prior data into subsequent prediction models. This guides the model to perceive and match the geological zoning logic of the monitoring points in advance when fitting continuous deformation values, thereby breaking the black box state of traditional models blindly fitting data and providing engineering theoretical support and interpretability for the final prediction results.

[0066] By implementing this inspection step, this disclosure achieves quantitative control over the quality of geological spatial zoning, eliminates unreasonable divisions lacking physical basis, and ensures that the final determined category labels can truly reflect the differences in the stress state and failure tendency of the surrounding rock in each section of the underground cavern, thus giving the prediction model engineering geological interpretability.

[0067] Furthermore, after completing spatial clustering and determining the category labels of each cluster, the centroid data of each cluster can be extracted and saved. This centroid data includes the spatial coordinates of each cluster at the spatial cluster center and geological feature data, such as the average maximum principal stress and average uniaxial compressive strength of the surrounding rock, to be used in the model inference stage to determine the category of the nodes to be predicted, ensuring that the clustering logic in the prediction stage and the training stage remains strictly consistent.

[0068] Continue to refer to Figure 1 As shown, in step 103, an input feature set is constructed based on the multi-source feature data of each monitoring point and the corresponding category label.

[0069] In this exemplary embodiment, the multi-source feature data of each monitoring point and its corresponding category label can be sequentially concatenated to obtain the input feature vector of each monitoring point. For any monitoring point in the target underground cavern, the spatial coordinate data, geological feature data, temporal deformation data, and construction feature data obtained in the previous steps, as well as the unique category label assigned to the monitoring point after spatial clustering, are extracted. At the data structure level, the algorithm performs a one-dimensional sequence concatenation operation along the feature dimension direction with the discrete label data representing spatial geological zoning attributes and the continuous numerical data representing the engineering physical state. After completing the feature sequence concatenation of all individual monitoring points, the input feature vectors of all valid monitoring points contained in the target underground cavern are collected and summarized. The one-dimensional input feature vectors are arranged in an ordered row order to construct a two-dimensional structured feature matrix. This two-dimensional structured feature matrix is ​​the final output input feature set and is directly used as the underlying input data for subsequent prediction model training.

[0070] The above method integrates complex, discrete, and heterogeneous multi-source engineering data of underground caverns with spatial category labels generated by the algorithm. By employing sequence concatenation and matrix assemblies, not only is the original engineering physical meaning of each dimension parameter fully preserved, but the computational barriers between different data modalities are also completely eliminated, transforming it into a unified standard format that can be directly and efficiently read by the subsequent initial gradient boosting tree model.

[0071] In step 104, the weighted loss function in the pre-constructed initial gradient boosting tree model is determined based on the geological spatial collaborative weights of various clusters, and the initial gradient boosting tree model is iteratively trained using the input feature set and the weighted loss function to obtain the cavern deformation prediction model.

[0072] In the exemplary implementation of this disclosure, conventional machine learning prediction models typically use a globally uniform standard mean squared error as the loss function during the training phase. Under this error penalty mechanism, because the amount of monitoring data in low-risk areas is much larger than that in high-risk areas, the model is easily affected by the monitoring data in low-risk areas during the iteration process. This causes the prediction errors of a few high-risk areas to be diluted or masked, leading to underfitting of the model in high-risk areas. To address this, this disclosure proposes a method for constructing a weighted loss function driven by geological spatial collaborative weights.

[0073] Specifically, the weighted loss function in the pre-constructed initial gradient boosting tree model is determined based on the geological spatial collaborative weights of various clusters, including the following steps: First, obtain the predicted and actual values ​​of the monitoring points within each cluster using the initial gradient boosting tree model.

[0074] In the initial gradient boosting tree model, such as the CatBoost model, during a single forward propagation iteration, the input feature set constructed in step 103 is input into the model, and the model will output the corresponding prediction results. Based on the pre-divided spatial cluster structure, the algorithm will extract the corresponding true labels and prediction results according to the clusters.

[0075] Secondly, based on the predicted and actual values, the mean squared error of the initial gradient boosting tree model in each cluster is calculated. The predicted and actual values ​​can be either the cumulative deformation or the deformation rate, which can be determined based on the specific circumstances.

[0076] Finally, the prediction mean square error of each cluster is summed with the corresponding geological spatial collaborative weights to obtain the weighted loss function.

[0077] In the exemplary implementation of this disclosure, the geological spatial synergy weights of each independent cluster extracted in the previous steps are introduced. The mean squared errors of each cluster are weighted and summed to construct the final weighted loss function, as shown in the following formula: in, This represents the value of the weighted loss function. C This represents the total number of clusters. For the first c The total number of monitoring points within each cluster. For the first c Within the cluster, the th i Predicted values ​​of time-series deformation data for each monitoring point For the first c Within the cluster, the th i The measured values ​​of time-series deformation data at each monitoring point, i.e., the true values. For the first c The geological spatial collaborative weights corresponding to each cluster are such that the larger the weight value, the greater the contribution of the prediction error of the corresponding cluster to the global loss function. This guides the model to prioritize parameter optimization and error correction for high-weight segments during iterative training, thereby improving the prediction accuracy of high-risk areas.

[0078] Through this step, the weighted loss function will dominate the calculation and update direction of the global gradient during subsequent model parameter updates. For clusters corresponding to high-risk areas with complex geological conditions and severe deformation, the collaborative weight is large, and the resulting prediction error will be amplified many times over, thus transmitting a stronger gradient penalty signal to the prediction model and forcibly guiding the node splitting of the prediction model to tilt towards fitting high-risk features. Conversely, for clusters corresponding to low-risk areas with stable deformation, the contribution of their error to the overall loss is reasonably converged. This approach, which combines the category labels obtained from geological spatial clustering with the gradient optimization mechanism of the underlying algorithm, not only avoids the problem of local high-risk geological features being submerged by massive amounts of globally stable data, achieving high-precision adaptive prediction under complex conditions, but also makes each parameter update within the model have engineering geological guidance, breaking the black-box barrier of traditional data-driven models and giving the final prediction decision strong theoretical transparency and engineering interpretability.

[0079] Corresponding to the embodiments of the aforementioned methods, this disclosure also provides embodiments of a cavern deformation prediction method based on geological spatial collaborative clustering, such as... Figure 2 Steps 201-203 are shown.

[0080] Step 201: Obtain multi-source feature data of the nodes at the locations to be predicted in the target underground cavern.

[0081] In the example implementation of this disclosure, the node to be predicted can be a newly added monitoring point at the engineering site, or a spatiotemporal extension node of an existing monitoring point at a future time.

[0082] Specifically, multi-source feature data of the node to be predicted is acquired. This multi-source feature data includes at least the spatial coordinate data and geological feature data of the node. The spatial coordinate data represents the physical location of the node; the geological feature data includes the maximum principal stress and uniaxial compressive strength of the surrounding rock obtained from field measurements or geological interpolation. Furthermore, to eliminate the influence of dimensions and ensure the accuracy of model inference, after acquiring the original multi-source feature data, dimensionless processing is performed on the multi-source feature data of the node to be predicted using the same standardized parameters as in the model training phase (i.e., reusing the mean and standard deviation of each feature calculated from the training set). By reusing the preprocessing parameters of the training set, the data distribution shift between the prediction and training phases can be effectively avoided, ensuring that the input features are within the numerical space that the cavern deformation prediction model can stably resolve.

[0083] Step 202: Construct the input vector to be predicted based on the multi-source feature data of the node to be predicted and its corresponding category label.

[0084] In the exemplary implementation of this disclosure, the pre-trained cavern deformation prediction model is a prediction algorithm architecture with geological spatial perception capabilities generated by iterative optimization of the weighted loss function based on the training method described above.

[0085] For example, during the model inference phase, the geological spatial collaborative clustering criteria generated during the model training phase can be reused to determine the category of the node to be predicted. This involves first extracting the centroid data of each cluster after spatial clustering during the model training phase. This centroid data includes the spatial coordinates of each centroid and geological feature data. Then, using the collaborative metric distance calculation logic defined in the preceding steps, the collaborative metric distance between the node to be predicted and each cluster centroid is calculated.

[0086] Specifically, the cluster identifier corresponding to the minimum value in the collaborative metric distance can be determined as the category label of the node to be predicted. Then, the multi-source feature data of the node to be predicted is concatenated with its determined category label to obtain the input vector for prediction. The category label is used to characterize the geological risk zone where the node to be predicted is located.

[0087] Step 203: Input the input vector to be predicted into the pre-trained cavern deformation prediction model to obtain the deformation prediction value of the node at the location to be predicted.

[0088] For example, during the prediction execution phase of model inference, the concatenated input vector to be predicted can be input into the data input layer of the cavern deformation prediction model. The model will then perform regression calculations based on the input vector to calculate the specific deformation prediction values ​​for the nodes at the predicted locations.

[0089] The cavern deformation prediction model is a model with geological spatial perception capabilities obtained through iterative optimization based on the aforementioned weighted loss function. The predicted deformation value includes the cumulative deformation or deformation rate of the node at the predicted location within a preset future period.

[0090] By implementing this step, this application overcomes the blind prediction mode of traditional data-driven algorithms. The category labels output by the model provide geological logic support for the final deformation values. Field engineers can not only obtain accurate deformation prediction values, but also trace the geomechanical background from which the predicted values ​​originated through the synchronously output category labels, such as belonging to the corresponding cluster of high-risk areas. This dual-track output mechanism gives the prediction results strong engineering transparency and theoretical interpretability, improving the model's field credibility and practical value in disaster prevention and mitigation decision-making for underground caverns in water conservancy and hydropower projects.

[0091] Corresponding to the embodiments of the foregoing methods, this disclosure also provides embodiments of the apparatus and the terminal to which it is applied.

[0092] like Figure 3 As shown, Figure 3 This is a schematic diagram of a training device for a cavern deformation prediction model based on geological spatial collaborative clustering, according to an exemplary embodiment of this disclosure. The device includes: a first acquisition module 310, a determination module 320, a construction module 330, and a training module 340. The first acquisition module 310 is used to acquire multi-source feature data of each monitoring point in the target underground cavern. The multi-source feature data includes at least the spatial coordinate data, geological feature data and temporal deformation data of each monitoring point. The determination module 320 is used to calculate the collaborative metric distance between any two monitoring points based on the spatial coordinate data and geological feature data of each monitoring point, and to perform spatial clustering of each monitoring point according to the collaborative metric distance to obtain multiple clusters and category labels of each cluster. Based on the temporal deformation data of the monitoring points in each cluster, the geological spatial collaborative weights corresponding to each cluster are determined. Module 330 is used to construct an input feature set based on the multi-source feature data and corresponding category labels of each monitoring point; Training module 340 is used to determine the weighted loss function in the pre-constructed initial gradient boosting tree model based on the geological spatial collaborative weights of various clusters, and to iteratively train the initial gradient boosting tree model using the input feature set and the weighted loss function to obtain the cavern deformation prediction model.

[0093] like Figure 4 As shown, Figure 4 This is a schematic diagram of a cavern deformation prediction device based on geological spatial collaborative clustering, according to an exemplary embodiment of this disclosure. The device includes: a second acquisition module 410, an input module 420, and a prediction module 430. The second acquisition module 410 is used to acquire multi-source feature data of the node to be predicted in the target underground cavern. The multi-source feature data includes at least the spatial coordinate data and geological feature data of the node to be predicted. The input module 420 is used to construct the input vector to be predicted based on the multi-source feature data of the location node to be predicted and the corresponding category label; Prediction module 430 is used to input the input vector to be predicted into a pre-trained cavern deformation prediction model to obtain the deformation prediction value of the node to be predicted. The cavern deformation prediction model was trained using the aforementioned training method for the cavern deformation prediction model based on geological spatial collaborative clustering.

[0094] The training device or embodiment of the cavern deformation prediction model based on geological spatial collaborative clustering disclosed herein can be applied to computer equipment, such as servers or terminal devices. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by its processor reading the corresponding computer program instructions from non-volatile memory into memory and running them. From a hardware perspective, such as... Figure 5 The diagram shown is a hardware structure diagram of a training device for a cavern deformation prediction model based on geological spatial collaborative clustering, or a computer device containing such a device, according to an embodiment of this disclosure. Except for... Figure 5 In addition to the processor 510, memory 530, network interface 520, and non-volatile memory 540 shown, the server or electronic device where the training device for the cave deformation prediction model based on geological spatial co-clustering or the cave deformation prediction device based on geological spatial co-clustering is located in the embodiment may also include other hardware depending on the actual function of the computer device, which will not be described in detail here.

[0095] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0096] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0097] In one example embodiment of this disclosure, a computer-readable storage medium is also provided, on which a program product capable of implementing the methods described above is stored. In some possible embodiments, various aspects of this disclosure may also be implemented as a program product comprising program code that, when the program product is run on a terminal device, causes the terminal device to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure.

[0098] The foregoing has described specific embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0099] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention applied herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not claimed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0100] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

[0101] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A training method for a cavern deformation prediction model based on geological spatial collaborative clustering, characterized in that, The method includes: Acquire multi-source feature data of each monitoring point in the target underground cavern. The multi-source feature data includes at least the spatial coordinate data, geological feature data, and temporal deformation data of each monitoring point. Based on the spatial coordinate data of each monitoring point and the geological feature data, the collaborative metric distance between any two monitoring points is calculated, and each monitoring point is spatially clustered according to the collaborative metric distance to obtain multiple clusters and category labels for each cluster. Based on the temporal deformation data of the monitoring points in each cluster, the geological spatial collaborative weights corresponding to each cluster are determined. Based on the multi-source feature data and corresponding category labels of each monitoring point, an input feature set is constructed; The weighted loss function in the pre-constructed initial gradient boosting tree model is determined based on the geological spatial collaborative weights of each of the aforementioned clusters. The initial gradient boosting tree model is then iteratively trained using the input feature set and the weighted loss function to obtain the cavern deformation prediction model.

2. The method according to claim 1, characterized in that, The multi-source feature data also includes construction feature data for each of the monitoring points; The spatial coordinate data is used to indicate the three-dimensional spatial position of each monitoring point in the target underground cavern; the geological feature data includes at least the maximum principal stress and uniaxial compressive strength of the surrounding rock in the area where each monitoring point is located; the time-series deformation data includes the cumulative deformation of each monitoring point in the historical observation period, and / or the deformation rate; the construction feature data includes the excavation disturbance parameters and support compensation parameters of the area where each monitoring point is located.

3. The method according to claim 1, characterized in that, The calculation of the collaborative metric distance between any two monitoring points based on the spatial coordinate data and geological feature data of each monitoring point includes: The distance between any two monitoring points is determined based on the spatial coordinate data of each monitoring point. Based on the geological characteristic data of each monitoring point, the ratio of the maximum principal stress of the surrounding rock to the uniaxial compressive strength of the surrounding rock at each monitoring point is determined to obtain the strength-stress ratio of the surrounding rock, and a geological difference penalty term is determined based on the difference between the strength-stress ratios of the surrounding rock at any two monitoring points. The collaborative measurement distance between any two monitoring points is determined based on the product of the distance and the geological difference penalty term.

4. The method according to claim 1, characterized in that, The determination of the geological spatial collaborative weights corresponding to each cluster based on the temporal deformation data of the monitoring points in each cluster includes: Based on the time-series deformation data of the monitoring points in each of the aforementioned clusters, the intra-cluster deformation standard deviation of each of the aforementioned clusters is calculated respectively; The maximum value among the intra-cluster deformation standard deviations of each of the aforementioned clusters is determined as the global deformation standard deviation; The ratio of the intra-cluster deformation standard deviation to the global deformation standard deviation of each cluster is determined as the geological spatial collaborative weight corresponding to each cluster.

5. The method according to claim 1, characterized in that, After performing spatial clustering of each monitoring point based on the collaborative metric distance to obtain multiple clusters and category labels for each cluster, the method further includes: Calculate the probability density distribution curves for each of the aforementioned clusters; If the probability density distribution curve of each cluster exhibits a single-peak distribution and the peak value is greater than or equal to a preset threshold, then the category label of each cluster is determined.

6. The method according to claim 1, characterized in that, The input feature set is constructed based on the multi-source feature data and corresponding category labels of each monitoring point, including: The multi-source feature data of each monitoring point and its corresponding category label are sequentially concatenated to obtain the input feature vector of each monitoring point. The input feature set is obtained based on the input feature vectors of each monitoring point.

7. The method according to claim 1, characterized in that, The determination of the weighted loss function in the pre-constructed initial gradient boosting tree model based on the geological spatial collaborative weights of each of the aforementioned clusters includes: Obtain the predicted and actual values ​​of the monitoring points within each cluster using the initial gradient boosting tree model; Based on the predicted and actual values, the mean square error of the initial gradient boosting tree model in each of the clusters is calculated. The weighted loss function is obtained by weighting and summing the prediction mean square error of each cluster with the corresponding geological spatial collaborative weight.

8. A method for predicting cavern deformation based on geological spatial collaborative clustering, characterized in that, The method includes: Acquire multi-source feature data of the node to be predicted in the target underground cavern, wherein the multi-source feature data includes at least the spatial coordinate data and geological feature data of the node to be predicted; Based on the multi-source feature data and corresponding category labels of the nodes to be predicted, construct the input vector to be predicted; The input vector to be predicted is input into the pre-trained cavern deformation prediction model to obtain the deformation prediction value of the node at the location to be predicted. The cavern deformation prediction model is trained based on the training method of the cavern deformation prediction model based on geological spatial collaborative clustering as described in any one of claims 1 to 7.

9. A training device for a cavern deformation prediction model based on geological spatial collaborative clustering, characterized in that, include: The first acquisition module is used to acquire multi-source feature data of each monitoring point in the target underground cavern. The multi-source feature data includes at least the spatial coordinate data, geological feature data and temporal deformation data of each monitoring point. The determination module is used to calculate the collaborative metric distance between any two monitoring points based on the spatial coordinate data of each monitoring point and the geological feature data, and to perform spatial clustering of each monitoring point according to the collaborative metric distance to obtain multiple clusters and category labels of each cluster. Based on the temporal deformation data of the monitoring points in each cluster, the geological spatial collaborative weights corresponding to each cluster are determined. The construction module is used to construct an input feature set based on the multi-source feature data of each monitoring point and the corresponding category label; The training module is used to determine the weighted loss function in the pre-constructed initial gradient boosting tree model based on the geological spatial collaborative weights of each of the aforementioned clusters, and to iteratively train the initial gradient boosting tree model using the input feature set and the weighted loss function to obtain a cavern deformation prediction model.

10. A cavern deformation prediction device based on geological spatial collaborative clustering, characterized in that, include: The second acquisition module is used to acquire multi-source feature data of the node to be predicted in the target underground cavern. The multi-source feature data includes at least the spatial coordinate data and geological feature data of the node to be predicted. The input module is used to construct the input vector to be predicted based on the multi-source feature data of the node to be predicted and the corresponding category label; The prediction module is used to input the input vector to be predicted into the pre-trained cavern deformation prediction model to obtain the deformation prediction value of the node to be predicted. The cavern deformation prediction model is trained based on the training method of the cavern deformation prediction model based on geological spatial collaborative clustering as described in any one of claims 1 to 7.

Images