An underground cavern group excavation sequence optimization method, device and processing equipment

Abstract

The application provides a method and device for optimizing a sequence of underground cavern group excavation and processing equipment, which is based on an underground cavern group graph structure, a physical response prediction model, a reinforcement learning excavation environment, a reward function and an excavation decision intelligent agent, builds a set of excavation sequence optimization design architecture capable of uniformly expressing the spatial topology, dynamic excavation state and engineering constraints of the underground cavern group and capable of taking into account the construction efficiency and the stability of surrounding rock, thereby good technical defects in the prior art, such as large underground cavern group excavation sequence scheme space, high evaluation cost, insufficient representation mode, and difficulty in jointly optimizing construction feasibility and surrounding rock stability, can be improved, and accurate, efficient and low-cost scheme design is provided for underground cavern group excavation work.

Description

Technical Field

[0001] This application relates to the field of geology, specifically to a method, apparatus, and processing equipment for optimizing the excavation sequence of underground cavern groups. Background Technology

[0002] Large-scale hydropower stations, underground energy storage projects, underground mines, and deep transportation projects typically include multiple interconnected, large-span underground caverns. Compared to single tunnels or regular underground structures, underground cavern groups often have more complex spatial relationships, and the excavation responses of different caverns and different layers influence each other, making the excavation sequence a crucial factor affecting the stability of the surrounding rock and construction safety.

[0003] In actual construction, large-span underground caverns typically cannot be excavated in one go. Instead, they require layered, zoned, and phased excavation based on factors such as cavern function, surrounding rock conditions, and construction organization. The design of the excavation sequence must satisfy various engineering constraints. For example, the excavation sequence is often from top to bottom to avoid instability; interconnected or adjacent caverns should avoid simultaneous excavation; the timing of excavation for auxiliary caverns such as busbar tunnels needs to be determined based on the excavation progress of the main powerhouse or main transformer room; and to meet schedule requirements, multiple cavern groups need to be advanced in parallel under safety constraints. Therefore, optimizing the excavation sequence of underground cavern groups is a large-scale combinatorial optimization problem influenced by spatial topology, mechanical response, and dynamic engineering constraints.

[0004] In current design processes, excavation schemes for underground cavern groups are typically determined based on engineering experience, engineering analogies, and 3D numerical simulations. While these methods can provide a reference for engineering design, candidate schemes are usually pre-defined manually and are difficult to cover the vast excavation sequence space. Performing 3D numerical simulations on a large number of candidate sequences one by one would be computationally expensive and difficult to meet the needs for rapid scheme generation and large-scale search.

[0005] Therefore, a technical solution is needed that can uniformly express the spatial topology, dynamic excavation state, and engineering constraints of underground cavern groups, and can take into account both construction efficiency and surrounding rock stability for the optimal design of excavation sequences. Summary of the Invention

[0006] This application provides a method, apparatus, and processing equipment for optimizing the excavation sequence of underground cavern groups. Based on the underground cavern group graph structure, physical response prediction model, reinforcement learning excavation environment, reward function, and excavation decision-making agent, it establishes an excavation sequence optimization design architecture that can uniformly express the spatial topology, dynamic excavation state, and engineering constraints of underground cavern groups, while taking into account construction efficiency and surrounding rock stability. This can effectively improve the technical defects of existing underground cavern group excavation sequence schemes, such as large space, high evaluation cost, insufficient representation mode, and difficulty in jointly optimizing construction feasibility and surrounding rock stability. It provides accurate, efficient, and low-cost scheme design for underground cavern group excavation work.

[0007] Firstly, this application provides a method for optimizing the excavation sequence of underground cavern groups, the method comprising: Under the local operation strategy, obtain the engineering data of the underground cavern group that needs to be optimized; Based on engineering data, a map structure of the underground cavern group corresponding to the underground cavern group to be optimized is constructed. In the map structure of the underground cavern group, the nodes correspond to the excavation layered units of the underground cavern group to be optimized, and the edges correspond to the spatial contact relationship between the excavation layered units. Configure a physical response prediction model to output corresponding predicted values ​​of the surrounding rock physical response based on the excavation sequence and underground cavern group diagram structure input to the model; Configure a reinforcement learning mining environment that includes a state space, action space, and reward function; In a reinforcement learning excavation environment, an excavation decision agent is trained to output joint excavation actions based on the state space. During the training process, the reward is evaluated by combining the predicted value of the surrounding rock physical response output by the physical response prediction model. Among the multiple candidate excavation sequences generated by the excavation decision-making agent, the target excavation sequence that meets the preset optimization conditions is selected as the excavation sequence optimization design result of the underground cavern group to be optimized.

[0008] Secondly, this application provides an optimization device for the excavation sequence of underground cavern groups, the device comprising: The data acquisition unit is used to acquire engineering data of the current underground cavern group to be optimized under the local operation strategy; The graph structure building unit is used to construct the underground cavern group graph structure corresponding to the underground cavern group to be optimized based on engineering data. In the underground cavern group graph structure, the nodes correspond to the excavation layer units of the underground cavern group to be optimized, and the edges correspond to the spatial contact relationship between the excavation layer units. The model configuration unit is used to configure a physical response prediction model that outputs corresponding predicted values ​​of the surrounding rock physical response based on the excavation sequence and underground cavern group diagram structure input to the model. The environment configuration unit is used to configure the reinforcement learning mining environment, including the state space, action space, and reward function. The agent training unit is used to train an excavation decision agent in a reinforcement learning excavation environment to output joint excavation actions based on the state space. During the training process, the reward is evaluated by combining the predicted value of the surrounding rock physical response output by the physical response prediction model. The optimization output unit is used to select the target excavation sequence that meets the preset optimization conditions from multiple sets of candidate excavation sequences generated by the excavation decision agent as the excavation sequence optimization design result of the underground cavern group to be optimized.

[0009] Thirdly, this application provides a processing device, including a processor and a memory, wherein a computer program is stored in the memory, and the processor executes the method provided in the first aspect of this application when it invokes the computer program in the memory.

[0010] Fourthly, this application provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute the method provided in the first aspect of this application.

[0011] From the above, it can be concluded that this application has the following beneficial effects: Under the objective of optimizing the excavation sequence of underground cavern groups, this application, based on the underground cavern group graph structure, physical response prediction model, reinforcement learning excavation environment, reward function, and excavation decision-making agent, establishes an excavation sequence optimization design architecture that can uniformly express the spatial topology, dynamic excavation state, and engineering constraints of underground cavern groups, while taking into account both construction efficiency and surrounding rock stability. This can effectively improve the technical defects of existing underground cavern group excavation sequence schemes, such as large space, high evaluation cost, insufficient representation mode, and difficulty in jointly optimizing construction feasibility and surrounding rock stability, providing accurate, efficient, and low-cost scheme design for underground cavern group excavation work. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating an optimization method for the excavation sequence of underground cavern groups in this application. Figure 2 This is a schematic diagram representing the structure of the diagram in this application; Figure 3This is a schematic diagram of the overall model architecture of the physical response prediction model of this application; Figure 4 This is a schematic diagram illustrating the interaction between the reinforcement learning excavation environment and the excavation decision-making agent in this application. Figure 5 This is a schematic diagram illustrating how the two-stage reward value changes during the training process in this application. Figure 6 This is a schematic diagram of a structure of the underground cavern group excavation sequence optimization device of this application; Figure 7 This is a schematic diagram of one type of processing equipment used in this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. The naming or numbering of steps appearing in this application does not imply that the steps in the method flow must be performed in the chronological / logical order indicated by the naming or numbering. The execution order of named or numbered process steps can be changed according to the desired technical purpose, as long as the same or similar technical effect is achieved.

[0016] The module division described in this application is a logical division. In practical applications, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual coupling, direct coupling, or communication connections may be through interfaces, and the indirect coupling or communication connections between modules may be electrical or other similar forms, none of which are limited in this application. Moreover, the modules or sub-modules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed across multiple circuit modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution in this application.

[0017] Before introducing the method for optimizing the excavation sequence of underground cavern groups provided in this application, we will first introduce the background information involved in this application.

[0018] The method, apparatus, and computer-readable storage medium for optimizing the excavation sequence of underground cavern groups provided in this application can be applied to processing equipment. Based on the underground cavern group graph structure, physical response prediction model, reinforcement learning excavation environment, reward function, and excavation decision-making agent, an excavation sequence optimization design architecture is built that can uniformly express the spatial topology, dynamic excavation state, and engineering constraints of underground cavern groups, while taking into account construction efficiency and surrounding rock stability. This can effectively improve the technical defects of existing underground cavern group excavation sequence schemes, such as large space, high evaluation cost, insufficient representation mode, and difficulty in jointly optimizing construction feasibility and surrounding rock stability, and provide accurate, efficient, and low-cost scheme design for underground cavern group excavation work.

[0019] The method for optimizing the excavation sequence of underground cavern groups mentioned in this application can be implemented by an underground cavern group excavation sequence optimization device, or by different types of processing devices such as servers, physical hosts, or user equipment (UE) that integrate the underground cavern group excavation sequence optimization device. The underground cavern group excavation sequence optimization device can be implemented in hardware or software. The UE can specifically be a terminal device such as a smartphone, tablet, laptop, desktop computer, or personal digital assistant (PDA). The processing devices can be configured in a device cluster.

[0020] It is understandable that the proposed solution is usually based on existing data or data that has already been collected. Therefore, the processing equipment that executes the proposed method for optimizing the excavation sequence of underground cavern groups or is equipped with the corresponding application service of the proposed method for optimizing the excavation sequence of underground cavern groups usually only needs to meet the required data processing capabilities. The specific equipment type and equipment deployment form are quite flexible.

[0021] If the direct collection of existing data mentioned above is also involved, then further hardware and software adaptation configurations of the processing equipment are obviously required to enable the corresponding data collection capabilities. For example, if it is necessary to collect engineering data or other input data involved in this application in real time, the corresponding data collection equipment can be incorporated into the equipment cluster of the processing equipment, or the processing equipment itself can be the control part of these data collection equipment, or the data collection equipment outside the processing equipment can be triggered by a third party to perform real-time data collection operations.

[0022] In addition, if there is a need to display the processing progress (including the processing results), the processing device itself can be configured with the required display screen (including touch screen) to display the specific content. Of course, the processing device can also display the specific content through an external display device or other devices with a display screen.

[0023] The following section introduces the method for optimizing the excavation sequence of underground cavern groups provided in this application.

[0024] First, refer to Figure 1 , Figure 1 This paper illustrates a flowchart of the underground cavern excavation sequence optimization method of this application. The underground cavern excavation sequence optimization method provided by this application may specifically include the following steps S101 to S106: Step S101: Under the local operation strategy, obtain the engineering data of the underground cavern group to be optimized. Understandably, given the shortcomings of existing methods for large-span underground caverns / cavern groups, which are accompanied by large space for excavation sequence schemes, high evaluation costs, insufficient characterization models, and difficulty in jointly optimizing construction feasibility and surrounding rock stability, the proposed solution can obtain engineering data for the underground cavern group to be optimized.

[0025] The engineering data is mainly used to construct the subsequent underground cavern group graph structure under the condition of introducing a graph structure in this application. Obviously, it may involve specific data types with multiple dimensions / multimodalities.

[0026] In terms of details, the engineering data may include geometric data of the cavern group, excavation layer information, rock mechanics parameters, geostress distribution and engineering constraints, etc. Furthermore, the geometric data of the cavern group may specifically include the spatial structure, size, and connectivity of the underground caverns.

[0027] In terms of details, the geometric data of the cavern complex can be imported from 3D design models, Building Information Modeling (BIM), or Computer-Aided Design (CAD) drawings.

[0028] Cavern groups can be classified according to their function, spatial location, construction working face, or excavation organization method. In different projects, cavern groups may include underground powerhouses, underground storage facilities, access tunnels, drainage tunnels, mining areas, connecting tunnels, or other underground spaces.

[0029] Furthermore, the excavation layer information may specifically include layer thickness, layer elevation, layer center coordinates, the cavern group to which it belongs, and the excavation direction.

[0030] Excavation layer data is mainly used to discretize a group of caverns into multiple excavation layer units. Each excavation layer unit can serve as a minimum excavation decision object.

[0031] The specific layer thickness can be determined based on the construction method, surrounding rock grade, tunnel span, equipment capacity, or construction experience. Each tunnel can use the same or different preset layer thicknesses, that is, the preset layer thicknesses of each tunnel are configured independently.

[0032] Furthermore, the rock mechanics parameters may specifically include rock density, elastic modulus, Poisson's ratio, cohesion, internal friction angle, tensile strength, rock mass classification, fault parameters, and weak interlayer parameters.

[0033] Furthermore, the distribution of geostress can specifically include the magnitude of geostress components and the direction of principal stresses.

[0034] Furthermore, the engineering constraints may specifically include elevation dependence constraints, spatial proximity constraints, excavation capacity constraints, and maximum number of excavation steps constraints.

[0035] It is understood that, regarding the specific engineering data mentioned above, considering that these are existing indicators / characteristics that can be involved in underground cavern excavation operations, the data collection methods can also be directly adopted using existing methods. Therefore, this application has not elaborated on this further. Of course, it should also be understood that how to calculate / obtain specific indicators may involve further optimization design in actual situations, and can be configured according to actual needs.

[0036] Thus, the aforementioned series of specific indicators laid a solid foundation for the subsequent construction of the underground cavern complex structure.

[0037] Step S102: Based on the engineering data, construct the underground cavern group map structure corresponding to the underground cavern group to be optimized. In the underground cavern group map structure, the nodes correspond to the excavation layered units of the underground cavern group to be optimized, and the edges correspond to the spatial contact relationship between the excavation layered units. As mentioned above, this application introduces a graph structure, which aims to express the specific characteristics of underground cavern groups, such as excavation stratification, cavern groups, rock mass parameters, spatial adjacency relationships, and constraint states. Compared with the fixed-dimensional vector expression method in existing methods, it can better describe the irregular spatial topology and excavation coupling relationship of underground cavern groups, and can better serve subsequent related data processing and analysis.

[0038] Furthermore, the node characteristics of nodes in the underground cavern group diagram structure may specifically include at least one of the following elements: excavation status, excavation time, spatial coordinates, excavation volume, cavern group identifier, rock mechanics parameters, ground stress value, number of adjacent nodes, and constraint satisfaction status.

[0039] Among them, the constraint satisfaction state is mainly used for subsequent reinforcement learning decisions, specifically to indicate whether the corresponding excavation layer unit currently meets the pre-excavation conditions, while the preceding series of indicators are mainly used for the physical response prediction model.

[0040] For the edges in the underground cavern group diagram structure, they can be divided into intra-group connecting edges and inter-group connecting edges. Intra-group connecting edges represent the vertical or front-to-back connection relationship between adjacent layers within the same cavern group, while inter-group connecting edges represent the contact relationship or spatial proximity relationship between different cavern groups. The specific edge weight can be determined based on the contact area, spatial distance, topological connection type, or mechanical coupling strength (or preset engineering experience coefficient) between excavated layer units.

[0041] Taking the contact area as an example, the larger the contact area, the higher the edge weight.

[0042] As an example, you can also refer to this. Figure 2 The diagram shown is a schematic representation of the structure of this application, for a more intuitive understanding.

[0043] Step S103: Configure a physical response prediction model to output corresponding surrounding rock physical response prediction values ​​based on the excavation sequence and underground cavern group diagram structure input to the model; Understandably, this application, based on the previously constructed underground cavern group map structure, introduces machine learning methods (including deep learning methods) to configure a physical response prediction model. This model is used to perform surrounding rock physical response prediction processing based on the excavation sequence and underground cavern group map structure input to the model, and outputs the corresponding surrounding rock physical response prediction value. Specifically, this physical response prediction model serves the training work of the subsequent excavation decision intelligence involved.

[0044] The model configuration work here mainly involves model training, which requires configuring and labeling the corresponding training samples before proceeding with the actual training. Model training typically includes: A training sample is input into the model, which then performs prediction processing to achieve forward propagation. Next, the loss function is calculated based on the model's prediction results and the annotations. The model parameters are then optimized based on the loss function calculation results to achieve backpropagation. In this way, when the model training requirements such as training time, number of training sessions, or prediction accuracy are met, the model training can be completed.

[0045] Once training is complete, the physical response prediction model can be used to quickly evaluate the final-state stability of candidate excavation sequences in subsequent reinforcement learning training or scheme comparison stages.

[0046] Furthermore, the training samples used to train the physical response prediction model may specifically include excavation sequence samples and surrounding rock physical response data samples used as annotations (that is, the true predicted values ​​that the model should theoretically predict).

[0047] The excavation sequence samples can be obtained through historical engineering schemes, manually designed schemes, random feasible schemes, or rule-generated schemes. The excavation sequence samples include a set of excavation layered units arranged according to the excavation steps.

[0048] The physical response data of the surrounding rock can be obtained by three-dimensional finite difference, finite element, discrete element or other numerical simulation software. The physical response data samples of the surrounding rock include surrounding rock deformation, surrounding rock stress, elastic energy, plastic zone, local stress concentration index, plastic zone volume and tensile stress zone volume, etc. Surrounding rock deformation includes the displacement of the sidewalls of the key cavern and the displacement of the arch, surrounding rock stress includes the maximum principal stress, and elastic energy includes total elastic energy and local elastic energy.

[0049] Furthermore, this application also specifically designed an optimization scheme for the specific model architecture of the physical response prediction model.

[0050] For details, please refer to Figure 3The diagram shown illustrates an overall model architecture for the physical response prediction model of this application. This model architecture can include several major components: a sequence feature extraction branch, a static graph feature extraction branch, a dynamic mask graph feature extraction branch, a fusion prediction branch, and a fully connected prediction network. The corresponding components are: 3.1) Sequence Feature Extraction Branch The sequence feature extraction branch is used to extract the corresponding excavation path dependency features (also known as the excavation state sequence in the figure) of the excavation state of the excavation sequence input to the model as the excavation state changes with the excavation step.

[0051] The sequence feature extraction branch specifically receives the sequence input formed by the change of excavation state with each excavation step. For each excavation step, the excavation state of each excavation layer unit is encoded into a state vector (such as a 0 / 1 sequence), which is then mapped through a linear layer and input into a Long Short-Term Memory (LSTM) network, a Bidirectional Long Short-Term Memory (LSTM) network, a Gated Recurrent Unit (GRU), a Transformer encoder, or other types of temporal models to extract excavation path-dependent features, or temporal features.

[0052] 3.2) Static graph feature extraction branch The static graph feature extraction branch is used to extract the corresponding static spatial features (also known as static node features in the graph) of the excavation sequence input to the model based on the underground cavern group graph structure.

[0053] The static graph feature extraction branch specifically receives static graph structure inputs such as rock mechanics parameters, spatial coordinates, cavern group identifiers, and edge connection relationships, and learns the spatial relationships between nodes through a static graph attention network or a graph convolutional network.

[0054] This branch is used to express the spatial distribution of rock mass parameters, the topology of cavern groups, and the adjacent disturbance relationships.

[0055] 3.3) Dynamic Mask Image Feature Extraction Branch The dynamic mask graph feature extraction branch is used to generate node masks or edge masks (also known as excavation mask features in the graph) corresponding to the excavation state of the excavation sequence input to the model. Based on the dynamic changes of the generated masks, the branch extracts the dynamic neighborhood interaction features formed by the change of effective neighborhood relationships over time during the excavation process.

[0056] The dynamic mask map feature extraction branch specifically generates node masks or edge masks based on the current excavation status. For the current edge... When the side When both ends of the node satisfy the preset unexcavated state or effective interaction state, the edge is retained. For use in graph information transmission, otherwise the edges are masked. Or lower the edge By assigning weights to the dynamic graph attention network, the network can effectively extract dynamic neighborhood interaction features formed by the change of effective neighborhood relationships over time during the excavation process.

[0057] 3.4) Fusion Prediction Branch The fusion prediction branch integrates the excavation path dependency features, static spatial features, and dynamic neighborhood interaction features to obtain multimodal fusion features.

[0058] The fusion method can specifically employ techniques such as splicing, multi-head self-attention, weighted summation, or gated fusion. 3.5) Fully Connected Predictive Network The fully connected prediction network maps multimodal fusion features into predicted values ​​of the surrounding rock physical response for output.

[0059] Step S104: Configure a reinforcement learning mining environment including a state space, an action space, and a reward function; Understandably, the reinforcement learning excavation environment constructed in this application is used to model the optimization process of the underground cavern group excavation sequence as a Markov decision process. Correspondingly, the excavation decision agent can receive the current state at each excavation decision step, output a joint excavation action, and the environment executes the action and returns a new state and reward. The interaction mode between the reinforcement learning excavation environment and the excavation decision agent can also be achieved through… Figure 4 The schematic diagram shown here illustrates the interaction between the reinforcement learning excavation environment and the excavation decision-making agent of this application, providing a more vivid understanding.

[0060] For reinforcement learning mining environments, the specific configuration details for the three main aspects—state space, action space, and reward function—are as follows: 4.1) State Space The state space includes graph structure state and global construction state. Graph structure state includes node features, edge connection relationships and edge weights. Global construction state includes excavation progress of each cavern group, remaining proportion of each cavern group, current excavation step proportion, overall excavation progress, safe excavation proportion of each cavern group, overall constraint satisfaction proportion and action mask. 4.2) Action Space The action space consists of the joint excavation actions of multiple cavern groups. Assume that the underground cavern group to be optimized includes k cavern groups, and the actions of each cavern group within one excavation decision step are as follows. The excavation volume level is represented by the preset excavation particle size. The excavation volume level is represented by no excavation, excavating one preset layer, or excavating multiple preset layers (the latter two can also be referred to as at least one excavation). Different cavern groups can have the same or different sets of excavation volume levels, that is, the sets of excavation volume levels of different cavern groups are independently configured. The joint excavation action can be represented as: ; in, This indicates a joint excavation operation. , This indicates the total number of actions involved in the joint excavation. The aforementioned combined action space can represent parallel construction on multiple work surfaces.

[0061] As an example, the excavation volume level can also be represented by the number of excavation layers, such as no excavation, excavating one preset layer, or excavating multiple preset layers; As another example, the excavation volume level can also be expressed based on the excavation volume, excavation length, or excavation advance.

[0062] In addition, it should be noted that this application does not limit the specific number of grades or the excavation volume corresponding to each grade.

[0063] An action mask matrix is ​​generated at each excavation decision step.

[0064] For a certain cavern group, if the number of remaining layers is insufficient to execute the decision-making excavation action, the corresponding action is marked as an inactive action.

[0065] In terms of details, during the action selection process of the intelligent agent, the action value of the non-actionable actions can be set to a minimum value, thereby effectively avoiding the selection of actions that exceed the remaining number of excavation layers.

[0066] 4.3) Reward Function The reward function used for reward feedback during training includes real-time step rewards and final-state stability evaluation rewards to achieve two-stage rewards.

[0067] Specifically, real-time step rewards can be used to evaluate the rationality of construction at each excavation decision step. Real-time step rewards can be determined based on at least one of the following: single-step excavation safety, single-step excavation progress, completion status, construction efficiency, elevation constraint violation, spatial proximity conflict, and invalid action.

[0068] The final-state stability assessment reward can be used to evaluate the surrounding rock stability corresponding to the complete excavation sequence. The final-state stability assessment reward can be determined based on the predicted value of the surrounding rock physical response output by the physical response prediction model.

[0069] Among them, for the first Predicted physical response values It can be normalized according to the training sample range or the engineering safety range, and can be expressed as: , in, The normalized value of the output. The first Reference lower and reference upper bounds for each physical response index The smaller the value, the better the corresponding physical response is for stability.

[0070] In this way, a stability base reward can be calculated based on the preset weights of the predicted values ​​of different types of physical responses, and an over-threshold penalty can be calculated when at least one key physical response predicted value exceeds the corresponding safety threshold.

[0071] Based on the above reward configuration, the total reward can be specifically expressed as follows: , in, Indicates the total reward. Indicates real-time step reward. This indicates a reward for assessing final state stability. This represents the reward weight for evaluating final state stability.

[0072] As an example, you can also refer to this. Figure 5 The diagram shown illustrates how the two-stage reward value of this application changes during the training process (the final average value is represented by a horizontal dashed line) to provide a more intuitive understanding.

[0073] Furthermore, corresponding to the constraint satisfaction state in the node features of the previously mentioned underground cavern group structure, for elevation-dependent constraints, the reinforcement learning excavation environment can encode the constraint satisfaction state in the node features and impose penalties on elevation constraint violations in the reward function. For spatial proximity constraints, the environment can determine whether there are excavation layer units with spatial proximity or contact relationships being excavated synchronously within the same excavation step, and impose penalties based on the number of conflicts, conflict edge weights, or conflict contact areas.

[0074] Step S105: In the reinforcement learning excavation environment, an excavation decision agent is trained to output joint excavation actions based on the state space. During the training process, the reward is evaluated by combining the predicted value of the surrounding rock physical response output by the physical response prediction model. Understandably, the excavation decision-making agent to be trained in this application is specifically used to output multiple sets of candidate excavation sequences for evaluating the optimal excavation sequence optimization design results. Its training work is specifically carried out on the basis of the reinforcement learning excavation environment and physical response prediction model configured above.

[0075] Specifically, the excavation decision-making agent can employ deep Q-networks, dual deep Q-networks, proximal policy optimization networks, or other types of deep reinforcement learning networks.

[0076] Taking the application of deep Q networks as an example, this application also specifically designed a set of network architecture configurations.

[0077] Specifically, the decision-making agent for excavation in this application can adopt a deep Q-network, which can be composed of several major network components, including graph attention feature branch, node statistical feature branch, group state feature branch, and action value output branch. The graph attention feature branch is used to aggregate neighborhood information of the underground cavern group graph structure in order to extract the spatial topological features between excavation layers; The node statistical feature branch is used to perform statistical analysis on the mean, maximum, minimum, and standard deviation of node features to extract the global excavation status distribution features; The grouping status feature branch is used to extract the excavation progress and inter-group coordination relationship of each cavern group according to the aggregation node features of the cavern group; The action value output branch is used to output the action value of different actions for each cave group.

[0078] During training, the excavation decision agent can also be configured to store the current state, action, reward, next state, and termination flag in the experience replay buffer, and sample small batches of data from the experience replay buffer to update the network parameters.

[0079] To reduce storage overhead, the excavation decision agent can also store only the list of excavated nodes, the global construction status, and action information, and reconstruct the complete graph structure state based on the list of excavated nodes during training sampling.

[0080] Step S106: Among the multiple candidate excavation sequences generated by the excavation decision-making agent, the target excavation sequence that meets the preset optimization conditions is selected as the excavation sequence optimization design result of the underground cavern group to be optimized.

[0081] Understandably, after the excavation decision-making agent completes its training, it can be further used to generate multiple sets of candidate excavation sequences based on the state space, corresponding to the agent's working logic during the training process, for the current underground cavern group to be optimized. These sequences include the target excavation sequence to be optimally determined, that is, the excavation sequence optimization design result of the underground cavern group to be optimized.

[0082] Specifically, after training, the trained excavation decision agent can generate multiple sets of candidate excavation sequences. For each set of candidate excavation sequences, the physical response prediction model or numerical simulation can be used to calculate the predicted value of the surrounding rock physical response to calculate the corresponding performance evaluation index. Specifically, the number of excavation steps, the number of parallel excavation faces, the number of constraint violations, the real-time step reward, the final state stability evaluation reward, and the total reward can be calculated.

[0083] As an example, based on preset screening rules, the candidate excavation sequence with the maximum total reward, the number of constraint violations meeting preset conditions, and the surrounding rock physical response value meeting the safety threshold can be selected from the candidate excavation sequences and output as the target excavation sequence.

[0084] Meanwhile, the output target excavation sequence may include specific content such as the excavation chamber group corresponding to each excavation step, the number of excavation layers in each chamber group within the excavation step, constraint check results, stability prediction results, and scheme evaluation indicators, so as to provide accurate and concise data reference.

[0085] After obtaining the processing results, including the target excavation sequence, it is understandable that specific data application processing can be performed, such as local storage, off-site storage, result display, result forwarding, result push, or further data analysis. In this process, it is obvious that adaptive processing can be performed according to the pre-configured and real-time configured data application strategies.

[0086] In addition, to better understand the above solutions, you can refer to the two engineering examples shown below.

[0087] In a specific engineering example, corresponding Figure 2 The underground cavern group to be optimized is the underground cavern group of a pumped storage power station. This underground cavern group of the pumped storage power station includes multiple caverns with different functions, such as the main powerhouse, main transformer room, and tailgate room. Each cavern is divided into multiple excavation layered units according to a preset thickness. Each excavation layered unit is a node in the diagram structure. Intra-group connection edges are set between adjacent layers within the group, and inter-group connection edges are set between layers that have contact or spatial proximity between different caverns.

[0088] In this example, multiple sets of different excavation sequence samples are generated through three-dimensional numerical simulation, and multiple physical response indicators of the surrounding rock after excavation are extracted. After the physical response prediction model is trained, it is used to predict indicators such as deformation, stress, elastic energy, plastic zone and tensile stress zone corresponding to the excavation sequence.

[0089] The reinforcement learning excavation environment treats multiple cavern groups as joint action objects. The action of each cavern group in a single excavation step is represented by a preset excavation volume level. The environment can generate action masks based on the remaining excavable volume of each cavern group and the preset excavation granularity, and calculate penalty terms based on elevation dependency constraints and spatial proximity constraints.

[0090] In the example data, the physical response prediction model averages 19 physical response indicators. The efficiency reached 0.9388, indicating that the optimized scheme generated by the trained agent reduced the number of excavation steps compared to the designed scheme and improved on multiple stability indicators.

[0091] Finally, regarding the above-mentioned solutions, in general, under the objective of optimizing the excavation sequence of underground cavern groups, this application, based on the underground cavern group graph structure, physical response prediction model, reinforcement learning excavation environment, reward function, and excavation decision-making agent, has built an excavation sequence optimization design architecture that can uniformly express the spatial topology, dynamic excavation state, and engineering constraints of underground cavern groups, while also taking into account construction efficiency and surrounding rock stability. This can effectively improve the technical defects of existing underground cavern group excavation sequence schemes, such as large spatial size, high evaluation cost, insufficient representation mode, and difficulty in jointly optimizing construction feasibility and surrounding rock stability, providing accurate, efficient, and low-cost scheme design for underground cavern group excavation work.

[0092] In terms of details, this application has the following specific beneficial effects: 1) The graph structure is used to express the excavation layers, cavern groups, rock mass parameters, spatial adjacency relationships and constraint states. Compared with the fixed-dimensional vector expression method, it is more suitable for describing the irregular spatial topology and excavation coupling relationship of underground cavern groups. 2) By using the physical response prediction model to quickly evaluate the stability of candidate excavation sequences, the number of repeated three-dimensional numerical simulations can be reduced, making large-scale interactive training and multi-scheme comparison of reinforcement learning feasible in engineering. 3) By using action masks, constraint state encoding, and penalty / reward methods to handle engineering constraints such as excavation capacity, elevation dependence, and spatial proximity, it is helpful to generate excavation sequences that conform to construction logic. 4) The training of the agent is guided by real-time step rewards and final stability assessment rewards, so that the optimization objective covers construction efficiency, constraint satisfaction and surrounding rock stability at the same time, avoiding excavation decisions based solely on short-term progress or rule satisfaction. 5) It can output candidate excavation sequences, target excavation sequences, stability evaluation indicators and constraint check results, providing efficient auxiliary decision-making basis for the generation and comparison of construction schemes in the preliminary design stage of underground cavern groups; 6) The solution's local operation strategy does not involve the public network or third-party servers, allowing for a high degree of independent control over the data. This ensures that there is no risk of data leakage and improves the system's security protection capabilities.

[0093] The above is an introduction to the method for optimizing the excavation sequence of underground cavern groups provided in this application. In order to facilitate the better implementation of the method for optimizing the excavation sequence of underground cavern groups provided in this application, this application also provides an apparatus for optimizing the excavation sequence of underground cavern groups from the perspective of functional modules.

[0094] See Figure 6 , Figure 6 This is a schematic diagram of a structure for the underground cavern group excavation sequence optimization device of this application. In this application, the underground cavern group excavation sequence optimization device 600 may specifically include the following structure: The data acquisition unit 601 is used to acquire the engineering data of the underground cavern group to be optimized under the local operation strategy. The graph structure building unit 602 is used to build the underground cavern group graph structure corresponding to the underground cavern group to be optimized based on engineering data. In the underground cavern group graph structure, the nodes correspond to the excavation layer units of the underground cavern group to be optimized, and the edges correspond to the spatial contact relationship between the excavation layer units. Model configuration unit 603 is used to configure a physical response prediction model that outputs corresponding surrounding rock physical response prediction values ​​based on the excavation sequence and underground cavern group diagram structure input to the model; Environment configuration unit 604 is used to configure a reinforcement learning mining environment including a state space, an action space, and a reward function; The agent training unit 605 is used to train an excavation decision agent in a reinforcement learning excavation environment to output joint excavation actions based on the state space. During the training process, the reward is evaluated by combining the predicted value of the surrounding rock physical response output by the physical response prediction model. The optimized output unit 606 is used to select the target excavation sequence that meets the preset optimization conditions from multiple sets of candidate excavation sequences generated by the excavation decision agent as the excavation sequence optimization design result of the underground cavern group to be optimized.

[0095] In one exemplary embodiment, the engineering data includes the geometric data of the cavern group, excavation layer information, rock mechanics parameters, geostress distribution, and engineering constraints. The geometric data of the cavern complex includes the spatial structure, dimensions, and connectivity of the underground caverns; Excavation layer information includes layer thickness, layer elevation, layer center coordinates, the cavern group to which the layer belongs, and the excavation direction; Rock mass mechanics parameters include rock mass density, elastic modulus, Poisson's ratio, cohesion, internal friction angle, tensile strength, rock mass classification, fault parameters, and weak interlayer parameters; The distribution of geostress includes the magnitude of geostress components and the direction of principal stresses; Engineering constraints include elevation dependence constraints, spatial proximity constraints, excavation capacity constraints, and maximum number of excavation steps constraints; For the node characteristics of nodes in the underground cavern group diagram structure, at least one of the following is considered: excavation status, excavation time, spatial coordinates, excavation volume, cavern group identifier, rock mechanics parameters, ground stress value, number of adjacent nodes, and constraint satisfaction status. For the edges in the underground cavern group diagram structure, they are divided into intra-group connecting edges and inter-group connecting edges. Intra-group connecting edges represent the vertical or front-to-back connection relationship between adjacent layers within the same cavern group, while inter-group connecting edges represent the contact relationship or spatial proximity relationship between different cavern groups. The edge weight is determined based on the contact area, spatial distance, topological connection type, or mechanical coupling strength between excavated layer units.

[0096] In yet another exemplary embodiment, the training samples used to train the physical response prediction model include excavation sequence samples and surrounding rock physical response data samples. The excavation sequence sample includes a set of excavation stratified units arranged according to excavation steps; The physical response data sample of the surrounding rock includes surrounding rock deformation, surrounding rock stress, elastic energy, plastic zone, local stress concentration index, plastic zone volume and tensile stress zone volume. Surrounding rock deformation includes the displacement of the sidewalls of the key cavern and the displacement of the arch. Surrounding rock stress includes the maximum principal stress. Elastic energy includes total elastic energy and local elastic energy.

[0097] In yet another exemplary embodiment, the model architecture for the physical response prediction model includes a sequence feature extraction branch, a static graph feature extraction branch, a dynamic mask graph feature extraction branch, a fusion prediction branch, and a fully connected prediction network. The sequence feature extraction branch is used to extract the corresponding excavation path dependency features of the excavation state of the excavation sequence input to the model as the excavation step changes; The static graph feature extraction branch is used to extract the corresponding static spatial features of the excavation sequence input to the model based on the underground cavern group graph structure; The dynamic mask graph feature extraction branch is used to generate node masks or edge masks corresponding to the underground cavern group graph structure based on the excavation state of the excavation sequence input to the model, and extract the dynamic neighborhood interaction features formed by the change of effective neighborhood relationships over time during the excavation process based on the dynamic changes of the generated masks. The fusion prediction branch fuses excavation path dependency features, static spatial features, and dynamic neighborhood interaction features to obtain multimodal fusion features; The fully connected prediction network maps multimodal fusion features into predicted values ​​of the surrounding rock physical response for output.

[0098] In another exemplary embodiment, for a reinforcement learning excavation environment, the state space includes a graph structure state and a global construction state. The graph structure state includes node features, edge connection relationships, and edge weights. The global construction state includes the excavation progress of each cavern group, the remaining proportion of each cavern group, the current excavation step proportion, the overall excavation progress, the safe excavation proportion of each cavern group, the overall constraint satisfaction proportion, and the action mask. The action space consists of the joint excavation actions of multiple cavern groups. Assume that the underground cavern group to be optimized includes k cavern groups, and the actions of each cavern group within one excavation decision step are as follows. The excavation volume level is represented by the preset excavation particle size. The excavation volume level can be represented by no excavation, excavating one preset layer, or excavating multiple preset layers. The excavation volume level sets of different cavern groups are configured independently. The joint excavation action is represented as follows: , in, This indicates a joint excavation operation. , This indicates the total number of actions involved in the joint excavation. Generate an action mask matrix at each excavation decision step; The reward function used for reward feedback during training includes real-time step rewards and final-state stability evaluation rewards to achieve two-stage rewards.

[0099] In yet another exemplary embodiment, the real-time step reward is used to evaluate the construction rationality of each excavation decision step. The real-time step reward is determined based on at least one of the following: single-step excavation safety, single-step excavation progress, completion status, construction efficiency, elevation constraint violation, spatial proximity conflict, and invalid action. The final-state stability assessment reward is used to evaluate the surrounding rock stability corresponding to the complete excavation sequence. The final-state stability assessment reward is determined based on the predicted value of the surrounding rock physical response output by the physical response prediction model. The total reward is expressed as follows: , in, Indicates the total reward. Indicates real-time step reward. This indicates a reward for assessing final state stability. This represents the reward weight for evaluating final state stability.

[0100] In yet another exemplary embodiment, the excavation decision agent employs a deep Q-network, which includes a graph attention feature branch, a node statistical feature branch, a group state feature branch, and an action value output branch. The graph attention feature branch is used to aggregate neighborhood information of the underground cavern group graph structure in order to extract the spatial topological features between excavation layers; The node statistical feature branch is used to perform statistical analysis on the mean, maximum, minimum, and standard deviation of node features to extract the global excavation status distribution features; The grouping status feature branch is used to extract the excavation progress and inter-group coordination relationship of each cavern group according to the aggregation node features of the cavern group; The action value output branch is used to output the action value of different actions for each cave group; During training, the mining decision agent stores the current state, action, reward, next state, and termination flag in the experience replay buffer, and samples small batches of data from the experience replay buffer to update the network parameters.

[0101] This application also provides a processing device from a hardware architecture perspective. As mentioned earlier, in practice, a processing device may exist as a device cluster. In this case, each device in the device cluster can also be referred to as a processing device. See [reference needed]. Figure 7 , Figure 7 This diagram illustrates a structural schematic of the processing device of this application. Specifically, the processing device may include a processor 701, a memory 702, and an input / output device 703. The processor 701 executes the computer program stored in the memory 702 to implement, for example... Figure 1 The corresponding steps of the underground cavern group excavation sequence optimization method in the embodiment; or, when the processor 701 executes the computer program stored in the memory 702, it implements as follows: Figure 6 Corresponding to the functions of each unit in the embodiment, the memory 702 is used to store the functions executed by the processor 701 as described above. Figure 1 The computer program required for the underground cavern group excavation sequence optimization method in the corresponding embodiment.

[0102] For example, a computer program may be divided into one or more modules / units, one or more of which are stored in memory 702 and executed by processor 701 to complete this application. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in a computer device.

[0103] The processing device may include, but is not limited to, processor 701, memory 702, and input / output device 703. Those skilled in the art will understand that the illustrations are merely examples of the processing device and do not constitute a limitation on the processing device. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the processing device may also include network access devices, buses, etc., and processor 701, memory 702, input / output device 703, etc., are connected via a bus.

[0104] The processor 701 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the processing device, connecting various parts of the device through various interfaces and lines.

[0105] The memory 702 can be used to store computer programs and / or modules. The processor 701 implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory 702 and by calling data stored in the memory 702. The memory 702 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function, etc.; the data storage area may store data created according to the use of the processing device, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0106] When processor 701 executes a computer program stored in memory 702, it can specifically perform the following functions: Under the local operation strategy, obtain the engineering data of the underground cavern group that needs to be optimized; Based on engineering data, a map structure of the underground cavern group corresponding to the underground cavern group to be optimized is constructed. In the map structure of the underground cavern group, the nodes correspond to the excavation layered units of the underground cavern group to be optimized, and the edges correspond to the spatial contact relationship between the excavation layered units. Configure a physical response prediction model to output corresponding predicted values ​​of the surrounding rock physical response based on the excavation sequence and underground cavern group diagram structure input to the model; Configure a reinforcement learning mining environment that includes a state space, action space, and reward function; In a reinforcement learning excavation environment, an excavation decision agent is trained to output joint excavation actions based on the state space. During the training process, the reward is evaluated by combining the predicted value of the surrounding rock physical response output by the physical response prediction model. Among the multiple candidate excavation sequences generated by the excavation decision-making agent, the target excavation sequence that meets the preset optimization conditions is selected as the excavation sequence optimization design result of the underground cavern group to be optimized.

[0107] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the underground cavern excavation sequence optimization device, processing equipment, and its corresponding units described above can be found in, for example... Figure 1 The description of the underground cavern group excavation sequence optimization method in the corresponding embodiment will not be repeated here.

[0108] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0109] Therefore, this application provides a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute the present application. Figure 1 The steps of the underground cavern group excavation sequence optimization method in the corresponding embodiment can be referred to as follows for specific operations. Figure 1 The description of the underground cavern group excavation sequence optimization method in the corresponding embodiment will not be repeated here.

[0110] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0111] Because of the instructions stored in the computer-readable storage medium, the present application can be executed as described above. Figure 1 The steps of the underground cavern group excavation sequence optimization method in the corresponding embodiment can therefore achieve the results of this application. Figure 1 The beneficial effects that the underground cavern group excavation sequence optimization method can achieve in the corresponding embodiment are detailed in the preceding description and will not be repeated here.

[0112] The above provides a detailed description of the method, apparatus, processing equipment, and computer-readable storage medium for optimizing the excavation sequence of underground cavern groups provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the core ideas of this application; furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for optimizing the excavation sequence of underground cavern groups, characterized in that, The method includes: Under the local operation strategy, obtain the engineering data of the underground cavern group that needs to be optimized; Based on the engineering data, a map structure of the underground cavern group corresponding to the underground cavern group to be optimized is constructed. In the underground cavern group map structure, the nodes correspond to the excavation layer units of the underground cavern group to be optimized, and the edges correspond to the spatial contact relationship between the excavation layer units. Configure a physical response prediction model to output corresponding predicted values ​​of the surrounding rock physical response based on the excavation sequence input to the model and the underground cavern group diagram structure; Configure a reinforcement learning mining environment that includes a state space, action space, and reward function; In the reinforcement learning excavation environment, an excavation decision agent is trained to output joint excavation actions based on the state space. During the training process, the reward is evaluated by combining the predicted value of the surrounding rock physical response output by the physical response prediction model. Among the multiple candidate excavation sequences generated by the excavation decision-making agent, the target excavation sequence that meets the preset optimization conditions is selected as the excavation sequence optimization design result of the underground cavern group to be optimized. For the reinforcement learning excavation environment, the state space includes a graph structure state and a global construction state. The graph structure state includes node features, edge connection relationships, and edge weights. The global construction state includes the excavation progress of each cavern group, the remaining proportion of each cavern group, the current excavation step proportion, the overall excavation progress, the safe excavation proportion of each cavern group, the overall constraint satisfaction proportion, and the action mask. The action space refers to the joint excavation actions of multiple cavern groups. Assuming the underground cavern group to be optimized includes k cavern groups, the actions of each cavern group within an excavation decision step... The excavation volume level is represented by a preset excavation particle size, whereby the excavation volume level is represented by no excavation, excavation of one preset layer, or excavation of multiple preset layers. The excavation volume level sets for different cavern groups are configured independently of each other. The joint excavation action is represented as follows: ， in, This indicates the joint excavation operation. , This represents the total number of actions involved in the combined excavation. Generate an action mask matrix at each excavation decision step; The reward function used for reward feedback during training includes real-time step rewards and final-state stability evaluation rewards to achieve a two-stage reward system.

2. The method according to claim 1, characterized in that, The engineering data includes the geometric data of the cavern group, excavation layer information, rock mechanics parameters, geostress distribution and engineering constraints; The geometric data of the cavern group includes the spatial structure, dimensions, and connectivity of the underground caverns; The excavation layer information includes layer thickness, layer elevation, layer center coordinates, the cavern group to which the layer belongs, and the excavation direction; The rock mass mechanical parameters include rock mass density, elastic modulus, Poisson's ratio, cohesion, internal friction angle, tensile strength, rock mass classification, fault parameters, and weak interlayer parameters. The geostress distribution includes the magnitude of the geostress components and the direction of the principal stresses; The engineering constraints include elevation dependence constraints, spatial proximity constraints, excavation capacity constraints, and maximum number of excavation steps constraints. The node characteristics of the nodes in the underground cavern group diagram structure include at least one of the following: excavation status, excavation time, spatial coordinates, excavation volume, cavern group identifier, rock mechanics parameters, ground stress value, number of adjacent nodes, and constraint satisfaction status. The edges in the underground cavern group structure are divided into intra-group connecting edges and inter-group connecting edges. Intra-group connecting edges represent the vertical or front-back connection relationship between adjacent layers within the same cavern group. Inter-group connecting edges represent the contact relationship or spatial proximity relationship between different cavern groups. The edge weight is determined based on the contact area, spatial distance, topological connection type, or mechanical coupling strength between the excavation layer units.

3. The method according to claim 1, characterized in that, Training samples used to train the physical response prediction model include excavation sequence samples and surrounding rock physical response data samples; The excavation sequence sample includes a set of excavation stratification units arranged according to excavation steps; The surrounding rock physical response data sample includes surrounding rock deformation, surrounding rock stress, elastic energy, plastic zone, local stress concentration index, plastic zone volume and tensile stress zone volume. The surrounding rock deformation includes the displacement of the key cavern sidewalls and the arch displacement. The surrounding rock stress includes the maximum principal stress. The elastic energy includes the total elastic energy and the local elastic energy.

4. The method according to claim 1, characterized in that, The model architecture of the physical response prediction model includes a sequence feature extraction branch, a static graph feature extraction branch, a dynamic mask graph feature extraction branch, a fusion prediction branch, and a fully connected prediction network. The sequence feature extraction branch is used to extract the corresponding excavation path dependency features of the excavation state of the excavation sequence input to the model as the excavation step changes; The static graph feature extraction branch is used to extract the corresponding static spatial features of the excavation sequence input to the model based on the underground cavern group graph structure; The dynamic mask map feature extraction branch is used to generate node masks or edge masks corresponding to the underground cavern group map structure based on the excavation state of the excavation sequence input to the model, and extract dynamic neighborhood interaction features formed by the change of effective neighborhood relationships over time during the excavation process based on the dynamic changes of the generated masks. The fusion prediction branch fuses the excavation path dependency features, the static spatial features, and the dynamic neighborhood interaction features to obtain multimodal fusion features; The fully connected prediction network maps the multimodal fusion features to the predicted values ​​of the surrounding rock physical response for output.

5. The method according to claim 1, characterized in that, The real-time step reward is used to evaluate the construction rationality of each excavation decision step. The real-time step reward is determined based on at least one of the following: single-step excavation safety, single-step excavation progress, completion status, construction efficiency, elevation constraint violation, spatial proximity conflict, and invalid action. The final-state stability assessment reward is used to evaluate the surrounding rock stability corresponding to the complete excavation sequence. The final-state stability assessment reward is determined based on the predicted value of the surrounding rock physical response output by the physical response prediction model. The total reward is expressed as follows: ， in, This represents the total reward. This refers to the real-time step reward. This represents the reward for the final state stability assessment. This represents the reward weight for evaluating final state stability.

6. The method according to claim 1, characterized in that, The excavation decision agent adopts a deep Q-network, which includes a graph attention feature branch, a node statistical feature branch, a group state feature branch, and an action value output branch. The graph attention feature branch is used to aggregate neighborhood information of the underground cavern group graph structure in order to extract the spatial topological features between excavation layers; The node statistical feature branch is used to perform statistical analysis on the mean, maximum, minimum and standard deviation of node features to extract the global excavation status distribution features; The grouping state feature branch is used to aggregate the node features according to the cavern group, and extract the excavation progress and inter-group coordination relationship of each cavern group; The action value output branch is used to output the action value of different actions for each of the cavern groups; During training, the excavation decision-making agent stores the current state, action, reward, next state, and termination flag in the experience replay buffer, and samples small batches of data from the experience replay buffer to update the network parameters.

7. A device for optimizing the excavation sequence of underground cavern groups, characterized in that, The device includes: The data acquisition unit is used to acquire engineering data of the current underground cavern group to be optimized under the local operation strategy; The graph structure construction unit is used to construct the underground cavern group graph structure corresponding to the underground cavern group to be optimized based on the engineering data. In the underground cavern group graph structure, the nodes correspond to the excavation layer units of the underground cavern group to be optimized, and the edges correspond to the spatial contact relationship between the excavation layer units. The model configuration unit is used to configure a physical response prediction model that outputs corresponding surrounding rock physical response prediction values ​​based on the excavation sequence input to the model and the underground cavern group diagram structure. The environment configuration unit is used to configure the reinforcement learning mining environment, including the state space, action space, and reward function. The agent training unit is used to train an excavation decision agent that outputs joint excavation actions according to the state space in the reinforcement learning excavation environment. During the training process, the reward is evaluated by combining the predicted value of the surrounding rock physical response output by the physical response prediction model. The optimization output unit is used to select the target excavation sequence that meets the preset optimization conditions from the multiple sets of candidate excavation sequences generated by the excavation decision-making agent as the excavation sequence optimization design result of the underground cavern group to be optimized. For the reinforcement learning excavation environment, the state space includes a graph structure state and a global construction state. The graph structure state includes node features, edge connection relationships, and edge weights. The global construction state includes the excavation progress of each cavern group, the remaining proportion of each cavern group, the current excavation step proportion, the overall excavation progress, the safe excavation proportion of each cavern group, the overall constraint satisfaction proportion, and the action mask. The action space refers to the joint excavation actions of multiple cavern groups. Assuming the underground cavern group to be optimized includes k cavern groups, the actions of each cavern group within an excavation decision step... The excavation volume level is represented by a preset excavation particle size, whereby the excavation volume level is represented by no excavation, excavation of one preset layer, or excavation of multiple preset layers. The excavation volume level sets for different cavern groups are configured independently of each other. The joint excavation action is represented as follows: ， in, This indicates the joint excavation operation. , This represents the total number of actions involved in the combined excavation. Generate an action mask matrix at each excavation decision step; The reward function used for reward feedback during training includes real-time step rewards and final-state stability evaluation rewards to achieve a two-stage reward system.

8. A processing device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the method as described in any one of claims 1 to 6 when it invokes the computer program in the memory.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the method of any one of claims 1 to 6.

Images