A method, device and medium for planning open-pit mining of polymetallic co-associated ores

Abstract

The application discloses a kind of multi-metal co-associated open-pit mine mining planning method, equipment and medium, and the application belongs to the field of intelligent mining technology.The application first constructs the economic geology model of multi-mining area multi-metal ore block;Second, the maximum net present value of mine full life cycle is taken as the core to construct objective function;Design high-grade ore dynamic ore blending strategy based on warehouse buffer, define constraint system, and introduce graph neural network to construct ore block topological graph structure, finally construct hierarchical reinforcement learning model coordinated by high-level manager and bottom-layer executor, output the global optimal time sequence mining ore blending scheme.The application proposes a kind of multi-mining area multi-metal open-pit mine full cycle mining ore blending optimization method combining graph neural network GNN and hierarchical reinforcement learning, the spatial neighborhood topology and grade correlation characteristics of ore block are mined by GNN, hierarchical reinforcement learning is used to split long-term strategic planning and short-term mining ore blending action, effectively solve the solution problem of long-term mining planning of multi-mining area multi-metal open-pit mine.

Description

Technical Field

[0001] This invention relates to intelligent mining technology, and more particularly to a method, equipment and medium for planning the mining of polymetallic co-existing open-pit mines. Background Technology

[0002] Long-term mining planning for multi-mining areas with polymetallic associated open-pit mines is a typical large-scale, multi-constraint sequential optimization decision problem. The solution results directly determine the economic benefits of the mine throughout its entire life cycle and the comprehensive utilization rate of associated metal resources.

[0003] Currently, mining planning optimization technologies in the industry mainly fall into two categories, both of which have significant drawbacks: First, mathematical modeling methods, including mixed integer programming models and integer linear programming models. These methods can find the theoretical optimal solution in small-scale models, but for large-scale engineering scenarios with multiple mining areas and massive amounts of mineral blocks, they are prone to the curse of dimensionality, which drastically increases computation time and may even fail to provide an effective solution, resulting in poor engineering practicality.

[0004] Second, metaheuristic intelligent optimization algorithms, such as particle swarm optimization and ant colony optimization. These algorithms are adaptable to complex constraints and have higher flexibility in solving problems, but they are memoryless black-box optimization mechanisms and lack the ability to learn state strategies autonomously or generalize across mining scenarios. At the same time, they are difficult to handle long-term reward sparsity problems and are prone to converge to local optima, making it difficult to maximize the net present value over the entire cycle.

[0005] Existing research has also introduced reinforcement learning to optimize short-term mine scheduling, but it still has inherent shortcomings in long-term time-series planning scenarios: First, it can only optimize short-term mining actions and lacks global long-term strategic control capabilities; second, the modeling process treats each ore block independently, losing information on spatial adjacency and step-level topology between ore blocks, making it impossible to accurately represent the spatial correlation of slope mining, and resulting in insufficient accuracy in ore blending and mining time-series planning.

[0006] Therefore, there is an urgent need for an intelligent mining optimization method for multi-metal open-pit mines in multiple mining areas that can meet the needs of long-term planning in open-pit mines for long-term decision-making. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method, equipment and medium for planning the mining of polymetallic co-existing open-pit mines, which addresses the deficiencies in the existing technology.

[0008] The technical solution adopted by this invention to solve its technical problem is: a method for planning the mining of polymetallic associated open-pit mines, comprising the following steps: 1) Construct a multi-mining area polymetallic ore block model to obtain the spatial coordinates, metal grade values, and ore block mass of each block; 2) Classify ore blocks according to their metal grade; Based on the grade thresholds of various metal industries, the ore blocks are divided into four categories: high-grade ore, general-grade ore, low-grade ore, and waste rock. 3) Establish a multi-mining area polymetallic mining planning model, wherein the objective function is to maximize the net present value (NPV) over the entire life cycle; 4) Establish a multi-mine-area polymetallic mining planning and ore blending strategy. Set up storage bins to buffer high-grade ore and mix high and low grades so that at least one metal reaches industrial grade after ore mixing. 5) Set constraints for multi-metal mining planning in multiple mining areas; 6) Construct a graph neural network (GNN); The input to a graph neural network is: node feature vectors and graph network topology. The output of the graph neural network is a shared intermediate feature set, which integrates block attributes and block neighborhood topology information. 7) Layered decomposition of long-term strategy and short-term mining operations to achieve intelligent decision-making for mining sequences in multiple mining areas; A two-layer hierarchical reinforcement learning model is constructed, consisting of a high-level manager and a low-level executor. The high-level manager receives shared intermediate features from the output of the GNN and generates a long-term mining strategic target vector through a dilated LSTM. The low-level executor fuses the shared intermediate features from the output of the GNN with the high-level target vector and generates a mining allocation action probability distribution based on LSTM. The high-level and low-level models update parameters using policy gradients. The high-level manager updates the policy gradient with the full-cycle NPV return as the objective, while the low-level executor updates the policy gradient with the current combined return of ore allocation and mining. 8) Complete model training and output a complete time series sequence of mining operations in multiple mining areas.

[0009] According to the above scheme, in step 2), High-grade ore means that the grade of any one metal in the ore block is higher than the dynamic high-grade multiple threshold; General grade ore indicates that the metal grade reaches the industrial grade, but does not reach the high grade multiple threshold. Low-grade ore means that the grade of all metals in the ore block is lower than the industrial grade. Waste rock indicates that the ore block contains no valuable metal grade.

[0010] According to the above scheme, in step 4), the ore blending strategy is as follows: a storage silo is defined to store high-grade ore, and low-grade ore is matched to achieve ore blending. If the matching fails for multiple rounds, the storage silo is emptied. Low-grade ore is selected to be matched with the current high-grade ore or with the high-grade ore in the storage silo. If the matching fails for multiple rounds, it is treated as waste rock. General-grade ore is selected to be mixed with ore of the same stage under the condition of meeting the constraints for ore blending in the beneficiation plant, or enters the beneficiation plant alone. High-grade ore is selected to be mixed with ore of the same stage under the condition of meeting the constraints for ore blending in the beneficiation plant, or a portion of it is selected to be stored and enters the remaining beneficiation plant, or all of it is entered into the beneficiation plant.

[0011] According to the above scheme, in step 4), the decision, state, and judgment functions for ore blending optimization are defined, and the specific process of the ore blending strategy is as follows: 4.1) Decision Variables Decision variables include storage decisions and ore blending decisions, for each time period t, t=1, 2…t n The definition is as follows: Store decision variables ,Decide Whether to store, where 0 means no storage and all processing, and 1 means 15% storage and 85% processing (mixed or direct selection). Decision option 0 indicates no ore blending; decision option 1 indicates... and Ore blending; Processing decision option 2 indicates With storage warehouse Ore blending; 4.2) State Variables The storage capacity of high-grade ore in the storage bin, with an initial value of 0; : Number of storage bin mixing failures at the start of time t; : Number of low-grade ore mixing failures at the start of time t; 4.3) Mixed judgment function Mixed decision function for processing decisions Calculate whether at least one metal reaches industrial grade after blending. Return 1 if blending is successful, otherwise return 0.

[0012] In the formula, The amount of high-grade ore selected for the time period t includes high-grade ore in the storage bin and high-grade ore directly processed; The high-grade ore grade selected for the time period t; The amount of general-grade ore selected for the time period t; The typical grade of ore selected for the time period t; The amount of low-grade ore selected for the time period t; The grade of low-grade ore selected for the time period t.

[0013] According to the above scheme, in step 5), the constraints include ore block reserve constraints, open-pit slope bench mining priority constraints, upper and lower limits of mine mining, storage, and beneficiation plant capacity constraints, and minimum grade constraints of polymetallic materials entering the plant.

[0014] According to the above scheme, the constraints in step 5) are as follows: 5.1) Reserve Constraints The production volume of the ore block shall not exceed the reserves, and the sum of the ore volume and waste rock volume shall not exceed the total reserves of the ore block itself in the mining area; 5.2) Priority constraints for slope mining; Before mining a block at a lower elevation, the mining of a group of overlying blocks above it must be completed first. 5.3) Capacity, crushing, storage, and beneficiation plant constraints; The mining output, storage warehouse, and feed into the beneficiation plant must not exceed the upper and lower limits of the equipment's capacity. 5.4) Grade constraints: The ore sent to the beneficiation plant must meet the minimum grade requirements: The grade of the i-th metal entering the beneficiation plant within the t mining cycle must meet the minimum grade requirements of the beneficiation plant.

[0015] In the above scheme, step 6) involves the following graph neural network construction process: Graph network topology construction: Each mineral block is a graph node, and the node features include coordinates, grade, and quality; an adjacency matrix is ​​constructed based on the Euclidean distance and distance threshold between mineral blocks, and edges are established to connect spatially adjacent mineral blocks; Encoding is performed through a multi-layer graph convolutional network, followed by global pooling and feature fusion to generate a shared intermediate feature set; Global average pooling is performed on the output of the last layer of the single-layer GCN to compress the features of all block nodes in the mining area into a one-dimensional vector. The graph embedding vectors of all mining areas are concatenated to perform feature fusion. The output feature set is a shared intermediate feature set, which is simultaneously supplied to the high-level and low-level reinforcement learning modules.

[0016] In the above scheme, in step 7), the hierarchical reinforcement learning model operates as follows: 7.1) High-level manager: The long-term strategic goal is to maximize NPV. It receives the intermediate feature set output by GNN, maps it to latent states through fully connected layers and ReLU, and then inputs it into the target vector output by dynamic dilated LSTM, which is then normalized to a unit vector. ; 7.2) Low-level executor: Responsible for executing specific low-level actions, output by the perception module. Target vector set by the high-level manager Action design, i.e., block selection, is based on data; target vector. Embedding is achieved by summing the objectives of the past c steps and passing them through an unbiased linear layer. Mapped to target embedding vector Action embedding outputs an action embedding matrix using a standard LSTM. Action probability distribution is generated by combining target embedding and action embedding through matrix multiplication and then passing the SoftMax function. ; ; ; ; In the formula, Embed the vector for the target; It is a linear layer without bias; Let t be the target vector at step t; Represents the hidden features in an RNN; Embed the action matrix; For standard LSTM; The output vector of the sensing module; This is the probability distribution vector of the action.

[0017] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described scheme.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method in the above-described scheme.

[0019] The beneficial effects of this invention are: This invention proposes a method for optimizing ore blending in multi-mine polymetallic open-pit mines throughout the entire mining cycle by integrating graph neural networks (GNN) and hierarchical reinforcement learning. By mining the topological and grade correlation features of the spatial neighborhood of ore blocks through GNN, and using hierarchical reinforcement learning to separate long-term strategic planning and short-term mining and ore blending actions, this method effectively solves the problem of solving long-term mining planning in multi-mine polymetallic open-pit mines.

[0020] This invention designs a complete dynamic ore blending mechanism for storage, sets up high-grade ore storage decisions and multiple types of ore blending action selection, and allows low-grade ore in the current period to be mixed with high-grade ore in real time and stored high-grade ore to meet the standards for beneficiation. Under the premise of meeting the minimum grade constraints of the beneficiation plant, it significantly improves the comprehensive utilization rate of polymetallic associated mineral resources. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the GNN structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the hierarchical reinforcement learning model structure according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the ore blending strategy according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the mining constraints of a ore block according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] like Figure 1 As shown, a method for planning the mining of polymetallic associated open-pit mines includes the following steps: S1) Construct a multi-mining area multi-metal block model to obtain the spatial coordinates, metal grade values, and block quality of each block; Based on a geological model containing metal grade information, the kriging method and the distance inverse weighted method are used to interpolate and assign values ​​to the characteristic parameters of the ore blocks. The surface topography and the final mine boundary constraints are superimposed to complete the model trimming and division to obtain standardized ore blocks. Each ore block attribute includes spatial (X, Y, Z) coordinates, each metal grade, ore block number, and ore block quality. All ore block grid cells are generated in batches using the ore block center coordinate formula. Each block contains a spatial location (X, Y, Z coordinates), and the formula for calculating the center coordinates of a block is: ; In the formula, i, j, and k represent the sequence numbers of the mineral blocks in the X, Y, and Z directions, respectively. , , Indicates the center coordinates of the origin ore block. , , These represent the lengths of the mineral block in the X, Y, and Z directions, respectively. S2) Classify ore blocks by metal grade; Based on the grade thresholds of various metal industries, the ore blocks are divided into four categories: high-grade ore, general-grade ore, low-grade ore, and waste rock. High-grade ore means that the grade of any one metal in the ore block is higher than the dynamic high-grade multiple threshold. General grade ore indicates that the metal grade reaches the industrial grade, but does not reach the high grade multiple threshold. Low-grade ore means that the grade of all metals in the ore block is lower than the industrial grade. Waste rock indicates that the ore block contains no valuable metals. S3) Establish a multi-mining area polymetallic mining planning model, wherein the objective function is to maximize the net present value (NPV) over the entire life cycle; ; in, ;

[0024] = ; In the formula, Let t be the cash inflow in period t. For the cash outflow in period t, based on the cash outflows of all periods within year T... and Receive cash inflow in year T With cash outflow ; Let T be the net cash flow in year T. Let K be the price of the Kth type of concentrate in year t. Let be the recovery rate of the i-th valuable ore in the concentrator; Let i be the grade of the i-th metal in the k-th concentrate produced in the t-th cycle. The amount of ore transported from the crushing station to the beneficiation plant in the j-th region of the b-th mining area during cycle t is represented by this value. Let i be the grade of the i-th ore contained in the k-th concentrate in the j-th region of the b-th mining area. Let be the amount of ore transported from the storage silo in the b-th mine to the processing plant during period t. To determine the grade of the i-th valuable ore required for transporting the k-th concentrate from the storage warehouse to the processing plant, Let t be the total cost of stripping and blasting in period t. The total cost of producing explosives in cycle t is... The total cost of loading and unloading in cycle t is... The transportation cost of the spoil heap in period t. The cost of breaking down a circuit in period t. Let t be the cost of mineral processing in period t. n is the number of years required for the mine to complete its operation; b is the mining area index, b=1 to nb, where nb is the number of mining areas; Let represent the set of mining areas j that are mined in period t; m is the discount rate. Set up a multi-mine-area multi-metal mining planning and ore blending strategy, matching high-grade ore with low-grade ore to achieve ore blending, and after blending, at least one metal is blended to reach industrial grade. like Figure 4 Ore blending strategy: A storage silo is defined to store high-grade ore, which is then blended with low-grade ore. If blending fails for multiple rounds, the storage silo is emptied. Low-grade ore is selected for blending with currently available high-grade ore or with high-grade ore in the storage silo. If blending fails for multiple rounds, it is treated as waste rock. General-grade ore can be blended with high-grade ore of the same stage under certain constraints for beneficiation, or it can be sent to the beneficiation plant alone. High-grade ore can be blended with ore of the same stage under certain constraints for beneficiation, or a portion can be stored and sent to the remaining beneficiation plant, or all of it can be sent to the beneficiation plant. Figure 4 As shown. The specific logical flow of the ore blending strategy is as follows.

[0025] (1) Decision variables Decision variables include storage decisions and ore blending decisions, for each time period t, t=1, 2…t n The definition is as follows: Store decision variables ,Decide Whether to store, where 0 means no storage and all processing, and 1 means 15% storage and 85% processing (mixed or direct selection plant).

[0026] Decision option 0 indicates no ore blending; decision option 1 indicates... and Ore blending; Processing decision option 2 indicates With storage warehouse Ore blending.

[0027] (2) Set state variables; The storage capacity of high-grade ore in the storage bin, with an initial value of 0; : Number of storage bin mixing failures at the start of time t; : Number of low-grade ore mixing failures at the start of time t; (3) Mixed judgment function; Mixed decision function for processing decisions Calculate whether at least one metal reaches industrial grade after blending. Return 1 if blending is successful, otherwise return 0.

[0028] In the formula, The amount of high-grade ore selected for the time period t includes high-grade ore in the storage bin and high-grade ore directly processed; The high-grade ore grade selected for the time period t; The amount of general-grade ore selected for the time period t; The typical grade of ore selected for the time period t; The amount of low-grade ore selected for the time period t; The low-grade ore grade selected for the time period t; Waste rock and ore processing flow: Multiple rounds of unmatched Entering the spoil heap; matching successful. , , Entering the beneficiation plant; Low-grade ore and waste rock that failed to blend are transported to the spoil heap; high-grade, general-grade and low-grade mixed ore that successfully blends, as well as high-grade ore and general-grade ore that meet the standards on their own, are transported to the concentrator.

[0029] 5) Set constraints for multi-metal mining planning in multiple mining areas; Constraints include constraints on ore block reserves; constraints on slope mining priority; constraints on mine capacity, crushing, storage, and beneficiation plants; and grade constraints. In multi-mining area grade constraints, only one type of metal needs to meet the requirements for polymetallic ores to enter the beneficiation plant.

[0030] 1) Reserve constraints The production volume of the ore block shall not exceed the reserves, and the total amount of ore and waste rock shall not exceed the total reserves of the j ore block in mining area b. This also avoids double counting or over-exploitation of resources and ensures that the model conforms to the actual reserve management logic of the mine.

[0031] 2) Priority constraints for slope mining Before mining a block at a lower elevation (or step), the extraction of the overlying blocks above it must be completed. For example, to mine the (2,2,1) block, the extraction of the nine overlying blocks above it must be completed. Figure 5 As shown.

[0032] 3) Constraints on production capacity, crushing, storage, and beneficiation plants The mining output, storage warehouse, and feed into the beneficiation plant must not exceed the maximum / minimum capacity of the equipment; 4) Grade constraints (multi-metal grade constraints) The ore sent to the processing plant must meet the minimum grade requirements: The grade of the i-th metal entering the beneficiation plant during the t mining cycle must meet the minimum grade requirements of the beneficiation plant. 6) Construct a Graph Neural Network (GNN), the structure of which is as follows: Figure 2 As shown; the graph neural network includes: The graph construction submodule is used to construct a block graph based on the open-pit mine block model. Each block corresponds to a node in the block graph, and the edges between blocks are determined based on spatial adjacency, mining priority of overlying strata, affiliation within the same mining area, and / or geological continuity. Each node's characteristics include at least the block's spatial coordinates, block quality, ore type, metal grade, economic value, mining status, and information about its associated mining area. The node feature encoding submodule is used to encode the original attributes of each mining block node to obtain an initial node embedding vector; the neighborhood message passing submodule is used to pass the feature information of adjacent mining block nodes to the current mining block node according to the edge connection relationship in the mining block graph; the feature aggregation submodule is used to aggregate the neighborhood features of the current mining block node and update the node embedding vector of the current mining block node. The shared feature output submodule is used to generate a shared intermediate feature set based on the updated node embedding vector. The shared intermediate feature set integrates block attribute information, block neighborhood topology information, and block mining priority relationship information, and serves as the common input of the high-level manager and the low-level executor in the hierarchical reinforcement learning model. This enables the high-level manager to generate a phased target vector based on the shared intermediate feature set, and enables the low-level executor to generate block selection actions, storage actions, and / or ore allocation actions based on the shared intermediate feature set and the phased target vector.

[0033] The input to a graph neural network is: node feature vectors and graph network topology. The node feature vector includes coordinates, polymetallic grade, and ore mass; Graph network topology construction: Each mineral block is a graph node, and the node features include coordinates, grade, and quality; an adjacency matrix is ​​constructed based on the Euclidean distance and distance threshold between mineral blocks, and edges are established to connect spatially adjacent mineral blocks; Encoding is performed through a multi-layer graph convolutional network, followed by global pooling and feature fusion to generate a shared intermediate feature set; Global average pooling is performed on the output of the last layer of a single-layer GCN to compress the features of all block nodes in the mining area into a one-dimensional vector. Multi-layer GCN feature encoding: Normalized graph convolutional layers are used to progressively transmit the neighborhood spatial and grade correlation features of the ore block; For the Lth layer GCN, the information propagation formula is: ; In the formula: Features of the Lth layer; It is the ReLU activation function; for The degree matrix, ; A is the adjacency matrix, and I is the identity matrix; Let L be the weight matrix of the Lth layer; Global feature fusion: The embedding vector of the mining area is obtained by global average pooling of the single mining area map. After concatenating all mining area vectors, they are compressed into a unified shared intermediate feature through a fully connected layer, and then simultaneously supplied to the high-level and low-level reinforcement learning modules.

[0034] 7) Layered decomposition of long-term strategy and short-term mining operations to achieve intelligent decision-making for mining sequences in multiple mining areas; A two-layer hierarchical reinforcement learning model is constructed, consisting of a high-level manager and a low-level actuator. The high-level manager receives the shared intermediate features from the GNN output and generates a long-term mining strategic target vector through a dilated LSTM. The low-level actuator fuses the shared intermediate features from the GNN output with the high-level target vector and generates a mining and ore-matching action probability distribution based on LSTM. The high-level manager and the low-level executor each use a strategy gradient to update parameters; the high-level manager updates the strategy gradient with the full-cycle NPV return as the target, while the low-level executor updates the strategy gradient with the current ore blending and mining comprehensive return. (1) High-level manager: Responsible for formulating long-term strategic goals (with the goal of maximizing NPV). Receives the intermediate feature set from the GNN output and maps it to latent states through fully connected layers and ReLU. The target vector is then input into the dynamically dilated LSTM. and normalized to a unit vector. ; (2) Low-level executor: responsible for executing specific low-level actions, output by the perception module. The target vector set by the Manager The target vector serves as the basis for action design, i.e., block selection. Embedding is achieved by summing the objectives of the past c steps and passing them through an unbiased linear layer. Mapped to target embedding vector Action embedding outputs an action embedding matrix using a standard LSTM. The action probability distribution is generated by combining target embedding and action embedding through matrix multiplication and then passing the result through the SoftMax function. .

[0035] During training, the Manager updates the policy gradient to maximize external rewards. The Worker updates the policy gradient to maximize combined rewards, thus achieving optimal block selection.

[0036] ; ; ; In the formula, Embed the vector for the target; It is a linear layer without bias; Let t be the target vector at step t; Represents the hidden features in an RNN; Embed the action matrix; For standard LSTM; The output vector of the sensing module; This is the probability distribution vector of the action.

[0037] 8) Model training, outputting a complete multi-mining time series (mining sequence of blocks, storage and ore allocation scheme, and mining volume allocation for each cycle).

[0038] The A3C algorithm is used for training, with a set training step size. The reward function includes an external reward (weekly net profit and annual NPV scaling) and an internal reward (cosine similarity between the state changes of the higher-level manager and the target direction of the lower-level executor). Policy entropy coefficients and gradient pruning thresholds are set. After model training, the initial block state is input, and the model outputs a weekly mining plan sequence.

[0039] This invention decomposes complex long-term open-pit mine planning into high-level strategic objectives and low-level tactical actions using a hierarchical reinforcement learning architecture. This effectively addresses problems faced by traditional methods, such as the curse of dimensionality, local optima, and difficulties in long-term credit allocation. The graph neural network perception module accurately models the spatial topology and geological constraints of the ore blocks, compensating for the information loss inherent in traditional encoders. The collaborative ore blending strategy, through inter-period mixing of high- and low-grade ores, transforms a large amount of low-grade associated resources into economically recoverable reserves, significantly improving resource utilization and mine economic benefits. Compared to traditional optimization methods such as mixed-integer programming and particle swarm optimization, this invention can quickly generate high-quality long-term mining plans that meet all constraints, avoiding constraint violations and local optima traps. In real-world mine cases, it demonstrates significant net present value improvement. Furthermore, the model possesses strong generalization ability and environmental adaptability, requiring no human intervention and adaptable to open-pit mine planning scenarios involving multiple mining areas, multiple metals, and complex geological conditions. This enables a shift from human experience-based decision-making to data-driven intelligent decision-making, providing reliable technical support for smart mine construction and the green and efficient development of mineral resources.

[0040] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for planning the mining of polymetallic associated open-pit mines, characterized in that, Includes the following steps: 1) Construct a multi-mining area polymetallic ore block model to obtain the spatial coordinates, metal grade values, and ore block mass of each block; 2) Classify ore blocks according to their metal grade; Based on the grade thresholds of various metal industries, the ore blocks are divided into four categories: high-grade ore, general-grade ore, low-grade ore, and waste rock. 3) Establish a multi-mining area polymetallic mining planning model, wherein the objective function is to maximize the net present value (NPV) over the entire life cycle; 4) Establish a multi-mine-area polymetallic mining planning and ore blending strategy. Set up storage bins to buffer high-grade ore and mix high and low grades so that at least one metal reaches industrial grade after ore mixing. 5) Set constraints for multi-mining area polymetallic mining planning; 6) Construct a graph neural network (GNN); The input to a graph neural network is: node feature vectors and graph network topology. The output of the graph neural network is a shared intermediate feature set, which integrates block attributes and block neighborhood topology information. 7) Layered decomposition of long-term strategy and short-term mining operations to achieve intelligent decision-making for mining sequences in multiple mining areas; A two-layer hierarchical reinforcement learning model is constructed, consisting of a high-level manager and a low-level executor. The high-level manager receives shared intermediate features from the output of the GNN and generates a long-term mining strategic target vector through a dilated LSTM. The low-level executor fuses the shared intermediate features from the output of the GNN with the high-level target vector and generates a mining allocation action probability distribution based on LSTM. The high-level and low-level models update parameters using policy gradients. The high-level manager updates the policy gradient with the full-cycle NPV return as the objective, while the low-level executor updates the policy gradient with the current combined return of ore allocation and mining. 8) Complete model training and output a complete multi-mining time series.

2. The method for planning the mining of polymetallic associated open-pit mines according to claim 1, characterized in that, In step 2), High-grade ore means that the grade of any one metal in the ore block is higher than the dynamic high-grade multiple threshold; General grade ore indicates that the metal grade reaches the industrial grade, but does not reach the high grade multiple threshold. Low-grade ore means that the grade of all metals in the ore block is lower than the industrial grade. Waste rock indicates that the ore block contains no valuable metal grade.

3. The method for planning the mining of polymetallic associated open-pit mines according to claim 1, characterized in that, In step 4), the ore blending strategy is as follows: a storage bin is defined to store high-grade ore, and low-grade ore is matched to achieve blending. If the matching fails for multiple rounds, the storage bin is emptied. Low-grade ore is selected to be matched with the current high-grade ore or with the high-grade ore in the storage bin. If the matching fails for multiple rounds, it is treated as waste rock. General-grade ore is selected to be mixed with ore of the same stage under the condition of meeting the constraints for beneficiation, or enters the beneficiation plant alone. High-grade ore is selected to be mixed with ore of the same stage under the condition of meeting the constraints for beneficiation, or a portion is selected to be stored and enters the remaining beneficiation plant, or all of it enters the beneficiation plant.

4. The method for planning the mining of polymetallic associated open-pit mines according to claim 3, characterized in that, In step 4), the decision, state, and judgment functions for ore blending optimization are defined. The specific process of the ore blending strategy is as follows: 4.1) Decision Variables Decision variables include storage decisions and ore blending decisions, for each time period t, t=1, 2…t n The definition is as follows: Store decision variables ,Decide Whether to store, where 0 means no storage and all processing, and 1 means 15% storage and 85% processing; Decision option 0 indicates no ore blending; decision option 1 indicates... and Ore blending; Processing decision option 2 indicates With storage warehouse Ore blending; 4.2) State Variables The storage capacity of high-grade ore in the storage bin, with an initial value of 0; : Number of storage bin mixing failures at the start of time t; : Number of low-grade ore mixing failures at the start of time t; 4.3) Mixed judgment function Mixed decision function for processing decisions Calculate whether at least one metal reaches industrial grade after blending. Return 1 if blending is successful, otherwise return 0. In the formula, The amount of high-grade ore selected for the time period t includes high-grade ore in the storage bin and high-grade ore directly processed; The high-grade ore grade selected for the time period t; The amount of general-grade ore selected for the time period t; The typical grade of ore selected for the time period t; The amount of low-grade ore selected for the time period t; The grade of low-grade ore selected for the time period t.

5. The method for planning the mining of polymetallic associated open-pit mines according to claim 1, characterized in that, In step 5), the constraints include ore block reserve constraints, open-pit slope bench mining priority constraints, upper and lower limits of mine mining, storage, and beneficiation plant capacity constraints, and minimum grade constraints of polymetallic materials entering the plant.

6. The method for planning the mining of polymetallic associated open-pit mines according to claim 1, characterized in that, In step 5), the constraints are as follows: 5.1) Reserve Constraints The production volume of the ore block shall not exceed the reserves, and the sum of the ore volume and waste rock volume shall not exceed the total reserves of the ore block itself in the mining area; 5.2) Priority constraints for slope mining; Before mining a block at a lower elevation, the mining of a group of overlying blocks above it must be completed first. 5.3) Capacity, crushing, storage, and beneficiation plant constraints; The mining output, storage warehouse, and feed into the beneficiation plant must not exceed the upper and lower limits of the equipment's capacity. 5.4) Grade constraints: The ore sent to the beneficiation plant must meet the minimum grade requirements: The grade of the i-th metal entering the beneficiation plant within the t mining cycle must meet the minimum grade requirements of the beneficiation plant.

7. The method for planning the mining of polymetallic associated open-pit mines according to claim 1, characterized in that, In step 6), the graph neural network construction process is as follows: Graph network topology construction: Each mineral block is a graph node, and the node features include coordinates, grade, and quality; an adjacency matrix is ​​constructed based on the Euclidean distance and distance threshold between mineral blocks, and edges are established to connect spatially adjacent mineral blocks; Encoding is performed through a multi-layer graph convolutional network, followed by global pooling and feature fusion to generate a shared intermediate feature set; including: Global average pooling is performed on the output of the last layer of a single-layer GCN to compress the features of all block nodes in the mining area into a one-dimensional vector. Global feature fusion: The embedding vector of the mining area is obtained by global average pooling of the single mining area map. After concatenating all mining area vectors, they are compressed into a unified shared intermediate feature through a fully connected layer and simultaneously fed into the two-layer reinforcement learning module.

8. The method for planning the mining of polymetallic associated open-pit mines according to claim 1, characterized in that, In step 7), the hierarchical reinforcement learning model operates as follows: 7.1) High-level manager: The long-term strategic goal is to maximize NPV. It receives the intermediate feature set output by GNN, maps it to latent states through fully connected layers and ReLU, and then inputs it into the target vector output by dynamic dilated LSTM, which is then normalized to a unit vector. ; 7.2) Low-level executor: Responsible for executing specific low-level actions, output by the perception module. Target vector set by the high-level manager Action design, i.e., block selection, is based on data; target vector. Embedding is achieved by summing the objectives of the past c steps and passing them through an unbiased linear layer. Mapped to target embedding vector Action embedding outputs an action embedding matrix using a standard LSTM. Action probability distribution is generated by combining target embedding and action embedding through matrix multiplication and then passing the SoftMax function. ; ； ； ； In the formula, Embed the vector for the target; It is a linear layer without bias; Let t be the target vector at step t; Represents the hidden features in an RNN; Embed the action matrix; For standard LSTM; The output vector of the sensing module; This is the probability distribution vector of the action.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 8.

Images