Coal bulk handling efficiency optimization method and system based on machine learning

Abstract

The application provides a coal bulk cargo loading and unloading efficiency optimization method and system based on machine learning, relating to the technical field of coal bulk cargo loading and unloading, and the method comprises the following steps: identifying key factors and quantitative association rules between the key factors and the coal bulk cargo loading and unloading efficiency by using a large language model; inputting a stockyard decision topology graph into a reinforcement learning model to determine a coal bulk cargo loading and unloading scheme by executing a state space coding, action selection and reward function calculation process, the reward function is constructed according to the key factors and the quantitative association rules and aims to maximize the predicted value of the coal bulk cargo loading and unloading efficiency in the current period; after the coal bulk cargo loading and unloading scheme is completed, the actual value of the coal bulk cargo loading and unloading efficiency in the current period is collected, the efficiency difference value is calculated, and the coal bulk cargo loading and unloading efficiency in the next period is optimized based on the efficiency difference value. The application can effectively improve the efficiency while accurately meeting the necessary conditions for loading and unloading vehicle operation.

Description

Technical Field

[0001] This application relates to the technical field of bulk coal loading and unloading, and in particular to a method and system for optimizing the efficiency of bulk coal loading and unloading based on machine learning. Background Technology

[0002] Coal bulk cargo loading and unloading is a crucial link in the coal logistics supply chain. Loading and unloading operations, as the core process connecting ports and railways, involve multi-dimensional coordination, including train scheduling, coal type matching, stacker / reclaimer collaboration, and yard stack management. In this scenario, necessary conditions must be met, such as matching coal type with stack location, ensuring no conflict between stacker / reclaimer operations, limiting stack capacity, and prohibiting empty stack locations. Simultaneously, it is essential to consider requirements for efficient equipment utilization (e.g., reducing waiting time), energy conservation (e.g., shortening equipment movement distance), and safety (e.g., high-temperature restrictions, collision prevention) to achieve the core objective of minimizing total operation time. With the increase in coal transportation volume and the expansion of yard size, the constraints of unloading operations are becoming increasingly complex, placing higher demands on the development of precise and efficient operational plans.

[0003] In existing technologies, loading and unloading operations largely rely on manual scheduling and experience-based decision-making. Scheduling personnel must manually plan the unloading sequence and allocate stacking positions and stackers based on train forecasts (such as train number, coal type, and tonnage), arrival times from the railway transport system, the current status of stockpile locations (such as coal type, capacity, and equipment association), and equipment operating status. During the decision-making process, constraints such as coal type matching and stacker-reclaimer conflicts must be manually verified, and operational modes must be selected based on experience (such as prioritizing smaller-capacity stacking positions or shortening equipment movement distances). Finally, an operational plan is generated and issued to the site for execution.

[0004] However, existing technologies rely too heavily on human experience, which has significant limitations: on the one hand, human decision-making is difficult to handle multi-dimensional constraints in real time (such as equipment failure, high temperature restrictions, stacking capacity limits, etc.), which can easily lead to problems such as equipment occupation conflicts and unreasonable stacking allocation, increasing waiting time; on the other hand, human decision-making lacks quantitative basis for efficiency targets, making it difficult to achieve global optimization, and in complex scenarios such as concentrated train arrivals and dynamic changes in equipment status, decision-making efficiency is low and error rate is high, which cannot meet the high-efficiency and precision requirements of large-scale unloading operations. Summary of the Invention

[0005] The purpose of this application is to provide a machine learning-based method and system for optimizing the efficiency of bulk coal loading and unloading, in order to solve the problems of low decision-making efficiency and high error rate in the existing technology under complex scenarios such as concentrated train arrivals at ports and dynamic changes in equipment status, which cannot meet the needs of high efficiency and precision for large-scale unloading operations.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a machine learning-based method for optimizing the efficiency of bulk coal loading and unloading, comprising:

[0007] Collect planning data, equipment status and location data, stockyard resource occupancy data and stockyard constraint data of bulk coal loading and unloading operations to form a data set. The stockyard constraint data includes: coal type matching requirements, stacker-reclaimer operation conflict constraints, stack capacity limit, empty stack space usage restrictions, equipment failure association constraints and high temperature operation restrictions.

[0008] Based on the dataset, a constrained stockpile decision topology map is generated;

[0009] We use large language models to identify key factors related to the efficiency of bulk coal loading and unloading, and to quantify the correlation rules between these key factors and the efficiency of bulk coal loading and unloading.

[0010] Using a deep double-Q network as the framework of a reinforcement learning model, the yard decision topology is input into the reinforcement learning model. By executing state space encoding, action selection, and reward function calculation, a coal bulk loading and unloading scheme is determined. The reward function aims to maximize the predicted value of coal bulk loading and unloading efficiency in the current period and is constructed based on the key factors and the quantitative association rules.

[0011] After the bulk coal loading and unloading plan is completed, the actual value of the bulk coal loading and unloading efficiency in the current cycle is collected, the efficiency difference is calculated, and the bulk coal loading and unloading efficiency in the next cycle is optimized based on the efficiency difference.

[0012] Optionally, the step of inputting the yard decision topology map into the reinforcement learning model and determining the bulk coal loading and unloading scheme by performing state-space encoding, action selection, and reward function calculation includes:

[0013] The stockpile decision topology is input into the reinforcement learning model. The spatiotemporal graph convolutional network of the reinforcement learning model performs multi-level neighborhood aggregation on the node features and edge weights of the stockpile decision topology to obtain a dynamic state vector that integrates the real-time location of the equipment, the coal type attributes of the stack, the task queue, and the equipment fault status.

[0014] Based on a deep dual-Q network, the action value vector corresponding to the dynamic state vector is calculated. According to the action value vector with the highest value, the corresponding action is selected. The action includes: the stacker position coordinate allocation sequence of the stacker-reclaimer, the coal type identification configuration scheme of the conveyor belt, and the equipment movement scheme.

[0015] After performing the aforementioned action, determine the observation data, which includes the actual loading and unloading volume, equipment movement distance, and violation event count;

[0016] The reward value is calculated based on the observed data and the reward function, which is constructed based on the key factors and the quantitative correlation rules, and includes efficiency reward items, compliance reward items, mobility cost penalty items and energy saving reward items;

[0017] Based on the reward value, the action value is updated using temporal difference error. When the fluctuation range of the action value of N consecutive decisions is less than the preset amplitude threshold, the reinforcement learning model is determined to have converged. The output is a coal bulk loading and unloading scheme that includes the stacker coordinate allocation sequence of the stacker reclaimer, the coal type identification configuration scheme of the conveyor belt, and the equipment movement scheme, where N is greater than or equal to 3.

[0018] Optionally, based on the deep dual-Q network, the action value vector corresponding to the dynamic state vector is calculated. The corresponding action is selected based on the action value vector with the highest value. The actions include: the stacker-reclaimer's stack position coordinate allocation sequence, the conveyor belt's coal type identification configuration scheme, and the equipment movement scheme, including:

[0019] The dynamic state vector is input into a deep dual-Q network to output multiple types of action values, including: the action value of each stack position coordinate as the matching position of the stacker-reclaimer, the action value of each coal type identifier being activated on the conveyor belt, and the action value of each path sequence as the equipment movement scheme.

[0020] Based on the countdown timer for train arrival or ship berthing in the planning data, and combined with the user level, task level, and yard operation mode in the planning data, the urgency weight of the task is calculated.

[0021] Adjust the action value corresponding to the stacking position coordinates based on the task urgency weight and the priority corresponding to the yard operation mode.

[0022] The coordinates of the stack position with the highest action value are selected as the matching position of the stacker-reclaimer, the coal type identifier with the highest adjusted action value is selected as the configuration instruction of the conveyor belt, and the path sequence with the highest action value is selected as the equipment movement scheme to obtain the corresponding action.

[0023] Optionally, the step of calculating the task urgency weight based on the train arrival countdown or ship berthing countdown in the planning data, combined with the user level, task level, and yard operation mode in the planning data, includes:

[0024] Calculate the first difference between the estimated time of train arrival and the current time, or the second difference between the estimated time of ship arrival and the current time;

[0025] The ratio of the first difference to the preset time window is used as the countdown to the train's arrival at the station, or the ratio of the second difference to the preset time window is used as the countdown to the ship's berthing.

[0026] The user level, the task level, and the yard operation mode are quantified respectively to obtain the quantified values ​​of the user level, the task level, and the yard operation mode.

[0027] The sum of the train berthing countdown, user level quantification value, task level quantification value, and yard operation mode quantification value is used as the emergency weight for the unloading task, and the sum of the ship berthing countdown, user level quantification value, task level quantification value, and yard operation mode quantification value is used as the emergency weight for the loading task.

[0028] Optionally, a large language model is used to identify key factors related to the efficiency of bulk coal handling, and quantitative correlation rules between these key factors and the efficiency of bulk coal handling, including:

[0029] Calculate the Pearson correlation coefficient between all fields in the dataset and the loading and unloading efficiency index, and filter out structured fields and unstructured text fields with Pearson correlation coefficients greater than a preset threshold.

[0030] The structured fields and the unstructured text fields are mapped to knowledge graph nodes according to spatiotemporal coordinates. The edge weights between nodes are calculated based on historical operation data to form a knowledge graph with the loading and unloading process chain as the skeleton and carrying the loading and unloading process dependencies and edge weights.

[0031] Based on the knowledge graph, combined with the domain knowledge base and the preset causal template, a multi-round chain reasoning guided by the efficiency of coal bulk cargo loading and unloading is performed to output a list of key factors and quantitative association rules. The list of key factors includes multiple key factors.

[0032] Optionally, based on the knowledge graph, combined with the domain knowledge base and a preset causal template, multi-round chain reasoning oriented towards the efficiency of bulk coal loading and unloading is performed to output a list of key factors and quantitative association rules, including:

[0033] The knowledge graph nodes are matched with preset causal templates for similarity to filter out a set of candidate nodes that affect the efficiency of bulk coal loading and unloading.

[0034] Calculate the efficiency impact weights of candidate nodes;

[0035] For candidate nodes whose efficiency impact weight is greater than a preset weight threshold, the corresponding gradient significance test value is calculated by difference based on historical operation data.

[0036] Candidate nodes whose gradient significance test values ​​are greater than a preset test threshold are selected as key factors to output a list of key factors;

[0037] The correlation rules between key factors and coal bulk cargo loading and unloading efficiency are quantified by pre-setting a fitting method, and the quantified correlation rules are obtained.

[0038] Optionally, optimizing the coal bulk cargo loading and unloading efficiency in the next cycle based on the efficiency difference includes:

[0039] When the efficiency difference is less than or equal to the preset difference, the weights of the efficiency reward item, compliance reward item, moving cost penalty item and energy saving reward item in the reward function are adjusted according to the offset direction and magnitude of the efficiency difference, so as to optimize the coal bulk cargo loading and unloading efficiency in the next cycle according to the adjusted reward function.

[0040] Alternatively, when the efficiency difference is greater than a preset difference, the large language model is adjusted based on the prediction deviation sample set to optimize the coal bulk cargo loading and unloading efficiency in the next cycle based on the adjusted large language model.

[0041] Secondly, this application provides a machine learning-based coal bulk cargo loading and unloading efficiency optimization system, comprising:

[0042] The data acquisition module is used to collect planned data, equipment status and location data, yard resource occupancy data, and yard constraint data for bulk coal loading and unloading operations, forming a data set.

[0043] The generation module is used to generate a constrained stockpile decision topology map based on the data set;

[0044] The identification module is used to identify key factors related to the efficiency of bulk coal loading and unloading using a large language model, as well as the quantitative correlation rules between key factors and the efficiency of bulk coal loading and unloading.

[0045] The reinforcement learning module is used to use a deep double-Q network as the framework of the reinforcement learning model. The decision topology map of the yard is input into the reinforcement learning model. The module performs state space encoding, action selection and reward function calculation. The reward function aims to maximize the predicted value of coal bulk loading and unloading efficiency in the current period and is constructed based on the key factors and the quantitative association rules to determine the coal bulk loading and unloading scheme.

[0046] The calculation and optimization module is used to collect the actual value of the coal bulk cargo loading and unloading efficiency in the current period after the coal bulk cargo loading and unloading plan is completed, calculate the efficiency difference, and optimize the coal bulk cargo loading and unloading efficiency in the next period based on the efficiency difference.

[0047] Thirdly, this application provides an electronic device, comprising:

[0048] Memory, used to store computer programs;

[0049] A processor, configured to execute the computer program to implement the steps of a machine learning-based method for optimizing the efficiency of bulk coal loading and unloading as described in the first aspect above.

[0050] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the machine learning-based coal bulk cargo loading and unloading efficiency optimization method described in the first aspect above.

[0051] This application provides a machine learning-based method for optimizing the efficiency of bulk coal loading and unloading, comprising: collecting planned data, equipment status and location data, yard resource occupancy data, and yard constraint data for bulk coal loading and unloading operations to form a dataset. The yard constraint data includes: coal type matching requirements, stacker-reclaimer operation conflict constraints, stack capacity limits, empty stack space usage restrictions, equipment failure correlation constraints, and high-temperature operation restrictions. Based on the dataset, a constrained yard decision topology map is generated. A large language model is used to identify key factors related to the efficiency of bulk coal loading and unloading, and the relationship between these key factors and the efficiency of bulk coal loading and unloading. Quantitative correlation rules between efficiencies are established. A deep double-Q network is used as the framework for a reinforcement learning model. The yard decision topology is input into the reinforcement learning model. By executing state-space encoding, action selection, and reward function calculation, a coal bulk loading and unloading scheme is determined. The reward function aims to maximize the predicted value of coal bulk loading and unloading efficiency in the current period and is constructed based on key factors and quantitative correlation rules. After the coal bulk loading and unloading scheme is completed, the actual value of coal bulk loading and unloading efficiency in the current period is collected, the efficiency difference is calculated, and the coal bulk loading and unloading efficiency in the next period is optimized based on the efficiency difference.

[0052] The technical solution adopted in this application has the following advantages:

[0053] (1) This application generates a constrained stockpile decision topology by collecting stockpile constraint data, including coal type matching requirements, stacker-reclaimer operation conflict constraints, stack capacity limit (27,000 tons), empty stack space usage restrictions, equipment failure association constraints, and high temperature operation restrictions (no unloading above 40 degrees Celsius), and makes scheme decisions based on this. This ensures that all unloading operations strictly follow the necessary conditions and avoids violations caused by human judgment oversights (such as loading coal into empty stack spaces, stacker-reclaimers operating simultaneously in the same stack space), fundamentally guaranteeing the compliance and stability of operations.

[0054] (2) Based on the key factors (such as stacker travel distance, stacking capacity, equipment occupancy conflict, etc.) and quantitative association rules identified by the large language model, combined with the reward function (covering dimensions such as efficiency, energy saving, and travel cost) aimed at maximizing loading and unloading efficiency, the reinforcement learning model can automatically generate the optimal unloading plan: prioritize the realization of "all trains operating at the same time" to reduce waiting time and shorten the total operation time, which is in line with the core indicator of "the shortest total operation time is optimal"; directly respond to energy saving needs and reduce equipment energy consumption and losses by minimizing stacker travel distance and prioritizing the smallest stacking; balance equipment occupancy conflict, avoid efficiency loss caused by resource competition, and improve equipment utilization.

[0055] (3) The reinforcement learning model of this application can be flexibly adapted to different operation modes according to the actual situation of the yard: in the preparation mode, priority is given to the stacking position with small yard capacity; in the balance mode, the cargo volume of the BQ line stacking position is uniform; in the high-production mode, the focus is on shortening the unloading time. There is no need to manually switch the decision logic, which significantly improves the adaptability to complex yard scenarios and ensures that efficient solutions can be output under different operation objectives.

[0056] (4) By collecting the actual efficiency value after the operation is completed, calculating the efficiency difference and feeding it back to the model optimization process (adjusting the reward function weight or large language model parameters), this application can realize the dynamic iteration of the decision-making scheme. In response to emergencies such as equipment failure and coal type changes, the model can continuously optimize the decision logic through continuous learning, avoid the limitations caused by fixed experience, and ensure the efficiency and adaptability of unloading operations in the long term.

[0057] (5) The plan incorporates safety conditions such as collision avoidance (e.g., a one-staple interval between stacking and picking) and high-temperature operation restrictions, and ensures implementation through automatic model verification, reducing potential safety hazards that may be overlooked by manual decision-making, and providing stable safety assurance for unloading operations.

[0058] (6) Furthermore, this method integrates dynamic information such as equipment, stacking position, and task, adjusts the action value according to the urgency weight of the task and the operation mode, selects the optimal stacking position allocation, conveyor belt configuration and equipment movement scheme, and optimizes the decision by combining a reward function that includes efficiency, compliance, movement cost and energy saving. It can meet the necessary conditions for unloading operation, reduce waiting and movement costs, adapt to multi-mode needs, and improve efficiency and energy saving effect. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0060] Figure 1 A flowchart illustrating a machine learning-based method for optimizing the loading and unloading efficiency of bulk coal, provided in an embodiment of this application;

[0061] Figure 2 A schematic diagram illustrating an application scenario of a machine learning-based method for optimizing the loading and unloading efficiency of bulk coal, provided in an embodiment of this application.

[0062] Figure 3 This is a schematic diagram of a machine learning-based coal bulk cargo loading and unloading efficiency optimization system provided in an embodiment of this application. Detailed Implementation

[0063] To address the problems of low decision-making efficiency and high error rates in existing technologies under complex scenarios such as concentrated train arrivals and dynamic changes in equipment status, which fail to meet the demands for high efficiency and precision in large-scale unloading operations, this application provides a machine learning-based method for optimizing the efficiency of bulk coal loading and unloading. This method employs the following concept: It collects operational plans, equipment status, yard resources, and various constraints (such as coal type matching, equipment conflicts, and high-temperature operation restrictions) to form an overall decision-making basis; then, it utilizes intelligent models to analyze key influencing factors and patterns, ultimately calculating the optimal loading and unloading scheme, and continuously optimizing it based on actual results. This approach, on the one hand, automates the handling of complex constraints, avoiding oversights in manual decision-making and reducing equipment conflicts and waiting times; on the other hand, it achieves globally optimal decision-making through data-driven quantitative analysis, maintaining high efficiency and accuracy even in complex scenarios, thus effectively compensating for the shortcomings of traditional manual scheduling.

[0064] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0065] The core of this application is to provide a machine learning-based method for optimizing the efficiency of bulk coal loading and unloading. A flowchart illustrating one specific implementation is shown below. Figure 1 As shown, the method includes:

[0066] S11. Collect planned data, equipment status and location data, yard resource occupancy data and yard constraint data for bulk coal loading and unloading operations to form a data set. The yard constraint data includes: coal type matching requirements, stacker-reclaimer operation conflict constraints, stack capacity limit, empty stack space usage restrictions, equipment failure association constraints and high temperature operation restrictions.

[0067] In step S11, the bulk coal loading and unloading operation includes bulk coal unloading from trains and bulk coal loading onto ships. Bulk coal unloading refers to the temporary storage of coal from trains to specific stacks within the stockyard using various intelligent devices. Bulk coal loading refers to the loading of coal from specific stacks within the stockyard onto ships using various intelligent devices. These intelligent devices include stacker cranes, reclaimers, conveyor belts, tippers, ship loaders, intelligent monitoring cameras, IoT sensors, and a central control system. Figure 2As shown, there are four tipplers, CD1, CD2, CD3, and CD4, which begin operation after the train arrives. There are also four belt conveyors (BF1, BF2, BF3, and BF4) below the tippler hoppers, used to transport coal. Each belt is equipped with a belt scale corresponding to the tippler, coded BSBF1, BSBF2, BSBF3, and BSBF4, used to obtain relevant electronic scale information for the corresponding tippler. This information includes instantaneous values, cumulative values, and the number of tippler sections. For safe transportation, this embodiment of the application provides belt conveyor transfer towers at different belt positions: T4, T5, T6, T7, T8, T9, T10, T11, etc.; BH1-3, BH1-4, BH2-3, BH2-4, BH3-3, BH3-4, BH4-3, BH4-4 represent the codes of the intermediate belts of different turning lines in the corresponding longitudinal direction. There are 3 intermediate belts for the loading and unloading lines: BJ1-1, BJ2-1, and BJ3-1; MSBQ1, MSBQ2, and MSBQ3 represent the corresponding electromagnetic separators. There are 4 stacker belts: BD1, BD2, BD3, and BD4. There are 4 reclaimer belts: BQ1, BQ2, BQ3, and BQ4. There are 3 ship loaders: SL1, SL2, and SL3; and 3 conveyor belts connecting the ship loaders: BM1, BM2, and BM3. This application includes four stacker cranes in the stockyard, namely S1, S2, S3, and S4. It also includes multiple reclaimers in the stockyard: R1-1, R1-2, R2-1, R2-2, R3-1, and R3-2. For example, the stockyard can have 6 rows and 4 columns, thus the number of stockyard positions can be 24, such as stockyard position 101, stockyard position 102, stockyard position 103, ..., stockyard position 604, etc. There are multiple types of coal, such as coal type A, coal type B, coal type C, coal type D, and coal type E. The system includes a stacker for storing unloaded coal to designated stacking positions; a reclaimer for retrieving coal from the stacking positions for loading onto ships; a conveyor belt for transporting coal between different equipment, stacking positions, and ships; a tipper for automatically unloading coal from train cars onto the conveyor belt; a ship loader for precisely loading the transported coal into the ship's hold; intelligent monitoring cameras for real-time monitoring of the work site; IoT sensors for collecting data such as equipment operating parameters, coal flow rate, and stacking capacity; and a central control system for unified coordination and remote control of all types of equipment, ensuring automated and intelligent operation of unloading and loading operations.

[0068] Furthermore, the planning data includes forecast information for trains / ships, such as train number, ship name, coal type, tonnage, and estimated arrival / berthing time; equipment status and location data can refer to the real-time operating status (e.g., normal / faulty) and location (e.g., stack coordinates, track position, etc.) of equipment such as stackers and reclaimers; stockpile resource occupancy data refers to information such as the current coal type, occupied quantity, and remaining capacity of each stack; stockpile constraint data are the rules that must be followed in operations, including coal type matching requirements (including: specific coal types can only be stacked in corresponding stacks, and only the same coal type can be stacked in the same stack), stacker and reclaimer operation conflict constraints (such as stacking and reclaiming cannot be carried out simultaneously in the same stack), stack capacity limit (i.e., the maximum coal loading capacity of a single stack, such as 27,000 tons), empty stack usage restrictions (such as stacks without associated coal types cannot be loaded), equipment fault association constraints (such as stacks corresponding to faulty equipment cannot be used), and high temperature operation restrictions (unloading is not allowed above a certain temperature, such as stacks above 40 degrees Celsius are not operated).

[0069] In this embodiment, firstly, planned data is collected from the train dispatching system and the ship forecasting system, such as the train number of train K, the coal type being bituminous coal, the tonnage being 5000 tons, and the estimated arrival time being 10:00. Secondly, the status (normal operation) and position (currently at stack position 2) of the stacker-reclaimer B and the status (normal operation) of the conveyor belt C are collected through equipment sensors. Next, stack position resource data is retrieved from the stockyard management system, such as stockyard resource occupancy data for stack position 1 (bituminous coal, 10000 tons already occupied, upper limit 27000 tons), stack position 2 (anthracite, 8000 tons already occupied), and stack position 3 (empty stack position, no associated coal type). Finally, preset stockyard constraint data is organized, such as "bituminous coal can only be stacked at stack positions 1, 4, and 5", "stacker-reclaimers cannot operate at the same stack position simultaneously", and "unloading is not allowed above 40°C", etc., and all data are summarized into a data set.

[0070] S12. Based on the dataset, generate a constrained yard decision topology diagram.

[0071] In step S12, the constrained stockpile decision topology diagram is a structure that graphically displays the elements and constraints related to stockpile operations. The “nodes” can represent train / ship tasks, equipment (stack reclaimers, conveyor belts), stack locations, etc., while the “edges” represent the associations or constraints between nodes (such as a stack reclaimer can reach a certain stack location, a certain type of coal matches a certain stack location, a certain piece of equipment conflicts with a certain stack location, etc.).

[0072] In this embodiment, based on the data set formed in S11, various elements are transformed into nodes of the topology graph. For example, train K (task node), stacker-reclaimer B (equipment node), and stack position 1 (stack position node) are used as nodes. Then, edges between nodes are drawn according to the constraints in the data. For example, the coal type (bituminous coal) of train K matches the coal type of stack position 1, and they are connected with solid lines and marked "coal type matched". Stacker-reclaimer B is currently at stack position 2, and there is no conflict in the movement path to stack position 1. They are connected with dashed lines and marked "movable". Stack position 3 is an empty stack position without associated coal type. It is connected to train K (bituminous coal) with an edge marked "×" and marked "empty stack position restricted". Finally, a topology graph containing all tasks, equipment, stack positions and constraints is formed.

[0073] S13. Use a large language model to identify key factors related to the efficiency of bulk coal loading and unloading, and quantitative correlation rules between key factors and the efficiency of bulk coal loading and unloading.

[0074] In step S13, the large language model is an artificial intelligence model that can process structured data and text information and identify patterns; key factors refer to factors that have a significant impact on loading and unloading efficiency (such as equipment moving distance, remaining capacity of stacking positions, urgency of tasks, etc.); quantitative association rules refer to the mathematical relationship between key factors and loading and unloading efficiency (such as "for every 50 meters increase in equipment moving distance, loading and unloading time increases by 3 minutes"); Pearson correlation coefficient is used to measure the degree of linear correlation between two variables (the closer the value is to 1 or -1, the stronger the correlation).

[0075] In the embodiments of this application, the specific description of step S13 can be referred to steps 131 to 133 below, and will not be repeated here.

[0076] S14. Using a deep double-Q network as the framework of a reinforcement learning model, the decision topology of the yard is input into the reinforcement learning model. By executing the process of state space encoding, action selection and reward function calculation, the coal bulk loading and unloading scheme is determined. The reward function aims to maximize the predicted value of coal bulk loading and unloading efficiency in the current cycle and is constructed based on key factors and quantitative association rules.

[0077] In S14, reinforcement learning is an artificial intelligence model that optimizes decision-making through interaction with the environment and trial-and-error learning. A deep double-Q network serves as its framework, consisting of two neural networks to reduce decision bias. The reward function evaluates the quality of actions, including efficiency rewards (points for shorter travel times), compliance rewards (points for adhering to constraints), movement cost penalties (points for longer travel distances), and energy-saving rewards (points for energy-efficient actions).

[0078] In the embodiments of this application, the specific description of step S14 can be referred to steps 141 to 145 below, and will not be repeated here.

[0079] S15. After the coal bulk cargo loading and unloading plan is completed, collect the actual value of the coal bulk cargo loading and unloading efficiency in the current cycle, calculate the efficiency difference, and optimize the coal bulk cargo loading and unloading efficiency in the next cycle based on the efficiency difference.

[0080] The preset difference value can be set comprehensively based on the accuracy requirements of the actual operation scenario, the fluctuation range of historical data, and business objectives, and is not a fixed value. For example, in coal bulk cargo loading and unloading operations, if "operation time" is used as an efficiency metric, combined with the industry's common operational accuracy requirements, the preset difference value may be set to 5-10 minutes (the specific value can be adjusted according to the port's average daily workload, equipment operational stability, etc.): when the difference between the actual operation time and the predicted time is within 5 minutes, the model prediction deviation is considered small, and only minor adjustments to the reward function weights are needed; if the difference exceeds 10 minutes, the deviation is considered large, and the key factor identification logic needs to be optimized by adjusting the large language model. The core is to ensure that the preset difference value can reflect the acceptable error range of the model prediction and trigger an effective correction mechanism in a timely manner, ensuring that the model continuously adapts to changes in actual operations.

[0081] Specifically, this step adjusts the weights of the efficiency reward item, compliance reward item, movement cost penalty item, and energy-saving reward item in the reward function according to the offset direction and magnitude of the efficiency difference when the efficiency difference is less than or equal to a preset difference. This optimizes the coal bulk loading and unloading efficiency in the next cycle based on the adjusted reward function. Alternatively, when the efficiency difference is greater than the preset difference, the large language model is adjusted based on the prediction deviation sample set to optimize the coal bulk loading and unloading efficiency in the next cycle. The preset deviation sample set includes samples corresponding to mode selection deviation, constraint satisfaction deviation, etc. It should be noted that the formulas or expressions for each item in the reward function are not specifically limited in this embodiment. Furthermore, the efficiency difference in the current cycle can optimize the coal bulk loading and unloading efficiency in the next cycle. Similarly, the reward function and large language model used in the current cycle are obtained after modification based on the efficiency difference in the previous cycle.

[0082] For example, if the actual efficiency is better than the predicted efficiency and is mainly due to the short equipment movement distance, the weight of the movement cost penalty term can be appropriately increased, so that the model will focus more on shortening the equipment movement distance in the next cycle. If the efficiency difference is greater than the preset difference, it indicates that the model prediction deviation is large. At this time, it is necessary to collect such samples with large prediction deviations to form a prediction deviation sample set. Based on this sample set, the large language model used to identify key factors and quantify association rules in S13 can be adjusted. For example, the model can be retrained to correct the judgment of key factors or the quantification relationship of association rules, so that the adjusted large language model can more accurately explore the factors and patterns affecting efficiency, thereby making the loading and unloading scheme optimization in the next cycle more in line with the actual operation, and finally achieving the optimization of coal bulk loading and unloading efficiency in the next cycle.

[0083] Based on the above process, this application provides the following specific example: Port A receives an unloading task from train K, train number A01, transporting bituminous coal in a quantity of 5,000 tons, expected to arrive at the port at 10:00. Subsequently, Port A initiates a machine learning-based coal bulk cargo loading and unloading efficiency optimization process.

[0084] First, comprehensive data collection was conducted in phase S11. Planned data for train K was obtained from the railway dispatching system. The status and location of stacker-reclaimers B and D were collected via IoT devices. Stacker-reclaimer B was in normal condition and located at stack position 2, while stacker-reclaimer D was malfunctioning and its corresponding stack position 6 was disabled. Stack position information was retrieved from the stockpile system: stack position 1 stored bituminous coal, with 10,000 tons already stored and a maximum capacity of 27,000 tons; stack position 4 stored bituminous coal, with 15,000 tons already stored; stack position 3 was empty and not associated with any coal type. Simultaneously, constraint rules were established: bituminous coal can only be stored at stack positions 1, 4, and 5; stacker-reclaimers cannot operate simultaneously at the same stack position; and operations are prohibited when the ambient temperature exceeds 40℃. This resulted in a complete data set.

[0085] Next, in stage S12, a constrained stockpile decision topology map is generated based on the data. Train K is designated as the task node, stacker-reclaimer B as the equipment node, and stack positions 1, 4, and 3 as stack position nodes, collectively forming the core nodes. A "coal type matching" edge connects train K to stack positions 1 and 4, a "vacant stack position restriction" edge connects train K to stack position 3, and a "movable" edge marks the path from stacker-reclaimer B to stack positions 1 and 4. The remaining capacity of stack positions 1 and 4 is also marked, ensuring that both meet the storage requirement of 5000 tons.

[0086] Subsequently, in stage S13, key factors were mined using a large language model. Based on historical bituminous coal unloading data from Port A, the Pearson correlation coefficients of "equipment movement distance," "remaining capacity of the stack," and total operation time were calculated to be 0.7 and 0.65, respectively, both exceeding the threshold of 0.6. After mapping these to knowledge graph nodes, multiple rounds of reasoning determined these two factors to be key factors. The quantified association rule was: Total operation time = 20 minutes + 0.1 minutes / meter × movement distance - 0.001 minutes / ton × remaining capacity, meaning that the shorter the movement distance and the larger the remaining capacity, the shorter the operation time.

[0087] In stage S14, the reinforcement learning model generates the optimal solution. First, a spatiotemporal graph convolutional network aggregates topological graph features to obtain a dynamic state vector. This vector contains the following information: stacker-reclaimer B is at position 2, position 1 has 17,000 tons of bituminous coal remaining, and train K has a 30-minute countdown. The deep double-Q network initially outputs "stacker-reclaimer B to position 1" with a value of 80 points and "to position 4" with a value of 70 points. After adjusting the emergency weight of train K by 0.8, the action value for position 1 increases to 85 points. After executing this action, the observed movement distance is 50 meters, there are no violations, and the loading / unloading volume is 5,000 tons. A reward of 30 points is calculated, including 15 points for efficiency, 10 points for compliance, and 5 points for movement cost. After three iterations, the model's decision fluctuation is less than 3%, and the converged output solution is for stacker-reclaimer B to move from position 2 to position 1 along a straight path, with the conveyor belt marked with bituminous coal.

[0088] Finally, dynamic optimization was performed in stage S15. The actual unloading time for train K was 28 minutes, which is -2 minutes less than the predicted value of 30 minutes. This difference is less than the preset difference of 5 minutes, so the weight of the efficiency reward term in the reward function was increased. In the next cycle, when processing the same type of task, the actual time of the model's output solution was further reduced to 26 minutes, achieving continuous efficiency optimization.

[0089] The method provided in this application, by performing all the above steps, can have the following advantages:

[0090] (1) Achieve intelligent decision-making throughout the entire process, breaking through the limitations of human experience. Specifically, through S11, comprehensive data collection (plans, equipment, yard, constraints) provides a complete information foundation for decision-making; S12 transforms the data into a visual topology map, intuitively presenting the relationship and constraints between tasks, equipment, and stacking positions; S13 uses a large language model to mine key factors (such as equipment movement distance and remaining stacking capacity) and quantification rules (such as "for every 50 meters the movement distance is shortened, the operation time is reduced by 5 minutes"), replacing human experience as the basis for decision-making; S14 automatically generates the optimal solution through a reinforcement learning model, comprehensively considering efficiency, compliance, energy saving, and other objectives to select actions (such as the optimal path for the stacker-reclaimer and the configuration of the conveyor belt). The entire process requires no human intervention, completely eliminating the dependence on experience and making decisions more accurate and scientific.

[0091] (2) Improve loading and unloading efficiency and resource utilization, and reduce operating costs. Specifically, through the reinforcement learning model in S14, the solution can prioritize efficient actions such as shortening equipment movement distance and matching stacks with sufficient remaining capacity (e.g., in the example, the stacker-reclaimer selects stack 1 with a shorter movement distance to reduce ineffective movement), directly shortening the total operation time. At the same time, the model adjusts the value of actions through the urgency weight of tasks to ensure that high-priority tasks (such as trains that are about to arrive) are processed first, reducing waiting time. Combined with the dynamic optimization in S15 (such as increasing the weight of efficiency reward items), the solution is continuously iterated, further shortening the operation time of subsequent similar tasks (in the example, it is optimized from 28 minutes to 26 minutes), significantly improving the overall loading and unloading efficiency. In addition, resource allocation is more reasonable (such as avoiding equipment conflicts and making full use of effective stacks), reducing equipment idleness and resource waste, and indirectly reducing operating costs.

[0092] (3) Strictly adhere to operational constraints to ensure operational safety and standardization. Specifically, the scheme explicitly includes constraint data such as coal type matching, stacker-reclaimer conflict, and high-temperature restrictions in S11. The topology map in S12 visually marks constraint relationships (such as empty stack space restrictions and unreachable paths for equipment). The reward function in S14 ensures that the scheme strictly complies with the rules (e.g., no violations in the example) through compliance reward items (points for compliance with constraints) and violation penalties. This effectively avoids the risk of violations caused by human oversight (such as loading coal into empty stack spaces or high-temperature operations) and ensures operational safety. At the same time, the unified constraint execution standard makes the operation process more standardized and reduces errors caused by inconsistent operations.

[0093] (4) Taking into account energy conservation goals and promoting green operation mode, the method is to set energy conservation reward items (such as bonus points for shorter equipment movement distance) in the reward function of S14, and the key factors identified in S13 also include energy conservation related indicators (such as equipment movement distance), guiding the model to prioritize energy conservation actions (such as the stacker-reclaimer choosing the shortest path in the example). By reducing ineffective equipment movement and optimizing resource allocation, energy consumption and equipment wear are reduced, promoting the transformation of port loading and unloading operations to a green and low-carbon mode.

[0094] (5) Dynamically adapts to complex scenarios and has continuous optimization capabilities. Specifically, in the face of complex scenarios such as concentrated train arrivals, sudden equipment failures, and changes in ambient temperature, the dynamic optimization mechanism of S15 can adjust the model according to the difference between actual efficiency and prediction: when the deviation is small, the weight of the reward function is fine-tuned (such as giving more emphasis to efficiency); when the deviation is large, the key factor identification logic of the large language model is corrected. This closed-loop optimization capability enables the solution to continuously adapt to changes in the scenario and maintain a high-efficiency decision-making state in the long term, solving the problem of slow response and poor adaptability of traditional manual scheduling in dynamic scenarios.

[0095] (6) It provides technical support for the construction of smart ports and has industry promotion value. Specifically, the solution achieves unmanned and refined operation of bulk coal loading and unloading through the whole-chain intelligentization of "data collection-intelligent analysis-automatic decision-making-dynamic optimization", which meets the needs of improving the level of intelligence in the construction of "smart ports". Its technical ideas (such as constraint visualization and machine learning-driven decision-making) can be extended to other bulk cargo ports or industrial scenarios (such as production organizations with "plan-execution" architecture), providing a reusable solution for the intelligent upgrading of the industry.

[0096] In summary, this method comprehensively improves the efficiency, safety, standardization, and energy conservation of bulk coal loading and unloading through an intelligent, data-driven, and dynamic decision-making mechanism. It also has strong adaptability and promotional value, providing strong support for the efficient operation and intelligent transformation of bulk cargo ports.

[0097] In some possible embodiments, S13, using a large language model to identify key factors related to the efficiency of bulk coal loading and unloading, and the quantitative correlation rules between key factors and the efficiency of bulk coal loading and unloading, includes the following process:

[0098] Step 131: Calculate the Pearson correlation coefficient between all fields in the dataset and the loading and unloading efficiency index, and filter out structured fields and unstructured text fields with Pearson correlation coefficients greater than the preset coefficient threshold.

[0099] In step 131, the loading and unloading efficiency indicators include total operation time, equipment energy consumption, etc. The Pearson correlation coefficient is an indicator that measures the degree of linear correlation between two variables (such as a field and loading and unloading efficiency); the closer the value is to 1 or -1, the stronger the correlation. Structured fields refer to quantifiable numerical data (such as equipment movement distance, remaining stack capacity, real-time equipment operating rate, conveyor belt switching time, yard turnover timestamps, temperature data, etc.); unstructured text fields refer to descriptive information (such as loading and unloading operation instructions, equipment fault records, operation logs, scheduling delay logs, ship loading feedback forms, etc.). The preset coefficient threshold is the standard for judging whether the correlation is significant (e.g., 0.6; values ​​exceeding this value are considered correlated). Loading and unloading efficiency indicators are quantitative indicators that measure operational efficiency (such as total operation time, loading and unloading volume per unit time).

[0100] In this embodiment, firstly, all fields (including structured fields such as "stack reclaimer travel distance" and "remaining capacity of stack position", and unstructured text fields such as "number of equipment failures" and "operating environment temperature record") are extracted from the data set formed in S11; secondly, the Pearson correlation coefficient between each field and the loading and unloading efficiency index (such as total operating time) is calculated. This embodiment does not specify the formula for calculating the Pearson correlation coefficient; finally, fields with correlation coefficients greater than a preset coefficient threshold (such as 0.6) are selected as candidate factors for subsequent analysis.

[0101] Step 132: Map structured fields and unstructured text fields to knowledge graph nodes according to spatiotemporal coordinates, calculate the edge weights between nodes based on historical operation data, and form a knowledge graph with the loading and unloading process chain as the skeleton, carrying the loading and unloading process dependencies and edge weights.

[0102] In step 132, spatiotemporal coordinates are identifiers combining time and spatial location (e.g., "the stacker-reclaimer is at stack position 2 at 9:30"); a knowledge graph is a graph that represents entities and relationships using nodes and edges. Here, nodes are selected fields, and edges represent dependencies and weights between fields. Knowledge graph nodes are mapped structured / unstructured fields (e.g., "stacker-reclaimer moving distance" and "remaining capacity of stack position"); edge weights are numerical values ​​that measure the strength of dependencies between nodes (higher weights indicate stronger dependencies); the loading and unloading process chain refers to the complete process of coal loading and unloading (e.g., unloading → stacking → reclaiming → loading onto ship). Loading and unloading process dependencies can include coal type matching dependencies, equipment conflict dependencies, etc.

[0103] In this embodiment, firstly, the structured and unstructured text fields selected in step 131 are mapped according to spatiotemporal coordinates. For example, "Stacker-reclaimer moving distance 50 meters" is mapped to the node "10:00, Stacker-reclaimer B, moving distance 50 meters", and "Remaining capacity of stack 1 17,000 tons" is mapped to the node "10:00, Stack 1, remaining capacity 17,000 tons". Secondly, the edge weights between nodes are calculated based on historical operation data. For example, by analyzing the co-occurrence frequency and influence of "Stacker-reclaimer moving distance" and "Total operation time" in historical data, the edge weight between the two nodes is determined to be 0.8. Finally, the nodes are connected according to the process flow using the loading and unloading process chain (such as "train unloading → stacker stacking → stack storage") as the skeleton to form a knowledge graph carrying process dependencies and edge weights.

[0104] Step 133: Based on the knowledge graph, combined with the domain knowledge base and the preset causal template, perform multi-round chain reasoning guided by the efficiency of coal bulk cargo loading and unloading, and output a list of key factors and quantitative association rules. The list of key factors includes multiple key factors.

[0105] In step 133, the domain knowledge base is a database storing professional knowledge of the coal loading and unloading industry, such as "the shorter the movement distance of the stacker-reclaimer, the higher the operation efficiency". It may also include equipment technical parameters, loading and unloading process rules, historical optimization cases, unloading energy-saving specifications, safety operation standards, etc. The preset causal template is a standardized sentence for reasoning causal relationships, such as "if factor X increases, then efficiency index Y changes", for example: [weather type] affects [equipment efficiency index, which in turn affects [total loading and unloading time]; multi-round chain reasoning refers to the process of gradually deriving deep associations based on initial relationships; gradient significance test value is a statistical measure of the significance of the factor's impact on efficiency; quantitative association rule is a mathematical expression between key factors and efficiency (such as "total operation time = a + b × factor 1 + c × factor 2"), where a, b, and c are coefficients and their values ​​are constants. This application embodiment does not specifically limit this. The list of key factors includes several key factors such as: reclaimer idle rate, conveyor belt switching frequency, coal type differences between adjacent stockpiles, equipment movement distance, waiting time for changing compartments, coal type matching degree, equipment movement distance, stacking capacity utilization rate, operating mode parameters, and ambient temperature.

[0106] Specifically, step 133 may include the following process: Step a1: Match the knowledge graph nodes with the preset causal templates to filter out a set of candidate nodes that affect the efficiency of bulk coal loading and unloading; Step a2: Calculate the efficiency impact weight of the candidate nodes; Step a3: For candidate nodes whose efficiency impact weight is greater than the preset weight threshold, calculate the corresponding gradient significance test value based on historical operation data; Step a4: Filter candidate nodes whose gradient significance test value is greater than the preset test threshold as key factors to output a list of key factors; Step a5: Quantify the association rules between key factors and bulk coal loading and unloading efficiency through a preset fitting method to obtain the quantified association rules.

[0107] In this embodiment, firstly, the knowledge graph nodes formed in step 132 are matched with a preset causal template (e.g., "If the stacker-reclaimer travels a greater distance, the total operation time will be extended") to screen out a set of candidate nodes that may affect loading and unloading efficiency. Secondly, the efficiency impact weight of the candidate nodes is calculated (based on the edge weights of the knowledge graph and the domain knowledge base). For example, the weight of "stacker-reclaimer travel distance" is 0.7, and the weight of "remaining capacity of the stacking position" is 0.6. Next, for candidate nodes with weights higher than a preset weight threshold (e.g., 0.5), the gradient significance test value is calculated based on the difference of historical operation data, and nodes with test values ​​greater than a preset test threshold (e.g., 1.96) are selected as key factors. Finally, the association rule between key factors and loading and unloading efficiency is quantified by a preset fitting method such as linear regression to obtain an expression such as "total operation time = 20 + 0.1 × stacker-reclaimer travel distance - 0.001 × remaining capacity of the stacking position". This embodiment does not specifically limit the calculation formula of the gradient significance test value.

[0108] In steps 131-133, the Pearson correlation coefficients of all fields in the dataset (such as equipment movement distance, stack capacity, coal type matching degree, etc.) with loading and unloading efficiency indicators (such as total operation time) are calculated, and fields with correlation coefficients greater than 0.6 (preset threshold) (such as equipment movement distance and remaining stack capacity) are selected. Next, these fields are mapped to knowledge graph nodes according to spatiotemporal coordinates (such as "stack reclaimer B is at stack 2 at 10:00"). The edge weights between nodes are calculated based on historical operation data (such as the edge weight between "equipment movement distance" and "total operation time" being 0.8), forming a knowledge graph based on loading and unloading efficiency indicators (such as equipment movement distance and remaining stack capacity). A knowledge graph with the unloading process chain as its backbone is used. Finally, based on this knowledge graph, combined with the port operation domain knowledge base (such as "the movement distance of the stacker-reclaimer affects efficiency") and the preset causal template (such as "if X increases, then Y changes"), multiple rounds of reasoning are performed: first, candidate node sets (such as equipment movement distance and stack capacity) are matched and screened, and their efficiency impact weights are calculated. For nodes with high weights, gradient significance test values ​​are calculated by difference. Nodes with test values ​​that meet the standards are screened as key factors. Then, quantitative association rules (such as "total operation time = base time + 0.06 × equipment movement distance") are obtained through linear fitting and other methods.

[0109] In a specific example, the Port A dataset contains fields such as "Stacker-Reclaimer travel distance", "Remaining stack space capacity", "Operating environment temperature", and "Number of equipment failures". Using total operating time as the efficiency indicator, the correlation coefficients between "Stacker-Reclaimer travel distance" and total operating time are calculated to be 0.7, "Remaining stack space capacity" to be 0.65, "Operating environment temperature" to be 0.3, and "Number of equipment failures" to be 0.55. The preset coefficient threshold is 0.6, so the fields "Stacker-Reclaimer travel distance" and "Remaining stack space capacity" are selected. Port A maps the selected "Stacker-Reclaimer Movement Distance" and "Remaining Capacity of Stacker Position" as nodes using spatiotemporal coordinates. Based on historical data, the edge weights of "Stacker-Reclaimer Movement Distance" and "Total Operation Time" are calculated to be 0.75, "Remaining Capacity of Stacker Position" and "Total Operation Time" to be 0.6, and "Stacker-Reclaimer Movement Distance" and "Remaining Capacity of Stacker Position" to be 0.2 (because both affect operation time but the direct correlation is weak). A knowledge graph is formed with the process chain of "Train Unloading → Stacker Movement → Stacking to Stacker Position" as the framework. The candidate node set for Port A consists of "Stacker-Reclaimer travel distance" and "Remaining capacity of stacking positions". After matching the causal template, the efficiency impact weights of the two are calculated to be 0.7 and 0.6 respectively (both > 0.5); the gradient significance test values ​​calculated by difference are 3.2 and 2.8 respectively (both > 1.96), which are identified as key factors; the quantitative association rule is obtained by fitting linear regression: Total operation time (minutes) = 20 + 0.1 × Stacker-Reclaimer travel distance (meters) - 0.001 × Remaining capacity of stacking positions (tons), that is, for every 10 meters increase in travel distance, the operation time increases by 1 minute; for every 1000 tons increase in remaining capacity, the operation time decreases by 1 minute.

[0110] By executing steps a1 to a5, this embodiment of the application can accurately locate key factors that significantly affect the efficiency of bulk coal loading and unloading from the knowledge graph, and establish quantitative correlation rules between key factors and efficiency. Specifically, candidate factors are initially identified through causal template matching, and then further filtered through weight calculation and gradient significance testing to ensure that the selected key factors have actual impact rather than being accidental correlations. Then, the relationship between key factors and efficiency is transformed into calculable quantitative rules through a preset fitting method. This provides a scientific basis for constructing the reward function of the subsequent reinforcement learning model, and also transforms the core logic affecting efficiency from empirical judgment to data-driven precise rules, thereby improving the pertinence and effectiveness of the entire loading and unloading efficiency optimization scheme.

[0111] By executing steps 131-133, this embodiment of the application can initially screen out fields significantly related to loading and unloading efficiency through quantitative analysis, eliminate irrelevant or weakly correlated factors, reduce redundant information in subsequent analysis, and lay the foundation for accurate identification of key factors. The screened fields are transformed into a structured knowledge graph, intuitively presenting the dependencies between each factor and efficiency, and between factors themselves, providing clear structural support for subsequent reasoning about key factors. Through multiple rounds of reasoning and quantitative analysis, key factors and their correlation patterns that significantly affect loading and unloading efficiency are accurately identified, providing quantitative basis for subsequent reinforcement learning model decisions, replacing empirical judgments, and improving the scientific nature of decision-making.

[0112] In some possible embodiments, in step S14, the yard decision topology map is input into the reinforcement learning model, and a coal bulk loading and unloading scheme is determined by performing state-space encoding, action selection, and reward function calculation.

[0113] Step 141: Input the stockpile decision topology map into the reinforcement learning model. Through the spatiotemporal graph convolutional network of the reinforcement learning model, perform multi-level neighborhood aggregation on the node features and edge weights of the stockpile decision topology map to obtain a dynamic state vector that integrates the real-time location of the equipment, the coal type attributes of the stack, the task queue, and the equipment fault status.

[0114] In step 141, the spatiotemporal graph convolutional network is used to process the spatiotemporal features of the topology graph, and multi-level neighborhood aggregation refers to the fusion of the features of a node and its surrounding multi-layer nodes. The dynamic state vector is a vector that fuses the real-time location of the equipment (such as the coordinates of the stacker-reclaimer), the coal type attributes of the stack position (such as coal type and capacity), the task queue (such as information on trains waiting to be unloaded), and the equipment fault status (such as normal / faulty), and is used to reflect the dynamic state of the current operation scenario.

[0115] In this embodiment, a constrained stockpile decision topology graph is input into a reinforcement learning model. Then, through a spatiotemporal graph convolutional network in the model, multi-level neighborhood aggregation is performed on the features of each node in the topology graph (e.g., the coal type of stack 1 is bituminous coal, and the remaining capacity is 17,000 tons) and the edge weights between nodes (e.g., the weight of the "coal type matching" edge is 0.9). This means that not only is the node's own information considered, but also the features of its neighboring nodes (e.g., the association between stacker-reclaimer B and stack 1) and the neighboring nodes of neighboring nodes (e.g., the indirect association between train K and stacker-reclaimer B). Finally, a dynamic state vector is obtained, which comprehensively reflects real-time information such as the current equipment location, stack attributes, task queue, and equipment faults.

[0116] Step 142: Based on the deep dual-Q network, calculate the action value vector corresponding to the dynamic state vector. Select the corresponding action based on the action value vector with the highest value. The actions include: the stacker position coordinate allocation sequence of the stacker-reclaimer, the coal type identification configuration scheme of the conveyor belt, and the equipment movement scheme.

[0117] In step 142, the deep dual-Q network is a reinforcement learning network composed of two independent Q networks (Q1, Q2) used to reduce action value estimation bias; the action value vector is a set of merit scores for each possible action (such as the stacker-reclaimer moving to certain stack positions). The stack position coordinate allocation sequence of the stacker-reclaimer refers to the order in which the stacker-reclaimer should go, satisfying the requirements of not having the stacker-reclaimer operate at the same stack position at the same time and the collision avoidance interval; the coal type identification configuration scheme of the conveyor belt refers to the coal type matching identification (such as "bituminous coal") set for the conveyor belt; the equipment movement scheme refers to the movement path planning of the stacker-reclaimer.

[0118] Specifically, step 142 may include the following process: Step b1, input the dynamic state vector into the deep dual-Q network, and output multiple types of action values, including: the action value of each stack coordinate as the matching position of the stacker-reclaimer, the action value of each coal type identifier being activated and configured on the conveyor belt, and the action value of each path sequence as the equipment movement scheme; Step b2, calculate the task urgency weight based on the train berthing countdown or ship berthing countdown in the planning data, combined with the user level, task level, and yard operation mode in the planning data; Step b3, adjust the action value corresponding to the stack coordinates according to the task urgency weight and the priority corresponding to the yard operation mode; Step b4, select the stack coordinate with the highest action value as the matching position of the stacker-reclaimer, select the coal type identifier with the highest adjusted action value as the configuration instruction of the conveyor belt, and select the path sequence with the highest action value as the equipment movement scheme to obtain the corresponding action.

[0119] In the above process, during unloading operations, the yard operation modes include preparation mode, balancing mode, and high-production mode. High-production mode prioritizes shortening unloading time, while preparation mode prioritizes small-capacity stacks. Furthermore, stack coordinates must meet constraints such as the disabling of empty stacks and the unavailability of stacks corresponding to faulty equipment, and must also meet anti-collision interval requirements. The value of various actions includes three categories: stack coordinate matching value (e.g., the value of stack 1 as a target for the stacker-reclaimer), coal type identification configuration value (e.g., the value of activating the "bituminous coal" identifier on the conveyor belt), and path sequence value (e.g., the value of the straight-line path from stack 2 to stack 1 for the stacker-reclaimer). Task urgency weight is a quantitative value measuring the urgency of a task (range 0-1, higher values ​​indicate greater urgency); train arrival countdown is the proportion of the difference between the estimated arrival time and the current time within a preset window; ship berthing countdown is similar; user level (e.g., VIP / normal), task level (e.g., urgent / normal), and yard operation mode (e.g., preparation / balancing / high-production) need to be quantified into numerical values ​​(e.g., 1-5 points). The priority corresponding to the stockpile operation mode is the decision tendency of different modes (e.g., high-yield mode prioritizes shortening time, balanced mode prioritizes even distribution); the adjusted action value is the action value after being corrected by combining urgency weight and mode priority (e.g., the value of the stacker position for urgent tasks is increased). Matching position is the stacker position that the stacker-reclaimer should go to for operation; configuration instruction is the coal type identifier that the conveyor belt needs to activate; movement scheme is the optimal movement path of the stacker-reclaimer.

[0120] Step b2 may include the following process: Step b21, calculate the first difference between the estimated train berthing time and the current time, or the second difference between the estimated ship berthing time and the current time; Step b22, use the ratio of the first difference to a preset time window as the train berthing countdown, or the ratio of the second difference to a preset time window as the ship berthing countdown; Step b23, quantify the user level, task level, and yard operation mode respectively to obtain the quantified values ​​of the user level, task level, and yard operation mode; Step b24, use the superposition result of the train berthing countdown, user level quantification value, task level quantification value, and yard operation mode quantification value as the urgency weight of the unloading task, and use the superposition result of the ship berthing countdown, user level quantification value, task level quantification value, and yard operation mode quantification value as the urgency weight of the loading task.

[0121] In the above process, the first difference is the difference between the estimated train arrival time and the current time (e.g., 30 minutes); the second difference is the difference between the estimated ship arrival time and the current time. The preset time window is a standard time range set according to the operational scenario (e.g., 60 minutes for train unloading, 120 minutes for ship loading); the train arrival countdown is the ratio of the first difference to the preset window (range 0-1, the smaller the value, the more urgent the situation); the ship arrival countdown is similar. The user level quantification value is the numerical value converted from user level (e.g., VIP, Normal) (e.g., VIP=5, Normal=3); the task level quantification value is the numerical value converted from task level (e.g., Urgent, Regular) (e.g., Urgent=5, Regular=2); the yard operation mode quantification value is the numerical value converted from operation mode (e.g., High-yield, Balanced) (e.g., High-yield=4, Balanced=3). The urgency weight of the unloading task is the quantified value of the urgency of the train unloading task (obtained by the sum of the train berthing countdown, user level quantification, etc.); the urgency weight of the loading task is the quantified value of the urgency of the ship loading task (obtained by the sum of the ship berthing countdown, etc.).

[0122] By executing steps b21 to b24, this embodiment of the application can provide basic data for subsequent measurement of task urgency by calculating time differences. Converting absolute time differences into relative proportions standardizes the measurement scale of urgency, facilitating subsequent weight calculations. Transforming non-numerical levels and patterns into calculable values ​​supports the quantitative calculation of urgency weights. Through multi-dimensional overlay calculations, the urgency of tasks is comprehensively measured, ensuring that highly urgent tasks are given priority.

[0123] By executing steps b1 to b4, the embodiments of this application can output the quantitative value of each action through a deep dual-Q network, providing a clear basis for subsequent action selection, reducing the estimation bias of a single network, and thus adjusting the priority of action values; by adjusting action values, the decision is made more in line with the urgency of the task and the goal of the work mode, improving the scenario adaptability of the solution; by selecting the highest value action, the decision is ensured to be optimal in the current scenario, taking into account efficiency, urgency and work mode.

[0124] Step 143: After the action is performed, determine the observation data including the actual loading and unloading volume, equipment movement distance, and violation event count.

[0125] In step 143, the observed data are the actual results generated after the action is executed, including the actual loading and unloading volume (e.g., 5000 tons of coal actually unloaded), the equipment movement distance (e.g., the stacker-reclaimer moves 50 meters), and the number of violations (e.g., 0 violations). Violations may include violations such as high-temperature operation, loading coal in empty stacks, and mismatch of coal types.

[0126] In this embodiment of the application, after the actions determined in step 142 are performed (such as the stacker-reclaimer B moving to stack position 1 and the conveyor belt being marked with bituminous coal), the actual results are collected through IoT sensors and monitoring systems: the actual tonnage of coal loaded and unloaded is recorded, the actual distance of equipment movement is counted, and any violations (such as coal type mismatch or equipment conflict) are checked and counted, and the observation data is summarized.

[0127] Step 144: Calculate the reward value based on the observation data and the reward function. The reward function is constructed based on key factors and quantitative correlation rules, and includes efficiency reward items, compliance reward items, mobility cost penalty items, and energy-saving reward items.

[0128] In step 144, the reward value is a quantitative score that measures the quality of an action; the efficiency reward is a bonus based on actual loading and unloading efficiency (e.g., reduced time, e.g., 15 points for completing ahead of schedule), focusing on optimizing total operation time; the compliance reward is a bonus for adhering to constraints (e.g., no violations, e.g., 10 points for zero violations), constraints include coal type matching compliance, equipment operation conflict avoidance, and prohibition of high-temperature operations; the movement cost penalty is a deduction for long movement distances (e.g., 5 points for short distances), conversely, minimum stacking priority and short-distance movement rewards; the energy-saving reward is a bonus for energy-saving actions (e.g., short paths, e.g., 5 points for short distances). All scores can be directly summed or weighted summed to obtain the reward value.

[0129] In this embodiment of the application, the reward value is calculated based on the observation data in step 143 and the key factors and quantitative correlation rules determined in S13: the efficiency reward is calculated based on the difference between the actual working time and the standard time (e.g., 15 points are added for working 2 minutes early); the compliance reward is calculated based on the number of violations (e.g., 10 points are added for 0 violations); the movement cost penalty is calculated based on the movement distance (e.g., 5 points are added for 50 meters, and 5 points are deducted for 100 meters); the energy saving reward is calculated by combining the movement distance and energy consumption correlation rules (e.g., 5 points are added for short distances). The sum of the four items is the total reward value.

[0130] Step 145: Based on the reward value, update the action value using temporal difference error. When the fluctuation range of the action value of N consecutive decisions is less than the preset amplitude threshold, the reinforcement learning model is determined to have converged. The output includes the stack position coordinate allocation sequence of the stacker-reclaimer, the coal type identification configuration scheme of the conveyor belt, and the equipment movement scheme of the bulk coal loading and unloading scheme, where N is greater than or equal to 3.

[0131] In step 145, the temporal difference error is the error value used to update the action value (reflecting the difference between the predicted value and the actual reward plus the next state value); the fluctuation amplitude value is the change in action value in continuous decision-making (such as the absolute value of the difference between two decision values); the preset amplitude threshold is the standard for judging model convergence (such as 3%); N is the number of consecutive convergences (≥3).

[0132] In this embodiment, firstly, the action value is updated using the time-series difference error formula; secondly, the fluctuation range of the action value for N consecutive decisions is calculated (e.g., 80 points for the first decision, 82 points for the second decision, with a fluctuation of 2.5%); finally, when the fluctuation range is less than a preset threshold (e.g., 3%) and N ≥ 3 consecutive decisions, the model is determined to have converged, and a loading and unloading scheme including stacker-reclaimer allocation, conveyor belt configuration, and equipment movement is output.

[0133] For example, regarding the unloading task of train K at port A, the yard decision topology graph includes nodes such as "Train K (5000 tons of bituminous coal, 30 minutes countdown to 9:30)," "Stacker-reclaimer B (normal, stack position 2)," and "Stack position 1 (bituminous coal, remaining 17000 tons)," with edge weights such as "Train K - Stack position 1 (coal type matching, 0.9)" and "Stacker-reclaimer B - Stack position 1 (movable, 0.8)." After aggregating this information, the spatiotemporal graph convolutional network outputs a dynamic state vector: "Stacker-reclaimer B is located at stack position 2 (normal), stack position 1 has 17000 tons of bituminous coal remaining, train K (5000 tons of bituminous coal, arriving in 30 minutes) is waiting to be unloaded, and there are no equipment malfunctions." After the dynamic state vector of Port A is input into a deep double-Q network, the output is: matching value of stack position 1 is 80 points, matching value of stack position 4 is 70 points; the configuration value of the "bituminous coal" label on the conveyor belt is 90 points, and the value of the "anthracite" label is 0 points; the value of the straight path of the stacker-reclaimer from stack position 2 to stack position 1 is 85 points, and the value of the detour path is 60 points. Train K is expected to arrive at Port A at 10:00, and the current time is 9:30, so the first difference is 10:00 - 9:30 = 30 minutes; if the loading task of ship B is to be handled, it is expected to berth at 14:00, and the current time is 13:00, so the second difference is 60 minutes. For train K at Port A, the first difference is 30 minutes, the preset window is 60 minutes, and the countdown timer for train berthing is 30 / 60 = 0.5; for ship B, the second difference is 60 minutes, the preset window is 120 minutes, and the countdown timer for ship berthing is 60 / 120 = 0.5. The user of train K at port A is a VIP (quantified value 4), the task level is relatively urgent (quantified value 3), and the current yard operation mode is high production (quantified value 4). Therefore, the user level quantified value = 4, the task level quantified value = 3, and the yard operation mode quantified value = 4. The train arrival countdown for train K at port A is 0.5 (taking 1-0.5=0.5 for calculation), the user level quantified value = 4, the task level quantified value = 3, the mode quantified value = 4, and the urgency weight = 0.25×0.5+0.25×4+0.25×3+0.25×4=0.125+1+0.75+1=2.875. At Port A, train K has an emergency weight of 2.875. In high-production mode, priority is given to shortening the time. The original value of stack position 1 is 80 points (closer distance, shorter time), and the adjusted value is 80 × (1 + 2.875 × 0.1) = 80 × 1.2875 = 103 points; the original value of stack position 4 is 70 points (farther distance), and the adjusted value is 70 × (1 + 2.875 × 0.1) = 89.125 points. After the adjustment in step b3 at Port A, the value of stack position 1 is 103 points (highest), so stacker-reclaimer B is matched with stack position 1; the value of the "bituminous coal" label on the conveyor belt is 90 points (highest), and the configuration instruction is "bituminous coal"; the value of the straight path from stack position 2 to stack position 1 for stacker-reclaimer B is 85 points (highest), and the movement scheme is a straight path.After Port A executed the action, it actually unloaded 5,000 tons of coal (consistent with the plan). The stacker-reclaimer B moved 50 meters. No violations such as coal type mismatch or equipment conflict occurred, and the violation count was 0. The observed data for Port A is (5,000 tons, 50 meters, 0 violations). The actual operation time was 28 minutes (standard 30 minutes). The efficiency bonus was +15 points; the compliance bonus was +10 points for 0 violations; the movement distance of 50 meters (shorter) resulted in a movement cost penalty of +5 points; and the energy saving bonus was +5 points for short-distance energy saving. The total bonus value was 15 + 10 + 5 + 5 = 35 points. Port A updated the action value with the bonus value of 35 points. After 3 iterations, the action value changed from 80 points to 82 points to 83 points, with fluctuation ranges of 2.5% and 1.2% respectively (both <3%). The model converged, and the output solution was: the stacker-reclaimer B moved from stack position 2 to stack position 1 along a straight path, and the conveyor belt was marked with bituminous coal.

[0134] By executing steps 141 to 145, this embodiment of the application can transform complex topological graph information into concise dynamic state vectors through multi-level aggregation of spatiotemporal graph convolutional networks. This not only preserves key scene information but also facilitates rapid processing by subsequent models, providing a foundation for accurate decision-making. It forms specific plans for stacker-reclaimer allocation, conveyor belt configuration, and equipment movement. Observational data reflects the actual effects of action execution, providing real feedback for subsequent reward calculation and model optimization. By quantifying action effects through reward values, the model is guided to prioritize efficient, compliant, and energy-saving actions, improving decision-making quality. By updating action values ​​and determining convergence, the model continuously optimizes decisions, ultimately outputting a stable and optimal loading and unloading plan, ensuring efficient operation.

[0135] Figure 3 A schematic diagram of a machine learning-based coal bulk cargo loading and unloading efficiency optimization system provided in this application embodiment is shown below. Figure 3 The system may include:

[0136] The data acquisition module 31 is used to collect planned data, equipment status and location data, yard resource occupancy data and yard constraint data of bulk coal loading and unloading operations, forming a data set.

[0137] The generation module 32 is used to generate a constrained yard decision topology map based on the data set.

[0138] The identification module 33 is used to identify key factors related to the efficiency of coal bulk cargo loading and unloading using a large language model, as well as the quantitative correlation rules between key factors and the efficiency of coal bulk cargo loading and unloading.

[0139] The reinforcement learning module 34 is used to use a deep double-Q network as the framework of the reinforcement learning model. The decision topology of the yard is input into the reinforcement learning model. By executing the process of state space encoding, action selection and reward function calculation, the reward function aims to maximize the predicted value of coal bulk loading and unloading efficiency in the current period. It is constructed based on key factors and quantitative association rules to determine the coal bulk loading and unloading scheme.

[0140] The calculation and optimization module 35 is used to collect the actual value of the coal bulk cargo loading and unloading efficiency in the current period after the coal bulk cargo loading and unloading plan is completed, calculate the efficiency difference, and optimize the coal bulk cargo loading and unloading efficiency in the next period based on the efficiency difference.

[0141] The machine learning-based coal bulk cargo loading and unloading efficiency optimization system of this application embodiment is used to implement the aforementioned machine learning-based coal bulk cargo loading and unloading efficiency optimization method. Therefore, the specific implementation of the machine learning-based coal bulk cargo loading and unloading efficiency optimization system can be seen in the embodiment section of the machine learning-based coal bulk cargo loading and unloading efficiency optimization method above. The specific implementation can be referred to the description of the corresponding embodiment, and will not be repeated here.

[0142] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described machine learning-based coal bulk cargo loading and unloading efficiency optimization method.

[0143] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described machine learning-based coal bulk cargo loading and unloading efficiency optimization methods.

[0144] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0145] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the machine learning-based coal bulk cargo loading and unloading efficiency optimization method described above.

[0146] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0147] The above provides a detailed description of a machine learning-based method and system for optimizing the loading and unloading efficiency of bulk coal. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A coal bulk handling efficiency optimization method based on machine learning, characterized by, The method comprises the following steps: Collecting plan data, equipment state position data, stockyard resource occupation data and stockyard constraint data of coal bulk cargo handling operation to form a data set, wherein the stockyard constraint data comprises coal type matching requirement, stacker-reclaimer operation conflict constraint, stack capacity upper limit, empty stack position usage limit, equipment failure correlation constraint and high temperature operation limit; Generating a stockyard decision topology graph with constraints according to the data set; Using a large language model to identify key factors related to coal bulk cargo handling efficiency and quantitative association rules between the key factors and the coal bulk cargo handling efficiency; Using a deep double Q network as a framework of a reinforcement learning model, inputting the stockyard decision topology graph into the reinforcement learning model, and determining a coal bulk cargo handling scheme by executing a state space coding, action selection and reward function calculation process, wherein the reward function aims to maximize the predicted value of the coal bulk cargo handling efficiency in the current period, and is constructed based on the key factors and the quantitative association rules; After the coal bulk cargo handling scheme is completed, collecting an actual value of the coal bulk cargo handling efficiency in the current period, calculating an efficiency difference value, and optimizing the coal bulk cargo handling efficiency in the next period based on the efficiency difference value; The method of inputting the stockyard decision topology graph into the reinforcement learning model, and determining the coal bulk cargo handling scheme by executing the state space coding, action selection and reward function calculation process comprises the following steps: Inputting the stockyard decision topology graph into the reinforcement learning model, and performing multi-order neighborhood aggregation on node features and edge weights of the stockyard decision topology graph by a spatio-temporal graph convolution network of the reinforcement learning model to obtain a dynamic state vector which integrates real-time positions of equipment, coal type attributes of stack positions, task queues and equipment failure states; Based on the deep double Q network, calculating an action value vector corresponding to the dynamic state vector, and selecting an action corresponding to a maximum value in the action value vector according to the maximum value, wherein the action comprises a stack position coordinate allocation sequence of a stacker-reclaimer, a coal type identification configuration scheme of a conveyor belt and an equipment movement scheme; After the action is executed, determining observation data comprising actual handling quantity, equipment movement distance and violation event count; According to the observation data and a reward function, calculating a reward value, wherein the reward function is constructed based on the key factors and the quantitative association rules, and comprises an efficiency reward item, a compliance reward item, a movement cost penalty item and an energy saving reward item; Based on the reward value, updating the action value using a time difference error, determining that the reinforcement learning model converges when a fluctuation amplitude value of action values of continuous N times of decision-making is less than a preset amplitude threshold, and outputting a coal bulk cargo handling scheme comprising the stack position coordinate allocation sequence of the stacker-reclaimer, the coal type identification configuration scheme of the conveyor belt and the equipment movement scheme, wherein N is greater than or equal to 3; The method of using a large language model to identify key factors related to coal bulk cargo handling efficiency and quantitative association rules between the key factors and the coal bulk cargo handling efficiency comprises the following steps: Calculating Pearson correlation coefficients of all fields in the data set and handling efficiency indicators, and screening structured fields and unstructured text fields with Pearson correlation coefficients greater than a preset coefficient threshold; Map the structured field and the unstructured text field to knowledge graph nodes according to a space-time coordinate, calculate edge weights between the nodes based on historical job data, and form a knowledge graph taking a loading and unloading process chain as a skeleton and carrying a loading and unloading process dependency relationship and edge weights; the loading and unloading process chain is a complete process including unloading, stacking, taking material and loading; Based on the knowledge graph, combined with a domain knowledge base and a preset causal template, perform multi-round chain reasoning oriented to coal bulk cargo loading and unloading efficiency, and output a key factor list and a quantitative association rule; the key factor list includes multiple key factors.

2. The method of claim 1, wherein, The deep double Q network is used to calculate an action value vector corresponding to the dynamic state vector, and an action corresponding to a maximum value in the action value vector is selected according to the maximum value; the action includes: a stack position coordinate allocation sequence of a stacker-reclaimer, a coal type identifier configuration scheme of a conveyor belt and a device movement scheme, which includes: The dynamic state vector is input into the deep double Q network, and multiple types of action values are output; the multiple types of action values include: an action value of each stack position coordinate as a matching position of the stacker-reclaimer, an action value of each coal type identifier activated and configured on the conveyor belt, and an action value of each path sequence as the device movement scheme; According to a train station arrival countdown or a ship berthing countdown in the plan data, combined with a user level, a task level and a yard operation mode in the plan data, a task emergency weight is calculated; According to the task emergency weight and a priority corresponding to the yard operation mode, an action value corresponding to a stack position coordinate is adjusted; A stack position coordinate with a highest action value is selected as a matching position of the stacker-reclaimer, a coal type identifier with an adjusted highest action value is selected as a configuration instruction of the conveyor belt, and a path sequence with a highest action value is selected as the device movement scheme, so as to obtain a corresponding action.

3. The method of claim 2, wherein, According to the train station arrival countdown or the ship berthing countdown in the plan data, combined with the user level, the task level and the yard operation mode in the plan data, the task emergency weight is calculated, including: A first difference between a train station arrival predicted time and a current time or a second difference between a ship berthing predicted time and the current time is calculated; The first difference is taken as the train station arrival countdown, or the second difference is taken as the ship berthing countdown; The user level, the task level and the yard operation mode are quantified respectively to obtain a user level quantified value, a task level quantified value and a yard operation mode quantified value; A superposition result of the train station arrival countdown, the user level quantified value, the task level quantified value and the yard operation mode quantified value is taken as an emergency weight of an unloading task, and a superposition result of the ship berthing countdown, the user level quantified value, the task level quantified value and the yard operation mode quantified value is taken as an emergency weight of a loading task.

4. The method of claim 1, wherein, Based on the knowledge graph, combined with a domain knowledge base and a preset causal template, perform multi-round chain reasoning oriented to coal bulk cargo loading and unloading efficiency, and output a key factor list and a quantitative association rule, including: The knowledge graph node is matched with a preset causal template for similarity, to screen out a candidate node set affecting coal bulk cargo loading and unloading efficiency; An efficiency influence weight of the candidate node is calculated; For a candidate node with an efficiency influence weight greater than a preset weight threshold, a gradient significance test value corresponding to the candidate node is differentially calculated based on historical operation data; A candidate node with a gradient significance test value greater than a preset test threshold is screened as a key factor, to output a key factor list; An association rule between the key factor and the coal bulk cargo loading and unloading efficiency is quantified through a preset fitting mode, to obtain a quantitative association rule.

5. The method of claim 1, wherein, The efficiency difference value is used to optimize the coal bulk cargo loading and unloading efficiency in the next period, including: When the efficiency difference value is less than or equal to a preset difference value, the weights corresponding to the efficiency reward item, the compliance reward item, the movement cost penalty item and the energy saving reward item in the reward function are adjusted according to the offset direction and size of the efficiency difference value, to optimize the coal bulk cargo loading and unloading efficiency in the next period according to the adjusted reward function; Or, when the efficiency difference value is greater than the preset difference value, the large language model is adjusted according to the prediction bias sample set, to optimize the coal bulk cargo loading and unloading efficiency in the next period according to the adjusted large language model.

6. A coal bulk handling efficiency optimization system based on machine learning, characterized by, It includes: A collection module is configured to collect plan data, equipment state position data, stockyard resource occupation data and stockyard constraint data of coal bulk cargo loading and unloading operation to form a data set; A generation module is configured to generate a stockyard decision topology graph with constraints based on the data set; An identification module is configured to identify key factors related to coal bulk cargo loading and unloading efficiency and quantitative association rules between the key factors and the coal bulk cargo loading and unloading efficiency by using a large language model; A reinforcement learning module is configured to use a deep double Q network as a framework of a reinforcement learning model, input the stockyard decision topology graph into the reinforcement learning model, and determine a coal bulk cargo loading and unloading scheme by executing a process of state space coding, action selection and reward function calculation, wherein the reward function aims to maximize a predicted value of the coal bulk cargo loading and unloading efficiency in a current period and is constructed based on the key factors and the quantitative association rules; A calculation and optimization module is configured to collect an actual value of the coal bulk cargo loading and unloading efficiency in the current period after the coal bulk cargo loading and unloading scheme is completed, calculate an efficiency difference value, and optimize the coal bulk cargo loading and unloading efficiency in the next period based on the efficiency difference value; The stockyard decision topology graph is input into the reinforcement learning model, and a dynamic state vector fused with real-time equipment positions, pile position coal type attributes, task queues and equipment fault states is obtained by performing multi-order neighborhood aggregation on node features and edge weights of the stockyard decision topology graph through a spatio-temporal graph convolution network of the reinforcement learning model, to determine the coal bulk cargo loading and unloading scheme. ​ The action value vector corresponding to the dynamic state vector is calculated based on a deep double Q network, and an action corresponding to a maximum value in the action value vector is selected according to the maximum value, the action including a stack position coordinate allocation sequence of the stacker-reclaimer, a coal type identification configuration scheme of the conveyor belt, and a device movement scheme; After the action is performed, observation data including an actual loading and unloading amount, a device movement distance, and a violation event count are determined; A reward value is calculated according to the observation data and a reward function, the reward function being constructed according to the key factors and the quantitative correlation rules and including an efficiency reward item, a compliance reward item, a movement cost penalty item, and an energy saving reward item; Based on the reward value, the action value is updated using a time difference error, and when a fluctuation amplitude value of the action values of N consecutive decisions is less than a preset amplitude threshold, the reinforcement learning model is determined to be converged, and a coal bulk cargo loading and unloading scheme including the stack position coordinate allocation sequence of the stacker-reclaimer, the coal type identification configuration scheme of the conveyor belt, and the device movement scheme is output, N being greater than or equal to 3; Key factors related to the coal bulk cargo loading and unloading efficiency and quantitative correlation rules between the key factors and the coal bulk cargo loading and unloading efficiency are identified using a large language model, including: Pearson correlation coefficients of all fields in the data set and the loading and unloading efficiency indicators are calculated, and structured fields and unstructured text fields with Pearson correlation coefficients greater than a preset coefficient threshold are screened; The structured fields and the unstructured text fields are mapped into knowledge graph nodes according to a space-time coordinate, edge weights between the nodes are calculated based on historical operation data, a knowledge graph carrying a loading and unloading process chain as a skeleton and a loading and unloading process dependency relationship and edge weights is formed, and the loading and unloading process chain is a complete process including unloading, stacking, reclamation, and loading; Based on the knowledge graph, a multi-round chain reasoning oriented to the coal bulk cargo loading and unloading efficiency is performed in combination with a domain knowledge base and a preset causal template, and a key factor list and quantitative correlation rules are output, the key factor list including multiple key factors.

7. An electronic device, comprising: including: a memory for storing a computer program; a processor for implementing the steps of the coal bulk cargo loading and unloading efficiency optimization method based on machine learning according to any one of claims 1 to 5 when the computer program is executed.

8. A computer-readable storage medium, characterized in that, The computer program is stored in the computer readable storage medium, and the computer program can implement the coal bulk cargo loading and unloading efficiency optimization method based on machine learning according to any one of claims 1 to 5 when executed by the processor.

Images