An industry chain scheduling decision system based on data and mechanism fusion

Abstract

The application relates to the technical field of industrial chain scheduling decision, and discloses an industrial chain scheduling decision system based on data and mechanism fusion, wherein a data fusion and structuring module receives multi-element heterogeneous real-time data of an industrial chain, and generates a global state at a current moment; a hierarchical decision module generates a joint action at the current moment based on the global state at the current moment through a hierarchical architecture, and feeds back to a digital twin environment construction module; the digital twin environment construction module constructs a digital twin environment, carries out calculation based on the global state at the current moment and the joint action at the current moment in a mode of fusion of a mechanism model and data correction, obtains a global state at a next moment, and sends the global state to the hierarchical decision module. Through the cooperative action of the modules, the application realizes full-link automatic scheduling decision, provides a continuously optimized decision instruction for the industrial chain, and improves the full-chain collaborative operation efficiency and dynamic environment adaptability.

Description

Technical Field

[0001] This invention relates to the field of supply chain scheduling and decision-making technology, and specifically to a supply chain scheduling and decision-making system based on the fusion of data and mechanisms. Background Technology

[0002] In large-scale industrial systems, especially in supply chain systems involving multiple links such as production, transportation, warehousing and sales, such as coal, oil and containers, achieving coordinated and optimized scheduling across the entire chain is of great significance for reducing operating costs, ensuring stable supply of materials and improving overall operational efficiency.

[0003] Currently, scheduling decisions for such complex supply chains primarily rely on human experience or traditional operations research optimization methods. Scheduling based on human experience is highly dependent on the individual skills and familiarity with the business by the scheduling personnel. As the scale of the supply chain expands and the market environment dynamically changes, human decision-making has limitations in both the breadth and depth of information processing. It struggles to make comprehensive, real-time judgments on the overall system status, leading to delayed decisions or optimizations only achieved in localized stages. This results in low efficiency in cross-departmental and cross-stage collaboration, causing problems such as resource waiting, capacity mismatch, or inventory backlog.

[0004] Mathematical programming methods in operations research (such as linear programming and mixed-integer programming) seek theoretically optimal solutions by establishing a mathematical model of the entire system under given constraints. However, the effectiveness of these methods highly depends on an accurate and comprehensive system model. For dynamic and complex supply chains, building and maintaining such a model is both challenging and time-consuming. Once established, the model's structure is relatively fixed, making it difficult to flexibly adapt to dynamic events in the physical world. Furthermore, as the number of nodes and decision variables in the supply chain increases, the solution dimension of the optimization model expands dramatically, leading to an exponential increase in computational complexity and excessively long solution times, making it difficult to meet the timeliness requirements of many scenarios. Summary of the Invention

[0005] This invention provides a supply chain scheduling decision system based on the fusion of data and mechanisms to solve one of the technical problems existing in the prior art.

[0006] This invention provides a supply chain scheduling decision-making system based on the fusion of data and mechanisms. The system includes a data fusion and structuring module, a hierarchical decision-making module, and a digital twin environment construction module. The data fusion and structuring module is used to receive diverse and heterogeneous real-time data from the industry chain and generate the global state at the current moment. The hierarchical decision-making module is used to generate joint actions for the current moment based on the global state at the current moment through a hierarchical architecture, and then feed them back to the digital twin environment construction module. The digital twin environment building module is used to build a digital twin environment that operates synchronously with the industry chain; The digital twin environment construction module is also used to calculate the global state at the next moment in the digital twin environment by combining the mechanism model with data correction, based on the global state at the current moment and the joint actions at the current moment, and then send it to the hierarchical decision module.

[0007] This invention unifies diverse and heterogeneous real-time data into a global state through data fusion and structuring modules, providing a complete and consistent data foundation for decision-making and overcoming the limitations of the breadth and depth of information processing in manual decision-making. By using a hierarchical decision-making module to generate joint actions based on a hierarchical architecture and feeding them back to the digital twin environment, it avoids the complexity explosion problem caused by solving all variables with a single model, thus improving the real-time performance of decision-making. Through a deductive approach that integrates mechanistic models and data correction within the digital twin environment, it uses data-driven correction of model biases while ensuring physical constraints, improving the accuracy of state prediction and enabling dynamic adaptation to sudden operating conditions such as equipment failures and weather changes, overcoming the shortcomings of traditional operations research models that have rigid structures and are difficult to dynamically adjust. The synergistic effect of these modules achieves fully automated scheduling and decision-making across the entire chain, from data acquisition, state fusion, hierarchical decision-making to simulation deduction and closed-loop feedback, providing continuously optimized decision instructions for the industry chain and improving the efficiency of collaborative operation and adaptability to dynamic environments.

[0008] In one alternative implementation, the data fusion and structuring module is specifically used for: Receive diverse and heterogeneous real-time data from the industry chain and preprocess it to obtain preprocessed data; The preprocessed data is divided into node state sub-vectors, transportation unit state sub-vectors, key equipment state sub-vectors, and external environment state sub-vectors, which constitute the global state at the current moment.

[0009] This invention improves the efficiency of subsequent processing by preprocessing multi-dimensional heterogeneous real-time data through data fusion and structuring modules. The preprocessed data is divided into multiple sub-vectors and combined into a global state, which structurally contains all the necessary information representing the complete operation of the industrial chain at the current moment, providing a data foundation for subsequent decision-making.

[0010] In one optional implementation, the system further includes a scheduling execution module; The hierarchical decision-making module is also used to feed back the joint actions at the current moment to the scheduling and execution module; The scheduling and execution module is used to schedule and control the supply chain based on the joint actions at the current moment, and to collect the actual evolution status of the supply chain after executing the joint actions at the current moment.

[0011] This invention provides feedback data from the real industrial chain to the digital twin environment through a scheduling and execution module, enabling the system to continuously optimize decisions based on the difference between reality and prediction, forming a complete closed loop of decision-making-execution-feedback.

[0012] In one alternative implementation, the hierarchical decision-making module includes a strategic agent and multiple tactical agents; The strategic agent is used to aggregate the current global state into a strategic state and input it into the target strategic policy network to obtain a strategic intent vector, which includes the decision weight of each macro-objective. Strategic agents are also used to broadcast strategic intent vectors to each tactical agent; Each tactical agent is used to extract the local state of its own business domain from the global state at the current moment, and input the local state and strategic intent vector into the target tactical policy network to obtain the coordinated fine-tuning action vector. The hierarchical decision-making module is also used to concatenate all collaborative fine-tuning action vectors and perform legality verification to form the joint action at the current moment.

[0013] This invention constructs a hierarchical architecture that coordinates strategic and tactical intelligent agents, decomposing the complex end-to-end optimization problem into two levels: macro-level intent formulation and micro-level action execution. This achieves quantitative alignment and collaborative optimization of macro-level strategy and micro-level tactics.

[0014] In one alternative implementation, the digital twin environment construction module is specifically used for: In a digital twin environment, for each collaborative fine-tuning action vector in the joint action at the current moment, the mechanism model corresponding to each collaborative fine-tuning action vector is invoked; The mechanism model corresponding to each collaborative fine-tuning action vector is adopted. Based on the global state at the current moment and the corresponding collaborative fine-tuning action vector, the calculation is performed to obtain the local state at the next moment corresponding to each collaborative fine-tuning action vector. All the local states at the next moment are used to form the theoretical prediction state vector. A deep learning model is used to generate a residual correction state vector based on the current global state and the current joint action. The deep learning model is trained based on the difference between the theoretically predicted state vector and the actual evolved state. The fusion weights are dynamically determined based on the residual correction state vector. The global state at the next time step is calculated based on the theoretically predicted state vector, the residual corrected state vector, and the fusion weights.

[0015] This invention integrates mechanism and data, which can satisfy the underlying objective mechanism constraints and accurately depict the actual working conditions under the influence of complex environment, thus providing an accurate and reliable data input source for subsequent multi-objective optimization and macro-benefit evaluation.

[0016] In one optional implementation, the system further includes an offline training module; the offline training module is used for: Set a preset strategic intent vector; The hierarchical decision-making module is invoked, and an initial tactical strategy network is adopted. Based on the preset strategic intent vector and the simulation local state of each tactical agent at the current simulation time step, a simulation collaborative fine-tuning action vector for each tactical agent is generated. All simulation collaborative fine-tuning action vectors are spliced ​​together and their legality is verified to form a simulation joint action. The digital twin environment construction module is invoked. In the digital twin environment, the global state of the simulation for the next simulation time step is generated based on the global state of the simulation and the joint simulation actions of the current simulation time step. Calculate the tactical reward for each tactical agent; Store the first interaction data of the current simulation time step into the experience replay pool. The first interaction data includes the simulation local state of the current simulation time step, the preset strategic intent vector, the simulation collaborative fine-tuning action vector, the tactical reward, and the simulation local state of the next simulation time step. Sample the first interaction data from the experience replay pool, update the initial tactical strategy network, and complete the current simulation time step; Returning to the call to the hierarchical decision-making module, the initial tactical strategy network is adopted. Based on the preset strategic intent vector and the simulation local state of each tactical agent at the current simulation time step, the simulation collaborative fine-tuning action vector of each tactical agent is generated. This process continues until the initial tactical strategy networks of all tactical agents converge, resulting in the target tactical strategy network of each tactical agent.

[0017] This invention, by fixing the high-level strategic intent, iteratively trains the underlying tactical agents in a digital twin environment, enabling each agent to independently learn the optimal local decision-making strategy under the guidance of clear macro-level preferences. Through repeated closed-loop iterations of action generation, state deduction, reward calculation, and experience sampling, the initial tactical policy network is continuously updated until convergence, allowing the tactical agents to learn through repeated trial and error, mastering how to make optimal micro-level decisions under given macro-level preferences.

[0018] In one alternative implementation, the output layer of the initial tactical policy network employs a multi-head architecture.

[0019] This invention, through a multi-head architecture design, enables the tactical strategy network to simultaneously process continuous and discrete action commands, fully covering the diverse operational needs in supply chain scheduling decisions, and improving the expressive power and decision-making accuracy of the strategy.

[0020] In one alternative implementation, the tactical reward is calculated using the following formula:

[0021] In the formula, Indexes representing tactical agents; Indicates the index of the simulation time step; Indicates the first The tactical agent in the 1st... Tactical rewards for each simulated time step; Indicates the balance coefficient; Indicates the first The tactical agent in the 1st... The intrinsic reward for each simulation time step; Indicates the first The tactical agent in the 1st... The first strategic reward for each simulation time step is calculated using the following formula:

[0022] In the formula, Represents the simulated strategic intent vector; Represents the contribution vector; An index representing macroeconomic objectives; Indicates the quantity of macroeconomic targets; Indicates the first The weights of macro-level objectives in the simulated strategic intent vector; The tactical agent is represented in the first... The simulation time step for the first The degree of contribution to the macro-level objectives.

[0023] This invention, by calculating tactical rewards, can provide each tactical agent with immediate and quantitative feedback signals, enabling them to clearly understand the merits and demerits of their current actions in terms of both local operational efficiency and macro-strategic alignment. This guides tactical agents to optimize local operational indicators within their own business domains, and transmits high-level decision preferences to the bottom layer through the inner product of strategic intent vector and contribution vector. This allows tactical agents to proactively serve global macro-goals while pursuing local efficiency, ensuring consistency between bottom-level decisions and high-level intentions.

[0024] In one alternative implementation, the offline training module is further configured to: A fixed-target tactical policy network for each tactical agent; In a digital twin environment, a hierarchical decision-making module is invoked, an initial strategic policy network is adopted, a simulation strategic intent vector is generated based on the simulation global state at the current simulation time step, and the simulation strategic intent vector is broadcast to each tactical agent. Each tactical agent is invoked to adopt a fixed target tactical strategy network. Based on the simulated strategic intent vector and the simulated local state at the current simulation time step, a simulated collaborative fine-tuning action vector is generated. All simulated collaborative fine-tuning action vectors are spliced ​​together and their legality is verified to form a simulated joint action. The digital twin environment construction module is invoked. In the digital twin environment, the global state of the simulation for the next simulation time step is generated based on the global state of the simulation and the joint simulation actions of the current simulation time step. The second strategic reward for computational strategic agents; Store the second interaction data of the current simulation time step into the experience replay pool. The second interaction data includes the simulation strategy state, simulation strategy intent vector, second strategy reward, and simulation strategy state of the next simulation time step. Sample the second interaction data from the experience replay pool, update the initial strategy network, and complete the current simulation time step; Returning to the digital twin environment, the hierarchical decision-making module is invoked, and the initial strategic policy network is used to generate the simulation strategic intent vector based on the simulation global state at the current simulation time step. This process continues until the initial strategic policy network converges, resulting in the target strategic policy network.

[0025] This invention trains a strategic agent to learn how to generate optimal strategic intent vectors based on different strategic states, building upon the already well-executed underlying strategies, in order to achieve a dynamic trade-off between macro-level objectives and maximize long-term gains.

[0026] In one alternative implementation, the second strategic reward is calculated using the following formula:

[0027] In the formula, This indicates the second strategic reward; Represents the macroeconomic return vector; Indicates the first The first macroeconomic goal in The achieved value of each simulation time step.

[0028] This invention, by calculating a second strategic reward, helps train a strategic agent to generate a strategic intent vector that is most beneficial to long-term macroeconomic gains based on different strategic states. Attached Figure Description

[0029] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of a supply chain scheduling decision-making system based on data and mechanism fusion according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the data fusion and structuring module according to an embodiment of the present invention; Figure 3 This is an execution flowchart of the hierarchical decision-making module and the digital twin environment construction module according to an embodiment of the present invention; Figure 4 This is an execution flowchart of the offline training module according to an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0033] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0034] This embodiment provides a supply chain scheduling decision-making system based on the fusion of data and mechanisms. Figure 1 This is a schematic diagram of a supply chain scheduling decision-making system based on data and mechanism fusion according to an embodiment of the present invention, such as... Figure 1As shown, the system includes a data fusion and structuring module, a hierarchical decision-making module, and a digital twin environment construction module. The data fusion and structuring module receives diverse and heterogeneous real-time data from the industry chain and generates the current global state. The hierarchical decision-making module generates joint actions based on the current global state through a hierarchical architecture and feeds them back to the digital twin environment construction module. The digital twin environment construction module constructs a digital twin environment that operates synchronously with the industry chain. The digital twin environment construction module is also used to perform calculations in the digital twin environment based on the current global state and the current joint actions through a fusion of mechanism models and data correction to obtain the global state at the next moment and send it to the hierarchical decision-making module.

[0035] Specifically, the data fusion and structuring module receives diverse and heterogeneous real-time data from the industry chain, including static business data, dynamic business data, and external environmental data. Static business data describes the infrequently changing physical attributes and constraints within the industry chain, including the rated capacity of each production unit, the rated capacity of each warehousing unit, the topology and rated capacity of each transportation route, and the number and technical parameters of each port berth. Dynamic business data reflects the operational status of each link in the industry chain in real-time or near real-time, including the current inventory of each production and warehousing unit, the real-time geographical location and current load of each transportation unit (e.g., trains, ships), and the operational status of key equipment (e.g., in operation, standby, faulty). External environmental data originates from outside the industry chain system but affects its operation, including Geographic Information System (GIS) data, meteorological data for specific regions (e.g., wind speed, rainfall), and macroeconomic indices. After receiving the above data, the data fusion and structuring module processes and outputs a structured global state at the current moment; that is, the diverse and heterogeneous real-time data is systematically organized to form a global state for simulation and deduction.

[0036] The hierarchical decision-making module, through its hierarchical architecture, generates joint actions for the current moment based on the global state at the current moment and feeds them back to the digital twin environment construction module. By decomposing high-dimensional decision problems through a hierarchical architecture, it avoids the complexity explosion problem caused by solving all variables with a single model, thereby improving the real-time performance of decision-making.

[0037] The digital twin environment construction module builds a high-fidelity digital twin environment that operates synchronously with the physical supply chain. As a computable simulator, this environment starts with the current global state and is driven by the current joint actions. Through a fusion of mechanistic models and data correction, it performs state transition simulations to calculate the next global state, thus completing a full decision-making cycle. This fusion of mechanistic and data approaches, while ensuring adherence to physical constraints, enables the system to flexibly adapt to dynamic events such as equipment failures and weather changes, overcoming the limitations of traditional operations research models that are structurally rigid and difficult to dynamically adjust.

[0038] The digital twin environment construction module sends the global state of the next moment to the hierarchical decision-making module to trigger the next decision cycle. At this time, the hierarchical decision-making module uses the global state of the next moment as the new global state of the current moment, and executes the decision and action generation process again, forming a continuous closed-loop iteration.

[0039] Through the above mechanism, the system realizes the fully automated operation of the entire chain from data acquisition, state fusion, hierarchical decision-making to simulation and closed-loop feedback, and achieves continuous optimization and closed-loop control of the industrial chain scheduling decision.

[0040] In some alternative implementations, Figure 2 This is a schematic diagram of the data fusion and structuring module according to an embodiment of the present invention, as shown below. Figure 2 As shown, the data fusion and structuring module is specifically used to: receive multi-dimensional heterogeneous real-time data from the industrial chain and preprocess it to obtain preprocessed data; divide the preprocessed data into node state sub-vectors, transportation unit state sub-vectors, key equipment state sub-vectors, and external environment state sub-vectors to form the global state at the current moment.

[0041] Specifically, the data fusion and structuring module receives diverse and heterogeneous real-time data from the industry chain, preprocesses it (e.g., data cleaning, spatiotemporal alignment, and feature fusion), and standardizes and normalizes it to eliminate the influence of dimensions, resulting in preprocessed data that improves subsequent processing efficiency. Based on the physical meaning of the data and the operational logic of the industry chain, this preprocessed data is divided into four sub-parts: node state sub-vectors, transportation unit state sub-vectors, key equipment state sub-vectors, and external environment state sub-vectors, which together constitute the global state at the current moment, which can be represented by the following formula:

[0042] In the formula, express The global state at any given moment; This represents a sub-vector of node states, containing all production, transit, and warehousing nodes. Material inventory levels at any given time, for example ,in Indicates the first Each node Material inventory levels at any given time; This represents the status sub-unit of a transportation unit, containing all transportation units (such as trains and ships) in... Information such as location, speed, and load at any given moment; This represents the state sub-vector of key equipment, which includes the state of key equipment (such as port machinery and belt conveyors) in each stage. Information such as real-time operating status and operation rate; This represents a sub-vector representing the external environment state, containing information about the environment state. Quantitative external factors that constantly affect the operation of the system, such as market price indices and meteorological data for specific regions.

[0043] The four sub-vectors mentioned above together constitute the global state at the current moment. As a high-dimensional numerical vector, it structurally contains all the necessary information to represent the complete operation of the industrial chain at the current moment, providing a data foundation for subsequent decision-making.

[0044] In some optional implementations, the system also includes a scheduling execution module; a hierarchical decision module, which is also used to feed back the joint actions at the current moment to the scheduling execution module; and a scheduling execution module, which is used to schedule and control the industrial chain according to the joint actions at the current moment, and to collect the actual evolution status of the industrial chain after executing the joint actions at the current moment.

[0045] Specifically, the hierarchical decision-making module feeds back the current joint actions to the scheduling and execution module, enabling the scheduling and execution module to control the supply chain based on these actions. The scheduling and execution module, acting as a bridge between decision-making and physical execution, is responsible for parsing the joint actions into specific equipment control commands and distributing them to the execution units in each business area of ​​the supply chain. While executing the aforementioned scheduling and control, the scheduling and execution module collects real-time feedback data on the actual evolution of the supply chain after executing the current joint actions, including changes in actual inventory at each node, the actual location and load of transportation units, and the actual operating efficiency of key equipment. By providing feedback data from the real supply chain to the digital twin environment through the scheduling and execution module, the system can continuously optimize decisions based on the discrepancy between actual and predicted data, forming a complete decision-making-execution-feedback closed loop.

[0046] In some alternative implementations, Figure 3 This is an execution flowchart of the hierarchical decision-making module and the digital twin environment construction module according to an embodiment of the present invention, such as... Figure 3As shown, the hierarchical decision-making module includes a strategic agent and multiple tactical agents. The strategic agent aggregates the current global state into a strategic state and inputs it into the target strategic policy network to obtain a strategic intent vector, which includes the decision weight of each macro-objective. The strategic agent also broadcasts the strategic intent vector to each tactical agent. Each tactical agent extracts the local state of its own business domain from the current global state and inputs the local state and the strategic intent vector into the target tactical policy network to obtain a collaborative fine-tuning action vector. The hierarchical decision-making module also concatenates all collaborative fine-tuning action vectors and performs legality verification to form the joint action at the current moment.

[0047] Specifically, the hierarchical decision-making module adopts a hierarchical architecture, consisting of a high-level strategic agent and multiple low-level tactical agents. The strategic agent is responsible for macro-level decision-making, while the tactical agents are responsible for micro-level coordinated execution. By decomposing the complex end-to-end optimization problem into two levels—macro-level intent formulation and micro-level action execution—the solution complexity of a single decision-making entity is reduced.

[0048] The high-level strategic agent first uses a state aggregation function to aggregate multiple discrete data points in the current global state into a few key performance indicators, forming a low-dimensional strategic state. This reduces the complexity of strategic decision-making. The calculation process can be represented as follows:

[0049] In the formula, Indicates a strategic state; Represents a state aggregation function; Depend on The sum of the material inventory of all nodes in the global state at any given time is obtained, which represents the aggregated inventory level of the entire industry chain; This represents the project fulfillment rate calculated based on planned data and actual execution data. The actual cumulative delivery volume at any given time is used as the actual execution data (this actual execution data is derived from the digital twin environment from the initial time to the present). The total amount of material inventory actually received by the final delivery node during this period (at this moment) will The phased pre-determined demand at any given time is used as the planned data (this planned data serves as the system's preset input condition, directly obtained from the production schedule issued by the upper-level business system or the pre-determined target quantity of customer orders). The actual execution data is divided by the planned data as the... The project's on-time performance rate; This refers to the safety stock indicator, which is determined by pre-setting safety stock lower limits for each key production, transit, and warehousing node in the supply chain to mitigate the risk of supply disruptions. The safety stock level is obtained in real-time from the current global status. The material inventory of each node at any given time is counted. The number of nodes whose material inventory is not less than the corresponding safety stock threshold is counted. This number of nodes is divided by the total number of nodes in the entire industry chain to obtain the safety stock index, which intuitively quantifies the ability of the entire chain to resist the risk of supply disruption at the current moment.

[0050] Then, the strategic agent inputs the strategic state into a pre-trained target strategy network to generate a strategic intent vector for the current moment. This vector contains dynamic decision weights for multiple macro-level objectives within the current decision-making cycle, which guide the underlying execution strategy at the current moment to maximize the achievement of the overall macro-level objectives of the industry chain in the next moment.

[0051] In the formula, express The strategic intent vector at any given moment; This indicates the quantity of macroeconomic targets, which include at least two or more of the following: supply guarantee targets, efficiency targets, safety stock targets, plan fulfillment targets, and timeliness targets. This represents the weight assigned to the first macro objective. The sum of the weights of all macro objectives is 1. To ensure that the sum of the weights is 1, a Softmax activation function is set at the end of the objective strategy network.

[0052] The high-level strategic agent dynamically assigns weights to various macro-level objectives through an adaptive weight adjustment mechanism. Specifically, the key performance indicators (KPIs) in the strategic state reflect the "current state" of the supply chain, while the macro-level objectives in the strategic intent vector represent the "direction" that the high-level management hopes to optimize. The core role of the strategic agent is to establish a mapping from the "current state" to the "direction": dynamically adjusting the decision weights of each macro-level objective based on the real-time performance of the KPIs. Since the strategic state directly reflects the real-time performance of the supply chain's KPIs, the target strategic strategy network is trained to autonomously identify the urgency of the KPIs. For example, if a safety stock indicator alarm is detected in the strategic state, indicating a risk of supply disruption in the supply chain, the target strategic strategy network will output a weight distribution tilted towards the supply guarantee objective, causing a sharp increase in the weight of the corresponding supply guarantee objective; if the inventory in each link is stable but the cost is too high, the efficiency objective will be adaptively assigned a higher weight. Through this mechanism, the weights of macro-level objectives can be flexibly switched according to the dynamic operating conditions of the supply chain, realizing the dynamic generation of high-level decision-making intentions.

[0053] Each tactical agent at the bottom layer is configured to be responsible for a specific business domain, which includes at least two or more of the following: coal source scheduling, railway transportation, port operations, and shipping scheduling. For example, in a coal production, transportation, and sales industry chain, tactical agents for coal source scheduling, railway transportation, port operations, and shipping scheduling can be set up. Each tactical agent extracts local states directly related to its business domain from the current global state through a state extraction function. Then, the tactical agent calls its pre-trained target tactical strategy network, taking its local state and strategic intent vector as input, and outputs a collaborative fine-tuning action vector, which is a vector containing specific operational instructions for the corresponding business domain. For example, for a railway transportation tactical agent, its collaborative fine-tuning action vector may include the following specific instructions: departure instructions (e.g., whether to depart, departure time), route selection instructions (e.g., which railway line to select), and marshalling instructions (e.g., the number of wagons to be attached).

[0054] Finally, the hierarchical decision-making module directly concatenates the collaborative fine-tuning action vectors generated by each tactical agent along the feature dimension to form an initial joint action at the global dimension. Subsequently, the built-in rule engine or constraint matrix performs a legality check on this initial joint action to identify and handle logical conflicts between actions in different business domains and operations that exceed physical and security limitations (e.g., a railway transport instruction setting an instantaneous shipment volume exceeding the current available storage capacity limit of the target port, or multiple adjacent agents simultaneously issuing mutually exclusive scheduling instructions to the same critical lifting equipment). When a conflict or violation is detected, the hierarchical decision-making module adaptively downgrades, corrects, or eliminates conflicting action instructions based on preset safety fallback strategies and business action priority rules, thereby obtaining a legal action that fully satisfies both the physical and business constraints of the industry chain—that is, the joint action at the current moment. Through this concatenation and legality check combination mechanism, the global execution conflicts that may arise from the excessive pursuit of local optima by multiple agents in distributed decision-making are effectively mitigated, fundamentally ensuring the physical security and high executability of the instructions output by the scheduling decision system.

[0055] By using strategic intent vectors as input, each tactical agent can perceive the macro-level preferences of higher levels when making decisions, thereby aligning its behavior with global macro-level goals while pursuing its own local efficiency.

[0056] It should be noted that the data flow within the hierarchical decision-making module is designed as a broadcast mechanism, meaning that the strategic intent vector generated by the high-level strategic agents is transmitted indiscriminately to all low-level tactical agents. This mechanism ensures that, at the same decision-making moment, although each tactical agent faces different local states, the macro-objective weights upon which they base their decisions are exactly the same. This guarantees that the cluster of low-level tactical agents works collaboratively under unified macro-guidance, avoiding global conflicts of interest caused by each agent acting independently.

[0057] Optionally, for the coal production, transportation, and sales industry chain scenario, it physically consists of multiple geographically dispersed coal sources, a railway transportation network connecting the coal sources to a core seaport, a seaport with multiple berths and storage yards, a fleet of ships for maritime transportation, and multiple downstream power plants as the final consumers. The core optimization problem in this scenario is: under the premise of meeting the coal demand contracts of downstream power plants, to coordinate and schedule the entire business flow from coal production, railway collection, port transshipment to maritime transportation to achieve one or more preset macro-operational objectives. To adapt to this scenario, the tactical agents are specifically configured as follows: a coal source scheduling tactical agent responsible for deciding the shipment volume of each coal source; a railway transportation tactical agent responsible for deciding the daily train dispatch plan; a port operation tactical agent responsible for deciding the berth allocation and loading / unloading operation sequence of ships; and a shipping scheduling tactical agent responsible for deciding the navigation and berthing plans of ships. In this scenario, the macro-objectives managed by the strategic agents are set as: supply guarantee fulfillment rate target, comprehensive operating cost target, and full-chain inventory health target. Therefore, the generated strategic intent vector has the following structure:

[0058] In the formula, This indicates the weight allocated to the supply guarantee and contract fulfillment targets; This indicates the weight assigned to the overall operating cost target; This represents the weights assigned to the inventory health target across the entire supply chain, and the sum of the three weights is 1.

[0059] In some alternative implementations, such as Figure 3As shown, the digital twin environment construction module is specifically used for: in the digital twin environment, for each collaborative fine-tuning action vector in the joint action at the current moment, calling the mechanism model corresponding to each collaborative fine-tuning action vector; using the mechanism model corresponding to each collaborative fine-tuning action vector, calculating based on the global state at the current moment and the corresponding collaborative fine-tuning action vector, obtaining the local state at the next moment corresponding to each collaborative fine-tuning action vector, and constructing a theoretical prediction state vector from all the local states at the next moment; using a deep learning model, generating a residual correction state vector based on the global state at the current moment and the joint action at the current moment, the deep learning model is trained based on the difference between the theoretical prediction state vector and the actual evolved state; dynamically determining the fusion weights based on the residual correction state vector; and calculating the global state at the next moment based on the theoretical prediction state vector, the residual correction state vector, and the fusion weights.

[0060] Specifically, the digital twin environment construction module integrates a mechanism model, which is a set of mathematical functions used to describe the physical processes and business rules in the industry chain. It consists of mechanism models from multiple different business areas, and each mechanism model is responsible for simulating the state transition logic of its corresponding link. Taking the coal production, transportation and sales industry chain as an example, it specifically includes mechanism models for coal source scheduling, railway transportation, port operations and shipping scheduling. The coal source scheduling mechanism model is used to calculate the actual output of each coal mine or the actual shipment of each transport station within a decision cycle. Its inputs are the current inventory and capacity constraints from the global state, as well as the production or shipment instructions from joint actions. The output is the updated coal source inventory. The railway transportation mechanism model is used to update the status of all trains in the railway network. Its inputs are the current position, speed, and cargo status of the trains from the global state, as well as the scheduling instructions from joint actions (e.g., departure, berthing, and marshalling instructions). The output is the expected position and status of the trains at the next moment. The port operation mechanism model is used to update the real-time inventory and berth status of the port yard. Its inputs are the current port inventory and real-time berth availability from the global state, as well as the loading, unloading, and berthing instructions from joint actions. The output is the updated absolute inventory of the port yard and the berthing status of the ships at the next moment. The shipping scheduling mechanism model is used to estimate the position, speed, and cargo status of ships at the next moment based on their current status and navigation instructions, combined with marine meteorological data.

[0061] The digital twin environment construction module simulates the state changes of the industry chain, concurrently calls all relevant mechanism models, and summarizes the local states output by each mechanism model to achieve a complete update of the entire global state vector from the current moment to the next moment, thus obtaining the theoretically predicted state vector.

[0062] For example, for a node representing inventory changes, its material inventory level at the next time step is calculated and updated using the following formula:

[0063] In the formula, Represents a node exist The material inventory level at any given time is One component; and They represent in Inflow and outflow nodes within a given time period The amount of material is determined by the module based on the node. It is calculated using related joint actions (e.g., upstream shipping instructions, downstream receiving instructions) and mechanistic models (e.g., transportation delay models, equipment loading and unloading efficiency models).

[0064] Regarding the implementation process of the above-mentioned model, it should be noted that the mechanistic model here is a set of mathematical equations built upon the laws of physics and the inherent logic of industrial processes (such as the laws of material balance and energy balance) to explicitly express the state transformations covering all aspects of the industrial chain, including procurement, production, and logistics. The inventory change nodes listed above... The model is a clear example. Based on the standard material balance law, it forms the core computational framework of the mechanism model by accurately summing the initial inventory of the node and the inflow within the time step and deducting the outflow.

[0065] However, in practical applications, purely mechanistic models often struggle to fully capture random errors and systemic biases caused by abnormal equipment wear, human intervention, or sudden changes in complex external environments. Therefore, this invention introduces a deep learning method, forming a coupled architecture of "mechanistic prediction + data correction" through residual modeling. Specifically, a deep learning model, such as a Long Short-Term Memory (LSTM) network or a multilayer perceptron, is introduced. During offline training, for each historical time step h, it acquires the global state and joint actions of historical time step h, outputs the corresponding theoretically predicted state vector using the mechanistic model, and acquires the actual evolution state of historical time step h+1, calculating the historical residual between the two. Using historical time steps... Using the global state and joint actions as samples, and historical residuals as labels, a deep learning model is trained using supervised learning, enabling it to learn the mapping relationship from the current state and actions to the residuals. After training, the current global state and joint actions are input into the deep learning model to generate a residual correction state vector, which is used to compensate for the prediction bias caused by complex factors in the mechanistic model.

[0066] Based on this, the system dynamically adjusts the fusion weights according to changes in operating conditions: when the magnitude or mean square error of the calculated residual correction state vector remains within a preset low threshold range, the prediction results of the characterization mechanism model are highly consistent with the actual physical environment. That is, when the system is in a stable operating period and meets physical expectations, the system will adaptively increase the value of the fusion weights corresponding to the theoretically predicted state vector to ensure the rigor of the theoretical laws. Conversely, if the value of the calculated residual correction state vector jumps abruptly or its volatility increases sharply and exceeds the preset safety threshold, that is, when the real environment encounters a sudden factor not covered by the characterization mechanism model (such as severe weather causing a sharp drop in the actual loading and unloading rate, causing the real state to completely deviate from the predicted trajectory of the baseline model's set parameters), the real-time error monitoring mechanism will prompt the system to adaptively increase the fusion weights corresponding to the residual correction state vector that represents the data-driven correction effect, thereby forcibly pulling the virtual state back to the real physical trajectory based on real-time residual data. Through the above mechanism, the system dynamically adjusts the fusion weights corresponding to the theoretically predicted state vector and the residual corrected state vector based on the difference between the actual evolution state in the real world and the predicted state by the mechanism model, and corrects the state transition calculation process. Finally, the global state at the next moment is calculated according to the following formula:

[0067] In the formula, express The global state at any given moment; This represents the fusion weights corresponding to the theoretically predicted state vector; Represents the theoretically predicted state vector; express The combined actions of moments; These represent the fusion weights corresponding to the residual corrected state vectors; Represents the residual correction state vector; This represents a deep learning model used to compute the residual correction state vector.

[0068] By integrating mechanisms and data, the system can satisfy both the underlying objective mechanism constraints and accurately depict the actual working conditions under the influence of complex environments, thus providing an accurate and reliable source of data input for subsequent multi-objective optimization and macro-benefit evaluation.

[0069] In some alternative implementations, Figure 4 This is an execution flowchart of the offline training module according to an embodiment of the present invention, such as... Figure 4As shown, the system also includes an offline training module; the offline training module is used for: setting a preset strategic intent vector; calling the hierarchical decision-making module, using an initial tactical policy network, generating simulation collaborative fine-tuning action vectors for each tactical agent based on the preset strategic intent vector and the simulation local state of each tactical agent at the current simulation time step, concatenating all simulation collaborative fine-tuning action vectors, and performing legality verification to form simulation joint actions; calling the digital twin environment construction module, generating the simulation global state for the next simulation time step in the digital twin environment based on the simulation global state and simulation joint actions at the current simulation time step; calculating the tactical reward for each tactical agent; and setting the first intersection of the current simulation time step... The interaction data is stored in the experience replay pool. The first interaction data includes the simulation local state of the current simulation time step, the preset strategic intent vector, the simulation collaborative fine-tuning action vector, the tactical reward, and the simulation local state of the next simulation time step. The first interaction data is sampled from the experience replay pool to update the initial tactical strategy network and complete the current simulation time step. The process returns to calling the hierarchical decision module, using the initial tactical strategy network, and generating the simulation collaborative fine-tuning action vector of each tactical agent based on the preset strategic intent vector and the simulation local state of each tactical agent in the current simulation time step. This process continues until the initial tactical strategy networks of all tactical agents converge, resulting in the target tactical strategy network of each tactical agent.

[0070] Specifically, the system's offline training module is divided into two stages. The first stage is used to train the target tactical policy network invoked by each tactical agent at the bottom layer. The core idea of ​​this stage is: under the premise that the high-level strategic intent is fixed, let the bottom-level tactical agents learn through repeated trial and error, and master how to make optimal micro-level decisions under given macro-level preferences.

[0071] In the first phase, the strategic intent vector of the strategic agent is fixed; that is, one or a series of preset constant values ​​are assigned to the strategic intent vector, and these values ​​remain unchanged throughout the first phase. Optionally, multiple training scenarios can be set, each corresponding to a fixed strategic intent vector, for example... This system is used to train tactical agents' execution capabilities under a single macroscopic objective. Then, simulation interactions are performed in a digital twin environment. At each simulation time step, each tactical agent receives its current simulation local state and the aforementioned preset strategic intent vector, and outputs simulation collaborative fine-tuning action vectors based on its initial tactical policy network for the current simulation time step. All simulation collaborative fine-tuning action vectors are combined and validated internally within the module to generate simulation joint actions, which are then input into the digital twin environment. The digital twin environment performs state transitions to obtain the simulation global state for the next simulation time step. Based on the state transition results, a tactical reward is calculated for each tactical agent to quantify the quality of the current action and guide the tactical policy network to update in a more optimal direction. The first interaction data of the current simulation time step, i.e., the experience tuple (the current simulation local state, preset strategic intent vector, simulation collaborative fine-tuning action vector, tactical reward, and the simulation local state for the next simulation time step), is stored in the experience replay pool corresponding to each tactical agent. Then, reinforcement learning algorithms (such as the Deep Deterministic Policy Gradient (DDPG) algorithm or the Soft Actor-Critic (SAC) algorithm) are used to sample a batch of first interaction data from the experience replay pool, calculate the loss function, and update the network parameters of the initial tactical policy network for each tactical agent to maximize the accumulated tactical reward. The updated network serves as the new initial tactical policy network, and the above process of interaction, storage, sampling, and updating is repeated until the initial tactical policy networks of all tactical agents converge, resulting in the target tactical policy network. The convergence criterion can be set as follows: in a preset number (e.g., 100) consecutive simulation interaction rounds (each round containing multiple consecutive simulation time steps), the variance of the moving average of the accumulated tactical reward obtained by the tactical agent is less than a preset threshold (e.g., 0.01), or the preset maximum number of simulation interaction rounds is reached.

[0072] In some alternative implementations, the output layer of the initial tactical policy network employs a multi-head architecture.

[0073] Specifically, since the simulated collaborative fine-tuning action vectors output by the tactical agent may simultaneously contain continuous variables (e.g., shipment volume, number of trains) and discrete variables (e.g., the on / off command for departure), and traditional reinforcement learning algorithms can typically only handle either continuous or discrete action spaces individually, they are ill-suited to directly address such mixed action scenarios. Therefore, the output layer of the initial tactical policy network is designed with a multi-head architecture: one branch is responsible for outputting continuous variables, using the Tanh activation function to map the output values ​​to a reasonable range, and optimizing them using reinforcement learning algorithms suitable for continuous action spaces, such as Deep Deterministic Policy Gradient (DDPG) or Soft Actor-Critic (SAC); the other branch is responsible for outputting discrete variables, using the Gumbel-Softmax activation function for differentiable approximation, making the discrete selection process differentiable, thus incorporating end-to-end gradient propagation, or directly updating network parameters using reinforcement learning algorithms supporting mixed action spaces, such as Proximal Policy Optimization (PPO). Through the aforementioned multi-head architecture design, the tactical strategy network can simultaneously process continuous and discrete action commands, fully covering the diverse operational needs in supply chain scheduling decisions, and improving the expressive power and decision-making accuracy of the strategy.

[0074] In some alternative implementations, the tactical reward is calculated using the following formula:

[0075] In the formula, Indexes representing tactical agents; Indicates the index of the simulation time step; Indicates the first The tactical agent in the 1st... Tactical rewards for each simulated time step; The balance coefficient is a hyperparameter between 0 and 1, used to balance local interests and global objectives. Indicates the first The tactical agent in the 1st... The intrinsic reward for each simulation time step is used to measure the local operational efficiency of the tactical agent within its own business domain. For example, for a railway transportation tactical agent, its intrinsic reward can be designed as a negative value for transportation costs or a positive value for punctuality rate; for a port operation tactical agent, its intrinsic reward can be designed as a negative value for ship waiting time or a positive value for equipment utilization rate. Indicates the first The tactical agent in the 1st... The first strategic reward for each simulation time step is calculated using the following formula:

[0076] In the formula, Represents the simulated strategic intent vector; Represents the contribution vector; An index representing macroeconomic objectives; Indicates the quantity of macroeconomic targets; Indicates the first The weights of macro-level objectives in the simulated strategic intent vector; The tactical agent is represented in the first... The simulation time step, after executing the simulation co-tuning motion vector, adjusts the first... The contribution of a macro-level objective can be calculated as follows: for example, for the supply guarantee objective, the contribution can be calculated as the amount of material delivered to downstream nodes directly caused by the simulation-coordinated fine-tuning of the action vector; for the benefit objective, the contribution can be calculated as the value of the immediate benefit generated by the simulation-coordinated fine-tuning of the action vector.

[0077] By simulating the inner product of strategic intent vector and contribution vector, the macro-level preferences of high-level strategic agents are directly transformed into a scalar reward signal, which guides tactical agents to generate actions consistent with the macro-strategy. Simultaneously, when optimizing their respective tactical rewards, each tactical agent uses the same simulated strategic intent vector to measure the contribution of its actions to various macro-level objectives in its first strategic reward component. This achieves action synergy at the underlying level, avoiding the risks of local overfitting and damage to global interests.

[0078] In some alternative implementations, such as Figure 4As shown, the offline training module is also used for: fixing the target tactical policy network for each tactical agent; in the digital twin environment, calling the hierarchical decision-making module, using the initial strategic policy network, generating a simulation strategic intent vector based on the simulation global state at the current simulation time step, and broadcasting the simulation strategic intent vector to each tactical agent; calling each tactical agent to use the fixed target tactical policy network, generating a simulation collaborative fine-tuning action vector based on the simulation strategic intent vector and the simulation local state at the current simulation time step, concatenating all simulation collaborative fine-tuning action vectors, and performing legality verification to form a simulation joint action; calling the digital twin environment construction module, in the digital twin environment, based on the simulation global state at the current simulation time step... The simulation process involves a combined action of the current state and simulation results to generate the global simulation state for the next simulation time step; calculating the second strategic reward for the strategic agent; storing the second interaction data of the current simulation time step into the experience replay pool, which includes the simulation strategic state, simulation strategic intent vector, second strategic reward, and simulation strategic state for the next simulation time step; sampling the second interaction data from the experience replay pool to update the initial strategic policy network, completing the current simulation time step; returning to the digital twin environment, invoking the hierarchical decision-making module, using the initial strategic policy network, and generating the simulation strategic intent vector based on the global simulation state of the current simulation time step, continuing until the initial strategic policy network converges to obtain the target strategic policy network.

[0079] Specifically, in the second stage of the offline training module, the system fixes the target tactical policy network of all tactical agents that have completed and converged in the first stage, so that its network parameters are no longer updated. Thus, based on the fact that the underlying policies have good execution capabilities, the system trains the high-level strategic agents to learn how to generate the optimal strategic intent vector according to different strategic states, so as to achieve dynamic trade-offs of macro objectives and maximize long-term benefits.

[0080] The initial strategy network's functional entity is a neural network model, which can employ a multilayer perceptron structure containing an input layer, multiple fully connected hidden layers (such as those based on the ReLU activation function), and an output layer. If processing time-series state features, an LSTM layer can also be introduced to extract the temporal state representation.

[0081] In the second phase of the offline training module, simulation interaction is performed in the digital twin environment. At each simulation time step, the strategic agent receives the global simulation state of the current simulation time step, aggregates it into a simulation strategic state, and outputs a simulation strategic intent vector based on its initial strategic policy network for the current simulation time step. This simulation strategic intent vector is broadcast to the fixed tactical agents. Each tactical agent outputs a joint simulation action based on its local simulation state and the aforementioned simulation strategic intent vector. The digital twin environment receives the joint simulation actions and performs a state transition to obtain the global simulation state for the next simulation time step. Based on the state transition result, the second strategic reward of the strategic agent is calculated to evaluate the long-term macroscopic effect of the current strategic decision. The second interaction data of the current simulation time step, i.e., the experience tuple (the simulation strategic state of the current simulation time step, the simulation strategic intent vector, the second strategic reward, and the simulation strategic state of the next simulation time step), is stored in the experience replay pool corresponding to the strategic agent. Then, reinforcement learning algorithms (such as the Deep Deterministic Policy Gradient (DDPG) algorithm or the Soft Actor-Critic (SAC) algorithm) are used to sample a batch of second interaction data from the experience replay pool, calculate the loss function, and update the network parameters of the strategic agent's initial strategy network to maximize the long-term accumulated second strategic reward. The updated network serves as the new initial strategy network, and the above process of interaction, storage, sampling, and updating is repeated until all the initial strategy networks of the strategic agent converge, resulting in the target strategic network. This network can generate the strategic intent vector that is most beneficial to long-term macroeconomic gains based on different strategic states. The convergence criterion can be set as follows: in a preset number of consecutive simulated interaction rounds, the variance of the moving average of the accumulated second strategic reward obtained by the strategic agent is less than a preset threshold.

[0082] After training, the system embeds the target tactical policy network of each tactical agent and the target strategic policy network of each strategic agent into a hierarchical decision-making module for online application and supports online fine-tuning to adapt to dynamically evolving environmental distributions. Optionally, these two stages can be jointly trained asynchronously or synchronously through reinforcement learning mechanisms.

[0083] In some alternative implementations, the second strategic reward is calculated using the following formula:

[0084] In the formula, This indicates the second strategic reward; Represents the macroeconomic return vector; Indicates the first The first macroeconomic goal in The achieved value of each simulation time step.

[0085] To quantify the achievement of macro-level objectives, taking the cost target as an example, the negative performance of the sum of all measurable cost items generated by the system can be measured by the following formula, which can be used as the achievement value:

[0086] In the formula, This indicates the value achieved in fulfilling the cost target; Indicates the unit transportation rate; Indicates the unit storage fee rate; Indicates the first The actual transport volume of each transport section at each simulation time step; Indicates the first The actual inventory levels at each warehouse node under each simulation time step. and All of them are directly derived from the global simulation state of the next simulation time step output after the state transition deduction of the digital twin environment model; This indicates the penalty cost incurred for failing to fulfill obligations on time. It is calculated by statistically analyzing all non-compliant orders within the current decision-making cycle, and then calculating the stockout quantity for each order (i.e., the target demand quantity versus the system's stockout quantity in the specified period). The positive part of the difference between the actual delivery volume at each simulation time step is multiplied by the preset unit default penalty rate and summed to obtain the result.

[0087] Since the above formula has a negative sign, meaning that the lower the cost, the greater the achievement value of the objective function, it guides the reinforcement learning agent to optimize the strategy in the direction of reducing the overall cost.

[0088] Taking the macro-level objective of fulfilling the planned contract target as an example, the progress of delivery completion can be measured using the following formula, which can be used as its achievement value:

[0089] In the formula, This indicates the achievement value of the planned performance targets; This indicates that a specific delivery cycle is extrapolated to the next stage. The actual delivery completion amount under each simulation time step is derived from the actual cumulative delivery status value of the end node (such as the target port or unloading terminal) output by the digital twin environment construction module after state transition simulation. This parameter represents the pre-determined demand (or target planned quantity) within the same specific delivery cycle. It serves as a preset boundary condition for the system and is directly obtained from the production schedule or customer order data issued by the upper-level business system.

[0090] By comparing the two, the system can quantitatively assess the contribution of the current scheduling strategy to the completion of the final delivery task.

[0091] Optionally, for other macro-level objectives, their achievement values ​​can be quantified.

[0092] The supply chain scheduling and decision-making system provided in this embodiment, based on data and mechanism fusion, unifies diverse and heterogeneous real-time data into a global state through data fusion and structuring modules, providing a complete and consistent data foundation for decision-making and overcoming the limitations of the breadth and depth of information processing in manual decision-making. The hierarchical decision-making module generates joint actions based on a hierarchical architecture and feeds them back to the digital twin environment, avoiding the complexity explosion problem caused by solving all variables with a single model and improving the real-time performance of decisions. Through a deduction method that integrates mechanism models and data correction in the digital twin environment, while ensuring physical constraints, data-driven correction of model biases is used, improving the accuracy of state prediction and enabling dynamic adaptation to sudden operating conditions such as equipment failures and weather changes, overcoming the shortcomings of traditional operations research models that have rigid structures and are difficult to dynamically adjust. The synergistic effect of these modules realizes fully automated scheduling and decision-making across the entire chain, from data acquisition, state fusion, hierarchical decision-making to simulation deduction and closed-loop feedback, providing continuously optimized decision instructions for the supply chain and improving the efficiency of collaborative operation and adaptability to dynamic environments.

[0093] The present invention has the following beneficial effects: (1) This invention decomposes the complex whole-chain optimization problem into two levels: macro-intention formulation and micro-coordinated execution by constructing a hierarchical decision-making module of high-level strategic intelligent agent and low-level tactical intelligent agent cluster; and by using strategic intention vector and strategic reward mechanism based on contribution vector calculation, a clear quantitative relationship is established between the two levels, which solves the technical problem that decision-making objectives are difficult to unify under different spatiotemporal scales, and realizes the decoupling and linkage of decision-making. (2) This invention ensures that the entire underlying tactical agent cluster works collaboratively under unified and dynamically changing macro guidance by dynamically generating strategic intent vectors based on real-time changing global states through the mechanism of high-level strategic agents and broadcasting them to all tactical agents. This enables the industry chain to respond quickly and in a coordinated manner to internal disturbances and changes in the external environment as a whole, thereby improving the system’s synergy and adaptability. (3) This invention structures the decision-making process into hierarchical decision-making modules and generates a clear intermediate output, the strategic intent vector, making the system's decision priorities at a specific moment transparent and analyzable. By examining the weight distribution of the strategic intent vector, the system's current decision preferences can be directly obtained. Compared with a single, end-to-end optimization model, this significantly enhances the structuring and interpretability of the decision-making process.

[0094] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A supply chain scheduling decision-making system based on data and mechanism fusion, characterized in that, The system includes a data fusion and structuring module, a hierarchical decision-making module, and a digital twin environment construction module; The data fusion and structuring module is used to receive diverse and heterogeneous real-time data from the industry chain and generate the global state at the current moment. The hierarchical decision-making module is used to generate joint actions for the current moment based on the global state at the current moment through a hierarchical architecture, and feed them back to the digital twin environment construction module. The digital twin environment construction module is used to build a digital twin environment that operates synchronously with the industry chain; The digital twin environment construction module is also used to calculate the global state at the next moment based on the current global state and the current joint actions in the digital twin environment by combining the mechanism model and data correction, and send it to the hierarchical decision module.

2. The system according to claim 1, characterized in that, The data fusion and structuring module is specifically used for: Receive and preprocess the diverse and heterogeneous real-time data from the aforementioned industry chain to obtain preprocessed data; The preprocessed data is divided into node state sub-vectors, transportation unit state sub-vectors, key equipment state sub-vectors, and external environment state sub-vectors to form the global state at the current moment.

3. The system according to claim 1, characterized in that, The system also includes a scheduling and execution module; The hierarchical decision-making module is also used to feed back the joint action at the current moment to the scheduling execution module; The scheduling and execution module is used to schedule and control the industrial chain according to the joint actions at the current moment, and to collect the actual evolution status of the industrial chain after executing the joint actions at the current moment.

4. The system according to claim 3, characterized in that, The hierarchical decision-making module includes a strategic intelligent agent and multiple tactical intelligent agents; The strategic agent is used to aggregate the current global state into a strategic state and input it into the target strategic policy network to obtain a strategic intent vector, which includes the decision weight of each macro-objective. The strategic agent is also used to broadcast the strategic intent vector to each tactical agent; Each tactical agent is used to extract the local state of its own business domain from the global state at the current moment, and input the local state and the strategic intent vector into the target tactical policy network to obtain a collaborative fine-tuning action vector. The hierarchical decision-making module is also used to concatenate all collaborative fine-tuning action vectors and perform legality verification to form the joint action at the current moment.

5. The system according to claim 4, characterized in that, The digital twin environment construction module is specifically used for: In the digital twin environment, for each collaborative fine-tuning action vector in the joint action at the current moment, the mechanism model corresponding to each collaborative fine-tuning action vector is invoked; The mechanism model corresponding to each collaborative fine-tuning action vector is adopted. Based on the global state at the current moment and the corresponding collaborative fine-tuning action vector, the calculation is performed to obtain the local state at the next moment corresponding to each collaborative fine-tuning action vector. All the local states at the next moment are used to form the theoretical prediction state vector. A deep learning model is used to generate a residual corrected state vector based on the global state and the joint action at the current time. The deep learning model is trained based on the difference between the theoretically predicted state vector and the actual evolution state. The fusion weights are dynamically determined based on the residual correction state vector. Based on the theoretically predicted state vector, the residual corrected state vector, and the fusion weights, the global state at the next time step is calculated.

6. The system according to claim 5, characterized in that, The system further includes an offline training module; the offline training module is used for: Set a preset strategic intent vector; The hierarchical decision-making module is invoked, and an initial tactical strategy network is adopted. Based on the preset strategic intent vector and the simulation local state of each tactical agent at the current simulation time step, a simulation collaborative fine-tuning action vector for each tactical agent is generated. All simulation collaborative fine-tuning action vectors are spliced ​​together and their legality is verified to form a simulation joint action. The digital twin environment construction module is invoked, and in the digital twin environment, the global state of the simulation for the next simulation time step is generated based on the global state of the simulation at the current simulation time step and the joint simulation action. Calculate the tactical reward for each tactical agent; The first interaction data of the current simulation time step is stored in the experience replay pool. The first interaction data includes the simulation local state of the current simulation time step, the preset strategic intent vector, the simulation collaborative fine-tuning action vector, the tactical reward, and the simulation local state of the next simulation time step. Sample the first interaction data from the experience replay pool, update the initial tactical strategy network, and complete the current simulation time step; Returning to the call to the hierarchical decision-making module, the initial tactical strategy network is adopted. Based on the preset strategic intent vector and the simulation local state of each tactical agent at the current simulation time step, the simulation collaborative fine-tuning action vector of each tactical agent is generated. This process continues until the initial tactical strategy networks of all tactical agents converge, resulting in the target tactical strategy network of each tactical agent.

7. The system according to claim 6, characterized in that, The output layer of the initial tactical strategy network adopts a multi-head architecture.

8. The system according to claim 6, characterized in that, The tactical reward is calculated using the following formula: In the formula, Indexes representing tactical agents; Indicates the index of the simulation time step; Indicates the first The tactical agent in the 1st... Tactical rewards for each simulated time step; Indicates the balance coefficient; Indicates the first The tactical agent in the 1st... The intrinsic reward for each simulation time step; Indicates the first The tactical agent in the 1st... The first strategic reward for each simulation time step is calculated using the following formula: In the formula, Represents the simulated strategic intent vector; Represents the contribution vector; An index representing macroeconomic objectives; Indicates the quantity of macroeconomic targets; Indicates the first The weights of macro-level objectives in the simulated strategic intent vector; The tactical agent is represented in the first... The simulation time step for the first The degree of contribution to each macroeconomic goal.

9. The system according to claim 6, characterized in that, The offline training module is also used for: A fixed-target tactical policy network for each tactical agent; In the digital twin environment, the hierarchical decision-making module is invoked, and an initial strategic policy network is adopted to generate a simulation strategic intent vector based on the simulation global state at the current simulation time step, and the simulation strategic intent vector is broadcast to each tactical agent. Each tactical agent is invoked to adopt a fixed target tactical strategy network. Based on the simulated strategic intent vector and the simulated local state at the current simulation time step, a simulated collaborative fine-tuning action vector is generated. All simulated collaborative fine-tuning action vectors are spliced ​​together and their legality is verified to form a simulated joint action. The digital twin environment construction module is invoked, and in the digital twin environment, the global state of the simulation for the next simulation time step is generated based on the global state of the simulation at the current simulation time step and the joint simulation action. Calculate the second strategic reward for the strategic agent; The second interaction data of the current simulation time step is stored in the experience replay pool. The second interaction data includes the simulation strategy state of the current simulation time step, the simulation strategy intent vector, the second strategy reward, and the simulation strategy state of the next simulation time step. Sample the second interaction data from the experience replay pool, update the initial strategy network, and complete the current simulation time step; Returning to the digital twin environment, the hierarchical decision-making module is invoked, and the initial strategic policy network is used to generate a simulation strategic intent vector based on the simulation global state at the current simulation time step, until the initial strategic policy network converges to obtain the target strategic policy network.

10. The system according to claim 9, characterized in that, The second strategic reward is calculated using the following formula: In the formula, This indicates the second strategic reward; Represents the macroeconomic return vector; Indicates the first The first macroeconomic goal in The achieved value of each simulation time step.

Images