Source network load storage collaborative industrial production line dynamic multi-game optimization scheduling method and device

By collecting data from industrial production lines and microgrids, extracting features using one-dimensional temporal convolutional networks and graph attention networks, constructing a master-slave game environment, and combining a meta-reinforcement learning model for multi-game optimization scheduling, the problems of privacy leakage, production cycle disruption, and carbon compliance in industrial park power load scheduling are solved, and global optimal scheduling that can quickly adapt to sudden disturbances is achieved.

CN123073804APending Publication Date: 2026-11-10BEIJING EASY TIMES DIGITAL TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610999609.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-07
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies for power load dispatching in industrial parks have problems such as privacy leakage risks, disruption of production cycle, difficulty in achieving carbon compliance, and insufficient generalization ability of new disturbances.

Method used

By collecting micro-production data from industrial production lines and macro-energy data from microgrids, feature vectors are extracted using one-dimensional temporal convolutional networks and graph attention networks. A Stackelberg master-slave game environment is constructed, and multi-game optimization scheduling is performed using a meta-reinforcement learning model to achieve synchronous output of continuous energy storage actions and discrete production scheduling actions.

Benefits of technology

It achieves data decoupling between microgrids and manufacturing execution systems, solves the problem of multi-timescale data fusion, quickly adapts to sudden disturbances, and balances global Pareto optimal scheduling for both new energy arbitrage and low-carbon production.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123073804A_ABST
    Figure CN123073804A_ABST
Patent Text Reader

Abstract

The embodiment of the specification provides an industrial production line dynamic multi-game optimization scheduling method and device of source network load storage cooperation, wherein the method realizes physical and information decoupling between the microgrid and the MES through a master-slave game model, breaks the privacy island; through a multi-scale graph time sequence fusion mechanism, macro and micro characteristics are aligned, and the problem of system scheduling time step misalignment is solved; and through the introduction of a meta-reinforcement learning framework and a double-branch mixed action space network, when the system faces a sudden heterogeneous disturbance, only a few samples are needed to complete local parameter online adaptive iteration within milliseconds, and finally the globally Pareto optimal scheduling of microgrid new energy arbitrage and production line low-carbon on-time production is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments in this specification relate to the field of power technology, and in particular to a dynamic multi-game optimization scheduling method for industrial production lines that integrates power generation, grid, load, and energy storage. Background Technology

[0002] Against the backdrop of global macroeconomic policies aimed at achieving carbon peaking and carbon neutrality, the penetration rate of renewable energy sources such as wind and solar power, characterized by intermittency and volatility, in modern power systems is growing exponentially. Traditional industrial park electricity loads are typically considered rigid loads, passively receiving data from the grid. However, with the evolution of the integrated energy system concept encompassing generation, grid, load, and storage, industrial parks are gradually transforming into highly resilient and self-regulating micro-energy internet nodes. Deeply exploring the flexible regulation potential of high-energy-consuming industrial loads can not only significantly improve the local absorption rate of new energy within regional microgrids but also effectively reduce the overall operating costs of the system.

[0003] Existing technologies have significant technical drawbacks in large-scale deployment in real-world industrial scenarios: First, existing centralized macro-level joint dispatching schemes require the MES (Manufacturing Execution System) to fully expose core business secrets (such as order data and process parameters) to the microgrid, which poses significant privacy risks and management challenges in commercial practice. Simultaneously, these schemes often crudely abstract the "load" end into power-consuming nodes that can be represented by continuous functions, completely stripping away the extremely complex discrete logic within the industrial production line (such as process sequence dependencies and machine tool non-preemption constraints). This results in the output control strategy being unable to be implemented without disrupting the production cycle, severely limiting its practicality.

[0004] Secondly, existing micro-reinforcement learning scheduling methods only focus on the internal optimization of production facilities and do not involve deep coupling with external energy systems (such as grid carbon emission factors and microgrid energy storage). Under increasingly stringent carbon quota systems, they cannot meet the dual carbon compliance requirements of enterprises.

[0005] Finally, existing microgrid scheduling methods based on traditional transfer learning or deep reinforcement learning have insufficient generalization ability when faced with new disturbance scenarios that deviate completely from the historical distribution. Transfer learning can usually only transfer knowledge unidirectionally between similar source and target domains, resulting in slow response speed and easy loss of local optimal solutions.

[0006] Therefore, a better solution is urgently needed. Summary of the Invention

[0007] In view of this, embodiments of this specification provide a dynamic multi-game optimization scheduling method for industrial production lines that integrates source-grid-load-storage coordination. One or more embodiments of this specification also relate to a dynamic multi-game optimization scheduling device for industrial production lines that integrates source-grid-load-storage coordination, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0008] According to a first aspect of the embodiments of this specification, a dynamic multi-game optimization scheduling method for industrial production lines with source-grid-load-storage coordination is provided, comprising: Collect micro-production data from industrial production lines and macro-energy data from microgrids. The micro-production data includes real-time power consumption of machine tools and topology dependencies of order processes, while the macro-energy data includes the state of charge of energy storage batteries, photovoltaic power output forecasts, grid time-of-use electricity prices, and carbon emission factors. Low-frequency temporal feature vectors of macroscopic energy data are extracted by a one-dimensional temporal convolutional network, and the topological graph containing machine tool status and order process topological dependencies is transformed into high-frequency graph topological feature vectors by a graph attention network. In the hidden layer space of the neural network, the low-frequency temporal feature vectors and high-frequency graph topological feature vectors are dimensionally aligned and spliced ​​to generate a unified joint Markov state representation. An interactive environment based on Stackelberg master-slave game is constructed, with the microgrid energy management system as the leader and the manufacturing execution system as the follower. The joint Markov state representation is input into the pre-trained meta-reinforcement learning model. The leader publishes the internal interactive electricity price, and the follower adjusts the production schedule and feeds back the expected load according to the internal interactive electricity price, so as to approach the Nash equilibrium point through multiple interactive iterations. Through the dual-branch hybrid action network of the meta-reinforcement learning model, a joint scheduling strategy including continuous energy storage actions and discrete production scheduling actions is output at the same frequency. The continuous energy storage actions include the energy storage charge and discharge rate, and the discrete production scheduling actions include the selection of machine tool processing steps. The continuous energy storage actions in the joint scheduling strategy are parsed into active power adjustment commands and sent to the energy storage battery management system and photovoltaic inverter. The discrete production scheduling actions are parsed into manufacturing work orders and sent to the workshop programmable logic controller. The feedback status after the physical equipment is executed is obtained to form a dynamic closed-loop control.

[0009] In one possible implementation, an interactive environment based on a Stackelberg master-slave game is constructed, which further includes: the leader aims to maximize its own net income, which includes revenue from selling electricity to the main grid, revenue from photovoltaic power generation, energy storage losses, and the cost of purchasing electricity from the main grid; the follower aims to minimize the overall production cost, which includes electricity costs, carbon emission quota exceeding costs, and penalties for order overdue defaults.

[0010] In one possible implementation, the pre-trained meta-reinforcement learning model is trained through an inner and outer double-layer cyclic differentiation mechanism. The inner loop uses a small amount of real-time data under perturbation scenarios to update local gradients to obtain temporary scheduling parameters; the outer loop collects the performance of all temporary scheduling parameters under the corresponding scenarios and updates the global meta-model parameters based on backpropagation of the multi-objective Pareto reward function.

[0011] In one possible implementation, the dual-branch hybrid action network includes: a continuous action branch for outputting the energy storage charge / discharge rate; and a discrete action branch for outputting the selection of machine tool processing steps; the energy storage charge / discharge rate and the selection of machine tool processing steps are output in tandem by the meta-reinforcement learning model.

[0012] In one possible implementation, low-frequency temporal feature vectors of macroscopic energy data are extracted through a one-dimensional temporal convolutional network, further including: extracting slowly changing features; the topological graph containing machine tool status and order process topological dependencies is transformed into a high-frequency graph topological feature vector through a graph attention network, further including: extracting rapidly changing features; the low-frequency temporal feature vectors and high-frequency graph topological feature vectors are dimensionally aligned and concatenated in the hidden layer space of the neural network, further including: fusing slowly changing features and rapidly changing features through an attention mechanism.

[0013] One possible implementation also includes: when an unknown disturbance such as a sudden change in the external spot market electricity price, a sharp drop in photovoltaic output, or an internal emergency order insertion is detected, the real-time observation status data at the current moment is captured, and local gradient updates are performed based on the pre-trained meta-model parameters to generate a locally optimal joint scheduling strategy for the current disturbance scenario.

[0014] In one possible implementation, the continuous energy storage actions in the joint scheduling strategy are parsed into active power adjustment commands and sent to the energy storage battery management system and the photovoltaic inverter, further including: sending commands via the Modbus TCP protocol; the discrete production scheduling actions are parsed into manufacturing work orders and sent to the shop floor programmable logic controller, further including: sending commands via the OPCUA protocol.

[0015] According to a second aspect of the embodiments of this specification, a dynamic multi-game optimization scheduling device for industrial production lines with source-grid-load-storage coordination is provided, comprising: The data acquisition module is configured to collect micro-production data from industrial production lines and macro-energy data from microgrids. The micro-production data includes real-time power consumption of machine tools and topology dependencies of order processes, while the macro-energy data includes the state of charge of energy storage batteries, photovoltaic power output forecasts, grid time-of-use electricity prices, and carbon emission factors. The feature determination module is configured to extract low-frequency temporal feature vectors from macroscopic energy data through a one-dimensional temporal convolutional network, and transform the topological graph containing machine tool status and order process topological dependencies into high-frequency graph topological feature vectors through a graph attention network. In the hidden layer space of the neural network, the low-frequency temporal feature vectors and high-frequency graph topological feature vectors are dimensionally aligned and concatenated to generate a unified joint Markov state representation. The environment construction module is configured to build an interactive environment based on Stackelberg master-slave game, with the microgrid energy management system as the leader and the manufacturing execution system as the follower. The joint Markov state representation is input into the pre-trained meta-reinforcement learning model. The leader publishes the internal interactive electricity price, and the follower adjusts the production schedule according to the internal interactive electricity price and feeds back the expected load, so as to approach the Nash equilibrium point through multiple interactive iterations. The strategy output module is configured to output a joint scheduling strategy that includes continuous energy storage actions and discrete production scheduling actions through a dual-branch hybrid action network of a meta-reinforcement learning model. The continuous energy storage actions include the energy storage charge and discharge rate, and the discrete production scheduling actions include the selection of machine tool processing steps. The instruction issuing module is configured to parse the continuous energy storage actions in the joint scheduling strategy into active power adjustment instructions and issue them to the energy storage battery management system and photovoltaic inverter, parse the discrete production scheduling actions into manufacturing work orders and issue them to the workshop programmable logic controller, and obtain the feedback status of the physical equipment after execution to form dynamic closed-loop control.

[0016] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the above-mentioned source-grid-load-storage coordinated dynamic multi-game optimization scheduling method for industrial production lines.

[0017] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the above-described source-grid-load-storage coordinated dynamic multi-game optimization scheduling method for industrial production lines.

[0018] According to a fifth aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, the computer is instructed to perform the steps of the above-described source-grid-load-storage coordinated dynamic multi-game optimization scheduling method for industrial production lines.

[0019] This specification provides a method and apparatus for dynamic multi-game optimization scheduling of industrial production lines with source-grid-load-storage coordination. The method achieves physical and information decoupling between the microgrid and the MES through a master-slave game model, breaking down privacy silos; it aligns macro and micro features through a multi-scale graph time-series fusion mechanism to solve the problem of misaligned system scheduling time steps; and by introducing a meta-reinforcement learning framework and a dual-branch hybrid action space network, the system can complete online adaptive iteration of local parameters within milliseconds with only a few samples when facing sudden heterogeneous disturbances, ultimately achieving global Pareto optimal scheduling that balances microgrid renewable energy arbitrage with low-carbon and on-time production of the production line. Attached Figure Description

[0020] Figure 1 This is a flowchart of a dynamic multi-game optimization scheduling method for industrial production lines that integrates source-grid-load-storage coordination, provided in one embodiment of this specification. Figure 2 This is a system architecture diagram of a source-grid-load-storage collaborative dynamic multi-game optimization scheduling method for industrial production lines, provided in one embodiment of this specification. Figure 3 This is a schematic diagram of the structure of a source-grid-load-storage coordinated dynamic multi-game optimization scheduling device for industrial production lines, provided in one embodiment of this specification. Figure 4 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0021] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0022] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0023] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0024] This specification provides a dynamic multi-game optimization scheduling method for industrial production lines that integrates source, grid, load, and storage. This specification also relates to a dynamic multi-game optimization scheduling device for industrial production lines that integrates source, grid, load, and storage, a computing device, and a computer-readable storage medium, which will be described in detail in the following embodiments.

[0025] See Figure 1 , Figure 1 A flowchart is shown for a source-grid-load-storage collaborative dynamic multi-game optimization scheduling method for industrial production lines according to an embodiment of this specification, which specifically includes the following steps.

[0026] Step 201: Collect micro-production data from industrial production lines and macro-energy data from microgrids. The micro-production data includes real-time power consumption of machine tools and order process topology dependencies, while the macro-energy data includes the state of charge of energy storage batteries, photovoltaic power output forecasts, grid time-of-use electricity prices, and carbon emission factors. Low-frequency temporal feature vectors of macroscopic energy data are extracted by a one-dimensional temporal convolutional network, and the topological graph containing machine tool status and order process topological dependencies is transformed into high-frequency graph topological feature vectors by a graph attention network. In the hidden layer space of the neural network, the low-frequency temporal feature vectors and high-frequency graph topological feature vectors are dimensionally aligned and spliced ​​to generate a unified joint Markov state representation. An interactive environment based on Stackelberg master-slave game is constructed, with the microgrid energy management system as the leader and the manufacturing execution system as the follower. The joint Markov state representation is input into the pre-trained meta-reinforcement learning model. The leader publishes the internal interactive electricity price, and the follower adjusts the production schedule and feeds back the expected load according to the internal interactive electricity price, so as to approach the Nash equilibrium point through multiple interactive iterations. Through the dual-branch hybrid action network of the meta-reinforcement learning model, a joint scheduling strategy including continuous energy storage actions and discrete production scheduling actions is output at the same frequency. The continuous energy storage actions include the energy storage charge and discharge rate, and the discrete production scheduling actions include the selection of machine tool processing steps. The continuous energy storage actions in the joint scheduling strategy are parsed into active power adjustment commands and sent to the energy storage battery management system and photovoltaic inverter. The discrete production scheduling actions are parsed into manufacturing work orders and sent to the workshop programmable logic controller. The feedback status after the physical equipment is executed is obtained to form a dynamic closed-loop control.

[0027] Among them, the micro-production data of industrial production lines can refer to a set of data characterizing the internal production status of the production line, such as the real-time operating power and processing status of equipment such as CNC machine tools and robotic arms, as well as the process topology describing the order processing sequence and resource dependencies. The macro-energy data of microgrids can refer to a set of data characterizing the energy supply and demand status of the microgrid side, such as the remaining percentage of energy storage batteries, the predicted power curve of photovoltaic power generation systems, the time-of-use price of electricity purchased from the main grid, and the carbon dioxide emission coefficient corresponding to the unit of electricity consumption.

[0028] One-dimensional temporal convolutional networks can refer to a type of convolutional neural network structure used to process temporal data. They effectively extract long-term dependency features from temporal data through causal convolution and dilated convolution techniques. Graph attention networks can refer to a type of neural network based on a graph structure. They effectively extract inter-node correlation features by assigning different weights to different neighboring nodes in the graph through an attention mechanism.

[0029] The joint Markov state representation can refer to the unified state vector formed by aligning the slowly changing characteristics of the macroscopic energy system with the rapidly changing characteristics of the microscopic production system in the hidden layer space. This vector satisfies the Markov property, that is, the future state depends only on the current state.

[0030] Stackelberg master-follower game can refer to a hierarchical game model where the leader acts first, and followers respond optimally based on the leader's actions. The leader anticipates the followers' responses and makes the optimal decision. A microgrid energy management system can refer to the core management system responsible for the coordinated control and optimal scheduling of power sources, grid, load, and storage within a microgrid. A manufacturing execution system can refer to the factory-level management system responsible for executing production plans, scheduling processes, and monitoring equipment status in the workshop. Internal interactive pricing can refer to the electricity trading price provided by the microgrid to the production line; this price is set and published by the microgrid to guide the production line's electricity consumption behavior. A Nash equilibrium point can indicate a stable state reached by both parties in the strategy space, where unilaterally changing the strategy of either party will not yield additional gains.

[0031] Meta-reinforcement learning models refer to reinforcement learning models trained using a meta-learning framework. By pre-training on massive amounts of tasks, the model gains the ability to quickly adapt to new tasks. Dual-branch hybrid action networks refer to neural network structures where the output layer contains two independent branches, one branch outputting continuous-value actions and the other branch outputting discrete-value actions.

[0032] A battery management system (BMS) can refer to a system used to monitor and manage the status of energy storage battery packs and control the charging and discharging process. A photovoltaic inverter can refer to a power electronic device that converts direct current (DC) generated by photovoltaic modules into alternating current (AC) for grid connection or supply to loads. A programmable logic controller (PLC) can refer to industrial computer equipment used for logic control and sequential control in industrial control systems.

[0033] The present application will be further described below through a detailed embodiment: This embodiment focuses on an industrial park scenario. The park includes a microgrid system equipped with photovoltaic arrays and energy storage battery packs, as well as a flexible manufacturing workshop containing multiple CNC machine tools and robotic arms. The microgrid is controlled by a microgrid energy management system, while the workshop is controlled by a manufacturing execution system. The two are physically isolated and exchange data through an encrypted communication interface.

[0034] First, the system collects micro-level production data from industrial production lines and macro-level energy data from the microgrid. Specifically, it acquires the real-time operating power of each CNC machine tool in the workshop through industrial IoT sensors, and simultaneously reads the process topology dependencies of the current production order from the Manufacturing Execution System (MES), including the process steps contained in each order, the sequential constraints between processes, and the machine tool resources available for each process. At the same time, the system obtains the state of charge of energy storage batteries, the output forecast curve of the photovoltaic power generation system for a future period, the next-day time-of-use electricity price published by the grid company, and the carbon emission coefficient per unit of electricity consumption calculated based on the grid carbon emission factor from the microgrid energy management system.

[0035] Subsequently, the system extracts low-frequency temporal feature vectors from macroeconomic energy data using a one-dimensional temporal convolutional network. This one-dimensional temporal convolutional network employs a dilated convolution structure, effectively capturing daily cyclical patterns and trend characteristics in electricity price sequences and photovoltaic power output prediction sequences. Simultaneously, the system transforms a topological graph containing machine tool status and order process topological dependencies into a high-frequency graph topological feature vector using a graph attention network. Specifically, each machine tool is considered a node, with node features including current operating power and idle status. Dependencies between processes are treated as directed edges. The graph attention network extracts high-dimensional feature representations reflecting process density and machine tool load balance by calculating attention weights between adjacent nodes. Then, in the hidden space of the neural network, the system uses an attention mechanism to dimensionally align and concatenate the aforementioned low-frequency temporal feature vectors with the high-frequency graph topological feature vectors, generating a unified joint Markov state representation. This representation simultaneously includes the changing trends of the external energy environment and the real-time status of the internal production system.

[0036] Next, the system constructs an interactive environment based on a Stackelberg master-slave game, with the microgrid energy management system as the leader and the manufacturing execution system as the follower. The previously generated joint Markov state representation is input into a pre-trained meta-reinforcement learning model. The microgrid energy management system, as the leader, first publishes a set of future internal interactive electricity price sequences, reflecting the incentive signals from the microgrid to guide production line electricity consumption. Upon receiving this electricity price sequence, the manufacturing execution system, as the follower, adjusts its production schedule while satisfying process sequence dependencies and machine resource constraints, scheduling high-energy-consuming processes during periods of lower electricity prices, and feeding back the adjusted expected load curve to the microgrid energy management system. Based on the feedback load curve, the microgrid energy management system reassesses its own revenue and adjusts the next published electricity price sequence. Through multiple rounds of interactive iteration, both parties approach the Nash equilibrium point of the Stackelberg game. At this equilibrium point, the microgrid energy management system can no longer increase its revenue by further adjusting electricity prices, and the manufacturing execution system can no longer reduce its overall costs by further adjusting its production schedule.

[0037] Subsequently, the system uses a dual-branch hybrid action network of a meta-reinforcement learning model to output a joint scheduling strategy that includes continuous energy storage actions and discrete production scheduling actions. The continuous action branch outputs the energy storage charge / discharge rate, for example, a value ranging from negative to positive, where negative values ​​indicate charging and positive values ​​indicate discharging. The discrete action branch outputs the selection of machine tool processing operations, for example, selecting one or more operations from the currently executable set and assigning them to a specified machine tool.

[0038] Finally, the system parses the continuous energy storage actions in the joint scheduling strategy into active power regulation commands, which are then sent to the energy storage battery management system and photovoltaic inverter via the Modbus TCP protocol. This controls the energy storage battery to charge and discharge at a specified rate and controls the photovoltaic inverter to regulate reactive power output. The system also parses the discrete production scheduling actions into manufacturing work orders, which are sent to the shop floor programmable logic controller (PLC) via the OPC UA protocol. This controls the start / stop, spindle speed adjustment, or standby / hibernation of CNC machine tools. After the physical equipment executes the commands, the actual power consumption and the actual processing progress are fed back to the system as new status inputs, forming a complete dynamic closed-loop control.

[0039] The beneficial effects of this embodiment are as follows: By constructing a dual-stream feature extraction mechanism that combines a one-dimensional temporal convolutional network and a graph attention network, the slow-changing characteristics of the macro-energy system and the fast-changing characteristics of the micro-production system are effectively aligned, solving the dimensionality curse problem of multi-timescale data fusion; by using the Stackelberg master-slave game model, data decoupling between the microgrid and the manufacturing execution system is achieved, approaching the global optimum without exchanging underlying order privacy and equipment-level parameters; by using a meta-reinforcement learning model and a dual-branch hybrid action network, the synchronous joint optimization of continuous energy storage actions and discrete production scheduling actions is achieved, providing a complete closed-loop solution for source-grid-load-storage coordinated scheduling.

[0040] In one possible implementation, an interactive environment based on a Stackelberg master-slave game is constructed, which further includes: the leader aims to maximize its own net income, which includes revenue from selling electricity to the main grid, revenue from photovoltaic power generation, energy storage losses, and the cost of purchasing electricity from the main grid; the follower aims to minimize the overall production cost, which includes electricity costs, carbon emission quota exceeding costs, and penalties for order overdue defaults.

[0041] The microgrid's net revenue can refer to the core optimization goal pursued by the microgrid energy management system as a leader. This can be calculated as the revenue from selling electricity to the main grid plus the savings in electricity purchase costs due to photovoltaic (PV) power being absorbed by production lines, minus the lifespan depreciation costs of energy storage batteries during charge-discharge cycles, and the cost of purchasing electricity from the main grid. Revenue from selling electricity to the main grid can refer to the revenue the microgrid receives from selling excess electricity to the main grid when there is excess PV power generation or when energy storage is discharging. PV absorption revenue can refer to the value corresponding to PV power being directly absorbed by production line loads, equivalent to the savings from avoiding purchasing electricity from the main grid at a higher price. Energy storage losses can refer to the cost reduction incurred by energy storage batteries during charge-discharge processes due to energy conversion efficiency and cycle life degradation. The cost of purchasing electricity from the main grid can refer to the cost paid when the microgrid's own power generation is insufficient to purchase electricity from the main grid.

[0042] Total production cost can refer to the core objective of a manufacturing execution system (MES) as a follower pursuing optimization. For example, it equals the electricity cost incurred during production, plus the cost of purchasing additional carbon emission reductions due to exceeding the free allowance, plus penalties for late order delivery. Electricity cost can refer to the total electricity bill calculated based on the microgrid's internal interactive electricity price and the actual electricity consumption of the production line. Carbon emission allowance overrun cost refers to the cost of purchasing additional allowances in the carbon trading market when the actual carbon emissions of the production line exceed the allocated free allowance. Order late delivery penalty can refer to the cost of contract penalties or reputational damage incurred due to production scheduling adjustments that prevent orders from being completed before the promised delivery date.

[0043] The present application will be further described below through a detailed embodiment, which further defines the objective function of both parties in the game based on the method described in claim 1.

[0044] Building upon the aforementioned embodiments, the microgrid energy management system, acting as the leader, aims to maximize its net revenue. Specifically, when formulating its internal interactive pricing sequence, the microgrid energy management system comprehensively considers the following four aspects of revenue and cost: When the photovoltaic (PV) power generation within the microgrid exceeds the production line load demand, the excess electricity can be sold to the main grid, generating revenue from selling electricity to the main grid; when the production line load directly consumes PV power generation, the microgrid avoids purchasing electricity from the main grid, and this saving is considered PV consumption revenue; energy storage batteries experience energy conversion efficiency losses during charging and discharging, and each charge-discharge cycle affects battery life, therefore, energy storage losses are included as a cost item; when PV output is insufficient and energy storage discharge is insufficient to meet the load, the microgrid needs to purchase electricity from the main grid, incurring costs associated with purchasing electricity from the main grid. By optimizing the internal interactive pricing sequence, the microgrid energy management system guides the production line load curve to better match the PV output curve, thereby maximizing its net revenue.

[0045] As a follower, the Manufacturing Execution System (MES) aims to minimize its overall production cost. Specifically, after receiving the internal interactive electricity price from the microgrid, the MES optimizes its production scheduling while meeting process and resource constraints. Its overall production cost comprises three components: electricity cost (calculated based on the internal interactive electricity price and the actual electricity consumption of the production line at each time period); carbon emission allowance exceeding limits (calculated by multiplying the electricity consumption of the production line at each time period by the corresponding carbon emission factor; if the free allowance is exceeded, the excess must be purchased at the carbon trading market price); and order overdue penalty (if production scheduling adjustments cause some orders to fail to be completed before the contractually agreed delivery date, corresponding penalties are paid according to the contract). The MES seeks to minimize the sum of these three cost items by optimizing its production schedule, guided by the internal interactive electricity price.

[0046] The microgrid energy management system and the manufacturing execution system approach the Nash equilibrium point of the Stackelberg game through multiple rounds of interactive iteration. At this equilibrium point, the microgrid energy management system cannot further increase its net revenue by adjusting the electricity price, and the manufacturing execution system cannot further reduce the overall production cost by adjusting the production schedule, thus achieving a Pareto improvement for the interests of both parties.

[0047] The beneficial effects of this embodiment are as follows: by clearly defining the multi-objective payoff functions of leaders and followers, the new energy arbitrage objective of microgrids and the low-carbon on-time production objective of production lines are incorporated into a unified game framework; in particular, the cost of exceeding carbon emission quotas is incorporated into the follower objective function, enabling production lines to proactively adjust their production schedules when facing carbon quota constraints, shifting high-energy-consuming processes to peak photovoltaic power generation periods to achieve low-carbon production; and by guiding both parties to make autonomous decisions through price signals, without exposing underlying order privacy and equipment-level parameters, it has stronger feasibility in commercial practice.

[0048] In one possible implementation, the pre-trained meta-reinforcement learning model is trained through an inner and outer double-layer cyclic differentiation mechanism. The inner loop uses a small amount of real-time data under perturbation scenarios to update local gradients to obtain temporary scheduling parameters; the outer loop collects the performance of all temporary scheduling parameters under the corresponding scenarios and updates the global meta-model parameters based on backpropagation of the multi-objective Pareto reward function.

[0049] The inner and outer loop differentiation mechanism can refer to a training paradigm for meta-reinforcement learning models. The inner loop learns rapid adaptation capabilities for a specific task, while the outer loop learns shared meta-knowledge across tasks. The perturbation scenario refers to anomalies that deviate from historical data distributions, such as sudden changes in external spot market electricity prices, a sharp drop in photovoltaic output due to severe weather, or emergency order insertions in internal production. A small amount of real-time data refers to several sets of observation samples that characterize the perturbation scenario, such as current equipment status, order information, and energy price data. Local gradient updates refer to performing only a few gradient descent calculations for the current task within the inner loop. Temporary scheduling parameters refer to model parameters obtained after adapting to the specific perturbation scenario within the inner loop, which achieve good scheduling performance under the current scenario.

[0050] A multi-objective Pareto reward function can refer to a reward function that comprehensively evaluates the merits of a scheduling strategy. This function considers multiple objective dimensions simultaneously, such as electricity costs, carbon emissions, and order delinquency rates, and follows the Pareto optimality principle, meaning that no objective can be further improved without compromising other objectives. Global meta-model parameters refer to the initial model parameters updated in the outer loop. After training on numerous different tasks, these parameters reside in a parameter space that is sensitive to various tasks, allowing the inner loop to adapt to new tasks with only minor updates.

[0051] The present application will be further described below through a detailed embodiment, which further defines the training mechanism of the meta-reinforcement learning model based on the method described in claim 1.

[0052] On the cloud-based digital twin platform, a training task set containing a large number of heterogeneous disturbance scenarios is first constructed. These disturbance scenarios are generated using the Latin hypercube sampling method and cover various types of external and internal disturbances, such as a sudden drop in photovoltaic output due to thunderstorms, a surge in electricity spot market prices, the insertion of urgent high-priority orders, and sudden failures of critical equipment. Each disturbance scenario constitutes an independent training task.

[0053] For each training task, the system executes an inner loop process. Specifically, the pre-trained global meta-model parameters are used as initial parameters. Utilizing the limited real-time observation data available for the current task, such as the energy storage state of charge at the moment of the disturbance, real-time photovoltaic output, and the current list of pending orders, a few gradient descent steps are performed to obtain a set of temporary scheduling parameters for the current disturbance scenario. The number of gradient update steps in the inner loop is typically set to a small value to maintain the model's rapid adaptability.

[0054] After completing the inner loop of all tasks, the system executes the outer loop. For each task, the scheduling strategy is evaluated in the task environment using the temporary scheduling parameters corresponding to that task, resulting in a series of scheduling results. These scheduling results are comprehensively evaluated based on a multi-objective Pareto reward function, which considers multiple objective dimensions such as electricity cost, total carbon emissions, and order delinquency rate, and calculates a comprehensive reward value according to the importance weight of each objective. Subsequently, the cumulative sum of the comprehensive reward values ​​of all tasks is calculated, and the global meta-model parameters are updated through backpropagation, ensuring that the scheduling performance of new tasks continues to improve after minor updates in the inner loop.

[0055] Through alternating iterations of inner and outer loops, the parameters of the global meta-model continuously converge towards a direction that is highly adaptable to various perturbation scenarios. The resulting pre-trained meta-reinforcement learning model, after training in the cloud, is distributed to the edge computing gateway for subsequent online scheduling decisions.

[0056] The beneficial effects of this embodiment are as follows: through the inner and outer double-layer cyclic differentiation mechanism, the meta-reinforcement learning model can extract shared meta-knowledge from a large number of heterogeneous perturbation scenarios; when facing new perturbation scenarios that have never been seen before after online deployment, only a small amount of gradient updates are needed using a small amount of real-time observation data to quickly generate high-quality scheduling strategies, avoiding the problem of traditional deep reinforcement learning requiring a large amount of sampling and long-term retraining in new scenarios, and significantly improving the system's response speed to unknown perturbations.

[0057] In one possible implementation, the dual-branch hybrid action network includes: a continuous action branch for outputting the energy storage charge / discharge rate; and a discrete action branch for outputting the selection of machine tool processing steps; the energy storage charge / discharge rate and the selection of machine tool processing steps are output in tandem by the meta-reinforcement learning model.

[0058] A dual-branch hybrid action network can refer to a neural network structure that includes two independent branches in parallel at the output layer. One branch handles the continuous-value action space, and the other handles the discrete-value action space. The continuous-value action branch can refer to a network branch whose output layer activation function is a linear function or a hyperbolic tangent function, and whose output value changes continuously within a specific range. The energy storage charge / discharge rate can be a physical quantity that indicates the charging and discharging rate of an energy storage battery. A negative charge rate indicates charging, and a positive discharge rate indicates discharging; the larger the absolute value, the greater the charging and discharging power.

[0059] Discrete action branches can refer to network branches whose output layer is a probability distribution that has undergone flexible maximum normalization. A specific action is selected from the set of discrete actions by sampling or taking the maximum value. Machine tool processing step selection can refer to the action of selecting one or more steps from the currently executable set and assigning them to a specified machine tool; for example, assigning a specific step of an order to an idle machine tool for processing. Synchronous output can refer to the simultaneous generation of action outputs by continuous action branches and discrete action branches at the same decision moment, with the two actions together forming a complete joint scheduling strategy.

[0060] The present application will be further described below through a detailed embodiment, which further defines the structure of the dual-branch hybrid action network based on the method described in claim 1.

[0061] In this embodiment, the actor network of the meta-reinforcement learning model adopts a dual-branch hybrid action network structure. The first few layers of this network are shared encoding layers, used to process the aforementioned generated joint Markov state representations and extract high-dimensional feature representations. After the shared encoding layers, the network splits into two independent branches.

[0062] The continuous action branch handles the control decisions of the energy storage system. The final layer of this branch uses a hyperbolic tangent activation function, mapping the output value to a range from negative to positive, corresponding to the energy storage charge / discharge rate. A negative output value indicates the energy storage battery is charging, with a larger absolute value indicating greater charging power; a positive output value indicates the energy storage battery is discharging, with a larger absolute value indicating greater discharging power. The output of this continuous action branch accurately reflects the specific extent to which the energy storage system participates in peak shaving and valley filling, and smoothing out fluctuations in renewable energy sources.

[0063] The discrete motion branch handles production scheduling decisions. The last layer of this branch is a fully connected layer, with an output dimension equal to the number of currently executable operations and machine tool combinations. The output is transformed into a probability distribution using a flexible maximum normalization function. A specific machining operation is selected from this probability distribution and assigned an action through probability sampling or maximization; for example, assigning operation B of order A to CNC machine tool number C. The output of this discrete motion branch reflects the specific adjustments to the production schedule.

[0064] Continuous and discrete action branches output results simultaneously at the same decision-making moment, forming a complete joint scheduling strategy. This joint scheduling strategy includes both continuous control commands for energy storage devices on the microgrid side and discrete scheduling commands for processing operations on the production line side, achieving coordinated optimization of multiple elements such as source, grid, load, and storage. In subsequent steps, energy storage charge / discharge rate commands are sent to the energy storage battery management system, and machine tool processing operation selection commands are parsed into manufacturing work orders and sent to the workshop programmable logic controller.

[0065] The beneficial effects of this embodiment are as follows: by constructing a dual-branch hybrid action network, a unified modeling of the hybrid action space containing continuous control variables and discrete decision variables is achieved, avoiding the accuracy loss and dimensionality explosion problems caused by discretizing continuous actions; the continuous action branch and the discrete action branch output at the same frequency, ensuring the coordination and consistency of energy storage control and production scheduling in the time dimension, and providing an effective action expression method for source-grid-load-storage coordinated optimization.

[0066] In one possible implementation, low-frequency temporal feature vectors of macroscopic energy data are extracted through a one-dimensional temporal convolutional network, further including: extracting slowly changing features; the topological graph containing machine tool status and order process topological dependencies is transformed into a high-frequency graph topological feature vector through a graph attention network, further including: extracting rapidly changing features; the low-frequency temporal feature vectors and high-frequency graph topological feature vectors are dimensionally aligned and concatenated in the hidden layer space of the neural network, further including: fusing slowly changing features and rapidly changing features through an attention mechanism.

[0067] Slow-changing characteristics refer to features extracted from macroeconomic energy data that change slowly over time, such as electricity price trends, daily cyclical patterns of photovoltaic power output, and long-term evolution trajectories of energy storage state of charge. These characteristics change at a low frequency. Fast-changing characteristics refer to features extracted from microeconomic production data that change rapidly over time, such as instantaneous fluctuations in machine tool power, dynamic changes in process allocation, and the suddenness of order arrivals. These characteristics change at a high frequency.

[0068] Dimension alignment refers to mapping feature vectors from different sources and with different dimensions to the same feature space dimension through linear or nonlinear transformations, making them comparable and composable. Attention mechanisms can refer to a mechanism that simulates human visual attention, focusing attentional resources on features more important to the current decision by calculating the importance weights of different features.

[0069] The present application will be further described below through a detailed embodiment. This embodiment further defines the specific implementation of feature extraction and fusion based on the method described in claim 1.

[0070] In the aforementioned embodiments, the system extracts low-frequency temporal feature vectors from macroscopic energy data using a one-dimensional temporal convolutional network. This process further includes extracting slowly varying features. Specifically, the one-dimensional temporal convolutional network employs a multi-scale dilated convolutional structure, gradually increasing the receptive field in the time dimension, thereby effectively capturing the long-term dependencies in the electricity price sequence and the photovoltaic output prediction sequence. For example, the network can identify the regular fluctuation patterns of electricity prices during daily peak and off-peak hours, as well as the diurnal cycle characteristic of photovoltaic output peaking at noon and approaching zero at night. These features with low frequency of change and strong regularity are extracted as slowly varying feature vectors.

[0071] Simultaneously, the system transforms the topology graph, which contains machine tool states and order process topology dependencies, into a high-frequency graph topology feature vector through a graph attention network. This process further includes extracting rapidly changing features. Specifically, a graph structure is constructed with each machine tool as a node. Node features include dynamic information such as the machine tool's real-time operating power, current processing status, and the length of the queue of processes to be processed. Edges represent the sequential dependencies between processes and the allocatable relationships between processes and machine tools. The graph attention network dynamically captures real-time changes in the process topology by calculating attention weights between nodes, such as dependency reconstruction caused by the insertion of emergency orders or resource constraint changes caused by sudden machine tool failures. These features with high frequency of change and strong transient characteristics are extracted as rapidly changing feature vectors.

[0072] In the hidden space of the neural network, low-frequency temporal feature vectors and high-frequency graph topological feature vectors are dimensionally aligned and concatenated. This process further includes fusing slowly changing and rapidly changing features through an attention mechanism. Specifically, the system first maps the slowly changing and rapidly changing feature vectors to the same dimensional space through a linear transformation layer, achieving dimensional alignment. Subsequently, an attention mechanism is introduced to calculate the importance weights of the slowly changing and rapidly changing features for the current decision task. The weighted feature vectors are then concatenated to form a unified joint Markov state representation. The attention mechanism enables the model to focus more on electricity price and photovoltaic information in the slowly changing features when the external energy environment changes drastically; and to focus more on topological changes in the rapidly changing features when the production system experiences emergency order insertions or equipment failures, thereby achieving adaptive feature fusion.

[0073] The beneficial effects of this embodiment are as follows: by extracting the slowly changing features of macroscopic energy data through a one-dimensional temporal convolutional network and extracting the rapidly changing features of microscopic production data through a graph attention network, efficient representation of heterogeneous data from different time scales and sources is achieved; by adaptively fusing the slowly changing and rapidly changing features in the hidden layer space through the attention mechanism, the system can dynamically adjust the degree of attention to different features according to the needs of the current decision-making task, providing a more accurate and comprehensive state representation for subsequent game optimization and strategy generation.

[0074] One possible implementation also includes: when an unknown disturbance such as a sudden change in the external spot market electricity price, a sharp drop in photovoltaic output, or an internal emergency order insertion is detected, the real-time observation status data at the current moment is captured, local gradient updates are performed based on the pre-trained meta-model parameters, and a locally optimal joint scheduling strategy for the current disturbance scenario is generated.

[0075] Among these, sudden changes in external spot market electricity prices refer to unexpected changes in real-time electricity prices published by the main grid, where these prices deviate significantly from the day-ahead forecasts. A sharp drop in photovoltaic output refers to a rapid decrease in the output power of photovoltaic power generation systems within a short period due to weather conditions such as thunderstorms or rapidly moving clouds. Internal emergency order insertion refers to situations where high-priority orders are received temporarily during production, requiring immediate insertion into the existing production schedule for processing.

[0076] Unknown disturbances can refer to new scenarios that deviate from historical distributions and have not appeared in the training data. Real-time observed state data refers to the latest system state information collected from physical devices at the moment the disturbance occurs, including operating parameters of each device, order information, energy data, etc. Local gradient update refers to performing a few steps of gradient descent calculations on the pre-trained meta-model parameters using only a small amount of data from the current disturbance scenario to adapt the model parameters to the current scenario. Locally optimal joint scheduling strategy refers to the scheduling strategy output by the model after local gradient update, which can achieve better overall benefits under the current disturbance scenario.

[0077] The present application will be further described below through a detailed embodiment, which further defines the adaptive response mechanism of the system to unknown disturbances based on the method described in claim 1.

[0078] Based on the aforementioned embodiments, suppose that during operation on a certain workday, the system detects an anomaly. Specifically, around 2 PM, meteorological monitoring data indicates that extreme severe convective weather is approaching the industrial park. The microgrid energy management system detects that the real-time output of the photovoltaic system drops sharply from 80% of its rated power to below 5% within five minutes. Simultaneously, external electricity spot market price information shows that the price has surged to three times its previous level. The system determines that the current state deviates significantly from the distribution of historical training data, triggering an adaptive response process for unknown disturbances.

[0079] The system immediately captures real-time observation data, including the remaining state of charge of the energy storage battery, the current operating power and processing status of each CNC machine tool, the list of orders to be processed and their process dependencies, and the current external electricity price and carbon emission factor. This data constitutes a small set of real-time observation samples.

[0080] The meta-reinforcement learning model deployed in the edge computing gateway initiates an inner-loop local gradient update mechanism based on pre-trained meta-model parameters. Specifically, the model uses the currently captured real-time observation data as training samples, calculates the loss function of the current policy under perturbation scenarios, and performs one or two gradient descent calculations to locally fine-tune the model parameters. Since the meta-model parameters have been pre-trained on a large number of heterogeneous perturbation scenarios and are in a globally optimal basin that is relatively sensitive to various scenarios, only a few gradient update steps are needed for the model to adapt to new perturbation scenarios.

[0081] After local gradient updates, the model outputs a locally optimal joint scheduling strategy for the current scenario of sudden photovoltaic power drop and soaring electricity prices. In this strategy, the continuous action branch commands the energy storage battery to discharge at its maximum safe rate to compensate for the photovoltaic power output shortfall and avoid purchasing electricity at high prices; the discrete action branch commands the manufacturing execution system to suspend the machine tools with the highest energy consumption but less urgent delivery dates, prioritizing the operation of low-energy-consuming processes for short-delivery orders. This joint scheduling strategy completes calculations and is deployed to physical equipment for execution within milliseconds, effectively mitigating the operational risks caused by sudden disturbances.

[0082] The beneficial effects of this embodiment are as follows: by using the local gradient update mechanism of the meta-reinforcement learning model, the system can generate an effective scheduling strategy within milliseconds when faced with new disturbances that deviate from the historical distribution, without retraining or large-scale sampling, but only by using a small amount of real-time observation data to perform a very small number of gradient updates. Compared with the traditional deep reinforcement learning method that requires a large number of trial and error samplings and long-term retraining in new scenarios, this method significantly improves the system's real-time response capability to unknown disturbances and ensures the operational reliability of industrial production lines and microgrids in complex environments.

[0083] In one possible implementation, the continuous energy storage actions in the joint scheduling strategy are parsed into active power adjustment commands and sent to the energy storage battery management system and the photovoltaic inverter, further including: sending commands via the Modbus TCP protocol; the discrete production scheduling actions are parsed into manufacturing work orders and sent to the shop floor programmable logic controller, further including: sending commands via the OPCUA protocol.

[0084] Among these, Modbus TCP protocol refers to an Ethernet-based industrial communication protocol widely used in industrial automation for data exchange between controllers and equipment. Active power regulation commands refer to digital instructions used to control the output or absorption of active power by energy storage battery management systems and photovoltaic inverters. OPC UA protocol refers to a secure and reliable data communication protocol for industrial automation, used for interoperability between equipment and systems from different manufacturers. Manufacturing work orders refer to production instruction files or data structures containing information such as product specifications, quantity, processing steps, process parameters, and delivery time. Shop floor programmable logic controllers (PLCs) refer to industrial controllers used to control the operation of shop floor equipment, receiving instructions from upper-level systems and converting them into executable electrical signals for the equipment.

[0085] The present application will be further described below through a detailed embodiment. This embodiment further defines the specific communication protocol for issuing instructions based on the method described in claim 1.

[0086] Based on the aforementioned embodiments, after the system generates a joint scheduling strategy, it needs to parse the actions in the strategy into instructions that can be executed by physical devices and send them out through a specific communication protocol.

[0087] For continuous energy storage operations in the joint dispatch strategy, the system parses the operation value into an active power adjustment command. Specifically, the energy storage charge / discharge rate is multiplied by the rated capacity of the energy storage battery to obtain the target charge / discharge power value. The system then converts this power value into a register write command conforming to the Modbus TCP protocol format. The command is sent to the energy storage battery management system via the network, and the management system controls the energy storage battery pack to charge or discharge at the specified power according to the command. Simultaneously, for the photovoltaic inverter, the system sends corresponding power adjustment commands via the Modbus TCP protocol to adjust the inverter's active power output to coordinate with the operation of the energy storage system.

[0088] For discrete actions in the joint scheduling strategy, the system parses these actions into manufacturing work orders. Specifically, based on the machine tool processing operation selected by the discrete action branch, the system generates a manufacturing work order containing information such as the order number, product type, processing parameters, target machine tool number, planned start time, and planned end time for that operation. This manufacturing work order is sent to the shop floor programmable logic controller (PLC) via the OPC UA protocol. The OPC UA protocol provides secure session establishment, data encryption, and access control mechanisms to ensure reliable transmission of production instructions. After receiving the work order, the PLC parses its contents and controls the corresponding CNC machine tool to execute the appropriate processing operations.

[0089] After the physical equipment executes the command, the actual power output value, actual production progress and other feedback data are uploaded to the system again through the above communication protocol to form a complete control closed loop.

[0090] The beneficial effects of this embodiment are as follows: by adopting the Modbus TCP protocol and the OPC UA protocol, two communication standards widely used in the industrial field, reliable interconnection between the upper-layer scheduling algorithm and the lower-layer physical devices is achieved; the Modbus TCP protocol ensures the real-time and deterministic nature of the power adjustment commands of energy storage devices and photovoltaic inverters, and the OPC UA protocol ensures the security and interoperability of manufacturing work order issuance, enabling the scheduling method of this application to be easily deployed in existing industrial systems and possessing good engineering applicability.

[0091] In one possible implementation, the system corresponding to this solution is as follows: Figure 2 As shown, the system principle block diagram is divided into the physical device layer, the edge control layer, and the cloud decision layer from bottom to top.

[0092] The physical equipment layer includes CNC machine tools, conveyor belts, and robotic arms (i.e., "loads") on the industrial production line side, and photovoltaic inverters (i.e., "sources"), energy storage battery packs (i.e., "storage"), and grid-connected gateways (i.e., "grids") on the microgrid side. The physical equipment layer is responsible for executing discrete and continuous commands such as start / stop and charge / discharge issued from the upper layer.

[0093] Edge control layer: This layer comprises the Manufacturing Execution System (MES) server and the Microgrid Energy Management System (MG-EMS) controller. The two are physically isolated and interact only with price signals and projected load data via standard encrypted protocols.

[0094] Cloud-based decision layer: Deploys cross-domain game theory decision-making and meta-reinforcement learning (MAML) models. It is responsible for receiving historical time-series data uploaded from the edge layer, performing offline pre-training of the model and extraction of meta-knowledge, and then distributing the lightweight model to the edge layer.

[0095] In one overall embodiment, to further illustrate the feasibility and technical details of this application, the following is a specific scheduling implementation example of an industrial park facing sudden external disturbances.

[0096] Implementation Scenario: A high-energy-consuming industrial park includes a microgrid system (equipped with a 10MW photovoltaic array and a 5MWh energy storage battery pack) and a flexible manufacturing workshop (containing multiple large machining centers and high-energy-consuming equipment such as polysilicon reduction furnaces). In terms of hardware topology, the microgrid is controlled by the MG-EMS (Microgrid Management System), and the workshop is controlled by the MES (Manufacturing Execution System), with the two physically isolated. The edge computing gateway deploys the MAML meta-reinforcement learning model described in this application and connects to the underlying PLC via the OPCUA protocol, and connects to the energy storage battery management system (BMS) and photovoltaic inverters via ModbusTCP.

[0097] Implementation steps and effects: Cloud-based pre-training phase: Before the system goes live, hundreds or thousands of heterogeneous historical disturbance scenarios, including thunderstorms, soaring spot electricity prices, and emergency order insertions, are constructed on the cloud-based digital twin platform using Latin hypercube sampling. Meta-knowledge is extracted through the inner and outer double-layer gradient loop mechanism of MAML to obtain a set of highly sensitive initial neural network parameters in the globally optimal basin, and then distributed to the edge computing gateway in the workshop.

[0098] Daily Game Scheduling Phase: On regular workdays, the MG-EMS (leader) releases an internal interactive electricity price sequence and carbon emission factor based on the day-ahead electricity price and photovoltaic forecast. Upon receiving this signal, the MES (follower), while adhering to the constraint of process sequence dependence, shifts the production schedule of energy-intensive machining centers to the midday period when photovoltaic power generation is high and internal electricity prices are lowest. Both parties interact through a price elasticity mechanism, approaching the Nash equilibrium point without exchanging core order data. Adaptive Phase to Sudden Disturbance (Core Effect): Suppose that at 14:00 in the afternoon, a sudden extreme convective weather event causes the photovoltaic output to drop sharply to zero, while the spot market electricity price on the main grid soars.

[0099] If a traditional offline centralized solver (such as MILP) is used, recalculating the production schedule and energy storage scheduling of the entire workshop would take tens of minutes, resulting in serious control lag.

[0100] However, in the system of this application, after the edge computing gateway detects an unknown disturbance, it immediately captures a very small amount of real observation data (Few-shot data). Based on the pre-trained meta-parameters, the model only needs to perform 1 to 2 local gradient updates within microseconds to milliseconds to quickly "mutate" and generate a locally optimal joint scheduling strategy network for the current sudden rainstorm situation.

[0101] Action issuance and closed-loop: The updated dual-branch hybrid action network outputs commands in sync: the continuous action branch commands the energy storage battery to discharge at the maximum safe rate to mitigate electricity price fluctuations; the discrete action branch commands the MES to immediately suspend and put into hibernation the machine tools with the highest energy consumption but less urgent delivery dates, prioritizing the operation of low-energy-consuming processes for short-delivery orders. Ultimately, guided by a multi-objective Pareto reward mechanism, the system resolved the crisis within milliseconds, perfectly balancing minimum late payment penalties, minimum electricity purchase costs, and carbon emission compliance.

[0102] In another overall embodiment, the solution is further illustrated by taking millisecond-level adaptive scheduling for emergency order insertion in the workshop (production-side disturbance) as an example.

[0103] Scenario: A high-end equipment manufacturing plant is operating smoothly according to its established production schedule (optimized based on periods of low electricity prices). At 10:00 AM, the sales department suddenly places a high-priority, urgent, and high-profit order, requiring delivery within 24 hours. At this time, the plant's photovoltaic output is at a moderate level, and it is also during peak electricity consumption periods on the external power grid (when electricity prices are high).

[0104] Process: 1. Feature Realignment: The Graph Attention Network (GAT) at the MES instantly captures changes in the order topology (addition of high-priority nodes) and re-integrates them with the current macro-level high electricity price characteristics. 2. Rapid MAML Fine-tuning: Faced with the new state of "high electricity price + urgent orders," the dual-branch MAML model deployed on the edge computing gateway starts its inner loop. Leveraging historically accumulated meta-knowledge, it updates parameters and generates a new scheduling strategy within milliseconds using only the current state data. 3. Game Theory and Action Execution: The discrete action branch of the MES (follower) outputs a decision: suspending two production lines originally scheduled to execute low-priority ordinary orders, forcing the freeing up high-energy-consuming machine tools to prioritize urgent orders; the MG-EMS (leader) detects the intention of a sudden increase in production load, and its continuous action branch immediately commands the energy storage system (BMS) to discharge at full capacity to "shaving and filling valleys."

[0105] Implementation results: Under the premise of not violating the maximum capacity limit of the main power grid and effectively controlling the peak demand electricity price, the system ensured the timely delivery of high-profit emergency orders, demonstrating the model's rapid response capability to sudden disturbances on the production side.

[0106] In another overall embodiment, the scheme is further illustrated by taking the dynamic trade-off between "carbon-energy-production" multi-objectives when the carbon emission quota approaches the critical value.

[0107] Implementation scenario: A heavy industrial park faces severe carbon emission assessment pressure at the end of the quarter, with its remaining free carbon emission allowances nearing the red line. Exceeding the allowance would require purchasing carbon emission reduction credits (CCERs) at a high price in the carbon trading market, resulting in extremely high additional costs.

[0108] The process is as follows: 1. Leader Dynamic Pressure: The microgrid MG-EMS (leader) senses insufficient remaining carbon allowances and automatically and significantly increases the "carbon emission factor penalty weight" in its internal interaction signals. 2. Follower Multi-Objective Optimization: The MES (follower) receives a high carbon emission cost signal. At this time, the system's Pareto multi-objective reward function tilts, with the weight of carbon emission reduction exceeding the reward for early order delivery. 3. Low-Carbon Production Restructuring: The algorithm controls the MES to proactively adjust processes, forcibly shifting large-scale thermal processing processes originally scheduled for nighttime (mainly relying on grid thermal power, with high carbon emission factors) to the following day's noon photovoltaic peak generation period (green electricity, with carbon emission factors approaching zero). Even if this results in some non-core orders being completed on the last day of the delivery deadline.

[0109] Implementation Results: This perfectly validates the closed-loop logic of "carbon-energy-production" synergy in this patent. Guided by internal game signals, the production line can proactively adapt to carbon quota constraints, achieving passive low-carbon production at zero cost while avoiding penalties for non-compliance.

[0110] In yet another overall embodiment, this scheme is further illustrated by taking flexible load deep arbitrage against extreme spot electricity price fluctuations as an example.

[0111] Implementation scenario: A chemical plant has multiple electrolyzers that can be started and stopped (a typical high-energy-consuming flexible load). The following day, affected by extreme weather, the day-ahead market price published by the external electricity spot market fluctuated wildly, with off-peak electricity prices extremely low and peak electricity prices soaring to three times the usual level.

[0112] The process is as follows: 1. Proactive Feature Extraction: A one-dimensional temporal convolutional network (1D-TCN) extracts and processes the time-series features of extreme electricity prices for the next day in advance. 2. Pre-game Theory and Equilibrium: Before production begins the next day, the cloud model has already deduced and found a new Stackelberg Nash equilibrium point. MG-EMS sets a highly attractive "valley-hour internal electricity price". 3. Collaborative Arbitrage Execution: MES concentrates the full-load production period of the electrolyzers during the valley hours, while maintaining only the minimum standby power for insulation during the peak hours; at the same time, the microgrid purchases a large amount of electricity from the grid during the valley hours to fill the energy storage, and during the peak hours, it not only meets the small load within the plant, but also sells the surplus energy from the energy storage back to the main grid.

[0113] Implementation Results: The system deeply integrates the production process with the electricity market. Instead of increasing production costs due to extremely high electricity prices, the system leverages the flexibility of industrial loads and the support of energy storage to achieve negative-cost (arbitrage profit) operation in the electricity spot market.

[0114] This application creatively integrates multi-agent master-slave game theory with a cross-timescale meta-reinforcement learning architecture, fundamentally overcoming many technical barriers faced by existing technologies when implemented in industrial settings, achieving the following significant technical effects: First, it completely breaks down information silos within the factory area, achieving extremely high data privacy protection. Traditional centralized scheduling schemes require the manufacturing system to completely expose highly confidential order data and process parameters to the microgrid, which poses a significant risk of privacy leakage and management resistance in commercial practice. This application introduces the Stackelberg master-slave game model, utilizing price elasticity mechanisms for information interaction. The leader (microgrid) only issues dynamic internal electricity and carbon prices, while the follower (MES) only feeds back the overall power demand curve, completely avoiding the reverse uploading of underlying process data, achieving data "usable but invisible," and significantly improving the feasibility of the solution for engineering implementation among different stakeholders. Second, it solves the "curse of dimensionality" caused by cross-timescale misalignment and endows the system with millisecond-level disturbance resistance. To address the time mismatch between hourly fluctuations in large power grids, minute-level responses in microgrids, and second-level execution in production lines, this application utilizes a one-dimensional temporal convolutional network and a graph attention network to successfully align macroscopic environmental variables with the microscopic, rapidly changing equipment topology in the hidden layer space, constructing a panoramic joint Markov state. More importantly, when faced with new disturbances deviating from historical distributions, such as sudden order insertions or severe weather, the MAML meta-learning framework employed in this application avoids the "catastrophic forgetting" and lengthy retraining process of traditional deep reinforcement learning (DRL). The algorithm only needs to extract a small amount of real-time data to complete local gradient iteration within microseconds to milliseconds, outputting high-quality hybrid scheduling instructions, enabling the entire industrial production line and energy system to exhibit unprecedented agility and robustness.

[0115] Corresponding to the above method embodiments, this specification also provides an embodiment of a dynamic multi-game optimization scheduling device for industrial production lines that integrates source-grid-load-storage coordination. Figure 3 This specification illustrates a schematic diagram of a dynamic multi-game optimization scheduling device for industrial production lines, based on an embodiment of this specification, that integrates source-grid-load-storage coordination. (See attached diagram.) Figure 3 As shown, the device includes: The data acquisition module 301 is configured to collect micro-production data from industrial production lines and macro-energy data from microgrids. The micro-production data includes real-time power consumption of machine tools and topology dependencies of order processes, while the macro-energy data includes the state of charge of energy storage batteries, photovoltaic power output prediction, time-of-use electricity price of the grid, and carbon emission factors. The feature determination module 302 is configured to extract low-frequency temporal feature vectors of macroscopic energy data through a one-dimensional temporal convolutional network, and transform the topological graph containing machine tool status and order process topological dependencies into high-frequency graph topological feature vectors through a graph attention network. In the hidden layer space of the neural network, the low-frequency temporal feature vectors and high-frequency graph topological feature vectors are dimensionally aligned and spliced ​​to generate a unified joint Markov state representation. The environment construction module 303 is configured to construct an interactive environment based on the Stackelberg master-slave game, with the microgrid energy management system as the leader and the manufacturing execution system as the follower. The joint Markov state representation is input into the pre-trained meta-reinforcement learning model. The leader publishes the internal interactive electricity price, and the follower adjusts the production schedule according to the internal interactive electricity price and feeds back the expected load, so as to approach the Nash equilibrium point through multiple interactive iterations. The strategy output module 304 is configured to output a joint scheduling strategy containing continuous energy storage actions and discrete production scheduling actions through a dual-branch hybrid action network of a meta-reinforcement learning model. The continuous energy storage actions include the energy storage charge and discharge rate, and the discrete production scheduling actions include the selection of machine tool processing steps. The instruction issuing module 305 is configured to parse the continuous energy storage actions in the joint scheduling strategy into active power adjustment instructions and issue them to the energy storage battery management system and photovoltaic inverter, parse the discrete production scheduling actions into manufacturing work orders and issue them to the workshop programmable logic controller, and obtain the feedback status of the physical equipment after execution to form dynamic closed-loop control.

[0116] In one possible implementation, an interactive environment based on a Stackelberg master-slave game is constructed, which further includes: the leader aims to maximize its own net income, which includes revenue from selling electricity to the main grid, revenue from photovoltaic power generation, energy storage losses, and the cost of purchasing electricity from the main grid; the follower aims to minimize the overall production cost, which includes electricity costs, carbon emission quota exceeding costs, and penalties for order overdue defaults.

[0117] In one possible implementation, the pre-trained meta-reinforcement learning model is trained through an inner and outer double-layer cyclic differentiation mechanism. The inner loop uses a small amount of real-time data under perturbation scenarios to update local gradients to obtain temporary scheduling parameters; the outer loop collects the performance of all temporary scheduling parameters under the corresponding scenarios and updates the global meta-model parameters based on backpropagation of the multi-objective Pareto reward function.

[0118] In one possible implementation, the dual-branch hybrid action network includes: a continuous action branch for outputting the energy storage charge / discharge rate; and a discrete action branch for outputting the selection of machine tool processing steps; the energy storage charge / discharge rate and the selection of machine tool processing steps are output in tandem by the meta-reinforcement learning model.

[0119] In one possible implementation, low-frequency temporal feature vectors of macroscopic energy data are extracted through a one-dimensional temporal convolutional network, further including: extracting slowly changing features; the topological graph containing machine tool status and order process topological dependencies is transformed into a high-frequency graph topological feature vector through a graph attention network, further including: extracting rapidly changing features; the low-frequency temporal feature vectors and high-frequency graph topological feature vectors are dimensionally aligned and concatenated in the hidden layer space of the neural network, further including: fusing slowly changing features and rapidly changing features through an attention mechanism.

[0120] One possible implementation also includes: when an unknown disturbance such as a sudden change in the external spot market electricity price, a sharp drop in photovoltaic output, or an internal emergency order insertion is detected, the real-time observation status data at the current moment is captured, local gradient updates are performed based on the pre-trained meta-model parameters, and a locally optimal joint scheduling strategy for the current disturbance scenario is generated.

[0121] In one possible implementation, the continuous energy storage actions in the joint scheduling strategy are parsed into active power adjustment commands and sent to the energy storage battery management system and the photovoltaic inverter, further including: sending commands via the Modbus TCP protocol; the discrete production scheduling actions are parsed into manufacturing work orders and sent to the shop floor programmable logic controller, further including: sending commands via the OPCUA protocol.

[0122] The above is a schematic scheme of a source-grid-load-storage coordinated dynamic multi-game optimization scheduling device for industrial production lines according to this embodiment. It should be noted that the technical solution of this source-grid-load-storage coordinated dynamic multi-game optimization scheduling device for industrial production lines belongs to the same concept as the technical solution of the source-grid-load-storage coordinated dynamic multi-game optimization scheduling method for industrial production lines described above. Details not described in detail in the technical solution of the source-grid-load-storage coordinated dynamic multi-game optimization scheduling device for industrial production lines can be found in the description of the technical solution of the source-grid-load-storage coordinated dynamic multi-game optimization scheduling method for industrial production lines described above.

[0123] Figure 4 A structural block diagram of a computing device 400 according to one embodiment of this specification is shown. The components of the computing device 400 include, but are not limited to, a memory 410 and a processor 420. The processor 420 is connected to the memory 410 via a bus 430, and a database 450 is used to store data.

[0124] The computing device 400 also includes an access device 440, which enables the computing device 400 to communicate via one or more networks 460. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 440 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0125] In one embodiment of this specification, the aforementioned components of the computing device 400 and Figure 4 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 4 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0126] The computing device 400 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 400 can also be a mobile or stationary server.

[0127] The processor 420 executes the following computer-executable instructions, which, when executed by the processor, implement the steps of the aforementioned source-grid-load-storage collaborative dynamic multi-game optimization scheduling method for industrial production lines. The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the aforementioned source-grid-load-storage collaborative dynamic multi-game optimization scheduling method for industrial production lines belong to the same concept. Details not described in detail in the technical solution of the computing device can be found in the description of the technical solution of the aforementioned source-grid-load-storage collaborative dynamic multi-game optimization scheduling method for industrial production lines.

[0128] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described source-grid-load-storage coordinated dynamic multi-game optimization scheduling method for industrial production lines.

[0129] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the above-mentioned source-grid-load-storage collaborative industrial production line dynamic multi-game optimization scheduling method. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the above-mentioned source-grid-load-storage collaborative industrial production line dynamic multi-game optimization scheduling method.

[0130] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described source-grid-load-storage coordinated dynamic multi-game optimization scheduling method for industrial production lines.

[0131] The above is an illustrative scheme of a computer program according to this embodiment. It should be noted that the technical solution of this computer program belongs to the same concept as the technical solution of the above-mentioned source-grid-load-storage coordinated industrial production line dynamic multi-game optimization scheduling method. For details not described in detail in the technical solution of the computer program, please refer to the description of the technical solution of the above-mentioned source-grid-load-storage coordinated industrial production line dynamic multi-game optimization scheduling method.

[0132] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0133] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0134] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0135] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0136] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A dynamic multi-game optimization scheduling method for industrial production lines with source-grid-load-storage coordination, characterized in that, include: Collect micro-production data from industrial production lines and macro-energy data from microgrids. The micro-production data includes real-time power consumption of machine tools and topology dependencies of order processes, while the macro-energy data includes the state of charge of energy storage batteries, photovoltaic power output forecasts, time-of-use electricity prices, and carbon emission factors. Low-frequency temporal feature vectors of the macroscopic energy data are extracted by a one-dimensional temporal convolutional network, and the topological graph containing the machine tool state and the topological dependence of the order process is transformed into a high-frequency graph topological feature vector by a graph attention network. The low-frequency temporal feature vectors and the high-frequency graph topological feature vectors are dimensionally aligned and concatenated in the hidden layer space of the neural network to generate a unified joint Markov state representation. An interactive environment based on Stackelberg master-slave game is constructed, with the microgrid energy management system as the leader and the manufacturing execution system as the follower. The joint Markov state representation is input into a pre-trained meta-reinforcement learning model. The leader publishes an internal interactive electricity price, and the follower adjusts its production schedule and feeds back the expected load based on the internal interactive electricity price, so as to approach the Nash equilibrium point through multiple interactive iterations. Through the dual-branch hybrid action network of the meta-reinforcement learning model, a joint scheduling strategy including continuous energy storage actions and discrete production scheduling actions is output at the same frequency. The continuous energy storage actions include the energy storage charge and discharge rate, and the discrete production scheduling actions include the selection of machine tool processing steps. The continuous energy storage actions in the joint scheduling strategy are parsed into active power adjustment commands and sent to the energy storage battery management system and photovoltaic inverter. The discrete production scheduling actions are parsed into manufacturing work orders and sent to the workshop programmable logic controller. The feedback status after the physical equipment is executed is obtained to form a dynamic closed-loop control.

2. The method according to claim 1, characterized in that, The construction of the interactive environment based on the Stackelberg master-slave game further includes: the leader aims to maximize its own net income, which includes revenue from selling electricity to the main grid, revenue from photovoltaic power generation, energy storage losses, and the cost of purchasing electricity from the main grid; the follower aims to minimize the overall production cost, which includes electricity costs, costs of exceeding carbon emission quotas, and penalties for order overdue defaults.

3. The method according to claim 1, characterized in that, The pre-trained meta-reinforcement learning model is trained through an inner and outer double-layer cyclic differentiation mechanism. The inner loop uses a small amount of real-time data under perturbation scenarios to update local gradients to obtain temporary scheduling parameters; the outer loop collects the performance of all the temporary scheduling parameters under the corresponding scenarios and updates the global meta-model parameters based on backpropagation of the multi-objective Pareto reward function.

4. The method according to claim 1, characterized in that, The dual-branch hybrid action network includes: a continuous action branch for outputting the energy storage charge / discharge rate; and a discrete action branch for outputting the selection of the machine tool processing procedure; the energy storage charge / discharge rate and the selection of the machine tool processing procedure are output synchronously by the meta-reinforcement learning model.

5. The method according to claim 1, characterized in that, The step of extracting low-frequency temporal feature vectors from the macroscopic energy data using a one-dimensional temporal convolutional network further includes: extracting slowly changing features; the step of converting a topological graph containing the machine tool state and the topological dependency of the order process into a high-frequency graph topological feature vector using a graph attention network further includes: extracting quickly changing features; the step of dimensionally aligning and concatenating the low-frequency temporal feature vectors and the high-frequency graph topological feature vectors in the hidden layer space of the neural network further includes: fusing the slowly changing features and the quickly changing features using an attention mechanism.

6. The method according to claim 1, characterized in that, Also includes: When an unknown disturbance is detected, such as a sudden change in the external spot market electricity price, a sharp drop in photovoltaic output, or an internal emergency order insertion, the real-time observation status data at the current moment is captured, and local gradient updates are performed based on the pre-trained meta-model parameters to generate a locally optimal joint scheduling strategy for the current disturbance scenario.

7. The method according to claim 1, characterized in that, The step of parsing the continuous energy storage actions in the joint scheduling strategy into active power adjustment commands and sending them to the energy storage battery management system and photovoltaic inverters further includes: sending commands via the Modbus TCP protocol; the step of parsing the discrete production scheduling actions into manufacturing work orders and sending them to the shop floor programmable logic controller further includes: sending commands via the OPC UA protocol.

8. A dynamic multi-game optimization scheduling device for industrial production lines with source-grid-load-storage coordination, characterized in that, include: The data acquisition module is configured to collect micro-production data from industrial production lines and macro-energy data from microgrids. The micro-production data includes real-time power consumption of machine tools and order process topology dependencies, while the macro-energy data includes the state of charge of energy storage batteries, photovoltaic power output prediction, grid time-of-use electricity price, and carbon emission factor. The feature determination module is configured to extract low-frequency temporal feature vectors of the macroscopic energy data through a one-dimensional temporal convolutional network, and to transform the topological graph containing the machine tool state and the topological dependency of the order process into a high-frequency graph topological feature vector through a graph attention network. In the hidden layer space of the neural network, the low-frequency temporal feature vectors and the high-frequency graph topological feature vectors are dimensionally aligned and concatenated to generate a unified joint Markov state representation. The environment construction module is configured to build an interactive environment based on Stackelberg master-slave game, with the microgrid energy management system as the leader and the manufacturing execution system as the follower. The joint Markov state representation is input into a pre-trained meta-reinforcement learning model. The leader publishes an internal interactive electricity price, and the follower adjusts its production schedule and feeds back the expected load according to the internal interactive electricity price, so as to approach the Nash equilibrium point through multiple interactive iterations. The strategy output module is configured to output a joint scheduling strategy that includes continuous energy storage actions and discrete production scheduling actions through the dual-branch hybrid action network of the meta-reinforcement learning model. The continuous energy storage actions include the energy storage charge and discharge rate, and the discrete production scheduling actions include the selection of machine tool processing steps. The instruction issuing module is configured to parse the continuous energy storage actions in the joint scheduling strategy into active power adjustment instructions and issue them to the energy storage battery management system and photovoltaic inverter, parse the discrete production scheduling actions into manufacturing work orders and issue them to the workshop programmable logic controller, and obtain the feedback status after the physical equipment is executed to form dynamic closed-loop control.

9. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the source-grid-load-storage coordinated dynamic multi-game optimization scheduling method for industrial production lines as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the source-grid-load-storage coordinated dynamic multi-game optimization scheduling method for industrial production lines according to any one of claims 1 to 7.