Carbon collaborative management method and system based on ai technology

Abstract

The application discloses an energy-carbon collaborative management method and system based on AI technology, relates to the technical field of computer information processing and energy management, and comprises the following steps: constructing an energy-carbon coupling heterogeneous graph according to multi-source heterogeneous data of an enterprise production system; obtaining the coupling sensitivity of each structural node in the energy-carbon coupling heterogeneous graph according to a time sequence graph attention network, and generating a global situation awareness report; introducing a strategy training gradient to obtain the decision action of an agent and form a global decision action set; performing decision execution on the agent based on the global decision action set, updating the energy-carbon coupling heterogeneous graph and the agent strategy network according to execution feedback data, and constructing a traceable decision ledger based on the energy-carbon coupling heterogeneous graph, the global situation awareness report, the global decision action set and the execution feedback data. The application is used to solve the problem that the existing technology cannot reflect the nonlinear energy-carbon coupling relationship between devices, and the collaborative optimization relationship between energy consumption and carbon emission is not fully considered.

Description

Technical Field

[0001] This invention relates to the fields of computer information processing and energy management technology, and more specifically, to an energy and carbon co-management method and system based on AI technology. Background Technology

[0002] Accurate carbon emission accounting and dynamic optimization have become core requirements for the green transformation of industrial enterprises. Enterprise energy and carbon management involves multiple aspects such as energy consumption monitoring, carbon emission accounting, and optimization and control decision-making, and is a typical complex system engineering project.

[0003] For example, Chinese invention patent application CN122027699A discloses a method, system, and storage medium for enterprise energy and carbon synergistic control. This method is applied to an energy and carbon management system consisting of a sensing end, an analysis end, an execution end, and a communication module. Specifically, the sensing end uses a dual-channel acquisition circuit with a coprime folding modulus to acquire and compress signals from electrical equipment with high dynamic range, and determines the timing and priority of data transmission based on state timeliness indicators. The communication module, based on the priority, uses power domain superposition technology to merge and transmit multiple types of data on the same physical channel. The analysis end uses serial interference cancellation technology to demodulate and restore the signal, performs white-box analysis using deterministic mathematical tool sequences, and generates equipment saturation state determination and standardized adjustment incremental instructions. The execution end parses and executes the instructions to achieve amplitude limiting isolation for saturated equipment and coordinated power allocation for unsaturated equipment.

[0004] However, the above method has the following shortcomings: (1) Using deterministic mathematical tools and fixed emission factors for carbon flow accounting cannot reflect the complex nonlinear energy-carbon coupling relationship between equipment, nor can it dynamically adjust with equipment aging, process changes and seasonal changes, resulting in limited accuracy of carbon emission accounting; (2) Carbon flow accounting to strategy matching is a one-way open-loop process, lacking a closed-loop mechanism for feedback of execution effect to decision data, resulting in the system being unable to self-correct and continuously evolve, and only focusing on single-objective control of carbon emissions, without fully considering the synergistic optimization relationship between energy consumption and carbon emissions.

[0005] To address the aforementioned shortcomings, there is an urgent need for an energy-carbon synergistic management method that can reflect the complex nonlinear energy-carbon coupling relationship between devices, enable system self-correction and continuous evolution, and fully consider the synergistic optimization relationship between energy consumption and carbon emissions. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an energy and carbon collaborative management method and system based on AI technology, in order to solve the problems that the prior art cannot reflect the complex nonlinear energy-carbon coupling relationship between devices, lacks a closed-loop mechanism for feedback of execution effect to decision data, and does not fully consider the collaborative optimization relationship between energy consumption and carbon emissions.

[0007] To achieve the above objectives, this application provides the following technical solution: The first aspect of this invention provides an energy and carbon collaborative management method based on AI technology, comprising: acquiring multi-source heterogeneous data from an enterprise production system; constructing an energy and carbon coupling heterogeneous graph based on the multi-source heterogeneous data; and introducing a dual-timescale update mechanism to update the energy and carbon coupling heterogeneous graph; constructing a time-series graph attention network containing a multi-dimensional attention mechanism; obtaining the coupling sensitivity of each structural node in the energy and carbon coupling heterogeneous graph based on the time-series graph attention network; and generating a global situational awareness report based on the coupling sensitivity; constructing an intelligent agent based on the enterprise production system; introducing the agent's policy training gradient; conducting collaborative game training on the agent based on the policy training gradient and the global situational awareness report to obtain the agent's corresponding decision actions; and forming a global decision action set based on the decision actions; executing decisions for each intelligent agent in the enterprise production system based on the global decision action set; acquiring the agent's execution feedback data; updating the energy and carbon coupling heterogeneous graph and policy training gradient based on the feedback loop generated by the execution feedback data; and storing the energy and carbon coupling heterogeneous graph, the global situational awareness report, the global decision action set, and the execution feedback data on the blockchain to construct a traceable decision ledger for the enterprise production system.

[0008] In a preferred embodiment, acquiring multi-source heterogeneous data of an enterprise production system and constructing an energy-carbon coupling heterogeneous graph based on the multi-source heterogeneous data includes: the multi-source heterogeneous data including energy data, process data, carbon emission data, topology data, and node data of the enterprise production system; acquiring node coupling relationships and node coupling strengths between system nodes in the enterprise production system based on the node data; generating structural nodes based on the mapping of system nodes in the enterprise production system, wherein the structural nodes include node type labels and real-time status attributes; generating node connection edges between structural nodes based on the node coupling relationships between system nodes, wherein the node connection edges include energy transmission edges, material supply edges, and carbon flow association edges, and forming an edge set based on the node connection edges between all structural nodes; generating edge weight values ​​for the node connection edges between structural nodes based on the node coupling strength between system nodes, constructing a dynamic weight matrix based on the edge weight values ​​of each node connection edge in the edge set; and constructing the energy-carbon coupling heterogeneous graph based on the structural nodes, the edge set, and the dynamic weight matrix.

[0009] In a preferred embodiment, a dual-timescale update mechanism is introduced to update the energy-carbon coupling heterogeneous graph, including: the dual-timescale update mechanism includes a slow timescale mechanism and a fast timescale mechanism; the dynamic weight matrix in the energy-carbon coupling heterogeneous graph is periodically updated according to the slow timescale mechanism; and the real-time state attributes of the structural nodes in the energy-carbon coupling heterogeneous graph are periodically updated according to the fast timescale mechanism.

[0010] In a preferred embodiment, a time-series graph attention network incorporating a multi-dimensional attention mechanism is constructed, comprising: the time-series graph attention network including a spatial attention mechanism, a temporal attention mechanism, and a multi-head attention mechanism; generating a historical time series corresponding to the current moment based on a preset time window; constructing a historical graph sequence based on the energy-carbon coupling heterogeneous graph at each moment in the historical time series; the spatial attention mechanism is used to calculate the attention scores of structural nodes to adjacent structural nodes based on the historical graph sequence, and to obtain the spatial aggregation features of structural nodes at each moment based on the attention scores; forming a spatial-temporal sequence based on the spatial aggregation features of structural nodes at each moment; the temporal attention mechanism is used to calculate the temporal attention weights of each moment in the historical time series to the current moment based on the spatial-temporal sequence; spatiotemporally fusing the temporal attention weights and spatial aggregation features to obtain the spatiotemporal fusion features of structural nodes; generating spatiotemporal processing units based on the energy-carbon coupling heterogeneous graph; the multi-head attention mechanism is used to divide the spatiotemporal processing units into at least one attention head, and to linearly transform the spatiotemporal fusion features of each attention head to obtain the final feature vector of each structural node in the energy-carbon coupling heterogeneous graph.

[0011] In a preferred embodiment, the coupling sensitivity of each structural node in the energy-carbon coupling heterogeneous graph is obtained based on the time-series graph attention network, and a global situational awareness report is generated based on the coupling sensitivity. This includes: the time-series graph attention network further having a preset network output layer; performing carbon emission prediction on the final feature vector of each structural node in the energy-carbon coupling heterogeneous graph based on the network output layer to obtain the global carbon emission prediction; obtaining the energy input of the structural node at the current moment; obtaining the coupling sensitivity of the structural node based on the energy input of the structural node at the current moment and the global carbon emission prediction; and generating a global situational awareness report based on the coupling sensitivity of each structural node in the energy-carbon coupling heterogeneous graph.

[0012] In a preferred embodiment, constructing an intelligent agent based on an enterprise production system includes: constructing an intelligent agent based on energy-consuming equipment in the enterprise production system. Each intelligent agent includes a local observation space, an action space, a policy network, and a local state memory. The local observation space includes a set of local observation information of the physical entity corresponding to the intelligent agent. The action space includes a set of control actions that the intelligent agent can execute. The policy network is used to map the local observation information into control actions. The local state memory is used to record the historical local observation information and historical control actions of the intelligent agent.

[0013] In a preferred embodiment, a policy training gradient for the agent is introduced. The agent is then trained through cooperative game theory based on this policy training gradient and a global situational awareness report. This includes: the policy training gradient comprising a judge network, an experience replay pool, and the agent's policy network; obtaining the enterprise's total carbon emissions and total energy consumption at the current moment; generating a penalty for high carbon emissions based on the total carbon emissions; generating a penalty for high energy consumption based on the total energy consumption; generating a positive incentive based on the coupling sensitivity of the agent's policy network and the corresponding structural nodes; and constructing a global reward function based on the penalty for high carbon emissions, the penalty for high energy consumption, and the positive incentive. The judge network is used to evaluate the cumulative expected reward of the agent's generated regulatory actions based on the global reward function. The network parameters of the agent's policy network are iteratively updated and trained based on the cumulative expected reward of the judge network to obtain the updated agent's policy network. The experience replay pool stores the global situational awareness report, regulatory actions, and the corresponding cumulative expected reward of the regulatory actions.

[0014] In a preferred embodiment, obtaining the decision actions corresponding to the agents and forming a global decision action set based on the decision actions includes: obtaining the coupling sensitivity of the structural nodes corresponding to each agent based on the global situational awareness report; inputting the local observation information of the agents and the coupling sensitivity of the structural nodes corresponding to the agents into the trained policy network in the agents, generating the control actions of the agents based on the policy network; and summarizing and processing the control actions of all agents in the enterprise production system to obtain the global decision action set.

[0015] In a preferred embodiment, the energy-carbon coupling heterogeneous graph and policy training gradient are updated based on the feedback loop generated from the execution feedback data. The energy-carbon coupling heterogeneous graph, global situational awareness report, global decision action set, and execution feedback data are then stored on the blockchain to construct a traceable decision ledger. This includes: converting the control actions in the global decision action set into equipment control instructions; executing corresponding control actions on the execution terminals of the enterprise production system according to the equipment control instructions; obtaining the actual operating data of each agent after the control actions are executed; forming execution feedback data based on the actual operating data; performing differential privacy processing on the execution feedback data of each agent to obtain a first feedback loop and a second feedback loop; updating the energy-carbon coupling heterogeneous graph based on the first feedback loop; updating the experience replay pool of the policy training gradient based on the second feedback loop; obtaining the hash values ​​of the energy-carbon coupling heterogeneous graph, global situational awareness report, global decision action set, and execution feedback data respectively; storing the hash values ​​on the blockchain; and forming a traceable decision ledger based on the hash values ​​stored on the blockchain.

[0016] A second aspect of this invention provides an AI-based energy-carbon synergistic management system, comprising: a coupled heterogeneous graph construction module, used to acquire multi-source heterogeneous data from an enterprise production system, construct an energy-carbon coupled heterogeneous graph based on the multi-source heterogeneous data, and introduce a dual-timescale update mechanism to update the energy-carbon coupled heterogeneous graph; a perception report generation module, used to construct a time-series graph attention network containing a multi-dimensional attention mechanism, obtain the coupling sensitivity of each structural node in the energy-carbon coupled heterogeneous graph based on the time-series graph attention network, and generate a global situational awareness report based on the coupling sensitivity; and a decision-making action generation module, used to construct an intelligent agent based on the enterprise production system and introduce intelligent... The system employs a strategy training gradient module, which trains agents in collaborative game theory based on the strategy training gradient and global situational awareness reports to obtain the corresponding decision actions of the agents. These decision actions form a global decision action set. The decision execution and execution feedback module is used to execute decisions for each agent in the enterprise production system based on the global decision action set. It also acquires the execution feedback data of the agents, updates the energy-carbon coupling heterogeneous graph and strategy training gradient based on the feedback loop generated from the execution feedback data, and stores the energy-carbon coupling heterogeneous graph, global situational awareness reports, global decision action set, and execution feedback data on the blockchain to construct a traceable decision ledger for the enterprise production system.

[0017] The beneficial effects of this invention are: 1. This invention acquires multi-source heterogeneous data from an enterprise's production system, constructs an energy-carbon coupling heterogeneity diagram based on the multi-source heterogeneous data to reflect the complex nonlinear energy-carbon coupling relationship between equipment, and introduces a dual time-scale update mechanism to update the energy-carbon coupling heterogeneity diagram, thereby realizing dynamic adjustment of the energy-carbon coupling heterogeneity diagram as operating conditions such as equipment aging and process changes change, and improving the accuracy of carbon emission accounting.

[0018] 2. This invention obtains the coupling sensitivity of each structural node in the energy-carbon coupling heterogeneous graph based on a time-series graph attention network, and generates a global situational awareness report based on the coupling sensitivity. The global situational awareness report reflects the synergistic optimization relationship between energy consumption and carbon emissions, and introduces the agent's policy training gradient. Based on the policy training gradient and the global situational awareness report, the agent is trained through collaborative game to obtain the agent's decision actions, and a global decision action set is formed based on the decision actions. Finally, based on the global decision action set, the agents in the enterprise production system execute decisions, obtain the agent's execution feedback data, and update the energy-carbon coupling heterogeneous graph and agent policy network based on the feedback loop generated by the execution feedback data, forming a closed-loop mechanism that feeds back the execution effect of decision actions to the decision data, realizing the system's self-correction and continuous evolution. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an AI-based energy and carbon co-management method. Figure 2 This is a schematic diagram of the structure of an AI-based energy and carbon co-management system. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1, please refer to Figure 1 , Figure 1This invention presents an AI-based method for coordinated energy and carbon management, comprising the following steps: acquiring multi-source heterogeneous data from an enterprise production system; constructing an energy-carbon coupling heterogeneous graph based on the multi-source heterogeneous data; and updating the energy-carbon coupling heterogeneous graph using a dual-timescale update mechanism; constructing a time-series graph attention network containing a multi-dimensional attention mechanism; obtaining the coupling sensitivity of each structural node in the energy-carbon coupling heterogeneous graph based on the time-series graph attention network; and generating a global situational awareness report based on the coupling sensitivity; constructing intelligent agents based on the enterprise production system; introducing policy training gradients for the intelligent agents; conducting collaborative game training on the intelligent agents based on the policy training gradients and the global situational awareness report to obtain the decision actions corresponding to the intelligent agents; and forming a global decision action set based on the decision actions; executing decisions for each intelligent agent in the enterprise production system based on the global decision action set; acquiring execution feedback data of the intelligent agents; updating the energy-carbon coupling heterogeneous graph and policy training gradient based on the feedback loop generated by the execution feedback data; and storing the energy-carbon coupling heterogeneous graph, the global situational awareness report, the global decision action set, and the execution feedback data on the blockchain to construct a traceable decision ledger for the enterprise production system.

[0022] S1. Obtain multi-source heterogeneous data from the enterprise's production system, construct an energy-carbon coupling heterogeneity diagram based on the multi-source heterogeneous data, and introduce a dual-timescale update mechanism to update the energy-carbon coupling heterogeneity diagram, including: S11. Obtain multi-source heterogeneous data from the enterprise's production system and construct an energy-carbon coupling heterogeneous diagram based on the multi-source heterogeneous data. By deploying an IoT sensing layer on-site at the enterprise's production system, multi-source heterogeneous data from the enterprise's production system can be collected in real time and synchronously. The IoT sensing layer includes, but is not limited to, sensing facilities such as smart meters, flow meters, temperature sensors, continuous emission monitoring systems for flue gas, and PLC controllers.

[0023] S111. Multi-source heterogeneous data includes energy data, process data, carbon emission data, topology data, and node data of the enterprise production system. Based on the node data, the node coupling relationship and node coupling strength between the system nodes in the enterprise production system are obtained. Energy data includes real-time data on the consumption of four types of energy—electricity, gas, cooling, and heating—by various energy-consuming equipment in the enterprise's production system; process data includes process parameters such as temperature, pressure, speed, and raw material ratios for each production step in the enterprise's production system; carbon emission data includes direct carbon emission data from each emission node in the enterprise's production system, such as monitoring data from a continuous emission monitoring system for flue gas; topology data includes the physical connections between various devices in the enterprise's production system, energy flow paths, and material transport routes; node data of the enterprise's production system includes, but is not limited to, source nodes in the enterprise's production system (such as power grid inlets, natural gas inlets, etc.). The system includes various types of system nodes, such as air intake, energy-consuming equipment nodes (e.g., motors, kilns, air conditioners), conversion nodes (e.g., transformers, cogeneration units), production line nodes (e.g., assembly lines, reactors), energy storage nodes (e.g., batteries, thermal storage tanks), and emission nodes (e.g., chimneys, wastewater treatment discharge outlets). It also includes the node coupling relationships and node coupling strengths of each system node. The node coupling relationships between system nodes include energy flow relationships, material flow relationships, and carbon flow coupling relationships. The node coupling strength refers to the rate of change in carbon emission output of other system nodes caused by changes in the energy input of a system node.

[0024] S112. Generate structural nodes based on the system node mapping in the enterprise production system. The structural nodes include node type labels and real-time status attributes. The generation of structural nodes based on the mapping of system nodes in the enterprise production system means that each system node in the enterprise production system corresponds to a structural node in the energy-carbon coupling heterogeneity diagram. Each structural node carries a node type label and real-time status attributes. The node type label specifically includes various node types such as source node, energy-consuming equipment node, conversion node, production line node, energy storage node, and emission node. The node type label of the structural node is obtained based on the node type of its corresponding system node, that is, the node type label of the structural node is consistent with the node type of its corresponding system node. The real-time status attributes refer to the real-time status data of the system node corresponding to the structural node, such as the current load, energy consumption rate, and carbon emission rate.

[0025] S113. Generate node connection edges between structural nodes based on the node coupling relationship between system nodes. The node connection edges include energy transmission edges, material supply edges, and carbon flow association edges, and form an edge set based on the node connection edges between all structural nodes. The node coupling relationships between system nodes include energy flow relationships, material flow relationships, and carbon flow coupling relationships. Energy transmission edges between corresponding structural nodes are generated based on the energy flow relationships between system nodes, such as "energy transmission edge from transformer to production line"; material supply edges between corresponding structural nodes are generated based on the material flow relationships between system nodes, such as "material supply edge from raw material warehouse to reactor"; carbon flow association edges between corresponding structural nodes are generated based on the carbon flow coupling relationships between system nodes, such as "carbon flow association edge from boiler to chimney"; the edge set contains node connection edges between all structural nodes.

[0026] S114. Generate edge weight values ​​for the nodes connecting the structural nodes based on the node coupling strength between system nodes, and construct a dynamic weight matrix based on the edge weight values ​​of the nodes connecting each node in the edge set. Each node connection edge is assigned a time-varying edge weight value by the node coupling strength between system nodes. The edge weight value specifically refers to the energy-carbon coupling strength of the structural node to other structural nodes at the current moment. This weight value can be defined as the rate of change in carbon emission output of other structural nodes caused by a unit change in the energy input of the structural node, that is, the degree of influence of a unit change in the energy input of the structural node on the carbon emission output of other structural nodes at the current moment. The dynamic weight matrix contains the edge weight values ​​of all node connection edges in the edge set.

[0027] S115. Construct an energy-carbon coupled heterogeneous graph based on structural nodes, edge sets, and dynamic weight matrices; The carbon-energy coupling isomer diagram can be represented as follows: ,in for Energy-carbon coupling isomer diagram at time t; For structural nodes; Let it be the set of edges; for The dynamic weight matrix at time step.

[0028] It should be noted that the energy-carbon coupling heterogeneous graph systematically and structurally maps the complex energy-carbon relationship in the enterprise production system into a machine-computable energy-carbon coupling heterogeneous graph representation, providing a physically interpretable modeling foundation for subsequent steps. Moreover, the dynamic weight matrix enables the energy-carbon coupling heterogeneous graph to have adaptive evolution capabilities, and can track medium- and long-term operating condition drift such as equipment aging, process changes, and seasonal changes.

[0029] S12. A dual-timescale update mechanism is introduced to update the energy-carbon coupling isomer diagram, including: S121. The dual time-scale update mechanism includes a slow time-scale mechanism and a fast time-scale mechanism; Slow timescale mechanisms are update mechanisms with timescales on the order of minutes or hours, while fast timescale mechanisms are update mechanisms with timescales on the order of seconds or milliseconds.

[0030] S122. The dynamic weight matrix in the energy-carbon coupling heterogeneity diagram is periodically updated based on the slow time scale mechanism. Based on the execution feedback data transmitted back through the second feedback loop in S4, the dynamic weight matrix in the carbon-coupling heterogeneous graph is periodically updated on a minute or hourly basis. The specific update logic is to compare the measured carbon emission change rate after the previous round of decision execution with the carbon emission change rate predicted by the attention network in the time series graph of step S2. If the measured change rate is lower than the predicted carbon emission change rate, it indicates that the coupling strength of the connection edge of the node is actually weak, and the edge weight of the connection edge of the node is automatically reduced; if the measured change rate is higher than the predicted carbon emission change rate, the edge weight of the connection edge of the node is strengthened. This update is usually triggered on a minute or hourly basis.

[0031] S123. The real-time state attributes of structural nodes in the energy-carbon coupling heterogeneous graph are periodically updated based on the fast timescale mechanism. When new multi-source heterogeneous data is acquired, the carbon-coupled heterogeneous graph updates the real-time status attributes of the corresponding structural nodes at the second or millisecond level (e.g., updating the "current load" to the latest value) based on a fast time scale mechanism without changing the edge set and dynamic weight matrix. This update is continuous and always maintains a response speed at the second or millisecond level.

[0032] It should be noted that the dual-timescale update mechanism balances the accuracy of the carbon-coupling heterogeneous graph with the real-time performance of decision-making, resolving the contradiction between "model accuracy" and "real-time performance of decision-making" in existing technologies. In addition, the slow timescale mechanism enables the carbon-coupling heterogeneous graph to have the intelligent characteristic of continuous self-evolution and to always keep in sync with the actual aging, wear and drift state of physical equipment, ensuring the accuracy of carbon emission accounting. The fast timescale mechanism enables the carbon-coupling heterogeneous graph to have millisecond-level situational awareness capabilities, enabling it to respond instantly to real-time situations such as grid voltage fluctuations or sudden speed changes in production lines.

[0033] S2. Construct a temporal graph attention network containing a multi-dimensional attention mechanism, obtain the coupling sensitivity of each structural node in the energy-carbon coupling heterogeneous graph based on the temporal graph attention network, and generate a global situational awareness report based on the coupling sensitivity. S21. Construct a temporal graph attention network that includes a multi-dimensional attention mechanism, including: The temporal graph attention mechanism includes the following learnable parameters: spatial attention-related parameters, temporal attention-related parameters, and multi-head attention fusion parameters; Spatial attention-related parameters include the weight matrix. Value weight matrix Attention weight vector The weight matrix The dimension is Used to input structure nodes Its adjacent structural nodes The execution state attribute features are linearly transformed from the initial feature space to the attention feature space. In this embodiment, represents the dimension of the state feature vector in the real-time state attributes of the structural node. The value is 64; Value weight matrix The dimension is This is used to extract state feature vectors from the real-time state attributes of adjacent structural nodes to generate spatial aggregation features; attention weight vectors The dimension is This is used to map the concatenated state feature vector to an attention score; Temporal attention-related parameters include the transformation matrix. Transformation matrix and attention weight vector Transformation matrix and transformation matrix All dimensions are These are used to map the spatial aggregated features of the current time and historical time to the same dimensional space for comparison; attention weight vector The dimension is This is used to map the fused state feature vector to a time correlation score.

[0034] Multi-head attention fusion parameters include the output transformation matrix. The dimension of the output transformation matrix is ,in The number of attention heads in this embodiment is [number]. The value is 4, and this is the output transformation matrix. Used to compress high-dimensional features from multiple attention heads back to the target dimension.

[0035] The above can be learned parameters: weight matrix Value weight matrix Attention weight vector Transformation matrix Transformation matrix Attention weight vector Output transformation matrix All of these can be obtained through the following training process: (1) Based on the input and output dimensions of the time-series attention network, a uniformly distributed sampling interval is formed, and the weight matrix is... Value weight matrix Transformation matrix Transformation matrix Output transformation matrix The weight matrix is ​​obtained by sampling from the sampling interval. Value weight matrix Transformation matrix Transformation matrix Output transformation matrix Initial values, attention weight vector and attention weight vector From the interval Random sampling is performed to obtain the attention weight vector. and attention weight vector The initial value; (2) Obtain historical operation data of the enterprise production system. The time span can be up to 12 months. Construct training samples based on the historical operation data. Each training sample is a graph sequence data within a training time window. The time step of the training time window can be 4. Each training sample contains input features and labels. The input features are the node state vector of the real-time state attribute of the structural node at each moment within the training time window and the edge weight of the node connection edge between the structural node and the adjacent structural node. The label is the actual carbon emission change at the next moment after the end of the training time window. (3) Forward propagation is performed based on the training samples. Specifically, all training samples are randomly divided into training set, validation set and test set in a ratio of 8:1:1. The training set is used for parameter update, the validation set is used for hyperparameter tuning and early stopping judgment, and the test set is used for final performance evaluation. The training set samples are input into the temporal graph attention network, and forward propagation is performed in the following order: spatial attention calculation: using the current weight matrix Value weight matrix Attention weight vector Calculate the spatial attention weights and spatial aggregation features at each time step, using the current transformation matrix. Transformation matrix and attention weight vector Calculate the spatiotemporal attention weights and spatiotemporal fusion features using the current output transformation matrix. The multiple attention points are fused into a final feature vector, and the carbon emission change is obtained based on the final feature vector. The mean square error between the carbon emission change and the actual carbon emission change is used as the loss function. (4) Backpropagation and parameter updates are performed using the Adam optimizer, which is well-known in the art. In each training round, the training set is divided into multiple sub-batches with a batch size of 32, and forward propagation, loss calculation, backpropagation and parameter updates are performed sequentially. The initial learning rate in the Adam optimizer is set to 0.001, the exponential decay rate of the first moment estimation is set to 0.9, the exponential decay rate of the second moment estimation is set to 0.999, and the numerical stability constant is set to 10. -8 Set the batch size to 32; (5) Set the maximum number of training rounds to 500. After each training round, calculate the loss value on the validation set. The training termination condition is: the maximum number of training rounds is reached (500 rounds), or the validation set loss does not decrease for 20 consecutive rounds. After training, the final weight matrix is ​​obtained. Value weight matrix Attention weight vector Transformation matrix Transformation matrix Attention weight vector Output transformation matrix .

[0036] S211. Sequence graph attention networks include spatial attention mechanisms, temporal attention mechanisms, and multi-head attention mechanisms. This temporal graph attention network is a deep learning model independently designed for energy-carbon coupled heterogeneous graphs. Its core processing logic includes three parallel attention mechanisms: spatial attention, temporal attention, and multi-head attention. The spatial attention mechanism is used to aggregate the neighboring structural node information of each structural node in the energy-carbon coupled heterogeneous graph, specifically by aggregating the neighboring structural nodes of a structural node within a single time step. The temporal attention mechanism is used to aggregate the historical node information of structural nodes, specifically by fusing information across time steps. The multi-head attention mechanism is used to aggregate the parallel spatiotemporal attention information of structural nodes from multiple perspectives, specifically by aggregating multiple sets of independent parameters of structural nodes in different dimensions.

[0037] S212. Generate the historical time series corresponding to the current moment based on the preset time window, construct the historical graph sequence based on the energy-carbon coupling heterogeneity graph at each moment in the historical time series, and use the spatial attention mechanism to calculate the attention score of the structural node to the adjacent structural node based on the historical graph sequence, and obtain the spatial aggregation characteristics of the structural node at each moment based on the attention score. The preset time window can be 4-12 time steps, with the specific value depending on the rate of change in operating conditions. A smaller preset time window value is used for scenarios with drastic changes to enable rapid response, while a larger time window value is used for scenarios with gradual changes to capture long-term patterns. In this embodiment, the preset time window is preferably 8 time steps. The historical time series can be specifically represented as follows: The value of K is 8, at each time step The corresponding historical time series are all .

[0038] The historical graph sequence can be specifically represented as ,in, The time step of the preset time window, for Energy-carbon coupling isomer diagram at time 10000. for Energy-carbon coupling isomer diagram at time 10000. for Energy-carbon coupling isomer diagram at time t.

[0039] The specific calculation process of the attention score of a structural node to its adjacent structural nodes is as follows: (1) Using the weight matrix Linear transformation of structure nodes respectively and adjacent structural nodes The corresponding state feature vector in the real-time state attributes, and Time structure node and adjacent structural nodes (2) Concatenate the three transformed vectors to obtain a concatenated long vector that integrates the information of the self-structure node, the adjacent structure node, and the physical association information between the self-structure node and the adjacent structure node; (3) Use the attention weight vector The inner product of the concatenated long vector is calculated to obtain a linear score. This linear score is then fed into the LeakyReLU activation function to introduce nonlinearity, ultimately yielding the time step. Lower structure node For adjacent structural nodes (4) Unnormalized attention raw scores; For adjacent structural nodes The original attention scores are used to perform natural exponentiation to obtain the structural nodes. For adjacent structural nodes The exponent value, and traversed the structure nodes. All adjacent structural nodes , structural nodes For all adjacent structural nodes The sum of the exponent values ​​is used as the denominator for the structural nodes. For a single adjacent structural node The exponent value is used as the numerator, and the structural nodes are obtained based on the numerator and denominator. For each single adjacent structural node Spatial attention weights, structural nodes For all adjacent structural nodes The sum of the spatial attention weights is 1; (5) Use the value weight matrix For all adjacent structural nodes Perform a linear transformation on the state feature vector in the corresponding real-time state attribute, based on the structure nodes. For adjacent structural nodes The spatial attention weights are applied to the linearly transformed state feature vectors, and the weighted state feature vectors of all adjacent structural nodes are summed to obtain... Time structure node Spatial aggregation features that integrate global adjacent structural node information.

[0040] S213. Based on the spatial aggregation characteristics of structural nodes at each time, a spatial temporal sequence is formed. The temporal attention mechanism is used to calculate the temporal attention weight of each time in the historical time series to the current time based on the spatial temporal sequence. Spatial time series can be specifically represented as ,in For structural nodes At any moment Spatial aggregation characteristics; For structural nodes At any moment Spatial aggregation characteristics; For structural nodes At any moment Spatial aggregation characteristics, The time step is the preset time window, and the value can be 8.

[0041] The specific calculation process of the time attention weight between the current time and each historical time in the historical time series is as follows: (1) Using the transformation matrix Linear transformation structure nodes At the present moment Spatial aggregation characteristics Using the transformation matrix Linear transformation structure nodes At a historical moment Node space features Transform the two sets of features and The vectors are summed element-wise to obtain a fused temporal feature vector that combines "current" and "historical" data. This temporal feature vector is then activated using the tanh activation function. Finally, the activated temporal feature vector is combined with the attention weight vector. Perform inner product calculations to obtain the unnormalized historical moments. For the current moment The time attention score is used to initially measure historical moments. For the current moment (2) The strength of the correlation; For the current moment The time attention score is used to perform natural exponentiation to obtain historical moments. For the current moment The exponential value is calculated, and all historical moments in the historical time series are iterated over to compare all historical moments with the current moment. The sum of the exponent values ​​is used as the denominator to represent a single historical moment. For the current moment The exponent value is used as the numerator, and a single historical moment is obtained based on the numerator and denominator. For the current moment Time attention weights, all historical moments in a historical time series For the current moment The sum of the time attention weights is 1.

[0042] It should be noted that the physical meaning of temporal attention weights is to represent the reference value of the energy-carbon coupling state at a historical moment to the current moment. For the current moment The smaller the time attention weight, the more historical the moment. The energy-carbon coupling state at the current moment The lower the reference value of a historical moment; For the current moment The greater the time attention weight, the more historical the moment. The energy-carbon coupling state at the current moment The higher the reference value, the better.

[0043] S214. Spatiotemporally fuse the temporal attention weights and spatial aggregation features to obtain the spatiotemporal fusion features of the structural nodes. By iterating through all historical moments in the historical time series, the temporal attention weight of each historical moment is multiplied element-wise by the spatial aggregation feature of that historical moment to obtain the spatial aggregation feature of each historical moment after being weighted by the temporal attention weight. The weighted spatial aggregation features of all historical moments are then summed to obtain the structural nodes. The spatiotemporal fusion characteristics at time t.

[0044] S215. Spatiotemporal processing units are generated based on the energy-carbon coupling heterogeneous graph. A multi-head attention mechanism is used to divide the spatiotemporal processing unit into at least one attention head and to linearly transform the spatiotemporal fusion features of each attention head to obtain the final feature vector of each structural node in the energy-carbon coupling heterogeneous graph. After inputting the carbon-energy coupling isomer diagram into the time-series graph attention network, the processing of spatial and temporal attention mechanisms is treated as a whole and denoted as a spatiotemporal processing unit. This complete spatiotemporal processing unit is then replicated using a multi-head attention mechanism. share, The value can be 4, with each part treated as an attention head. Each attention head has independently learnable parameters (including the weight matrix). Value weight matrix Attention weight vector Transformation matrix Transformation matrix Attention weight vector Output transformation matrix This allows us to focus on the feature patterns of different semantic dimensions in the energy-carbon coupling relationship: for example, attention head 1 focuses on the coupling path from "electricity consumption to indirect carbon emissions"; attention head 2 focuses on the coupling path from "fuel combustion to direct carbon emissions"; attention head 3 focuses on the coupling path from "material flow to implicit carbon emissions"; and attention head 4 focuses on the dynamic response pattern from "equipment load changes to carbon emission fluctuations". Each attention head independently performs a complete "space-to-time" attention mechanism calculation and outputs the spatiotemporal fusion features under that attention head. The specific process of linearly transforming the spatiotemporal fusion features of each attention head to obtain the final feature vectors of each structural node in the energy-carbon coupling heterogeneity graph is as follows: The spatiotemporal fusion features of each attention head are concatenated... The spatiotemporal fusion features from different attention heads are concatenated end to end to form a high-dimensional long vector. This high-dimensional long vector is then combined with the output transformation matrix. Element-wise multiplication of vectors is performed to achieve a linear transformation of high-dimensional long vectors, and the vector output by the linear transformation is used as a structural node. At any moment The final eigenvector.

[0045] It should be noted that by concatenating and merging all the multi-view features learned by all attention heads, the unique information captured by each attention head is preserved, and the transformation matrix is ​​output. By performing a linear transformation, the concatenated high-dimensional vector is adjusted to the dimension required by the downstream task. On the other hand, the network automatically learns the importance of different attention head information and merges them to obtain a unified and complete node representation. Finally, the feature vector integrates spatial and temporal coupled information from multiple perspectives, making it richer in information and more resistant to interference.

[0046] S22. Based on the temporal graph attention network, obtain the coupling sensitivity of each structural node in the energy-carbon coupling heterogeneity graph, and generate a global situational awareness report based on the coupling sensitivity, including: S221. The temporal graph attention network also has a pre-set network output layer. Based on the network output layer, carbon emission prediction is performed on the final feature vector of each structural node in the carbon-energy coupling heterogeneous graph to obtain the global carbon emission prediction. The network output layer consists of pooling layers and fully connected layers. The final feature vector output by the temporal graph attention network is first aggregated into a graph-level feature h_graph(t) (64 dimensions) through an attention pooling layer. Subsequently, h_graph(t) is fed into a three-layer multilayer perceptron (MLP) to predict global carbon emissions. Based on the final feature vectors of each structural node in the energy-carbon coupling heterogeneous graph from the network output layer, carbon emission prediction is performed, and the specific calculation process for obtaining the global carbon emission prediction is as follows: (1) Obtain the carbon emission prediction of each structural node in the energy-carbon coupling heterogeneous graph. The final feature vector at time t is obtained, and the pooling layer performs the known Pool graph pooling operation (such as average pooling or max pooling) on ​​the final feature vectors of all structural nodes, and the final feature vectors corresponding to the scattered structural nodes are fused into a global graph feature representing the state of the entire energy-carbon coupling heterogeneous graph; (2) The fused global graph feature is input into a multi-source perceptron composed of fully connected layers, and the global graph feature is nonlinearly mapped by the multi-source perceptron to obtain the current time t. The global carbon emission prediction is achieved using a multi-source sensor structure: the first fully connected layer maps the 64-dimensional input to 128 dimensions and connects it to a ReLU activation function and a Dropout layer with a dropout rate of 0.2; the second fully connected layer maps the 128-dimensional input to 64 dimensions and connects it to a ReLU activation function; the third layer is the output layer, which maps the 64-dimensional input to a 1-dimensional scalar and has no activation function. It should be noted that before carbon emission prediction is performed at the network output layer, the historical graph sequence of time t and the K consecutive times before time t in the enterprise's historical operating data is used as input, and the actual global carbon emission at time t+1 is used as the label. The entire network, including the temporal graph attention network, pooling layer and multi-source perceptron, is trained end-to-end using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. The optimization objective is to minimize the mean square error between the global carbon emission prediction and the actual global carbon emission. The training of the Adam optimizer is terminated when the training stops when the loss on the preset validation set does not decrease for 20 consecutive rounds. At this point, the weight parameters of each fully connected layer of the multi-source perceptron are determined.

[0047] S222. Obtain the energy input of the structural node at the current moment. Based on the energy input of the structural node at the current moment and the global carbon emission prediction, obtain the coupling sensitivity of the structural node. The specific calculation process for the coupling sensitivity of structural nodes is as follows: taking structural nodes as an example... At the present moment Using energy input as the independent variable and global carbon emission forecast as the dependent variable, calculate the rate of change of global carbon emission forecast with respect to energy input, and use this rate of change as the structural node. The coupling sensitivity at time t. The physical meaning of coupling sensitivity is... At any given time, for structural nodes When the energy input is adjusted unit by unit, the rate of change in global carbon emissions, and the higher the coupling sensitivity of the structural node, the greater the leverage effect of adjusting its energy input on carbon emission reduction. For example, if This indicates that for every 1 unit increase in the energy input of the structural node, Global carbon emissions increase ;like This indicates that for every 1 unit increase in the energy input of this node... Global carbon emissions increase ,in For structural nodes exist Moment-time coupling sensitivity.

[0048] S223. Generate a global situational awareness report based on the coupling sensitivity of each structural node in the energy-carbon coupling heterogeneous diagram; The coupling sensitivity of each structural node in the integrated energy-carbon coupling isomorphism diagram ,in, For structural nodes exist Moment-time coupling sensitivity, The range of values ​​is , where n is the total number of structural nodes in the carbon-coupled heterogeneous graph.

[0049] The global situation awareness report includes current situation information and future trend predictions; The current situation information includes the total global carbon emissions of the enterprise's production system at the current moment, as well as the carbon emissions of each system node; a complete list and ranking of the coupling sensitivity of each structural node, with the ranking sorted from high to low coupling sensitivity, and the structural nodes with the top 10% coupling sensitivity designated as high-leverage control nodes; and visualization of the critical path of the energy-carbon coupling relationship of the system at the current moment, i.e., highlighting the structural node with the highest coupling sensitivity and its node connection edges on the energy-carbon coupling heterogeneous graph. The future trend prediction includes a carbon emission prediction sequence for the enterprise's production system over the next T time periods. This carbon emission prediction sequence reflects the carbon emission change trend over the next T time periods. Specifically, after processing the historical graph sequence, the time-series graph attention network outputs a hidden state vector that encodes historical energy-carbon coupling information. This vector is used as the initial hidden state input to a single-layer LSTM decoder in the prediction head. The single-layer LSTM decoder generates the decoded hidden states for the next T time periods in an autoregressive manner. Each decoded hidden state is mapped to the corresponding carbon emission prediction value through a fully connected layer in the prediction head, thus obtaining a carbon emission prediction sequence containing the next T time periods. It should be noted that the training phase of the prediction head adopts a supervised learning approach, using historical graph sequences as input and corresponding future actual carbon emission sequences as labels. The optimization objective is to minimize the mean squared error between the predicted and actual values. The prediction head is trained using an Adam optimizer with a learning rate of 0.001 and a batch size of 32. The prediction head can be jointly trained end-to-end with the time-series graph attention network, or it can be fine-tuned separately after the encoder training is completed. This embodiment prefers the end-to-end joint training approach. To quantify the prediction uncertainty, the inference phase of the prediction head adopts the Monte Carlo Dropout method: keeping the Dropout layer active, performing 50 forward propagations on the same input, calculating the mean and standard deviation of the predicted values ​​at each time step, using the mean as the final predicted value at each time step, and using the mean ± 1.96 times the standard deviation as the 95% confidence interval.

[0050] S3. Construct an intelligent agent based on the enterprise production system, and introduce the agent's policy training gradient. Based on the policy training gradient and the global situational awareness report, conduct collaborative game training on the agent to obtain the agent's corresponding decision actions, and form a global decision action set based on the decision actions. S31. Construct intelligent agents based on the enterprise's production system, including: S311. Construct intelligent agents based on the energy-consuming equipment in the enterprise's production system. Each intelligent agent includes a local observation space, an action space, a policy network, and a local state memory. It should be noted that, in addition to energy-consuming equipment, each workshop and production line in the enterprise production system can also be defined as an intelligent agent. An intelligent agent is a software entity that can autonomously perceive, make autonomous decisions, and execute autonomously; and the intelligent agent corresponds to the structural node in the energy-carbon coupling heterogeneous diagram that can perform regulation.

[0051] S312. The local observation space includes the set of local observation information of the physical entity corresponding to the agent, and the action space includes the set of control actions that the agent can perform. The local observation space defines the set of information that an agent can obtain from local sensors, including but not limited to: the real-time load rate, instantaneous power, and cumulative energy consumption of the equipment / workshop / production line; key process parameters such as temperature, pressure, and rotation speed of the equipment / workshop / production line; the current operating status of the equipment / workshop / production line (e.g., running / standby / fault / maintenance); and the carbon emission rate of the equipment / workshop / production line (e.g., the equipment / workshop / production line is an emission node). The action space defines the set of control actions that an agent can perform. Its design varies depending on the type of equipment, specifically including: for continuous-action equipment (such as frequency converters and regulating valves), the action space is a continuous interval, such as the control action interval for load adjustment. Discrete-action equipment (such as switching equipment and multi-position equipment) has a finite set of action space, such as "{start, stop}" or "{position 1, position 2, position 3}". Composite-action equipment (such as combined heat and power units) has a mixed action space of continuous and discrete actions. It should be noted that the boundaries of the action space are limited by physical safety constraints. For example, the temperature of the reactor must not exceed the safety threshold, and the motor load must not exceed 120% of the rated power. These constraints are encoded during the action space design to ensure that the agent does not generate dangerous commands.

[0052] S313, The policy network is used to map local observation information into control actions, and the local state storage space is used to store the agent's historical local observation information and historical control actions. Each agent has a deep neural network as its policy network, which maps the agent's local observation information into action probability distributions (discrete actions) or deterministic action values ​​(continuous actions). It should be noted that the policy network is structured as a multilayer perceptron, and the specific number of layers and neurons is determined according to the complexity of device control. In this embodiment, the agent's policy network is a dual-input structure, that is, it simultaneously receives the agent's local observation information and the global situational awareness report from step S2. The global situational awareness report provides the sensitivity ranking of each structural node and the carbon emission trend prediction, enabling the agent to have a global perspective when making decisions.

[0053] In addition, each agent maintains a local state memory as its local state storage space, which is used to record its historical local observation information and historical control actions, and is used for online training and experience playback of the policy network.

[0054] It should be noted that the physical characteristics, control methods and safety constraints of different devices vary greatly. Instead of forcing uniformity, the local observation space and action space of each device type are customized through the intelligent agent architecture, so that the same management framework can be adapted to various types of equipment, from simple lighting switches to complex cogeneration units.

[0055] S32. Introduce the agent's policy training gradient, and train the agent in cooperative game theory based on the policy training gradient and the global situational awareness report, including: S321. Obtain the total carbon emissions and total energy consumption of the enterprise's production system at the current moment. Generate a penalized high carbon emission item based on the total carbon emissions and a penalized high energy consumption item based on the total energy consumption. Generate a positive incentive item based on the coupling sensitivity of the policy network and structural nodes, and construct a global reward function based on the penalized high carbon emission item, the penalized high energy consumption item, and the positive incentive item. Preset weight values Weight value These represent the degree of emphasis that a company's production system places on the three goals of carbon reduction, energy conservation, and precise control. Increasing the emphasis reflects this. The system will take more proactive measures to reduce carbon emissions; adjust... The system will be more inclined to choose an energy-saving operating mode; increase The system will respond more proactively to the "high carbon reduction leverage" nodes identified in step two, striving for precise regulation and pre-setting weight values. The value needs to be determined based on the company's specific strategic goals (whether to focus on energy conservation or carbon reduction) and current operating conditions. In this embodiment, a preset weight value is used. The preferred value is ; Set the preset weight value The negative value of the product of the total carbon emissions of the system is used as a penalty for high carbon emission items. The higher the total carbon emissions of the system, the larger the negative value of the penalty for high carbon emission items, thus guiding each agent to act in the direction of carbon reduction. Set the preset weight value The negative value of the product of the total system energy consumption and the total system energy consumption is used as a penalty for high energy consumption. The higher the total system energy consumption, the larger the negative value of the penalty for high energy consumption, thus guiding each agent to act in the direction of energy saving. Based on the policy network, the amplitude of the agent's control action at the current moment is obtained. The product of the amplitude of the control action of each agent in the enterprise production system and the coupling sensitivity of the corresponding structural node of the agent is accumulated and calculated. The accumulated calculation result is then compared with the preset weight value. The product of these factors serves as a positive incentive term. This positive incentive means that larger-scale regulatory actions are taken at highly sensitive nodes. For example, the higher the coupling sensitivity of the structural node corresponding to the agent, the greater the impact of adjusting the agent on global carbon emissions; therefore, the agent is actively regulated. Conversely, the lower the coupling sensitivity of the structural node corresponding to the agent, the smaller the impact of adjusting the agent on global carbon emissions; therefore, regulation of the agent is avoided. By setting positive incentive terms, each agent, in pursuing its own reward maximization, will be naturally guided towards "prioritizing the regulation of highly sensitive structural nodes," thereby achieving the greatest carbon reduction benefit with the least energy adjustment cost. The global reward function is obtained by summing the penalties for high carbon emissions, high energy consumption, and positive incentives. Based on this function, the global reward value corresponding to the agent's regulatory actions can be calculated. It should be noted that all agents in the enterprise production system share this global reward function, which determines the behavioral orientation of each agent.

[0056] S322, The policy training gradient includes the evaluation network, the experience replay pool, and the agent's policy network; A policy training gradient, comprising a policy network, a judge network, and an experience replay pool, is introduced to enable cyclical updating training of the judge network and the policy network. The training objective is to obtain an updated policy network, so that the regulatory actions generated by the policy network in the agent are more in line with the evaluation mechanism of the judge network.

[0057] Obtaining the updated policy network involves two phases: training and execution. The training phase involves each agent's policy network generating regulatory actions based on local observations. However, the evaluation network, during training, accesses the states, actions, and global reward information of all agents, guiding their regulatory actions towards the cooperative optimal direction to obtain the cumulative reward expectation. The network parameters are then adjusted based on this cumulative reward expectation to obtain the updated policy network. Specifically, if the cumulative reward expectation is greater than zero, the network parameters are fine-tuned along the positive gradient direction, making the policy network more inclined to output larger action values ​​next time; if the cumulative reward expectation is less than zero, the network parameters are fine-tuned along the negative gradient direction, making the policy network more inclined to output smaller action values ​​next time. The execution phase involves retaining only the policy networks of each agent after training for online execution to obtain their regulatory actions.

[0058] S323. The evaluation network is used to evaluate the cumulative expected value of the regulatory actions generated by the agent based on the global reward function; The training objective of the evaluation network is to more accurately assess the cumulative expected reward value of each agent's corresponding regulatory actions, and to judge whether the regulatory actions generated by the policy network are good or bad by the cumulative expected reward value.

[0059] S324. Based on the cumulative reward expectation value of the evaluation network, iteratively update and train the network parameters of the policy network in the agent to obtain the updated policy network of the agent. The training objective of the policy network is to maximize the cumulative expected reward value corresponding to the regulatory actions generated by each agent by continuously adjusting its internal network parameters.

[0060] S325. The training method of the experience replay pool is to randomly sample historical data from the experience replay pool during training, instead of using the online data at the current moment. This "experience replay" mechanism breaks the temporal correlation between data, making training more stable and avoiding training oscillations.

[0061] S33. Obtain the decision actions corresponding to the agent, and form a global decision action set based on the decision actions, including: S331. Obtain the coupling sensitivity of the structural nodes corresponding to each agent based on the global situational awareness report; The coupling sensitivity of the corresponding structural nodes of the agent is input into the policy network trained in the agent based on the local observation information and global situational awareness report of the agent, and the agent's control actions are generated based on the policy network.

[0062] S332. Summarize and process the control actions of all intelligent agents in the enterprise production system to obtain a global decision action set; Global decision action set Specifically, it can be represented as ,in For intelligent agents The regulatory actions, The value is , This represents the total number of intelligent agents in the enterprise's production system.

[0063] S4. Based on the global decision action set, make decisions and execute them for each agent in the enterprise production system, and obtain the execution feedback data of the agents. Update the energy-carbon coupling heterogeneous graph and policy training gradient based on the feedback loop generated by the execution feedback data. Then, put the energy-carbon coupling heterogeneous graph, global situational awareness report, global decision action set and execution feedback data on the blockchain for evidence storage, and build a traceable decision ledger for the enterprise production system.

[0064] S41. Convert the control actions in the global decision-making action set into equipment control instructions, execute the corresponding control actions on the execution terminal of the enterprise production system according to the equipment control instructions, and obtain the actual operation data of each intelligent agent after the control actions are executed, and form execution feedback data based on the actual operation data.

[0065] The centralized control actions of global decision-making are converted into equipment control commands, such as adjusting load, changing temperature setpoints, and starting or stopping equipment. These commands are then sent to the execution terminals of each piece of equipment / workshop / production line via industrial communication protocols to perform the corresponding control operations. After the control is executed, the IoT sensing layer collects the actual operating data of each node, such as changes in energy consumption, carbon emissions, and process parameters, to form execution feedback data.

[0066] S42. Perform differential privacy processing on the execution feedback data of each agent to obtain the first feedback loop and the second feedback loop, update the energy-carbon coupling heterogeneous graph according to the first feedback loop, and update the experience replay pool of the policy training gradient according to the second feedback loop. Differential privacy processing of the execution feedback data of each agent refers to the following: For continuous numerical data in the execution feedback data, such as changes in energy consumption and carbon emission rates, calibrated noise is added using a Laplace mechanism known in the art. Specifically, this involves: acquiring continuous numerical data in the execution feedback data; calculating a scale parameter based on the ratio of global sensitivity to privacy budget; randomly sampling noise values ​​from the Laplace distribution of this scale parameter; and adding the continuous numerical data and noise values ​​to obtain the differentially privacy-processed continuous numerical data. Here, global sensitivity is the maximum value of the change corresponding to the continuous numerical data in the agent, such as the maximum value of energy consumption change. The smaller the privacy budget value, the stronger the privacy protection of the data. In this embodiment, the preferred value for the privacy budget is 1. For discrete data in the execution feedback data, such as changes in device start / stop status and gear shifting records, a well-known exponential mechanism is used for privacy protection, outputting the true state with a high probability and other states with a low probability, ensuring differential privacy protection.

[0067] The first feedback loop uses the execution feedback data as update data for the energy-carbon coupling heterogeneity diagram to update the dynamic weight matrix in the diagram; the second feedback loop uses the execution feedback data as new samples for the experience replay pool to update the experience replay pool.

[0068] S43. Obtain the hash values ​​of the energy-carbon coupling heterogeneous diagram, global situational awareness report, global decision action set and execution feedback data respectively, store the hash values ​​on the blockchain, and form a traceable decision ledger based on the hash values ​​after storage on the blockchain. The energy-carbon coupling heterogeneous graph, global situational awareness report, global decision action set, and execution feedback data are stored on the blockchain using hash digests. Specifically, the hash values ​​of the energy-carbon coupling heterogeneous graph, global situational awareness report, global decision action set, and execution feedback data are stored on the blockchain. The construction of a traceable decision ledger makes the entire decision-making process auditable, meaning that each round of the "perception → decision → execution → feedback" process has an immutable record.

[0069] Example 2, Figure 2 This invention presents an AI-based energy and carbon synergistic management system, comprising: a coupled heterogeneous graph construction module, used to acquire multi-source heterogeneous data from an enterprise production system, construct an energy and carbon coupled heterogeneous graph based on the multi-source heterogeneous data, and introduce a dual-timescale update mechanism to update the energy and carbon coupled heterogeneous graph; a perception report generation module, used to construct a time-series graph attention network containing a multi-dimensional attention mechanism, obtain the coupling sensitivity of each structural node in the energy and carbon coupled heterogeneous graph based on the time-series graph attention network, and generate a global situational awareness report based on the coupling sensitivity; and a decision-making action generation module, used to construct an intelligent agent based on the enterprise production system and introduce the intelligent agent's policy... The system employs a training gradient approach, using policy training gradients and global situational awareness reports to train agents through collaborative game theory. This yields the agent's corresponding decision actions, which are then used to form a global decision action set. The decision execution and feedback module executes decisions for each agent in the enterprise production system based on the global decision action set. It also acquires the agent's execution feedback data, updates the energy-carbon coupling heterogeneous graph and policy training gradient based on the feedback loop generated from the execution feedback data, and stores the energy-carbon coupling heterogeneous graph, global situational awareness reports, global decision action set, and execution feedback data on the blockchain to construct a traceable decision ledger for the enterprise production system.

[0070] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0071] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0072] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0073] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0074] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0075] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI technology-based carbon collaborative management method, characterized by, Includes the following steps: Acquire multi-source heterogeneous data from the enterprise's production system, construct an energy-carbon coupling heterogeneity diagram based on the multi-source heterogeneous data, and introduce a dual-timescale update mechanism to update the energy-carbon coupling heterogeneity diagram; A time-series graph attention network with a multi-dimensional attention mechanism is constructed. The coupling sensitivity of each structural node in the energy-carbon coupling heterogeneous graph is obtained based on the time-series graph attention network, and a global situational awareness report is generated based on the coupling sensitivity. Intelligent agents are constructed based on enterprise production systems, and policy training gradients are introduced for the agents. The agents are trained through collaborative game based on the policy training gradients and global situational awareness reports to obtain the decision actions corresponding to the agents. A global decision action set is then formed based on the decision actions. Based on the global decision action set, decisions are made and executed by each agent in the enterprise production system, and the execution feedback data of the agents is obtained. The energy-carbon coupling heterogeneous graph and policy training gradient are updated based on the feedback loop generated by the execution feedback data. The energy-carbon coupling heterogeneous graph, global situational awareness report, global decision action set and execution feedback data are stored on the blockchain to build a traceable decision ledger for the enterprise production system. 2.The AI technology-based carbon synergic management method according to claim 1, characterized in that, Acquire multi-source heterogeneous data from the enterprise's production system, and construct an energy-carbon coupling heterogeneity diagram based on the multi-source heterogeneous data, including: The multi-source heterogeneous data includes energy data, process data, carbon emission data, topology data, and node data of the enterprise production system. Based on the node data, the node coupling relationship and node coupling strength between the system nodes in the enterprise production system are obtained. A structural node is generated based on the system node mapping in the enterprise production system. The structural node includes a node type label and a real-time status attribute. Based on the node coupling relationship between system nodes, node connection edges between structural nodes are generated. These node connection edges include energy transmission edges, material supply edges, and carbon flow association edges, and an edge set is formed based on all the node connection edges between structural nodes. The edge weights of the nodes connecting the system nodes are generated based on the node coupling strength between the system nodes, and a dynamic weight matrix is ​​constructed based on the edge weights of the nodes connecting each node in the edge set. Based on the structural nodes, edge set, and dynamic weight matrix, an energy-carbon coupled heterogeneous graph is constructed. 3.The AI technology-based carbon synergic management method according to claim 2, characterized in that, A dual-timescale update mechanism is introduced to update the energy-carbon coupling isomer diagram, including: The dual time-scale update mechanism includes a slow time-scale mechanism and a fast time-scale mechanism; The dynamic weight matrix in the energy-carbon coupling heterogeneity diagram is periodically updated based on the aforementioned slow timescale mechanism. The real-time state attributes of structural nodes in the energy-carbon coupling heterogeneous graph are periodically updated based on the fast timescale mechanism described above. 4.The AI technology-based carbon synergic management method of claim 1, wherein, Construct a temporal graph attention network incorporating a multidimensional attention mechanism, including: The temporal graph attention network includes spatial attention mechanisms, temporal attention mechanisms, and multi-head attention mechanisms. Based on a preset time window, a historical time series corresponding to the current moment is generated. A historical graph sequence is constructed based on the energy-carbon coupling heterogeneity graph at each moment in the historical time series. The spatial attention mechanism is used to calculate the attention score of the structural node to the adjacent structural node based on the historical graph sequence, and to obtain the spatial aggregation characteristics of the structural node at each moment based on the attention score. A spatial temporal sequence is formed based on the spatial aggregation characteristics of structural nodes at each time point. The temporal attention mechanism is used to calculate the temporal attention weight of each time point in the historical time sequence to the current time point based on the spatial temporal sequence. The temporal attention weights and spatial aggregation features are spatiotemporally fused to obtain the spatiotemporal fusion features of the structural nodes. Spatiotemporal processing units are generated based on the energy-carbon coupling heterogeneous graph. The multi-head attention mechanism is used to divide the spatiotemporal processing unit into at least one attention head and to linearly transform the spatiotemporal fusion features of each attention head to obtain the final feature vector of each structural node in the energy-carbon coupling heterogeneous graph. 5.The AI technology-based carbon synergic management method according to claim 4, characterized in that, The coupling sensitivity of each structural node in the energy-carbon coupling heterogeneity graph is obtained based on the time-series graph attention network, and a global situational awareness report is generated based on the coupling sensitivity, including: The time-series graph attention network also has a preset network output layer. Based on the network output layer, carbon emission prediction is performed on the final feature vector of each structural node in the carbon-energy coupling heterogeneous graph to obtain the global carbon emission prediction. Obtain the energy input of the structural node at the current moment, and based on the energy input of the structural node at the current moment and the global carbon emission prediction, obtain the coupling sensitivity of the structural node. A global situational awareness report is generated based on the coupling sensitivity of each structural node in the energy-carbon coupling heterogeneous diagram. 6.The AI technology-based carbon synergic management method according to claim 1, characterized in that, Intelligent agents are constructed based on the enterprise's production system, including: Intelligent agents are constructed based on the energy-consuming equipment in the enterprise's production system. Each intelligent agent includes a local observation space, an action space, a policy network, and a local state memory. The local observation space includes the set of local observation information of the physical entity corresponding to the agent, and the action space includes the set of control actions that the agent can perform. The policy network is used to map local observation information into regulatory actions, and the local state storage space is used to record the agent's historical local observation information and historical regulatory actions. 7.The AI technology-based carbon synergic management method according to claim 6, characterized in that, Introducing a policy training gradient for the agent, and conducting cooperative game training of the agent based on the policy training gradient and global situational awareness report, including: The policy training gradient includes a judge network, an experience replay pool, and a policy network for the agent. Obtain the total carbon emissions and total energy consumption of the enterprise's production system at the current moment. Generate a penalized high carbon emission item based on the total carbon emissions and a penalized high energy consumption item based on the total energy consumption. Generate a positive incentive item based on the coupling sensitivity of the agent's policy network and the corresponding structural nodes of the agent. Construct a global reward function based on the penalized high carbon emission item, the penalized high energy consumption item, and the positive incentive item. The evaluation network is used to evaluate the cumulative expected value of the regulatory actions generated by the agent based on the global reward function; The network parameters of the policy network in the agent are iteratively updated and trained based on the cumulative reward expectation value of the evaluation network to obtain the updated policy network of the agent. The experience replay pool is used to store global situational awareness reports, control actions, and the expected cumulative reward values ​​corresponding to the control actions. 8.The AI technology-based carbon synergic management method according to claim 1, wherein, The process of obtaining the decision actions corresponding to the intelligent agent and assembling a global decision action set based on these decision actions includes: The coupling sensitivity of the structural nodes corresponding to each agent is obtained based on the global situational awareness report; The local observation information of the agent and the coupling sensitivity of the structural nodes corresponding to the agent are input into the trained policy network of the agent, and the agent's control actions are generated according to the policy network. The control actions of all intelligent agents in the enterprise production system are aggregated and processed to obtain a global decision action set. 9.The AI technology-based carbon synergic management method according to claim 1, characterized in that, The feedback loop generated based on the execution feedback data updates the energy-carbon coupling heterogeneity graph and policy training gradients. The energy-carbon coupling heterogeneity graph, global situational awareness report, global decision action set, and execution feedback data are then stored on the blockchain to construct a traceable decision ledger, including: The control actions in the global decision-making action set are converted into equipment control instructions. Based on the equipment control instructions, the corresponding control actions are executed on the execution terminal of the enterprise production system. The actual operating data of each intelligent agent after the control actions are executed are obtained, and execution feedback data is formed based on the actual operating data. Differential privacy processing is performed on the execution feedback data of each agent to obtain the first feedback loop and the second feedback loop. The energy-carbon coupling heterogeneous graph is updated according to the first feedback loop, and the experience replay pool of the policy training gradient is updated according to the second feedback loop. The hash values ​​of the energy-carbon coupling heterogeneous diagram, global situational awareness report, global decision action set, and execution feedback data are obtained respectively. The hash values ​​are stored on the blockchain for notarization, and a traceable decision ledger is formed based on the notarized hash values.

10. A system using the AI-based energy and carbon synergistic management method as described in any one of claims 1-9, comprising: The coupling heterogeneity graph construction module is used to acquire multi-source heterogeneous data from the enterprise's production system, construct an energy-carbon coupling heterogeneity graph based on the multi-source heterogeneous data, and introduce a dual time-scale update mechanism to update the energy-carbon coupling heterogeneity graph. The perception report generation module is used to construct a time-series graph attention network containing a multi-dimensional attention mechanism, obtain the coupling sensitivity of each structural node in the energy-carbon coupling heterogeneous graph based on the time-series graph attention network, and generate a global situational awareness report based on the coupling sensitivity. The decision action generation module is used to build intelligent agents based on the enterprise production system, introduce the policy training gradient of the intelligent agents, conduct collaborative game training on the intelligent agents based on the policy training gradient and the global situational awareness report, obtain the decision actions corresponding to the intelligent agents, and form a global decision action set based on the decision actions. The decision execution and execution feedback module is used to execute decisions for each agent in the enterprise production system based on the global decision action set, obtain the execution feedback data of the agents, update the energy-carbon coupling heterogeneous graph and policy training gradient based on the feedback loop generated by the execution feedback data, and put the energy-carbon coupling heterogeneous graph, global situational awareness report, global decision action set and execution feedback data on the blockchain for evidence storage, so as to build a traceable decision ledger for the enterprise production system.

Images