A low-carbon park adaptive regulation method based on multi-source data fusion and energy-carbon double-flow coupling

Abstract

The present application relates to a kind of based on multi-source data fusion and energy-carbon double flow coupling low-carbon park adaptive control method. The multi-modal data of the operation of each flexible resource in low-carbon park is collected and the dynamic characteristics are extracted;Based on the composite similarity measure, the flexible resource is adaptively layered and partitioned, and the virtual flexible resource is obtained;Each virtual flexible resource is constructed as an independent reinforcement learning agent, and the layered coordination optimization of attention mechanism weighted reinforcement learning multi-agent is carried out;Adopt double-dimensional attention prediction model to predict the generation park net load value and tie line exchange power, use dynamic graph attention carbon potential perception model to generate each node carbon potential prediction value, and feedback to global coordination layer to form closed loop, finally obtain the optimal operation strategy trajectory of park and energy-saving and carbon-reducing potential.Compared with prior art, the present application realizes the scalable collaborative optimization of large-scale system and the real-time closed-loop coupling regulation of energy flow-carbon flow.

Description

Technical Field

[0001] This invention relates to the field of low-carbon regulation technology for integrated energy systems in industrial parks, and in particular to an adaptive regulation method for low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling. This method is especially suitable for real-time optimization scheduling and refined carbon emission management of complex industrial park energy systems that include distributed renewable energy, diversified energy storage, combined cooling, heating and power (CCHP), and adjustable loads. Background Technology

[0002] Integrated energy systems at the park level, due to their diverse energy sources, complex energy consumption behaviors, and centralized and controllable carbon emissions, have become a key scenario for achieving energy structure optimization and carbon emission reduction. Typical low-carbon parks usually integrate various heterogeneous energy units such as photovoltaics, wind power, gas turbines, energy storage, heat pumps, and adjustable loads. Each unit has significant differences in physical characteristics, response speed, operational constraints, and carbon emission intensity, forming a complex system with high-dimensional nonlinearity, strong uncertainty, and multi-timescale coupling characteristics.

[0003] Currently, the park's energy management and carbon emission reduction technologies mainly face the following three unresolved technical challenges: Challenge 1: Difficulty in the unified representation and dynamic aggregation of heterogeneous flexible resources. Most existing methods employ simplified steady-state models to describe equipment characteristics, such as treating energy storage charging and discharging efficiency as constants, the adjustment range of adjustable loads as fixed intervals, and the power generation efficiency of gas turbines as rated values. This simplification leads to a significant deviation of the aggregated virtual resource external characteristics from reality. Scheduling strategies based on distorted models often encounter problems such as power limit exceeding, response lag, and even equipment protection actions during actual execution. Although some studies have attempted to extract equipment characteristics using data-driven methods, there is a lack of differentiated modeling strategies for the different operating mechanisms of various resource types, making it difficult to achieve an accurate mapping from micro-level equipment characteristics to macro-level cluster characteristics.

[0004] Challenge Two: The Curse of Dimensionality and Convergence in Multi-Agent Collaborative Decision-Making. A park system contains dozens or even hundreds of controllable devices. If each device is directly modeled as an independent decision-making unit, the state-action space of reinforcement learning will expand exponentially, making training difficult to converge. Existing centralized optimization methods (such as mixed integer programming) offer high accuracy, but their computational complexity increases dramatically with system size, making them unsuitable for online real-time scheduling. While hierarchical control can reduce complexity to some extent, the information transfer and goal coordination mechanisms between upper and lower layers are still imperfect, easily leading to a disconnect between upper-level goals and lower-level feasible regions.

[0005] Challenge Three: The Lag in Carbon Emission Accounting and the Disconnect Between Optimization Decision-Making. Current carbon accounting in industrial parks generally uses monthly or annual static statistics based on average emission factors, which cannot reflect minute-level changes in carbon intensity caused by fluctuations in new energy sources, equipment start-ups and shutdowns, and load changes during system operation. More importantly, carbon flow information and the optimization scheduling process are independent. The optimization model cannot know the real-time carbon potential of each node when making decisions. This leads to scheduling strategies that, while meeting economic objectives, may increase electricity load during high-carbon periods, thus pushing up actual carbon emissions. Although some literature proposes using graph neural networks for carbon potential prediction, the prediction results are only output as an independent module and are not embedded in the closed-loop feedback of optimization decision-making, failing to achieve real-time linkage between perception, decision-making, execution, and evaluation. Summary of the Invention

[0006] The purpose of this invention is to address the difficulties in the unified representation and dynamic aggregation of heterogeneous and flexible resources in existing technologies, the dimensionality curse and convergence problems in multi-agent collaborative decision-making, and the lag in carbon emission accounting and the disconnect between optimization decision-making. This invention provides an adaptive regulation method for low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling.

[0007] The objective of this invention can be achieved through the following technical solutions: A method for regulating low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling, comprising the following steps: Collect multi-source heterogeneous operation data of various flexible resources in the low-carbon park and extract feature vectors; The feature vectors of each flexibility resource are input into a three-layer aggregation architecture that includes the device layer, cluster layer and system layer. Based on the composite similarity metric, the flexibility resources are hierarchically and partitioned to aggregate virtual flexibility resources. Each virtual flexibility resource is constructed as an independent reinforcement learning agent, and hierarchical coordination optimization based on attention mechanism weighted reinforcement learning is carried out. The global coordination layer aims at the overall economic efficiency and low carbon emissions of the park and generates coordination and control instructions for each virtual flexibility resource. In the local execution layer, each agent takes its own local observation state and coordination and control instructions as input and outputs decision actions. A dual-dimensional attention prediction model integrating temporal and spatial attention is adopted to generate predicted net load values ​​and tie-line switching power values ​​for multiple future sampling periods; a dynamic graph attention carbon potential perception model is constructed based on the actual topology of the park to generate predicted carbon potential values ​​for each node. The predicted net load of the park and the predicted switching power of the tie line are used as the forward input of the global coordination layer, and the real-time carbon potential prediction of each node is used as the basis for real-time updating of the low-carbon reward item of the global reward function. A closed-loop feedback link is formed and a shrinking time-domain model predictive control framework is used for rolling optimization. Finally, the flexibility boundary of the park system, the optimal operating strategy trajectory, and the energy-saving and carbon-reducing potential are output.

[0008] As a preferred technical solution, the flexible resources include: adjustable and non-adjustable loads, energy conversion equipment, energy storage devices, and renewable energy power generation units; For adjustable loads, their physical adjustment parameters are jointly modeled with user behavior uncertainties to construct a feature vector that includes rated power, adjustment depth, response delay, continuous adjustment capability, energy state boundary, and comfort relaxation factor. For non-adjustable loads, the power sequence is decomposed into trend components, periodic components, and random residual components using the local weighted regression scatter smoothing method, and the statistical moments of the residual components are used as uncertainty features. For energy conversion equipment, the load efficiency curve and the full-condition efficiency map are fitted with historical operating data. The full-condition efficiency map uses input power and output power as coordinate axes to represent the equipment efficiency as a binary quadratic response surface function. At the same time, the minimum stable output, maximum ramp rate, start-up and shutdown time and start-up and shutdown energy consumption of the energy conversion equipment are extracted as dynamic features. For energy storage devices, the nonlinear coupling relationship between the state of charge, health state and charge / discharge efficiency of the energy storage device is modeled as a third-order state-space equation. In the third-order state-space equation, the charging efficiency and discharging efficiency are nonlinear functions of the current state of charge, charge / discharge rate and operating temperature, respectively, and the health state is the superposition of the cumulative effect of cyclic aging and the cumulative effect of calendar aging. For a renewable energy power generation unit, variational mode decomposition is used to decouple the output sequence of the renewable energy power generation unit into multiple intrinsic mode components with different center frequencies. Fluctuation statistical features are extracted for each intrinsic mode component, and the instantaneous frequency and instantaneous amplitude of each component constitute a fluctuation mode feature vector.

[0009] As a preferred technical solution, in the device-cluster-system three-layer aggregation architecture: At the device layer, the maximum information coefficient is used to filter the extracted feature vectors of each flexibility resource, and redundant features are removed before being used as the clustering input matrix. At the cluster layer, a two-stage clustering strategy based on composite similarity metrics is used for resource aggregation. The first stage is to improve density peak clustering. The composite similarity between devices is used as a distance measure to calculate the local density of each device and its minimum composite distance with higher density devices. The product of local density and minimum composite distance is used as the cluster center selection index to identify cluster centers and complete the initial resource partitioning. Devices with a density lower than the global density threshold and a composite distance from each cluster center exceeding a set multiple are marked as boundary devices. The second stage is fuzzy subspace clustering, which flexibly corrects the initial resource partition boundaries formed in the first stage. In fuzzy subspace clustering, each feature dimension is assigned an independent weight coefficient, allowing different virtual flexibility resources to have differentiated levels of attention across different feature dimensions. Through iterative optimization of the membership degree of each device to various virtual flexibility resources and the weight coefficients of each feature dimension, devices within the same virtual flexibility resource tend to have consistent key adjustment characteristics, while maximizing the differences in adjustment characteristics between different virtual flexibility resources. For devices that simultaneously belong to multiple virtual flexibility resources and whose membership degrees are all higher than a set threshold, their adjustment capacity is allocated to the corresponding virtual flexibility resources according to their membership degree ratio. The objective function of the fuzzy subspace clustering is: in, For equipment i Belongs to Virtual Flexibility Resources c Membership degree; For virtual flexibility resources c In the f Cluster centers on dimensional features; The feature weight of the virtual flexibility resource c on the f-th feature dimension; m For fuzzy index; C Total number of virtual flexibility resources; N Total number of devices; F The number of feature dimensions; This is the weight entropy regularization coefficient; membership degree The iterative update formula is: Virtual flexibility resources c Upper limit of equivalent power regulation for: Equivalent response delay for: in, For equipment i At any moment t Maximum adjustable power; For equipmenti Self-adjusting response delay; Indicates equipment i At any moment t It is in an adjustable operating state; The iteration terminates when both the change in cluster centers and the change in the objective function are below the convergence threshold, and outputs several virtual flexibility resources and their corresponding device member membership matrices. At the system layer, the external characteristics of each virtual flexibility resource are weighted and aggregated according to the membership matrix. The equivalent power adjustment upper bound, equivalent power adjustment lower bound, equivalent ramp rate, equivalent response delay, equivalent continuous adjustment time and equivalent energy state range of each virtual flexibility resource are calculated to form a set of virtual flexibility resources with a unified interface description.

[0010] As a preferred technical solution, the composite similarity is composed of six sub-similarity factors: spatial Euclidean distance, type attribute matching degree, adjustment capacity difference, response time difference, runtime overlap rate, and functional coupling strength, linearly weighted by adaptive weights. The adaptive weights are dynamically adjusted based on the importance ratio of the adjustment capacity dimension to the spatial dimension in the current clustering target. Represented as: in, For the current clustering objective, the first m Sensitivity coefficient for moderating the similarity of items. This refers to the temperature parameter.

[0011] As a preferred technical solution, the hierarchical coordination optimization of multi-agent systems based on attention mechanism weighted reinforcement learning is specifically as follows: The global coordination layer aims at the overall economic efficiency and low carbon emissions of the park. It integrates ultra-short-term forecasts of total load, ultra-short-term forecasts of renewable energy, time-of-use electricity price sequences, carbon price sequences, tie-line power exchange plans, and real-time feedback of carbon potential information from each node. Through a global attention mechanism, it weights the adjustable capacity boundaries reported by each virtual flexibility resource based on their importance, and generates coordinated control instructions for each virtual flexibility resource. The global attention mechanism dynamically allocates attention weights based on the adjustment capacity, response speed, and carbon potential of each virtual flexibility resource at its node. Each agent in the local execution layer takes its own local observation state and the coordinated control command as input. The local observation state includes the current output power, the remaining adjustable upper and lower bounds, the current ramp margin, the remaining sustainable adjustment time, the current state of charge of the energy storage and the real-time carbon potential of the node. Each agent outputs the power adjustment amount, the charge and discharge power set value, the start and stop control signal or the load transfer amount through the strategy network.

[0012] As a preferred technical solution, the reinforcement learning agent is jointly trained using a multi-agent proximal policy optimization algorithm with priority experience replay: During the training phase, the global coordination layer calculates the global reward based on the system response after each agent performs an action. The global reward is a weighted average of three components: an economic cost reward, a low-carbon reward, and a safety constraint reward. The economic cost reward consists of negative values ​​of electricity purchase cost, gas purchase cost, and equipment operation and maintenance cost. The low-carbon reward consists of the negative cumulative value of the product of each node's carbon potential and net load, and is dynamically updated based on real-time feedback of each node's carbon potential information. The safety constraint reward consists of negative quadratic penalties for power balance deviation, equipment output exceeding limits, energy storage state of charge exceeding limits, and ramping exceeding limits. The global reward function... Represented as: in, , , Time periods t The costs of electricity purchase, gas purchase, and operation and maintenance. For nodes n During the period t Real-time carbon potential. For nodes n During the period t The load power, The total number of nodes. Q The total number of constraint terms. No. q Quadratic penalty for violating the constraint. , , The weighting coefficients for the three awards: economic, low-carbon, and safety. No. i The shearing loss function for updating the policy network of each agent is: in, For probability ratios, For the dominant function, For shear hyperparameters; Advantage function Based on global rewards by the global coordination layer With global value function V(s) calculate: in, Let V(s) be the discount factor, and V(s) be the global state value function. During the distributed execution phase, each agent makes independent decisions based on its local observations and the current coordination instructions issued by the global coordination layer. The global coordination layer verifies the feasibility of each agent's decision-making actions and performs minimum adjustment projection corrections on actions that violate power balance constraints or equipment operation constraints, forming an executable sequence of scheduling instructions.

[0013] As a preferred technical solution, the dual-dimensional attention prediction model that integrates temporal attention and spatial attention adopts an encoder-decoder architecture; The encoder consists of a temporal attention module and a spatial attention module in parallel. The temporal attention module captures long-range temporal dependencies in the input feature sequence based on a multi-head self-attention mechanism, while the spatial attention module captures spatial coupling relationships between different energy devices based on a graph attention mechanism. The outputs of the temporal attention module and the spatial attention module are fused through cross-attention to form a comprehensive encoded feature. The decoder takes the comprehensive coding features as input and generates predicted net load values ​​and tie-line switching power values ​​for multiple future sampling periods in an autoregressive manner.

[0014] As a preferred technical solution, the dynamic graph attention carbon potential perception model constructed based on the actual topology of the park is specifically implemented as follows: The topology of the power grid in the park is defined as a dynamic graph structure. Graph nodes include busbars, feeders and key equipment connection points, and graph edges include transmission and distribution lines and transformer branches. The initial feature vector of each graph node consists of its injected active power, injected reactive power, node type code and the number of edges connected to the node. The initial feature vector of each graph edge consists of its resistance, reactance and current transmission power. The dynamic graph attention carbon potential sensing model adopts a multi-layer temporal graph attention convolutional architecture. In each layer, each node aggregates the feature information of its neighboring nodes through a graph attention mechanism. The graph attention mechanism calculates an independent attention coefficient for each pair of adjacent nodes. The attention coefficient is jointly determined by the feature vectors of the two nodes, the feature vector of the connecting edge, and the current time period. A gated recurrent unit is introduced to recursively update the hidden state of each node along the time dimension, so that the carbon potential sensing result evolves dynamically with the time-varying system operating state. After multi-layer graph attention convolution and gated recurrent unit recursion, the carbon potential prediction value of each node in the current and several future time periods is obtained through the fully connected regression output layer. The real-time carbon emissions corresponding to the power consumption of any group of load nodes in any time period are given by the product of the node's carbon potential prediction value and the load power.

[0015] As a preferred technical solution, the shrinking time-domain model prediction control framework undergoes rolling optimization, as follows: The shrinking time-domain model prediction control framework is optimized in a rolling manner. At each scheduling moment, a fixed length of the prediction time domain is truncated forward from the current moment. The distributed execution process is performed to obtain the scheduling instruction sequence for each time period within the prediction time domain. Only the instruction for the current scheduling time period in the scheduling instruction sequence is sent to the field execution terminal. After the current scheduling period ends, the actual operating status of each device in the park, actual load demand, actual renewable energy output, actual node carbon potential and the latest electricity price and carbon price signals are collected again, the prediction time domain starting point is updated, and the above rolling optimization process is repeated. During the rolling optimization process, the policy networks of each agent that have been trained are fine-tuned online: a small batch of samples is formed using the actual running data of several recent scheduling periods, and the temporal difference error between the cumulative reward after actual execution and the predicted reward is calculated; the temporal difference error is used to update the policy network parameters with a small step gradient, so that the agent policy can continuously adapt to the changes in the current operating conditions.

[0016] As a preferred technical solution, the flexibility boundary of the park system is the power limit and duration that can be adjusted upwards / downwards after the aggregation of each virtual flexibility resource; The optimal operating strategy trajectory consists of the output setpoints of each device in each scheduling period, the energy storage charging and discharging plan, the adjustable load adjustment amount, and the tie-line power exchange plan. The energy-saving and carbon-reduction potential optimized scheduling scheme, compared with the baseline operation scheme, includes reductions in energy costs, primary energy consumption, renewable energy curtailment rates, and total carbon emission reductions. The total carbon emission reduction is statistically analyzed for each load node, forming a node-level carbon emission reduction contribution distribution map. The total carbon emission reduction... The calculation is as follows: in, , Nodes in the baseline scenario n During the period t Carbon potential and load power, , Nodes in optimized scenarios n During the period t Carbon potential and load power; This represents the total number of scheduling periods; This represents the total number of nodes.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention achieves accurate multimodal modeling of heterogeneous and flexible resources. Differentiated feature extraction models are constructed for four types of resources: load, energy storage, energy conversion equipment, and renewable energy. For the first time, the nonlinear coupling relationship of energy storage efficiency, the user comfort relaxation factor of adjustable load, and the full-condition efficiency response surface of energy conversion equipment are introduced into the park control model. Compared with the existing approach of directly setting equipment as intelligent agents, this invention has a significant advantage in the accuracy of equipment characteristic characterization.

[0018] 2) This invention achieves efficient and scalable aggregation of large-scale resources. By improving a two-stage strategy that combines density peak clustering and fuzzy subspace clustering, and by using adaptive weight adjustment based on composite similarity, it achieves accurate mapping from a massive number of dispersed devices to a limited number of virtual flexible resources. Compared to existing static partitioning methods that rely solely on hierarchical data association, the dynamic adaptive clustering of this invention can adapt to aggregation needs in different scenarios, and the fuzzy membership mechanism avoids capacity truncation errors caused by hard clustering.

[0019] 3) This invention achieves a deep closed-loop coupling between carbon flow perception and optimization decision-making. Existing technologies treat carbon flow prediction and scheduling optimization as independent modules. This invention, for the first time, directly embeds the real-time perceived node carbon potential into the low-carbon term of the reward function of multi-agent reinforcement learning, and forms a closed-loop feedback of perception → decision → execution → re-perception through rolling optimization and online fine-tuning mechanisms. This enables the scheduling strategy to proactively adjust electricity consumption behavior based on carbon potential signals, rather than calculating carbon emissions after the fact.

[0020] 4) This invention significantly improves the spatiotemporal accuracy of prediction and perception. The proposed temporal-spatial dual-dimensional attention fusion prediction model simultaneously captures temporal patterns and spatial topological relationships, achieving better prediction accuracy than models using only a single attention mechanism. The dynamic graph attention carbon potential perception model introduces temporal recursion and edge feature-dependent attention coefficient calculation on the basis of traditional graph neural networks, which can more accurately reflect the nonlinear transmission characteristics of carbon flow in complex topologies.

[0021] 5) This invention forms an engineering-deployable rolling optimization closed-loop system. By organically combining a shrinking time-domain model predictive control framework with an online policy fine-tuning mechanism, this invention ensures both the real-time nature of optimization decisions (distributed execution) and the adaptability for long-term operation (online fine-tuning), avoiding the problem of offline training models becoming disconnected from the actual operating environment in existing methods, and has direct engineering application value. Attached Figure Description

[0022] Figure 1 This is a flowchart of an adaptive regulation method for low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling, according to the present invention.

[0023] Figure 2This is a schematic diagram of a flexible resource feature extraction process driven by multimodal data.

[0024] Figure 3 This is a schematic diagram of a two-stage clustering aggregation process based on composite similarity metrics.

[0025] Figure 4 This is a diagram of a hierarchical coordination optimization architecture based on attention-weighted multi-agent reinforcement learning.

[0026] Figure 5 This is a schematic diagram of the temporal-spatial dual-dimensional attention fusion prediction model.

[0027] Figure 6 This is a schematic diagram illustrating message passing and updating in a dynamic graph attention carbon potential perception model.

[0028] Figure 7 A schematic diagram of the closed-loop rolling optimization and adaptive correction process for energy-carbon dual-flow coupling.

[0029] Figure 8 The configuration parameters for the simulation system are shown in the diagram.

[0030] Figure 9 This is the result of load pattern analysis.

[0031] Figure 10 Models for contributing to renewable energy.

[0032] Figure 11 The results are from the VFR clustering analysis.

[0033] Figure 12 To match renewable energy with load and optimize dispatch results.

[0034] Figure 13 Details of power output and load regulation for each piece of equipment.

[0035] Figure 14 This relates to the composition of operating costs and the distribution of carbon emissions.

[0036] Figure 15 This represents the time series of carbon potential and the distribution of the carbon footprint of the load.

[0037] Figure 16 The composition ratio of energy supply and consumption.

[0038] Figure 17 This is the result of carbon price sensitivity analysis.

[0039] Figure 18 The results are from a sensitivity analysis of renewable energy penetration. Detailed Implementation

[0040] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0041] Example 1 This invention proposes an adaptive regulation method for low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling. It aims to achieve precise modeling of flexible resources, scalable collaborative optimization of large-scale systems, and real-time closed-loop coupled regulation of energy and carbon flows through differentiated dynamic feature extraction, improved composite similarity clustering, attention-weighted multi-agent reinforcement learning, and spatiotemporal attention fusion-based carbon flow perception. For example... Figure 1 As shown, the steps of this method include: Step S1: Multimodal data-driven flexible resource operation modality identification and dynamic feature extraction.

[0042] like Figure 2 As shown, this invention collects multi-source heterogeneous operational data from all energy units within the park using a unified sampling period. Data sources include: power / flow time-series data collected by smart meters, heat meters, and gas meters; operational status codes fed back by equipment controllers; environmental data such as irradiance, wind speed, and temperature provided by weather stations; and time-of-use electricity and carbon price signals provided by market interfaces.

[0043] The collected raw data has issues such as inconsistent timestamps, different units, and outliers, requiring preprocessing to form a standardized time-series feature set. The preprocessing process includes: aligning all data to a unified time axis using cubic spline interpolation; removing outliers based on physically reasonable ranges (e.g., power not exceeding 1.2 times the equipment's rated value, state of charge between 0 and 1); and converting different physical quantities to similar numerical ranges using min-max normalization or Z-score standardization.

[0044] After preprocessing, this invention constructs differentiated feature extraction models for four types of core flexibility resources: (1) Adjustable load and non-adjustable load.

[0045] For adjustable loads (such as air conditioners, electric vehicle charging stations, and industrial interruptible loads), this invention not only extracts their physical adjustment parameters (rated power, adjustment depth, response delay, sustainable adjustment time, and energy state boundary), but also innovatively introduces a user comfort relaxation factor to quantify the degree of acceptable comfort deviation for users during demand response. This relaxation factor is embedded as a continuous variable in the clustering and optimization process, enabling the aggregated virtual resources to reflect the probability distribution of actual callable capacity, rather than simple deterministic upper and lower limits.

[0046] By jointly modeling its physical adjustment parameters with user behavior uncertainties, a feature vector is constructed that includes rated power, adjustment depth, response delay, continuous adjustment capability, energy state boundary, and comfort relaxation factor. This results in a feature vector for the adjustable load. Represented as: in, Rated power, , These represent the maximum adjustable depth upwards and downwards, respectively. To respond to the delay, The duration under maximum adjustable power, , For the energy state boundary, For user comfort relaxation factor, and The larger the value, the greater the deviation from the acceptable level of comfort.

[0047] The equivalent external characteristic parameters after adjustable load cluster aggregation are: in The equipment is in operation. The availability coefficient is based on the user comfort relaxation factor. .

[0048] For non-adjustable loads, the power sequence is decomposed into trend components, periodic components, and random residual components using the local weighted regression scatter smoothing method, and the statistical moments of the residual components are used as uncertainty characteristics.

[0049] (2) Energy conversion equipment.

[0050] For energy conversion equipment such as CCHP, gas boilers, and heat pumps, this invention abandons the traditional constant efficiency assumption and constructs a full-condition efficiency response surface function. It expresses the equipment efficiency as a binary quadratic response surface form of input power and output power, thereby accurately characterizing the efficiency changes of the equipment under partial load conditions. Simultaneously, it extracts dynamic features such as minimum stable output, maximum ramp rate, start-up and shutdown time, and start-up and shutdown energy consumption, enabling the equipment model to possess both steady-state accuracy and dynamic response capability.

[0051] Historical operating data was used to fit the partial load efficiency curve and the full-condition efficiency map. The full-condition efficiency map uses input power and output power as coordinate axes, representing the equipment efficiency as a binary quadratic response surface function. Simultaneously, its minimum stable output, maximum ramp rate, start-up and shutdown time, and start-up and shutdown energy consumption were extracted as dynamic features. The full-condition efficiency response surface function of the energy conversion equipment is as follows: in, For equipment j Operating efficiency For input power, For output power, These are the response surface coefficients obtained by least-squares fitting from historical operating data.

[0052] The dynamic characteristic vector of the energy conversion device is: in, , For minimum and maximum stable output, , The maximum upward and downward gradients, , To minimize continuous operation and downtime, , Energy consumption for a single start-up and shutdown.

[0053] (3) Energy storage device.

[0054] The core challenge of energy storage devices lies in the strong nonlinear coupling between their state of charge (SOC), state of health (SOH), and charge / discharge efficiency. This invention models this coupling relationship as a third-order state-space equation, where charging and discharging efficiency are both nonlinear functions of the current SOC, charge / discharge rate, and operating temperature. The decay of SOH considers both cyclic aging and calendar aging cumulative effects. The state of health is the superposition of cyclic aging and calendar aging cumulative effects. Compared to traditional constant efficiency models, this model more realistically reflects the performance evolution of energy storage systems during long-term operation, avoiding suboptimal scheduling strategies due to efficiency estimation biases. The third-order state-space equation of the energy storage device is as follows: State of charge update: Charge / discharge efficiency function: Health status update: in, For rated capacity, As a benchmark efficiency, For the optimal state of charge, I is the charging and discharging current. Where is the rated current and T is the operating temperature. For optimal operating temperature, The efficiency attenuation coefficient, The aging factor is... , R is the activation energy, and R is the gas constant.

[0055] (4) Renewable energy power generation unit.

[0056] The strong fluctuations in photovoltaic and wind power output pose a major challenge to system regulation. This invention employs variational mode decomposition (VMD) to adaptively decompose the original power output sequence into several eigenmode components with different center frequencies. Fluctuation statistical features are extracted for each component, and a fluctuation mode feature vector is constructed using the instantaneous frequency and instantaneous amplitude of each component. Compared to empirical mode decomposition (EMD), VMD has a more robust mathematical theoretical foundation, effectively avoids mode aliasing problems, and the extracted features are more physically interpretable. The VMD expression for the renewable energy generation unit is: in, For the original output sequence of renewable energy, K The number of intrinsic modal components. For the first k Each intrinsic mode component For its center frequency, For the Dirac function, This is a convolution operation.

[0057] The wave eigenvector of each intrinsic mode component is: in, Standard deviation, The mean of the absolute values ​​of the instantaneous rates of change. The maximum absolute value of the instantaneous rate of change. , These represent the upward and downward ramp rates at a 99% confidence level. This represents the average instantaneous amplitude.

[0058] Step S2: Flexible resource adaptive hierarchical partitioning aggregation based on composite similarity metric.

[0059] After completing the feature extraction of a single device, this invention adopts a three-layer aggregation architecture of device-cluster-system to aggregate a large number of dispersed devices into a limited number of virtual flexibility resources (VFRs) to reduce the dimensionality of subsequent optimization decisions.

[0060] At the device layer, the feature vectors of each flexibility resource extracted in step S1 are filtered by the maximum information coefficient, and redundant features are removed to form a standardized clustering input matrix.

[0061] At the cluster layer, a two-stage clustering strategy based on a composite similarity metric is used for resource aggregation: like Figure 3 As shown, the first stage is to improve density peak clustering, using the composite similarity between devices as a distance measure. This invention defines a composite similarity as the distance measure between devices, which is composed of six sub-similarity factors: spatial Euclidean distance, type attribute matching degree, adjustment capacity difference, response time difference, runtime overlap rate, and functional coupling strength, linearly weighted by adaptive weights. Unlike traditional fixed weights, the weight coefficients in this invention are dynamically and adaptively adjusted according to the importance ratio of the adjustment capacity dimension to the spatial dimension in the current clustering objective, allowing the clustering results to flexibly adapt to different application scenarios (such as regional autonomy scenarios biased towards spatial proximity or scheduling aggregation scenarios biased towards adjustment capacity consistency).

[0062] Composite similarity The calculation formula is: The six sub-similarity values ​​are as follows: in, .

[0063] The adaptive weights are dynamically adjusted based on the ratio of the importance of the adjustment capability dimension to the spatial dimension in the current clustering objective. The adjustment mechanism is as follows: in, Let be the sensitivity coefficient for moderating the similarity of the m-th sub-item under the current clustering objective. This refers to the temperature parameter.

[0064] Based on the composite similarity, the local density of each device and its minimum composite distance with higher density devices are calculated. The product of local density and minimum composite distance is used as the cluster center selection index. The cluster centers are automatically identified and the initial resource partitioning is completed. Devices with a density lower than the global density threshold and a composite distance from each cluster center exceeding a set multiple are marked as boundary devices.

[0065] The second stage involves fuzzy subspace clustering. For the initial resource partition boundaries formed in the first stage, this invention further introduces fuzzy subspace clustering for flexible boundary correction. Unlike traditional fuzzy C-means, which assigns equal weights to all feature dimensions, this invention independently learns weight coefficients for each feature dimension of each cluster, allowing different VFRs to have differentiated focuses on different feature dimensions. For example, one VFR might prioritize capacity consistency, while another might prioritize response time consistency. This differentiated feature weighting makes the clustering results more closely match actual scheduling requirements. During iterative optimization, the membership degree of each device to each VFR is continuously updated. For devices that belong to multiple VFRs and have high membership degrees, their capacity is allocated to each VFR according to the membership degree ratio, thus avoiding capacity truncation errors caused by hard clustering.

[0066] In fuzzy subspace clustering, each feature dimension is assigned an independent weight coefficient, allowing different virtual flexibility resources to have differentiated levels of attention across different feature dimensions. By iteratively optimizing the membership degree of each device to various virtual flexibility resources and the weight coefficients of each feature dimension, devices within the same virtual flexibility resource tend to have consistent key adjustment characteristics, while maximizing the differences in adjustment characteristics between different virtual flexibility resources. For devices that belong to multiple virtual flexibility resources and whose membership degrees are all higher than a set threshold, their adjustment capacity is allocated to the corresponding virtual flexibility resources according to their membership degree ratio. The objective function of fuzzy subspace clustering is: in, Let be the membership matrix, satisfying and ; This is the cluster center matrix; Let be the feature weight matrix, satisfying and m is the fuzziness index; F is the number of feature dimensions; is the weight entropy regularization coefficient.

[0067] Membership degree The iterative update formula is: Upper bound of equivalent power regulation for virtual flexibility resource c for: Equivalent response delay for: The iteration terminates when both the change in cluster centers and the change in the objective function are below the convergence threshold, and outputs several virtual flexibility resources and their corresponding device member membership matrices.

[0068] At the system level, after clustering convergence, the external characteristics of each virtual flexibility resource are weighted and aggregated according to the membership matrix. The equivalent power adjustment upper bound, equivalent power adjustment lower bound, equivalent ramp rate, equivalent response delay, equivalent duration adjustment time, and equivalent energy state interval of each virtual flexibility resource are calculated to form a set of virtual flexibility resources with a unified interface description, providing a standardized agent definition for subsequent multi-agent reinforcement learning.

[0069] Step S3: Hierarchical coordination optimization based on attention mechanism-weighted multi-agent reinforcement learning.

[0070] like Figure 4 As shown, this invention constructs each virtual flexibility resource (VFR) obtained by aggregation into a reinforcement learning agent with an independent policy network and a value network, and adopts a two-layer architecture of global coordination layer and local execution layer for collaborative optimization.

[0071] The global coordination layer does not directly control equipment. Instead, it aims at the overall economic efficiency and low-carbon operation of the park. It integrates ultra-short-term forecasts of the park's total load, ultra-short-term forecasts of renewable energy, time-of-use pricing sequences, carbon price sequences, tie-line power exchange plans, and real-time carbon potential information from each node as reported in step S4. Through a global attention mechanism, it weights the adjustable capacity boundaries reported by each virtual flexibility resource based on their importance, generating coordinated control instructions for each virtual flexibility resource. The global attention mechanism dynamically allocates attention weights based on the adjustment capacity, response speed, and carbon potential of each virtual flexibility resource's node. Virtual flexibility resources (VFRs) with large adjustment capacities, fast response speeds, and high carbon potential at their respective nodes receive higher attention weights and are therefore prioritized during target allocation. This attention-weighted mechanism makes the coordination layer's decisions more targeted, avoiding the pitfalls of traditional average allocation strategies that force a fit.

[0072] In the local execution layer, each agent takes its own local observation state and coordination and control instructions as input. The local observation state includes the current output power, the remaining adjustable upper and lower bounds, the current ramp margin, the remaining sustainable adjustment time, the current state of charge of the energy storage and the real-time carbon potential of the node. Each agent outputs power adjustment amount, charge and discharge power setpoint, start and stop control signal or load transfer amount through the policy network.

[0073] During the training phase, a multi-agent proximal policy optimization algorithm with priority experience replay is used for joint training. During training, the global coordination layer calculates the global reward based on the system response after each agent performs an action. The global reward consists of a weighted average of three items: economic cost reward, low-carbon reward, and safety constraint reward. The economic cost reward consists of the negative values ​​of electricity purchase cost, gas purchase cost, and equipment operation and maintenance cost. The low-carbon reward consists of the negative cumulative value of the product of each node's carbon potential and the node's net load, and is dynamically refreshed with the carbon potential sensed in real time during step S4. This allows the agents to not only learn how to reduce energy costs during training but also how to proactively adjust their electricity consumption behavior based on carbon potential signals—this is one of the core innovations that distinguishes this invention from existing technologies.

[0074] The safety constraint reward consists of negative quadratic penalties for power balance deviation, equipment output exceeding limits, energy storage state of charge exceeding limits, and ramp exceeding limits.

[0075] Global reward function Represented as: in, , , These represent the electricity purchase cost, gas purchase cost, and operation and maintenance cost for time period t. Let be the real-time carbon potential of node n in time period t. Let n be the load power of node n in time period t. Let Q be the total number of nodes and Q be the total number of constraint terms. The penalty for violating the constraint in the q-th term is a quadratic form. , , The weighting coefficients for the three awards: economy, low carbon, and safety.

[0076] In the multi-agent proximal policy optimization algorithm with priority experience replay, the first i The shearing loss function for updating the policy network of each agent is: in, For probability ratios, For the dominant function, This refers to the shear hyperparameter.

[0077] Advantage function Based on global rewards by the global coordination layer Calculated with the global value function V(s): in, V(s) is the discount factor, and V(s) is the global state value function.

[0078] After training converges, the process transitions to the distributed execution phase. In this phase, each agent makes independent decisions based on its local observations and the current coordination instructions issued by the global coordination layer. The global coordination layer only verifies the feasibility of each agent's decision and performs minimum adjustment projection corrections on actions that violate power balance constraints or equipment operation constraints, forming an executable sequence of scheduling instructions to ensure execution safety.

[0079] Step S4: Interactive behavior prediction and real-time carbon flow perception based on temporal-spatial dual-dimensional attention fusion.

[0080] like Figure 5 As shown, this invention constructs a two-dimensional attention prediction model that integrates temporal attention and spatial attention to perform ultra-short-term predictions of campus net load and tie-line switching power. The prediction model uses an encoder-decoder architecture, with the encoder consisting of a temporal attention module and a spatial attention module operating in parallel.

[0081] The temporal attention module captures long-range temporal dependencies in the input feature sequence based on a multi-head self-attention mechanism. The multi-head self-attention calculation of the temporal attention module is as follows: in, Q, K, V These are the query matrix, key matrix, and value matrix, respectively. Let be the dimension of the key vector. H For the number of attention heads.

[0082] The spatial attention module captures the spatial coupling relationships between different energy devices based on a graph attention mechanism. The graph attention coefficients of the spatial attention module... for: in, , For nodes i, j The hidden state, For the edge ij eigenvectors, , The weight matrix is ​​a learnable matrix. For attention parameter vectors, This represents a vector concatenation operation. For nodes i The set of neighbors.

[0083] The outputs of the two modules are fused through cross-attention to form a comprehensive encoded feature. The decoder uses this comprehensive encoded feature as input and generates predicted net load and tie-line switching power for multiple future sampling periods using an autoregressive approach. Compared to models using only temporal attention or only spatial attention, the two-dimensional fusion model can simultaneously utilize temporal patterns and spatial topological information, significantly improving prediction accuracy.

[0084] Meanwhile, this invention constructs a dynamic graph attention carbon potential sensing model based on the actual topology of the park. For example... Figure 6 As shown, the topology of the park's power grid is defined as a dynamic graph structure. Graph nodes include busbars, feeders, and key equipment connection points, while graph edges include transmission and distribution lines and transformer branches. The initial feature vector of each graph node consists of its injected active power, injected reactive power, node type code, and the number of edges connected to that node. The initial feature vector of each graph edge consists of its resistance, reactance, and current transmission power.

[0085] The dynamic graph attention carbon potential sensing model employs a multi-layer temporal graph attention convolutional architecture. In each layer, each node aggregates the feature information of its neighboring nodes through a graph attention mechanism. The graph attention mechanism calculates independent attention coefficients for each pair of adjacent nodes. These coefficients are determined by the feature vectors of the two nodes, the feature vectors of the connecting edges, and the current time period, enabling differentiated modeling of the carbon flow's propagation direction and intensity within the topology. Unlike traditional graph convolution, which assigns equal weights to all neighbors, the graph attention coefficients in this invention simultaneously depend on the features of the two endpoints, the features of the connecting edges, and the current time period, resulting in refined modeling of the carbon flow's propagation direction and intensity within the topology. Furthermore, a gated recurrent unit is introduced to recursively update the hidden states of each node along the time dimension, allowing the carbon potential sensing results to dynamically evolve with the system's operating state, rather than relying solely on static calculations at the current cross-section. The dynamic graph attention carbon potential sensing model includes nodes... i exist l The update for +1 level is: in, Update function for gated loop unit Let be the graph attention coefficient of the l-th layer. For the first l The transformation matrix of the layer.

[0086] After multi-layer graph attention convolution and gated recurrent units, the carbon potential prediction values ​​of each node for the current and several future time periods are obtained through a fully connected regression output layer. The real-time carbon emissions corresponding to the electricity consumption of any group of load nodes in any time period are given by the product of the node's carbon potential prediction value and the load power, thereby realizing the end-to-end carbon emission responsibility tracing from the generation side to the load side and providing node-level fine data for subsequent carbon emission reduction contribution assessment.

[0087] Step S5: Closed-loop rolling optimization and adaptive correction of energy-carbon dual-flow coupling.

[0088] This invention embeds the output of the above prediction and perception modules into the optimization decision module in real time, forming a complete closed-loop feedback chain: the carbon potential perception module feeds back the current and future carbon potential of each node to the low-carbon term of the reward function of multi-agent reinforcement learning; the net load prediction module feeds back the future net load curve to the constraints and target settings of the global coordination layer.

[0089] like Figure 7 As shown, the predicted net load of the park and the predicted switching power of the tie line obtained in step S4 are used as the forward input of the global coordination layer in step S3. The predicted carbon potential of each node in real time in step S4 is used as the basis for real-time refresh of the low-carbon reward item of the global reward function in step S3, forming a closed-loop feedback link of "carbon potential perception → optimization decision → action execution → carbon potential re-perception".

[0090] Based on the aforementioned closed-loop feedback loop, a shrinking time-domain model predictive control framework is used for rolling optimization. At each scheduling moment, a fixed-length prediction time domain is extracted from the current moment, and the distributed execution process in step S3 is executed to obtain the scheduling instruction sequence for each time period within this prediction time domain. However, only the instructions for the current scheduling time period in the scheduling instruction sequence are sent to the field execution terminal. After the current scheduling time period ends, the actual operating status of each device in the park, actual load demand, actual renewable energy output, actual node carbon potential, and the latest electricity price and carbon price signals are collected again to update the starting point of the prediction time domain, and the above rolling optimization process is repeated. This rolling execution method can continuously use the latest measured data to correct prediction deviations and enhance the system's adaptability to new energy fluctuations and load uncertainties.

[0091] This invention further introduces an online policy fine-tuning mechanism during the rolling optimization process, performing online fine-tuning on the policy networks of each agent trained in step S3: using actual operating data from several recent scheduling periods to form a small batch of samples, the temporal difference error between the actual reward after execution and the predicted reward before execution is calculated. This error is then used to update the policy network parameters with gradients of a small step size, enabling the agent policy to continuously adapt to changes in the current operating conditions. This online fine-tuning allows the pre-trained policy network to continuously adapt to slow changes in the current operating conditions (such as equipment aging, seasonal load changes, etc.) without frequent offline retraining.

[0092] Among them, timing difference error for: in, This refers to the actual reward obtained after execution during the current period. Output to the target value network. This is the output of the current value network.

[0093] The online update step size for the policy network parameters is: in, For online learning rate, The value of is between 1 / 10 and 1 / 5 of the learning rate during the training phase.

[0094] Ultimately, this method outputs three types of quantization results: (1) The flexibility boundary of the park system, namely the power limit and duration that can be adjusted upward / downward after the aggregation of each virtual flexibility resource, provides a capacity basis for the park to participate in the grid demand response.

[0095] (2) The optimal operation strategy trajectory of the park, namely the output setting value of each equipment, the energy storage charging and discharging plan, the adjustable load adjustment amount and the tie line power exchange plan for each scheduling period, can be directly issued and executed.

[0096] (3) Assessment of the park's node-level energy conservation and carbon reduction potential, namely, the reduction in energy costs, primary energy consumption, renewable energy curtailment rate, and total carbon emission reduction compared to the baseline operation plan under the optimized scheduling plan. The total carbon emission reduction is calculated separately for each load node, forming a node-level carbon emission reduction contribution distribution map, providing refined data support for the park's carbon asset management. Total carbon emission reduction The calculation method is as follows: in, , Nodes in the baseline scenario n During the period t Carbon potential and load power, , These are the corresponding values ​​for the optimized scenarios.

[0097] Example 2 This embodiment uses a comprehensive low-carbon industrial park encompassing industrial manufacturing, commercial offices, and residential services as its application scenario. The park's main equipment configuration is as follows: 500 kW photovoltaic capacity, 300 kW wind power capacity, a CCHP unit with a maximum generating capacity of 300 kW (minimum stable output of 50 kW), a gas-fired boiler with a rated thermal power of 400 kW, an electrochemical energy storage system with a rated capacity of 400 kWh (maximum charge / discharge power of 100 kW), a base load peak of approximately 800 kW, and an adjustable load capacity of approximately 200 kW. The park is connected to a 10 kV distribution network. The time-of-use electricity price is 0.80 yuan / kWh during peak hours (08:00-11:00, 18:00-21:00), 0.25 yuan / kWh during off-peak hours (22:00-06:00), and 0.50 yuan / kWh during normal hours. The carbon market reference price is 80 yuan / ton CO2. Figure 8 The system equipment capacity configuration, time-of-use pricing mechanism, and key operating parameters were showcased. Figure 8 (a) The peak-valley price difference is 3.2 times, providing an arbitrage opportunity of about RMB 2,800 per day for energy storage systems; Figure 8 (b) Provide the power range and efficiency benchmarks for each device; Figure 8 (c) The average carbon emission factor of the power grid is 0.80 kgCO2 / kWh, and the carbon cost accounts for about 14.3% of the total cost under the carbon price level, which is enough to have a significant impact on dispatch decisions; Figure 8 (d) The maximum ramp rate of CCHP is 60 kW / 15min, which limits the adjustment speed when the load fluctuates rapidly.

[0098] Step 1: Multimodal data-driven flexible resource operation modality identification and dynamic feature extraction.

[0099] Operational data was collected continuously for 30 days with a sampling period of 15 minutes. After timestamp alignment, outlier removal, and unit normalization of the raw data, features were extracted according to equipment type.

[0100] (1) Adjustable load feature extraction. Feature vector of adjustable load. The structure is as follows: In this embodiment, the maximum upward adjustment depth is 3 kW, and the downward adjustment depth is 2 kW. The response delay is approximately 2 minutes, and it can operate continuously for 30 minutes. The user comfort relaxation factor is also considered. The equivalent external characteristic parameters after aggregation of multiple adjustable load clusters are: (2) Energy storage characteristic extraction. The rated capacity of the energy storage system is 400 kWh, and the current SOC is 65%. Calculations are made based on the efficiency function: Pick =0.94, =0.5, λ1=0.3, λ2=0.1, λ3=0.05, , The current charging efficiency is calculated to be approximately 87.3%, significantly better than the 94% assumed by the traditional constant efficiency model. The SOC for the next period is updated to 69.4%. (3) Extraction of new energy fluctuation characteristics. Using the photovoltaic power output sequence (96 sampling points) as input, variational mode decomposition is adopted, and the optimization model is as follows: By setting K=4, the intrinsic mode components with different center frequencies are obtained by decomposition, and the wave characteristic vectors of each component are extracted.

[0101] Figure 9 The results of the load pattern analysis are presented. Figure 9 (a) The weekday load shows a significant bimodal characteristic, with a peak of approximately 800 kW in the morning (08:00-11:00) and a peak of approximately 750 kW in the evening (18:00-21:00), dropping back to approximately 600 kW at midday. Weekend load is about 25% lower than the weekday average. Figure 9 (b) The total load is decomposed into non-adjustable load (approximately 80%, daily electricity consumption 8,800 kWh) and adjustable load baseline (approximately 20%, daily electricity consumption 2,400 kWh). The adjustable load has a maximum upward adjustment potential of approximately 22 kW and a reduction capacity of approximately 150 kW during peak electricity price periods.

[0102] Figure 10 It showcased renewable energy power output models. Figure 10 (a) It shows that the photovoltaic power reached its peak of about 480 kW around 12:00, with a daily power generation of about 3,200 kWh; the wind power was relatively stable throughout the day, with a higher output at night, peaking at about 270 kW, with a daily power generation of about 2,200 kWh. Figure 10 (b) The stacked area diagram shows the complementary characteristics of the two on the time scale, which narrows the fluctuation of the overall output throughout the day from ±50% of the single energy source to ±25%. Figure 10 (c)(d) shows that the total daily power generation from renewable energy sources is approximately 5,400 kWh, accounting for 35.2% of the park's total daily electricity consumption, of which photovoltaic power accounts for 62.5% and wind power accounts for 37.5%.

[0103] Step 2: Flexible resource adaptive hierarchical partitioning aggregation based on composite similarity metric.

[0104] There are 62 controllable devices in the park, which were uniformly reduced to 10 core features after screening based on the maximum information coefficient. Improved density peak clustering was used for the first stage of clustering, and the composite similarity was calculated as follows: The six sub-similarity metrics are spatial Euclidean distance, type attribute matching degree, adjustment capacity difference, response time difference, runtime overlap rate complementarity, and functional coupling strength complementarity. This embodiment emphasizes the consistency of adjustment capacity, with adaptive weights set to w1=0.15, w2=0.10, w3=0.30, w4=0.20, w5=0.15, and w6=0.10. The density threshold is set to 0.55, automatically identifying five initial cluster centers.

[0105] After initial partitioning, the fuzzy subspace clustering stage begins, with the objective function as follows: The membership degree iterative update formula is as follows: Calculate the upper bound of the equivalent power regulation of VFR and the equivalent response delay respectively.

[0106] Figure 11 The cluster analysis results are presented. Figure 11 (a) Three main clusters were formed in the dimensionality reduction scatter plot: VFR1 (adjustable load cluster) has a fast response speed (2-3 minutes) and a short adjustment time (15-30 minutes); VFR2 (energy storage and CCHP combination) has a large capacity adjustment capability (aggregate power of about 120 kW) and medium speed response characteristics; VFR3 (new energy and gas boiler combination) has a slower response speed but a long sustainable adjustment time. Figure 11 (b) The characteristic parameters of each VFR were quantified: VFR1 has a combined regulation capacity of 150 kW / 100 kW and a response delay of 3 minutes; VFR2 has a capacity of 120 kW / 80 kW and a response delay of 5 minutes; VFR3 has a capacity of 80 kW / 50 kW and a response delay of 8 minutes. Figure 11 (c) and (d) show that the aggregated capacities of each VFR are 200 kW, 180 kW, and 120 kW, with available confidence levels of 0.90, 0.85, and 0.80, respectively. Through VFR aggregation, the optimization problem of 62 distributed devices is simplified to a collaborative decision-making problem of 5 VFR agents, reducing the state space dimension by approximately 90%.

[0107] Step 3: Hierarchical coordination optimization based on attention mechanism-weighted multi-agent reinforcement learning.

[0108] The five VFRs are modeled as five agents, employing a two-layer architecture consisting of a global coordination layer and local execution layers. The global reward function is as follows: set up , , A multi-agent proximal policy optimization algorithm with priority experience replay is adopted, and the policy network shearing loss function is as follows: Advantage function The discount factor γ = 0.95 and the shearing hyperparameter ε = 0.2. After 1200 rounds of training, the cumulative reward of each agent tends to stabilize.

[0109] Figure 12 The results of matching renewable energy with load and optimizing scheduling were presented. Figure 12 (a) It shows that during the midday period from 12:00 to 14:00, the output of new energy exceeds 600 kW, and the load is about 620 kW, with the two being close to balanced; there is a power gap of about 300-400 kW during the morning and evening peak periods. Figure 12 (b) shows the output stacking diagram of each device after optimized scheduling: CCHP increases its output to 200-250 kW during peak electricity price periods; photovoltaic power is fully absorbed during the day; energy storage charges during valley periods and discharges during peak periods. Figure 12 (c) It shows that the power purchased during the peak photovoltaic period during the day drops to a minimum of about 100 kW, and even a short-term power purchase of about 20 kW occurs between 12:30 and 13:00. The total power purchased throughout the day is about 6,200 kWh, accounting for 42.3% of the total power consumption of the park. Figure 12 (d) is the energy storage SOC curve. The system is charged at about 80 kW during the valley period from 02:00 to 06:00, and the SOC rises from 50% to 85%. It is discharged to 55% during the peak period and discharged again to 30% during the evening peak, completing two complete charge and discharge cycles throughout the day.

[0110] Figure 13 Further refine the output and load adjustment details of each piece of equipment. Figure 13 (a) The stacked diagram breaks down the power output of each device: CCHP maintains 220 kW from 08:00 to 11:00 and reaches 260 kW from 18:00 to 20:00; PV reaches its peak of 480 kW at 12:00; and energy storage discharges 90 kW from 18:00 to 19:00. Figure 13 (b) Demonstrates that the adjustable load is reduced by 28 kW and 30 kW during the morning and evening peak hours, respectively, with a response delay of less than 2 minutes, a power tracking error of less than 5 kW, and a total daily load adjustment of approximately 420 kWh. Figure 13(c) Displays precise control of energy storage charging and discharging power, with a power tracking error of less than 3 kW during frequent switching. Figure 13 (d) shows the CCHP operating power curve. The morning peak ramp rate is approximately 56 kW / 15min, and the evening peak is approximately 54 kW / 15min, both meeting the 60 kW / 15min ramp constraint. The total daily power generation is approximately 4,300 kWh, and the gas consumption is approximately 1,230 m³. 3 .

[0111] Step 4: Interactive behavior prediction and real-time carbon flow perception based on temporal-spatial dual-dimensional attention fusion.

[0112] In the two-dimensional attention prediction model, the multi-head self-attention of the temporal attention module is calculated using the following formula: The graph attention coefficient of the spatial attention module is calculated using the following formula: The model outputs a mean absolute percentage error of 1.8% for the net load forecast for the next 4 hours, which is better than the baseline model (4.5%).

[0113] In the carbon potential sensing model, node updates are calculated using the following formula: Figure 15 The time series of carbon potential and the distribution of carbon footprint load are presented. Figure 15 (a) The data shows that the carbon potential is lowest at about 0.45 kgCO2 / kWh during the trough period from 02:00 to 06:00, rises to 0.65 kgCO2 / kWh during the early morning peak from 08:00 to 11:00, and reaches a peak of 0.75 kgCO2 / kWh during the evening peak from 18:00 to 21:00, with a peak-to-trough difference of 0.30 kgCO2 / kWh. Figure 15 (b) shows that industrial load accounts for 48.4% of carbon emissions (carbon potential 0.62), commercial load accounts for 24.6% (carbon potential 0.58), and office load accounts for 27.0% (carbon potential 0.55). Figure 15 (c) The differences in carbon potential at each node were verified. Figure 15 (d) is the cumulative carbon emission curve for the whole day, which increases monotonically from 0 to 8,300 kgCO2. The slope increases significantly during the morning and evening peak hours and slows down at noon.

[0114] Step 5: Closed-loop rolling optimization and adaptive correction of energy-carbon dual-flow coupling.

[0115] Starting at 10:00, a 4-hour time domain prediction is performed. After optimization, only instructions for the current time period are issued. At the next moment, the actual state is re-acquired, the prediction time domain starting point is updated, and the solution is recalculated. The time-series difference error during online fine-tuning is calculated using the following formula: The online update step size of the policy network is calculated using the following formula: The total reduction in carbon emissions is calculated using the following formula: After continuous optimization throughout the day, the results are as follows: Figure 14 As shown in Table 1. Figure 14 (a) The total daily operating cost was RMB 11,200, with net electricity purchase cost accounting for 55.8% (approximately RMB 6,250), gas purchase cost accounting for 20.1% (approximately RMB 2,250), carbon cost accounting for 14.3% (approximately RMB 1,600), and operation and maintenance cost accounting for 9.8% (approximately RMB 1,100). Compared with the baseline scenario, the total cost decreased by 18.7%, mainly due to peak-valley arbitrage of energy storage (contributing approximately RMB 2,800) and substitution of electricity purchase during high-price periods of CCHP (contributing approximately RMB 1,200). Figure 14 (b) Total carbon emissions amounted to 8,300 kg CO2 / day, of which 6,200 kg CO2 / day (74.7%) was indirectly emitted from grid-purchased electricity, 2,100 kg CO2 / day (25.3%) was directly emitted from CCHP, and approximately 2,800 kg CO2 / day was avoided from solar and wind power. Table 1 shows that the renewable energy penetration rate increased from 28% to 35.2%, the grid-purchased electricity ratio decreased from 55% to 42.3%, and the carbon emission intensity decreased from 0.80 to 0.58 kg CO2 / kWh.

[0116] Table 1 Comparison of key system performance indicators before and after optimization Figure 16 It shows the composition ratio of energy supply and consumption. Figure 16 (a) The energy supply composition is as follows: renewable energy 35.2% (approximately 5,400 kWh), CCHP 28.5% (approximately 4,300 kWh), grid-purchased electricity 31.8% (approximately 4,800 kWh), and energy storage discharge 4.5% (approximately 680 kWh), with clean energy supply accounting for a total of 63.7%. Figure 16 (b) Energy consumption is comprised of: base load 42.3% (approximately 6,500 kWh), adjustable load 25.8% (approximately 3,900 kWh), energy storage charging 18.2% (approximately 2,800 kWh), and grid sales 13.7% (approximately 2,100 kWh), with total daily supply and demand both at 15,180 kWh, which is strictly balanced.

[0117] Carbon price sensitivity, such as Figure 17 As shown. Figure 17 (a)(b) shows that when the carbon price rises from 0 yuan / ton to 200 yuan / ton, the operating cost increases from about 12,000 yuan to about 18,000 yuan, and carbon emissions decrease from 9,000 kgCO2 / day to about 7,200 kgCO2 / day. The emission reduction effect is most significant in the range of 50-100 yuan / ton. Figure 17 (c)(d) show that renewable energy penetration increased from 30% to approximately 42%, while carbon intensity decreased from 0.75 kgCO2 / kWh to 0.45 kgCO2 / kWh. A carbon price of 80-120 yuan / ton is recommended.

[0118] Renewable energy penetration sensitivity, such as Figure 18 As shown. Figure 18 (a) It shows that when the penetration rate increases from 20% to 60%, the operating cost first decreases and then increases, reaching its lowest point in the 35%-40% range (approximately RMB 11,000). Figure 18 (b) Shows that carbon emissions decreased linearly from 9,500 kgCO2 / day to approximately 6,200 kgCO2 / day; Figure 18 (c)(d) show that the electricity purchase ratio decreased from 70% to approximately 28%, and the carbon intensity decreased from 0.74 to 0.42 kgCO2 / kWh. It is recommended that the penetration rate be controlled within the range of 35%-45%.

[0119] The sensitivity of energy storage capacity is shown in Table 2. When energy storage increases from 200 kWh to 800 kWh, the total cost first decreases and then increases, reaching its lowest point at 600 kWh (RMB 10,950 / day); carbon emissions decrease from 8,750 kgCO2 / day to 7,950 kgCO2 / day; and the renewable energy utilization rate increases from 88.5% to 97.5%. It is recommended that energy storage capacity be configured in the range of 400-600 kWh.

[0120] Table 2 Energy Storage Capacity Sensitivity Analysis In summary, the adaptive regulation method for low-carbon industrial parks proposed in this invention, based on multi-source data fusion and energy-carbon dual-flow coupling, achieves a technological leap in the park's energy system from "static modeling → open-loop scheduling → delayed accounting" to "dynamic perception → closed-loop optimization → real-time source tracing" through refined feature extraction, adaptive clustering and aggregation, attention-weighted multi-agent reinforcement learning, dual-dimensional attention prediction and dynamic graph attention carbon potential perception, as well as closed-loop rolling optimization and online fine-tuning. It has achieved significant results in reducing energy costs, reducing carbon emissions, and improving the absorption of renewable energy.

[0121] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0122] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for regulating low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling, characterized by the following steps: include: Collect multi-source heterogeneous operation data of various flexible resources in the low-carbon park and extract feature vectors; The feature vectors of each flexibility resource are input into a three-layer aggregation architecture that includes the device layer, cluster layer and system layer. Based on the composite similarity metric, the flexibility resources are hierarchically and partitioned to aggregate virtual flexibility resources. Each virtual flexibility resource is constructed as an independent reinforcement learning agent, and hierarchical coordination optimization based on attention mechanism weighted reinforcement learning is carried out. The global coordination layer aims at the overall economic efficiency and low carbon emissions of the park and generates coordination and control instructions for each virtual flexibility resource. In the local execution layer, each agent takes its own local observation state and coordination and control instructions as input and outputs decision actions. A dual-dimensional attention prediction model integrating temporal and spatial attention is adopted to generate predicted net load values ​​and tie-line switching power values ​​for multiple future sampling periods; a dynamic graph attention carbon potential perception model is constructed based on the actual topology of the park to generate predicted carbon potential values ​​for each node. The predicted net load of the park and the predicted switching power of the tie line are used as the forward input of the global coordination layer, and the real-time carbon potential prediction of each node is used as the basis for real-time updating of the low-carbon reward item of the global reward function. A closed-loop feedback link is formed and a shrinking time-domain model predictive control framework is used for rolling optimization. Finally, the flexibility boundary of the park system, the optimal operating strategy trajectory, and the energy-saving and carbon-reducing potential are output.

2. The low-carbon industrial park regulation method based on multi-source data fusion and energy-carbon dual-flow coupling according to claim 1, characterized in that, The flexible resources include: adjustable and non-adjustable loads, energy conversion equipment, energy storage devices, and renewable energy generation units; For adjustable loads, their physical adjustment parameters are jointly modeled with user behavior uncertainties to construct a feature vector that includes rated power, adjustment depth, response delay, continuous adjustment capability, energy state boundary, and comfort relaxation factor. For non-adjustable loads, the power sequence is decomposed into trend components, periodic components, and random residual components using the local weighted regression scatter smoothing method, and the statistical moments of the residual components are used as uncertainty features. For energy conversion equipment, the load efficiency curve and the full-condition efficiency map are fitted with historical operating data. The full-condition efficiency map uses input power and output power as coordinate axes to represent the equipment efficiency as a binary quadratic response surface function. At the same time, the minimum stable output, maximum ramp rate, start-up and shutdown time and start-up and shutdown energy consumption of the energy conversion equipment are extracted as dynamic features. For energy storage devices, the nonlinear coupling relationship between the state of charge, health state and charge / discharge efficiency of the energy storage device is modeled as a third-order state-space equation. In the third-order state-space equation, the charging efficiency and discharging efficiency are nonlinear functions of the current state of charge, charge / discharge rate and operating temperature, respectively, and the health state is the superposition of the cumulative effect of cyclic aging and the cumulative effect of calendar aging. For a renewable energy power generation unit, variational mode decomposition is used to decouple the output sequence of the renewable energy power generation unit into multiple intrinsic mode components with different center frequencies. Fluctuation statistical features are extracted for each intrinsic mode component, and the instantaneous frequency and instantaneous amplitude of each component constitute a fluctuation mode feature vector.

3. The low-carbon industrial park regulation method based on multi-source data fusion and energy-carbon dual-flow coupling according to claim 1, characterized in that, In the aforementioned three-tier aggregation architecture of device-cluster-system: At the device layer, the maximum information coefficient is used to filter the extracted feature vectors of each flexibility resource, and redundant features are removed before being used as the clustering input matrix. At the cluster layer, a two-stage clustering strategy based on composite similarity metrics is used for resource aggregation. The first stage is to improve density peak clustering. The composite similarity between devices is used as a distance measure to calculate the local density of each device and its minimum composite distance with higher density devices. The product of local density and minimum composite distance is used as the cluster center selection index to identify cluster centers and complete the initial resource partitioning. Devices with a density lower than the global density threshold and a composite distance from each cluster center exceeding a set multiple are marked as boundary devices. The second stage is fuzzy subspace clustering, which flexibly corrects the initial resource partition boundaries formed in the first stage. In fuzzy subspace clustering, each feature dimension is assigned an independent weight coefficient, allowing different virtual flexibility resources to have differentiated levels of attention across different feature dimensions. Through iterative optimization of the membership degree of each device to various virtual flexibility resources and the weight coefficients of each feature dimension, devices within the same virtual flexibility resource tend to have consistent key adjustment characteristics, while maximizing the differences in adjustment characteristics between different virtual flexibility resources. For devices that simultaneously belong to multiple virtual flexibility resources and whose membership degrees are all higher than a set threshold, their adjustment capacity is allocated to the corresponding virtual flexibility resources according to their membership degree ratio. The objective function of the fuzzy subspace clustering is: in, For equipment i Belongs to Virtual Flexibility Resources c Membership degree; For virtual flexibility resources c In the f Cluster centers on dimensional features; The feature weight of the virtual flexibility resource c on the f-th feature dimension; m For fuzzy index; C Total number of virtual flexibility resources; N Total number of devices; F The number of feature dimensions; This is the weight entropy regularization coefficient; membership degree The iterative update formula is: Virtual flexibility resources c upper limit of equivalent power regulation for: Equivalent response delay for: in, For equipment i At any moment t Maximum adjustable power; For equipment i Self-adjusting response delay; Indicates device i At any moment t It is in an adjustable operating state; The iteration terminates when both the change in cluster centers and the change in the objective function are below the convergence threshold, and outputs several virtual flexibility resources and their corresponding device member membership matrices. At the system layer, the external characteristics of each virtual flexibility resource are weighted and aggregated according to the membership matrix. The equivalent power adjustment upper bound, equivalent power adjustment lower bound, equivalent ramp rate, equivalent response delay, equivalent continuous adjustment time and equivalent energy state range of each virtual flexibility resource are calculated to form a set of virtual flexibility resources with a unified interface description.

4. A method for regulating low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling according to claim 3, characterized in that, The composite similarity is composed of six sub-similarity factors: spatial Euclidean distance, type attribute matching degree, adjustment capacity difference, response time difference, runtime overlap rate, and functional coupling strength, linearly weighted by adaptive weights. These adaptive weights are dynamically adjusted based on the importance ratio of the adjustment capacity dimension to the spatial dimension in the current clustering target. Represented as: in, For the current clustering objective, the first m Sensitivity coefficient for moderating the similarity of items. This refers to the temperature parameter.

5. A method for regulating low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling according to claim 1, characterized in that, The hierarchical coordination optimization of multi-agent systems based on attention mechanism weighted reinforcement learning is as follows: The global coordination layer aims at the overall economic efficiency and low carbon emissions of the park. It integrates ultra-short-term forecasts of total load, ultra-short-term forecasts of renewable energy, time-of-use electricity price sequences, carbon price sequences, tie-line power exchange plans, and real-time feedback of carbon potential information from each node. Through a global attention mechanism, it weights the adjustable capacity boundaries reported by each virtual flexibility resource based on their importance, and generates coordinated control instructions for each virtual flexibility resource. The global attention mechanism dynamically allocates attention weights based on the adjustment capacity, response speed, and carbon potential of each virtual flexibility resource at its node. Each agent in the local execution layer takes its own local observation state and the coordinated control command as input. The local observation state includes the current output power, the remaining adjustable upper and lower bounds, the current ramp margin, the remaining sustainable adjustment time, the current state of charge of the energy storage and the real-time carbon potential of the node. Each agent outputs the power adjustment amount, the charge and discharge power set value, the start and stop control signal or the load transfer amount through the strategy network.

6. A method for regulating low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling according to claim 1, characterized in that, The reinforcement learning agents are jointly trained using a multi-agent proximal policy optimization algorithm with priority experience replay. During the training phase, the global coordination layer calculates the global reward based on the system response after each agent performs an action. The global reward is a weighted average of three components: an economic cost reward, a low-carbon reward, and a safety constraint reward. The economic cost reward consists of negative values ​​of electricity purchase cost, gas purchase cost, and equipment operation and maintenance cost. The low-carbon reward consists of the negative cumulative value of the product of each node's carbon potential and net load, and is dynamically updated based on real-time feedback of each node's carbon potential information. The safety constraint reward consists of negative quadratic penalties for power balance deviation, equipment output exceeding limits, energy storage state of charge exceeding limits, and ramping exceeding limits. The global reward function... Represented as: in, , , Time periods t The costs of electricity purchase, gas purchase, and operation and maintenance. For nodes n During the period t Real-time carbon potential. For nodes n During the period t The load power, The total number of nodes. Q The total number of constraint terms. No. q Quadratic penalty for violating the constraint. , , The weighting coefficients for the three awards: economic, low-carbon, and safety. No. i The shearing loss function for updating the policy network of each agent is: in, For probability ratios, For the dominant function, For shear hyperparameters; Advantage function Based on global rewards by the global coordination layer With global value function V(s) calculate: in, Let V(s) be the discount factor, and V(s) be the global state value function. During the distributed execution phase, each agent makes independent decisions based on its local observations and the current coordination instructions issued by the global coordination layer. The global coordination layer verifies the feasibility of each agent's decision-making actions and performs minimum adjustment projection corrections on actions that violate power balance constraints or equipment operation constraints, forming an executable sequence of scheduling instructions.

7. A method for regulating low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling according to claim 1, characterized in that, The dual-dimensional attention prediction model that integrates temporal attention and spatial attention adopts an encoder-decoder architecture; The encoder consists of a temporal attention module and a spatial attention module in parallel. The temporal attention module captures long-range temporal dependencies in the input feature sequence based on a multi-head self-attention mechanism, while the spatial attention module captures spatial coupling relationships between different energy devices based on a graph attention mechanism. The outputs of the temporal attention module and the spatial attention module are fused through cross-attention to form a comprehensive encoded feature. The decoder takes the comprehensive coding features as input and generates predicted net load values ​​and tie-line switching power values ​​for multiple future sampling periods in an autoregressive manner.

8. A method for regulating low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling according to claim 1, characterized in that, The dynamic graph attention carbon potential perception model based on the actual topology of the park is implemented as follows: The topology of the power grid in the park is defined as a dynamic graph structure. Graph nodes include busbars, feeders and key equipment connection points, and graph edges include transmission and distribution lines and transformer branches. The initial feature vector of each graph node consists of its injected active power, injected reactive power, node type code and the number of edges connected to the node. The initial feature vector of each graph edge consists of its resistance, reactance and current transmission power. The dynamic graph attention carbon potential sensing model adopts a multi-layer temporal graph attention convolutional architecture. In each layer, each node aggregates the feature information of its neighboring nodes through a graph attention mechanism. The graph attention mechanism calculates an independent attention coefficient for each pair of adjacent nodes. The attention coefficient is jointly determined by the feature vectors of the two nodes, the feature vector of the connecting edge, and the current time period. A gated recurrent unit is introduced to recursively update the hidden state of each node along the time dimension, so that the carbon potential sensing result evolves dynamically with the time-varying system operating state. After multi-layer graph attention convolution and gated recurrent unit recursion, the carbon potential prediction value of each node in the current and several future time periods is obtained through the fully connected regression output layer. The real-time carbon emissions corresponding to the power consumption of any group of load nodes in any time period are given by the product of the node's carbon potential prediction value and the load power.

9. A method for regulating low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling according to claim 1, characterized in that, The shrinking time-domain model prediction control framework undergoes rolling optimization, as detailed below: The shrinking time-domain model prediction control framework is optimized in a rolling manner. At each scheduling moment, a fixed length of the prediction time domain is truncated forward from the current moment. The distributed execution process is performed to obtain the scheduling instruction sequence for each time period within the prediction time domain. Only the instruction for the current scheduling time period in the scheduling instruction sequence is sent to the field execution terminal. After the current scheduling period ends, the actual operating status of each device in the park, actual load demand, actual renewable energy output, actual node carbon potential and the latest electricity price and carbon price signals are collected again, the prediction time domain starting point is updated, and the above rolling optimization process is repeated. During the rolling optimization process, the policy networks of each agent after training are fine-tuned online: small batches of actual running data from several recent scheduling periods are used to calculate the temporal difference error between the cumulative reward after actual execution and the predicted reward. The time-series difference error is used to perform gradient updates on the policy network parameters with a small step size, so that the agent policy can continuously adapt to changes in the current operating conditions.

10. A method for regulating low-carbon industrial parks based on multi-source data fusion and energy-carbon dual-flow coupling according to claim 1, characterized in that, The flexibility boundary of the park system is the power limit and duration that can be adjusted upwards / downwards after the aggregation of each virtual flexibility resource; The optimal operating strategy trajectory consists of the output setpoints of each device in each scheduling period, the energy storage charging and discharging plan, the adjustable load adjustment amount, and the tie-line power exchange plan. The energy-saving and carbon-reduction potential optimized scheduling scheme, compared with the baseline operation scheme, includes reductions in energy costs, primary energy consumption, renewable energy curtailment rates, and total carbon emission reductions. The total carbon emission reduction is statistically analyzed for each load node, forming a node-level carbon emission reduction contribution distribution map. The total carbon emission reduction... The calculation is as follows: in, , Nodes in the baseline scenario n During the period t Carbon potential and load power, , Nodes in optimized scenarios n During the period t Carbon potential and load power; This represents the total number of scheduling periods; This represents the total number of nodes.

Images