A method and system for optimizing the scheduling of distributed resources in a park

By using cloud-edge collaborative scenario-based modeling and lightweight deployment, the edge side utilizes a lightweight simulation proxy model for rapid pre-simulation and autonomous decision-making, solving the problem of candidate control action decision-making for edge computing nodes under limited computing power and storage conditions, and realizing real-time and accurate optimized scheduling of distributed resources in the park.

CN123117620APending Publication Date: 2026-11-17国网河北省电力有限公司营销服务中心 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611256866.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

With limited computing power and storage, existing edge computing nodes struggle to make rapid, accurate, and real-time decisions on candidate control actions in the optimized scheduling of distributed resources in industrial parks. This is especially true when there are significant fluctuations in renewable energy output and load, which can easily lead to control results that exceed limits or have poor economic efficiency.

Method used

Through cloud-edge collaborative scenario-based modeling and lightweight deployment, a lightweight simulation proxy model is generated on the cloud side and distributed to the edge side. The edge side uses the lightweight simulation proxy model for rapid pre-simulation and autonomous decision-making, and combines model error feedback for online fine-tuning and version rollback to ensure the accuracy and adaptability of the pre-simulation.

Benefits of technology

It enables rapid and accurate simulation and pre-playback of candidate control actions at the edge under limited computing power and storage conditions, ensuring the real-time performance and long-term operational accuracy of edge decision-making, adapting to the complex operating conditions of distributed resources in the park, and improving the adaptability and accuracy of autonomous decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117620A_ABST
    Figure CN123117620A_ABST
Patent Text Reader

Abstract

This invention relates to the field of power system technology, and more particularly to a method and system for optimized scheduling of distributed resources in industrial parks. The invention constructs an edge-side operational status representation and a set of candidate control actions, generates and distributes scenario-specific lightweight simulation proxy models on the cloud side, and performs rapid pre-simulation and autonomous decision-making based on these models at the edge. Based on model error feedback, the simulation proxy models are collaboratively updated, enabling edge computing nodes to accurately pre-simulate and safely decide on candidate control actions under limited computing power and storage conditions. Furthermore, the lightweight simulation proxy models are continuously and collaboratively updated with operational data, maintaining long-term pre-simulation accuracy. This ensures real-time decision-making while improving the accuracy, security, and adaptability of autonomous decision-making at the edge.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a method and system for optimizing the scheduling of distributed resources in a park. Background Technology

[0002] With the continuous growth in the scale of distributed resources such as distributed photovoltaics, energy storage, electric vehicle charging piles, and flexible loads within the park, real-time optimization and scheduling of these distributed resources are necessary to improve the park's operational economy, security, and low-carbon performance. Under a cloud-edge collaborative scheduling architecture, the cloud side typically handles global optimization and sends boundary policies such as power, voltage, and state of charge as operational constraints to the edge side. The edge side, within these boundary constraints, needs to make local decisions and execute candidate control actions such as energy storage charging and discharging, charging power adjustment, load shedding and transfer, and inverter reactive power regulation. The key to ensuring the real-time performance and operational quality of the park's distributed resource scheduling lies in how to quickly and reliably select the optimal action that balances safety and economy from a large number of candidate control actions at the edge side.

[0003] However, existing edge-side autonomous decision-making methods still have many shortcomings. Decision-making methods based on rules or empirical thresholds are simple to implement, but they struggle to account for the coupling relationships between distributed resources and the comprehensive impact of candidate control actions on power deviation, voltage overruns, energy storage state of charge, operating costs, and carbon emissions. Under conditions of significant fluctuations in renewable energy output and load, they are prone to overruns or poorly economical control results. Decision-making methods based on complete physical simulation models or detailed mechanistic models can accurately predict the execution effects of candidate control actions, but they have high computational complexity and consume large amounts of storage and computing resources. Edge computing nodes have very limited computing power, storage capacity, and power consumption budgets, making it difficult to run them locally in real time. This results in long prediction times and fails to meet the time requirements for real-time decision-making at the edge. Purely data-driven decision-making methods directly learn control action mappings from historical data, lacking explicit prediction of the system state after the execution of candidate control actions, which poses control risks under extreme or rare conditions. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the existing technology in that it is difficult to balance the accuracy of candidate control action pre-simulation and the long-term accuracy of decision-making under the limited computing power and storage conditions of edge computing nodes.

[0005] To address the aforementioned technical problems, this invention provides a method for optimizing and scheduling distributed resources in a park, comprising: Upload the edge-side operational status representation vector and candidate control action set of each edge computing node at each historical moment to the cloud side; Receive and store the lightweight simulation proxy models corresponding to each running scenario issued by the cloud side. At each decision time, match the lightweight simulation proxy model corresponding to each edge computing node at the current decision time according to the running scenario of each edge computing node at the current decision time. Input the edge-side operational status representation vector of each edge computing node at the current decision time and the set of candidate control actions into the corresponding lightweight simulation proxy model, and output the pre-simulation output vector corresponding to the execution of each candidate control action by each edge computing node at the current decision time; Based on the pre-simulated output vectors corresponding to each candidate control action of each edge computing node at the current decision time, the target control action of each edge computing node at the current decision time is selected, and it is converted into a device-level control command and sent to the distributed resources under the jurisdiction of the corresponding edge computing node for execution.

[0006] Preferably, the pre-simulated output vectors corresponding to the candidate control actions executed by each edge computing node at the current decision-making moment include: At the current decision-making moment, the common coupling point power deviation, common coupling point voltage prediction value, energy storage state of charge change, operating cost increment, carbon emission increment, and control risk assessment quantity are corresponding to each candidate control action of each edge computing node.

[0007] Preferably, the method for selecting the target control action of each edge computing node at the current decision time based on the pre-simulated output vector corresponding to each candidate control action of each edge computing node at the current decision time includes: Based on the pre-simulated voltage value of the common connection point after each edge computing node executes each candidate control action at the current decision time, calculate the voltage over-limit risk assessment quantity corresponding to each candidate control action of each edge computing node at the current decision time. Based on the common connection point power deviation, voltage over-limit risk assessment, operating cost increment, carbon emission increment and control risk assessment of each candidate control action of each edge computing node at the current decision time, the comprehensive evaluation value of each candidate control action of each edge computing node at the current decision time is obtained. The candidate control action with the smallest comprehensive evaluation value is taken as the target control action of each edge computing node at the current decision moment.

[0008] Preferably, based on the predicted voltage values ​​of the point of common connection after each edge computing node executes each candidate control action at the current decision-making time, the formula for calculating the voltage exceedance risk assessment quantity corresponding to each candidate control action of each edge computing node at the current decision-making time is as follows: , in, Indicates the moment of decision-making Below, the voltage over-limit risk assessment quantity corresponding to the m-th candidate control action of the e-th edge computing node. This indicates taking the maximum value. and This indicates the upper and lower limits of the allowable voltage. Indicates the moment of decision-making Below, the pre-simulated value of the common connection point voltage corresponding to the m-th candidate control action of the e-th edge computing node.

[0009] Preferably, after selecting the target control actions of each edge computing node at the current decision time, the control risk of each edge computing node is verified, and the target control actions of each edge computing node at the current decision time that pass the verification are taken as the final target control actions of each edge computing node at the current decision time. When the target control action of the edge computing node at the current decision-making moment satisfies the following formula, it is determined to have passed the control risk verification: , in, Indicates the moment of decision-making Below, the control risk assessment quantity corresponding to the target control action of the e-th edge computing node. This indicates the control risk threshold. Indicates the moment of decision-making Below, the voltage over-limit risk assessment quantity corresponding to the target control action of the e-th edge computing node; If the target control action of the edge computing node fails the verification at the current decision moment, the conservative safety action corresponding to the edge computing node is executed, and the decision period corresponding to the current decision moment is marked as a high-risk period.

[0010] Preferably, after converting the target control actions of each edge computing node at the current decision-making moment into device-level control commands and issuing them to the distributed resources managed by each edge computing node for execution, the method further includes: Based on the error between the pre-simulated output vector and the actual output vector corresponding to the target control action executed by each edge computing node at the current decision moment, a model error summary of each edge computing node at the current decision moment is constructed. Determine whether the current decision moment is the last decision moment within the feedback cycle. If so, fine-tune the parameters of the lightweight simulation proxy model corresponding to each edge computing node at the current decision moment to obtain the model parameters of the target lightweight simulation proxy model corresponding to each edge computing node. Based on the change in model parameters before and after fine-tuning, construct the compressed model increment for each edge computing node in the feedback cycle. The model error feedback quantity of each edge computing node in the feedback cycle is constructed by using model error summary, key feature summary, scene mismatch marking times, running scene, high-risk cycle marking times, and compressed model increment of each edge computing node in the feedback cycle, and uploaded to the cloud side. This enables the cloud side to adjust the parameters of the lightweight simulation proxy model corresponding to the running scene of each edge computing node.

[0011] This invention also provides a method for optimizing and scheduling distributed resources in a campus, for use on the cloud side, comprising: Receive edge-side operational status representation vectors and candidate control action sets corresponding to each edge computing node at each historical moment uploaded by the edge side, and classify them according to the operational scenario; Using the edge-side operational status representation vectors and corresponding candidate control actions of each edge computing node at each historical moment under the same operational scenario as training samples, and the actual output vectors obtained by the corresponding edge computing node after executing the candidate control actions at the corresponding moment as sample labels, the base simulation model is trained in different scenarios to obtain the base simulation model corresponding to each operational scenario. The base simulation model under each operating scenario is lightweighted to obtain a lightweight simulation proxy model under each operating scenario. The version information of each lightweight simulation proxy model is marked and stored in the model version library, and then distributed to the edge side.

[0012] The model error feedback of each edge computing node in the feedback period is received from the edge side. The model error summary of each edge computing node in the feedback period is weighted and summed with the error weight coefficient of each component in the model error summary to obtain the comprehensive model error index of each edge computing node in the feedback period. If the comprehensive model error index of the feedback cycle edge computing node exceeds the accuracy threshold but does not exceed the retraining threshold, the compressed model increment of the feedback cycle edge computing node is aggregated, and the lightweight simulation proxy model in the running scenario of the feedback cycle edge computing node is fine-tuned and updated online. If the comprehensive model error index of the feedback cycle edge computing node exceeds the retraining threshold, or the number of scene mismatch markings of the feedback cycle edge computing node exceeds the preset upper limit, the lightweight simulation proxy model of the running scenario to which the feedback cycle edge computing node belongs is regenerated by combining the key feature summary of the feedback cycle edge computing node.

[0013] Preferably, after obtaining the lightweight simulation proxy model of the running scenario to which the feedback cycle edge computing node belongs after online fine-tuning and updating, or the regenerated lightweight simulation proxy model of the running scenario to which the feedback cycle edge computing node belongs, it is verified offline. If the comprehensive model error index on the verification dataset is larger than that of the lightweight simulation proxy model before adjustment, it is not sent to the edge side. When the lightweight simulation proxy model of the feedback cycle edge computing node under the running scenario, which has been fine-tuned and updated online, or the lightweight simulation proxy model of the feedback cycle edge computing node under the running scenario, is sent to the edge side, if the increase in the number of high-risk cycle markers within the observation window compared to the previous version of the lightweight simulation proxy model under the running scenario exceeds a set threshold, then the lightweight simulation proxy model of the previous version under the running scenario is retrieved from the model version library and sent to the edge side as the final lightweight simulation proxy model under that running scenario.

[0014] This invention also provides a distributed resource optimization and scheduling system for industrial parks, comprising: On the edge side, it is used to implement the above-mentioned distributed resource optimization scheduling method in the park; On the cloud side, it is used to implement the above-mentioned distributed resource optimization and scheduling method for the park.

[0015] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention discloses a method and system for optimizing and scheduling distributed resources in a park. It achieves efficient and accurate autonomous edge decision-making through cloud-edge collaborative scenario-based modeling and lightweight deployment. On the cloud side, corresponding base simulation models are independently trained for different operating scenarios within the park. These base models are then lightweighted to generate multi-scenario-specific lightweight simulation proxy models, which are then distributed to the edge side. This approach effectively reduces the local computation and storage pressure on edge computing nodes, enabling them to perform complete, explicit, and rapid simulations of the system state after candidate control actions, even under limited computing, storage, and power resources. This completely avoids the problem of difficulty in running complete physical mechanism models in real-time at the edge. The edge side matches the corresponding scenario's lightweight simulation proxy model to the real-time operating conditions, inputs the locally constructed operating status representation vector and the set of candidate control actions, and can quickly perform effect simulations of multiple control actions. Based on the explicit simulation results, on-site autonomous decision-making is achieved. Furthermore, relying on the adaptability of scenario-specific models, it effectively adapts to complex operating conditions with significant fluctuations in new energy output and load, ensuring the real-time performance, simulation accuracy, and long-term operational precision of edge decisions.

[0016] Furthermore, as the characteristics, operating conditions, and environmental conditions of distributed resources within the park change over time, a fixed simulation model will gradually deviate from the actual operating conditions at the edge, leading to inaccurate simulation results and a decline in decision-making quality. Existing methods lack effective monitoring and collaborative updating mechanisms for edge-side simulation model deviations, making it difficult to ensure that the simulation model maintains continuous accuracy during long-term operation. To address this, this invention uses model error feedback to perform online fine-tuning, retraining, or version rollback of the lightweight simulation proxy model corresponding to each edge computing node. This allows for continuous collaborative updating based on operating conditions, maintaining long-term simulation accuracy and improving the adaptability of edge autonomous decision-making. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a method for optimizing and scheduling distributed resources in a park, as described in this invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0019] like Figure 1 As shown, Figure 1 This is a flowchart illustrating a method for optimizing and scheduling distributed resources in a park, as described in this invention.

[0020] This embodiment provides a method for optimizing and scheduling distributed resources in a campus, used at the edge, including: Obtain the operation and environmental data of the distributed resources managed by each edge computing node in the park at different historical times, construct the edge-side operation status representation vector of each edge computing node for simulation and pre-running at each historical time, and upload the edge-side operation status representation vector and candidate control action set of each edge computing node at each historical time to the cloud. The edge-side operational status representation vector and candidate control action set serve as inputs for subsequent rapid pre-simulation and autonomous decision-making using a lightweight simulation proxy model. The boundary policy issued from the cloud side serves only as input constraints for local edge-side decision-making.

[0021] Each edge computing node within the park collects operational data from its distributed resources, including distributed photovoltaic systems, energy storage devices, electric vehicle charging piles, and flexible loads, as well as environmental data such as irradiance and temperature. After protocol adaptation and time alignment, it constructs an operational status representation vector for each edge computing node at each decision-making moment. , in, In order to make decisions Below, the operational status representation vector of the e-th edge computing node, In order to make decisions Below, the distributed photovoltaic output corresponding to the e-th edge computing node, In order to make decisions Below, the flexible load power corresponding to the e-th edge computing node, In order to make decisions Below, the state of charge of the energy storage device corresponding to the e-th edge computing node, In order to make decisions Below, the power of the electric vehicle charging pile corresponding to the e-th edge computing node, In order to make decisions Below, the common connection point exchange power corresponding to the e-th edge computing node, In order to make decisions Below, the voltage of the common connection point corresponding to the e-th edge computing node, Indicates the decision moment The environmental state sub-vector consists of irradiance, ambient temperature, etc., corresponding to the e-th edge computing node, where e represents the edge computing node index. Indicates the moment of decision-making.

[0022] While constructing the operational status representation vector, the edge computing nodes extract the adjustability of each distributed resource at the current decision-making moment, including the charging power margin and discharging power margin of energy storage devices, the adjustable power range of electric vehicle charging piles, the reduceable and transferable capacity of flexible loads, and the available reactive power capacity of inverters, denoted as the adjustable capability set. This is used to determine the feasible domain of subsequent candidate control actions; Edge computing nodes define control action vectors for the distributed resources under their jurisdiction. The control action vectors consist of energy storage charging and discharging actions, charging power adjustment actions, load reduction / transfer actions, and inverter reactive power adjustment actions.

[0023] The control action vector of each edge computing node at each decision moment consists of the energy storage charging and discharging action, charging power adjustment action, load reduction / transfer action, and inverter reactive power adjustment action corresponding to each edge computing node at each decision moment: , in, In order to make decisions Below, the control action vector of the e-th edge computing node, In order to make decisions Below, the energy storage charging and discharging power adjustment amount corresponding to the e-th edge computing node, with positive values ​​representing charging and negative values ​​representing discharging. In order to make decisions Below, the charging power adjustment amount of the charging pile corresponding to the e-th edge computing node. In order to make decisions Below, the amount of flexible load reduction or transfer power corresponding to the e-th edge computing node. In order to make decisions Below, the reactive power adjustment of the inverter corresponding to the e-th edge computing node.

[0024] Determine the feasible range of the control action vector of each edge computing node at each decision time, and perform discrete sampling within the feasible range of each control action component to obtain the initial control action set of each edge computing node at each decision time; The feasible range of each component in the control action vector of each edge computing node at each decision moment is determined by the physical constraints and adjustability set of the corresponding distributed resources. The following was jointly determined: Energy storage charging and discharging operation constraints: , In the formula, and They represent the time of decision-making, respectively. Below, the rated power of the energy storage device corresponding to the e-th edge computing node and the maximum chargeable power and maximum dischargeable power determined by the current state of charge.

[0025] Charging power adjustment action constraints: , In the formula, and They represent the decision time respectively. Below, the lower and upper limits of the charging power are determined by the charging demand of electric vehicles and the rated power of charging piles corresponding to the e-th edge computing node.

[0026] Load shedding / transfer action constraints: , In the formula, Indicates the decision-making moment The maximum power that can be reduced or transferred is determined by the ratio of flexible load interruptibility / transferability corresponding to the e-th edge computing node and user comfort constraints.

[0027] Inverter reactive power regulation action constraints: , In the formula, This indicates the inverter's rated apparent capacity. Indicates the decision-making moment The current active power output of the inverter corresponding to the e-th edge computing node is given by the formula, which indicates that the reactive power regulation is constrained by the apparent capacity of the inverter and the current active power output.

[0028] Based on the boundary policy constraints issued by the cloud side, the feasibility of the initial control action set of each edge computing node at each decision time is screened to obtain the candidate control action set of each edge computing node at each decision time. The feasibility screening only retains control actions that are expected to bring the power, voltage, and state of charge of the point of common coupling (PCC) into the allowable range issued by the cloud side after the application of the control action. The feasibility screening criteria are as follows: , in, Indicates the moment of decision-making Next, the m-th candidate control action of the e-th edge computing node, , and They respectively represent the application Estimates of power, voltage, and state of charge of energy storage at the post-point of common coupling. , , , , and The cloud-side boundary policy specifies the allowable range of power, voltage, and energy storage state of charge for the common connection point of the e-th edge computing node.

[0029] The complete set of initial candidate control actions retained after feasibility screening constitutes the candidate control action set for each edge computing node at each decision time: , In the formula, Indicates the moment of decision-making Below, the candidate control action set for the e-th edge computing node, This represents the number of initial candidate control actions retained after feasibility screening, i.e. The total number of candidate control actions for the e-th edge computing node at time e. Indicates the moment of decision-making Below, the e-th edge computing node's... Candidate control actions.

[0030] Thus, the edge computing nodes obtain operational status representation vectors for simulation pre-running. and candidate control action sets that have undergone preliminary feasibility screening Both methods utilize a lightweight simulation proxy model to rapidly simulate and input candidate control actions for autonomous decision-making. Currently, the candidate control action set has only undergone preliminary feasibility screening based on steady-state predictions. Evaluation metrics for each candidate control action during the dynamic process, such as power deviation, voltage exceedance risk, state of charge change, cost and carbon emission increments, and control risk, will be further simulated using the lightweight simulation proxy model.

[0031] The edge computing node receives and stores lightweight simulation proxy models corresponding to various operating scenarios from the cloud. At each decision point, based on the operating scenario of each edge computing node at that time, it matches the corresponding lightweight simulation proxy model from the locally stored lightweight simulation proxy models. The edge computing node also stores scenario-specific lightweight simulation proxy models from the cloud locally. In each decision cycle, the edge computing node first matches the operating scenario based on the current operating scenario and selects the corresponding lightweight simulation proxy model. Then, it uses the selected lightweight simulation proxy model to quickly pre-simulate each candidate control action in the generated candidate control action set, outputting evaluation quantities such as power deviation, voltage limit exceedance risk, energy storage state of charge change, cost and carbon emission increment, and control risk for each candidate control action. Finally, based on multi-objective comprehensive evaluation, the optimal control action is selected from the candidate control action set, and after control risk verification, it is issued for execution. The entire pre-simulation and decision-making process is completed locally on the edge computing node without real-time interaction with the cloud.

[0032] The edge computing node extracts the runtime scenario vector at the current decision moment, determines the runtime scenario label for the current runtime scenario, and selects a lightweight simulation proxy model for the corresponding scenario from the locally stored scenario-specific simulation proxy models. , in, express The runtime scenario label for the e-th edge computing node at time e. This represents a set of scenario-based simulation agent models. express The lightweight simulation proxy model corresponding to the e-th edge computing node at time e. The model selection operator is used to select a lightweight simulation proxy model for the corresponding running scenario from the set of lightweight simulation proxy models for each running scenario, based on the running scenario label.

[0033] The operational scenarios include photovoltaic output level, load level, energy storage charge state range, and electric vehicle access scale.

[0034] When no lightweight simulation proxy model matches the current running scenario label, the edge computing node selects the lightweight simulation proxy model that is closest to the running scenario as a replacement, and marks this decision cycle as a scenario mismatch for use in model error feedback.

[0035] Input the edge-side operational status representation vector of each edge computing node at the current decision time and the set of candidate control actions into the corresponding lightweight simulation proxy model, and output the pre-simulation output vector corresponding to the execution of each candidate control action by each edge computing node at the current decision time; At the current decision-making moment, each edge computing node uses the selected lightweight simulation proxy model to process the candidate control action set. Each candidate control action in Perform a rapid pre-simulation to obtain the pre-simulation output vectors corresponding to the candidate control actions executed by each edge computing node at the current decision moment: , , In the formula, Indicates the moment of decision-making Below, the pre-simulated output vector of the m-th candidate control action of the e-th edge computing node. Indicates the moment of decision-making Below, the common connection point power deviation after applying the candidate control action corresponding to the e-th edge computing node m-th action is... Indicates the moment of decision-making Below, the pre-simulated voltage value of the common connection point corresponding to the m-th candidate control action of the e-th edge computing node; Indicates the moment of decision-making Below, the change in the state of charge of the energy storage corresponding to the m-th candidate control action of the e-th edge computing node; Indicates the moment of decision-making Below, the incremental operating cost corresponding to the m-th candidate control action of the e-th edge computing node is... Indicates the moment of decision-making Below, the carbon emission increment corresponding to the m-th candidate control action of the e-th edge computing node. Indicates the moment of decision-making Below, the control risk assessment quantity corresponding to the m-th candidate control action of the e-th edge computing node. This indicates the number of candidate control actions.

[0036] Because of the small size and fast inference speed of the lightweight simulation agent model, the edge computing node can complete the pre-rehearsal of all candidate control actions within a single decision cycle.

[0037] In this embodiment, specifically, the pre-simulation output vectors corresponding to the candidate control actions executed by each edge computing node at the current decision-making moment include: At the current decision-making moment, the common coupling point power deviation, common coupling point voltage prediction value, energy storage state of charge change, operating cost increment, carbon emission increment, and control risk assessment quantity are corresponding to each candidate control action of each edge computing node.

[0038] Based on the pre-simulated output vectors corresponding to each candidate control action of each edge computing node at the current decision time, the target control action of each edge computing node at the current decision time is selected, and it is converted into a device-level control command and sent to the distributed resources under the jurisdiction of the corresponding edge computing node for execution.

[0039] In this embodiment, preferably, the method for selecting the target control action of each edge computing node at the current decision time based on the pre-simulated output vector corresponding to each candidate control action of each edge computing node at the current decision time includes: Based on the pre-simulated voltage value of the common connection point after each edge computing node executes each candidate control action at the current decision time, calculate the voltage over-limit risk assessment quantity corresponding to each candidate control action of each edge computing node at the current decision time. The formula for calculating the voltage over-limit risk assessment quantity corresponding to each candidate control action of each edge computing node at the current decision moment is: , in, Indicates the moment of decision-making Below, the voltage over-limit risk assessment quantity corresponding to the m-th candidate control action of the e-th edge computing node. This indicates taking the maximum value. and This indicates the upper and lower limits of the allowable voltage. Indicates the moment of decision-making Below, the pre-simulated value of the common connection point voltage corresponding to the m-th candidate control action of the e-th edge computing node.

[0040] When the simulated voltage is within the allowable range, the voltage over-limit risk assessment quantity is zero; the greater the degree of over-limit, the greater the voltage over-limit risk assessment quantity.

[0041] Based on the common connection point power deviation, voltage over-limit risk assessment, operating cost increment, carbon emission increment and control risk assessment of each candidate control action of each edge computing node at the current decision time, the comprehensive evaluation value of each candidate control action of each edge computing node at the current decision time is obtained. In this embodiment, specifically, the power deviation of the common connection point, the voltage over-limit risk assessment quantity, the increase in operating cost, the increase in carbon emission, and the control risk assessment quantity corresponding to each candidate control action of each edge computing node at the current decision time are normalized and then weighted and summed to obtain the comprehensive evaluation value of each candidate control action of each edge computing node at the current decision time: , in, Indicates the moment of decision-making Below, the comprehensive evaluation value of the m-th candidate control action of the e-th edge computing node. to These represent the weighting coefficients for power deviation, voltage exceedance risk, cost increment, carbon emission increment, and control risk, respectively. All weighting coefficients are non-negative and... It can be configured according to the emphasis of the park's operational goals. The smaller the comprehensive evaluation value, the better the overall performance of the candidate control action in terms of safety, economy and low carbon emissions.

[0042] The candidate control action with the smallest comprehensive evaluation value is taken as the target control action for each edge computing node at the current decision time, as shown in the formula: , in, Indicates the moment of decision-making Next, the target control action of the e-th edge computing node.

[0043] In this embodiment, preferably, in order to avoid over-limit or unsafe control caused by the pre-simulation error of the lightweight simulation proxy model, after selecting the target control actions of each edge computing node at the current decision time, the control risk is verified, and the target control actions of each edge computing node at the current decision time that have passed the verification are taken as the final target control actions of each edge computing node at the current decision time. When the target control action of the edge computing node at the current decision-making moment satisfies the following formula, it is determined to have passed the control risk verification: , in, Indicates the moment of decision-making Below, the control risk assessment quantity corresponding to the target control action of the e-th edge computing node. This indicates the control risk threshold. Indicates the moment of decision-making Below, the voltage over-limit risk assessment quantity corresponding to the target control action of the e-th edge computing node; If the target control action of the edge computing node fails the verification at the current decision time, the conservative safety action corresponding to the edge computing node is executed, and the decision time corresponding to the current decision time is marked as a high-risk period.

[0044] In this embodiment, specifically, when the optimal control action passes the control risk verification, the edge computing node decomposes the target control action into device-level control commands such as energy storage charging and discharging, charging power adjustment, load reduction / transfer, and inverter reactive power adjustment, and sends them to the corresponding distributed resources for execution. When the target control action fails the control risk verification, the edge computing node executes according to the preset conservative control strategy: selects the control action that keeps the power of the common connection point near the midpoint of the cloud-side downlink power allowable interval and has the lowest control risk assessment value from the candidate control action set as the safe action to be issued and executed, and sets the decision period corresponding to the current decision time as the high-risk period for use in model error feedback.

[0045] Therefore, edge computing nodes, with the help of lightweight simulation proxy models that feature scale compression and fast inference, can perform explicit pre-simulation and safety verification of candidate control actions locally in real time, and autonomously select the optimal control action that balances safety, economy and low carbon emissions. At the same time, the deviation between the pre-simulation output and the actual execution result, as well as the markers for scenario mismatch and high risk, will serve as the basis for collaborative updating of the proxy model based on model error feedback.

[0046] In each feedback cycle, edge computing nodes statistically analyze the model error between the pre-run output and the actual execution result of the lightweight simulation proxy model, and upload it to the cloud platform along with key feature summaries, scene mismatch and high-risk markers, and compressed incremental model data. Based on the model error feedback uploaded by each edge computing node, the cloud platform determines the accuracy status of the corresponding lightweight simulation proxy model, and then performs online fine-tuning updates, version rollbacks, or retraining and distribution of the lightweight simulation proxy model in the model version library, so that the lightweight simulation proxy model on the edge side can maintain its pre-run accuracy in long-term operation.

[0047] In this embodiment, preferably, after converting the target control actions of each edge computing node at the current decision moment into device-level control commands and issuing them to the distributed resources managed by each edge computing node for execution, the method further includes: Based on the error between the pre-simulated output vector and the actual output vector corresponding to the target control action executed by each edge computing node at the current decision-making moment, a model error summary for each edge computing node at the current decision-making moment is constructed, as shown in the formula: , , in, Indicates the moment of decision-making Below, the single-step model error vector of the e-th edge computing node, Indicates the moment of decision-making Below, the pre-simulated output vector corresponding to the target control action executed by the e-th edge computing node. Indicates the moment of decision-making Below, the actual output vector corresponding to the target control action executed by the e-th edge computing node. Indicates the first The model error summary of the e-th edge computing node in the feedback cycle, i.e. In the pre-simulated output vector corresponding to the target control action executed by the e-th edge computing node in the feedback cycle, the first... The root mean square error of each component Indicates the moment of decision-making Below, the model error vector's first... One portion, This represents the total number of decision-making moments within the feedback cycle. This represents the dimension of the pre-show output.

[0048] Determine whether the current decision moment is the last decision moment within the feedback cycle. If so, fine-tune the parameters of the lightweight simulation proxy model corresponding to each edge computing node at the current decision moment to obtain the model parameters of the target lightweight simulation proxy model corresponding to each edge computing node at the current decision moment. Based on the changes in model parameters between the target lightweight simulation proxy model and the lightweight simulation proxy model corresponding to each edge computing node at the current decision moment, the compressed model increment for each edge computing node during the feedback cycle is constructed, as follows: , in, For the first The compression model increment of the e-th edge computing node in each feedback cycle. This represents sparsification and quantization compression operations. For the first The model parameters of the target lightweight simulation proxy model corresponding to the e-th edge computing node in each feedback cycle. For the first Model parameters of the lightweight simulation proxy model corresponding to the e-th edge computing node in each feedback cycle; The method for fine-tuning the parameters of the lightweight simulation proxy model corresponding to each edge computing node at the current decision time includes: accumulating local samples during the feedback period, wherein the local samples are the edge-side operation status representation vector, candidate control actions, pre-simulation output vector of the lightweight simulation proxy model, and actual output vector fed back after the distributed resource executes the target control action at each decision time of the corresponding edge computing node; and fine-tuning the parameters of the lightweight simulation proxy model using the local samples accumulated during the feedback period.

[0049] The compressed model increment is the result of sparsification and quantization compression of the difference between the model parameters before and after local fine-tuning on the edge side. Only the parameter components that have changed significantly are uploaded to reduce the amount of uplink communication data.

[0050] The model error feedback quantity of each edge computing node in the feedback cycle is constructed by using the model error summary, key feature summary, scene mismatch marking times, running scenario, high-risk cycle marking times, and compressed model increment of each edge computing node in the feedback cycle. The model error feedback amount for each edge computing node during the feedback period is: In the formula, This indicates that the e-th edge computing node is at the . The model error feedback amount per feedback cycle; This indicates the main operational scenario tags involved in this feedback cycle. This represents a summary of the model error, composed of the root mean square error of each pre-simulation output. This represents a summary of key features, consisting of the operating scenarios and their statistics corresponding to typical working conditions with large model errors. Indicates the number of times the scene mismatch marker is displayed. Indicates the number of times a high-risk cycle is marked. This represents the compressed model increment obtained by locally fine-tuning the lightweight simulation proxy model based on a small number of local samples on the edge side.

[0051] The model error feedback of each edge computing node during the feedback cycle is uploaded to the cloud side, so that the cloud side can adjust the parameters of the lightweight simulation proxy model corresponding to the running scenario of each edge computing node according to the model error feedback of each edge computing node during the feedback cycle.

[0052] This second embodiment provides a method for optimizing and scheduling distributed resources in a park, used on the cloud side, including: Receive edge-side operational status representation vectors and candidate control action sets corresponding to each edge computing node at each historical moment uploaded by the edge side, and classify them according to the operational scenario; Using the edge-side operational status representation vectors and corresponding candidate control actions of each edge computing node at each historical moment under the same operational scenario as training samples, and the actual output vectors obtained by the corresponding edge computing node after executing the candidate control actions at the corresponding moment as sample labels, the base simulation model is trained for each scenario to obtain the base simulation model corresponding to each operational scenario: , , in, This refers to the actual output vector obtained after executing the candidate control action. This is the edge-side operational status representation vector. Indicates the parameters of the base simulation model. Indicates the power deviation at the point of common coupling. Indicates the voltage at the point of common coupling. This represents the change in the state of charge of the energy storage. Indicates the increase in operating costs; Indicates the increase in carbon emissions. This indicates the amount of risk assessment to be controlled. This represents the base simulation model.

[0053] The actual output vector obtained after executing the candidate control action For high-precision physical simulation results or actual measurement results after candidate actions are executed, labels can be generated by high-precision simulation for candidate control actions that have not been executed in the past.

[0054] The pre-simulated output vector obtained by the base simulation model after executing the candidate control action. Compared with high-precision simulation or measured results The goal is to minimize the error between the two sides. The base simulation model is trained using the following formula: , in, This represents the optimal base simulation model parameters obtained through training. This represents the total number of training samples. This represents the input of the nth training sample. This indicates the corresponding high-precision simulation or measured output.

[0055] The base simulation model has high simulation accuracy, but its parameters are large and its computation and storage costs are high, making it unsuitable for direct deployment on edge computing nodes.

[0056] The base simulation model under each operating scenario is lightweighted to obtain a lightweight simulation proxy model under each operating scenario. The version information of each lightweight simulation proxy model is marked and stored in the model version library, and then distributed to the edge side.

[0057] In this embodiment, in order to obtain a lightweight simulation proxy model that can run on edge computing nodes, the cloud platform performs lightweight processing on the base simulation model under each running scenario, including knowledge distillation, network pruning, and parameter quantization.

[0058] Knowledge distillation uses a base simulation model as the teacher model and a lightweight simulation proxy model as the student model. This allows the student model to approximate the pre-reading capability of the teacher model while significantly reducing the parameter size. Its training loss function is: , In the formula, This represents the knowledge distillation loss. This represents a lightweight simulation proxy model. This represents the parameters of the lightweight simulation proxy model. This represents the distillation weighting coefficient, with a value range of [value range missing]. The first term makes the student model output approximate the teacher model output, and the second term makes the student model output approximate the true label.

[0059] Network pruning removes connection weights in a lightweight simulation surrogate model whose absolute value is less than a pruning threshold, further reducing the model size. , In the formula, In the lightweight simulation proxy model, the first... The first neuron and the second Connection weights between neurons Indicates the pruning threshold. Indicates the first One neuron, Indicates the first One neuron.

[0060] Parameter quantization maps the floating-point parameters of the lightweight simulation proxy model to low-bit-width fixed-point numbers, reducing model storage footprint and computational overhead during edge-side inference. , In the formula, This represents the floating-point parameters before quantization. This represents the quantized parameters. Indicates the quantization step size. and These represent the upper and lower bounds of the parameter value, respectively. Indicates the quantization bit width.

[0061] After knowledge distillation, network pruning, and parameter quantization, the parameter size and storage footprint of the lightweight simulation proxy model are significantly reduced compared to the base simulation model, thus enabling it to run in real time under the limited computing power, storage, and power consumption conditions of edge computing nodes.

[0062] Considering the differences in the state distribution and control response characteristics of distributed resources in different operating scenarios of the park, the cloud platform divides historical operating data into scenarios according to operating scenarios, and generates a dedicated lightweight simulation proxy model for each type of operating scenario, so as to further reduce the size of a single model while ensuring the accuracy of the simulation.

[0063] The cloud-based platform clusters operating conditions based on factors such as photovoltaic power output, load level, energy storage state of charge range, and electric vehicle access scale, resulting in several operating scenarios. Each operating scenario is then assigned a label. , In the formula, express Time of the first The runtime scenario label of each edge computing node. This refers to a runtime scenario comprised of the aforementioned runtime scenarios. The scene partitioning function is implemented by clustering or preset rules. This indicates the total number of running scenario categories.

[0064] For each operational scenario, the cloud platform trains and compresses training samples belonging to that scenario to obtain a lightweight simulation proxy model corresponding to that scenario. Therefore, the lightweight simulation proxy models corresponding to all operating scenarios constitute a scenario-specific simulation proxy model library. : , In the formula, Indicates the first The lightweight simulation proxy model corresponding to the class runtime scenario label.

[0065] The cloud platform assigns version information to each lightweight simulation proxy model in the scenario-based simulation proxy model library. The version information includes the model version number, the corresponding running scenario label, the applicable working condition range, the pre-simulation accuracy index, and the generation time. A model version library is established for unified management to support subsequent model updates, rollbacks, and re-deployments based on model error feedback.

[0066] Based on the historical distribution of typical operating scenarios for each edge computing node, the cloud platform selects lightweight simulation proxy models corresponding to several operating scenarios that the edge computing node may be involved in from the model version library, and sends them along with the corresponding operating scenario tags and version information to the local storage of the edge computing node. When the local storage capacity of the edge computing node is limited, the cloud platform prioritizes sending lightweight simulation proxy models corresponding to high-frequency scenarios based on the frequency of each operating scenario appearing on the node.

[0067] As a result, the cloud platform generates and distributes lightweight simulation agent models for different scenarios that are significantly compressed in size and can run in real time at the edge. These models and their version information serve as the basis for edge computing nodes to quickly pre-simulate and make autonomous decisions on candidate control actions, and provide a basis for version management for collaborative updates of agent models based on model error feedback.

[0068] The cloud-side platform trains a base simulation model for pre-simulating the execution effects of candidate control actions based on historical operational data and high-precision simulation samples uploaded by each edge computing node. Subsequently, the base simulation model undergoes model compression and parameter quantization to obtain a lightweight simulation proxy model suitable for operation under the limited computing power and storage conditions of edge computing nodes. Furthermore, it generates scenario-specific lightweight simulation proxy models according to operational scenario tags, and distributes them to the corresponding edge computing nodes after model version management. The cloud-side platform is only responsible for the generation and distribution of lightweight simulation proxy models and does not participate in real-time decision-making at the edge.

[0069] In this embodiment, preferably, the method for adjusting the parameters of the lightweight simulation proxy model corresponding to the running scenario of each edge computing node based on the model error feedback amount of each edge computing node in the feedback cycle includes: The model error feedback quantities of each edge computing node in the feedback period are received from the edge side. The model error summary of each edge computing node in the feedback period is then weighted and summed with the error weight coefficient of each component in the model error summary to obtain the comprehensive model error index of each edge computing node in the feedback period. The formula is as follows: , in, This represents the error index of the integrated model. The first in the model error summary Each component error weighting coefficient, and .

[0070] If the comprehensive model error index of the edge computing node in the feedback cycle exceeds the accuracy threshold However, it does not exceed the retraining threshold. At that time, the compressed model increment of the aggregate feedback cycle edge computing node is used to perform online fine-tuning and updating of the lightweight simulation proxy model in the running scenario to which the feedback cycle edge computing node belongs. The formula is as follows: , in, Indicates the updated number Lightweight simulation proxy model parameters for class runtime scenarios. Indicates being in the first The set of edge computing nodes for the class running scenario The aggregation weight is determined based on the number of samples or data quality at each edge computing node. This indicates the update step size.

[0071] If the overall model error index of the edge computing nodes in the feedback cycle exceeds the retraining threshold... Or the number of scene mismatch markers for edge computing nodes in the feedback cycle. When the preset upper limit is exceeded, the lightweight simulation agent model of the running scenario to which the edge computing node belongs is regenerated by combining the key feature summary of the edge computing node in the feedback cycle. If necessary, a new running scenario category is added, and the new version of the lightweight simulation agent model is generated and then sent to the relevant edge computing nodes.

[0072] After obtaining the lightweight simulation proxy model of the running scenario of the feedback cycle edge computing node after online fine-tuning or the regenerated lightweight simulation proxy model of the running scenario of the feedback cycle edge computing node, perform offline verification. If the comprehensive model error index on the verification dataset increases compared with the lightweight simulation proxy model before adjustment, then do not send it to the edge side. When the lightweight simulation proxy model of the feedback cycle edge computing node under the running scenario, which has been fine-tuned and updated online, or the newly generated lightweight simulation proxy model of the feedback cycle edge computing node under the running scenario, is sent to the edge side, if the increase in the number of high-risk cycle markers within the observation window compared to the previous version of the lightweight simulation proxy model under the running scenario exceeds a set threshold, then the lightweight simulation proxy model of the previous version under the running scenario is retrieved from the model version library and sent to the edge side as the final lightweight simulation proxy model under that running scenario, in order to ensure the availability and security of the simulation proxy model on the edge side.

[0073] After the cloud platform completes model updates, retraining, or rollback, it updates the version information of the corresponding lightweight simulation proxy model in the model version library and distributes the new version of the lightweight simulation proxy model and its version information to the relevant edge computing nodes. The edge computing nodes replace their local original models with the new version of the lightweight simulation proxy model and continue to perform rapid pre-simulation and autonomous decision-making in subsequent decision cycles. At the end of the next feedback cycle, the model error feedback is uploaded again, thus forming a closed loop of proxy model collaborative update of "edge pre-simulation decision-making - model error feedback - cloud-side collaborative update - model re-distribution".

[0074] In this invention, the cloud side compresses the high-precision base simulation model into a scenario-specific lightweight simulation proxy model through knowledge distillation, network pruning, and parameter quantization, and then distributes it to the edge side. This enables edge computing nodes to explicitly pre-simulate candidate control actions under limited computing power, storage, and power consumption conditions, solving the problem that the complete physical simulation model is difficult to run in real time at the edge.

[0075] The edge-side uses a lightweight simulation proxy model to quickly pre-simulate candidate actions. It comprehensively evaluates power deviation, voltage over-limit risk, SOC change, cost and carbon emission increment, and control risk before issuing the best action. After risk verification, it takes into account the real-time performance, security, and economic and low-carbon nature of the decision, overcoming the shortcomings of rule-based decision-making, which is rough and purely data-driven without verification.

[0076] By using model error feedback to fine-tune, retrain, or roll back the proxy model online, it can continuously update collaboratively with the operating conditions and maintain the pre-simulation accuracy over a long period of time, thereby improving the adaptability of edge autonomous decision-making.

[0077] Based on Embodiments 1 and 2, this Embodiment 3 provides a distributed resource optimization scheduling system for a park, including: On the edge side, it is used to implement the above-mentioned distributed resource optimization scheduling method in the park; On the cloud side, it is used to implement the above-mentioned distributed resource optimization and scheduling method for the park.

[0078] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0079] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for optimizing and scheduling distributed resources in a park, characterized in that, For the edge side, including: Upload the edge-side operational status representation vector and candidate control action set of each edge computing node at each historical moment to the cloud side; receive and store the lightweight simulation proxy model corresponding to each operational scenario issued by the cloud side. At each decision point, based on the operating scenario of each edge computing node at the current decision point, a lightweight simulation proxy model corresponding to each edge computing node at the current decision point is matched. Input the edge-side operational status representation vector of each edge computing node at the current decision time and the set of candidate control actions into the corresponding lightweight simulation proxy model, and output the pre-simulation output vector corresponding to the execution of each candidate control action by each edge computing node at the current decision time; Based on the pre-simulated output vectors corresponding to each candidate control action of each edge computing node at the current decision time, the target control action of each edge computing node at the current decision time is selected, and it is converted into a device-level control command and sent to the distributed resources under the jurisdiction of the corresponding edge computing node for execution.

2. The method for optimizing and scheduling distributed resources in a park according to claim 1, characterized in that, The pre-simulated output vectors corresponding to each candidate control action executed by each edge computing node at the current decision-making moment include: At the current decision-making moment, the common coupling point power deviation, common coupling point voltage prediction value, energy storage state of charge change, operating cost increment, carbon emission increment, and control risk assessment quantity are corresponding to each candidate control action of each edge computing node.

3. The method for optimizing and scheduling distributed resources in a park according to claim 2, characterized in that, The method for selecting the target control action of each edge computing node at the current decision time, based on the pre-simulated output vectors corresponding to each candidate control action of each edge computing node at the current decision time, includes: Based on the pre-simulated voltage value of the common connection point after each edge computing node executes each candidate control action at the current decision time, calculate the voltage over-limit risk assessment quantity corresponding to each candidate control action of each edge computing node at the current decision time. Based on the common connection point power deviation, voltage over-limit risk assessment, operating cost increment, carbon emission increment and control risk assessment of each candidate control action of each edge computing node at the current decision time, the comprehensive evaluation value of each candidate control action of each edge computing node at the current decision time is obtained. The candidate control action with the smallest comprehensive evaluation value is taken as the target control action of each edge computing node at the current decision moment.

4. The method for optimizing and scheduling distributed resources in a park according to claim 3, characterized in that, Based on the predicted voltage values ​​of the point of common connection after each edge computing node executes each candidate control action at the current decision-making time, the formula for calculating the voltage exceedance risk assessment quantity corresponding to each candidate control action of each edge computing node at the current decision-making time is as follows: , in, Indicates the moment of decision-making Below, the voltage over-limit risk assessment quantity corresponding to the m-th candidate control action of the e-th edge computing node. This indicates taking the maximum value. and This indicates the upper and lower limits of the allowable voltage. Indicates the moment of decision-making Below, the pre-simulated voltage value of the common connection point corresponding to the m-th candidate control action of the e-th edge computing node.

5. The method for optimizing and scheduling distributed resources in a park according to claim 2, characterized in that, After selecting the target control actions of each edge computing node at the current decision time, the control risk of each edge computing node is verified. The target control actions of each edge computing node at the current decision time that pass the verification are taken as the final target control actions of each edge computing node at the current decision time. When the target control action of the edge computing node at the current decision moment satisfies the following formula, it is determined to have passed the control risk verification: , in, Indicates the moment of decision-making Below, the control risk assessment quantity corresponding to the target control action of the e-th edge computing node. This indicates the control risk threshold. Indicates the moment of decision-making Below, the voltage over-limit risk assessment quantity corresponding to the target control action of the e-th edge computing node; If the target control action of the edge computing node fails the verification at the current decision moment, the conservative safety action corresponding to the edge computing node is executed, and the decision period corresponding to the current decision moment is marked as a high-risk period.

6. The method for optimizing and scheduling distributed resources in a park according to claim 5, characterized in that, After converting the target control actions of each edge computing node at the current decision-making moment into device-level control commands and issuing them to the distributed resources managed by each edge computing node for execution, it also includes: Based on the error between the pre-simulated output vector and the actual output vector corresponding to the target control action executed by each edge computing node at the current decision moment, a model error summary of each edge computing node at the current decision moment is constructed. Determine whether the current decision moment is the last decision moment within the feedback cycle. If so, fine-tune the parameters of the lightweight simulation proxy model corresponding to each edge computing node at the current decision moment to obtain the model parameters of the target lightweight simulation proxy model corresponding to each edge computing node. Based on the change in model parameters before and after fine-tuning, construct the compressed model increment for each edge computing node in the feedback cycle. The model error feedback quantity of each edge computing node in the feedback cycle is constructed by using model error summary, key feature summary, scene mismatch marking times, running scene, high-risk cycle marking times, and compressed model increment of each edge computing node in the feedback cycle, and uploaded to the cloud side. This enables the cloud side to adjust the parameters of the lightweight simulation proxy model corresponding to the running scene of each edge computing node.

7. A method for optimizing and scheduling distributed resources in a park, characterized in that, For cloud-side applications, including: Receive edge-side operational status representation vectors and candidate control action sets corresponding to each edge computing node at each historical moment uploaded by the edge side, and classify them according to the operational scenario; Using the edge-side operational status representation vectors and corresponding candidate control actions of each edge computing node at each historical moment under the same operational scenario as training samples, and the actual output vectors obtained by the corresponding edge computing node after executing the candidate control actions at the corresponding moment as sample labels, the base simulation model is trained in different scenarios to obtain the base simulation model corresponding to each operational scenario. The base simulation model under each operating scenario is lightweighted to obtain a lightweight simulation proxy model under each operating scenario. The version information of each lightweight simulation proxy model is marked and stored in the model version library, and then distributed to the edge side.

8. The method for optimizing and scheduling distributed resources in a park according to claim 7, characterized in that, The methods for adjusting the parameters of the lightweight simulation proxy model corresponding to the running scenario of each edge computing node based on the model error feedback amount of each edge computing node in the feedback cycle include: The model error feedback of each edge computing node in the feedback period is received from the edge side. The model error summary of each edge computing node in the feedback period is weighted and summed with the error weight coefficient of each component in the model error summary to obtain the comprehensive model error index of each edge computing node in the feedback period. If the comprehensive model error index of the feedback cycle edge computing node exceeds the accuracy threshold but does not exceed the retraining threshold, the compressed model increment of the feedback cycle edge computing node is aggregated, and the lightweight simulation proxy model in the running scenario of the feedback cycle edge computing node is fine-tuned and updated online. If the comprehensive model error index of the feedback cycle edge computing node exceeds the retraining threshold, or the number of scene mismatch markings of the feedback cycle edge computing node exceeds the preset upper limit, the lightweight simulation proxy model of the running scenario to which the feedback cycle edge computing node belongs is regenerated by combining the key feature summary of the feedback cycle edge computing node.

9. A method for optimizing and scheduling distributed resources in a park according to claim 7, characterized in that, After obtaining the lightweight simulation proxy model of the running scenario of the feedback cycle edge computing node after online fine-tuning or the regenerated lightweight simulation proxy model of the running scenario of the feedback cycle edge computing node, perform offline verification. If the comprehensive model error index on the verification dataset increases compared with the lightweight simulation proxy model before adjustment, then do not send it to the edge side. When the lightweight simulation proxy model of the feedback cycle edge computing node under the running scenario, which has been fine-tuned and updated online, or the lightweight simulation proxy model of the feedback cycle edge computing node under the running scenario, is sent to the edge side, if the increase in the number of high-risk cycle markers within the observation window compared to the previous version of the lightweight simulation proxy model under the running scenario exceeds a set threshold, then the lightweight simulation proxy model of the previous version under the running scenario is retrieved from the model version library and sent to the edge side as the final lightweight simulation proxy model under that running scenario.

10. A distributed resource optimization scheduling system for industrial parks, characterized in that, include: On the edge side, it is used to implement the campus distributed resource optimization scheduling method as described in any one of claims 1-6; The cloud side is used to implement the campus distributed resource optimization scheduling method as described in any one of claims 7-9.