A method and system for dynamic scheduling and collaborative control of grid-microgrid distributed resources

Abstract

This invention relates to the field of control of novel power systems and distribution systems, and discloses a method for dynamic scheduling and collaborative control of distributed resources in a power grid microgrid. This method first constructs multi-dimensional evaluation indicators, and uses a subjective-objective combined weighting and modularity maximization algorithm to achieve dynamic reconstruction and flexible networking of the microgrid aggregate. It then constructs a timely decentralized multi-agent control architecture, utilizing reinforcement learning to output the adjustable active power boundary of the aggregate. Next, it constructs a cloud-edge-device horizontal privacy isolation architecture, introducing a dynamic weight adjustment mechanism based on data quality for federated learning collaborative modeling. Finally, it constructs a non-cooperative-cooperative coupled electricity price game model, using deep Q-learning to generate collaborative scheduling instructions, and performing real-time closed-loop compensation and correction of power tracking deviations at the edge service layer. This invention achieves an engineering closed loop of game optimization—instruction issuance—real-time correction, significantly improving the executability of scheduling and the system's ability to maintain operational constraints while protecting data privacy.

Description

Technical Field

[0001] This invention relates to the field of optimized operation and coordinated control of new power systems and urban power distribution systems, and particularly to a method and system for dynamic scheduling and coordinated control of distributed resources in microgrids for multi-level coordinated operation scenarios of main grid-distribution grid-microgrid in large urban power grids. Background Technology

[0002] With the large-scale integration of distributed resources such as distributed photovoltaics, distributed wind power, user-side energy storage, electric vehicle charging and discharging facilities, and controllable loads into large urban power grids, the operation mode of urban power grids has shifted from the traditional "one-way power supply, centralized dispatch" model to a complex model of "multi-entity participation, multi-timescale fluctuations, and multi-regional coupling." Simultaneously, problems such as voltage sensitivity at the end of urban distribution networks, line load rate fluctuations, local congestion, and reverse power flow have significantly increased, and the contradiction between the autonomous operation of microgrids and park energy systems and the overall safe and economical operation of the urban power grid has become more prominent. To unleash the regulation potential of distributed resources and support coordinated control of main distribution and microgrid systems, there is an urgent engineering need for a feasible dynamic aggregation, autonomous decision-making, privacy-preserving collaborative optimization, and game-driven collaborative dispatch closed-loop control method.

[0003] In the process of realizing this invention, the inventors discovered that the prior art has at least the following drawbacks and shortcomings: 1. Existing distributed resource aggregation methods often rely on fixed partitions or static aggregation, lacking flexible networking mechanisms for "plug-and-play" scenarios and continuously changing operating states. They are difficult to simultaneously consider resource adjustment characteristics and network sensitivity, resulting in unstable aggregation boundaries, large deviations in adjustability assessment, and consequently, problems such as scheduling instructions becoming unexecutable or inconsistent execution effects. For example, the patent application with publication number CN121332760A, entitled "A Distributed Resource Aggregation Control Method and System", achieves efficient and fair aggregation control of distributed resources. Through the coordinated operation of each level of "main network-distribution network-balancing unit-transformer area or microgrid", it solves the problem of high peak-shaving and valley-filling pressure faced by the main network. However, it still does not solve the problem of scenarios with continuously changing operating status, and is essentially still a static aggregation.

[0004] 2. Existing microgrid / park autonomous control mostly adopts centralized or weakly coordinated strategies, which makes it difficult to quickly output executable aggregate adjustable power boundaries under heterogeneous constraints of multiple subjects and multiple types of resources. Furthermore, it is not sufficiently coupled with distribution network side operation constraints such as voltage / power flow / congestion, resulting in conflicts between local optima and system constraints. For example, the patent application with publication number CN121710158A, entitled "A Two-Stage Robust Cooperative Optimization Method and Device for Multi-Microgrid Distribution Systems", uses the day-ahead scheduling plan and the acquired basic operating data of the multi-microgrid distribution system, as well as the power coupling constraints of the tie lines between the distribution network and the multi-microgrid, as inputs to a pre-constructed two-layer cooperative optimization model to obtain an intraday optimized scheduling scheme that meets the requirements of good cooperative operation between the distribution network and the multi-microgrid. This ensures that the optimization results under the worst-case scenario meet the robustness requirements while guaranteeing the economy of the system. However, it still lacks a solution to the problem of insufficient coupling of operating constraints on the distribution network side, which leads to conflicts between local optima and system constraints.

[0005] 3. Cross-aggregate collaborative optimization often relies on centralized data aggregation, which presents a contradiction between data sharing and privacy compliance. Even when federated learning is adopted, traditional approaches often use static weights such as the number of samples, lacking explicit modeling of data integrity, communication timeliness, and data consistency drift. This can easily lead to problems such as low-quality data polluting the global model, slow convergence, or insufficient robustness. For example, the patent application with publication number CN121584688A, entitled "A Collaborative Optimization Scheduling Method for Electric Vehicles Based on a Layered Architecture of 'Network-Station-Vehicle'", introduces a comprehensive contribution degree after the initial distribution of benefits among members within the aggregate, constructs a three-dimensional comprehensive weight model consisting of group contribution degree, dynamic response degree, and controllable reputation degree, and uses the analytic hierarchy process to determine the weight coefficients. However, it still does not process the data and cannot solve the problems of low-quality data polluting the global model, slow convergence, or insufficient robustness.

[0006] 4. In the context of conflicting interests among multiple stakeholders, existing price incentives and coordinated scheduling often remain at the level of simple optimization or single game, lacking an engineering closed-loop mechanism that connects "electricity price game - coordinated scheduling - real-time execution deviation monitoring - closed-loop correction", making it difficult to suppress the spread of command tracking deviation and maintain the continuous satisfaction of operational constraints.

[0007] For example, the patent application with publication number CN121863475A, entitled "Multi-layer Optimization Method for User-Side Shared Energy Storage Considering Joint Pricing and Demand Response", collects basic operating parameter data of regional integrated energy systems, determines the relationship between the power of shared energy storage power stations and the power used by each user in the shared energy storage power stations, calculates the price incentive demand response revenue and peak-shaving cost of the user group, and then constructs and solves a multi-layer optimization model for user-side shared energy storage. However, its main purpose is to reduce the overall operating cost of the user group and effectively reduce the grid power purchase demand during peak hours, but it still cannot solve the above-mentioned problems.

[0008] Therefore, there is an urgent need to study a comprehensive method that can realize flexible networking of distributed resources, autonomous decision-making of aggregates, privacy-preserving collaborative modeling, and form executable collaborative scheduling and real-time closed-loop correction under the framework of electricity price game, so as to support the coordinated operation of main, distribution and micro-distribution systems in large urban power grids. Summary of the Invention

[0009] This application provides a method and system for dynamic scheduling and collaborative control of distributed resources in power grid microgrids, which is used to solve the technical problems existing in the background art.

[0010] The first aspect of this application provides a method for dynamic scheduling and coordinated control of distributed resources in a power grid microgrid, comprising: Based on the micro-topology and operation data of the main and distribution networks of large urban power grids, a multi-dimensional evaluation index set is established for distributed resource nodes to obtain the comprehensive aggregation potential score of the current node. Based on the comprehensive aggregation potential score, the modularity is maximized to realize flexible networking and obtain a set of microgrid aggregates. For each microgrid aggregate, an autonomous operation strategy is determined, and then the adjustable active power boundary that the aggregate can execute within the scheduling cycle is output. Determine the global model parameter vector after federated aggregation based on the model parameter vectors trained locally at the device layer or edge layer of the microgrid aggregate; Based on the global model parameter vector and the adjustable active power boundary, a price game-based collaborative scheduling framework is constructed between the main grid and the microgrid aggregate. The main grid generates dynamic electricity price signals based on supply and demand balance and revenue objectives, while the microgrid aggregate generates response power strategies under the constraints of the electricity price signals and the adjustable boundary. The collaborative scheduling framework is then learned to obtain collaborative scheduling strategies and form executable aggregate command power.

[0011] Furthermore, including: The process of maximizing modularity based on the comprehensive aggregation potential score to achieve flexible networking and obtain a set of microgrid aggregates includes: Based on the comprehensive aggregation potential score and real-time load distribution, a weighted graph model is constructed with distributed resources as nodes and electrical coupling relationships as edges. For any node... and Construct weighted edge weights : ; in, The electrical coupling strength, or equivalent electrical correlation, is estimated from power flow sensitivity, impedance coupling, co-feeder topology distance, or PCC voltage sensitivity matrix to reflect the combined effect of coordinated regulation between the two nodes on system constraints. For nodes The overall aggregation potential score, For nodes The overall aggregation potential score; The modularity is maximized to achieve flexible networking and output the microgrid aggregate partitioning results. Defined as: ; And, it satisfies: ; in, For nodes The weighting degree, For nodes The weighting degree, It is half the sum of the edge weights in the weighted graph model. For nodes The identifier of the aggregate to which it belongs. For nodes The identifier of the aggregate to which it belongs. For indicator functions, when Set the value to 1 if the condition is met, otherwise set it to 0, by maximizing... Obtain the aggregate set It outputs the aggregate member set and aggregate identifier for subsequent autonomous decision-making and collaborative scheduling.

[0012] Furthermore, including: The process of determining an autonomous operation strategy for each microgrid aggregate, thereby outputting the adjustable active power boundary that the aggregate can execute within the scheduling cycle, includes: For each microgrid aggregate Configure a multi-agent autonomous decision-making structure to map distributed power sources, energy storage, controllable loads, and flexible energy-consuming units within the aggregate as a set of agents. Based on the local state information and distribution network-side constraint information of each agent, the aggregate decision state is constructed. And construct a joint reward function based on the operating objectives of the aggregate layer. The autonomous operation strategy of the aggregate is obtained through multi-agent reinforcement learning. ; The net regulated active power of the polymer layer, as a response quantity of the polymer layer, is expressed as: ; in, For discrete-time indexing; As an aggregate At any moment The aggregate decision state is constructed based on local state information and distribution network-side constraint information. ; Joint reward function of aggregates It should include at least operating costs, penalties for exceeding limits, carbon emission constraints, and energy supply reliability constraints, and be expressed as a weighted combination: ; in, This is the operating cost item for the polymer. This is a penalty item for exceeding the limit; For carbon emissions or carbon intensity constraints; This is a penalty for power supply reliability. For the corresponding weight coefficients Based on the autonomous strategy obtained through training, the physical constraints of each unit within the aggregate, and the operational constraint margin of the distribution network, the executable boundary of the aggregate's net regulated active power is output: ; in, and They are polymers The lower and upper limits of adjustable active power are determined by a combination of available resource capacity, operational constraints, and safety margins within the aggregate.

[0013] Furthermore, including: The process of determining the global model parameter vector after federated aggregation based on the model parameter vectors trained locally at the device layer or edge layer of the microgrid aggregate includes: Let the first The global model parameter vector after round federation is The polymer The model parameter vector trained locally at the device layer or edge layer is Then the global model aggregation satisfies: ; And satisfy the weight constraints: ; in, The number of aggregates participating in federated learning. For the first Federal aggregation of aggregates Aggregate weights; The aggregate weight It is explicitly calculated from the confidence level: ; in, The confidence coefficient, calculated through completeness, timeliness, and consistency, is expressed as: ; in, As an aggregate In the statistics window Integrity rate within, As an aggregate In the statistics window Timeliness indicators within the period As an aggregate In the statistics window Internal consistency indicators These are the weighting coefficients for the corresponding indicators.

[0014] Furthermore, including: The polymer In the statistics window Integrity rate Represented as: ; in, The number of valid data entries. This represents the total number of data entries collected. polymer In the statistics window Timeliness indicators within Represented as: ; in, This refers to the data latency statistics from end to cloud or end to edge. For time delay scale parameters; polymer In the statistics window Internal consistency indicators , is represented as: ; in, This is a data feature deviation vector, used to characterize data distribution drift or consistency deviation. For consistency scale parameters, It is a 2-norm.

[0015] Furthermore, including: The framework for coordinated scheduling based on the global model parameter vector and the adjustable active power boundary, constructing a price game between the main grid and the microgrid aggregate, includes: Define the single-moment utility on the mainnet side as follows:

[0016] in, Mainnet-side revenue items; Main grid-side energy supply / peak shaving cost function; System constraint risk penalty items; The single-moment utility on the microgrid aggregate side is defined as follows: ; in, Dynamic electricity price signals generated and published by the main grid side; For the energy consumption benefit or comfort benefit function of the polymer; The internal adjustment cost function of the polymer; This is a collaborative incentive item used to reflect the collaborative benefits of the consortium's participation in system-level congestion mitigation, backup support, and low-carbon contributions.

[0017] Furthermore, including: The step of learning the electricity price game-based collaborative scheduling framework to obtain a collaborative scheduling strategy and form an executable aggregate command power includes: Let the global state be It includes at least the electricity price. Main network operating status, key node voltage / line load rate, and boundary margins of each aggregate; assuming the action is an aggregate command power vector: The reward is defined as a weighted average of the mainnet utility and the utility of each aggregate: ; in, If the weights are the weighting coefficients, then the action value function satisfies the Bellman relation:

[0018] The optimal policy is obtained after training converges: ; This generates the aggregate nominal command power for the coordinated scheduling phase. ; Coordinated scheduling strategy At any moment Output nominal command power The command is then sent to the aggregate, which, under local control and equipment constraints, tracks and executes the command to generate measured power, specifically including: Let the polymer be At any moment The measured net regulated active power is The nominal active power of the instructions issued during the coordinated scheduling phase is , it is The There are several components, and the power tracking error is defined as: ; in, The power of the instructions actually executed by the edge service layer is initially taken as... When the trigger condition is met At this time, the edge service layer uses a proportional compensator to calculate the compensation amount: ; And perform closed-loop updates on the command power: ; in, As the trigger threshold, To compensate for the gain proportionally; To ensure that the compensation is within physically feasible limits, the compensation amount must be constrained by the remaining adjustment capacity and subject to saturation limits. The executable boundary based on the output is: and The feasible range of compensation amount is expressed as follows:

[0019] in, For compensation amount The minimum value, For compensation amount The maximum value; And further ensure that the updated instruction power meets the following requirements: ; When the above executable boundaries are exceeded, the edge service layer... and Implement saturation limiting to bring the compensated instruction power back to the executable range.

[0020] A second aspect of this application provides a dynamic scheduling and collaborative control system for distributed resources in a power grid microgrid, comprising: The aggregate construction module is used to establish a multi-dimensional evaluation index set for distributed resource nodes based on the main and distribution micro-topology and operation data of large urban power grids, obtain the comprehensive aggregation potential score of the current node, and realize flexible networking by maximizing the modularity according to the comprehensive aggregation potential score to obtain the microgrid aggregate set. The operation strategy determination module is used to determine the autonomous operation strategy of each microgrid aggregate, and then output the adjustable active power boundary that the aggregate can execute within the scheduling cycle. The global model building module is used to determine the global model parameter vector after federated aggregation based on the model parameter vectors trained locally at the device layer or edge layer of the microgrid aggregate. The execution module is used to construct a price game-based collaborative scheduling framework between the main grid and the microgrid aggregate based on the global model parameter vector and the adjustable active power boundary. The main grid generates a dynamic price signal based on supply and demand balance and revenue objectives, while the microgrid aggregate generates a response power strategy under the constraints of the price signal and the adjustable boundary. The module also learns the price game-based collaborative scheduling framework to obtain a collaborative scheduling strategy and form an executable aggregate command power.

[0021] A third aspect of this application provides a computer-readable storage medium, characterized in that the storage medium stores a computer program, which, when executed by a processor, implements the method of the first aspect.

[0022] A fourth aspect of this application provides an electronic device including a processor and a memory, the processor being electrically connected to the memory, the memory being used to store instructions and data, and the processor being used to perform the method of the first aspect.

[0023] One or more technical solutions provided in this application have at least the following technical effects or advantages: This invention constructs a multi-dimensional evaluation index system that characterizes resource regulation characteristics and network sensitivity, and adopts flexible networking by combining weighting and maximizing modularity to realize the dynamic reconstruction of distributed resource aggregates. It can adapt to "plug and play" and rapidly changing operating state scenarios, and improve the stability and executability of the aggregation boundary. This invention constructs a timely decentralized multi-agent collaborative control architecture, outputs an adjustable active power boundary for the aggregate, provides executable capability constraints for master-distributor micro-cooperative scheduling, and reduces the risk that the scheduling scheme cannot be implemented. This invention achieves cross-aggregate collaborative modeling without uploading raw data through a collaborative computing architecture with a three-level vertical layering of "cloud-edge-device" and horizontal privacy isolation, thereby improving privacy compliance and data security. This invention introduces a dynamic weight adjustment mechanism based on data integrity, timeliness, and consistency to explicitly calculate aggregate weights and suppress the negative impact of low-quality data on the global model, thereby improving the robustness and convergence accuracy of the collaborative model. Furthermore, this invention constructs a non-cooperative-cooperative coupled electricity price game model and employs deep Q-learning to solve the collaborative scheduling strategy. Simultaneously, it performs deviation monitoring and closed-loop compensation updates at the edge service layer, achieving an engineering closed loop of game optimization, instruction issuance, and real-time correction, thus enhancing the system's constraint maintenance and anti-disturbance capabilities. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the distributed resource dynamic scheduling and collaborative control system architecture for a large-scale urban power grid microgrid provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the dynamic scheduling and collaborative control method for distributed resources in a large-scale urban power grid microgrid provided in an embodiment of the present invention; Figure 3 This is a comparative diagram showing whether or not the residual load of the main network is scheduled according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the tracking error before and after edge closed-loop compensation as described in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0027] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0028] Example 1, as Figure 2 As shown, this application provides a method for dynamic scheduling and collaborative control of distributed resources in a large urban power grid microgrid, the method comprising: Step S1: Construct a multi-dimensional evaluation index system to characterize resource regulation characteristics and network sensitivity; Step S2: Construct a timely, decentralized, multi-agent collaborative control architecture adapted to microgrid aggregates; Step S3: Construct a collaborative computing architecture with a three-tiered vertical structure of "cloud-edge-device" and horizontal privacy isolation; Step S4: Based on the global collaborative model, construct a non-cooperative-cooperative coupled electricity price game model between the main grid and the microgrid aggregate.

[0029] Specifically, step S1, which involves constructing a multi-dimensional evaluation index system to characterize resource regulation characteristics and network sensitivity, includes: Based on the micro-topology and operation data of large urban power grids, a multi-dimensional evaluation index set is established for distributed resource nodes. The indicators include at least regulation response speed, regulation economic cost, electrical coupling (sensitivity to PCC voltage / power flow), and power support sustainability, and each indicator is normalized. For resource nodes... In the Original values ​​of the dimension index Perform max-min normalization: ; in, The normalized index value, and The first in the statistics window The minimum and maximum values ​​of the dimensional indicator can be set to a sliding window. To adapt to changes in resource status and to enable "plug and play" access and exit.

[0030] The comprehensive weights are obtained by combining the Analytic Hierarchy Process (AHP) and the entropy weighting method. Let the subjective weight vector obtained by the AHP be... The objective weight vector obtained by the entropy weight method is The combined weight vector obtained by combining weights satisfy: And its first Each component satisfy:

[0031] in, The subjective and objective weighting fusion coefficient is used to balance between expert experience and data-driven approaches. In engineering, it can be determined by historical operation verification or cross-validation.

[0032] Calculate the nodes based on this. Comprehensive aggregation potential score :

[0033] in, The larger the value, the higher the overall value of the node in the selected adjustment dimension. It can be used as a priority reference for aggregate construction and can be further used to determine the allocation weight of resources participating in different services (peak shaving, frequency regulation, voltage regulation, backup, etc.).

[0034] Based on the comprehensive aggregation potential score and real-time load distribution, a weighted graph model is constructed with distributed resources as nodes and electrical coupling relationships as edges. For any node and Construct weighted edge weights :

[0035] in, The electrical coupling strength or equivalent electrical correlation can be estimated from power flow sensitivity, impedance coupling, co-feeder topology distance, or PCC voltage sensitivity matrix to reflect the combined effect of two-node coordinated regulation on system constraints (voltage / power flow).

[0036] Utilizing maximum modularity enables flexible network topology and outputs microgrid aggregate partitioning results. Modularity Defined as:

[0037] And satisfy:

[0038] in, For nodes The weighting degree, It is half the sum of the edge weights in the graph. For nodes The identifier of the aggregate to which it belongs. For indicator functions, when Set the value to 1 if the condition is met, otherwise set it to 0. Maximize... Obtain the aggregate set It outputs the aggregate member set and aggregate identifier for subsequent autonomous decision-making and collaborative scheduling.

[0039] Specifically, S2 constructs a timely decentralized multi-agent collaborative control architecture adapted to the microgrid aggregate as follows: For each microgrid aggregate Configure a multi-agent autonomous decision-making structure to map distributed power sources, energy storage, controllable loads, and flexible energy-consuming units within the aggregate as a set of agents. Each agent constructs the aggregate decision state based on its local state information and distribution network-side constraint information. And construct a joint reward function based on the operating objectives of the aggregate layer. The autonomous operation strategy of the aggregate is obtained through multi-agent reinforcement learning. This outputs the adjustable active power boundary that the aggregate can execute within the scheduling cycle, providing an executable adjustment range for the master-supplier micro-coordinated scheduling.

[0040] The autonomous strategy and control actions of the aggregate satisfy the following:

[0041] in, For discrete-time indexing; As an aggregate At any moment The aggregate decision-making state is composed of three types of information: the internal resource state of the aggregate, the operational constraint state of the distribution network side, and the external boundary conditions. Among them, the internal resource state of the aggregate includes at least the state of charge (SOC) of energy storage, the available charge and discharge capacity of energy storage, the available output of distributed power sources, the adjustable capacity of controllable loads, and the ramp-up margin of each resource. The operational constraint state of the distribution network side includes at least the voltage of critical nodes, the load rate of critical lines, the power flow limit violation indicator, the local congestion indicator, and the switching power margin of the point of common coupling (PCC). The external boundary conditions include at least the predicted load, the predicted output of renewable energy, the prior knowledge of time-of-use pricing or dynamic pricing, and meteorological forecast information. For aggregate autonomy strategy; For the aggregate at time The net regulated active power serves as the external response of the polymer layer.

[0042] The aggregate autonomous decision objective is expressed in the form of expected discounted return. The aggregate autonomous strategy is obtained by maximizing the cumulative long-term return under a given strategy, specifically as follows:

[0043] in, In order to implement autonomous strategies The expected operator below; The length of the decision time domain; It is a discount factor, and satisfies ; For the aggregate at time The single-step joint reward. The aggregate joint reward function includes at least operating costs, over-limit penalties, carbon emission constraints, and energy supply reliability constraints, and is expressed as a weighted combination:

[0044] in, This is the operating cost item for the aggregate, specifically including electricity purchase and sale costs, energy storage cycle losses, start-up and shutdown losses, etc. The penalties for exceeding limits include voltage limits, power flow limits, and congestion risks. For carbon emissions or carbon intensity constraints; Penalties for power supply reliability include load reduction and failure to meet critical load requirements. The corresponding weighting coefficients can be obtained through engineering tuning or historical data calibration.

[0045] During training, multi-agent reinforcement learning uses the single-step joint reward function as a feedback signal to maximize the expected discounted reward. To obtain a polymer autonomous strategy that meets the requirements of economy, safety, low carbon emissions and reliability. After training convergence, based on the obtained autonomous policy... Furthermore, by combining the physical constraints of each unit within the aggregate with the operational constraints of the distribution network, rolling simulations or online evaluations are performed on the achievable net regulated active power set of the aggregate within the scheduling cycle, thereby determining the executable boundary of the aggregate's net regulated active power.

[0046] in, and These are the lower and upper bounds of the aggregate's adjustable active power, determined jointly by the aggregate's internal resource availability capacity, energy storage SOC constraints, ramping constraints, minimum start-stop time constraints, distribution network-side node voltage constraints, line load rate constraints, and safety margins. In other words, the executable boundary is the feasible output range formed by the aggregate's autonomous strategy under the guidance of the joint reward function and jointly limited by resource physical constraints and distribution network operation constraints. This boundary serves as a hard constraint input for coordinated scheduling in subsequent step S4 and as a saturation limit basis in the closed-loop compensation during the execution phase to ensure the executability of scheduling instructions.

[0047] Specifically, the S3 architecture for constructing a collaborative computing system with a three-tiered vertical hierarchy ("cloud-edge-device") and horizontal privacy isolation is as follows: like Figure 1 The aforementioned architecture constructs a three-tiered, vertically layered collaborative computing architecture consisting of a cloud service layer, an edge service layer, and a device layer, namely, a cloud-edge-device layer. Figure 1 The cloud layer, edge layer, and endpoint layer correspond to each other, and cross-aggregate collaborative modeling is achieved through a horizontal privacy isolation mechanism. Among them, the cloud service layer is deployed on the main network side scheduling center or central server, which is responsible for global model aggregation, global collaborative optimization, electricity price signal generation, and policy distribution. The edge service layer is deployed on the edge computing nodes, park energy management nodes, or aggregation control terminals on the side of each microgrid aggregate, which is responsible for local state aggregation, local model training and updating, instruction execution coordination, power tracking deviation monitoring, and closed-loop compensation correction. The device layer includes distributed resource terminals such as photovoltaic arrays, energy storage systems, smart charging piles, and controllable loads, which are responsible for executing control instructions and uploading operational measurement data.

[0048] Therefore, the relationship between the main network side and the micro-network aggregate side mentioned in this embodiment and the "cloud-edge-device" three-level architecture is as follows: In this embodiment, the edge service layer is located between the main network side and the internal resource devices of the aggregate. It is the entity that carries local control and real-time correction on the aggregate side, rather than a functional module or device itself on the main network side. Federated learning is used to achieve cross-aggregate model collaborative training without uploading the original data; and a dynamic weight adjustment mechanism based on data quality is introduced to evaluate the data integrity, timeliness, and consistency of each micro-network aggregate participating in federated aggregation in real time, calculate the confidence coefficient, and dynamically adjust its federated aggregation weight accordingly, thereby suppressing the negative impact of low-quality data on global model updates.

[0049] Let the first The global model parameter vector after round federation is polymer The model parameter vector trained locally at the device layer or edge layer is Then the global model aggregation satisfies:

[0050] And satisfy the weight constraints:

[0051] in, The number of aggregates participating in federated learning. For the first Federal aggregation of aggregates Aggregate weights.

[0052] To achieve dynamic weight adjustment based on data quality, the aggregate is calculated. Completeness rate within the statistics window :

[0053] in, As an aggregate In the statistics window The number of valid data entries within. As an aggregate In the statistics window Total number of data entries collected within; In this embodiment, the collected data includes at least three categories: the first category is resource-side operational data, including energy storage SOC, energy storage charging and discharging power, active / reactive power output of distributed power sources, controllable load operating status, available equipment capacity, and ramp margin; the second category is network-side measurement data, including key node voltage, line power flow or line load rate, PCC switching power, and congestion and limit exceedance indicators; the third category is external auxiliary data, including load forecast values, renewable energy forecast values, time-of-use pricing or dynamic pricing information, meteorological information, and timestamp information of corresponding data packets. The number of valid data entries refers to the number of data entries remaining after removing missing, duplicate, out-of-bounds, format-abnormal, timestamp-abnormal, and communication verification-failed data.

[0054] Calculate aggregates in the statistics window For timeliness metrics, first define the average latency statistics from end to edge or end to cloud. for:

[0055] in, For the first The sampling timestamp of each data item This refers to the timestamp when the corresponding data is received by the edge service layer or cloud service layer. Based on the average latency statistics, the timeliness index is defined as:

[0056] in, This is a latency scale parameter used to characterize the system's tolerable baseline latency. It can be engineered and tuned based on communication network design specifications, service time constraints, or the latency distribution of historical normal operation samples. For example, the average latency or the 95th percentile latency of historical normal samples can be used as the calibration value. Therefore... This refers to quantities that can be directly calculated based on timestamps. These are pre-set or calibrated scale parameters, not abstract variables that are unavailable.

[0057] Calculate aggregates in the statistics window When defining consistency metrics, first define the data feature deviation vector. for:

[0058] in, As an aggregate In the statistics window Inner The sample mean of each feature, and These are the mean and standard deviation of the corresponding feature in the reference dataset, respectively. The reference dataset can be selected from historical high-quality samples, statistics from the previous round of global aggregation samples, or a validated benchmark sample library. The features include at least key variables such as load power, renewable energy output, energy storage SOC, node voltage, and line load factor. Further, the consistency index is defined as:

[0059] in, It is a 2-norm. The consistency scaling parameter is used to characterize the allowable data distribution deviation benchmark. It can be engineered based on the deviation norm statistics of historical high-quality samples, for example, by taking the mean or 95th percentile of the deviation norm as the calibration value; therefore, the data feature deviation vector Calculated from the current window sample statistic and the reference statistic. The pre-calibrated scale parameters all have clear engineering acquisition methods.

[0060] Confidence coefficients are calculated based on completeness, timeliness, and consistency. :

[0061] in, For the corresponding weight coefficients, satisfying and And the federal aggregation weights It is explicitly calculated from the confidence level:

[0062] This allows for dynamic adjustment of the federated aggregation weights, ensuring that high-quality data aggregates have a higher weight in global model updates, while low-quality data aggregates are either downweighted or removed under engineering strategies. Set the weight to zero at the time.

[0063] Specifically, the S4 model, which relies on a global collaborative model to construct a non-cooperative-cooperative coupled electricity price game, is as follows: Based on the global model formed in step S3 And the adjustable active power boundary output in step S2, to construct a price game-based collaborative scheduling framework between the main grid and the microgrid aggregate; wherein, the global model This is used to integrate the local training results uploaded by each aggregate to form a unified estimation or prediction model for load evolution, renewable energy output changes, price response characteristics, constraint risk levels, and aggregate state transition patterns; in step S4, the global model... This is used to generate or modify the global state representation, utility function parameters, constraint risk assessment quantities, and reward calculation basis required for coordinated scheduling, thereby providing environmental modeling support for the construction of the electricity price game model and deep Q-learning training. That is, the global model obtained in step S3 is not an intermediate result independent of subsequent scheduling, but rather the basis for state perception, benefit assessment, and policy learning in step S4.

[0064] To characterize the "non-cooperative-cooperative coupling" mechanism, utility functions are defined on the main network side and the aggregate side respectively, and feasible cooperative items such as congestion mitigation rewards or ancillary service compensation are introduced at the system level to form cooperative-competitive coupling.

[0065] The single-moment utility on the main network side can be expressed as:

[0066] in, Main grid-side revenue items (such as power supply revenue or system service revenue). Main grid-side energy supply / peak shaving cost function; This is a system constraint risk penalty item.

[0067] As an example of an implementable construction, It can be constructed from the over-limit values ​​of voltage and line congestion:

[0068] in, For the set of key nodes, For the set of critical paths; For node voltage measurements or estimates, Set the threshold for exceeding the allowed upper / lower limits; For line flow, Limits on line capacity; This is the penalty coefficient.

[0069] polymer The single-moment utility can be expressed as:

[0070] in, Dynamic electricity price signals generated and published by the main grid side; For the energy consumption benefit or comfort benefit function of the polymer; The internal adjustment cost function of the polymer; This serves as a collaborative incentive to reflect the benefits of the consortium's participation in system-level congestion mitigation, backup support, and low-carbon contributions. To ensure review readability, a set of feasible construction examples is provided:

[0071] in, The desired power reference trajectory can be obtained from the planned curve or prediction; For coefficients; These represent the improvements in risk of exceeding limits, carbon emission targets, and reliability risks after the aggregate participates in collaboration, compared to the baseline scenario (no collaboration or only local optimization). For example, they can be written as... .

[0072] And polymer response power The executable boundary constraints output in step S2 must be satisfied, i.e., the above. and The defined scope will not be elaborated further here.

[0073] A deep Q-learning algorithm is used to learn the electricity price game-based collaborative scheduling environment, obtain the collaborative scheduling strategy, and form an executable aggregate command power. Let the global state be... It may include at least the electricity price. Main network operating status, key node voltage / line load rate, boundary margins of each aggregate, etc.; assuming the action is the aggregate command power vector.

[0074] in, This means "defined as". For command power vector, This is the command power vector corresponding to the first aggregate. This is the command power vector corresponding to the Gth aggregate; In an alternative embodiment, the action can also be extended to a combined electricity price and power action. The following text is incomplete and cannot be translated. For example, the reward is defined as a weighted average of the mainnet-side utility and the utility of each aggregate:

[0075] in, and These are the weighting coefficients. Based on the reward function, the action value function satisfies the Bellman relation:

[0076] in, This represents the candidate state of the system at the next time step. Candidate actions for the next time step, i.e., the set of possible actions for the next time step. Any action in the list; in this application, the candidate action corresponds to the aggregate instruction power vector that satisfies the executable boundary constraints of each aggregate. The cooperative scheduling strategy is iteratively learned through the above Bellman recursion until the optimal strategy is obtained.

[0077] The optimal policy is obtained after training converges:

[0078] in, The current state;

[0079] This generates the aggregate nominal command power for the coordinated scheduling phase. .

[0080] During the execution phase, power tracking deviations are monitored in real time. When the deviation exceeds a preset threshold, the edge service layer triggers real-time closed-loop correction to compensate and update the collaborative scheduling instructions, thereby suppressing the spread of deviations and maintaining operational constraints. The "collaborative scheduling phase" refers to the phase after cloud / edge collaborative optimization, where the collaborative scheduling strategy... At any moment Output nominal command power The command is then sent to the aggregate; the "execution phase" refers to the aggregate tracking and executing the command under local control and equipment constraints and generating measured power.

[0081] Let the polymer be At any moment The measured net regulated active power is The nominal active power of the instructions issued during the coordinated scheduling phase is , it is The One component. The power tracking error is defined as:

[0082] in, The power of instructions actually issued and executed by the edge service layer is initially taken as the nominal instruction power. The edge service layer consists of edge computing and control nodes deployed on the aggregate side. It receives nominal scheduling instructions and model parameters from the cloud service layer or the main network, and communicates with equipment resources such as energy storage, photovoltaics, charging piles, and controllable loads to acquire real-time measurement data. Therefore, it serves as the aggregate-local execution and closed-loop correction carrier between the main network and the equipment layer. When the triggering condition is met... At that time, the edge service layer calculates the compensation amount using a proportional compensator based on the real-time power tracking deviation:

[0083] And perform closed-loop updates on the actual instruction power executed:

[0084] in, As the trigger threshold, The gain is compensated proportionally. To ensure that the compensation process is carried out within physically feasible limits, the compensation amount must be constrained by the polymer's remaining adjustment capacity and subject to saturation limiting.

[0085] Let the executable boundary of the aggregate output in step S2 be... and The feasible range of compensation amount can then be written as:

[0086] And further ensure that the updated instruction power meets the following requirements:

[0087] When the above executable boundaries are exceeded, the edge service layer... and By implementing saturation limiting, the compensated instruction power returns to the executable range, thus forming a " Issue nominal instructions—execute locally —Calculation deviation —A closed-loop engineering mechanism that compensates for updates and saturates constraints. This closed loop is used to suppress command-measurement deviations caused by prediction errors, equipment nonlinearity, and unmodeled disturbances, ensuring that the command power always meets the executable boundary given in step S2, thereby guaranteeing that the closed-loop correction is achievable and does not introduce new risks of exceeding limits.

[0088] Based on the above method, this embodiment provides the following implementation scheme: In a preferred embodiment of this invention, a method for dynamic scheduling and coordinated control of distributed resources in a large urban power grid microgrid is provided. This method includes the following steps: Step 101: Construct a multi-dimensional evaluation index system to characterize resource regulation characteristics and network sensitivity, and form a weighted graph model; Step 102: Based on the weighted graph model, flexible networking is achieved by maximizing modularity, and the microgrid aggregate partitioning results are output; Step 103: Construct a timely decentralized multi-agent cooperative control architecture for each microgrid aggregate, train it to obtain an autonomous policy, and output an adjustable active power boundary. Step 104: Construct a federated learning collaborative optimization architecture with a three-level vertical layering of "cloud-edge-device" and horizontal privacy isolation to form a global collaborative model; Step 105: Construct a non-cooperative-cooperative coupled electricity price game model based on the global collaborative model, and use deep Q-learning to obtain the collaborative scheduling strategy and generate aggregate command power; Step 106: During the execution phase, the edge service layer monitors the power tracking deviation, triggers closed-loop compensation updates, and maintains the operational constraints satisfied.

[0089] The following section provides a further description of the scheme in Example 1, using specific calculation formulas, architecture, and examples. See the description below for details.

[0090] Step 201: Construct a multi-dimensional evaluation index system and a weighted graph model, and complete the flexible networking of the aggregate. This embodiment selects a typical area of ​​a large urban power grid as the implementation object. This area includes a receiving-end load center on the main grid side, multiple distribution zones, and several park / transformer microgrids. Distributed resources include photovoltaics, energy storage, electric vehicle charging facilities, and controllable loads. Based on the historical operation data and online measurement data of distributed resources in the area, a set of evaluation dimensions is selected. According to the statistics window Performing max-min normalization yields And calculate the combined weights. With comprehensive aggregation potential score The corresponding formulas are as follows:

[0091] Based on the scoring results and real-time load distribution, a weighted graph model is constructed with distributed resources as nodes and electrical coupling relationships as edges. For any node and Construct weighted edge weights :

[0092] in, For nodes With nodes The electrical coupling strength is constructed and normalized using the power flow sensitivity to the point of common coupling voltage. Let the PCC voltage be... ,node With nodes The net active power injections are respectively , Then define sensitivity

[0093] The electrical coupling strength is defined as the normalized value of the sensitivity product:

[0094] in, It can be obtained by linearizing the power flow, sensitivity matrix, or distribution network equivalent Jacobian matrix; the normalization makes This ensures that the edge weights have consistent dimensions and comparability, so that the networking results simultaneously reflect the value of resource regulation and the combined impact on PCC constraints.

[0095] And solve the aggregate partitioning based on maximizing modularity:

[0096] The symbols used have the same meaning as those in S1 above. This embodiment maximizes... Output aggregate set And aggregate member mapping, used for subsequent autonomous decision-making and collaborative scheduling.

[0097] Step 202: Construct a decentralized multi-agent autonomous decision-making system for aggregates and output adjustable power boundaries. For each aggregate A collection of key resource allocation agents within the system Distributed training or online updates are completed under the coordination of the edge service layer. The training objective is:

[0098] The reward function should at least include cost, penalty for exceeding limits, carbon emissions, and reliability terms:

[0099] Net adjusted power boundary of the output aggregate after training convergence:

[0100] in, and The engineering calculations can be obtained by combining the energy storage SOC constraints within the aggregate, the adjustable range of controllable loads, the available output of distributed power sources, and the operational constraint margins on the distribution network side. The calculations are updated in a timely manner based on events (such as resource access / exit, critical line congestion, PCC voltage margin decrease, etc.).

[0101] Step 203: Construct a cloud-edge-device collaborative architecture and form a global collaborative model. Maintain the global model at the cloud service layer. At the edge service layer, it is responsible for aggregate model management, quality assessment, and policy distribution; at the device layer, it completes local training and data preprocessing. Federated learning aggregation is employed.

[0102] It also performs real-time assessments of data quality, calculating integrity rate, timeliness, and consistency.

[0103] Based on this, the confidence level is calculated and the dynamic weights are explicitly obtained:

[0104] This ensures that the aggregated weights are interpretable and traceable, and can suppress the adverse effects of low-quality data on global model updates in engineering.

[0105] Step 204: Constructing a collaborative scheduling and real-time closed-loop correction system based on electricity price game theory. Based on the global collaborative model obtained in step 203 and the adjustable boundaries of each aggregate, the main network side at time... Release dynamic electricity price signals And based on returns and risks, construct the single-time utility of the mainnet side:

[0106] in, Mainnet-side revenue items, Main grid-side energy supply / peak shaving costs This is a penalty item for system operation risks (used to characterize constraint risks such as voltage exceeding limits and line congestion).

[0107] For any polymer The aggregate responds within its executable boundary and constructs a single-moment utility:

[0108] in, The net regulated active power of the polymer, and satisfies:

[0109] For the energy consumption benefit or comfort benefit function of the polymer, The internal adjustment cost function of the polymer. This is an incentive for cooperation.

[0110] During the coordinated scheduling phase, a global state is constructed. With action The action is defined as the aggregate nominal command power vector:

[0111] The main network-side utility and the aggregate-side utility are weighted together to construct the reward for scheduling learning:

[0112] in, These are the weight coefficients. Deep Q-learning is used to solve the collaborative strategy, and the action-value function satisfies the Bellman relation:

[0113] The optimal policy is obtained after training converges. And generate the nominal command power issued during the coordinated scheduling phase:

[0114] During the execution phase, the aggregate tracks and executes the nominal instructions and generates measured power. Let the initial value of the actual instruction power executed by the edge service layer be:

[0115] Define power tracking deviation:

[0116] When the trigger condition is met:

[0117] Real-time closed-loop compensation is then triggered by the edge service layer, and the compensation amount is calculated using a proportional compensator:

[0118] To ensure the physical feasibility of compensation and avoid introducing new risks of exceeding limits, a saturation limit derived from the executable boundary is imposed on the compensation amount:

[0119] Based on this, the actual instruction execution power of the edge service layer is updated as follows:

[0120] After updating, verify and ensure that:

[0121] in, The calculation can be determined by combining the remaining adjustment capacity of the aggregate, the constraint margin and the real-time state; the closed-loop mechanism of "nominal command - deviation monitoring - compensation update - boundary saturation" is used to suppress the spread of command tracking deviation and maintain the continuous satisfaction of operating constraints.

[0122] To further verify the feasibility of the dynamic scheduling and collaborative control method for distributed resources in large-scale urban power grid microgrids described in this invention, a set of simplified simulation examples is constructed based on the above method. The examples select a local urban power distribution scenario, including 6 distributed resource nodes, which are dynamically networked to form 2 microgrid aggregates. The scheduling cycle is 24 discrete time periods, each lasting 1 hour. Each resource node includes distributed power sources, energy storage, and flexible adjustable loads. The main grid side publishes dynamic electricity price signals, and the aggregate side responds under executable power boundary constraints, with closed-loop compensation and correction performed by the edge service layer.

[0123] During the network deployment phase, four evaluation indicators were constructed for the six resource nodes: adjustment response speed, economic efficiency, electrical interconnection potential, and sustainability. Specifically, the adjustment response speeds for the six nodes were set to 0.92, 0.85, 0.78, 0.60, 0.66, and 0.55, respectively; the corresponding economic cost indicators were set to 0.30, 0.40, 0.35, 0.65, 0.55, and 0.70, with economic efficiency positively rounded down to 1 minus cost; the sustainability indicators were set to 0.88, 0.80, 0.72, 0.58, 0.62, and 0.50, respectively; and the electrical interconnection potential indicators were set to 0.90, 0.86, 0.75, 0.64, 0.60, and 0.56, respectively. A combined subjective and objective weighting method was used for comprehensive scoring, with the subjective weight vector set to [0.30, 0.20, 0.30, 0.20] and the objective weight vector set to [0.25, 0.25, 0.30, 0.20], and the fusion coefficient set to 0.60. After normalization and weighted calculation, the comprehensive aggregation potential scores of the six nodes were 1.0000, 0.8146, 0.6500, 0.1780, 0.2642, and 0, respectively. Furthermore, given the electrical coupling matrix between nodes, a weighted graph model was established according to the edge weight construct formula; the optimal modularity Q was obtained by maximizing the modularity, resulting in a value of 0.1082. The corresponding dynamic networking results were: nodes 1, 2, and 3 forming aggregate G1, and nodes 4, 5, and 6 forming aggregate G2. These results demonstrate that the present invention can dynamically aggregate distributed resources based on the comprehensive value of resources and their electrical coupling relationship, thereby forming an aggregate structure suitable for subsequent autonomous decision-making and collaborative scheduling.

[0124] During the autonomous operation and executable boundary output phase of the aggregates, both aggregates are equipped with energy storage and flexible load resources. Aggregate G1 has a rated energy storage capacity of 4.0 MWh, an initial state of charge (SOC) of 0.55, a lower and upper limit of SOC of 0.20 and 0.90 respectively, a maximum discharge power of 1.8 MW and a maximum charge power of 1.5 MW, and flexible load adjustment capabilities of 1.0 MW and 0.6 MW respectively. Aggregate G2 has a rated energy storage capacity of 3.0 MWh, an initial SOC of 0.60, a lower and upper limit of SOC of 0.20 and 0.90 respectively, a maximum discharge power of 1.2 MW and a maximum charge power of 1.0 MW, and flexible load adjustment capabilities of 0.8 MW and 0.5 MW respectively. The energy storage charge / discharge efficiency is assumed to be 0.95. Based on the energy storage SOC, charging and discharging power constraints, and the adjustability of flexible loads, the executable active power regulation boundary of the aggregate is calculated hourly, yielding Pgmin(t) and Pgmax(t) for each time period. This boundary is then used as the hard constraint input for subsequent coordinated scheduling. Taking the 19th time period as an example, the executable boundary for aggregate G1 is [-2.1MW, 1.0MW], and the executable boundary for aggregate G2 is [-1.5MW, 0.8MW]. Therefore, this invention can output executable power boundaries for aggregates that meet resource physical constraints and operational margin requirements, thereby ensuring the feasibility of subsequent scheduling commands.

[0125] In the federated collaborative modeling phase, let the model parameter vectors obtained after local training of aggregates G1 and G2 be θ1=[0.15, 0.97, -0.88] and θ2=[0.22, 1.03, -0.82], respectively. To reflect the dynamic weight adjustment mechanism based on data quality, let the integrity rates of aggregates G1 and G2 be 0.96 and 0.83, respectively, and the latency statistics from end to edge or end to cloud be 0.12 and 0.32, respectively. With the corresponding latency scale parameter d0 set to 0.25, the timeliness indices are 0.6188 and 0.2780, respectively. Furthermore, let the data feature deviations be 0.18 and 0.45, respectively, and the consistency scale parameter κ0 set to 0.50, then the consistency indices are 0.6977 and 0.4066, respectively. With confidence weight coefficients λ1, λ2, and λ3 set to 0.4, 0.3, and 0.3 respectively, the confidence coefficients of the two aggregates were calculated to be 0.7900 and 0.5458, respectively. This yielded federated aggregation weights β1 = 0.5918 and β2 = 0.4082, and further, the global model parameter vector θ = [0.1786, 0.9945, -0.8555]. These results demonstrate that this invention can differentiate the weighting of aggregates participating in federated aggregation based on completeness, timeliness, and consistency, allowing aggregates with higher data quality to occupy higher weights in global model updates, thereby improving the robustness and reliability of the global model.

[0126] During the coordinated scheduling phase of the main grid and microgrid aggregations, the dynamic electricity price signal is constructed according to typical peak-valley characteristics, and the equivalent safe capacity threshold of the main grid is taken as 6.8MW; the internal adjustment cost coefficients of aggregations G1 and G2 are taken as 0.07 and 0.08, respectively. Based on the global model, the system load and risk are estimated, and combined with the aggregation power boundary, the aggregate command power is solved hourly to obtain the nominal scheduling commands within 24 time periods. Figure 3 As shown, without scheduling, the number of periods where the system residual load exceeds the main network safety capacity threshold is 19. After adopting the nominal cooperative scheduling strategy of this invention, the number of periods exceeding the safety capacity threshold is reduced to 9. After further combining edge execution correction, the overall main network residual load curve is smoother and closer to the operating range below the safety threshold. This result shows that this invention can effectively utilize the aggregate regulation capability to reduce the number of high-risk periods on the main network and improve system operation safety.

[0127] During the execution phase, the tracking error of the aggregates to the nominal command is further considered, and closed-loop compensation correction is implemented by the edge service layer. Let the deviation trigger thresholds for aggregates G1 and G2 be 0.12MW and 0.10MW, respectively, and the proportional compensation gain Kp be 0.70. When the deviation between the measured net regulating power and the commanded power exceeds the threshold, the edge service layer calculates the compensation amount according to the compensation formula and applies a saturation limit to the compensation amount based on the aforementioned executable power boundary, thereby updating the actual command execution power of the edge service layer. Figure 4 As shown, without closed-loop compensation, the mean absolute error (MAE) of the aggregate power tracking error is 0.0655 MW, and the root mean square error (RMSE) is 0.0900 MW. After closed-loop compensation, the MAE decreases to 0.0224 MW, and the RMSE decreases to 0.0304 MW. Taking the 19th time period as an example, the nominal command power is G1: 1.000 MW and G2: 0.800 MW. After compensation, the command power is corrected to G1: 0.8954 MW and G2: 0.7139 MW, respectively. The corresponding measured power after compensation is G1: 0.9423 MW and G2: 0.7523 MW, and the compensated command power always satisfies its respective boundary constraints. This result shows that the edge service layer closed-loop compensation mechanism described in this invention can effectively suppress the spread of power tracking deviation without breaking the aggregate's executable boundary, and improve the tracking accuracy of scheduling commands in the actual execution process.

[0128] In summary, the simplified simulation examples above demonstrate that the method described in this invention can achieve dynamic networking of distributed resources, output of executable power boundaries of aggregates, federated collaborative modeling based on data quality, and collaborative scheduling and closed-loop correction between the main network and microgrid aggregates. Furthermore, while ensuring that scheduling instructions are within the executable range, it effectively reduces the number of high-risk periods in the system and improves instruction tracking performance, thus demonstrating clear engineering feasibility.

[0129] Example 2 is based on the same inventive concept as the method for dynamic scheduling and collaborative control of distributed resources in a large urban power grid microgrid in the previous example, such as... Figure 1 As shown, this application provides a dynamic scheduling and collaborative control system for distributed resources in a large-scale urban power grid microgrid. The system and method embodiments in this application are based on the same inventive concept. The system includes a cloud layer, an edge layer, and an end layer. The cloud layer mainly includes the main grid; the edge layer includes edge nodes of the microgrid aggregates; and the end layer includes photovoltaic arrays, energy storage systems, and smart charging piles. This is the "cloud-edge-end" three-level vertically layered collaborative computing architecture consisting of the cloud service layer, edge service layer, and device layer mentioned above in this embodiment, and cross-aggregate collaborative modeling is achieved through a horizontal privacy isolation mechanism. The cloud service layer is deployed at the main grid-side scheduling center or central server, responsible for global model aggregation, global collaborative optimization, electricity price signal generation, and strategy distribution. The edge service layer is deployed at edge computing nodes, park energy management nodes, or aggregation control terminals on each microgrid aggregate side, responsible for local state aggregation, local model training and updating, instruction execution coordination, power tracking deviation monitoring, and closed-loop compensation correction. The device layer includes distributed resource terminals such as photovoltaic arrays, energy storage systems, smart charging piles, and controllable loads, responsible for executing control instructions and uploading operational measurement data. Therefore, the edge service layer is located between the main network side and the internal resource devices of the aggregate, serving as the carrier entity for local control and real-time correction on the aggregate side, rather than a functional module or device itself on the main network side. Federated learning is employed to achieve cross-aggregate model collaborative training without uploading raw data; and a dynamic weight adjustment mechanism based on data quality is introduced to evaluate the data integrity, timeliness, and consistency of each micro-network aggregate participating in federated aggregation in real time, calculate confidence coefficients, and dynamically adjust their federated aggregation weights accordingly, thereby suppressing the negative impact of low-quality data on global model updates. Specifically, the collaborative control structure implemented by this system includes: The aggregate construction module is used to establish a multi-dimensional evaluation index set for distributed resource nodes based on the main and distribution micro-topology and operation data of large urban power grids, obtain the comprehensive aggregation potential score of the current node, and realize flexible networking by maximizing the modularity according to the comprehensive aggregation potential score to obtain the microgrid aggregate set. The operation strategy determination module is used to determine the autonomous operation strategy of each microgrid aggregate, and then output the adjustable active power boundary that the aggregate can execute within the scheduling cycle. The global model building module is used to determine the global model parameter vector after federated aggregation based on the model parameter vectors trained locally at the device layer or edge layer of the microgrid aggregate. The execution module is used to construct a price game-based collaborative scheduling framework between the main grid and the microgrid aggregate based on the global model parameter vector and the adjustable active power boundary. The main grid generates a dynamic price signal based on supply and demand balance and revenue objectives, while the microgrid aggregate generates a response power strategy under the constraints of the price signal and the adjustable boundary. The module also learns the price game-based collaborative scheduling framework to obtain a collaborative scheduling strategy and form an executable aggregate command power.

[0130] In Example 3, based on the same inventive concept as the distributed resource dynamic scheduling and collaborative control method for large urban power grid microgrids in the foregoing examples, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method as described in Example 1.

[0131] A fourth aspect of this application provides an electronic device including a processor and a memory, the processor being electrically connected to the memory, the memory being used to store instructions and data, and the processor being used to perform the method described in Embodiment 1.

[0132] Through the foregoing detailed description of the dynamic scheduling and collaborative control method for distributed resources in large urban power grid microgrids, those skilled in the art will clearly understand that this application provides a platform for dynamic scheduling and collaborative control of distributed resources in large urban power grid microgrids. Therefore, for the sake of brevity, further details are omitted here. Regarding the platform disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant details can be found in the method section.

[0133] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0134] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0135] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0136] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for dynamic scheduling and collaborative control of grid microgrid distributed resources, characterized in that, include: Based on the main and distribution micro-topology and operation data of the power grid in large cities, a multi-dimensional evaluation index set is established for distributed resource nodes to obtain the comprehensive aggregation potential score of the current node. Based on the comprehensive aggregation potential score, the modularity is maximized to realize flexible networking and obtain a set of microgrid aggregates. For each microgrid aggregate, an autonomous operation strategy is determined, and then the adjustable active power boundary that the aggregate can execute within the scheduling cycle is output. Determine the global model parameter vector after federated aggregation based on the model parameter vectors trained locally at the device layer or edge service layer of the microgrid aggregator. Based on the global model parameter vector and the adjustable active power boundary, a price game-based collaborative scheduling framework is constructed between the main grid and the microgrid aggregate. The main grid generates dynamic electricity price signals based on supply and demand balance and revenue objectives, while the microgrid aggregate generates response power strategies under the constraints of the electricity price signals and the adjustable boundary. The collaborative scheduling framework is then learned to obtain collaborative scheduling strategies and form executable aggregate command power.

2. The grid-microgrid distributed resource dynamic scheduling and co-ordination control method of claim 1, wherein, The process of maximizing modularity based on the comprehensive aggregation potential score to achieve flexible networking and obtain a set of microgrid aggregates includes: Based on the comprehensive aggregation potential score and real-time load distribution, a weighted graph model is constructed with distributed resources as nodes and electrical coupling relationships as edges. For any node... and Construct weighted edge weights : ； in, The electrical coupling strength, or equivalent electrical correlation, is estimated from power flow sensitivity, impedance coupling, co-feeder topology distance, or PCC voltage sensitivity matrix to reflect the combined effect of coordinated regulation between the two nodes on system constraints. For nodes The overall aggregation potential score, For nodes The overall aggregation potential score; The modularity is maximized to achieve flexible networking and output the microgrid aggregate partitioning results. Defined as: ; And, satisfies: ; in, For nodes The weighting degree, For nodes The weighting degree, It is half the sum of the edge weights in the weighted graph model. For nodes The identifier of the aggregate to which it belongs. For nodes The identifier of the aggregate to which it belongs. For indicator functions, when Set the value to 1 if the condition is met, otherwise set it to 0, by maximizing... Obtain the aggregate set It outputs the aggregate member set and aggregate identifier for subsequent autonomous decision-making and collaborative scheduling.

3. The method for dynamic scheduling and collaborative control of distributed resources in a power grid microgrid as described in claim 2, characterized in that, The process of determining an autonomous operation strategy for each microgrid aggregate, and then outputting the adjustable active power boundary that the aggregate can execute within the scheduling cycle, includes: For each microgrid aggregate Configure a multi-agent autonomous decision-making structure to map distributed power sources, energy storage, controllable loads, and flexible energy-consuming units within the aggregate as a set of agents. Based on the local state information and distribution network-side constraint information of each agent, the aggregate decision state is constructed. And construct a joint reward function based on the operating objectives of the aggregate layer. The autonomous operation strategy of the aggregate is obtained through multi-agent reinforcement learning. ; Paggnet represents the net regulated active power of the aggregate, which is the aggregate layer's response to the external quantities, and is expressed as: ; wherein, is a discrete time index; is a polymer at time an aggregated decision state constructed based on the local state information and the constraint information on the side of the network configuration, ; Joint reward function for a portfolio of aggregations At least comprising the running cost, the out-of-limit penalty, the carbon emission constraint and the power supply reliability constraint, and expressed by a weighted combination as: ; in, This is the operating cost item for the polymer. This is a penalty item for exceeding the limit; For carbon emissions or carbon intensity constraints; This is a penalty for power supply reliability. These are the weighting coefficients for the polymer operating cost item, the over-limit penalty item, the carbon emission or carbon intensity constraint item, and the energy supply reliability penalty item, respectively. Based on the autonomous strategy obtained through training, the physical constraints of each unit within the aggregate, and the operational constraint margin of the distribution network, the executable boundary of the aggregate's net regulated active power is output: ; wherein, with are polymer The lower and upper bounds of the adjustable active power are determined by the available capacity of the resources in the polymer, the operational constraints and the safety margin.

4. The grid-microgrid distributed resource dynamic scheduling and co-ordination control method of claim 3, wherein, The process of determining the global model parameter vector after federated aggregation based on the model parameter vectors trained locally at the device layer or edge service layer for the microgrid aggregate includes: Let the first The global model parameter vector after round federation is The polymer The model parameter vector trained locally at the device layer or edge service layer is Then the global model aggregation satisfies: ; and satisfying the weight constraint: ; wherein, is the number of aggregators participating in federated learning, is the number of federated aggregations, is the aggregation weight of the aggregator at the th federated aggregation. the aggregated weights explicitly computed from the confidence ; wherein, is the confidence coefficient calculated by the integrity, timeliness and consistency, expressed as: ; in, As an aggregate In the statistics window Integrity rate within, As an aggregate Statistics Window Timeliness indicators within the period As an aggregate In the statistics window t Internal consistency indicators These are the weighting coefficients for completeness rate, timeliness index, and consistency index, respectively.

5. The method for dynamic scheduling and collaborative control of distributed resources in a power grid microgrid as described in claim 4, characterized in that, The polymer In the statistics window t Integrity rate Represented as: ; wherein, is the number of effective data, is the total number of collected data; Polymer In a statistical window t The timeliness indicator Is expressed as: ; in, This refers to the data latency statistics from the device layer to the cloud service layer or from the device layer to the edge service layer. For time delay scale parameters; Polymer In a statistical window t of consistency within , expressed as: ; wherein, is a data feature bias vector, used to characterize data distribution shift or consistency bias, is a consistency scale parameter, is a two-norm.

6. The grid-microgrid distributed resource dynamic scheduling and co-ordination control method of claim 5, wherein, The framework for coordinated scheduling based on the global model parameter vector and the adjustable active power boundary, constructing a price game between the main grid and the microgrid aggregate, includes: The main network side single time utility is defined as: ; in, Mainnet-side revenue items; Main grid-side energy supply / peak shaving cost function; System constraints include risk penalty items; The single-time utility on the microgrid aggregator side is defined as ; in, Dynamic electricity price signals generated and published by the main grid side; For the energy consumption benefit or comfort benefit function of the polymer; The internal adjustment cost function of the polymer; This is a collaborative incentive item used to reflect the collaborative benefits of the consortium's participation in system-level congestion mitigation, backup support, and low-carbon contributions.

7. The method for dynamic scheduling and coordinated control of distributed resources in a power grid microgrid as described in claim 6, characterized in that, The step of learning the electricity price game-based collaborative scheduling framework to obtain a collaborative scheduling strategy and form an executable aggregate command power includes: Let the global state be which contains at least electricity price , main grid operating state, key node voltage / line load rate, each aggregation body boundary margin; Let the action be the aggregation body instruction power vector: The reward is defined as a weighted combination of the main network utility and the utility of each aggregator: ; wherein are weight coefficients, then the action value function satisfies the Bellman relationship: ; wherein, represents a candidate state of the system at the next time, is a candidate action at the next time, and the optimal policy is obtained after training convergence: ; wherein indicates t the action of selecting the time instant; and thereby generating an aggregate nominal instruction power of the cooperatively scheduled phase ; Coordinated scheduling strategy At any moment Output nominal command power The command is then sent to the aggregate, which, under local control and equipment constraints, tracks and executes the command to generate measured power, specifically including: Let the polymer be At any moment The measured net regulated active power is The nominal active power of the instructions issued during the coordinated scheduling phase is , it is The There are several components, and the power tracking error is defined as: ; wherein, is the actual executed command power of the microgrid aggregate, initially taken as When the trigger condition is satisfied The compensation amount is calculated using a proportional compensator: ; And the instruction power is updated in closed loop: ; in, As the trigger threshold, To compensate for the gain proportionally; To ensure that the compensation remains within physically feasible limits, the compensation amount is constrained by the remaining adjustment capacity and subject to saturation limits. The executable boundary based on the output is... and The feasible range of compensation amount is expressed as follows: ; ； wherein is a minimum value of the compensation amount is a maximum value of the compensation amount is a minimum value of the compensation amount is a maximum value of the compensation amount And further ensure that the updated instruction power meets: ; When the above executable boundary is exceeded, the microgrid aggregator side With The saturation limit is executed so that the compensated instruction power is brought back into the executable interval.

8. A dynamic scheduling and collaborative control system for distributed resources in a power grid microgrid, characterized in that, include: The aggregate construction module is used to establish a multi-dimensional evaluation index set for distributed resource nodes based on the main and distribution micro-topology and operation data of large urban power grids, obtain the comprehensive aggregation potential score of the current node, and realize flexible networking by maximizing the modularity according to the comprehensive aggregation potential score to obtain the microgrid aggregate set. The operation strategy determination module is used to determine the autonomous operation strategy of each microgrid aggregate, and then output the adjustable active power boundary that the aggregate can execute within the scheduling cycle. The global model building module is used to determine the global model parameter vector after federated aggregation based on the model parameter vectors trained locally at the device layer or edge service layer of the microgrid aggregate. The execution module is used to construct a price game-based collaborative scheduling framework between the main grid and the microgrid aggregate based on the global model parameter vector and the adjustable active power boundary. The main grid generates a dynamic price signal based on supply and demand balance and revenue objectives, while the microgrid aggregate generates a response power strategy under the constraints of the price signal and the adjustable boundary. The module also learns the price game-based collaborative scheduling framework to obtain a collaborative scheduling strategy and form an executable aggregate command power.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the power grid microgrid distributed resource dynamic scheduling and collaborative control method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, The device includes a processor and a memory, the processor being electrically connected to the memory, the memory being used to store instructions and data, and the processor being used to execute the steps of the power grid microgrid distributed resource dynamic scheduling and collaborative control method according to any one of claims 1 to 7.

Images