Intelligent knowledge collaboration method, device and equipment based on multilateral decision, and medium

Abstract

The invention provides an intelligent knowledge collaboration method, device and equipment based on multilateral decision, and a medium, and the method comprises the steps: firstly defining a global power distribution model based on multiple agents, then achieving the cross-agent knowledge fusion through federated learning, generating a knowledge enhancement model with global cognition, and in a real-time operation stage, achieving the multi-agent knowledge fusion. And dynamically solving an optimal power distribution scheme in combination with a distributed optimization algorithm and a knowledge enhanced model. The method comprises the following steps: defining a global power distribution model based on a plurality of agents, wherein each agent is an autonomous region in the power distribution network; determining a knowledge enhanced model corresponding to each agent by using federal learning according to the global power distribution model through a central server; and determining an optimal power distribution scheme of the intelligent agent through the intelligent agent by utilizing a local knowledge enhanced model and a distributed optimization algorithm.

Description

Methods, devices, equipment, and media based on intelligent knowledge collaboration for multilateral decision-making. Technical Field

[0001] This application relates to the technical field of power allocation, and in particular to a method, apparatus, device and medium based on intelligent knowledge collaboration for multilateral decision-making. Background Technology

[0002] With the rapid development of new power systems, a large number of distributed power sources (such as photovoltaic and wind power), energy storage systems, and flexible loads have been integrated into the distribution network, making it a multi-party decision-making entity. This has transformed the power system from a centralized, top-down control system into a complex system with multiple stakeholders participating and decision-making power decentralized. These entities are geographically dispersed, have independent ownership, and their power output and consumption behaviors are highly random and volatile. Traditional centralized control methods require the collection of global information from all nodes, resulting in high computational burden, high communication bandwidth requirements, and privacy risks.

[0003] Existing technologies include distributed optimization methods such as the Alternating Directional Multiplier Method (ADMM), which can achieve optimization objectives while protecting some privacy. However, these methods typically assume that the data distribution and knowledge level of all participating entities are uniform, ignoring the problem of "uneven knowledge distribution" in reality. For example, some energy storage systems in certain regions possess abundant historical regulation data, while newly connected photovoltaic units lack operational experience. In traditional distributed optimization, this knowledge disparity leads to inefficient collaborative decision-making, with local decisions often deviating from the optimal solution, resulting in slower overall system convergence and even trapping the system in local optima.

[0004] Therefore, there is an urgent need for an intelligent decision-making method that can achieve multilateral knowledge complementarity and collaboration, thereby enhancing the perception and collaborative control capabilities of the entire network's operational status. Summary of the Invention

[0005] This application provides a method, device, equipment, and medium for intelligent knowledge collaboration in multilateral decision-making. The method first defines a global power allocation model based on multiple agents, then achieves cross-agent knowledge fusion through federated learning to generate a knowledge-enhanced model with global cognition. In the real-time operation phase, the optimal power allocation scheme is dynamically solved by combining distributed optimization algorithms and the knowledge-enhanced model.

[0006] This application provides a method for intelligent knowledge collaboration based on multilateral decision-making, comprising: defining a global power allocation model based on multiple agents, wherein each agent is an autonomous region in a power distribution network; determining a knowledge-enhanced model for each agent using federated learning through a central server based on the global power allocation model; and determining the optimal power allocation scheme for each agent using its local knowledge-enhanced model and a distributed optimization algorithm. In some embodiments, before performing the step of determining the knowledge-enhanced model for each agent using federated learning through the central server based on the global power allocation model, the method further comprises: calculating a corresponding knowledge level index for each agent and uploading it to the central server; determining a knowledge distribution entropy based on all the knowledge level indices through the central server; wherein the knowledge distribution entropy is the degree of uneven distribution of global knowledge; and when the knowledge distribution entropy is greater than a preset knowledge distribution entropy threshold, performing the step of determining the knowledge-enhanced model for each agent using federated learning through the central server based on the global power allocation model.

[0007] In some embodiments, the step of determining the knowledge-enhanced model corresponding to each agent using federated learning via a central server based on the global power allocation model includes: distributing the global power allocation model to all agents via the central server, so that the agents can train the global power allocation model using local datasets to obtain local loss values ​​corresponding to local loss functions and model parameter update amounts of the global power allocation model; encrypting the model parameter update amounts; uploading the encrypted model parameter update amounts and the local loss values ​​to the central server; calculating the target model parameter update amount using a weighted average method based on the encrypted model parameter update amounts uploaded by all agents via the central server; and updating the global power allocation stored in the central server using the target model parameter update amount. The model parameters of the model are determined; based on the update amount of the target model parameters, it is determined whether the global power allocation model stored in the central server has converged, and based on the received local loss value, it is determined whether it meets the objective of federated learning; if it converges and meets the objective of federated learning, the knowledge alignment level of the global power allocation model stored in the central server is evaluated; if the evaluation passes, the agent's current global power allocation model is determined to be a knowledge-enhanced model; if the evaluation fails, a secondary penalty term is added to the local loss function; the steps of using the central server to distribute the global power allocation model to all agents are re-executed.

[0008] In some embodiments, the step of determining whether the received local loss value meets the objective of federated learning includes processing the local loss values ​​corresponding to all agents using the agent's knowledge weights and local dataset to obtain a target loss value; if the target loss value is less than a preset loss value, then it is determined that the objective of federated learning is met.

[0009] In some embodiments, the distributed optimization algorithm includes an iterative formula of ADMM; the step of determining the optimal power allocation scheme of the agent by utilizing the local knowledge-enhanced model and the distributed optimization algorithm includes: at the beginning of the current scheduling cycle, the agent uses the knowledge-enhanced model to calculate the initial exchange power with neighboring agents based on the agent's state vector; based on the initial exchange power and the exchange power with neighboring agents in the previous scheduling cycle, the agent determines the exchange power with neighboring agents in the current cycle; the agent sends the exchange power with neighboring agents in the current cycle to the neighboring agents, so that the neighboring agents can utilize the iterative formula of ADMM based on the received exchange power in the current cycle. The power allocation scheme is determined by substituting the formula, wherein the power allocation scheme includes its own decision variables, auxiliary variables, and dual variables; the steps of calculating the initial exchange power with neighboring agents based on the agent's state vector using the knowledge-enhanced model at the beginning of the current scheduling cycle are repeated until the exchange power with neighboring agents in the current cycle is sent to the neighboring agents so that the neighboring agents can determine the corresponding power allocation scheme using the iterative formula of ADMM based on the received exchange power in the current cycle, until the power allocation schemes of all agents converge; the power allocation scheme of each agent at the time of convergence of the power allocation schemes of all agents is determined as the corresponding optimal power allocation scheme.

[0010] In some embodiments, the method further includes: detecting whether it is necessary to switch to an emergency operation mode; if a switch is required, constructing an emergency operation island; wherein the emergency operation island includes communicable neighboring agents; adjusting the local optimization objective function to an emergency operation optimization objective; adjusting the state vector of the agent to a lightweight knowledge vector; defining the update rule of the lightweight knowledge vector; and determining the optimal power allocation scheme for the agents within the emergency operation island using the local knowledge-enhanced model and the iterative formula of ADMM.

[0011] This application embodiment also provides a multi-party decision-oriented intelligent knowledge collaboration device, comprising: a definition unit, used to define a global power allocation model based on multiple agents, the global power allocation model including input quantities, output quantities, local optimization objective functions and local operating constraints, the agents being autonomous regions in a power distribution network, the input quantities including the state vectors of the agents, and the output quantities including the decision vectors of the agents; a first determination unit, used by a central server to determine the knowledge-enhanced model corresponding to each agent based on the global power allocation model using federated learning; and a second determination unit, used by the agents to determine the optimal power allocation scheme for the agents based on the state vectors using the local knowledge-enhanced model and a distributed optimization algorithm.

[0012] In some embodiments, the state vector includes a knowledge level index; the device further includes: the agent calculating the corresponding knowledge level index and uploading it to a central server; the central server determining the knowledge distribution entropy based on all the knowledge level indices; wherein the knowledge distribution entropy is the degree of uneven distribution of global knowledge; when the knowledge distribution entropy is greater than a preset knowledge distribution entropy threshold, the central server performs the step of determining the knowledge-enhanced model corresponding to each agent using federated learning according to the global power allocation model.

[0013] In some embodiments, the first determining unit includes: a distributing unit, configured to: distribute the global power allocation model to all agents via the central server, so that the agents can train the global power allocation model using local datasets to obtain local loss values ​​corresponding to local loss functions and model parameter update amounts of the global power allocation model; encrypt the model parameter update amounts; and upload the encrypted model parameter update amounts and the local loss values ​​to the central server; a calculation unit, configured to: calculate a target model parameter update amount using a weighted average method based on the encrypted model parameter update amounts uploaded by all agents via the central server; and update the model parameters of the global power allocation model stored in the central server using the target model parameter update amount; and a judging unit, configured to: judge the... The process includes: a central server unit to evaluate the knowledge alignment level of the global power allocation model stored in the central server if it converges and meets the federated learning objective; a third determination unit to determine that the agent's current global power allocation model is a knowledge-enhanced model if the evaluation passes; a re-execution unit to add a quadratic penalty term to the local loss function if the evaluation fails; and re-execution of the steps of using the central server to distribute the global power allocation model to all agents, and determining whether the global power allocation model stored in the central server converges based on the target model parameter update amount, and whether it meets the federated learning objective based on the received local loss value.

[0014] In some embodiments, the determination unit is specifically used to: process the local loss values ​​corresponding to all agents using the knowledge weights of the agents and the local dataset to obtain the target loss value; if the target loss value is less than a preset loss value, then it is determined that the target of federated learning is met.

[0015] In some embodiments, the distributed optimization algorithm includes the iterative formula of ADMM; the second determining unit specifically performs the following: at the beginning of the current scheduling cycle, the agent uses a knowledge-enhanced model to calculate the initial exchange power with neighboring agents based on the agent's state vector; based on the initial exchange power and the exchange power with neighboring agents in the previous scheduling cycle, the exchange power with neighboring agents in the current cycle is determined; the exchange power with neighboring agents in the current cycle is sent to the neighboring agents so that the neighboring agents, according to the received exchange power in the current cycle, use the iterative formula of ADMM to determine the corresponding power allocation scheme, wherein... The power allocation scheme includes its own decision variables, auxiliary variables, and dual variables; repeating the steps from the beginning of the current scheduling cycle, where the agent uses a knowledge-enhanced model to calculate the initial exchange power with neighboring agents based on the agent's state vector, to the step of sending the exchange power with neighboring agents for the current cycle to the neighboring agents so that the neighboring agents can determine the corresponding power allocation scheme based on the received exchange power for the current cycle using the iterative formula of ADMM, until the power allocation schemes of all agents converge; the power allocation scheme of each agent at the point of convergence of the power allocation schemes of all agents is determined as the corresponding optimal power allocation scheme.

[0016] In some embodiments, the apparatus further includes: a detection unit, configured to detect whether a switch to an emergency operation mode is required; if a switch is required, construct an emergency operation island; wherein the emergency operation island includes communicable neighboring agents; adjust the local optimization objective function to an emergency operation optimization objective; adjust the state vector of the agent to a lightweight knowledge vector; define the update rule of the lightweight knowledge vector; and determine the optimal power allocation scheme for the agents within the emergency operation island using a local knowledge-enhanced model and the iterative formula of ADMM.

[0017] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method based on intelligent knowledge collaboration for multilateral decision-making.

[0018] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for intelligent knowledge collaboration based on multilateral decision-making.

[0019] The above embodiments provide a method, apparatus, device, and medium for intelligent knowledge collaboration based on multilateral decision-making. The method first defines a global power allocation model based on multiple agents, then achieves cross-agent knowledge fusion through federated learning to generate a knowledge-enhanced model with global cognition. During real-time operation, the optimal power allocation scheme is dynamically solved by combining a distributed optimization algorithm with the knowledge-enhanced model. The method includes defining a global power allocation model based on multiple agents, where each agent is an autonomous region in a power distribution network; a central server determines the knowledge-enhanced model corresponding to each agent using federated learning based on the global power allocation model; and each agent determines its optimal power allocation scheme using its local knowledge-enhanced model and a distributed optimization algorithm. Attached Figure Description

[0020] Figure 1 illustrates a flowchart of a method for intelligent knowledge collaboration based on multilateral decision-making according to some embodiments; Figure 2 illustrates a schematic diagram of a device for intelligent knowledge collaboration based on multilateral decision-making according to some embodiments. Detailed Implementation

[0021] To make the objectives and implementation methods of this application clearer, the exemplary implementation methods of this application will be clearly and completely described below with reference to the accompanying drawings of the exemplary embodiments of this application. Obviously, the exemplary embodiments described are only some embodiments of this application, and not all embodiments.

[0022] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0023] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence, unless otherwise specified. It should be understood that such terms are interchangeable where appropriate.

[0024] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0025] To address the aforementioned technical problems, embodiments of this application provide a method, apparatus, device, and medium for intelligent knowledge collaboration in multilateral decision-making. This method first defines a global power allocation model based on multiple agents, then achieves cross-agent knowledge fusion through federated learning to generate a knowledge-enhanced model with global cognition. During real-time operation, the optimal power allocation scheme is dynamically solved by combining a distributed optimization algorithm with the knowledge-enhanced model.

[0026] The method in this application embodiment is applied in the technical field of power system operation control and artificial intelligence, and in particular relates to a new type of power system that includes multiple stakeholders such as power sources, grids, loads and storage.

[0027] Figure 1 illustrates, exemplarily, a flowchart of a method for intelligent knowledge collaboration in multilateral decision-making, according to some embodiments. The method includes steps S100-S300.

[0028] S100. Define a global power allocation model based on multiple agents. The global power allocation model includes inputs, outputs, local optimization objective functions, and local operating constraints. The agent is an autonomous region in the power distribution network. The inputs include the state vector of the agent, and the outputs include the decision vector of the agent.

[0029] In this embodiment, the autonomous region can be a transformer substation or a microgrid, etc. Each agent is responsible for managing its internal distributed photovoltaic, energy storage systems, and flexible loads. Each agent locally maintains a global power allocation model, and its local optimization objective is to minimize its own operating costs and power imbalance penalties.

[0030] In some embodiments, the step of defining a global power allocation model based on multiple agents includes: defining input quantities in the global power allocation model, wherein the input quantities include the state vectors of the agents, and the state vectors include knowledge level indicators.

[0031] Specifically, the state vector of agent i Defined as: ;in, The state of charge of the energy storage system of agent i. The real-time photovoltaic power output of energy body i The load power demand of energy body i It is the amount of electricity that agent i purchases from the main power grid. Let be the node voltage amplitude of agent i. Let i be the node voltage phase angle of agent i. Hourly electricity price Let be the state vector of the neighboring agents of agent i; specifically, the state vector of the neighboring agent j of agent i can be: , The exchange power between agent i and its neighboring agent j. The state of charge of the energy storage system of neighboring intelligent agent j; Here, is a knowledge level index for agent i, used to quantify the richness of agent i's knowledge. .

[0032] Specifically, The expression is as follows: ;in, For the size of the local dataset, To balance the weights, a constant between 0 and 1 is used to balance the influence of knowledge quantity and model quality on the knowledge level metric. For example, when... When it is 1, It is entirely determined by the amount of data, when When it is 0, It is entirely determined by the accuracy of the model, which determines when When the value is 0.5, both are equally important. The normalized mean square error of the agent's local global power allocation model on the validation set is expressed as follows: ;in, It is the size of the validation set. It is the k-th state sample in the validation set. This is the corresponding optimal historical decision, which is the preset decision; The prediction decision of the local global power allocation model of energy entity i under the k-th state sample; It is the mean of all historical best decisions in the validation set.

[0033] Define the output quantity in the global power allocation model, whereby the output quantity includes the decision variables of the agent; specifically, the decision variables of agent i are defined as: ;in The charging and discharging power of the energy storage system of agent i is given. When the value is positive, the energy storage system is discharging; conversely, when the value is negative, the energy storage system is charging. The amount of photovoltaic curtailment power of energy source i The load reduction power of energy body i Power exchange between agent i and its neighbor agent j.

[0034] Define the local optimization objective function of the agent in the global power allocation model.

[0035] In this embodiment of the application, the local optimization objective function is to minimize its own operating cost and power imbalance penalty.

[0036] Specifically, the local optimization objective function is defined as: ;in, As a weighted average of energy costs, As a weight for energy storage losses, Weighting comfort costs, Weighting for power imbalance penalty. Weighting coefficient. - It can be trained using historical data or set based on expert experience, or it can be dynamically adjusted online, for example, based on real-time electricity price fluctuations. . Let i be the decision variable of agent i. Let be the decision variables of the neighboring agent j. The set of neighbors of agent i For energy costs. For energy storage loss costs, For comfort costs, This represents the power imbalance. Among these, comfort cost... It can effectively reflect the degree of user dissatisfaction caused by load reduction, and this quantity can be obtained through user satisfaction surveys or fitting historical behavior data.

[0037] The last term in the local optimization objective function is a cooperative penalty term, which encourages the agent's decisions to align with those of its neighbors, thereby achieving multilateral cooperation. (Weight of the multilateral cooperation penalty term) It can be dynamically adjusted according to the communication status.

[0038] This local optimization objective function balances economy, user experience, and system security, guiding the agent to minimize its own cost while taking into account the overall power balance.

[0039] The expressions for energy cost, energy storage loss cost, comfort cost, and power imbalance are as follows: ; ; ; ; ;in, The amount of electricity that energy entity i purchases from the main grid. It is the length of the scheduling period. It is the loss coefficient. For comfort cost coefficient, For the number of loads, This is the set value of load k. This is the actual power supply.

[0040] Establish the local operational constraints of the agents in the global power allocation model.

[0041] In this embodiment of the application, the local operating constraints include power balance constraints, energy storage constraints, photovoltaic output constraints, load shedding constraints, and power exchange constraints.

[0042] Power balance constraints: Energy storage constraints: ; Photovoltaic output constraints: Load reduction constraints: Power exchange constraints: ;in, The minimum state of charge constraint for the energy storage system of agent i; The maximum state of charge limit for the energy storage system of agent i; The minimum charging power of the energy storage system for agent i; The maximum discharge power of the energy storage system of agent i; Let be the maximum value of the exchange power between agent i and its neighbor agent j.

[0043] S200: The central server uses federated learning to determine the knowledge-enhanced model corresponding to each agent based on the global power allocation model.

[0044] In this embodiment of the application, the federated learning process is executed under the coordination of the central server.

[0045] In some embodiments, before executing the step of the central server determining the knowledge-enhanced model corresponding to each agent using federated learning based on the global power allocation model, the method further includes: the agent calculating the corresponding knowledge level index and uploading it to the central server; in this embodiment, according to the above-mentioned... The expression is used to calculate the knowledge level index. Understandably, each agent calculates its corresponding knowledge level metric and uploads it to the central server.

[0046] The central server determines the knowledge distribution entropy based on all the aforementioned knowledge level indicators; wherein, the knowledge distribution entropy represents the degree of unevenness in the global knowledge distribution; specifically, the knowledge distribution entropy... for: When the knowledge distribution entropy is greater than the preset knowledge distribution entropy threshold, the central server executes the step of determining the knowledge-enhanced model corresponding to each agent using federated learning based on the global power allocation model.

[0047] In this embodiment, the preset knowledge distribution entropy threshold is set according to the regional differences in the power system.

[0048] In this embodiment, when the knowledge distribution entropy is greater than a preset knowledge distribution entropy threshold, it indicates that the current global knowledge distribution is uneven. For example, agent A has extensive operational experience, while agent B lacks operational experience.

[0049] In some embodiments, the step of the central server determining the knowledge-enhanced model corresponding to each agent using federated learning based on the global power allocation model includes: the central server distributing the global power allocation model to all agents, so that the agents can train the global power allocation model using local datasets to obtain local loss values ​​corresponding to local loss functions and model parameter update amounts of the global power allocation model; encrypting the model parameter update amounts; and uploading the encrypted model parameter update amounts and the local loss values ​​to the central server.

[0050] In this embodiment, the data security of the multilateral agents is protected. Throughout the process, each agent does not need to upload sensitive local raw operational data; they only exchange model update quantities or power plans, effectively protecting the data privacy and trade secrets of the multilateral agents.

[0051] In one example, a local dataset contains several training data points, including information on light intensity, load demand, and electricity prices.

[0052] In this embodiment, the step of the central server distributing the global power allocation model to all the agents includes the central server broadcasting the global power allocation model to all online and communicating agents. Each agent uses its local dataset to train the received global power allocation model locally, obtaining the updated model parameters of the global power allocation model and the local loss value corresponding to the local loss function.

[0053] Specifically, the process of each agent training the received global power allocation model locally using its local dataset can be represented as follows: ;in, The learning rate is determined using an exponential decay strategy. The expression for the learning rate is: initial value ; These are the updated model parameters; These are the current model parameters; This represents the number of iterations. The gradient of the local loss function Represents the gradient symbol.

[0054] The model parameter update amount can be obtained by subtracting the updated model parameters from the current model parameters of the global power allocation model. minus The expression is: ;in This represents the amount of time the model parameters are updated.

[0055] In this embodiment of the application, the local loss function is: ,in , It is a loss function. This is the historically optimal decision, which is predetermined. It is a model-based prediction and decision-making process.

[0056] The central server calculates the target model parameter update amount by using a weighted average method on the encrypted model parameter update amounts uploaded by all agents, and uses the target model parameter update amount to update the model parameters of the global power allocation model stored in the central server.

[0057] In some embodiments, the central server calculates the target model parameter update amount by using a weighted average method for the encrypted model parameter updates uploaded by all agents, which can be accomplished using the following formula: ; This weighted strategy ensures that the experience of knowledge-rich areas has a higher weight in the global power allocation model stored in the central server, realizing the directional flow of knowledge from rich areas to barren areas, which is the core of solving the problem of uneven knowledge distribution.

[0058] Based on the target model parameter update amount, it is determined whether the global power allocation model stored in the central server has converged, and based on the received local loss value, it is determined whether it meets the objective of federated learning. In some embodiments, the step of determining whether the global power allocation model stored in the central server has converged based on the target model parameter update amount includes: determining whether the target model parameter update amount is not greater than a preset update amount; specifically, it is completed according to the following formula: If it is not greater than the preset update amount If so, it is determined that the global power allocation model stored in the central server has converged.

[0059] In some embodiments, the step of determining whether the received local loss value meets the objective of federated learning includes: processing the local loss values ​​corresponding to all agents using the agent's knowledge weights and the local dataset to obtain the target loss value; specifically, processing the local loss values ​​corresponding to all agents using the agent's knowledge weights and the local dataset to obtain the target loss value includes weighted summation of the local loss values ​​corresponding to all agents using the agent's knowledge weights and the local dataset, and then averaging the sums to obtain the target loss value.

[0060] The target loss value can be calculated using the following formula: Wherein, the parameters of the global power allocation model are assumed to be... Each agent has a local dataset. , It is a local loss function. Knowledge weights are used for weighted aggregation, making model updates in knowledge-rich regions more influential. Based on knowledge level indicators Considering the importance of power system regions, its expression is: ;in, This is an importance index for agent i, used to measure its criticality and influence in the operation of the power system. Independent of knowledge level, it is an index based on the physical characteristics and operational needs of the power system, effectively reflecting weight differences in dynamic decision-making. It prevents weights from being entirely dominated by knowledge level, ensuring that even if the knowledge level of a certain area is temporarily low, as long as it plays a key role in the power network (e.g., connecting critical loads or possessing large-scale generation capacity), its decision-making needs can still be fully reflected in the global power allocation model. This demonstrates a deep alignment with the research object of power systems.

[0061] Importance indicators It should be defined in close conjunction with the characteristics of multiple entities such as "source", "grid", "load" and "storage" in the new power system. As a composite index, it can be constructed and calculated from the following aspects: (1) Load level and criticality. Ensuring the power supply of critical loads is one of the highest priority tasks of the power system. The correctness of decision-making in a region that bears more critical loads is crucial to system security, and therefore should be given higher importance. The load level and criticality are defined as the proportion of critical loads within the jurisdiction of agent i to the total load. Its expression is: ;in, The critical load power for agent i; Let be the total load power of agent i.

[0062] (2) Generation / Storage Capacity and Regulation Capability: Regions with higher generation or larger storage capacity have a stronger ability to maintain system power balance and provide peak-shaving and frequency regulation services, and their behavior has a greater impact on the entire network. Therefore, the criticality of generation / storage capacity is defined as the proportion of the rated power of distributed power sources and the rated capacity of energy storage systems within the jurisdiction of agent i to the total system capacity. Its expression is as follows: ;in, Let i be the photovoltaic capacity of intelligent agent i; Let i be the energy storage capacity of agent i.

[0063] (3) Network Topology and Hub Level: Nodes located at the hub of the power grid (such as important substations or nodes connecting multiple microgrids) have a greater impact on the overall network structure due to their power fluctuations. Their stability is crucial to the overall network security. Based on the power grid topology, the proportion of the node degree of agent i to the total node degree is calculated. The expression for network topology and hub level can then be written as follows: ;in Let be the degree of the node where agent i is located.

[0064] The above multiple dimensions are combined using a weighted linear combination to form a unified [dimensionality / structure]. Its expression is: ;in , , The importance weight coefficients are, and satisfy the following conditions: .

[0065] In this embodiment, federated learning is not an independent prediction task; its training objective is consistent with the operational cost objective of subsequent distributed optimization. This is achieved by minimizing the local loss function. Its essence is to make the model Learn to approach the historical optimal decision This allows the initial exchange power between the agent and neighboring agents calculated using the knowledge-enhanced model to be closer to the global optimum, thus accelerating convergence.

[0066] If the target loss value is less than the preset loss value, then it is determined that the target of federated learning is met.

[0067] In this embodiment of the application, the target loss value is less than the preset loss value, that is, the target loss value is minimized, which meets the goal of federated learning.

[0068] If convergence is achieved and the federated learning objective is met, the knowledge alignment level of the global power allocation model stored in the central server is evaluated.

[0069] Specifically, in order to quantify the degree of convergence of the global knowledge distribution from uneven to aligned during the federated learning process, a knowledge alignment level assessment is conducted.

[0070] The knowledge alignment level is defined as follows: ;in, For knowledge alignment level; The global average knowledge level index is expressed as follows: .

[0071] If the evaluation passes, the agent's current local global power allocation model is determined to be a knowledge-enhanced model.

[0072] Specifically, when the knowledge alignment level is not less than 0.95, the knowledge alignment is considered complete and the evaluation is passed.

[0073] In this embodiment of the application, during the federated learning process, each agent obtains a knowledge-enhanced model that contains multilateral knowledge from the entire network and can adapt well to local conditions.

[0074] If the evaluation fails, a secondary penalty term is added to the local loss function; the steps of distributing the global power allocation model to all agents using the central server, determining whether the global power allocation model stored in the central server has converged based on the target model parameter update amount, and determining whether it meets the objective of federated learning based on the received local loss value are re-executed.

[0075] Specifically, in order to constrain the local global power allocation model The global power allocation model stored in the central server deviates from the previous round. Add a secondary penalty item: .

[0076] S300. Based on the state vector, the agent uses a local knowledge-enhanced model and a distributed optimization algorithm to determine the optimal power allocation scheme for the agent.

[0077] Within each real-time scheduling cycle, each agent uses a distributed optimization algorithm to perform iterative calculations based on its local knowledge-enhanced model, coordinating the output and power consumption behavior of its internal photovoltaic, flexible load, and energy storage systems. It also achieves power mutual assistance between regions by exchanging power with neighboring agents, decomposing the global optimization problem into multiple sub-problems and solving for the optimal power allocation scheme of the system.

[0078] In this embodiment, intelligent knowledge collaboration is achieved based on distributed optimization algorithms and federated learning to realize knowledge complementarity and alignment among multiple intelligent agents, as well as an efficient and fair method for allocating device power.

[0079] In some embodiments, the distributed optimization algorithm includes the iterative formula of ADMM; the step of the agent determining the optimal power allocation scheme based on the state vector using a local knowledge-enhanced model and the distributed optimization algorithm includes: at the beginning of the current scheduling period, the agent uses the knowledge-enhanced model and the agent's state vector to calculate the initial exchange power with neighboring agents; specifically, at the beginning of each scheduling period, each agent calculates the initial exchange power with neighboring agents based on the knowledge-enhanced model and the current state vector. ;in, Bang Learning's knowledge enhancement model; This represents the initial switching power.

[0080] Based on the initial exchange power and the exchange power with neighboring agents in the previous scheduling cycle, the exchange power with neighboring agents in the current cycle is determined.

[0081] Specifically, it can be done using the following formula: ; Among them, regular terms Limit power switching from the initial recommendations given by federated learning to maintain system stability. This represents all local operational constraints, which ensure that all decisions are solved within the safe operation specifications of the power system, and k represents the k-th scheduling cycle.

[0082] The knowledge-guided weights decrease as the scheduling cycle increases, allowing knowledge-poor agents to rely more on federated learning guidance in the early stages and gradually rely on local real-time information over time, thus achieving a smooth transition. The expression is: ;in, The initial value for knowledge-guided weights; is the time constant.

[0083] As can be seen from the above formula, the knowledge guidance weight changes dynamically with the scheduling cycle and the real-time knowledge level. This means that the more knowledge-scarce a node is, the more it relies on the guidance of federated learning in the early stages, and the more it gradually weakens over time.

[0084] Decision variables Other variables in ( , , This is determined by solving the following subproblems:

[0085] The exchange power between the agent and its neighboring agents in the current cycle is sent to the neighboring agents so that the neighboring agents can determine the corresponding power allocation scheme based on the received exchange power in the current cycle using the iterative formula of ADMM. The power allocation scheme includes its own decision variables, auxiliary variables, and dual variables. Specifically, each agent i sends its exchange power with its neighboring agents. It broadcasts to neighboring agents j within its communication range. Each agent i receives the exchanged power from its neighboring agents j. Each agent i is determined based on local source, load, and storage constraints, as well as received neighbor information, i.e., the exchange power with neighboring agent j. The ADMM iterative formula updates its Lagrange multipliers and power allocation decision variables. This process, through repeated coordination of decisions among power sources, grid, load, and storage, aims to balance the total power supply and demand of the entire network while minimizing overall operating costs. In other words, the overall network optimization objective of the ADMM iterative formula is to minimize total operating costs while satisfying the power balance condition. ; The power allocation scheme includes its own decision variables, which can be specifically represented as: ;in, Let be the decision variable for agent i in the (k+1)th scheduling cycle. For the global consistency auxiliary variable of the k-th scheduling period, Let be the dual variable of the corresponding constraint in the k-th scheduling period. This is the penalty parameter.

[0086] Auxiliary variables can be specifically represented as: The dual variable can be specifically represented as: .

[0087] Repeat the steps from the beginning of the current scheduling cycle, whereby the agent uses the knowledge-enhanced model to calculate the initial exchange power with neighboring agents based on the agent's state vector, to the step of sending the exchange power with neighboring agents for the current cycle to the neighboring agents so that the neighboring agents can determine the corresponding power allocation scheme based on the received exchange power for the current cycle using the iterative formula of ADMM, until the power allocation schemes of all agents converge.

[0088] The power allocation scheme of each agent at the point where all power allocation schemes converge is the corresponding optimal power allocation scheme.

[0089] In this embodiment of the application, the convergence of the power allocation schemes of all intelligent agents and the satisfaction of the system power balance constraints indicate that the multi-party entities such as source, network, load, and storage have reached a collaborative optimal decision, i.e., the optimal power allocation scheme, within the current scheduling cycle.

[0090] To effectively integrate the federated learning framework with real-time distributed optimization and power system fault emergency response mechanisms, and to adapt to extreme scenarios such as communication outages, this application's embodiments construct effective emergency operational islands in the event of communication network interruptions due to faults (i.e., faulty networking). In particular, it maintains robustness even under non-ideal communication conditions, forming a three-layer architecture of 'offline knowledge pre-training + online collaborative decision-making + fault emergency autonomy'.

[0091] Specifically, in some embodiments, the method further includes: detecting whether it is necessary to switch to emergency operation mode; specifically, detecting whether the central server has failed or whether there is a large-scale interruption of the communication network; if it fails or is interrupted, then determining that it is necessary to switch to emergency operation mode.

[0092] Detecting whether the central server is malfunctioning can be done by determining whether the signal loss time exceeds a time threshold. If it exceeds the time threshold... If so, then the central server is confirmed to be faulty.

[0093] Detecting a large-scale communication network outage can be accomplished by determining whether the network connectivity between agents falls below a threshold. If it does, the network will be considered open. If more than a certain percentage of communication links fail, then a large-scale communication network outage is determined.

[0094] If a switch is required, an emergency operation island is constructed; wherein the emergency operation island includes neighboring agents that can communicate; in the embodiments of this application, each agent, based on the knowledge-enhanced model obtained from the last federated learning, uses the established distributed optimized links to form an emergency operation island with neighboring agents that still maintain communication.

[0095] The local optimization objective function is adjusted to the emergency operation optimization objective. Specifically, within the emergency operation island, all agents continue to execute the steps of determining the agent's optimal power allocation scheme using the local knowledge-enhanced model and distributed optimization algorithm. However, the local optimization objective function of the local knowledge-enhanced model is adjusted to prioritize ensuring the power supply of critical loads on the load side, and to use energy storage on the storage side and interruptible loads on the load side for internal power balancing, which is the emergency operation optimization objective.

[0096] The modified local optimization objective function adds a knowledge consistency penalty term, actively guiding the decisions of each agent to align with the overall operational behavior within the isolated system. Therefore, the modified local optimization objective function is as follows: ;in, - These are the weighting coefficients. For the critical load power of agent i, It is the rated voltage. This represents the actual power supplied. This represents the average value of the knowledge vectors within the isolated island. The fourth term in the modified local optimization objective function. As a knowledge consistency penalty, its main function is to actively guide the decisions of each agent to align with the overall operational posture within the isolated system, thereby enhancing the stability of knowledge collaboration.

[0097] Adjust the agent's state vector to a lightweight knowledge vector.

[0098] In this embodiment, within the isolated system, intelligent agents share minimal and necessary operational status information through point-to-point communication, achieving a "lightweight" knowledge interaction to collaboratively maintain the stable operation of the isolated system. The core purpose of this knowledge interaction is to directly coordinate the behavior of "storage" and "load." Necessary operational status information includes: remaining power. Load reduction power Available power generation margin Adjustable load margin To achieve a lightweight "knowledge interaction" mechanism to collaboratively maintain the stable operation of isolated systems.

[0099] Lightweight knowledge vectors can be represented as: .

[0100] Define the update rules for the lightweight knowledge vector.

[0101] Specifically, the update rule for lightweight knowledge vectors in isolated systems is as follows: ;in, The knowledge weight factor is a constant between 0 and 1, used to balance the weights of one's own knowledge and the knowledge of its neighbors. The set of neighboring agents of agent i refers to the set of all agents that can communicate directly with agent i in the current emergency island. The size of the set of neighboring agents is defined as the number of neighboring agents. The update rule is a distributed weighted average consensus algorithm that enables lightweight knowledge vectors for all agents within the isolated island. Ultimately, they converge to a consensus value, that is This rule is the core of collaborative perception in emergency mode, enabling all nodes within the island to reach a consensus on the operational status (such as the degree of power shortage and energy storage capacity), providing a unified decision-making basis for subsequent fair rotation and energy sharing, thereby achieving collaborative perception of operational status.

[0102] This information is used for coordination in distributed optimization, where each agent continues to use the ADMM algorithm for distributed optimization and variable updates. However, its optimization objective is an emergency operation optimization objective, and the communication range and knowledge interaction are limited to neighbors within the current isolated area, thus maintaining basic coordination and stable operation of the source, network, load, and storage even in the event of a failure.

[0103] The optimal power allocation scheme for the agents within the emergency operation island is determined by using the local knowledge-enhanced model and the iterative formula of ADMM.

[0104] The optimal power allocation scheme in this application embodiment also includes decision variables, auxiliary variables, and dual variables.

[0105] The decision variables are represented as follows: .

[0106] Auxiliary variables are represented as: ; A set of intelligent agents for maintaining communication.

[0107] The dual variable is represented as: .

[0108] In this embodiment, the lightweight knowledge interaction and coordination mechanism within the emergency operation island mainly includes: 1. Storage-to-storage coordination: Each intelligent agent periodically broadcasts its state of charge information. When a certain intelligent agent When the energy level is too low, other intelligent agents within the isolated island with surplus energy storage will automatically reduce their power consumption from the node through optimization algorithms, or even supply power to it, thereby preventing the node from running out of energy and achieving mutual sharing of energy storage resources. 2. Load-load coordination: Each intelligent agent periodically broadcasts its load reduction power. Available power generation margin The system uses a consensus algorithm to ensure that the load shedding rate within the emergency island is consistent. This avoids the unfair situation where some nodes have their load completely cut off while others remain unaffected, achieving fair rotational shedding of critical loads. 3. Source-Load-Storage Coordination: Combining and The system can dynamically decide whether to use the surplus power of distributed photovoltaic power generation to charge the energy storage or to directly use it to restore some interruptible loads, thus achieving optimal dynamic allocation of limited energy.

[0109] This application embodiment constructs a three-layer collaborative architecture of "federated learning knowledge alignment - distributed real-time optimization - lightweight emergency interaction". For the first time, federated learning is used to solve the problem of uneven knowledge distribution in multilateral decision-making in new power systems. Through knowledge-guided distributed optimization and lightweight knowledge interaction mechanism, efficient, fair and robust decision-making is achieved in all scenarios from normal to fault.

[0110] The method in this embodiment dynamically triggers the federated learning process by introducing a knowledge level assessment mechanism and knowledge distribution entropy; it employs a weighted knowledge aggregation strategy to make model updates in knowledge-rich regions more influential, achieving complementarity and alignment among multilateral agents in the case of uneven knowledge distribution; furthermore, a lightweight knowledge vector and its consistency update rules are designed in emergency mode to achieve rapid self-organizing collaborative operation in the event of a fault. Ultimately, this method can achieve efficient and fair power allocation while protecting data privacy, and maintain the system's emergency operation capability in fault scenarios such as communication interruptions, comprehensively improving the decision-making efficiency and operational resilience of the new power system.

[0111] In this embodiment, the experience of knowledge-rich regions is transferred to knowledge-poor regions in the form of model parameters through a federated learning pre-process, solving the problem of uneven knowledge distribution in multiple marginal regions, raising the decision-making starting point for all agents, and accelerating the convergence of subsequent distributed optimization; it also protects the data security of the multilateral agents. Throughout the process, each agent does not need to upload sensitive local raw operating data, but only exchanges model update quantities or power plans, effectively protecting the data privacy and trade secrets of the multilateral agents; it achieves a closed-loop fusion of offline knowledge and online decision-making. Combining the "offline knowledge pre-training" of federated learning and the "online real-time decision-making" of distributed optimization, the system can utilize global historical knowledge and adapt to real-time fluctuations, achieving efficient and fair power allocation; it also constructs a resilient fault emergency response system. The designed fault networking emergency mechanism allows the system to degrade into multiple self-organized distributed collaborative islands when the central node fails or communication is interrupted, continuing to maintain critical operations, greatly improving the reliability of the new power system.

[0112] In this embodiment, a global power allocation model based on multiple agents is first established. Cross-regional knowledge fusion is achieved through federated learning, generating a knowledge-enhanced model with global cognition. During real-time operation, the optimal power allocation scheme is dynamically solved by combining distributed optimization algorithms and the knowledge-enhanced model. When a communication failure occurs, the system automatically switches to emergency operation mode, constructing autonomous islands based on local communication networks. Lightweight knowledge interaction ensures continuous power supply to critical loads and stable system operation. This invention effectively overcomes the problems of uneven knowledge among multiple agents, difficulty in protecting data privacy, and insufficient collaborative decision-making capabilities under non-ideal communication conditions, significantly enhancing the system's adaptability and resilience in complex operating environments.

[0113] This application also provides a device for intelligent knowledge collaboration based on multilateral decision-making. Figure 2 exemplarily illustrates a structural schematic diagram of a device for intelligent knowledge collaboration based on multilateral decision-making according to some embodiments. The device includes: a definition unit 201, used to define a global power allocation model based on multiple agents, the global power allocation model including input quantities, output quantities, local optimization objective functions and local operating constraints, the agents being autonomous regions in a power distribution network, the input quantities including the state vectors of the agents, and the output quantities including the decision vectors of the agents; a first determination unit 202, used by a central server to determine the knowledge-enhanced model corresponding to each agent based on the global power allocation model using federated learning; and a second determination unit 203, used by the agents to determine the optimal power allocation scheme for the agents based on the state vectors using the local knowledge-enhanced model and a distributed optimization algorithm.

[0114] In some embodiments, the state vector includes a knowledge level index; the device further includes: the agent calculating the corresponding knowledge level index and uploading it to a central server; the central server determining the knowledge distribution entropy based on all the knowledge level indices; wherein the knowledge distribution entropy is the degree of uneven distribution of global knowledge; when the knowledge distribution entropy is greater than a preset knowledge distribution entropy threshold, the central server performs the step of determining the knowledge-enhanced model corresponding to each agent using federated learning according to the global power allocation model.

[0115] In some embodiments, the first determining unit includes: a distributing unit, configured to: distribute the global power allocation model to all agents via the central server, so that the agents can train the global power allocation model using local datasets to obtain local loss values ​​corresponding to local loss functions and model parameter update amounts of the global power allocation model; encrypt the model parameter update amounts; and upload the encrypted model parameter update amounts and the local loss values ​​to the central server; a calculation unit, configured to: calculate a target model parameter update amount using a weighted average method based on the encrypted model parameter update amounts uploaded by all agents via the central server; and update the model parameters of the global power allocation model stored in the central server using the target model parameter update amount; and a judging unit, configured to: judge the... The process includes: a central server unit to evaluate the knowledge alignment level of the global power allocation model stored in the central server if it converges and meets the federated learning objective; a third determination unit to determine that the agent's current global power allocation model is a knowledge-enhanced model if the evaluation passes; a re-execution unit to add a quadratic penalty term to the local loss function if the evaluation fails; and re-execution of the steps of using the central server to distribute the global power allocation model to all agents, and determining whether the global power allocation model stored in the central server converges based on the target model parameter update amount, and whether it meets the federated learning objective based on the received local loss value.

[0116] In some embodiments, the determination unit is specifically used to: process the local loss values ​​corresponding to all agents using the knowledge weights of the agents and the local dataset to obtain the target loss value; if the target loss value is less than a preset loss value, then it is determined that the target of federated learning is met.

[0117] In some embodiments, the distributed optimization algorithm includes the iterative formula of ADMM; the second determining unit specifically performs the following: at the beginning of the current scheduling cycle, the agent uses a knowledge-enhanced model to calculate the initial exchange power with neighboring agents based on the agent's state vector; based on the initial exchange power and the exchange power with neighboring agents in the previous scheduling cycle, the exchange power with neighboring agents in the current cycle is determined; the exchange power with neighboring agents in the current cycle is sent to the neighboring agents so that the neighboring agents, according to the received exchange power in the current cycle, use the iterative formula of ADMM to determine the corresponding power allocation scheme, wherein... The power allocation scheme includes its own decision variables, auxiliary variables, and dual variables; repeating the steps from the beginning of the current scheduling cycle, where the agent uses a knowledge-enhanced model to calculate the initial exchange power with neighboring agents based on the agent's state vector, to the step of sending the exchange power with neighboring agents for the current cycle to the neighboring agents so that the neighboring agents can determine the corresponding power allocation scheme based on the received exchange power for the current cycle using the iterative formula of ADMM, until the power allocation schemes of all agents converge; the power allocation scheme of each agent at the point of convergence of the power allocation schemes of all agents is determined as the corresponding optimal power allocation scheme.

[0118] In some embodiments, the apparatus further includes: a detection unit, configured to detect whether a switch to an emergency operation mode is required; if a switch is required, construct an emergency operation island; wherein the emergency operation island includes communicable neighboring agents; adjust the local optimization objective function to an emergency operation optimization objective; adjust the state vector of the agent to a lightweight knowledge vector; define the update rule of the lightweight knowledge vector; and determine the optimal power allocation scheme for the agents within the emergency operation island using a local knowledge-enhanced model and the iterative formula of ADMM.

[0119] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method based on intelligent knowledge collaboration for multilateral decision-making.

[0120] This application also provides a computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program implements the steps of the method for intelligent knowledge collaboration based on multilateral decision-making.

[0121] The above embodiments provide a method, apparatus, device, and medium for intelligent knowledge collaboration based on multilateral decision-making. The method first defines a global power allocation model based on multiple agents, then achieves cross-agent knowledge fusion through federated learning to generate a knowledge-enhanced model with global cognition. During real-time operation, the optimal power allocation scheme is dynamically solved by combining a distributed optimization algorithm with the knowledge-enhanced model. The method includes defining a global power allocation model based on multiple agents, where each agent is an autonomous region in a power distribution network; a central server determines the knowledge-enhanced model corresponding to each agent using federated learning based on the global power allocation model; and each agent determines its optimal power allocation scheme using its local knowledge-enhanced model and a distributed optimization algorithm.

[0122] It will be readily understood by those skilled in the art that the above-described advantageous methods can be freely combined and superimposed without conflict. The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application. The above are merely preferred embodiments of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of this application, and these improvements and modifications should also be considered within the protection scope of this application.

Claims

1. A method for intelligent knowledge collaboration in multilateral decision-making, characterized in that, include: A global power allocation model based on multiple agents is defined. This model includes inputs, outputs, a local optimization objective function, and local operational constraints. Each agent is an autonomous region within the power distribution network. The inputs include the agent's state vector, and the outputs include the agent's decision vector. A central server, based on the global power allocation model, uses federated learning to determine a knowledge-enhanced model for each agent. Each agent, based on its state vector, uses its local knowledge-enhanced model and a distributed optimization algorithm to determine its optimal power allocation scheme.

2. The method according to claim 1, characterized in that, The state vector includes a knowledge level index. Before the step of the central server determining the knowledge-enhanced model corresponding to each agent using federated learning based on the global power allocation model, the method further includes: the agent calculating the corresponding knowledge level index and uploading it to the central server; the central server determining the knowledge distribution entropy based on all the knowledge level indices; wherein, the knowledge distribution entropy is the degree of uneven distribution of global knowledge; when the knowledge distribution entropy is greater than a preset knowledge distribution entropy threshold, the step of the central server determining the knowledge-enhanced model corresponding to each agent using federated learning based on the global power allocation model is executed.

3. The method according to claim 1, characterized in that, The steps by which the central server determines the knowledge-enhanced model corresponding to each agent using federated learning based on the global power allocation model include: the central server distributing the global power allocation model to all agents, enabling each agent to train the global power allocation model using a local dataset to obtain the local loss value corresponding to the local loss function and the model parameter update amount of the global power allocation model; encrypting the model parameter update amount; uploading the encrypted model parameter update amount and the local loss value to the central server; the central server calculating the target model parameter update amount using a weighted average method on the encrypted model parameter update amounts uploaded by all agents, and updating the model parameters of the global power allocation model stored in the central server using the target model parameter update amount. The process involves: determining whether the global power allocation model stored in the central server has converged based on the target model parameter update amount, and determining whether it conforms to the goal of federated learning based on the received local loss value; if converged and conforming to the goal of federated learning, evaluating the knowledge alignment level of the global power allocation model stored in the central server; if the evaluation passes, determining that the agent's current global power allocation model is a knowledge-enhanced model; if the evaluation fails, adding a secondary penalty term to the local loss function; and re-executing the steps of using the central server to distribute the global power allocation model to all agents, and determining whether the global power allocation model stored in the central server has converged based on the target model parameter update amount, and determining whether it conforms to the goal of federated learning based on the received local loss value.

4. The method according to claim 3, characterized in that, The step of determining whether the received local loss value meets the goal of federated learning includes: processing the local loss values ​​corresponding to all agents using the agent's knowledge weights and local dataset to obtain the target loss value; if the target loss value is less than a preset loss value, then it is determined that the goal of federated learning is met.

5. The method according to claim 1, characterized in that, The distributed optimization algorithm includes the iterative formula of ADMM; the steps for the agent to determine the optimal power allocation scheme based on the state vector, using a local knowledge-enhanced model and the distributed optimization algorithm, include: at the beginning of the current scheduling cycle, the agent uses the knowledge-enhanced model to calculate the initial exchange power with neighboring agents based on the agent's state vector; based on the initial exchange power and the exchange power with neighboring agents in the previous scheduling cycle, the agent determines the exchange power with neighboring agents in the current cycle; the agent sends the exchange power with neighboring agents in the current cycle to the neighboring agents, so that the neighboring agents can use the iterative formula of ADMM to determine the optimal power allocation scheme based on the received exchange power in the current cycle. The process involves determining a corresponding power allocation scheme, where each power allocation scheme includes its own decision variables, auxiliary variables, and dual variables; repeating the steps from the beginning of the current scheduling cycle, where the agent uses a knowledge-enhanced model to calculate the initial exchange power with neighboring agents based on the agent's state vector, to the step of sending the exchange power with neighboring agents for the current cycle to the neighboring agents so that the neighboring agents can determine the corresponding power allocation scheme based on the received exchange power for the current cycle using the iterative formula of ADMM, until the power allocation schemes of all agents converge; and determining the power allocation scheme of each agent at the point of convergence of the power allocation schemes of all agents as the corresponding optimal power allocation scheme.

6. The method according to claim 1, characterized in that, Also includes: The system detects whether a switch to emergency operation mode is needed; if a switch is needed, an emergency operation island is constructed; wherein the emergency operation island includes communicable neighboring agents; the local optimization objective function is adjusted to the emergency operation optimization objective; the state vector of the agent is adjusted to a lightweight knowledge vector; the update rule of the lightweight knowledge vector is defined; and the optimal power allocation scheme of the agents in the emergency operation island is determined by using the local knowledge-enhanced model and the iterative formula of ADMM.

7. A device for intelligent knowledge collaboration in multilateral decision-making, characterized in that, include: A definition unit is used to define a global power allocation model based on multiple agents. The global power allocation model includes inputs, outputs, local optimization objective functions, and local operating constraints. Each agent is an autonomous region in the power distribution network. The inputs include the state vectors of the agents, and the outputs include the decision vectors of the agents. A first determination unit is used by the central server to determine the knowledge-enhanced model corresponding to each agent based on the global power allocation model using federated learning. The second determining unit is used by the agent to determine the optimal power allocation scheme of the agent based on the state vector, using a local knowledge-enhanced model and a distributed optimization algorithm.

8. The apparatus according to claim 7, characterized in that, The state vector includes a knowledge level index; the device further includes: the agent calculating the corresponding knowledge level index and uploading it to a central server; the central server determining the knowledge distribution entropy based on all the knowledge level indices; wherein the knowledge distribution entropy is the degree of uneven distribution of global knowledge; when the knowledge distribution entropy is greater than a preset knowledge distribution entropy threshold, the central server executes the step of determining the knowledge-enhanced model corresponding to each agent using federated learning based on the global power allocation model.

9. The apparatus according to claim 7, characterized in that, The first determining unit includes: a distribution unit, configured to distribute the global power allocation model to all agents via the central server, so that the agents can train the global power allocation model using local datasets to obtain local loss values ​​corresponding to local loss functions and model parameter update amounts of the global power allocation model; encrypt the model parameter update amounts; and upload the encrypted model parameter update amounts and the local loss values ​​to the central server; a calculation unit, configured to calculate a target model parameter update amount using a weighted average method based on the encrypted model parameter update amounts uploaded by all agents, and update the model parameters of the global power allocation model stored in the central server using the target model parameter update amount; and a judgment unit, configured to judge the central server based on the target model parameter update amount. The process includes: a central server unit for evaluating the knowledge alignment level of the global power allocation model stored in the central server if it converges and meets the federated learning objective; a third determination unit for determining that the agent's current global power allocation model is a knowledge-enhanced model if the evaluation passes; a re-execution unit for adding a quadratic penalty term to the local loss function if the evaluation fails; and re-execution of the steps of using the central server to distribute the global power allocation model to all agents, determining whether the global power allocation model stored in the central server converges based on the target model parameter update amount, and determining whether it meets the federated learning objective based on the received local loss value.

10. The apparatus according to claim 9, characterized in that, The judgment unit is specifically used to: process the local loss values ​​corresponding to all agents using the knowledge weights of the agents and the local dataset to obtain the target loss value; if the target loss value is less than the preset loss value, then it is determined that it meets the target of federated learning.

11. The apparatus according to claim 7, characterized in that, The distributed optimization algorithm includes the iterative formula of ADMM; the second determining unit specifically executes: at the beginning of the current scheduling cycle, the agent uses the knowledge-enhanced model to calculate the initial exchange power with neighboring agents based on the agent's state vector; based on the initial exchange power and the exchange power with neighboring agents in the previous scheduling cycle, the exchange power with neighboring agents in the current cycle is determined; the exchange power with neighboring agents in the current cycle is sent to neighboring agents so that the neighboring agents can determine the corresponding power allocation scheme based on the received exchange power in the current cycle using the iterative formula of ADMM, wherein the power allocation scheme includes its own decision variables, auxiliary variables, and dual variables; repeating the step from the beginning of the current scheduling cycle where the agent uses the knowledge-enhanced model to calculate the initial exchange power with neighboring agents based on the agent's state vector to the step of sending the exchange power with neighboring agents in the current cycle to neighboring agents so that the neighboring agents can determine the corresponding power allocation scheme based on the received exchange power in the current cycle using the iterative formula of ADMM, until the power allocation schemes of all agents converge; The power allocation scheme of each agent at the point where all power allocation schemes converge is the corresponding optimal power allocation scheme.

12. The apparatus according to claim 7, characterized in that, Also includes: The detection unit detects whether a switch to emergency operation mode is needed; if a switch is needed, an emergency operation island is constructed; wherein the emergency operation island includes communicable neighboring agents; the local optimization objective function is adjusted to the emergency operation optimization objective; the state vector of the agent is adjusted to a lightweight knowledge vector; the update rule of the lightweight knowledge vector is defined; and the optimal power allocation scheme of the agents in the emergency operation island is determined by using the local knowledge-enhanced model and the iterative formula of ADMM.

13. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent knowledge collaboration method based on multilateral decision-making as described in any one of claims 1 to 6.

14. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent knowledge collaboration method based on multilateral decision-making as described in any one of claims 1 to 6.

Images