Edge calculation federated aggregation optimization method for multi-region power distribution network

Abstract

The invention discloses an edge computing federated aggregation optimization method for a multi-region power distribution network, and belongs to the technical field of power distribution network collaborative optimization and federated learning. The method is based on a'cloud-edge-end 'three-level architecture, an improved artificial bee colony (ABC) algorithm and an optimal transmission (OT) mechanism are fused, and an AOT aggregation optimization scheme is constructed: an end side collects and preprocesses operation time sequence data, and an edge node completes local model training and uploads parameters; the cloud searches the optimal aggregation weight through an improved ABC algorithm, adapts the difference between node data and the model, and realizes parameter distribution alignment and weighted fusion by means of an OT mechanism; and efficient aggregation is guaranteed through double-trigger cooperative communication scheduling. According to the method, the problems of fixed weight and insufficient parameter alignment of traditional federal aggregation are solved, the global model precision, stability and convergence speed are remarkably improved in a data non-independent identically distributed scene, the communication overhead is reduced, and reliable support is provided for intelligent collaborative scheduling of the multi-region power distribution network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution network collaborative optimization and federated learning, in particular to an edge computing federated aggregation optimization method for multi-region power distribution networks. BACKGROUND

[0002] In the development process of distributed energy systems and smart parks, more and more independent subjects (such as zero-carbon factories, green parks, and industrial clusters) form interconnected regional power grids. These subjects have obvious differences in energy production, consumption, and dispatching modes, and their load characteristics, operation strategies, and equipment types are different, but they need to carry out collaborative management and optimal dispatching under the premise of safe and stable power grid.

[0003] The traditional centralized energy management mode relies on collecting the operation data (such as real-time power load, equipment operating status, energy use plan, etc.) of each subject to a unified central platform for modeling and decision-making. However, in the multi-subject scenario, this mode has outstanding problems. First, the operation data involves sensitive content such as production processes, energy utilization efficiency, and market transaction information. Directly sharing raw data can easily cause commercial secrets to be leaked and privacy compliance risks. Second, when the number of subjects is large and the data scale is large, centralized computing and storage can easily form a performance bottleneck, requiring high communication bandwidth and being prone to data transmission delays and network congestion. Third, the energy use modes of different subjects differ significantly, and the collected data exhibits Non-IID characteristics. Directly using a fixed weighted global modeling method can easily lead to slow model convergence and accuracy decline.

[0004] To solve these problems, federated learning (FL) provides a new technical path. This method allows each subject to complete model training locally and only uploads model parameters or gradients to the upper layer for aggregation, thereby realizing joint modeling across subjects while ensuring that data does not leave the local area. However, the traditional federated learning method still has two major limitations in the multi-subject power distribution network scenario. First, it cannot dynamically adjust the aggregation weights according to the data quality, quantity, and model performance of each subject, leading to insufficient or excessive magnification of the contribution of some subjects to the global model. Second, it lacks an effective alignment mechanism for the distribution differences of model parameters from different sources in the model fusion stage, making it difficult to maintain stable global performance in a highly heterogeneous environment.

[0005] In the collaborative operation of multi-regional power distribution networks, multiple independent entities (such as zero-carbon factories, green parks, industrial parks, etc.) need to share operation information and make joint decisions under the premise of ensuring data privacy to achieve the coordination and optimization of source-grid-load-storage. Traditional centralized computing mode usually uploads raw data to the cloud for unified processing, which not only causes cross-regional communication traffic and bandwidth pressure, but also increases the risk of data leakage and network attacks, making it difficult to meet the high security requirements of power grid dispatching scenarios. The introduction of edge computing provides the ability to complete data processing and model training locally in multi-regional power distribution networks, reducing the need for cross-regional transmission of raw data and improving real-time performance and response speed. However, in the cloud-edge-end collaborative architecture, how to efficiently aggregate models distributed on various edge nodes while ensuring privacy still faces many technical bottlenecks.

[0006] Existing federated learning aggregation methods (such as FedAvg) usually perform unified fixed-weight aggregation in the cloud, lacking dynamic awareness of the quality differences of edge node models and data distribution heterogeneity. When the data between edge nodes exhibits strong non-independent and identical distribution (Non-IID) characteristics, fixed weights can cause global model parameter update bias, resulting in slow convergence, accuracy decline, and even unstable training. At the same time, due to the significant differences in edge node model parameter distribution, the lack of effective parameter alignment mechanism can lead to a decline in model generalization performance during cross-regional collaborative aggregation.

[0007] In addition, in the hierarchical cloud-edge-end system, there are often difficulties in time synchronization, communication delay, and uneven computing resources between edge layer aggregation and cloud global aggregation, especially in the complex operating environment of multi-regional power grids, these problems will further amplify, limiting the overall dispatching efficiency and intelligent level of the system. Therefore, there is an urgent need for a federated aggregation method that can balance dynamic weight optimization and parameter distribution alignment in the edge computing scenario, to fully release the performance potential of the cloud-edge-end collaborative architecture and improve the intelligent dispatching and operation stability of multi-regional power distribution networks. SUMMARY

[0008] In view of the current multi-regional power distribution network in the cloud-edge-end collaborative operation, there are systematic defects such as insufficient model aggregation adaptability, low parameter alignment accuracy, and limited convergence performance under data heterogeneity conditions.

[0009] To achieve the above object, the application provides a multi-region power distribution network-oriented edge computing federated aggregation optimization method AOT (Aggregated Optimal Transport), which combines an improved artificial bee colony ABC (Artificial Bee Colony) algorithm and an optimal transport (OT) mechanism to realize dynamic weight search and model parameter alignment in the model aggregation process of federated learning in a three-level architecture scenario of "cloud-edge-end". The improved AOT adaptively adjusts the aggregation weight according to the data distribution and model performance of each node, and the OT optimally maps the parameter distribution in the aggregation process to reduce the difference, thereby improving the accuracy, stability and convergence speed of the global model under the condition of data heterogeneity, and breaking through the performance bottleneck of traditional fixed weighted aggregation in cross-region cooperation.

[0010] More importantly, the application not only focuses on the improvement of the global model accuracy, but also considers multi-dimensional optimization objectives such as convergence speed, communication cost and system robustness, forming a closed-loop aggregation system of "dynamic weight optimization-parameter distribution alignment-multi-index collaborative improvement". In the complex power distribution network operation scenario, the method significantly improves the performance of the global model under the condition of high heterogeneity and multi-task, reduces the communication cost while maintaining the advantage of local data processing, and provides a new technical path for efficient cooperative scheduling and intelligent decision-making of multi-region power distribution networks.

[0011] In one aspect of the application, a multi-region power distribution network-oriented edge computing federated aggregation optimization method is provided, comprising the following steps: S1 Construct a three-layer architecture federated modeling system of cloud-edge-end: the end-side device collects time series data and pre-processes to form a sample set, the edge node trains a local model based on the sample set, and iteratively updates the model parameters to the cloud; S2 Aggregation weight search space modeling: define a weight vector satisfying the non-negativity and normalization constraints, construct a K-dimensional simplex search space, and calculate an adaptability function representing the aggregation effect based on a shared validation set; S3 Search for optimal aggregation weight by improved artificial bee colony algorithm: search for the optimal weight vector by population initialization, employed bee local search, observer bee chaotic disturbance, scout bee re-initialization and elite reservation strategy; S4 Realize parameter aggregation based on optimal transport mechanism: construct a cost matrix of model parameters, set edge distribution, solve the entropy regularization optimal transport mechanism, and generate a global model by weighted fusion; S5 Cloud-edge-end communication scheduling: adopt a periodic and event-driven dual scheduling mechanism, combine priority strategy and closed-loop control, and realize efficient cooperation of data transmission and parameter uploading.

[0012] Furthermore, in S1, the time-series data collected by the end-side device includes voltage, current, power, equipment operating status, and energy consumption plan data. Preprocessing includes data compression, anonymization, and outlier removal. The local model training of the edge node adopts an iterative update mechanism based on gradient descent.

[0013] Furthermore, in S2, the aggregate weight vector Among them, satisfying , K The number of edge nodes participating in the aggregation; the fitness function is constructed based on the prediction accuracy of the shared validation set, and the formula is:

[0014] in, To share the verification set, For the aggregated global model, This is an indicator function.

[0015] Furthermore, S3, based on the improved artificial bee colony algorithm, searches for the optimal aggregation weight, specifically as follows: Dynamic weight search employs an improved artificial bee colony algorithm in the cloud to search for the optimal aggregated weight vector in simplex space. ,include: Hired bee stage: Generate candidate solutions based on power-law decay adaptive perturbation factor; Observation phase: Select solutions based on fitness probability and introduce logistic chaotic perturbation; Reconnaissance Bee Phase: Reinitialize the unimproved solution.

[0016] Furthermore, the hired bee phase is based on an adaptive perturbation factor with power-law decay, as follows: , During the observation phase, a logistic chaotic perturbation is introduced, and the chaotic sequence formula is: , in, t This represents the current iteration number. z t For the first t The initial value of the logistic chaotic variable z0 ∈ (0,1) is randomly given. Iteration using the above formula yields a one-dimensional chaotic sequence located in the interval (0,1). μ is the control parameter of the logistic mapping, determining the dynamic behavior of the sequence. When μ = 4, the system is in a fully chaotic state, exhibiting the best ergodicity and randomness of the sequence. T max Where γ is the maximum number of iterations, and γ is the decay coefficient. In dimension Random perturbation coefficients on the upper side, used to control the perturbation amplitude and direction.

[0017] Further, the S4, based on the optimal transmission mechanism parameter aggregation includes: The difference between different client models is calculated using the Euclidean distance, and the transmission cost matrix of the parameters is constructed; According to the optimal aggregation weight vector obtained by the improved artificial bee colony algorithm, the source distribution is defined as the optimal weight vector obtained by the ABC algorithm, and the target distribution is defined as the uniform distribution; An entropy regularization term is introduced to enhance numerical stability, and the optimal transport plan is solved by Sinkhorn-Knopp iteration; Using the optimal transport plan for weighted fusion, a new round of global model parameters is constructed.

[0018] Further, the S5, cloud-edge-end communication scheduling specifically includes: Edge-cloud asynchronous aggregation and scheduling trigger: the edge node adopts a periodic and event-driven double scheduling mechanism to upload model parameters to the cloud, and the cloud automatically starts the aggregation process after receiving model parameters uploaded by more than a preset number of edge nodes; Heterogeneous bandwidth and scheduling priority strategy: according to the node bandwidth, response delay and model importance, the scheduling priority is calculated, and the cloud waits for the high-priority node to upload the results before each round of aggregation; Communication scheduling system closed-loop control: the end dynamically adjusts the scheduling parameters according to the convergence speed of each round of aggregation, the global loss curve fluctuation, and the edge-end communication failure rate, and the edge node autonomously selects to participate or skip a certain round of upload according to the local communication load and training progress.

[0019] In another aspect of the present application, a multi-region power distribution network-oriented edge computing federal aggregation optimization system is protected, which includes: The end-side acquisition module is used for collecting time series data and performing preprocessing operations. The edge training module is deployed on each regional edge node and is used for training local models based on local sample sets and outputting model parameters after iterative updates; The cloud aggregation module includes a weight searching unit and a parameter alignment unit, the weight searching unit is used for searching the optimal aggregation weight through the improved artificial bee colony algorithm, and the parameter alignment unit is used for realizing model parameter distribution alignment and weighted fusion through the optimal transport mechanism to generate a global model; The communication scheduling module is used for executing cloud-edge-end communication scheduling, adopts a periodic and event-driven double scheduling mechanism, and combines priority strategy and closed-loop control of dynamic adjustment of scheduling parameters.

[0020] Further, the ABC weight search unit of the cloud aggregation module supports parameter configuration of population size and iteration round number, and the OT parameter alignment unit supports entropy regularization parameter adjustment.

[0021] In another aspect of the present application, a computer-readable storage medium is provided, which stores a computer program, and the computer program, when executed by a processor, implements the edge computing federated aggregation optimization method for a multi-region power distribution network.

[0022] Compared with the prior art, the present application has the following remarkable beneficial effects: Compared with the traditional fixed weighting federated aggregation method, the adaptive aggregation mechanism established by the method can adjust the weight distribution in real time according to the node data distribution and model contribution, significantly improving the precision and convergence speed of the global model in a multi-region complex scene, and effectively suppressing the negative impact of low-quality updates on the global performance. The parameter mapping technology based on optimal transmission breaks through the precision bottleneck of existing methods in the parameter alignment of heterogeneous nodes, significantly reduces the difference between the model parameter distributions across regions, and thus improves the generalization ability and stability of the global model. By sinking part of the computing and aggregation tasks to the edge layer and introducing dynamic weights and chaotic disturbance mechanisms, the system realizes synchronous optimization in terms of communication delay, aggregation response speed and robustness. Even under adverse conditions such as node failure and communication fluctuations, the global performance remains stable. This technical system provides key technical support for building an intelligent cloud-edge-end collaborative learning platform for multi-region power distribution networks with autonomous aggregation decision-making, dynamic optimization and cross-region collaboration capabilities.

[0023] Other features and advantages of the present application will be described in the following description, and become apparent from the description, or be learned by practice of the present application.

[0024] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0025] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings: Figure 1 A flowchart of a federated aggregation optimization method for edge computing of a multi-region power distribution network in an embodiment of the present application is shown in the figure. Figure 2 A block diagram of a federated aggregation optimization method for edge computing of a multi-region power distribution network in an embodiment of the present application is shown in the figure. Figure 3This is a flowchart of a simplified AOT aggregation optimization algorithm in an edge computing federated aggregation optimization method for multi-regional distribution networks according to an embodiment of the present invention; Figure 4 This is a diagram of an edge computing federated aggregation optimization system for multi-regional power distribution networks in an embodiment of the present invention. Detailed Implementation

[0026] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0027] The core idea of ​​this invention is to address the challenges of heterogeneous data and collaborative modeling in multi-regional power distribution networks by constructing a federated learning system oriented towards a "cloud-edge-device" architecture. It proposes an AOT federated aggregation optimization method that combines an improved artificial bee colony (ABC) algorithm with an optimal transmission (OT) mechanism for weight optimization and distribution alignment of heterogeneous edge node models, significantly improving global model performance and aggregation robustness.

[0028] Figure 1 As shown, an edge computing federated aggregation optimization method for multi-regional distribution networks includes the following steps: Step 1: Construct a federated modeling system under a three-tier architecture of "cloud-edge-device". Figure 2 The diagram shows an architecture of an edge computing federated aggregation optimization method for multi-regional power distribution networks, including the cloud, edge side, and end side.

[0029] Step 1.1: End-side data acquisition and preprocessing mechanism In distributed scenarios of multi-regional power distribution networks, end-side devices, as the lowest-level nerve endings of the system, undertake the fundamental task of operational status perception. Deployed in locations such as smart meters, feeder terminal units (FTUs), switch controllers, and distributed energy interface devices, end-side devices possess high-frequency sampling and anomaly detection capabilities. Their primary responsibility is to collect local time-series data such as voltage, current, phase angle, and frequency in real time, and to mark disturbance signals through event-triggered mechanisms.

[0030] The edge device performs edge compression, encoding, and anonymization on the raw signal before storing it in a structured format in its local cache. This mechanism ensures that the data achieves a closed-loop perception without leaving the local storage and provides a compliant, secure, and complete sample set for training upper-layer models. D i .

[0031] Step 1.2: Edge-side local model training mechanism Edge computing nodes are deployed at regional master stations, station control layers, or high-computing-power feeder terminals, and possess independent model training and iteration capabilities. Each edge node is treated as a federation participant, maintaining a local model. And based on data within its jurisdiction Perform several rounds of local training and updates. Training employs a gradient descent mechanism, with the following local iterative form: in, For the first The local loss function for each edge node. For learning rate, Indicates the first In the first round of Federal Communications The model parameters are updated locally. Edge nodes execute at most [number] times per round of communication. Second local iteration; when the decrease in local loss between adjacent iterations is less than a threshold. Local training can be stopped early. At the global level, federated training can execute a maximum of [number] cycles. In the round, when global verification performance is at its best in recent times In-wheel improvement less than the threshold The training is terminated when convergence is determined.

[0032] In addition, edge nodes can adaptively select training frequency and upload cycle according to communication conditions and scheduling strategies, supporting asynchronous or quasi-synchronous update modes, enhancing the system's adaptability to unstable links and heterogeneous resources.

[0033] Step 1.3: Cloud-based model aggregation and global knowledge fusion mechanism Cloud-side nodes, acting as the coordination center for the entire system, are typically deployed at power distribution stations, operation and maintenance centers, or cloud platform servers, and are responsible for receiving the model parameter sets uploaded by edge nodes. To improve the convergence performance and generalization ability of the model in heterogeneous data environments, the cloud adopts the AOT (Aggregated Optimal Transport) aggregation mechanism proposed in this invention to achieve dynamic alignment and weighted fusion of model parameters.

[0034] Figure 3 The diagram shows a simplified AOT aggregation optimization algorithm in an edge computing federated aggregation optimization method for multi-regional distribution networks. This provides an introduction to... Figure 2 The framework diagram illustrates the combined working mechanism of the artificial bee colony algorithm and the optimal transmission mechanism.

[0035] Based on the differences in parameter distribution among edge models, construct the optimal transmission loss:

[0036] in, denotes the transmission plan from parameter a to b, For its cost function, OT is used to alleviate the aggregation bias problem caused by model misalignment.

[0037] Improved artificial bee colony optimization module (ABC) is used to search for optimal aggregation weights to maximize the performance indicator of the global model under the objective function. The final aggregation model expression is:

[0038] where is the parameter representation after optimal transmission, is the optimal weight obtained by artificial bee colony search. This approach is significantly better than traditional fixed weighting strategies such as FedAvg, and can dynamically adapt to the data heterogeneity and model training ability differences of each edge node.

[0039] Step 2: Constructing the aggregation weight search space modeling mechanism Step 2.1: Definition of weight vector In the federated learning system of the "cloud-edge-end" architecture, it is assumed that the current is the t+1 communication round, and there are edge nodes participating in the aggregation process, and each node After completing local training, the model parameters are generated. In order to achieve personalized and collaborative global aggregation, a non-negative weight needs to be assigned to each node's model update to measure its contribution to the aggregation process. Let:

[0040] is the weight vector corresponding to this round of aggregation, which needs to satisfy the following constraint conditions: (1) Each node weight is non-negative, i.e. .

[0041] (2) The sum of weights is 1, i.e. .

[0042] This vector characterizes the parameter update proportion of each edge node in the global model and is the core variable of the optimization objective.

[0043] Step 2.2: Search space construction The above constraints constitute a standard K-dimensional simplex space, i.e.:

[0044] From a geometric perspective, each point in this space represents a legitimate aggregation scheme. By searching for the optimal weight combination within this space, adaptive integration of client data quality, model performance, and distribution characteristics can be achieved, which is the key foundation for breaking traditional static weighting (such as FedAvg).

[0045] Step 2.3: Evaluation mechanism of aggregation effect In order to evaluate each candidate weight vector , the invention introduces an accuracy rate based on a shared small-scale validation set as the fitness function:

[0046] where, represents the prediction result of the global model aggregated using the current weight combination on the input sample , is an indicator function (takes 1 if true, otherwise takes 0).

[0047] This function provides evaluation basis for subsequent optimization search, and at the same time, due to the small size of the validation set, it can realize fast feedback and dynamic adjustment every round.

[0048] Step 3: Aggregation weight search mechanism based on improved artificial bee colony algorithm Step 3.1: Population initialization In the process of federated aggregation optimization, in order to obtain the optimal aggregation weight vector , the invention first constructs an initial search population in the defined -dimensional simplex search space . Let the search population size be , then aggregation weight vectors that meet the constraint conditions need to be initialized and used as candidate solutions.

[0049] For each individual , a set of uniform random numbers is generated from the interval , and it is mapped to the simplex space through normalization operation:

[0050] where a i,k represents the aggregation weight allocated to the i th client in the federated system in the k th candidate solution, the value range of k =1, …, K , i.e. KThe first simplex in the 3D k Each dimension. All initial weight vectors. All satisfy the nonnegativity and normalization constraints, that is This ensures that each candidate solution is a valid aggregation scheme.

[0051] For each initialized individual, calculate its value in the shared validation set. The classification accuracy of the aggregated global model is used as the fitness. The specific definition is as follows:

[0052] in, It is an indicator function. Indicates weight The global model after aggregating the edge node models.

[0053] Step 3.2: Candidate solution generation and local search update in the hired bee phase In each search iteration, the hired bee individuals orbit around the current solution. Perform local perturbation to generate candidate solution vectors We are exploring better aggregation solutions. Its update mechanism is as follows:

[0054] in, In dimension The random perturbation coefficient is used to control the amplitude and direction of the perturbation. To improve local search stability and search accuracy, this invention designs an adaptive perturbation factor based on power-law decay:

[0055] in This represents the current iteration number. The maximum number of iteration rounds, This is the attenuation coefficient. This mechanism allows for a wide-ranging search in the early stages and fine-tuning in the later stages.

[0056] Newly generated candidate vectors The projection operation will renormalize the model to simplex space. Internally, ensure legality:

[0057] Next, calculate its fitness and execute the following update strategy:

[0058] Step 3.3: Selection and Optimization Update in the Observation Bee Phase. In the observation bee phase, a probabilistic selection strategy is adopted based on the fitness distribution of candidate solutions:

[0059] To further prevent falling into local optimum, a chaotic disturbance mechanism is introduced to the optimal solution. The logistic map is used to generate chaotic sequences:

[0060] where, z t is the logistic chaotic variable at the t th iteration, the initial value z 0 is randomly given, and a one-dimensional chaotic sequence located in the interval (0, 1) is obtained by iteration. μ is the control parameter of the logistic map, which determines the dynamic behavior of the sequence. When μ= 4, the system is in a fully chaotic state, and the sequence has the best ergodicity and randomness.

[0061] The disturbance form is:

[0062] where ε ∈ (0, 1) is the chaotic disturbance strength parameter, which is used to scale the disturbance amplitude of the logistic chaotic sequence to the current solution. The larger the ε, the larger the search step, and the stronger the global exploration ability. The smaller the ε, the more conservative the update, which is conducive to fine local search.

[0063] This mechanism expands the search space while maintaining the ability of local search.

[0064] Step 3.4: Scout Bee Mechanism and Elite Preservation Strategy For continuous unimproved candidate solutions, the scout bee mechanism is used to reinitialize in the search space to avoid early stagnation of the search. At the same time, the elite preservation strategy is adopted to save the current optimal solution in each iteration to ensure that the global optimum is not lost. The weight update rule is:

[0065] Through the above multi-mechanism collaborative optimization, the aggregated weight takes into account data heterogeneity, model performance, and convergence speed, laying the foundation for subsequent optimal transport aggregation.

[0066] Step 4: Model Aggregation Strategy Based on Optimal Transport Mechanism Step 4.1: Construct the Parameter Distribution Distance Matrix between Client Models Let the local model parameters obtained by the th client at the end of the th communication be

[0067] The Euclidean distance is used to depict the difference between different client models, and a parameter transmission cost matrix is constructed :

[0068] wherein, represents the unit transmission cost of moving the "quality" from the model parameter distribution of the client to the model parameter distribution of the client , which provides a geometric measurement basis for subsequent OT optimization.

[0069] Step 4.2: Set the edge distribution of transmission Let the aggregation weight vector optimized by improving the artificial bee colony algorithm be

[0070] The application takes it as the source distribution in the optimal transmission problem:

[0071] Target distribution Can be flexibly set according to application requirements. By default, q can be taken as a uniform distribution, that is,

[0072] To achieve balanced aggregation of each parameter component; when it is necessary to highlight some parameter subspaces or specific regional nodes, q can also be designed according to the global model structure characteristics and scheduling strategy, as long as the non-negativity and normalization constraints

[0073] Therefore, p and q jointly constitute the edge constraints of the optimal transmission problem, respectively reflecting the aggregation contribution of each edge node and the global model target structure.

[0074] Step 4.3: Solve the entropy-regularized optimal transport plan Solving the transmission plan means that under the conditions of given source distribution p , target distribution q and cost matrix C, the transmission plan matrix γ that satisfies the edge constraints and minimizes the entropy-regularized transmission cost is obtained by Sinkhorn-Knopp iteration.

[0075] In order to enhance numerical stability and improve solving efficiency, entropy regularization is introduced into the classical optimal transmission model, and the transmission problem is modeled as:

[0076] wherein: is the transport plan matrix, denotes the mass assigned from the k-th component of the source distribution to the j-th component of the target distribution in the transport scheme:

[0077] is the set of feasible transport plans satisfying the marginal constraints, is the total transport cost; is the entropy of the transport plan, is the regularization parameter used to adjust the weight of the entropy term, balancing the minimum cost and the smoothness of the distribution.

[0078] To efficiently solve the above optimization problem, Sinkhorn-Knopp iterative algorithm is adopted. Denote the kernel matrix

[0079] then the transport plan matrix can be expressed as

[0080] wherein is the scaling vector, wherein, denotes the operator that constructs a diagonal matrix from a vector. The satisfaction of the marginal constraints is achieved by alternating normalization, whose iterative form is:

[0081] wherein " / " denotes element-wise division, is the number of iterations. As the iteration proceeds,

[0082] the row and column sums of converge to p and q, respectively. When the marginal constraint error is below a pre-set threshold, the iteration is considered to be converged, and the transport plan matrix obtained at this time is denoted as Γ, which is the entropy-regularized optimal transport plan.

[0083] Step 4.4: Perform parameter alignment and generate global model After obtaining the optimal transport plan matrix Γ, the parameters of the client models are weighted and fused to achieve parameter distribution alignment and generate the global model parameters of the next round. Define the effective aggregation weight of the client:

[0084] then the global model parameter is updated as:

[0085] wherein, For the elements in the optimal transmission plan matrix, it represents the quality proportion of the kth client model parameter distribution to the jth component of the global target model under the optimal transmission scheme. It can be understood as a modified aggregation weight after considering the OT geometry alignment. Through this mechanism, in the high-heterogeneous scenario of multi-area power distribution networks, the cross-node model parameter distribution deviation can be significantly reduced, and the precision, stability and convergence speed of the global model can be improved on the premise that the data does not leave the local.

[0086] Step 5: Cloud-edge-end communication scheduling mechanism design Step 5.1: Edge-end data uploading and model feedback mechanism The terminal device mainly undertakes the task of real-time acquisition of operating state data such as load, voltage and power, and the sampling period is short, but the computing power is limited. The application adopts an end-side lightweight preprocessing + cache uploading strategy. The end node periodically packages the sampling data and uploads it to the corresponding edge computing node via the edge gateway. The edge node deploys a cache queue and a data filtering module, and performs batch reception and verification according to the time window . For data with large model prediction error or outstanding sample representativeness, the edge node will return the label and feedback to the end side to support local adaptive calibration and sample resampling mechanism.

[0087] Step 5.2: Asynchronous aggregation and scheduling trigger mechanism between edge and cloud The edge node is responsible for local model training and parameter updating, and adopts a periodic and event-driven dual scheduling mechanism to upload model parameters to the cloud. When the local training reaches the set iteration number or the model convergence threshold , it triggers to send to the cloud. After the end receives the model uploaded by the edge node exceeding , the aggregation process is automatically started, and steps three (ABC) and four (OT) are called to complete global update. The end has a waiting timeout time for the edge node that does not respond in time to avoid communication blockage causing global update to stop.

[0088] Step 5.3: Heterogeneous bandwidth and scheduling priority strategy Considering the large difference in network conditions of nodes in different areas, the system introduces a bandwidth-aware scheduling module. According to the node bandwidth , response delay and model importance , the scheduling priority is calculated:

[0089] The cloud prioritizes waiting for the results uploaded by high-priority nodes before each round of aggregation, ensuring that the contributions of core nodes of the system are fully integrated. For low-priority nodes, an empirical value backfilling strategy or a delayed merging mechanism can be used to maintain a dynamic balance between communication efficiency and global consistency.

[0090] Step 5.4: Communication scheduling system closed-loop control mechanism To improve system stability and adaptability, a scheduling closed-loop feedback mechanism is designed. The end dynamically adjusts the scheduling parameters such as according to the convergence speed of each round of aggregation, global loss curve fluctuation, edge-end communication failure rate, etc. The edge node autonomously selects to participate or skip uploading in a certain round according to the local communication load and training progress, supporting the "partial participation aggregation" strategy. The system log module records the communication status and model changes each time, providing data support for subsequent system evolution.

[0091] The application provides a multi-region power distribution network-oriented edge computing federated aggregation optimization method AOT, which aims at the problems of traditional federated aggregation algorithms, such as insufficient dynamic adaptation of cross-region data heterogeneity, low parameter alignment accuracy, and the like, and constructs an innovative architecture combining intelligent optimization search and optimal transmission alignment.

[0092] Figure 4 As shown in the drawings, another embodiment of the application provides a multi-region power distribution network-oriented edge computing federated aggregation optimization system, which comprises: The end-side acquisition module is used for acquiring time series data and performing preprocessing operations. The end-side acquisition module is composed of smart meters, FTUs and other devices, and has a sampling frequency of 50Hz, which is used for acquiring time series data and performing compression and anonymization processing.

[0093] The edge training module is deployed on each regional edge node and is used for training a local model based on a local sample set and iteratively updating the model parameters. The edge training module supports asynchronous / quasi-synchronous update modes.

[0094] The cloud aggregation module comprises a weight searching unit and a parameter alignment unit. The weight searching unit is used for searching optimal aggregation weights by using an improved artificial bee colony algorithm. The parameter alignment unit is used for realizing model parameter distribution alignment and weighted fusion by using an optimal transmission mechanism to generate a global model. The communication scheduling module is used for performing cloud-edge-end communication scheduling, adopts a periodic and event-driven dual scheduling mechanism, and combines a priority strategy and a closed-loop control of dynamic adjustment of scheduling parameters.

[0095] The storage module is used for storing a local sample set, model parameters, a verification set and communication logs.

[0096] Further, the ABC weight search unit of the cloud aggregation module supports the configuration of population size and iteration round number parameters, and the OT parameter alignment unit supports the adjustment of entropy regularization parameters.

[0097] Under the cloud-edge-end three-level collaborative architecture of the multi-region power distribution network, the existing methods mostly use fixed weighting or simple adaptive aggregation strategies, which cannot fully adapt to the differences in data distribution and node computing performance among different regions, and easily cause the global model precision to decline, slow convergence, and insufficient stability. The present application introduces an improved artificial bee colony algorithm, performs dynamic weight search during cloud aggregation, takes the global performance index as the optimization target, combines chaos disturbance and adaptive parameter adjustment, accurately evaluates and weights the contribution of each edge node, thereby improving the generalization ability and adaptability of the global model under data heterogeneity conditions. In terms of aggregation parameter distribution alignment, the traditional method directly weights the parameters, which introduces significant distribution deviation, leading to model degradation. The present application introduces an optimal transport mechanism in the cloud aggregation stage, accurately maps the parameter distribution of each edge node to the global optimal reference distribution by establishing a minimum transmission cost mapping, effectively reduces the distribution deviation, and ensures the stable convergence and performance maintenance of the global model under high heterogeneity conditions. In terms of system architecture, the present application constructs a cloud-edge-end three-level aggregation collaborative mechanism: the end-side nodes perform local data training and feature extraction, reducing the amount of raw data transmission; the edge side performs intra-regional rapid aggregation and local optimization, reducing cross-region communication delay; the cloud end completes global aggregation and parameter alignment, and realizes performance maximization through ABC and OT collaborative optimization.

[0098] In summary, the present application breaks through the performance bottleneck of traditional federated aggregation in multi-region heterogeneous power distribution networks, realizes the organic integration of intelligent optimization search and optimal transport alignment, and constructs a high-precision, high-stability, and low-delay cloud-edge-end layered aggregation collaborative architecture in engineering implementation, providing a new technical solution for intelligent power distribution network construction.

[0099] The present application also provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the above-mentioned federated aggregation optimization method.

[0100] The present application is described with reference to flowcharts and / or block diagrams according to the method, device (system), and computer program product of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device realize the functions described in the flowcharts and / or block diagrams.Figure 1 apparatuses that implement a process or processes and / or functions specified in one or more blocks. Figure 1 apparatuses that implement a process or processes and / or functions specified in one or more blocks.

[0101] The above-described embodiments of the present application have been described with respect to the technical solutions, purposes and advantages of the present application. It should be understood that the above-described embodiments are merely specific embodiments of the present application and are not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An edge computing federated aggregation optimization method for a multi-area power distribution network, characterized in that, The method comprises the following steps: S1: constructing a cloud-edge-end three-layer architecture federation modeling system: the end-side device collects time series data and pre-processes to form a sample set, the edge node trains a local model based on the sample set, and iteratively updates the model parameters to the cloud; S2: aggregate weight search space modeling: define a weight vector that satisfies the non-negativity and normalization constraints, construct a K-dimensional simplex search space, and calculate an adaptability function representing the aggregation effect based on a shared validation set; S3: searching for optimal aggregation weights by improving artificial bee colony algorithm: through population initialization, employed bee local search, onlooker bee chaotic disturbance, scout bee re-initialization and elite reservation strategy, the optimal weight vector is searched; S4: parameter aggregation based on optimal transport mechanism: constructing a cost matrix of model parameters, setting edge distribution, solving the entropy regularization optimal transport mechanism, and weighted fusion to generate a global model; S5: cloud-edge-end communication scheduling: adopting a periodic and event-driven dual scheduling mechanism, combining priority strategy and closed-loop control, to realize efficient coordination of data transmission and parameter uploading.

2. The method of claim 1, wherein, In S1, the time series data collected by the end-side device includes voltage, current, power, device operating state and energy consumption plan data, and the preprocessing includes data compression, anonymization processing and outlier removal; the edge node local model training adopts gradient descent mechanism for iterative update.

3. The method of claim 1, wherein, Wherein, The S2, the aggregation weight vector Wherein, satisfy , K The number of edge nodes participating in aggregation; The fitness function is constructed based on the prediction accuracy of the shared validation set, and the formula is: ； wherein, is a shared validation set, is an aggregated global model, is an indicator function.

4. The method of claim 1, wherein, In S3, the optimal aggregation weights are searched based on the improved artificial bee colony algorithm, specifically: Dynamic weight search, the cloud uses an improved artificial bee colony algorithm to search for an optimal aggregation weight vector in a simplex space comprising: Employed bee stage: generating candidate solutions based on power attenuation adaptive disturbance factor; Onlooker bee stage: selecting solutions according to fitness probability and introducing logistic chaotic disturbance; Scout bee stage: re-initializing the unimproved solutions.

5. The method of claim 4, wherein, The employed bee stage is based on the power attenuation adaptive disturbance factor, and the formula is: ； The onlooker bee stage introduces logistic chaotic disturbance, and the chaotic sequence formula is: ； wherein, t is the current iteration number, z t is the logistic chaotic variable at the t th iteration, T max is the maximum iteration number, and γ is the decay coefficient, is the random perturbation coefficient in dimension , which is used to control the perturbation amplitude and direction, and μ is the control parameter of the logistic map, which determines the dynamic behavior of the sequence.

6. The method of claim 1, wherein, In S4, the parameter aggregation based on the optimal transport mechanism includes: Using Euclidean distance to calculate the difference between different client models and constructing the transmission cost matrix of parameters; According to the optimal aggregation weight vector obtained by the improved artificial bee colony algorithm, the source distribution is defined as the optimal weight vector obtained by the ABC algorithm, and the target distribution is defined as the uniform distribution; Introducing an entropy regularization term to enhance numerical stability, and solving the optimal transport plan by Sinkhorn-Knopp iteration; Using the optimal transport plan for weighted fusion to construct a new round of global model parameters.

7. The method of claim 1, wherein, In S5, the cloud-edge-end communication scheduling specifically includes: Asynchronous aggregation and scheduling trigger between edge and cloud: the edge node uses a periodic and event-driven dual scheduling mechanism to upload model parameters to the cloud, and the cloud automatically starts the aggregation process after receiving model parameters uploaded by more than a preset number of edge nodes; Heterogeneous bandwidth and scheduling priority strategy: according to the node bandwidth, response delay and model importance, the scheduling priority is calculated, and the cloud waits for the high-priority node to upload the results before each round of aggregation. Communication scheduling system closed-loop control: according to the convergence speed of each round of aggregation, the global loss curve fluctuation, the edge-end communication failure rate, dynamically adjust the scheduling parameters at all levels, the edge node according to the local communication load and the training progress, independently select to participate or skip a round of upload.

8. An edge computing federated aggregation optimization system for multi-zone power distribution network, characterized in that, It comprises: an end-side acquisition module for collecting time series data and performing preprocessing operations; an edge training module deployed on each regional edge node for training a local model based on a local sample set and iteratively updating and outputting model parameters; a cloud-side aggregation module including a weight searching unit and a parameter alignment unit, the weight searching unit being configured to search for optimal aggregation weights through an improved artificial bee colony algorithm, and the parameter alignment unit being configured to realize model parameter distribution alignment and weighted fusion through an optimal transmission mechanism to generate a global model; a communication scheduling module for performing cloud-edge-end communication scheduling, adopting a periodic and event-driven dual scheduling mechanism, and combining a priority strategy and closed-loop control of dynamic adjustment of scheduling parameters.

9. The system of claim 8, wherein, The ABC weight searching unit of the cloud-side aggregation module supports parameter configuration of population size and iteration number, and the OT parameter alignment unit of the optimal transmission mechanism supports adjustment of entropy regularization parameters.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the multi-region power distribution network oriented edge computing federal aggregation optimization method of any one of claims 1-7.

Images