An edge computing federated aggregation optimization method for a multi-region power distribution network

Abstract

This invention discloses an edge computing federated aggregation optimization method for multi-regional distribution networks, belonging to the field of distribution network collaborative optimization and federated learning technology. This method is based on a three-tiered "cloud-edge-device" architecture, integrating an improved artificial bee colony (ABC) algorithm and an optimal transport (OT) mechanism to construct an AOT aggregation optimization scheme: the edge side collects and preprocesses runtime sequence data, and edge nodes complete local model training and upload parameters; the cloud searches for optimal aggregation weights using the improved ABC algorithm, adapting to differences between node data and models, and uses the OT mechanism to achieve parameter distribution alignment and weighted fusion; efficient aggregation is ensured through dual-trigger collaborative communication scheduling. This invention solves the problems of fixed weights and insufficient parameter alignment in traditional federated aggregation, significantly improving the global model accuracy, stability, and convergence speed in scenarios where data is not independent and identically distributed, reducing communication overhead, and providing reliable support for intelligent collaborative scheduling of multi-regional distribution networks.

Description

Technical Field

[0001] This invention relates to the field of distribution network collaborative optimization and federated learning technology, specifically to an edge computing federated aggregation optimization method for multi-region distribution networks. Background Technology

[0002] In the development of distributed energy systems and smart parks, an increasing number of independent entities (such as zero-carbon factories, green parks, and industrial clusters) are forming interconnected regional power grids. These entities have significant differences in energy production, consumption, and dispatching models, as well as different load characteristics, operating strategies, and equipment types, but they all need to carry out collaborative management and optimized dispatching under the premise of power grid safety and stability.

[0003] Traditional centralized energy management relies on aggregating operational data from various entities (such as real-time power load, equipment operating status, and energy consumption plans) onto a unified central platform for modeling and decision-making. However, this model presents significant challenges in multi-entity scenarios. First, operational data involves sensitive information such as production processes, energy efficiency, and market transaction details; directly sharing raw data can easily lead to the leakage of trade secrets and privacy compliance risks. Second, when the number of entities is large and the data volume is massive, centralized computing and storage can easily become performance bottlenecks, demanding high communication bandwidth and prone to data transmission delays and network congestion. Third, the energy usage patterns of different entities vary significantly, and the collected data exhibits non-independent identically distributed (Non-IID) characteristics; directly employing a fixed-weighted global modeling method can easily result in slow model convergence and decreased accuracy.

[0004] To address these issues, Federated Learning (FL) offers a novel technical approach. This method allows each agent to train its model locally, uploading only model parameters or gradients to the upper layer for aggregation, thus achieving joint modeling across agents while ensuring data remains local. However, traditional federated learning methods still suffer from two major limitations in multi-agent distribution network scenarios. First, they cannot dynamically adjust aggregation weights based on the data quality, quantity, and model performance of each agent, leading to insufficient or excessive contributions from some agents to the global model. Second, they lack an effective alignment mechanism for differences in parameter distributions from different sources during the model fusion phase, making it difficult to maintain stable global performance in highly heterogeneous environments.

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

[0006] Existing federated learning aggregation methods (such as FedAvg) typically perform uniform, fixed-weight aggregation in the cloud, lacking dynamic awareness of differences in model quality and heterogeneity in data distribution among edge nodes. When the data between edge nodes exhibits strong non-independent identically distributed (Non-IID) characteristics, fixed weights can lead to biases in global model parameter updates, resulting in slow convergence, decreased accuracy, and even training instability. Simultaneously, due to significant differences in the distribution of model parameters among edge nodes, the lack of an effective parameter alignment mechanism can lead to decreased model generalization performance during cross-regional collaborative aggregation.

[0007] Furthermore, in a layered cloud-edge-device architecture, there are often challenges in time synchronization, communication latency, and uneven distribution of computing resources between edge-layer aggregation and cloud-based global aggregation. These problems are amplified, especially in the complex operating environment of multi-regional power grids, limiting the overall scheduling efficiency and intelligence 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 edge computing scenarios, in order to fully unleash the performance potential of the cloud-edge-device collaborative architecture and improve the intelligent scheduling and operational stability of multi-regional distribution networks. Summary of the Invention

[0008] This addresses the systemic shortcomings of current multi-regional power distribution networks in cloud-edge-device collaborative operation, such as insufficient model aggregation adaptability, low parameter alignment accuracy, and limited convergence performance under data heterogeneity conditions.

[0009] To achieve the above objectives, this invention proposes an edge computing federated aggregation optimization method, AOT (Aggregated Optimal Transport), for multi-regional power distribution networks. It combines the improved Artificial BeeColony (ABC) algorithm with the Optimal Transport (OT) mechanism to achieve dynamic weight search and model parameter alignment during model aggregation in a three-tiered "cloud-edge-device" architecture. The improved AOT adaptively adjusts the aggregation weights based on the data distribution and model performance of each node. During aggregation, OT performs optimal mapping of parameter distribution to reduce discrepancies, thereby improving the accuracy, stability, and convergence speed of the global model under conditions of data heterogeneity. This overcomes the performance bottleneck of traditional fixed-weighted aggregation in cross-regional collaboration.

[0010] More importantly, this invention not only focuses on improving the accuracy of the global model, but also considers multiple optimization objectives such as convergence speed, communication overhead, and system robustness, forming a closed-loop aggregation system of "dynamic weight optimization – parameter distribution alignment – ​​multi-indicator synergistic improvement". In complex distribution network operation scenarios, this method significantly improves the performance of the global model under highly heterogeneous and multi-tasking conditions, and maintains the advantage of local data processing while reducing communication costs, providing a new technical path for achieving efficient collaborative scheduling and intelligent decision-making in multi-regional distribution networks.

[0011] One aspect of the present invention protects an edge computing federated aggregation optimization method for multi-regional distribution networks, comprising the following steps:

[0012] S1 constructs a three-layer cloud-edge-device federated modeling system: edge devices collect time-series data and preprocess it to form a sample set, edge nodes train local models based on the sample set, and upload the model parameters to the cloud after iterative updates;

[0013] S2 Aggregation Weight Search Space Modeling: Define weight vectors that satisfy nonnegativity and normalization constraints, construct a K-dimensional simplex search space, and calculate the fitness function representing the aggregation effect based on a shared validation set;

[0014] S3 Improved Artificial Bee Colony Algorithm for Searching Optimal Aggregation Weights: The optimal weight vector is obtained by population initialization, local search by hired bees, chaotic perturbation by observer bees, re-initialization by scout bees, and elite retention strategies.

[0015] S4 Parameter aggregation based on optimal transmission mechanism: construct the cost matrix of model parameters, set the marginal distribution, solve the entropy regularization optimal transmission mechanism, and generate a global model by weighted fusion;

[0016] S5 Cloud-Edge-Device Communication Scheduling: Employs a dual scheduling mechanism of periodicity and event-driven scheduling, combined with priority strategies and closed-loop control, to achieve efficient coordination of data transmission and parameter uploading.

[0017] 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.

[0018] 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:

[0019]

[0020] in, To share the verification set, For the aggregated global model, This is an indicator function.

[0021] Furthermore, S3, based on the improved artificial bee colony algorithm, searches for the optimal aggregation weight, specifically as follows:

[0022] 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:

[0023] Hired bee stage: Generate candidate solutions based on power-law decay adaptive perturbation factor;

[0024] Observation phase: Select solutions based on fitness probability and introduce logistic chaotic perturbation;

[0025] Reconnaissance Bee Phase: Reinitialize the unimproved solution.

[0026] Furthermore, the hired bee phase is based on an adaptive perturbation factor with power-law decay, as follows:

[0027] ,

[0028] During the observation phase, a logistic chaotic perturbation is introduced, and the chaotic sequence formula is:

[0029] ,

[0030] in, t This represents the current iteration number. z t For the firstt 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 The random disturbance coefficient is used to control the amplitude and direction of the disturbance.

[0031] Furthermore, S4, which implements parameter aggregation based on the optimal transmission mechanism, includes:

[0032] The differences between different client models are calculated using Euclidean distance, and the transmission cost matrix of the parameters is constructed.

[0033] Based on the optimal aggregated 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 a uniform distribution.

[0034] An entropy regularization term is introduced to enhance numerical stability, and the optimal transmission plan is solved by Sinkhorn-Knopp iteration.

[0035] We use the optimal transmission plan to perform weighted fusion and construct a new round of global model parameters.

[0036] Furthermore, S5, cloud-edge-device communication scheduling specifically includes:

[0037] Asynchronous aggregation and scheduling triggering between edge and cloud: Edge nodes use a dual scheduling mechanism of periodicity and event-driven to upload model parameters to the cloud. After the cloud receives more than a preset number of uploaded models from edge nodes, it automatically starts the aggregation process.

[0038] Heterogeneous bandwidth and scheduling priority strategy: Calculate the scheduling priority of nodes based on their bandwidth, response latency and model importance, and the cloud will wait for the results uploaded by high-priority nodes before each round of aggregation;

[0039] The communication scheduling system implements closed-loop control: the endpoint dynamically adjusts the scheduling parameters at each level based on the convergence speed of each round of aggregation, the fluctuation of the global loss curve, and the edge-end communication failure rate. The edge nodes autonomously choose to participate in or skip a round of uploading based on the local communication load and training progress.

[0040] In another aspect, this invention protects an edge computing federated aggregation optimization system for multi-regional distribution networks, comprising:

[0041] The edge-side acquisition module is used to acquire time-series data and perform preprocessing operations;

[0042] The edge training module is deployed at edge nodes in various regions to train local models based on local sample sets and output model parameters after iterative updates.

[0043] The cloud aggregation module includes a weight search unit and a parameter alignment unit. The weight search unit is used to search for the optimal aggregation weights through an improved artificial bee colony algorithm. The parameter alignment unit is used to align and weightedly fuse model parameters through an optimal transmission mechanism to generate a global model.

[0044] The communication scheduling module is used to perform cloud-edge-device communication scheduling. It adopts a dual scheduling mechanism of periodicity and event-driven scheduling, combined with closed-loop control that dynamically adjusts priority strategies and scheduling parameters.

[0045] Furthermore, the ABC weight search unit of the cloud aggregation module supports parameter configuration for population size and number of iterations, and the optimal transmission mechanism OT parameter alignment unit supports entropy regularization parameter adjustment.

[0046] In another aspect, the present invention protects a computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the aforementioned edge computing federated aggregation optimization methods for multi-regional distribution networks.

[0047] Compared with the prior art, the present invention has the following significant advantages:

[0048] Compared to traditional fixed-weighted federated aggregation, this method establishes an adaptive aggregation mechanism that adjusts weight allocation in real time based on node data distribution and model contribution, significantly improving the accuracy and convergence speed of the global model in complex multi-regional scenarios, while effectively suppressing the negative impact of low-quality updates on global performance. Based on optimal transmission parameter mapping technology, it overcomes the accuracy bottleneck of existing methods in heterogeneous node parameter alignment, significantly reducing the difference in cross-regional model parameter distribution, thereby improving the generalization ability and stability of the global model. By offloading some computation and aggregation tasks to the edge layer and introducing dynamic weights and chaotic perturbation mechanisms, the system achieves simultaneous optimization in communication latency, aggregation response speed, and robustness, maintaining stable global performance even under adverse conditions such as node failure and communication fluctuations. This technical system provides key technical support for building an intelligent cloud-edge-device collaborative learning platform with autonomous aggregation decision-making, dynamic optimization, and cross-regional collaborative capabilities for multi-regional power distribution networks.

[0049] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention.

[0050] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a schematic diagram of an edge computing federated aggregation optimization method for multi-regional power distribution networks in an embodiment of the present invention;

[0053] Figure 2 This is an architecture diagram of an edge computing federated aggregation optimization method for multi-regional power distribution networks in an embodiment of the present invention;

[0054] Figure 3 This 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;

[0055] 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

[0056] 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.

[0057] 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.

[0058] Figure 1 As shown, an edge computing federated aggregation optimization method for multi-regional distribution networks includes the following steps:

[0059] 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.

[0060] Step 1.1: End-side data acquisition and preprocessing mechanism

[0061] 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.

[0062] 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 .

[0063] Step 1.2: Edge-side local model training mechanism

[0064] 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:

[0065] 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 When convergence is detected, training is terminated.

[0066] 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.

[0067] Step 1.3: Cloud-based model aggregation and global knowledge fusion mechanism

[0068] 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.

[0069] 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.

[0070] Based on the differences in parameter distribution among edge models, construct the optimal transmission loss:

[0071]

[0072] in, This represents the transmission plan from parameter a to b. OT is used as its cost function to alleviate the aggregation offset problem caused by model misalignment.

[0073] The improved artificial bee colony optimization module (ABC) is used to search for the optimal aggregation weights. This is done to maximize the performance metrics of the global model under the objective function. The final aggregated model expression is:

[0074]

[0075] in This is the parameter representation after optimal transmission. The optimal weights are obtained through artificial bee colony search. This method significantly outperforms traditional fixed-weighting strategies such as FedAvg, and can dynamically adapt to the heterogeneity of data at each edge node and the differences in model training capabilities.

[0076] Step 2: Constructing an aggregated weight search space modeling mechanism

[0077] Step 2.1: Define the weight vector

[0078] In a federated learning system with a "cloud-edge-device" architecture, assuming the current communication round is t+1, there are a total of Edge nodes participate in the aggregation process, each node After local training is complete, generate model parameters. To achieve global aggregation that balances personalization and collaboration, a non-negative weight needs to be assigned to the model update of each node to measure its contribution to the aggregation process. Let:

[0079]

[0080] The weight vector corresponding to this round of aggregation must satisfy the following constraints:

[0081] (1) The weight of each node is non-negative, i.e. .

[0082] (2) The weight sum is 1, that is .

[0083] The vector It represents the proportion of parameter updates for each edge node in the global model and is a core variable for the optimization objective.

[0084] Step 2.2: Search Space Construction

[0085] The above constraints constitute a standard K-dimensional simplex space, namely:

[0086]

[0087] From a geometric perspective, each point in this space represents a valid 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 a key foundation for breaking away from traditional static weighting (such as FedAvg).

[0088] Step 2.3: Evaluation Mechanism of Aggregation Effect

[0089] To assign weight vectors to each candidate To conduct performance evaluation, this invention introduces a method based on a shared small-scale validation set. Accuracy is used as the fitness function:

[0090]

[0091] in, Indicates using the current weight combination The aggregated global model for input samples The prediction results This is an indicator function (it takes 1 if the condition is true, otherwise it takes 0).

[0092] This function provides an evaluation basis for subsequent optimization searches, and because the validation set is small, it enables rapid feedback and dynamic adjustment in each round.

[0093] Step 3: Aggregated Weight Search Mechanism Based on Improved Artificial Bee Colony Algorithm

[0094] Step 3.1: Population Initialization

[0095] In the process of federated aggregation optimization, in order to obtain the optimal aggregation weight vector... The present invention first defines a pre-defined... 3D simplex search space An initial search population is constructed. Let the size of the search population be... Then initialization is required. An aggregate weight vector that satisfies the constraints. And as a candidate solution.

[0096] For each individual ,from Generate a set of uniformly random numbers within the interval And it is mapped to the simplex space through a normalization operation:

[0097]

[0098] in a i,k Indicates the first i Among the candidate solutions, the first one in the federated system k The aggregate weight assigned to each client has a value range of [value range missing]. k =1, …, K That is, in K The 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.

[0099] 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:

[0100]

[0101] in, It is an indicator function. Indicates weight The global model after aggregating the edge node models.

[0102] Step 3.2: Candidate solution generation and local search update in the hired bee phase

[0103] 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:

[0104]

[0105] 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:

[0106]

[0107] 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.

[0108] Newly generated candidate vectors The projection operation will renormalize the model to simplex space. Internally, ensure legality:

[0109]

[0110] Next, calculate its fitness and execute the following update strategy:

[0111]

[0112] 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:

[0113]

[0114] To further prevent getting trapped in local optima, a chaotic perturbation mechanism is introduced into the optimal solution. A chaotic sequence is generated using a logistic mapping:

[0115]

[0116] in, z t For the first t The initial value of the logistic chaotic variable in the next iteration. z 0 Given a random value ∈(0,1), the above formula is used to iterate and obtain a one-dimensional chaotic sequence located in the interval (0,1). μ The control parameters for the logistic mapping determine the dynamic behavior of the sequence. μ=At time 4, the system is in a fully chaotic state, exhibiting the best sequence ergodicity and randomness.

[0117] The disturbance takes the form of:

[0118]

[0119] Where ε∈(0,1) is the chaotic perturbation intensity parameter, used to scale the logistic chaotic sequence. The larger the perturbation amplitude ε of the current solution, the larger the search step size and the stronger the global exploration capability. The smaller ε is, the more conservative the update is, which is conducive to fine local search.

[0120] This mechanism expands the search space while maintaining local search capabilities.

[0121] Step 3.4: Reconnaissance Bee Mechanism and Elite Retention Strategy

[0122] For consecutive, unimproved candidate solutions, a scout bee mechanism is used in the search space. The algorithm reinitializes the solution to avoid early stagnation in the search. It also employs an elitist retention strategy, saving the current optimal solution in each iteration. This ensures that the global optimum is not lost. The final weight update rule is:

[0123]

[0124] Through the above multi-mechanism synergistic optimization, the following results were obtained Aggregated weights take into account data heterogeneity, model performance, and convergence speed, laying the foundation for subsequent optimal transport aggregation.

[0125] Step 4: Model aggregation strategy based on optimal transmission mechanism

[0126] Step 4.1: Construct the parameter distribution distance matrix between client models

[0127] Let the first The client in the first The local model parameters obtained at the end of the round communication are

[0128]

[0129] The differences between different client models are characterized by Euclidean distance, and a parameter transmission cost matrix is ​​constructed. :

[0130]

[0131] in, This indicates that during optimal transmission, "quality" is transferred from the client... The model parameter distribution is transferred to the client. The unit transmission cost when the model parameters are distributed provides a geometric metric basis for subsequent OT optimization.

[0132] Step 4.2: Set up edge distribution for transmission

[0133] Let the aggregate weight vector obtained by optimizing the improved artificial bee colony algorithm be denoted as .

[0134]

[0135] This invention uses it as the source distribution in the optimal transmission problem:

[0136]

[0137] Target distribution It can be flexibly configured according to application requirements. By default, q can be set to a uniform distribution, i.e.

[0138]

[0139] This achieves a balanced aggregation of contributions from all parameter components. When it is necessary to highlight certain parameter subspaces or nodes in specific regions, q can also be customized according to the global model structure characteristics and scheduling strategy, as long as the non-negativity and normalization constraints are satisfied.

[0140]

[0141] Therefore, p and q together constitute the edge constraints of the optimal transmission problem, reflecting the aggregate contribution of each edge node and the global model objective structure, respectively.

[0142] Step 4.3: Solve for the entropy-regularized optimal transmission plan

[0143] Solving the transmission plan refers to solving a problem given a source distribution. p Target distribution q Given the cost matrix C, the transmission plan matrix γ that satisfies the edge constraints and minimizes the entropy regularization transmission cost is obtained through Sinkhorn–Knopp iteration.

[0144] To enhance numerical stability and improve solution efficiency, entropy regularization is introduced into the classical optimal transport model, and the transport problem is modeled as follows:

[0145]

[0146] in: For the transmission plan matrix, This represents the quality allocated from the k-th component of the source distribution to the j-th component of the target distribution in the transmission scheme:

[0147]

[0148] To satisfy the set of feasible transmission plans, Total transmission cost; For the entropy of the transmission plan, This is a regularization parameter used to adjust the weight of the entropy term, balancing minimum cost and distribution smoothness.

[0149] To efficiently solve the above optimization problem, the Sinkhorn–Knopp iterative algorithm is adopted. Let the kernel matrix be...

[0150]

[0151] The transmission plan matrix can then be represented as

[0152]

[0153] in Let be a scaling vector, where, This represents the operator for constructing a diagonal matrix from vectors. It achieves the satisfaction of edge constraints through alternating normalization, and its iterative form is:

[0154]

[0155] Where “ / ” represents element-wise division. This represents the number of iterations. As the iterations proceed...

[0156]

[0157] The row and column sums gradually converge to p and q. When the edge constraint error is lower than a preset threshold, the iteration is considered to have converged, and the transmission plan matrix obtained at this time is denoted as Γ, which is the entropy-regularized optimal transmission plan.

[0158] Step 4.4: Perform parameter alignment and generate a global model

[0159] After obtaining the optimal transmission plan matrix Γ, the parameters of each client model are weighted and fused to align the parameter distribution and generate a new round of global model parameters. The effective aggregation weights for each client are defined as follows:

[0160]

[0161] The global model parameters are then updated as follows:

[0162]

[0163] in, For each element in the optimal transmission plan matrix, represents the proportion of quality that the parameter distribution of the k-th client model is allocated to the j-th component of the global target model under the optimal transmission scheme. This can be understood as a modified aggregation weight after considering OT geometric alignment. Through this mechanism, in highly heterogeneous scenarios of multi-regional distribution networks, the deviation of cross-node model parameter distribution can be significantly reduced, improving the accuracy, stability, and convergence speed of the global model while ensuring that the data does not leave the local area.

[0164] Step 5: Design of Cloud-Edge-Device Communication Scheduling Mechanism

[0165] Step 5.1: Edge-to-Device Data Upload and Model Feedback Mechanism

[0166] Terminal devices primarily handle real-time data collection of operational status data such as load, voltage, and power. While the sampling period is short, computing power is limited. This invention employs a lightweight preprocessing + cache upload strategy on the edge. The edge node periodically packages the sampled 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, filtering data according to time windows. Batch reception and verification are performed. For data with large model prediction errors or outstanding sample representativeness, edge nodes transmit them back with tags and feed them back to the end side to support local adaptive calibration and sample resampling mechanisms.

[0167] Step 5.2: Edge-Cloud Asynchronous Aggregation and Scheduling Trigger Mechanism

[0168] Edge nodes are responsible for local model training and parameter updates, employing a dual scheduling mechanism of periodicity and event-driven methods to upload model parameters to the cloud. Local training continues until a set number of iterations is reached. Or model convergence threshold At that time, it triggers the sending to the cloud. The end received more than After the edge node uploads the model, the aggregation process is automatically initiated, calling steps three (ABC) and four (OT) to complete the global update. The client sets a timeout period for edge nodes that do not respond in a timely manner. This avoids communication blockage that could cause global updates to stall.

[0169] Step 5.3: Heterogeneous Bandwidth and Scheduling Priority Strategy

[0170] Considering the significant differences in network conditions across different regions, the system introduces a bandwidth-aware scheduling module. This module is based on node bandwidth... Response delay With model importance Calculate its scheduling priority:

[0171]

[0172] Before each round of aggregation, the cloud prioritizes waiting for results uploaded by high-priority nodes to ensure that the contributions of core system nodes 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.

[0173] Step 5.4: Closed-loop control mechanism of communication dispatch system

[0174] To improve system stability and adaptability, a closed-loop feedback mechanism for scheduling is designed. The endpoint dynamically adjusts scheduling parameters at each level based on factors such as the convergence speed of each aggregation round, fluctuations in the global loss curve, and the edge-to-end communication failure rate. Each node can autonomously choose to participate in or skip a round of uploads based on its local communication load and training progress, supporting a "partially participatory aggregation" strategy. The system log module records the status of each communication and model changes, providing data support for subsequent system evolution.

[0175] This invention proposes an edge computing federated aggregation optimization method (AOT) for multi-regional distribution networks. It addresses the problems of traditional federated aggregation algorithms, such as cross-regional data heterogeneity, insufficient dynamic adaptation of aggregation weights, and low parameter alignment accuracy, by constructing an innovative architecture that combines intelligent optimization search with optimal transmission alignment.

[0176] Figure 4 As shown, another embodiment of the present invention provides an edge computing federated aggregation optimization system for multi-regional distribution networks, comprising:

[0177] The edge acquisition module is used to collect time-series data and perform preprocessing operations. The edge acquisition module consists of devices such as smart meters and FTUs, with a sampling frequency of 50Hz, and is used to collect time-series data and perform compression and anonymization processing.

[0178] The edge training module is deployed at edge nodes in various regions to train local models based on local sample sets and output model parameters after iterative updates; it supports asynchronous / quasi-synchronous update modes.

[0179] The cloud aggregation module includes a weight search unit and a parameter alignment unit. The weight search unit is used to search for the optimal aggregation weights through an improved artificial bee colony algorithm. The parameter alignment unit is used to align and weightedly fuse model parameters through an optimal transmission mechanism to generate a global model.

[0180] The communication scheduling module is used to perform cloud-edge-device communication scheduling. It adopts a dual scheduling mechanism of periodicity and event-driven scheduling, combined with closed-loop control that dynamically adjusts priority strategies and scheduling parameters.

[0181] Storage module: Used to store local sample sets, model parameters, validation sets, and communication logs.

[0182] Furthermore, the ABC weight search unit of the cloud aggregation module supports parameter configuration for population size and number of iterations, and the optimal transmission mechanism OT parameter alignment unit supports entropy regularization parameter adjustment.

[0183] In the cloud-edge-device three-level collaborative architecture of multi-regional power distribution networks, most existing methods employ fixed weighting or simple adaptive aggregation strategies, which cannot fully adapt to the differences in data distribution and node computing performance between different regions. This easily leads to decreased global model accuracy, slow convergence, and insufficient stability. This invention introduces an improved artificial bee colony algorithm, performing dynamic weight search during cloud aggregation. Using global performance indicators as the optimization objective, and combining chaotic perturbations with adaptive parameter adjustment, it accurately evaluates and weights the contribution of each edge node, thereby improving the generalization ability and adaptability of the global model under heterogeneous data conditions. Regarding the alignment of aggregation parameter distribution, traditional methods directly weight parameters, introducing significant distribution shifts and causing model degradation. This invention introduces an optimal transmission mechanism in the cloud aggregation stage. By establishing a minimum transmission cost mapping, it accurately maps the parameter distribution of each edge node to the globally optimal reference distribution, effectively reducing distribution bias and ensuring stable convergence and performance maintenance of the global model under highly heterogeneous conditions. In terms of system architecture, this invention constructs a three-level aggregation and collaboration mechanism of cloud-edge-device: the device-side node performs local data training and feature extraction to reduce the amount of raw data transmission; the edge side performs rapid aggregation and local optimization within the region to reduce cross-regional communication latency; the cloud completes global aggregation and parameter alignment, and maximizes performance through ABC and OT collaborative optimization.

[0184] In summary, this invention breaks through the performance bottleneck of traditional federated aggregation in multi-regional heterogeneous distribution networks, realizes the organic integration of intelligent optimization search and optimal transmission alignment, and constructs a high-precision, high-stability, and low-latency cloud-edge-device hierarchical aggregation collaborative architecture in engineering implementation, providing a new technical solution for the construction of smart distribution networks.

[0185] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described federated aggregation optimization method.

[0186] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0187] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An edge computing federated aggregation optimization method for multi-regional distribution networks, characterized in that, Includes the following steps: S1 constructs a three-layer cloud-edge-device federated modeling system: edge devices collect time-series data and preprocess it to form a sample set, edge nodes train local models based on the sample set, and upload the model parameters to the cloud after iterative updates; S2 Aggregation Weight Search Space Modeling: Define weight vectors that satisfy nonnegativity and normalization constraints, construct a K-dimensional simplex search space, and calculate the fitness function representing the aggregation effect based on a shared validation set; S3 Improved Artificial Bee Colony Algorithm for Searching Optimal Aggregate Weights: The optimal weight vector is obtained through population initialization, local search by hired bees, chaotic perturbation by observer bees, re-initialization by scout bees, and elite retention strategies. Dynamic weight search is performed in the cloud using the improved artificial bee colony algorithm to search for the optimal aggregate weight vector in simplex space. Specifically: 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; The hired bee phase is based on an adaptive perturbation factor with power-law decay, and the formula is: ； 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 Logistic chaotic variables at the next iteration T max Where γ is the maximum number of iterations, and γ is the decay coefficient. In dimension The random perturbation coefficients on the logistic map are used to control the amplitude and direction of the perturbation, and μ is the control parameter of the logistic map, which determines the dynamic behavior of the sequence. S4 Parameter aggregation based on the optimal transmission mechanism: Constructing the cost matrix of model parameters, setting marginal distributions, solving for the entropy-regularized optimal transmission mechanism, and weighted fusion to generate a global model; including: The differences between different client models are calculated using Euclidean distance, and the transmission cost matrix of the parameters is constructed. Based on the optimal aggregated 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 a uniform distribution. An entropy regularization term is introduced to enhance numerical stability, and the optimal transmission plan is solved by Sinkhorn-Knopp iteration. We use the optimal transmission plan to perform weighted fusion and construct a new round of global model parameters; S5 Cloud-Edge-Device Communication Scheduling: Employs a dual scheduling mechanism of periodicity and event-driven scheduling, combined with priority strategies and closed-loop control, to achieve efficient coordination of data transmission and parameter uploading.

2. The method according to claim 1, characterized in that, 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.

3. The method according to claim 1, characterized in that, in, In S2, the aggregated 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: ； in, To share the verification set, For the aggregated global model, This is an indicator function.

4. The method according to claim 1, characterized in that, The S5, cloud-edge-device communication scheduling, specifically includes: Asynchronous aggregation and scheduling triggering between edge and cloud: Edge nodes use a dual scheduling mechanism of periodicity and event-driven to upload model parameters to the cloud. After the cloud receives more than a preset number of uploaded models from edge nodes, it automatically starts the aggregation process. Heterogeneous bandwidth and scheduling priority strategy: Calculate the scheduling priority of nodes based on their bandwidth, response latency and model importance, and the cloud will wait for the results uploaded by high-priority nodes before each round of aggregation; The communication scheduling system implements closed-loop control: the endpoint dynamically adjusts the scheduling parameters at each level based on the convergence speed of each round of aggregation, the fluctuation of the global loss curve, and the edge-end communication failure rate. The edge nodes autonomously choose to participate in or skip a round of uploading based on the local communication load and training progress.

5. An edge computing federated aggregation optimization system for multi-regional distribution networks, characterized in that, The method applied to any one of claims 1-4 includes: The edge-side acquisition module is used to acquire time-series data and perform preprocessing operations; The edge training module is deployed at edge nodes in various regions to train local models based on local sample sets and output model parameters after iterative updates. The cloud aggregation module includes a weight search unit and a parameter alignment unit. The weight search unit is used to search for the optimal aggregation weights through an improved artificial bee colony algorithm. The parameter alignment unit is used to align and weightedly fuse model parameters through an optimal transmission mechanism to generate a global model. The communication scheduling module is used to perform cloud-edge-device communication scheduling. It adopts a dual scheduling mechanism of periodicity and event-driven scheduling, combined with closed-loop control that dynamically adjusts priority strategies and scheduling parameters.

6. The system according to claim 5, characterized in that, The ABC weight search unit of the cloud aggregation module supports parameter configuration for population size and number of iterations, while the OT parameter alignment unit of the optimal transmission mechanism supports parameter adjustment for entropy regularization.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the edge computing federated aggregation optimization method for multi-regional distribution networks as described in any one of claims 1-4.

Images