Power distribution network cluster local control method, device, equipment and medium

Abstract

The invention discloses a power distribution network cluster local control method and device, equipment and a medium. The method adopts a cloud edge end three-level collaborative architecture, and comprises the following steps: acquiring electrical parameters through a terminal equipment layer; constructing a graph model based on an electrical distance at the edge calculation layer; performing initial cluster division and issuing on the cloud computing layer; dynamically optimizing cluster division in an edge calculation layer through a lightweight deep reinforcement learning model; local control strategies such as voltage reactive power cooperative control, frequency active power balance control and fault rapid isolation and self-healing are executed based on the final division. The system comprises a terminal device layer, an edge computing layer and a cloud computing layer. According to the method, decoupling of global optimization and local real-time control is realized through cloud-side cooperation, the problems of high data processing pressure, high communication delay, expanded local fault influence range and the like of centralized control are effectively solved, and the real-time performance, reliability and disaster resistance of the power distribution network are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of smart distribution networks, specifically relating to a method, device, equipment, and medium for local control of a distribution network cluster. Background Technology

[0002] With the widespread adoption of distributed resources such as photovoltaic power generation, wind power generation, and electric vehicles in power distribution networks, traditional centralized power distribution network management models face significant challenges. The surge in data volume and computational complexity, driven by the massive influx of distributed energy resources and user-side equipment, has led to an exponential increase in power distribution network data, making it difficult for traditional centralized data processing methods to meet real-time requirements. Communication latency and reliability issues are prominent; centralized control relies on communication between the master station and terminal equipment, and communication delays and network failures can impact the real-time performance and reliability of control. Furthermore, the impact of localized faults expands; under centralized control models, localized faults can easily spread to a wider area, affecting the safe and stable operation of the power distribution network. Summary of the Invention

[0003] The purpose of this invention is to provide a method, apparatus, equipment and medium for local control of a power distribution network cluster, so as to at least solve or improve one of the problems in the prior art.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for local control of a power distribution network cluster, employing a three-level collaborative architecture of cloud, edge, and terminal, comprising: Electrical parameters of multiple nodes in the power distribution network are collected by the terminal device layer and sent to the edge computing layer. In the edge computing layer, the electrical distance between nodes is calculated based on electrical parameters, and a graph model of the distribution network is constructed based on the electrical distance, where nodes represent distribution network nodes and edge weights are based on electrical distance; At the cloud computing layer, a graph partitioning algorithm is used to initially partition the graph model into clusters, and the partitioning results are sent down to the edge computing layer. At the edge computing layer, a lightweight deep reinforcement learning model is used to dynamically optimize the initial cluster partitioning based on real-time network state and global constraints issued from the cloud computing layer, resulting in the final cluster partitioning. At the edge computing layer, local control strategies are executed based on the final cluster partitioning. These local control strategies include voltage and reactive power coordinated control, frequency and active power balance control, and rapid fault isolation and self-healing.

[0005] This invention decouples global optimization from local real-time control by constructing a "cloud-edge-device three-level collaborative architecture." Local data processing and rapid response are achieved through the edge computing layer, avoiding the uploading of massive amounts of data to the cloud, significantly reducing the burden on cloud computing centers and meeting the millisecond-level real-time requirements of the distribution network. Core control logic is decentralized to the edge, reducing reliance on cloud master station communication. Even when cloud-edge communication is temporarily interrupted, the edge layer can still maintain autonomous cluster operation, greatly improving system reliability and disaster recovery capabilities. Dynamic cluster partitioning decouples the large power grid into multiple autonomous clusters, achieving localized fault isolation and self-healing, effectively preventing fault propagation across regions and improving the overall safety and stability of the distribution network.

[0006] Furthermore, electrical parameters of multiple nodes in the distribution network are collected through the terminal device layer and sent to the edge computing layer, including the following steps: Multi-source sensors deployed at distribution network nodes are used to collect electrical measurement data, environmental data, and equipment status data in real time. The electrical measurement data includes voltage, current, and power; the environmental data includes temperature and light intensity; and the equipment status data includes the state of charge of the energy storage system and the photovoltaic output status. The collected data is aggregated and cleaned by edge computing nodes, and features are extracted to generate standardized input vectors. The standardized input vector is sent to the edge computing layer.

[0007] By collecting multi-source data, including electrical, environmental, and equipment status data, richer and more accurate input information is provided for subsequent cluster partitioning and control decisions, laying a data foundation for precision control. Through data aggregation, cleaning, and feature extraction at the edge, noise in the original data is eliminated, and standardized input vectors are generated, reducing the processing burden on upper-layer models and improving the overall system efficiency.

[0008] Furthermore, at the edge computing layer, the electrical distance between nodes is calculated based on electrical parameters, and a graph model of the distribution network is constructed based on the electrical distance, where nodes represent distribution network nodes and edge weights are based on electrical distance. This includes the following steps: Based on electrical parameters, the active voltage sensitivity matrix and reactive voltage sensitivity matrix of the distribution network are calculated. Calculate the electrical distance between any two nodes based on the active voltage sensitivity matrix and the reactive voltage sensitivity matrix. The edge weight between node i and node j is defined based on the electrical distance. Based on all distribution network nodes and edges defined by edge weights, a weighted graph model for cluster partitioning is constructed.

[0009] The above scheme uses "electrical distance" rather than simply physical distance or communication quality as the basis for defining edge weights. This results in tight electrical coupling and frequent power exchange within the partitioned clusters, which better reflects the actual physical characteristics of the distribution network and facilitates power self-balancing within the clusters. Calculating electrical distance based on the active and reactive voltage sensitivity matrices accurately quantifies the degree of mutual influence between nodes, providing a scientific and reliable objective function for subsequent optimization algorithms, leading to better partitioning results.

[0010] Furthermore, at the cloud computing layer, a graph partitioning algorithm is used to initially partition the graph model into clusters, and the partitioning results are sent down to the edge computing layer, including the following steps: A device topology graph is constructed based on a weighted graph model. In this model, distribution network nodes are abstracted as weighted nodes, and the node weights include device computing power, storage capacity, and remaining power consumption. The connections between nodes are abstracted as weighted edges, and the edge weights are calculated using a composite function based on electrical distance and incorporating communication delay and bandwidth. With minimizing the total cost of inter-cluster communication as the primary objective and balancing the computing resources of each cluster as a collaborative objective, a multi-objective optimization function for graph partitioning is constructed. The computing power and storage capacity of the devices are injected as constraints into the graph partitioning algorithm, and the k-way partitioning function is called to partition the device topology graph and generate an initial cluster partitioning scheme. The initial cluster partitioning scheme is distributed to the edge computing layer.

[0011] By constructing a multi-objective function encompassing device computing power, communication costs, and load balancing, the initial partitioning scheme is not only electrically reasonable but also near-optimal in terms of computing and communication resources, providing a high-quality foundational framework for dynamic optimization at the edge. By injecting constraints such as device computing power and storage capacity, it is ensured that the actual resources of the edge devices in each partitioned cluster can support subsequent model inference and control tasks, avoiding the problem of theoretical partitioning results being unenforceable.

[0012] Furthermore, at the edge computing layer, a lightweight deep reinforcement learning model is used to dynamically optimize the initial cluster partitioning based on real-time network state and global constraints issued from the cloud computing layer, resulting in the final cluster partitioning. This includes the following steps: Constructing a deep reinforcement learning model includes: a state space, an action space, and a reward function; among which, The state space includes device state, network state, and task state. Device state includes the real-time computing power utilization, storage occupancy, and device health status of each node. Network state includes the inter-cluster communication latency matrix, bandwidth utilization, and link stability index. Task state includes the queue length, priority distribution, and average response time of model inference tasks. The action space includes expert module migration, cluster boundary adjustment, and communication path optimization. Expert module migration involves migrating specific expert modules in the hybrid expert model to high-computing-power nodes based on load conditions. Cluster boundary adjustment includes dynamically merging low-load clusters or splitting overloaded clusters. Communication path optimization involves selecting low-latency, high-bandwidth redundant links to transmit model parameters and inference results. The reward function outputs a load balancing score calculated based on the variance of computing power between clusters, a communication cost calculated by weighting the latency and bandwidth consumption of cross-cluster communication, and an exponential penalty term for the number of timed-out tasks, which are obtained by weighted summation using weighting coefficients. At the edge computing layer, the current environmental state is collected at a set period and input into the deep reinforcement learning model; the deep reinforcement learning model outputs the current optimal action command from the action space based on the input environmental state. Execute action instructions to dynamically adjust the distribution of expert modules, cluster boundaries, or communication paths, thereby achieving real-time optimization of the initial cluster partitioning and forming the final cluster partitioning at the current moment.

[0013] By using a deep reinforcement learning model to perceive network and device status in real time and dynamically adjust cluster boundaries, the system can adapt to complex scenarios such as load fluctuations, device failures, and network topology changes, overcoming the rigidity and lag of static partitioning algorithms. Through specific actions such as "expert module migration," "cluster boundary adjustment," and "communication path optimization," the system achieves on-demand and flexible allocation of resources such as computing power, storage, and bandwidth, significantly improving the overall efficiency and task response speed of the distributed computing framework.

[0014] Furthermore, at the edge computing layer, local control strategies are executed based on the final cluster partitioning. These local control strategies include voltage-reactive power coordinated control, frequency-active power balance control, and rapid fault isolation and self-healing, comprising the following steps: The voltage-reactive power coordinated control process specifically includes: Within the cluster, the voltage amplitude, photovoltaic output, load power, and topological connections of the nodes within the cluster are collected as model input data. The lightweight DeepSeek model deployed in the cluster is used to infer the input data, predict future voltage change trends, and generate the optimal reactive power regulation strategy suitable for the cluster. Control actions are executed, including adjusting the reactive power output of the photovoltaic inverters within the cluster, controlling the charging and discharging power of the energy storage converters within the cluster to participate in reactive power compensation, and switching the switching status of the capacitor banks within the cluster. The frequency active power balance control process specifically includes: Within the cluster, frequency deviation, spinning reserve capacity, and load forecasting error are collected as model input data. The lightweight DeepSeek model deployed in the cluster is used to infer the input data and calculate the required energy storage charging and discharging power, adjustable load reduction, and photovoltaic curtailment ratio for the cluster. Control actions are then executed, including triggering rapid frequency regulation of the energy storage within the cluster, adjusting the setpoint of the electric heating load within the cluster, and dynamically adjusting the output of the diesel generators within the cluster. The rapid fault isolation and self-healing control process specifically includes: Within the cluster, fault current waveforms, protection device action signals, and equipment health status are collected as model input data. The lightweight DeepSeek model deployed in the cluster is used to infer the input data, identify the fault type, locate the fault range, and generate the minimum power outage isolation scheme within the cluster. Control actions are executed, including tripping the circuit breakers upstream and downstream of the fault point within the cluster, and starting the backup power supply and tie switch within the cluster to restore power supply to the non-faulty area.

[0015] By implementing closed-loop control based on a lightweight model within the cluster, the response speed far exceeds that of traditional cloud-based centralized control. This effectively suppresses voltage exceedances and frequency fluctuations caused by a high proportion of distributed energy access, ensuring power quality. Fault handling tasks are delegated to the edge cluster, leveraging the rapid inference capabilities of the local model to locate and isolate faults in an extremely short time, restoring power to non-faulty areas. This minimizes the impact of faults and power outage time, significantly improving power supply reliability.

[0016] Furthermore, the model iterative update is performed between the cloud computing layer and the edge computing layer, specifically including the following steps: The cloud computing layer aggregates key feature data and runtime data uploaded from multiple edge computing layers to train and optimize the complete DeepSeek model; The cloud computing layer performs lightweight processing on the optimized model and generates an incremental update package for the model. The cloud computing layer distributes incremental model update packages to the edge computing layer to update the lightweight DeepSeek models deployed there.

[0017] By retraining models using data from the entire network in the cloud, more universal and superior strategies can be learned. Update packages are then distributed to ensure that all edge-side models evolve synchronously, thereby continuously improving the operational efficiency and control performance of the entire distribution network system over time. Through regular updates, the lightweight edge-side models can adapt to long-term dynamics such as changes in grid structure and load growth, avoiding performance degradation due to environmental changes and ensuring long-term operational reliability.

[0018] Furthermore, the edge computing layer is deployed on edge AI hardware devices that integrate neural network processing units (NPUs). The lightweight DeepSeek model is converted into a hardware-executable format through the NPU-specific inference framework; The edge computing layer utilizes the integer computing capabilities of the NPU to perform inference tasks for lightweight DeepSeek models.

[0019] By accelerating the model through dedicated NPU hardware and inference framework, the stringent millisecond-level response time requirements of power distribution network control tasks are met, which is impossible for general-purpose CPUs. Dedicated AI hardware, while providing high performance, typically features low power consumption, high integration, and industrial-grade reliability, making it ideal for deployment in space-constrained and harsh environments such as power distribution substations and outdoor ring main units.

[0020] Furthermore, the method also includes edge autonomy and cache synchronization mechanisms: When a communication interruption is detected between the edge computing layer and the cloud computing layer, the following steps are performed: The edge computing layer switches to autonomous mode, maintaining the independent operation of its managed clusters based on the locally deployed lightweight DeepSeek model and the last received global constraints. Key data and control instructions generated by the edge computing layer during local caching operation; Once communication is restored, perform the following steps: The edge computing layer synchronizes cached data to the cloud computing layer; The edge computing layer receives and applies update instructions and model parameters issued by the cloud computing layer, returning to the cloud-edge collaborative optimization mode.

[0021] In extreme cases of cloud-edge communication disruption, the system will not be paralyzed but will automatically degrade to edge autonomous mode, relying on local capabilities to maintain critical operations. This is a significant advantage that traditional centralized systems cannot match. Through a data caching mechanism, data from the period of interruption can be retransmitted to the cloud after communication is restored, ensuring the integrity of data center data and enabling the cloud to quickly understand the edge status, achieving a seamless and smooth switch from autonomous mode to collaborative mode.

[0022] In a second aspect, the present invention provides a local control device for a distribution network cluster, used to implement a local control method for a distribution network cluster. The device adopts a three-level collaborative architecture of cloud, edge, and terminal, and includes: The terminal equipment layer includes multiple sensors and actuators deployed at distribution network nodes to collect electrical parameters in real time and execute control commands. The edge computing layer, which communicates with the terminal device layer, includes: The graph modeling module is used to calculate the electrical distance between nodes based on electrical parameters and to construct a graph model of the distribution network. The dynamic optimization module is used to run a lightweight deep reinforcement learning model. It dynamically optimizes the initial cluster partitioning based on real-time network state and global constraints to obtain the final cluster partitioning. The local control module is used to execute local control strategies such as voltage and reactive power coordinated control, frequency and active power balance control, and rapid fault isolation and self-healing based on the final cluster division. The cloud computing layer, which communicates with the edge computing layer, includes: The global optimization module is used to construct a global optimization model with the goals of overall network voltage stability, minimizing system network losses, and improving fault recovery efficiency. The initial partitioning module is used to perform initial cluster partitioning of the graph model using a graph partitioning algorithm and to send the partitioning results to the edge computing layer. The model management module is used to distribute the parameters and global constraints of lightweight deep reinforcement learning models to the edge computing layer. Attached Figure Description

[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a local control method for a power distribution network cluster according to an embodiment of the present invention; Figure 2 This is a simplified structural diagram of a local control device for a power distribution network according to an embodiment of the present invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0025] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0026] Example 1 like Figure 1 As shown, a local control method for a power distribution network cluster adopts a three-level collaborative architecture of cloud, edge, and terminal, including: S1. Collect electrical parameters of multiple nodes in the power distribution network through the terminal equipment layer and send the electrical parameters to the edge computing layer; S2. In the edge computing layer, the electrical distance between nodes is calculated based on electrical parameters, and a graph model of the distribution network is constructed based on the electrical distance. In this model, nodes represent distribution network nodes, and edge weights are based on electrical distance. S3. At the cloud computing layer, a graph partitioning algorithm is used to perform initial cluster partitioning of the graph model, and the partitioning results are sent down to the edge computing layer. S4. At the edge computing layer, a lightweight deep reinforcement learning model is used to dynamically optimize the initial cluster partitioning based on the real-time network state and global constraints issued from the cloud computing layer, so as to obtain the final cluster partitioning. S5. At the edge computing layer, local control strategies are executed based on the final cluster partitioning. These local control strategies include voltage and reactive power coordinated control, frequency and active power balance control, and rapid fault isolation and self-healing.

[0027] This invention adopts a three-tiered collaborative architecture of cloud, edge, and terminal. The terminal device layer collects electrical parameters in real time and executes control commands through sensors and actuators. The edge computing layer deploys a lightweight DeepSeek model to dynamically update cluster boundaries, isolate faults, and extract data features. The cloud computing layer, based on the complete DeepSeek model, achieves long-term optimization through model iteration and cross-cluster coordination. The edge layer uploads key feature data such as cluster power variance to the cloud, while the cloud distributes lightweight model parameters and global constraints. This solution reduces cross-regional power transmission loss through dynamic cluster partitioning, and, combined with millisecond-level local response and global strategy coordination, significantly improves fault isolation speed and power supply reliability, optimizes resource utilization, and ensures topology adaptation capabilities in complex operating environments.

[0028] In one embodiment, electrical parameters of multiple nodes in the power distribution network are collected through the terminal device layer and sent to the edge computing layer. This includes the following steps: Real-time collection of electrical measurement data, environmental data, and equipment status data is achieved using multi-source sensors deployed at the power distribution network nodes. The electrical measurement data includes voltage, current, and power; the environmental data includes temperature and light intensity; and the equipment status data includes the state of charge of the energy storage system and the photovoltaic output status. The collected data is aggregated and cleaned by the edge computing nodes, and features are extracted to generate a standardized input vector. The standardized input vector is then sent to the edge computing layer.

[0029] Specifically, a multi-objective optimization model is established at the edge, focusing on maximizing cluster electrical density, optimizing power self-balancing, and minimizing control response delay. At the cloud, a global optimization model is constructed, aiming to optimize overall network voltage stability, minimize system network losses, and maximize fault recovery efficiency, while also considering global constraints such as branch power thermal stability limits, node voltage amplitude fluctuations, and cluster reserve capacity. Each edge cluster constructs an independent "sensing-decision-execution" closed-loop control system. Its multi-source sensing layer, deployed on terminal devices, uses sensors to collect real-time electrical parameters (voltage, current, power), environmental data (temperature, illumination), and equipment status (energy storage SOC, photovoltaic output) within the cluster. The data is aggregated, cleaned, and feature-extracted through edge computing nodes to generate standardized input vectors.

[0030] In one embodiment, at the edge computing layer, the electrical distance between nodes is calculated based on electrical parameters, and a graph model of the distribution network is constructed based on the electrical distance, where nodes represent distribution network nodes and edge weights are based on electrical distances; the steps include: calculating the active voltage sensitivity matrix and reactive voltage sensitivity matrix of the distribution network based on electrical parameters; calculating the electrical distance between any two nodes based on the active voltage sensitivity matrix and reactive voltage sensitivity matrix; defining the edge weight of the edge between node i and node j based on the electrical distance; and constructing a weighted graph model for cluster partitioning based on all distribution network nodes and the edges defined by the edge weights.

[0031] To reflect the role of cluster partitioning in distribution network systems, this invention defines edge weights as the electrical distance between distribution network nodes, which characterizes the degree of coupling between nodes in the distribution network system. This electrical distance is characterized by electrical distance indices derived from the active power voltage sensitivity matrix and the reactive power voltage sensitivity matrix. The results of cluster partitioning are measured and determined by evaluation metrics. Modularity is a commonly used metric in network community planning. The definition is as follows:

[0032] In the formula: For connecting nodes and nodes The weight of the edge. In a general network, when a node... and nodes When connected, It equals 1, otherwise it is 0; For nodes The sum of the weights of all connected edges. For nodes The sum of the weights of all connected edges; It is the sum of the edge weights of all edges. This represents the Kronecker delta function, when the node... and nodes The value is 1 if they belong to the same community, and 0 otherwise.

[0033] To reflect the role of cluster partitioning in the distribution network system, in this invention, edge weight is defined as the electrical distance between distribution network nodes. Electrical distance characterizes the degree of coupling between nodes in the distribution network system. It is characterized by electrical distance indices in the active power voltage sensitivity matrix and the reactive power voltage sensitivity matrix, defining a weight with... In a system of nodes, there are nodes. and nodes Active voltage sensitivity for:

[0034] In the formula: This is the active voltage sensitivity matrix; the first element in the matrix... OK Column elements Represents a node Injected unit active power and node The relationship between voltage amplitude changes. Similarly, the reactive voltage sensitivity can be obtained. :

[0035] In the formula: This is the reactive voltage sensitivity matrix; the first element in the matrix... OK Column elements Represents a node Injected unit reactive power and node The relationship between voltage amplitude changes.

[0036] Since changes in node voltage in a distribution network are related to changes in both active and reactive power, the electrical distance between nodes is defined in this invention as follows: The corresponding edge weight is defined as follows: , Let be the set of all electrical distances in the network. In this case, the larger the electrical distance between two nodes, the larger the edge weight.

[0037] In one embodiment, at the cloud computing layer, a graph partitioning algorithm is used to perform initial cluster partitioning of the graph model, and the partitioning results are sent to the edge computing layer. This includes the following steps: constructing a device topology graph based on a weighted graph model; wherein, distribution network nodes are abstracted as weighted nodes, and node weights include device computing power, storage capacity, and remaining power consumption; connections between nodes are abstracted as weighted edges, and edge weights are calculated using a composite function based on electrical distance and incorporating communication latency and bandwidth; a multi-objective optimization function for graph partitioning is constructed with minimizing the total communication cost between clusters as the primary objective and balancing the computing resources of each cluster as a collaborative objective; device computing power and storage capacity are injected as constraints into the graph partitioning algorithm, and the k-way partitioning function is called to partition the device topology graph, generating an initial cluster partitioning scheme; the initial cluster partitioning scheme is then sent to the edge computing layer.

[0038] When deploying large-scale MoE (Hybrid Expert) models such as DeepSeek in edge computing scenarios, the expert modules of the model need to be dynamically allocated to distributed device clusters to achieve load balancing and reduce communication overhead. Traditional cluster partitioning algorithms (such as K-means clustering) are usually geared towards static data center environments and are difficult to adapt to the heterogeneity of edge devices, dynamic network topology, and resource fluctuations. This invention proposes a hybrid dynamic cluster partitioning algorithm that combines classic graph partitioning algorithms with deep reinforcement learning to address the challenges posed by device heterogeneity, dynamic changes in network topology, and resource fluctuations in the dynamic deployment requirements of DeepSeek models in edge computing scenarios. This algorithm adopts a hierarchical architecture that combines static pre-partitioning with dynamic real-time adjustment, and consists of two key stages: 1) Static pre-partitioning stage: Based on classic graph partitioning algorithms (such as METIS), the initial device cluster is coarsely partitioned to build a basic load balancing framework. 2) Dynamic optimization stage: Through a deep reinforcement learning (DRL) model, the network state changes are perceived in real time, and the cluster partitioning strategy is dynamically adjusted to achieve adaptive resource allocation and task scheduling.

[0039] Specifically, the static pre-partitioning stage includes: Define input parameters, including the device topology graph and edge weights. The device topology graph abstracts edge devices as weighted nodes and communication links as weighted edges. Node weights are: device computing power (TOPS), storage capacity (GB), and remaining power consumption (W). Edge weights are: communication latency (ms) and bandwidth (Mbps).

[0040] Load requirements: computational cost (FLOPs / task) and memory usage (MB / task) of DeepSeek model inference tasks. Real-time task queue length and priority constraints.

[0041] Graph structure modeling involves constructing an adjacency matrix between devices and defining the communication cost between nodes. For delay With bandwidth Composite functions:

[0042] in, α , β These are weighting coefficients, which are dynamically adjusted based on the actual network environment.

[0043] The objective function is to minimize the total cost of inter-cluster communication while balancing the computing resources of each cluster.

[0044] in, Indicates a cross-cluster communication link. gamma is the balancing coefficient, and Var is the variance of the cluster's computing power.

[0045] The METIS partitioning process includes graph partitioning initialization, constraint injection, and partitioning result generation. Graph partitioning initialization involves calling the k-way partitioning function of the METIS library and specifying the number of clusters (e.g., partitioning by physical regions). Constraint injection sets upper and lower limits for device computing power and storage to ensure that the resource capacity of a single cluster meets task requirements. Partitioning result generation outputs the initial cluster partitioning scheme, serving as the basic framework for dynamic optimization.

[0046] Specifically, deep reinforcement learning models in the dynamic optimization phase include: The state space includes: device status (real-time computing power utilization, storage utilization, and device health status (online / offline) for each node); network status (inter-cluster communication latency matrix, bandwidth utilization, and link stability index); and task status (queue length, priority distribution, and average response time for DeepSeek model inference tasks).

[0047] The Action Space includes: Expert Module Migration, which migrates specific expert modules from DeepSeek's MoE (Hybrid Expert) model to high-performance nodes based on load conditions; Cluster Boundary Adjustment, which dynamically merges low-load clusters or splits overloaded clusters, for example, moving 50% of the nodes from one cluster to an adjacent cluster; and Communication Path Optimization, which selects low-latency, high-bandwidth redundant links to transmit model parameters and inference results.

[0048] The reward function includes:

[0049] Where: Load_Balance_Score is calculated based on the variance of computing power between clusters, with a higher score for smaller variance. Communication_Cost is the weighted sum of latency and bandwidth consumption for cross-cluster communication. Task_Timeout_Penalty is an exponential penalty term for the number of timeout tasks. w1, w3, and w2 are weight coefficients, determined through offline training optimization.

[0050] The training and deployment process is as follows: Digital twin environment training includes: building a digital twin platform for the power distribution network in the cloud to simulate various edge scenarios such as equipment failure, load surges, and network jitter. The PPO (Proximal Policy Optimization) algorithm is used to train the DRL (Deep Reinforcement Learning) policy network until it converges to a stable policy.

[0051] Model lightweighting and edge deployment include: quantizing and compressing the trained DRL model (INT8) and pruning it to adapt to the computing resources of the edge controller. The lightweight model is then deployed to the edge layer master node to receive sensor data in real time and output optimized actions.

[0052] The online inference and feedback loop includes: the edge controller collects the environmental state every 100ms and inputs it into the DRL model to generate action commands. Based on the action execution results (such as task latency and device load), the reward function parameters are dynamically updated to achieve model self-optimization.

[0053] Large-scale artificial intelligence (AI) models such as DeepSeek-V3 and DeepSeek-R1 have demonstrated powerful inference capabilities in the cloud, but running these models on edge devices or AI hardware still faces numerous challenges. High computational resource requirements: DeepSeek-R1, as a large model optimized based on reinforcement learning, possesses powerful inference capabilities, but its computational demands are enormous, typically requiring high-end GPUs or TPUs. DeepSeek-V3 activates up to 37 billion parameters per inference iteration, placing extremely high demands on computing power, making it difficult to run directly on ordinary edge AI devices. Limited storage space: Large models typically require tens or even hundreds of gigabytes of storage, while edge devices have extremely limited storage resources. For example, common AIoT devices are usually equipped with only 2GB-8GB of RAM, insufficient to accommodate the parameters and intermediate computation results of large-scale models. Power consumption limitations: Edge AI devices (such as Rockchip RK3588 and NVIDIA Jetson Orin) typically operate in low-power environments, with power budgets far lower than cloud servers. Directly executing large-scale model inference tasks would cause a sharp increase in device power consumption, affecting the device's battery life and stability.

[0054] To adapt large-scale AI models such as DeepSeek to edge devices, a series of model optimization techniques are needed to reduce computational requirements, reduce storage footprint, and optimize power efficiency.

[0055] 1. Model Quantization Model quantization is an effective technique that can significantly reduce inference computation requirements and storage footprint. DeepSeek uses INT8 (8-bit integer) and FP16 (16-bit floating-point) quantization techniques to convert 32-bit floating-point calculations into 8-bit integer or 16-bit floating-point calculations, thereby greatly reducing computational complexity and storage space.

[0056] INT8 quantization converts 32-bit floating-point weights and activation values ​​into 8-bit integers, reducing computation to 1 / 4 of the original while maintaining almost no loss of precision, and reducing storage requirements to 1 / 4 of the original.

[0057] FP16 quantization: While maintaining high precision, it converts 32-bit floating-point calculations into 16-bit floating-point calculations, which is suitable for scenarios with high precision requirements.

[0058] Table 1 Comparison of Model Quantization Methods

[0059] Furthermore, DeepSeek can leverage inference acceleration frameworks such as NVIDIA TensorRT or RKNN (Rockchip NPU runtime library) to further optimize the inference efficiency of quantized models. These frameworks are deeply optimized for specific hardware platforms, significantly improving inference speed and reducing power consumption.

[0060] 2. Model Pruning Model pruning reduces the number of parameters and computational cost by removing redundant parameters or neurons. DeepSeek can employ a combination of structured and unstructured pruning to significantly reduce computational complexity and storage requirements while maintaining model performance. Structured pruning involves removing entire neurons or convolutional kernels, suitable for hardware acceleration. Unstructured pruning involves removing individual weights, suitable for further model compression.

[0061] 3. Knowledge Distillation Knowledge distillation enables efficient inference on edge devices by transferring knowledge from large-scale models to small models. DeepSeek achieves high-performance inference on edge devices by training a small student model to mimic the behavior of a large-scale teacher model.

[0062] 4. Dynamic Computation DeepSeek's MoE (Mixture of Experts) architecture supports dynamic computation, meaning that only some expert modules are activated to complete specific tasks. This mechanism can significantly reduce computational load, making it particularly suitable for resource-constrained edge devices. By optimizing the activation strategy of expert modules, computational complexity and power consumption can be further reduced.

[0063] 5. Hardware acceleration and collaborative optimization Edge AI devices are typically equipped with dedicated AI accelerators (such as NPUs and TPUs), and DeepSeek can fully leverage the computing power of these accelerators through co-optimization with the hardware platform. For example, utilizing NVIDIA Jetson Orin's Tensor Cores or Rockchip's RK3588 NPU can significantly improve inference efficiency and reduce power consumption.

[0064] Table 2 AI Hardware Performance Comparison

[0065] In one embodiment, at the edge computing layer, a lightweight deep reinforcement learning model is used to dynamically optimize the initial cluster partitioning based on real-time network state and global constraints issued from the cloud computing layer, resulting in the final cluster partitioning. This includes the following steps: Constructing a deep reinforcement learning model includes: a state space, an action space, and a reward function; among which, The state space includes device state, network state, and task state. Device state includes the real-time computing power utilization, storage occupancy, and device health status of each node. Network state includes the inter-cluster communication latency matrix, bandwidth utilization, and link stability index. Task state includes the queue length, priority distribution, and average response time of model inference tasks. The action space includes expert module migration, cluster boundary adjustment, and communication path optimization. Expert module migration involves migrating specific expert modules in the hybrid expert model to high-computing-power nodes based on load conditions. Cluster boundary adjustment includes dynamically merging low-load clusters or splitting overloaded clusters. Communication path optimization involves selecting low-latency, high-bandwidth redundant links to transmit model parameters and inference results. The reward function outputs a load balancing score calculated based on the variance of computing power between clusters, a communication cost calculated by weighting the latency and bandwidth consumption of cross-cluster communication, and an exponential penalty term for the number of timed-out tasks, which are obtained by weighted summation using weighting coefficients. At the edge computing layer, the current environmental state is collected at a set period and input into the deep reinforcement learning model; the deep reinforcement learning model outputs the current optimal action command from the action space based on the input environmental state. Execute action instructions to dynamically adjust the distribution of expert modules, cluster boundaries, or communication paths, thereby achieving real-time optimization of the initial cluster partitioning and forming the final cluster partitioning at the current moment.

[0066] Specifically, the local control method proposed in this invention, based on the edge computing cluster partitioning results and combined with the real-time inference capability of the DeepSeek lightweight model, realizes the distributed autonomous operation of the power distribution network cluster. Through localized decision-making and control, communication latency is significantly reduced, and system response speed and disturbance resistance are improved. The specific technical solution is as follows: The local control architecture includes: each edge cluster constructs an independent "perception-decision-execution" closed-loop control system, comprising the following core modules: The multi-source sensing layer consists of sensors deployed on terminal devices that collect real-time electrical parameters (voltage, current, power), environmental data (temperature, illumination), and equipment status (energy storage SOC, photovoltaic output) within the cluster. The data is aggregated, cleaned, and feature-extracted through edge computing nodes to generate standardized input vectors.

[0067] The local decision-making layer runs a lightweight DeepSeek model (such as DeepSeek-Ctrl-Lite) to predict the cluster's operating status based on real-time data and generate control commands. It supports multi-objective optimization decisions, including voltage / frequency regulation, power balancing, and fault isolation.

[0068] The execution control layer sends control commands to distributed energy sources (photovoltaic inverters, energy storage converters PCS), smart switches (interlocking switches, circuit breakers, etc.), and electric heating load controllers. This achieves millisecond-level closed-loop control, ensuring dynamic stability within the cluster.

[0069] Key control strategies include: 1. Voltage and reactive power coordinated control High-proportion distributed photovoltaic (PV) grid integration leads to increased voltage fluctuations in the distribution network, and traditional centralized AVC (Automatic Voltage Control) systems are insufficient in response speed. Model input data includes node voltage amplitudes, PV output, load power, and topology connections within the cluster.

[0070] Real-time reasoning: Predicts voltage change trends over the next 5 minutes and generates the optimal reactive power regulation strategy. Control actions: Adjusts the reactive power output of the photovoltaic inverter (QV droop control); controls the charging and discharging power of the energy storage converter PCS (participates in reactive power compensation); switches the switching status of the capacitor bank (discrete control).

[0071] 2. Frequency active power balance control In islanded operation or weak grid scenarios, load fluctuations can easily cause frequency deviations. In this case, the model input data includes cluster frequency deviation, spinning reserve capacity, and load forecasting error.

[0072] Real-time inference: Calculates energy storage charging and discharging power, adjustable load reduction, and photovoltaic curtailment ratio. Control actions: Triggers energy storage fast frequency regulation mode (response time <100ms). Adjusts the electric heating load temperature setpoint (flexible load participates in frequency regulation). Dynamically adjusts diesel generator output (backup power supply coordination control).

[0073] 3. Rapid fault isolation and self-healing In traditional distribution networks, fault location relies primarily on centralized fault location at the main substation, resulting in processing delays of several minutes. The model input data in this case includes fault current waveforms, protection device action signals, and equipment health status.

[0074] Real-time reasoning: Identifies fault types (short circuit, grounding, etc.) based on waveform characteristics; locates fault zones (accuracy up to 99%); generates a minimum power outage isolation scheme. Control actions: trips upstream and downstream circuit breakers at the fault point (action time <20ms); activates backup power and interconnection switches to restore power to non-faulty areas; issues inspection commands to drones / robots to pinpoint the fault location.

[0075] In one embodiment, at the edge computing layer, a local control strategy is executed based on the final cluster partitioning. The local control strategy includes voltage and reactive power coordinated control, frequency active power balance control, and rapid fault isolation and self-healing, including the following steps: The voltage-reactive power coordinated control process specifically includes: Within the cluster, the voltage amplitude, photovoltaic output, load power, and topological connections of the nodes within the cluster are collected as model input data. The lightweight DeepSeek model deployed in the cluster is used to infer the input data, predict future voltage change trends, and generate the optimal reactive power regulation strategy suitable for the cluster. Control actions are executed, including adjusting the reactive power output of the photovoltaic inverters within the cluster, controlling the charging and discharging power of the energy storage converters within the cluster to participate in reactive power compensation, and switching the switching status of the capacitor banks within the cluster. The frequency active power balance control process specifically includes: Within the cluster, frequency deviation, spinning reserve capacity, and load forecasting error are collected as model input data. The lightweight DeepSeek model deployed in the cluster is used to infer the input data and calculate the required energy storage charging and discharging power, adjustable load reduction, and photovoltaic curtailment ratio for the cluster. Control actions are then executed, including triggering rapid frequency regulation of the energy storage within the cluster, adjusting the setpoint of the electric heating load within the cluster, and dynamically adjusting the output of the diesel generators within the cluster. The rapid fault isolation and self-healing control process specifically includes: Within the cluster, fault current waveforms, protection device action signals, and equipment health status are collected as model input data. The lightweight DeepSeek model deployed in the cluster is used to infer the input data, identify the fault type, locate the fault range, and generate the minimum power outage isolation scheme within the cluster. Control actions are executed, including tripping the circuit breakers upstream and downstream of the fault point within the cluster, and starting the backup power supply and tie switch within the cluster to restore power supply to the non-faulty area.

[0076] In a preferred embodiment, the model iterative update is performed between the cloud computing layer and the edge computing layer, specifically including the following steps: The cloud computing layer aggregates key feature data and runtime data uploaded from multiple edge computing layers to train and optimize the complete DeepSeek model. The cloud computing layer then performs lightweight processing on the optimized model to generate an incremental update package. Finally, the cloud computing layer distributes the incremental update package to the edge computing layer to update the lightweight DeepSeek model deployed therein.

[0077] Specifically, to address complex faults or cross-cluster coordination needs, this invention designs a hierarchical collaborative control mechanism: In normal operation, each cluster relies entirely on its local DeepSeek model for closed-loop control, without the need for cloud intervention. Edge nodes periodically upload critical operational data (such as control logs and model performance metrics) to the cloud.

[0078] In an emergency takeover, when a cross-cluster fault is detected (such as a multi-cluster voltage collapse), the cloud initiates a global coordination mode: global constraints (such as tie-line power limits) are issued to the edge layer. Critical resources are reallocated (such as calling remote energy storage to support the faulty cluster).

[0079] Model iteration involves aggregating data from multiple clusters in the cloud to train a globally optimized model and generate incremental update packages. Edge nodes download these update packages during idle periods, enabling online model upgrades.

[0080] To address the need to balance complex computing tasks with resource constraints, this invention constructs a hierarchical computing architecture: The edge device performs high-frequency sampling of the voltage / current waveforms acquired by the terminal equipment. Fundamental components, harmonic distortion rates, and transient characteristics are extracted using wavelet transform and Fast Fourier Transform (FFT) analysis to generate a standardized electrical feature vector. This process utilizes the parallel computing capabilities of the NPU, achieving millisecond-level real-time control requirements with a single processing latency.

[0081] On the edge side, a lightweight DeepSeek model (optimized with INT8 quantization and pruning) analyzes feature data in real time, combining topological connectivity and power balance status to update the cluster boundary partitioning scheme once according to the set accuracy. For example, when a sudden increase in photovoltaic output in a certain area is detected, causing the local load to be unable to absorb it, the model dynamically adjusts the cluster range, prioritizing the inclusion of energy storage nodes in that area to balance the power.

[0082] Based on historical operational data and weather forecasts, the cloud-based system uses a mixed-integer nonlinear programming (MINLP) algorithm to generate a 24-hour power flow distribution forecast for the entire network, which is then distributed to the edge layer as a baseline strategy for cluster partitioning. For example, when predicting the peak photovoltaic output at noon the following day, the power limits of inter-cluster interconnects are optimized in advance to reduce network losses.

[0083] The cloud aggregates operational data uploaded from multiple edge nodes (such as fault event logs and control effect feedback) to perform full parameter fine-tuning and incremental training on the original DeepSeek model. After training, the model is anonymized using differential privacy technology, compressed into a lightweight version, and distributed to edge devices, enabling continuous evolution of model capabilities.

[0084] In a preferred embodiment, the edge computing layer is deployed on an edge AI hardware device that integrates a neural network processing unit (NPU); the lightweight DeepSeek model is converted into a hardware-executable format through the NPU's dedicated inference framework; the edge computing layer utilizes the NPU's integer computing capabilities to perform inference tasks on the lightweight DeepSeek model.

[0085] This invention selects the Rockchip RK3588 chip as the core AI processor on the edge side. Its technical characteristics are highly compatible with the requirements of power distribution network cluster partitioning and control. Specific advantages are as follows: It adopts a dedicated Neural Processing Unit (NPU) architecture, integrating a high-performance NPU module that supports INT8 quantization computation and achieves a peak computing power of 6 TOPS, meeting the millisecond-level inference requirements of DeepSeek's lightweight models. It also supports parallel computing of multiple models (such as simultaneously running cluster partitioning models and voltage control models), enabling multi-task collaborative processing. It features a deeply adapted AI development framework, with a built-in RKNN (Rockchip Neural Network) inference framework, supporting the direct conversion of DeepSeek's ONNX models into NPU executable formats, improving inference efficiency through graph optimization, operator fusion, and other techniques. It also provides a cross-platform deployment toolchain, compatible with mainstream deep learning frameworks such as TensorFlow and PyTorch. The chip features a low-power industrial-grade design, with typical power consumption below 5W, supporting operation in wide-temperature (-40℃~85℃) and high-humidity environments, meeting the deployment requirements of demanding scenarios such as power distribution substations and outdoor ring main units. It also provides a hardware-level security encryption module to ensure the security of power distribution network control commands and data communication.

[0086] DeepSeek model deployment process: Leveraging the AI ​​inference acceleration capabilities of the RK3588 chip, this invention enables efficient edge deployment of DeepSeek models. The specific process includes: 1) Model lightweighting and format conversion: The original DeepSeek model trained in the cloud (such as a cluster partitioning model based on the Transformer architecture) is subjected to INT8 quantization and pruning, compressing the model size to less than 1 / 4 of the original size. The optimized model is then converted to the RK3588 specific format (.rknn) using the RKNN-Toolkit2 toolchain, and NPU instruction set optimization parameters are injected.

[0087] 2) Hardware acceleration configuration: By leveraging the NPU's independent memory allocation mechanism, model weights and intermediate computation results are directly mapped to the NPU's dedicated storage area, reducing CPU-GPU data transfer latency. A mixed-precision computing mode (FP16 + INT8) is enabled to balance computational accuracy and energy efficiency requirements.

[0088] 3) Real-time inference optimization: Deploy a multi-threaded inference engine to achieve pipelined parallel processing of data acquisition, preprocessing, and model inference, with a single inference cycle controlled within 20ms. Dynamically allocate NPU computing resources to prioritize the real-time requirements of critical tasks (such as fault detection).

[0089] In a preferred embodiment, the method further includes an edge autonomy and cache synchronization mechanism: When a communication interruption between the edge computing layer and the cloud computing layer is detected, the following steps are performed: the edge computing layer switches to autonomous mode, maintaining the independent operation of its managed cluster based on the locally deployed lightweight DeepSeek model and the last received global constraints; the edge computing layer caches key data and control instructions generated during operation locally. Once communication is restored, the following steps are performed: the edge computing layer synchronizes the cached data to the cloud computing layer; the edge computing layer receives and applies the update instructions and model parameters issued by the cloud computing layer, and returns to the cloud-edge collaborative optimization mode.

[0090] Model application examples: The cloud-edge collaborative dynamic cluster partitioning and optimization control model proposed in this invention achieves dynamic optimization of distribution network cluster partitioning and precise execution of local control through real-time data processing at the edge and global strategy collaboration at the cloud. This model aims to solve problems such as slow response, low resource utilization efficiency, and large fault impact range in traditional cluster partitioning methods, and improves system flexibility and robustness through a hierarchical optimization mechanism. Specific implementation scenarios and technical advantages are as follows: Application Scenario Example 1: When severe weather events, such as a strong typhoon, cause communication disruptions to 10 edge nodes in a coastal area, the affected edge nodes automatically switch to autonomous mode, maintaining a minimal cluster operation based on historical data and local models. The cloud re-plans the cluster partitioning of the remaining healthy nodes and coordinates energy storage in unaffected areas to provide power support. After the typhoon passes, the cloud issues a global recovery command, and the edge nodes synchronize their cluster status and return to normal optimization.

[0091] Application Scenario Example 2: When midday photovoltaic (PV) output fluctuations cause local voltage exceedances, the edge model detects that a cluster's voltage has exceeded the limit. It immediately splits the cluster into two sub-clusters and adjusts the reactive power output of the PV inverters. Simultaneously, the cloud updates the network-wide optimization constraints, limiting the power transmission from adjacent clusters to the faulty area. After PV output returns to normal, the edge layer automatically merges the sub-clusters and restores the original interconnection strategy.

[0092] The present invention proposes a method for rapid partitioning and local control of distribution network clusters based on edge computing, which has the following advantages: 1) Layered optimization improves system reliability Through cloud-based global optimization, this invention generates a cluster partitioning baseline that balances economy and security based on hybrid integer programming and deep learning algorithms, reducing the risk of network-wide operation (such as distributing vulnerable nodes to different clusters).

[0093] By using real-time edge correction, this invention dynamically corrects partitioning bias through a lightweight DeepSeek model (inference latency <50ms), which can significantly reduce local power imbalance.

[0094] 2) Resource efficiency and response speed optimization This invention employs a hierarchical computing load strategy, with non-real-time tasks (such as long-term topology optimization and model training) processed in the cloud, with computation cycles on the order of hours. The edge layer performs millisecond-level control (such as fault isolation and voltage regulation), which can significantly reduce the amount of data communication in the cloud.

[0095] This invention employs a differential data transmission mechanism to compress communication overhead, uploading only key feature vectors (such as cluster power variance and node voltage extreme values) to reduce bandwidth usage.

[0096] 3) Adaptive to complex operating environments If the actual framework supports real-time changes in the distribution network topology in the future (such as adding photovoltaic power plants or energy storage nodes), the model can automatically adjust the cluster partitioning logic without manual reconfiguration, thus improving dynamic topology compatibility.

[0097] This invention incorporates resources such as photovoltaics, energy storage, and flexible loads into cluster partitioning decisions, optimizes local energy consumption and supply-demand balance, and improves dynamic topology compatibility.

[0098] In summary, the above scheme provides a local control method for distribution network clusters, aiming to achieve distributed autonomous operation of the distribution network and improve the system's flexibility, reliability, and disaster resistance. By introducing edge computing technology, it effectively solves problems such as data transmission delays, high computational resource consumption, and expanded impact range of local faults in traditional centralized control modes, achieving more refined and efficient distribution network management.

[0099] Example 2 like Figure 2 As shown, based on the same inventive concept as the above embodiments, the present invention also provides a distribution network cluster local control device for the above-described distribution network cluster local control method. The device adopts a cloud-edge-device three-level collaborative architecture and includes: The terminal equipment layer includes multiple sensors and actuators deployed at distribution network nodes to collect electrical parameters in real time and execute control commands. The edge computing layer, which communicates with the terminal device layer, includes: The graph modeling module is used to calculate the electrical distance between nodes based on electrical parameters and to construct a graph model of the distribution network. The dynamic optimization module is used to run a lightweight deep reinforcement learning model. It dynamically optimizes the initial cluster partitioning based on real-time network state and global constraints to obtain the final cluster partitioning. The local control module is used to execute local control strategies such as voltage and reactive power coordinated control, frequency and active power balance control, and rapid fault isolation and self-healing based on the final cluster division. The cloud computing layer, which communicates with the edge computing layer, includes: The global optimization module is used to construct a global optimization model with the goals of overall network voltage stability, minimizing system network losses, and improving fault recovery efficiency. The initial partitioning module is used to perform initial cluster partitioning of the graph model using a graph partitioning algorithm and to send the partitioning results to the edge computing layer. The model management module is used to distribute the parameters and global constraints of lightweight deep reinforcement learning models to the edge computing layer.

[0100] Example 3 The electronic device of this embodiment may include: at least one processor, a memory, and a computer program stored in the memory and capable of running on at least one processor.

[0101] When a processor executes a computer program, it implements the steps in the above method embodiments, including: Electrical parameters of multiple nodes in the power distribution network are collected by the terminal device layer and sent to the edge computing layer. In the edge computing layer, the electrical distance between nodes is calculated based on electrical parameters, and a graph model of the distribution network is constructed based on the electrical distance, where nodes represent distribution network nodes and edge weights are based on electrical distance; At the cloud computing layer, a graph partitioning algorithm is used to initially partition the graph model into clusters, and the partitioning results are sent down to the edge computing layer. At the edge computing layer, a lightweight deep reinforcement learning model is used to dynamically optimize the initial cluster partitioning based on real-time network state and global constraints issued from the cloud computing layer, resulting in the final cluster partitioning. At the edge computing layer, local control strategies are executed based on the final cluster partitioning. These local control strategies include voltage and reactive power coordinated control, frequency and active power balance control, and rapid fault isolation and self-healing.

[0102] Optionally, the electronic device can be a server, industrial control computer, edge computing gateway, or any other form of computing device with data processing capabilities. When deployed in a cloud computing center, the electronic device acts as the cloud computing layer, performing global optimization, initial partitioning, and model management functions. When deployed in field locations such as substations and switch stations, the electronic device acts as the edge computing layer, performing graph modeling, dynamic optimization, and local control functions.

[0103] Memory can be an internal storage unit of an electronic device, such as the hard drive or RAM of the electronic device. Memory can also be an external storage device of an electronic device, such as a plug-in hard drive, smart memory card, flash memory card, etc., that is equipped with on the electronic device.

[0104] Example 4 This embodiment provides a computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the steps of a local control method for a power distribution network cluster.

[0105] Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections having one or more wires, portable computer disks, hard disks, read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0106] When at least one instruction is executed by the processor, it enables the processor to: Coordinate the data flow and control flow of the cloud-edge-device three-tier architecture; Implement graph model construction and dynamic cluster partitioning based on electrical distance at the edge; Perform real-time optimization based on a lightweight deep reinforcement learning model; Local control strategies such as voltage-reactive power coordination, frequency-active power balance, and rapid fault isolation and self-healing can be implemented.

[0107] Those skilled in the art will understand that the division of modules in the above embodiments does not constitute a limitation on the structure of electronic devices and storage media, and modules can be added, deleted, or combined according to business needs. The functional modules in the above embodiments can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0108] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for local control of a power distribution network cluster, characterized in that, It adopts a three-tier collaborative architecture of cloud, edge, and device, including: Electrical parameters of multiple nodes in the power distribution network are collected by the terminal device layer and sent to the edge computing layer. In the edge computing layer, the electrical distance between nodes is calculated based on electrical parameters, and a graph model of the distribution network is constructed based on the electrical distance, where nodes represent distribution network nodes and edge weights are based on electrical distance; At the cloud computing layer, a graph partitioning algorithm is used to initially partition the graph model into clusters, and the partitioning results are sent down to the edge computing layer. At the edge computing layer, a lightweight deep reinforcement learning model is used to dynamically optimize the initial cluster partitioning based on real-time network state and global constraints issued from the cloud computing layer, resulting in the final cluster partitioning. At the edge computing layer, local control strategies are executed based on the final cluster partitioning. These local control strategies include voltage and reactive power coordinated control, frequency and active power balance control, and rapid fault isolation and self-healing.

2. The local control method for a power distribution network cluster according to claim 1, characterized in that, The process involves collecting electrical parameters from multiple nodes in the power distribution network at the terminal device layer and then sending these parameters to the edge computing layer, including the following steps: Multi-source sensors deployed at distribution network nodes are used to collect electrical measurement data, environmental data, and equipment status data in real time. The electrical measurement data includes voltage, current, and power; the environmental data includes temperature and light intensity; and the equipment status data includes the state of charge of the energy storage system and the photovoltaic output status. The collected data is aggregated and cleaned by edge computing nodes, and features are extracted to generate standardized input vectors. The standardized input vector is sent to the edge computing layer.

3. The local control method for a power distribution network cluster according to claim 2, characterized in that, In the edge computing layer, the electrical distance between nodes is calculated based on electrical parameters, and a graph model of the distribution network is constructed based on the electrical distance. Nodes represent distribution network nodes, and edge weights are based on the electrical distance. This process includes the following steps: Based on electrical parameters, the active voltage sensitivity matrix and reactive voltage sensitivity matrix of the distribution network are calculated. Calculate the electrical distance between any two nodes based on the active voltage sensitivity matrix and the reactive voltage sensitivity matrix. The edge weight between node i and node j is defined based on the electrical distance. Based on all distribution network nodes and edges defined by edge weights, a weighted graph model for cluster partitioning is constructed.

4. The local control method for a power distribution network cluster according to claim 3, characterized in that, At the cloud computing layer, a graph partitioning algorithm is used to initially partition the graph model into clusters, and the partitioning results are then sent to the edge computing layer, including the following steps: A device topology graph is constructed based on a weighted graph model. In this model, distribution network nodes are abstracted as weighted nodes, and the node weights include device computing power, storage capacity, and remaining power consumption. The connections between nodes are abstracted as weighted edges, and the edge weights are calculated using a composite function based on electrical distance and incorporating communication delay and bandwidth. With minimizing the total cost of inter-cluster communication as the primary objective and balancing the computing resources of each cluster as a collaborative objective, a multi-objective optimization function for graph partitioning is constructed. The computing power and storage capacity of the devices are injected as constraints into the graph partitioning algorithm, and the k-way partitioning function is called to partition the device topology graph and generate an initial cluster partitioning scheme. The initial cluster partitioning scheme is distributed to the edge computing layer.

5. The local control method for a power distribution network cluster according to claim 4, characterized in that, At the edge computing layer, a lightweight deep reinforcement learning model is used to dynamically optimize the initial cluster partitioning based on real-time network state and global constraints issued from the cloud computing layer, resulting in the final cluster partitioning. This includes the following steps: Constructing a deep reinforcement learning model includes: a state space, an action space, and a reward function; among which, The state space includes device state, network state, and task state. Device state includes the real-time computing power utilization, storage occupancy, and device health status of each node. Network state includes the inter-cluster communication latency matrix, bandwidth utilization, and link stability index. Task state includes the queue length, priority distribution, and average response time of model inference tasks. The action space includes expert module migration, cluster boundary adjustment, and communication path optimization. Expert module migration involves migrating specific expert modules in the hybrid expert model to high-computing-power nodes based on load conditions. Cluster boundary adjustment includes dynamically merging low-load clusters or splitting overloaded clusters. Communication path optimization involves selecting low-latency, high-bandwidth redundant links to transmit model parameters and inference results. The reward function outputs a load balancing score calculated based on the variance of computing power between clusters, a communication cost calculated by weighting the latency and bandwidth consumption of cross-cluster communication, and an exponential penalty term for the number of timed-out tasks, which are obtained by weighted summation using weighting coefficients. At the edge computing layer, the current environmental state is collected at a set period and input into the deep reinforcement learning model; the deep reinforcement learning model outputs the current optimal action command from the action space based on the input environmental state. Execute action instructions to dynamically adjust the distribution of expert modules, cluster boundaries, or communication paths, thereby achieving real-time optimization of the initial cluster partitioning and forming the final cluster partitioning at the current moment.

6. The local control method for a power distribution network cluster according to claim 5, characterized in that, At the edge computing layer, local control strategies are executed based on the final cluster partitioning. These local control strategies include voltage-reactive power coordinated control, frequency-active power balance control, and rapid fault isolation and self-healing, comprising the following steps: The voltage-reactive power coordinated control process specifically includes: Within the cluster, the voltage amplitude, photovoltaic output, load power, and topological connections of the nodes within the cluster are collected as model input data. The lightweight DeepSeek model deployed in the cluster is used to infer the input data, predict future voltage change trends, and generate the optimal reactive power regulation strategy suitable for the cluster. Control actions are executed, including adjusting the reactive power output of the photovoltaic inverters within the cluster, controlling the charging and discharging power of the energy storage converters within the cluster to participate in reactive power compensation, and switching the switching status of the capacitor banks within the cluster. The frequency active power balance control process specifically includes: Within the cluster, frequency deviation, spinning reserve capacity, and load forecasting error are collected as model input data. The lightweight DeepSeek model deployed in the cluster is used to infer the input data and calculate the required energy storage charging and discharging power, adjustable load reduction, and photovoltaic curtailment ratio for the cluster. Control actions are then executed, including triggering rapid frequency regulation of the energy storage within the cluster, adjusting the setpoint of the electric heating load within the cluster, and dynamically adjusting the output of the diesel generators within the cluster. The rapid fault isolation and self-healing control process specifically includes: Within the cluster, fault current waveforms, protection device action signals, and equipment health status are collected as model input data. The lightweight DeepSeek model deployed in the cluster is used to infer the input data, identify the fault type, locate the fault range, and generate the minimum power outage isolation scheme within the cluster. Control actions are executed, including tripping the circuit breakers upstream and downstream of the fault point within the cluster, and starting the backup power supply and tie switch within the cluster to restore power supply to the non-faulty area.

7. The local control method for a power distribution network cluster according to claim 6, characterized in that, The iterative update of the model between the cloud computing layer and the edge computing layer includes the following steps: The cloud computing layer aggregates key feature data and runtime data uploaded from multiple edge computing layers to train and optimize the complete DeepSeek model; The cloud computing layer performs lightweight processing on the optimized model and generates an incremental update package for the model. The cloud computing layer distributes incremental model update packages to the edge computing layer to update the lightweight DeepSeek models deployed there.

8. The local control method for a power distribution network cluster according to claim 6, characterized in that, The edge computing layer is deployed on edge AI hardware devices that integrate neural network processing units (NPUs); The lightweight DeepSeek model is converted into a hardware-executable format through the NPU-specific inference framework; The edge computing layer utilizes the integer computing capabilities of the NPU to perform inference tasks for lightweight DeepSeek models.

9. The local control method for a power distribution network cluster according to claim 1, characterized in that, The method also includes edge autonomy and cache synchronization mechanisms: When a communication interruption is detected between the edge computing layer and the cloud computing layer, the following steps are performed: The edge computing layer switches to autonomous mode, maintaining the independent operation of its managed clusters based on the locally deployed lightweight DeepSeek model and the last received global constraints. Key data and control instructions generated by the edge computing layer during local caching operation; Once communication is restored, perform the following steps: The edge computing layer synchronizes cached data to the cloud computing layer; The edge computing layer receives and applies update instructions and model parameters issued by the cloud computing layer, returning to the cloud-edge collaborative optimization mode.

10. A local control device for a distribution network cluster, used to implement the local control method for a distribution network cluster as described in claim 1, characterized in that, The device adopts a three-tier collaborative architecture of cloud, edge, and device, including: The terminal equipment layer includes multiple sensors and actuators deployed at distribution network nodes to collect electrical parameters in real time and execute control commands. The edge computing layer, which communicates with the terminal device layer, includes: The graph modeling module is used to calculate the electrical distance between nodes based on electrical parameters and to construct a graph model of the distribution network. The dynamic optimization module is used to run a lightweight deep reinforcement learning model. It dynamically optimizes the initial cluster partitioning based on real-time network state and global constraints to obtain the final cluster partitioning. The local control module is used to execute local control strategies such as voltage and reactive power coordinated control, frequency and active power balance control, and rapid fault isolation and self-healing based on the final cluster division. The cloud computing layer, which communicates with the edge computing layer, includes: The global optimization module is used to construct a global optimization model with the goals of overall network voltage stability, minimizing system network losses, and improving fault recovery efficiency. The initial partitioning module is used to perform initial cluster partitioning of the graph model using a graph partitioning algorithm and to send the partitioning results to the edge computing layer. The model management module is used to distribute the parameters and global constraints of lightweight deep reinforcement learning models to the edge computing layer.

Images