Power distribution network dynamic cluster division method based on multi-objective optimization

Abstract

The invention discloses a power distribution network dynamic cluster division method based on multi-objective optimization, and aims to solve the problems that an existing power distribution network cluster division method is rigid in multi-objective optimization logic, poor in dynamic adaptability and disjointed with actual demands. The method comprises the following steps: firstly, acquiring and preprocessing related data of the power distribution network, and constructing a node time sequence statistical abstract vector and a similarity attraction index; then, establishing a multi-objective optimization model aiming at minimizing cluster internal power loss, maximizing power supply reliability and minimizing operation cost, and introducing a node similarity attraction mechanism; solving by using a non-dominated sorting genetic algorithm fused with a guide strategy to obtain a Pareto optimal solution; performing multi-attribute decision screening based on operation requirements and the like; and finally, the optimal scheme is used for partition management, and full-process dynamic updating is triggered when the running state changes. The method can improve the operation efficiency and stability of the power distribution network and the comprehensive utilization rate of renewable energy sources, reduces the operation cost, can precisely adapt to a complex operation scene, and enhances the dynamic response capability of the system.

Description

A Dynamic Cluster Partitioning Method for Distribution Networks Based on Multi-Objective Optimization Technical Field

[0001] This invention relates to the field of power system distribution network control technology, and in particular to a dynamic cluster partitioning method for distribution networks based on multi-objective optimization. Background Technology

[0002] As the global energy structure accelerates its transition towards cleaner and lower-carbon energy, the large-scale integration of distributed renewable energy sources, such as wind and solar power, into the distribution network has become a core development trend. Simultaneously, user-side load types are becoming increasingly diversified, encompassing industrial loads, commercial loads, and smart home loads. This shift in energy structure and load types places more stringent demands on the operation and control of the distribution network.

[0003] As a crucial link directly connecting the power system and end users, the distribution network's operational efficiency, power supply reliability, and cost control directly determine the quality of power supply and the overall efficiency of energy utilization. Efficient and reliable distribution network operation is key to ensuring energy security. Therefore, improving the operational capabilities of the distribution network in complex energy environments has become a critical issue that urgently needs to be addressed.

[0004] Cluster partitioning technology, as one of the core means of active distribution network management, provides an effective way to address the above challenges. This technology divides the distribution network into several compact sub-clusters with consistent operating characteristics according to specific rules, enabling precise adjustment of line status, optimized allocation of renewable energy, and flexible load dispatching within each sub-cluster.

[0005] Clustering effectively addresses the volatility challenges posed by distributed power generation, enhancing the controllability and economic efficiency of distribution network operation. For instance, when distributed power output fluctuates, adjusting line status and load distribution within the cluster ensures power supply stability. Simultaneously, optimizing renewable energy allocation improves its absorption capacity, reduces wind and solar curtailment, and further enhances energy utilization efficiency.

[0006] Currently, several methods for dividing distribution network clusters have been proposed. For example, CN117595245 discloses a method and system for dividing distribution network clusters. This method obtains the spatial structure and operation data of the low-voltage distribution network, sets comprehensive performance indicators, and obtains clustering results through steps such as grid division, candidate center grid screening, and incremental clustering. Based on considering the cluster structure strength, internal power balance, and voltage control requirements, it achieves cluster size balance and improved voltage regulation capability.

[0007] However, existing distribution network cluster partitioning methods still suffer from two key flaws, making it difficult to adapt to the complex and ever-changing operational needs of distribution networks: 1. Rigid multi-objective optimization logic: Existing solutions often integrate multiple optimization indicators such as electrical modularity and source-load matching degree into a single comprehensive indicator through fixed weights (such as α, β, γ, δ). This approach cannot take into account the priority differences under different application scenarios. For example, in areas with high renewable energy penetration, to maximize the absorption of renewable energy, it is necessary to prioritize ensuring source-load matching degree; while in heavily loaded areas, to ensure power supply security, it is necessary to focus on reducing power loss. The fixed weight mode inevitably leads to the sacrifice of local performance, making the adaptability of the cluster partitioning scheme seriously insufficient.

[0008] 2. Insufficient robustness and dynamic adaptability of the algorithm: Existing solutions rely on a static process of grid partitioning and incremental clustering. Grid partitioning requires a pre-defined number of categories, k, which cannot automatically adjust the number of clusters when facing dynamic conditions such as fluctuations in distributed power generation output and sudden load changes. Furthermore, the selection of candidate center grids is based solely on the number of sample points, ignoring key differences such as node load characteristics and renewable energy output characteristics. This leads to clustering results that are easily deviated from actual operational needs, making it difficult to cope with the volatility and uncertainty challenges brought about by a high proportion of distributed power sources.

[0009] In summary, existing distribution network cluster partitioning methods have significant shortcomings in terms of multi-objective optimization logic and dynamic adaptability, failing to meet the demands of efficient distribution network operation and management in the context of large-scale distributed renewable energy integration and diversified user-side loads. Therefore, developing a distribution network cluster partitioning method that can synergistically optimize power loss, power supply reliability, and operating costs, and fully adapt to differences in node characteristics and dynamic operating conditions, is both necessary and urgent. This method will provide solid technical support for the efficient operation and management of distribution networks, promoting the development of distribution networks towards greater intelligence, efficiency, and reliability, and better adapting to future changes in energy structure and load types. Summary of the Invention

[0010] The technical problem this invention aims to solve is to provide a dynamic distribution network clustering method based on multi-objective optimization, addressing the technical issues in distribution network control technology where clustering methods struggle to simultaneously address multi-objective optimization, lack dynamic adaptability, and are disconnected from actual operational needs. Specifically, existing distribution network clustering methods often focus on a single optimization objective, failing to simultaneously consider the coordinated optimization of power loss, power supply reliability, and operating costs. Furthermore, these methods do not adequately consider the differences in node load characteristics and renewable energy output characteristics, leading to insufficient robustness and poor dynamic adaptability in actual operation. This invention aims to overcome these limitations and achieve efficient operation and management of distribution network zones.

[0011] To achieve the above technical objectives, the present invention adopts the following technical solution: The present invention provides a dynamic distribution network cluster partitioning method based on multi-objective optimization. It constructs a complete distribution network cluster partitioning and operation management system through six closely connected steps. The core innovation points run through the entire process of data processing, model construction, algorithm optimization, and solution implementation: (I) Data acquisition and preprocessing step 1: Acquire basic data of distribution network operation, including topology data (node ​​connection relationship, line impedance), node load data (active / reactive power demand at different times), line parameter data (rated capacity, maximum allowable current) and renewable energy access data (access location, capacity, output fluctuation characteristics).

[0012] Perform refined preprocessing on the raw data: according to a uniform time step High- and low-frequency data are aggregated or interpolated to form a unified time grid; robust interpolation with time windows and Kalman smoothing are used to repair isolated missing points to ensure data continuity; a descriptive feature set is constructed and normalized for each node, and a node time-series statistical summary vector is defined. : (1); where, Represents a node Average active load within the observation window This represents the standard deviation of the active power load. This indicates the peak-to-valley ratio to characterize the temporal instability of the load. Represents the average power factor. Indicates the installed capacity of renewable energy at the node. The relative volatility of renewable energy output is defined as the ratio of the output variance to the mean. This represents the normalized value of the equivalent impedance of a node to the common busbar of the distribution network.

[0013] The node dissimilarity metric is transformed into a similarity driver through a covariance-weighted formula: (2); where Σ is the feature covariance matrix of the entire network, This represents the inverse of the covariance matrix Σ, which counteracts the effects of feature scale and correlation; it is then mapped to an attractiveness index using a soft-gate function. : (3); parameters By controlling the rate of attraction decay, a balance between diversity and cohesion is achieved through small step-size tuning.

[0014] (ii) Multi-objective optimization model construction step 2: Construct a multi-objective optimization model with "minimizing power loss within the cluster, maximizing power supply reliability, and minimizing operating costs" as the core, set multiple constraints and introduce a node similarity attraction mechanism.

[0015] Optimize the target expression: (14); (15); (16); among which, For power loss within the cluster ( For line current, (Line resistance) The power supply reliability index is a weighted average of the average outage time (SAIDI), the average outage frequency (SAIFI), and the power supply reliability index (RLI). ); Operating costs (including cluster operating costs) Distributed energy dispatch costs Loss and cost ).

[0016] Constraints: including line capacity constraints Node voltage constraints Cluster size constraints In addition, cluster connectivity and radial structure constraints are used to ensure that the cluster can operate stably after partitioning.

[0017] Node similarity attraction mechanism: through load feature similarity Similarity of renewable energy characteristics Electrical distance similarity The formula for the overall attractiveness of the quantified node aggregation tendency is: (17); among which The similarity calculation methods are as follows: Load feature similarity: (4); among which, The Pearson correlation coefficient for the load curve. This represents the peak-to-valley ratio difference. For power factor difference, Similarity in characteristics of renewable energy: (5); among which, The correlation coefficient of the output curve. For the variance of fluctuation, Electrical distance similarity: (6); among which, For nodes and The equivalent resistance.

[0018] (III) Step 3 of the improved non-dominated sorting genetic algorithm: The non-dominated sorting genetic algorithm (NSGA-II) with a similarity attraction-guided strategy is used to solve the model, resulting in multiple Pareto optimal cluster partitioning schemes: Population initialization: Weighted initialization based on a node similarity scoring model, ensuring that nodes with similar characteristics have high fitness in the initial population; Fitness evaluation: Introducing similarity attraction weighting, the formula is: (7); among which, For traditional fitness functions, As an attractiveness weighting factor, Node similarity attraction score; Elite crossover operator: prioritizes individuals with high weighted fitness, the formula for weighted fitness ranking is: (18); Bundled crossover is used for sets of highly similar nodes: (19); Controlling the crossover range to ensure crossover only occurs within similar node groups; dynamic adjustment mechanism: dynamically adjusting the attraction coefficient during evolution to balance exploration and focusing. (10); among which, It is in the The attractiveness factor of the generation It is the initial value. For the current algebra, It is the largest algebra.

[0019] (iv) Multi-attribute decision-making to select the optimal solution Step 4: Based on the distribution network operation requirements, load forecasting and renewable energy penetration rate, a multi-attribute decision-making method is used to select the Pareto optimal solution: Operation requirements assessment: Consider the overall load, network topology, power supply capacity and stability under load fluctuations; Load forecasting adaptation: Combine historical and real-time data to evaluate the adaptability of the solution to future load changes; Renewable energy penetration rate adaptation: Evaluate the stability and energy utilization efficiency of the solution under different penetration rates; Through a weighted decision model, the optimal solution that adapts to the current operating environment is determined.

[0020] (V) Step 5 of the partitioned operation management implementation: Transform the optimal solution into a dynamic operation control strategy, and achieve energy balance among multiple clusters through collaborative control: Construct a cluster-level operation state vector: (11); among which, Represents a cluster Total active load, Represents a cluster The combined output of renewable energy and distributed power sources, For the reactive power requirements of the cluster, The average node voltage. For equivalent impedance; define a dynamically adjusted comprehensive optimization objective: (12); Control variables Includes circuit breaker status, energy storage power commands, etc. , , These are weighting coefficients, used to suppress power imbalance, reduce line losses, and maintain voltage stability, respectively; attraction-based load dispatching: node power complementarity adjustment is achieved through formulas, forming a locally self-stabilizing structure. (13); among which, For nodes The active power adjustment amount, Represents a set of nodes within the same cluster. This is the scheduling gain coefficient; Represents a cluster Total active load, Represents a cluster The combined output of domestic renewable energy and distributed power sources.

[0021] (vi) Dynamic update mechanism step 6: When the distribution network experiences significant changes in load demand, large fluctuations in renewable energy output, or sudden failures, the entire process update from step 1 to step 5 will be automatically triggered to re-optimize the cluster partitioning scheme and ensure that the cluster partitioning is always adapted to the operating environment.

[0022] The dynamic cluster partitioning method for distribution networks based on multi-objective optimization provided by this invention has the following beneficial effects: 1. This invention effectively solves the problem that existing distribution network cluster partitioning methods are difficult to take into account multi-objective optimization, realizes the coordinated optimization of power loss, power supply reliability and operating cost, and avoids the contradiction among the three.

[0023] 2. This invention overcomes the limitations of existing methods that do not fully consider the differences in node load characteristics and renewable energy output characteristics. Through the node similarity attraction mechanism, it improves the robustness of the cluster partitioning results in actual operation.

[0024] 3. This invention overcomes the shortcomings of existing methods that lack a dynamic update mechanism, enabling cluster partitioning to adapt to changes in the operating status of the power distribution network in real time, ensuring close alignment with actual operating needs.

[0025] 4. This invention provides a solid data foundation for subsequent optimization by accurately collecting power distribution network operation data, ensuring the accuracy and reliability of the optimization process.

[0026] 5. The multi-objective optimization model constructed in this invention takes minimizing power loss within the cluster, maximizing power supply reliability, and minimizing operating costs as its core objectives, and uses quantitative formulas to precisely control the core indicators.

[0027] 6. The node similarity attraction mechanism introduced in this invention quantifies node similarity from three dimensions: load characteristics, renewable energy characteristics, and electrical distance, promoting the natural aggregation of similar nodes to form a cluster with a compact structure and coordinated operation.

[0028] 7. The non-dominated sorting genetic algorithm with similarity attraction guided by this invention accelerates the convergence of the algorithm to a low-loss, low-cost Pareto optimal solution through operations such as weighted initialization and dynamic adjustment of attraction coefficient.

[0029] 8. This invention uses multi-attribute decision-making to comprehensively consider the operational needs of the distribution network, the adaptability of load forecasting, and the adaptability of renewable energy penetration, thereby avoiding local performance sacrifices and improving scenario adaptability.

[0030] 9. The dynamic update mechanism established by this invention can automatically trigger the re-optimization of cluster partitioning when there are load fluctuations, changes in renewable energy output, or sudden failures, thereby enhancing the dynamic response capability of the system.

[0031] 10. This invention further enhances the stability of cluster operation and improves power supply reliability by dynamically adjusting the internal line status of the cluster, optimizing the configuration of renewable energy, and precisely scheduling loads.

[0032] 11. This invention ultimately achieves a dual reduction in power loss and operating costs, significantly improving the economic efficiency and stability of the power distribution network and providing strong technical support for the construction of a green power grid.

[0033] 12. This invention maximizes the absorption of renewable energy, reduces energy waste, improves the comprehensive utilization rate of renewable energy, and promotes the development of power distribution networks towards high efficiency and low carbon through similarity aggregation and dynamic configuration optimization. Attached Figure Description

[0034] The present invention will be further described below with reference to the accompanying drawings and implementation examples: Figure 1 is an overall flowchart of the method of the present invention; Figure 2 is a bar chart comparing the multi-attribute decision screening indicators of the present invention; Figure 3 is a bar chart of the response time of the dynamic update mechanism of the present invention; Figure 4 is a logic diagram of the multi-objective optimization model of the present invention; Figure 5 is a logic diagram of the cluster dynamic update of the present invention. Detailed Implementation

[0035] The technical solution of the present invention will be further described below with reference to the embodiments and accompanying drawings: Embodiment 1 As shown in Figures 1 to 5, this embodiment provides a dynamic cluster partitioning method for distribution networks based on multi-objective optimization, including the following steps: Step 1, acquire the basic operating data of the distribution network, including the topology data, node load data, line parameter data, and renewable energy access data of the distribution network. The topology data includes node connection relationships and line impedance. The node load data includes active power and reactive power demand at different time periods. The line parameter data includes rated capacity and maximum allowable current. The renewable energy access data includes access location, capacity, and output fluctuation characteristics. Step 1 undertakes the key task of mapping the physical operating world into a high-dimensional digital state that can be used by the optimizer in the overall scheme. First, the data source and time semantic requirements are clarified. Topology snapshots of circuit breakers and buses are collected from the field telemetry system. Line resistance, impedance, and rated capacity are collected from power flow monitoring and line terminals. Acquisition (monitoring and data acquisition system) or smart meters obtains active and reactive power time-series data of nodes and marks the sampling timestamp; obtains the installed capacity and output time sequence of renewable energy access locations from grid-connected inverters and weather stations and records measurement uncertainty indicators; the raw data is first time-aligned and processed according to a unified time step. (Time Step) High-frequency and low-frequency data are aggregated or interpolated downwards to form a unified time grid. Isolated missing points are replaced with robust interpolation based on time windows and Kalman smoothing to maintain sequence continuity. Then, a descriptive feature set is constructed for each node and normalized to eliminate the influence of dimensions. The node time-series statistical summary vector is defined as: (1); where Represents a node Average active load within the observation window This represents the standard deviation of the active power load. This indicates the peak-to-valley ratio to characterize the temporal instability of the load. Represents the average power factor. Indicates the installed capacity of renewable energy at the node. The relative volatility of renewable energy output is defined as the ratio of the output variance to the mean. This represents the normalized equivalent impedance of a node to the common bus of the distribution network. To transform these characteristics into similarity drivers that can be directly used by the optimizer, a covariance-weighted node dissimilarity metric is proposed. (2); where the matrix The feature covariance matrix of the entire network is used to offset the effects of different feature scales and correlations. This distance metric takes into account both statistical differences and electrical coupling information, so that nodes with high coordination naturally move closer together in subsequent aggregation or optimization processes. The similarity weights calculated based on this metric are mapped to an attraction index through a soft gate function. (3); where the parameters The decay rate of attraction is controlled and fine-tuned in small steps during model initialization to balance diversity and aggregation. Data quality control is achieved by calculating the signal-to-noise ratio for each variable and adding confidence weights to the feature vectors for weighted processing in subsequent objective function calculations. The entire preprocessing process outputs two types of objects: a node-level feature matrix F used to construct a similarity scoring network and a time series matrix X used for scenario-based power flow calculations in constraint and objective evaluation. These outputs will serve as inputs to the multi-objective optimization model in the next step to directly invoke the feature-driven similarity attraction mechanism in the objective function and constraints, thereby achieving cluster partitioning based on operational characteristics and electrical coupling.

[0036] The innovation of this step lies in constructing node time-series statistical summary vectors through refined data preprocessing and employing a covariance-weighted node dissimilarity metric to transform multi-source heterogeneous data into a unified feature space, thereby providing high-precision input for subsequent optimization. This process not only eliminates the influence of dimensions but also achieves data-driven node aggregation guidance through similarity attraction mapping, ensuring the coordination of cluster partitioning based on operational characteristics and electrical coupling.

[0037] Step two involves constructing a multi-objective optimization model for distribution network cluster partitioning. The optimization objectives are minimizing power loss within the cluster, maximizing power supply reliability, and minimizing operating costs. Constraints such as line capacity, voltage stability, and cluster size are set, and a node similarity attraction mechanism is introduced into the model. In step two, the node similarity attraction mechanism is implemented by constructing a node similarity scoring model. The scoring model uses the attraction between nodes as the optimization driving force and is divided into three dimensions: load characteristic similarity, renewable energy characteristic similarity, and electrical distance similarity. The similarity between nodes is quantified in each dimension, making nodes with similar electricity consumption behavior or consistent load trends more likely to be partitioned into the same cluster.

[0038] Assume the distribution network contains a set of nodes. With line set The cluster set is obtained by partitioning. The multi-objective optimization model is established as follows: (14); (15); (16); among which, Represents a cluster set. For power loss within the cluster, For line current, Line resistance; The power supply reliability index is determined by the average outage time. (System Average Interruption Duration Index, Average Interruption Frequency) (System Average Interruption Frequency Index) and Power Supply Reliability Index (Reliability Index) weighted composition; Operating costs, including cluster operating costs Distributed energy dispatch costs With loss cost , Let be the weighting coefficient, satisfying Constraints include line capacity constraints. Node voltage constraints Cluster size constraints Furthermore, cluster connectivity and radial structure constraints are incorporated to ensure that each cluster structure is reasonable and operational after partitioning. Additionally, a node similarity attraction mechanism is introduced, using node attraction as an implicit driving force of the model through the attraction term. The optimization objective is weighted and adjusted to make nodes with similar structures and operational characteristics easier to aggregate.

[0039] In step 2, the node similarity attraction mechanism is implemented based on a similarity scoring model, by calculating the similarity between any two nodes. and Overall attractiveness This is used to influence the aggregation tendency of nodes in the optimization model; the overall attractiveness is determined by the similarity of load features. Similarity of renewable energy characteristics and electrical distance similarity Its composition, and its calculation method are as follows: (17); among which, The load characteristic similarity is determined by the Pearson correlation coefficient of the nodal daily load curve, and the weighting coefficient is adjustable. Peak-to-valley ratio difference power factor difference Based on comprehensive assessment: (4); among which, Let be the weighting coefficient, satisfying , The Pearson correlation coefficient for the daily load curve at the node. For nodes and The difference between peak and valley ratios For nodes and The power factor difference.

[0040] The similarity of renewable energy characteristics is assessed based on the correlation and fluctuation consistency of distributed power output at nodes: (5); among which, The correlation coefficient of the output curve. For nodes and The variance of output fluctuation, Let be the weighting coefficient, satisfying Electrical distance similarity is represented using nodal impedance distance normalization: (6); among which, For nodes and Equivalent impedance and combined attractive force During the optimization process, the target model is added in the form of a reward function or bias factor, so that nodes with high similarity are more likely to be assigned to the same cluster during iteration, thereby forming a highly coordinated cluster structure with consistent operating characteristics and tight electrical coupling.

[0041] Step 2 constructs a multi-objective optimization model to achieve efficient clustering of the distribution network by minimizing power loss, maximizing power supply reliability, and minimizing operating costs. By introducing a node similarity attraction mechanism, the similarity between nodes is quantified based on three dimensions: load characteristics, renewable energy characteristics, and electrical distance, so that nodes with similar operating characteristics can naturally aggregate, thereby improving the operational coordination and stability of the cluster.

[0042] The innovation of this step lies in constructing a multi-objective optimization model centered on minimizing power loss, maximizing power supply reliability, and minimizing operating costs, and for the first time introducing a node similarity attraction mechanism as an implicit driving force. This mechanism quantifies node similarity from three dimensions: load characteristics, renewable energy characteristics, and electrical distance, making the optimization process naturally tend to form a cluster structure with consistent operating characteristics and tight electrical coupling, thus breaking through the limitations of traditional single-objective optimization.

[0043] Step 3 involves using a non-dominated sorting genetic algorithm to solve the multi-objective optimization model. Similarity attraction is introduced into the population initialization, fitness evaluation, and evolutionary operations to guide the process, enabling nodes to gradually form clusters with similar characteristics and cohesiveness during iteration, resulting in multiple Pareto-optimal cluster partitioning schemes. In Step 3, the non-dominated sorting genetic algorithm actively guides the search direction in the solution space by introducing an elite crossover operator guided by similarity attraction, allowing individuals to automatically adjust gene crossover probability and crossover method based on node similarity attraction during genetic evolution. In the population initialization stage, individuals are weighted and initialized according to the aforementioned node similarity scoring model, ensuring that nodes with similar characteristics have higher fitness in the initial population. During fitness evaluation, an individual's fitness is determined not only by its multi-objective optimization values ​​(such as power loss, power supply reliability, and operating cost) but also by the weighted influence of node similarity attraction, defined as: (7); among which, For individuals fitness value, It is the fitness function of the traditional genetic algorithm. It is a weighting factor for similarity attraction. It is a node The similarity attraction score ensures that similar nodes are preferentially selected throughout the evolutionary process and gradually aggregate into a compact cluster structure. To enhance the convergence of the algorithm, the elite crossover operator in the non-dominated sorting genetic algorithm adopts similarity attraction-based guidance. When selecting parent individuals, individuals with high similarity attraction weighted fitness are prioritized. This is achieved by calculating the weighted fitness of each individual and sorting them. Specifically, the fitness of individuals is weighted and then non-dominated sorted, and the top-ranked individuals are selected for the crossover operation. (8); among which, Individual The weighted fitness ranking value, Based on fitness value The sorting, It is an adjustment coefficient for similarity attraction. In this way, we can prioritize the selection of node combinations that meet the similarity attraction, increase the probability of retaining nodes with high similarity, and thus make the cluster partitioning more natural and coordinated.

[0044] In crossover operations, for sets of nodes with high similarity attraction, a bundled inheritance approach is used for crossover. This means that similar nodes will crossover together at the gene level, forming new individuals. In structural crossover, the crossover operation takes place between nodes within a "cluster," rather than randomly selecting parent individuals. This effectively maintains the clustering of similar nodes and strengthens the transmissibility of highly similar gene segments. The specific crossover method is described by the following formula: (9); among which, It is a new individual after the crossover. and They are the parent individuals, This is a factor controlling the scope of crossover, ensuring that crossover only occurs within similar node groups. To ensure that the attraction-guided mechanism plays an appropriate role at different stages of the genetic algorithm, a strategy for dynamically adjusting the attraction influence coefficient is introduced in the elite crossover operator. In the early stages of genetic evolution, the attraction influence coefficient is small, aiming to preserve population diversity and avoid early over-convergence to certain local optima. In the later stages of evolution, the attraction influence coefficient gradually increases, causing highly similar nodes to cluster more, improving global convergence efficiency, and ensuring a more optimized cluster partitioning scheme. The dynamic adjustment of the attraction influence coefficient follows these rules: (10); among which, It is in the The attractiveness factor of the generation It is the initial value. For the current algebra, This algorithm utilizes a maximum algebraic approach. Through this dynamic adjustment mechanism, it balances initial exploration with later focusing, enabling the algorithm to both fully explore the solution space and quickly converge to the global optimum. By introducing an elite crossover operator guided by similarity attraction and a dynamic adjustment mechanism, the enhanced non-dominated sorting genetic algorithm not only searches the solution space more effectively but also ensures the generation of highly similar cluster partitioning schemes within a multi-objective optimization framework. This method significantly improves the algorithm's convergence speed and the quality of cluster partitioning.

[0045] The innovation of this step lies in deeply integrating the similarity attraction-guided strategy into the non-dominated sorting genetic algorithm. Through weighted initialization, attraction-weighted fitness evaluation, and elite crossover operators, it actively guides the search direction in the solution space. The attraction influence coefficient is dynamically adjusted. It balances exploration and focus, enabling the algorithm to converge quickly to the Pareto front and ensuring that the generated cluster partitioning schemes are both diverse and coordinated.

[0046] Step four involves comprehensively evaluating the candidate cluster partitioning schemes using a multi-attribute decision analysis method, based on the multiple Pareto optimal solutions obtained through the Non-dominated Sorting Genetic Algorithm II (NSGA-II). This step selects the optimal cluster partitioning scheme best suited to the current operating environment, considering factors such as the distribution network's operational requirements, load forecasting, and renewable energy penetration. Distribution network operational requirements: Taking into account factors such as the overall load, network topology, and load fluctuations of the distribution network, it is necessary to evaluate the performance of each cluster partitioning scheme under different operating conditions. For example, during high-load periods, can the cluster partitioning provide sufficient power supply capacity and stability? Load forecasting: Through trend analysis based on load forecasting, combined with historical load data and real-time data, the adaptability of each scheme to future load changes is evaluated. For example, if a cluster experiences significant load fluctuations... The section carries the risk of overload and should be excluded as a priority. Regarding renewable energy penetration: considering the integration of renewable energy, the stability of each scheme under different penetration conditions is evaluated. Renewable energy is highly volatile; if cluster partitioning fails to effectively allocate new energy resources, it may lead to energy waste or unstable power supply. Therefore, it is necessary to evaluate the performance of cluster partitioning in improving the efficiency of renewable energy use. After comprehensively analyzing these factors, a weighted decision model is used to select the optimal cluster partitioning scheme, which is then further verified and adjusted to ensure it meets both long-term and short-term distribution network operation needs. The innovation of this step lies in using a multi-attribute decision-making method to comprehensively evaluate the Pareto optimal solution. It dynamically weights and selects the optimal scheme by combining multiple dimensions such as distribution network operation needs, load forecasting, and renewable energy penetration. This method avoids insufficient scenario adaptability caused by fixed weights and improves the applicability and robustness of cluster partitioning in actual operating environments.

[0047] Step 5: Apply the optimal cluster partitioning scheme to the distribution network operation. This involves adjusting the internal line status of the clusters, optimizing renewable energy configuration, and load dispatching to achieve zoned operation management. Specific applications include: adjusting the internal line status of the clusters: dynamically adjusting the line topology, circuit breaker, and switch status within each cluster based on real-time operating status and load demand. This adjustment effectively distributes the load, preventing overload or power waste in certain areas, while improving power transmission efficiency and stability; optimizing renewable energy configuration: adjusting the energy configuration in different clusters based on the fluctuation characteristics of renewable energy access. For example, during periods of strong winds, more wind energy is connected to the clusters that need it, and during periods of strong solar energy... During peak power generation periods, the impact of volatility is mitigated by scheduling energy storage systems or optimizing load allocation. Load scheduling: Through precise load scheduling algorithms, each cluster can achieve adaptive power allocation based on its actual load demand and power supply. During the scheduling process, the accuracy of load forecasting is also considered, prioritizing the provision of more power support to areas with heavier loads. Step five plays a crucial role in the entire distribution network cluster partitioning system by transforming the static optimal partitioning results into dynamic operation control strategies. Its core lies in achieving steady-state and transient energy balance among multiple clusters through the coordinated control of reconfiguration of line states within the cluster, renewable energy power allocation, and adaptive load scheduling. First, based on the node feature matrix F obtained in Step one and the cluster partitioning results, a cluster-level operating state vector is constructed: (11); among which, Represents a cluster Total active load, This represents the combined output of renewable energy and distributed power sources within the cluster. For the reactive power requirements of the cluster, The average node voltage. As the equivalent impedance, the real-time control objective is to achieve voltage stability and optimal energy transfer by minimizing the weighted objective function of power imbalance and line loss within the cluster. The comprehensive optimization objective for dynamic adjustment is defined as follows: (12); where the control variable vector It includes circuit breaker status, energy storage power commands, and distributed power output settings. These are weighting coefficients. The first term is used to suppress power imbalance within the cluster, the second term represents line losses, and the third term is used to minimize voltage deviation. For line current, For line resistance, The reference voltage is used. Based on real-time load forecasting and renewable energy output forecasting data, the objective function is solved using quadratic programming within each scheduling cycle to obtain the optimal control command, thereby driving the reconfiguration of switch states and the redistribution of output within the cluster.

[0048] Building upon this, a coordination control algorithm based on attraction correction is introduced to enhance the similarity attraction between nodes. It plays a guiding role in the load dispatching process through an attraction-weighted load redistribution formula: (13); among which, For nodes The active power adjustment amount, Represents a set of nodes within the same cluster. For scheduling gain coefficient, attractive force Derived from the feature measurement results of step one and dynamically updated, when two nodes have high similarity in the feature space, their attraction weight is greater, thus favoring complementary power flow in scheduling. This process forms a local synergistic effect of load balancing, enabling highly volatile nodes and stable nodes to form an inherent self-stabilizing structure in the same cluster. The above two-layer optimization realizes dynamic closed-loop control from global power coordination to local self-balancing. Through real-time iterative optimization within the scheduling cycle, the system maintains a state of minimum power loss, voltage stability, and optimal power supply reliability even under fluctuating operating conditions. The adjustment vector and control commands output by this process are fed back to the dynamic update mechanism to realize the adaptive cluster reconstruction and continuous optimization in step six.

[0049] The innovation of this step lies in transforming the static optimal partitioning result into a dynamic operation control strategy. Steady-state and transient energy balance among multiple clusters is achieved through cluster-level operating state vectors and comprehensive optimization objectives. An attraction-based coordination control algorithm is introduced, allowing node similarity to play a guiding role in load scheduling, forming a locally self-stabilizing structure, which significantly improves system operating efficiency and voltage stability.

[0050] Step 6: When the operating status of the distribution network changes, the system will trigger steps one through five above to perform dynamic updates based on the changes. This means that when load demand or renewable energy output changes significantly, the system will reassess the operating status of the distribution network and adjust the cluster partitioning accordingly to ensure that the system can adapt to the new operating environment. For example, when renewable energy output fluctuates greatly, the system may adjust the cluster partitioning, recombining nodes that can better utilize new energy resources into the same cluster. At the same time, in the event of a surge in load demand or a sudden failure, the system will quickly respond and adjust the power supply path by repartitioning the clusters to ensure the safety and reliability of the system.

[0051] The introduction of a dynamic update mechanism enables the distribution network to adaptively adjust to real-time and anticipated demand fluctuations, enhancing the system's ability to respond to uncertainties and unforeseen events. This setup not only improves the flexibility of the distribution network but also effectively reduces operational risks, ensuring a stable power supply under any circumstances.

[0052] The innovation of this step lies in establishing a dynamic update mechanism throughout the entire process. When the operating status of the distribution network changes, the system automatically triggers a re-optimization from data acquisition to solution application. This adaptive capability ensures that cluster partitioning always adapts to the current operating environment, effectively copes with load fluctuations and uncertainties in renewable energy output, and improves the system's flexibility and reliability.

[0053] Example 2 In another preferred embodiment, based on Example 1, this embodiment provides a dynamic clustering method for distribution networks based on multi-objective optimization. Taking a 10kV distribution network in a certain city as the application object, the distribution network includes 50 nodes, 62 lines, 12 distributed photovoltaic power sources and 8 distributed wind power sources. The user side covers industrial load, commercial load and residential load. The load peak and valley differences are significant, and the output of renewable energy fluctuates greatly. It is necessary to achieve efficient operation and management through accurate clustering.

[0054] Step 1: Data Acquisition and Preprocessing. As shown in Figure 1, the data acquisition process is initiated first. A snapshot of the distribution network topology is collected through the on-site telemetry system to obtain node connection relationships and line impedance data. Active power and reactive power time-series data of 50 nodes for 72 consecutive hours are collected through the distribution-level SCADA (Supervisory Control And Data Acquisition) system and smart meters, with a sampling interval of 5 minutes. Parameters such as line rated capacity (all 1000A) and maximum allowable current (all 1200A) are collected through the power flow monitoring terminal. The access location, installed capacity (500kW for a single photovoltaic unit and 1000kW for a single wind power unit) and output time-series data of 12 photovoltaic power sources and 8 wind power sources are collected through the grid-connected inverter and meteorological station.

[0055] Preprocessing of raw data: 5-minute high-frequency load data and 1-hour low-frequency renewable energy output data are aggregated downwards at a unified time step ∆t=15 minutes to form a unified time grid; robust interpolation based on a 2-hour time window and Kalman smoothing are used to repair 3 isolated and missing load data points; a descriptive feature set is constructed for each node and normalized, and the node time series statistical summary vector is defined according to formula (1): (1); where the average active load of node 3 is... Peak-to-valley ratio of load Renewable energy installed capacity Equivalent impedance normalized value .

[0056] The node difference degree is calculated using formula (2) (the feature covariance matrix Σ of the whole network is obtained by fitting the feature data of 50 nodes), and then mapped to the attractiveness index according to formula (3). The parameter λ=0.8 is set. After small step size optimization, the attractiveness between node 3 and node 5 is calculated. The attraction between node 10 and node 12 After preprocessing, the node-level feature matrix F and the time series matrix X are output.

[0057] Step 2: Construction of the multi-objective optimization model. As shown in Figure 4, the construction of the multi-objective optimization model is started according to the model construction logic diagram: First, the three major optimization objectives are determined, and the objective functions are quantified according to formulas (14) to (16), where the weight coefficient of the power supply reliability index is set to , , Cluster operating costs Calculated based on a base of 50,000 yuan per cluster per year.

[0058] Set constraints: Line capacity constraint Node voltage constraints Cluster size constraints (That is, each cluster contains 5 to 10 nodes), while satisfying cluster connectivity and radial structure constraints.

[0059] A node similarity attraction mechanism is introduced, and the comprehensive attraction is calculated according to formula (17), with adjustable weight coefficients set. , , The similarity of load characteristics is calculated according to formula (4), and the Pearson correlation coefficient between node 3 and node 5 is calculated. Peak-to-valley ratio difference Power factor difference ,have to ; Calculate the similarity of renewable energy characteristics according to formula (5), and the correlation coefficient of photovoltaic output between node 8 and node 9. Fluctuation variance ,have to ; Calculate the electrical distance similarity according to formula (6), and the equivalent impedance of node 15 and node 16. ,have to The combined attractiveness of the final nodes 3 and 5 .

[0060] Step 3: Solve the model using the improved NSGA-II algorithm based on the preprocessed feature matrix F and time series matrix X. The model is solved using the NSGA-II algorithm with a similarity attraction-guided strategy: the population size is set to 200, the maximum number of generations T=100, the crossover probability is 0.8, and the mutation probability is 0.1.

[0061] During the population initialization phase, weighted initialization is performed based on a node similarity scoring model, enabling... The node combination increases the fitness by 20% in the initial population; fitness assessment is calculated according to formula (7), setting... Individual 5's traditional fitness score (Fitness(5)) = 0.85, and its node similarity attraction score is... ,have to Parent selection calculates the weighted fitness ranking value according to formula (8), and sets... Individual 5 ,have to The top 30% of individuals will participate in cross-operation.

[0062] Crossover operation on similarity attraction The node set adopts a bundled inheritance, set according to formula (9). Ensure that intersections only occur within similar node groups; the attraction influence coefficient is dynamically adjusted according to formula (10), with an initial value of In the 50th generation 90th generation To balance population diversity and convergence efficiency, 15 Pareto-optimal cluster partitioning schemes were obtained after 100 generations of evolution.

[0063] Step 4: Multi-attribute decision screening. As shown in Figure 2, based on the comparison bar chart of multi-attribute decision screening indicators, 15 candidate schemes were evaluated from three dimensions: operational demand adaptation, load forecast matching, and renewable energy penetration rate adaptation (maximum score 100 points). Scheme 3 scored 92 points in operational demand adaptation (sufficient power supply capacity during high load periods with no overload risk), 88 points in load forecast matching (adapting to a scenario of a 10% load increase in the next 3 days), and 90 points in renewable energy penetration rate adaptation (renewable energy consumption rate reaches 95%), ranking first in overall score.

[0064] Considering the current total load of the distribution network of 18MW, the radial network topology, the predicted load fluctuation of ±8% in the next week, and the renewable energy penetration rate of 35%, the comprehensive score was calculated by a weighted decision model (with weights of 0.4, 0.3, and 0.3 respectively). Scheme 3 scored 90.2 points and was determined as the optimal cluster partitioning scheme. This scheme divides 50 nodes into 6 clusters, with each cluster having 7-9 nodes, which meets the scale constraint.

[0065] Step 5, Zoned Operation Management: Apply the optimal solution to the distribution network operation and implement management according to the collaborative control strategy in Step 5: First, construct the operation state vectors of 6 clusters according to formula (11), and the total active load of cluster 2. Total output of renewable energy Average node voltage equivalent impedance .

[0066] Define the dynamically adjusted comprehensive optimization objective according to formula (12), and set... , , Control variable vector It includes 32 circuit breaker states, 20 energy storage power commands, and 20 distributed power output setpoints. The optimal control command is solved using quadratic programming in each scheduling cycle (15 minutes). A coordination control algorithm based on attraction correction is introduced, and the scheduling gain coefficient is set according to formula (13). In cluster 2, nodes 3 and 5 Active power adjustment at node 3 This enables complementary power flow within the cluster, forming a locally self-stabilizing structure.

[0067] Statistics after one week of operation show that the average power loss within the cluster decreased by 18%, the line loss rate was controlled within 3.2%, and the node voltage deviation was ≤±3%, meeting the voltage stability requirements.

[0068] Step 6: Dynamic Update. As shown in Figure 5, the cluster dynamic update monitoring mechanism is activated to monitor the distribution network operation status in real time. Updates are triggered in the following scenarios: Load Fluctuation Scenario: The industrial load of node 23 suddenly increases by 50%, exceeding the preset threshold of 30%, as shown in Figure 3. In this scenario, the dynamic update response time is 60 minutes. After re-executing steps 1 to 5, node 23 is reorganized with adjacent nodes 22 and 24 into cluster 4, adjusting the power supply path to avoid overload. Renewable Energy Fluctuation Scenario: Photovoltaic power output drops sharply by 60% due to rainy weather, with a response time of 70 minutes. After re-optimizing the cluster division, cluster 6, with concentrated photovoltaic access, shares energy storage resources with cluster 3, which has stable wind power output, improving energy absorption rate. Line Fault Scenario: A single-phase ground fault occurs on line 18, with a response time of 90 minutes. After isolating the faulty line, the cluster boundaries are redefined to ensure that the fault does not spread to other clusters and that power supply reliability is not affected.

[0069] After dynamic updates, the power distribution network still maintains a state of minimum power loss, stable voltage, and reliable power supply, verifying the strong adaptability of the invention to dynamic operating conditions.

[0070] In the preferred embodiment, step 1 includes topology data such as node connection relationships and line impedance; node load data including active and reactive power demands at different times; line parameter data including rated capacity and maximum allowable current; and renewable energy access data including access location, capacity, and output fluctuation characteristics. These settings comprehensively reflect the grid's operating status, providing accurate data support for subsequent power flow calculations and stability analyses. Based on this data, a refined grid model can be constructed to accurately simulate different operating conditions, thereby enabling the development of scientifically sound operating strategies and planning schemes.

[0071] In a preferred embodiment, step 1, processing the basic operational data of the distribution network, includes time-aligning the raw data according to a uniform time step. High-frequency and low-frequency data are aggregated or interpolated downwards to form a unified time grid. Robust interpolation based on time windows and Kalman smoothing are used to replace isolated missing points. A descriptive feature set is constructed for each node and normalized. Node variability is calculated using a covariance-weighted node variability metric, and an attractiveness index is obtained through soft-gate mapping. These settings ensure the effective fusion of data from different time scales, improving data quality and consistency. By constructing descriptive feature sets and normalizing, the influence of dimensions is eliminated. The calculation of node variability and attractiveness index provides a reliable basis for subsequent accurate analysis of the distribution network operation status and the formulation of scientific decisions.

[0072] In the preferred embodiment, step 2 includes constraints such as line capacity constraints, node voltage constraints, cluster size constraints, cluster connectivity constraints, and radial structure constraints. The node similarity attraction mechanism is implemented by constructing a node similarity scoring model, and the comprehensive attraction is composed of a weighted average of load characteristic similarity, renewable energy characteristic similarity, and electrical distance similarity. The above settings ensure that cluster partitioning is both economical and stable, load characteristic similarity reflects the degree of matching of electricity consumption patterns, renewable energy characteristic similarity ensures the complementarity of distributed power sources, and electrical distance similarity optimizes the grid structure. Through multi-dimensional constraints and dynamic attraction mechanisms, intelligent cluster aggregation is achieved.

[0073] In the preferred embodiment, the load characteristic similarity is determined by comprehensively considering the Pearson correlation coefficient of the node's daily load curve, the peak-to-valley ratio difference, and the power factor difference. The renewable energy characteristic similarity is assessed based on the correlation and fluctuation consistency of the distributed power generation output at the nodes. The electrical distance similarity is normalized using node impedance distance. These settings comprehensively consider various characteristics of the power system and accurately measure the degree of similarity between nodes. In this way, energy allocation and dispatch can be rationally planned based on these similarities, improving energy utilization efficiency, ensuring the stable operation of the power system, and achieving more scientific and efficient energy management.

[0074] In the preferred embodiment, in step 3, the non-dominated sorting genetic algorithm introduces an elite crossover operator guided by similarity attraction. During population initialization, a weighted initialization based on a node similarity scoring model is performed. Fitness evaluation incorporates node similarity attraction weighting. In the crossover operation, a bundled inheritance method is used for node sets with high similarity attraction. These settings effectively improve the algorithm's convergence speed and solution set diversity, avoiding getting trapped in local optima. Guided by similarity attraction, the population can more efficiently cluster towards high-quality solution regions during evolution, while maintaining a broad exploration capability of the solution space, ultimately obtaining a better Pareto front solution set.

[0075] In the preferred scheme, the elite crossover operator incorporates a strategy of dynamically adjusting the attraction influence coefficient. The attraction influence coefficient is small in the early stages of genetic evolution to preserve population diversity, and gradually increases in the later stages to improve global convergence efficiency. This setting effectively balances the algorithm's exploration and development capabilities, avoiding premature entrapment in local optima. Simultaneously, by combining adaptive mutation probability with dynamic adjustment of the mutation magnitude based on individual fitness, the algorithm's robustness is further enhanced, ensuring rapid and stable convergence in complex optimization problems.

[0076] In the preferred embodiment, step 4 involves evaluating factors for multi-attribute decision-making screening, including overall distribution network load, network topology, load fluctuation adaptability, future load change adaptability, renewable energy access stability, and renewable energy utilization efficiency. The optimal solution is determined after a comprehensive evaluation of candidate solutions using a weighted decision model. This setup ensures that the screening process considers both current operating conditions and long-term development needs. By quantifying the weight values ​​of each factor, subjective biases are eliminated. Especially for renewable energy access scenarios, the dynamic weight adjustment mechanism can respond in real time to changes in the power grid structure, ensuring the full-cycle applicability of the solution.

[0077] In the preferred embodiment, step 5, the optimal cluster partitioning scheme includes constructing a cluster-level operating state vector, defining a dynamically adjusted comprehensive optimization objective, achieving voltage stability and optimal energy transmission by minimizing internal power imbalance, line losses, and voltage deviations, and introducing a coordination control algorithm based on attraction correction to guide load dispatch. These settings effectively improve the flexibility and economy of power grid operation and reduce the fluctuation impact of large-scale renewable energy integration. By monitoring the cluster state vector parameters in real time and dynamically adjusting the control strategy weights, the multi-objective optimization process is ensured to balance safety constraints and efficiency indicators, achieving coordinated optimization of power generation, grid, load, and storage.

[0078] In the preferred embodiment, step 6 involves changes in the distribution network's operating status, including significant changes in load demand, significant fluctuations in renewable energy output, and sudden faults. These settings enable real-time capture of the distribution network's dynamic characteristics and rapid location of anomalies through multi-source data fusion analysis. Combined with machine learning algorithms to predict state evolution trends, optimal adjustment strategies are automatically generated to ensure the system maintains safe and economical operation even under complex conditions.

[0079] In summary, this invention proposes a dynamic distribution network cluster partitioning method based on multi-objective optimization, aiming to solve many problems existing in current distribution network cluster partitioning methods. Existing methods often focus on a single optimization objective, making it difficult to simultaneously optimize power loss, power supply reliability, and operating costs, and they do not fully consider node differences, resulting in a disconnect between cluster partitioning and actual operational needs.

[0080] This invention first collects and preprocesses data on the distribution network topology, load, line parameters, and renewable energy access. A multi-objective optimization model is then constructed, with optimization objectives encompassing minimizing power loss within the cluster, maximizing power supply reliability, and minimizing operating costs. Simultaneously, a node similarity attraction mechanism is introduced. This mechanism quantifies node similarity from multiple dimensions, including load characteristics, renewable energy characteristics, and electrical distance. For example, load characteristic similarity is quantified through peak-to-valley ratio, renewable energy characteristic similarity is measured based on output patterns, and electrical distance similarity is quantified based on impedance distance.

[0081] In terms of algorithms, the NSGA-II algorithm is adopted, and the cluster partitioning is optimized through the elite crossover operator. During the genetic algorithm evolution, the similarity attraction guides the operation to adjust the probability and method of gene crossover, so that similar node combinations have a higher inheritance probability, gradually forming clusters with strong similarity and compact structure, obtaining multiple Pareto optimal solutions. These solutions are then screened using a multi-attribute decision-making method, comprehensively considering factors such as load forecasting, renewable energy penetration rate, and system operation requirements, to select the optimal solution.

[0082] When the solution is implemented, the status of internal lines within the cluster is adjusted in real time, and the configuration and load scheduling of renewable energy are optimized to achieve zoned operation management and improve resource utilization and stability. When the operating status of the distribution network changes, the system automatically triggers dynamic updates to ensure flexible response to load fluctuations or unstable renewable energy generation, thereby improving emergency response and reliability.

[0083] This invention, through precise division, dynamic adjustment, and intelligent algorithms, can significantly improve the operating efficiency, stability, and comprehensive utilization rate of renewable energy in power distribution networks, reduce operating costs, and promote the construction of green power grids.

Claims

1. A dynamic cluster partitioning method for distribution networks based on multi-objective optimization, characterized in that, Includes the following steps: Step 1: Acquire and process the basic operational data of the distribution network, including topology data, node load data, line parameter data, and renewable energy access data. Step 2: Construct a multi-objective optimization model for distribution network cluster partitioning, with the optimization objectives of minimizing power loss within the cluster, maximizing power supply reliability, and minimizing operating costs. Constraints are set, and a node similarity attraction mechanism is introduced. Step 3: Solve the multi-objective optimization model using a non-dominated sorting genetic algorithm. A similarity attraction guidance strategy is introduced in population initialization, fitness evaluation, and evolutionary operations to obtain multiple Pareto optimal cluster partitioning schemes. Step 4: Perform multi-attribute decision screening on the Pareto optimal solutions. Based on distribution network operation requirements, load forecasting, and renewable energy penetration rate, determine the optimal cluster partitioning scheme. Step 5: Apply the optimal cluster partitioning scheme to distribution network operation. Implement zoned operation management by adjusting the line status within the cluster, optimizing renewable energy configuration, and load scheduling. Step 6: When the distribution network operation status changes, trigger dynamic updates from Step 1 to Step 5.

2. The distribution network cluster partitioning method based on multi-objective optimization according to claim 1, characterized in that: In step 1, the topology data includes node connection relationships and line impedance; the node load data includes active power and reactive power demand at different times; the line parameter data includes rated capacity and maximum allowable current; and the renewable energy access data includes access location, capacity, and output fluctuation characteristics.

3. The distribution network cluster partitioning method based on multi-objective optimization according to claim 1, characterized in that: In step 1, processing the basic operational data of the distribution network includes time alignment of the raw data, adjusting it to a uniform time step. High-frequency and low-frequency data are aggregated or interpolated downwards to form a unified time grid. Robust interpolation based on time windows and Kalman smoothing are used to replace isolated missing points. A descriptive feature set is constructed for each node and normalized. A node time-series statistical summary vector is defined. for: (1); Node dissimilarity is calculated using a covariance-weighted node dissimilarity metric formula. ： (2); then, the attractiveness index is obtained through soft-gate function mapping. ： (3); where, Represents a node Average active load within the observation window This represents the standard deviation of the active power load. This indicates the peak-to-valley ratio to characterize the temporal instability of the load. Represents the average power factor. Indicates the installed capacity of renewable energy at the node. The relative volatility of renewable energy output is defined as the ratio of the output variance to the mean. This represents the normalized value of the equivalent impedance of a node to the common busbar of the distribution network. and Representing nodes respectively and nodes Descriptive feature vectors; superscript Indicates the transpose operator; Let represent the inverse of the covariance matrix Σ, where Σ is the covariance matrix of the characteristics of all nodes in the network; Subscript This indicates the confidence weight in the quality control of associated data; Parameters used to control the rate of attractive force decay.

4. The distribution network cluster partitioning method based on multi-objective optimization according to claim 1, characterized in that: In step 2, the constraints include line capacity constraints, node voltage constraints, cluster size constraints, cluster connectivity constraints, and radial structure constraints. The node similarity attraction mechanism is implemented by constructing a node similarity scoring model, and the comprehensive attraction is composed of a weighted average of load characteristic similarity, renewable energy characteristic similarity, and electrical distance similarity.

5. The distribution network cluster partitioning method based on multi-objective optimization according to claim 4, characterized in that, The load characteristic similarity is determined by combining the Pearson correlation coefficient of the nodal daily load curves, the peak-to-valley ratio difference, and the power factor difference. (4); The similarity of renewable energy characteristics is assessed based on the correlation and fluctuation consistency of the output of distributed power sources at each node: (5); Electrical distance similarity is represented by nodal impedance distance normalization: (6); where, For load feature similarity; 、 、 For three different load characteristic weighting coefficients, satisfying ； Pearson correlation coefficient for the daily load curve of the node; The difference between peak and valley ratios; For nodes and The degree of difference in power factor; Similarity in characteristics of renewable energy; The correlation coefficient of the output curve; For nodes and The variance of output fluctuation; 、 These are two different renewable energy characteristic weighting coefficients, satisfying... ； Electrical distance similarity; For nodes and The equivalent impedance.

6. The distribution network cluster partitioning method based on multi-objective optimization according to claim 1, characterized in that: In step 3, the non-dominated sorting genetic algorithm introduces an elite crossover operator guided by similarity attraction. During population initialization, a weighted initialization is performed based on a node similarity scoring model. In fitness evaluation, node similarity attraction weighting is introduced, defined as: (7); When selecting parent individuals, the weighted fitness ranking value is calculated using the following formula: (8); In the crossover operation, a bundled inheritance method is adopted for the set of nodes with high similarity attraction, specifically: (9); where, For individuals fitness value, It is the crossover operation function in the genetic algorithm. It controls the intersection range. It is a node Similarity attraction score; Individual The weighted fitness ranking value, Based on fitness value The sorting, It is an adjustment factor for similarity attraction; It is a new individual after the crossover. and These are nodes and nodes The parent individual, It is a factor that controls the scope of crossovers, ensuring that crossovers occur only within similar node groups.

7. The distribution network cluster partitioning method based on multi-objective optimization according to claim 6, characterized in that, The elite crossover operator introduces a strategy for dynamically adjusting the attraction influence coefficient, following the rules below: (10); The attraction coefficient is relatively small in the early stages of genetic evolution to preserve population diversity, and gradually increases in the later stages to improve global convergence efficiency; where, It is in the The attractiveness coefficient of the generation It is the initial value. For the current algebra, It is the largest algebra.

8. The distribution network cluster partitioning method based on multi-objective optimization according to claim 1, characterized in that: In step 4, the evaluation factors for multi-attribute decision screening include the overall load of the distribution network, network topology, load fluctuation adaptability, adaptability to future load changes, stability of renewable energy access, and renewable energy utilization efficiency. The optimal solution is determined after a comprehensive evaluation of the candidate solutions using a weighted decision model.

9. The distribution network cluster partitioning method based on multi-objective optimization according to claim 1, characterized in that, In step 5, the optimal cluster partitioning scheme includes constructing a cluster-level runtime state vector. ： (11); Define a dynamically adjusted comprehensive optimization objective. ： (12); By minimizing power imbalance, line loss and voltage deviation within the cluster to achieve voltage stability and optimal energy transmission, a coordination control algorithm based on attraction correction is introduced to guide load scheduling and redistribution: (13); where, Represents a cluster Total active power load, Represents a cluster The combined output of renewable energy and distributed power sources, For the reactive power requirements of the cluster, The average node voltage. Equivalent impedance; For the control variable vector; 、 、 These are the weighting coefficients for minimizing power imbalance within the cluster, line loss, and voltage maintenance deviation, respectively. For line current, For line resistance, Reference voltage; For nodes The active power adjustment amount, Represents a set of nodes within the same cluster. This is the scheduling gain coefficient; Represents a cluster Total active power load, Represents a cluster The combined output of domestic renewable energy and distributed power sources.

10. The distribution network cluster partitioning method based on multi-objective optimization according to claim 1, characterized in that, In step 6, the changes in the operating status of the distribution network include significant changes in load demand, significant fluctuations in renewable energy output, and sudden faults.

Images