Zone area photovoltaic multi-dimensional model and data aggregation method

Abstract

The invention discloses a transformer area photovoltaic multi-dimensional model and data aggregation method, which comprises the following steps: S1, establishing a five-dimensional digital twinborn model: carrying out structured modeling on distributed photovoltaic physical characteristics, operating environment and regulation and control constraints, generating standardized feature vectors required by subsequent clustering aggregation, and providing a unified analysis dimension for a core link; s2, horizontal federation learning clustering based on a hierarchical region: completing data aggregation and decision grouping according to a hierarchical regulation and control demand by relying on a standardized feature vector output by the five-dimensional model in the S1, and completing cross-subject data interaction by relying on a subsequent transmission system; s3, building a cloud-edge-end data aggregation transmission system: realizing real-time acquisition of five-dimensional model data, secure interaction of federal clustering data and closed-loop issuing of a final regulation and control strategy through local area networking; the method has the advantages of being high in power grid operation safety, high in photovoltaic absorption capacity, high in system regulation and control efficiency and capable of protecting data privacy.

Description

Technical Field

[0001] This invention belongs to the field of high-proportion photovoltaic regulation and consumption technology, specifically involving a multi-dimensional photovoltaic model and data aggregation method for a distribution area. Background Technology

[0002] Existing distributed photovoltaic (PV) systems are characterized by numerous locations, diverse component characteristics, different installation methods, and varying energy storage configurations. Under the influence of factors such as installation environment and climate conditions, their processing and control characteristics are complex and diverse. Furthermore, the data from each distribution area is scattered, making centralized management and efficient control difficult. At the same time, data privacy protection issues must also be considered, making it difficult to achieve efficient control and absorption of high-proportion PV systems. To solve the above problems, it is necessary to develop a multi-dimensional model and data aggregation method for distributed PV systems. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-dimensional photovoltaic model and data aggregation method for distribution areas that has high grid operation security, strong photovoltaic absorption capacity, high system regulation efficiency, and can protect data privacy.

[0004] The objective of this invention is achieved as follows: a multi-dimensional model and data aggregation method for photovoltaic power distribution areas, comprising the following steps:

[0005] S1. Establish a five-dimensional digital twin model: This serves as the core data input for the entire method. By structurally modeling the physical characteristics, operating environment, and control constraints of distributed photovoltaics, it generates standardized feature vectors required for subsequent clustering and aggregation, providing a unified analytical dimension for the core process.

[0006] S2, horizontal federated learning clustering based on hierarchical regions: used as the core transformation link connecting the five-dimensional model data and actual control applications, relying on the standardized feature vector output by the five-dimensional model in step S1, data aggregation and decision grouping are completed according to hierarchical control requirements, while relying on the subsequent transmission system to complete cross-subject data interaction;

[0007] S3 establishes a cloud-edge-device data aggregation and transmission system: serving as the technical foundation for ensuring the efficient implementation of steps S1 and S2, enabling real-time acquisition of five-dimensional model data, secure interaction of federated clustering data, and closed-loop distribution of final control strategies through local area networking.

[0008] Preferably, the five-dimensional model of step S1 is as follows:

[0009] S11, Structural Model: The installation method and energy storage configuration of distributed photovoltaics are encoded to form a quantifiable structural feature code, providing a unified feature dimension of system hardware configuration for subsequent clustering;

[0010] S12, Geographic Model: Obtain the geographic coordinates of the geometric center of each photovoltaic installation area, combine them with the hierarchical regional boundaries of the transformer area and the county, generate geographic feature parameters, and provide geographic positioning support for the regional association analysis of hierarchical clustering;

[0011] S13, Meteorological Model: Real-time data of the installation area is collected through micro-meteorological devices to construct a meteorological feature dataset, providing a data foundation for power model output calculation and environmental impact similarity analysis during clustering;

[0012] S14, Power Model: Combining the hardware parameters of the structural model and the environmental parameters of the meteorological model, the core operating characteristics of photovoltaics and energy storage are output through the component model, battery model and comprehensive efficiency model, becoming the key quantitative indicator for clustering;

[0013] S15, Regulation Model: Clarifies the upper and lower limits of response capability and response cost constraints at each level, which not only supplements the five-dimensional feature vector with regulation constraint parameters, but also provides a basis for setting the target of subsequent hierarchical clustering.

[0014] Preferably, step S2 specifically includes:

[0015] S21, Determine the hierarchical regions and control requirements: Divide the physical regions according to the hierarchy of transformer area-feeder area-substation supply area-county area, combine the constraints of the control model in the five-dimensional model, clarify the core control objectives of each level, and set the grouping direction and evaluation criteria for clustering;

[0016] S22, Beam Federated Learning Clustering:

[0017] S221, Feature Filtering: Based on the feature vectors generated by the five-dimensional model, filter the common features of power companies and photovoltaic / storage operators to eliminate private data;

[0018] S222, Data Preprocessing: Based on the feature data of the five-dimensional model, each participant performs normalization processing locally to eliminate differences in units and provide input data of a uniform scale for clustering;

[0019] S223, Federated Clustering Training: By coordinating the distributed interaction between nodes and participating nodes, and aiming at hierarchical regulation needs, the five-dimensional feature vector is clustered, and the results of each level of optical / storage clusters are finally output, providing a grouping basis for the implementation of regulation strategies.

[0020] Preferably, the specific steps of step S223, federated clustering training, are as follows:

[0021] S2231, Initialization: Determine the number of clusters K according to the hierarchical requirements, select typical samples that meet the hierarchical requirements from the five-dimensional feature vector as the initial cluster centers, and coordinate the nodes to send the K value, initial cluster centers, and distance calculation rules to all participating nodes in the hierarchical region through cloud-edge-end encrypted transmission;

[0022] S2232, Local clustering calculation of participating nodes: Each participating node loads its own five-dimensional feature vector locally, and completes the clustering classification determination based on the rules issued by the coordinating node. The original data does not flow out of the local area throughout the process.

[0023] S2233, Participating nodes upload local results: After completing local clustering, participating nodes transmit three types of non-original data to the coordinating node only through the edge device-cloud platform in encrypted form to avoid privacy leakage. The three types of non-original data are the local clustering result, the mean of the five-dimensional features of the local cluster, and the hierarchical association label.

[0024] S2234, Coordination Node Updates Global Cluster Center: After receiving the uploaded data from all participating nodes, the coordination node filters the data according to the effectiveness of hierarchical regulation and updates the global center to ensure that the new center is more in line with the regulation objectives.

[0025] S2235, Iterative convergence verification: The coordinating node sends the new global cluster center to all participating nodes through encrypted transmission, and repeats steps S2232 → S2233 → S2234 until the dual convergence condition is met, that is, the data and the control target are both met.

[0026] S2236, Output hierarchical light / storage cluster results: The coordinating node outputs the final clustering results of the hierarchical region. The results clearly define the five-dimensional features and hierarchical regulation attributes, directly providing a grouping basis for subsequent regulation strategies.

[0027] Preferably, in step S2235, the data convergence target is that the change in the five-dimensional feature vector between the new global center and the previous round center is ≤ a preset threshold, to ensure the stability of data features; the regulation convergence target is that the core regulation indicators of each cluster achieve a compliance rate of ≥90%, to ensure that the clustering results can support regulation requirements.

[0028] Preferably, step S3 specifically includes:

[0029] S31, Terminal Equipment Deployment: Deploy terminal sensors in each photovoltaic power station area to collect the raw data required for the five-dimensional model in real time and complete the initial data collection;

[0030] S32, Side device convergence: Install side devices at the center of the photovoltaic power station to receive the collected data from the surrounding transformer substations, process the feature data of the five-dimensional model locally, and temporarily store the normalized feature vectors of each participant and the local cluster mean during the federated clustering process.

[0031] S33, Cloud-Edge Collaborative Transmission: Through a 5G secure network, edge devices encrypt and upload standardized feature data of the five-dimensional model and interactive data of federated clustering to the cloud platform to support clustering training; at the same time, the cluster control strategy generated by the cloud platform is sent back to the edge devices.

[0032] Due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0033] (1) By effectively aggregating and regulating photovoltaic power in the distribution area, this invention can monitor and control photovoltaic output in real time, avoid problems such as transformer reverse overload and voltage fluctuation caused by photovoltaic backfeeding, improve the safety of power grid operation, and ensure the safe and stable operation of the power system.

[0034] (2) This invention can better predict photovoltaic output through an accurate five-dimensional model and a reasonable data aggregation method. At the same time, combined with an energy storage system, it realizes the reasonable storage and release of photovoltaic power, improves the photovoltaic consumption level, and reduces the phenomenon of curtailment.

[0035] (3) This invention enables the power grid to perform hierarchical and graded management and control of distributed photovoltaic power through hierarchical aggregation and horizontal federated learning, thereby improving the accuracy and efficiency of control and reducing control costs.

[0036] (4) Through horizontal federated learning, this invention does not require sharing the original data during the data aggregation process, but only sharing the processed data such as feature vectors, thus protecting the data privacy of each participant.

[0037] In summary, this invention has the advantages of high grid operation safety, strong photovoltaic absorption capacity, high system regulation efficiency, and protection of data privacy. Detailed Implementation

[0038] The technical solution of the present invention will be further described in detail below through embodiments.

[0039] This invention provides a multi-dimensional model and data aggregation method for photovoltaic power distribution areas, comprising the following steps:

[0040] I. Establish a five-dimensional digital twin model: This serves as the core data input for the entire method. By structurally modeling the physical characteristics, operating environment, and control constraints of distributed photovoltaics, it generates standardized feature vectors required for subsequent clustering and aggregation, providing a unified analytical dimension for the core process.

[0041] 1. Structural Model: The installation method (rooftop / ground, angle / orientation) and energy storage configuration (type / capacity / charging / discharging power) of distributed photovoltaics are encoded to form quantifiable structural feature codes, providing a unified feature dimension of system hardware configuration for subsequent clustering.

[0042] 2. Geographic Model: Obtain the geographic coordinates of the geometric center of each photovoltaic installation area, and combine them with the hierarchical regional boundaries of the transformer area and the county to generate geographic feature parameters, providing geographic positioning support for the regional association analysis of hierarchical clustering.

[0043] 3. Meteorological Model: Real-time data such as irradiance and temperature in the installation area are collected by micro-meteorological devices to construct a meteorological feature dataset, providing a data foundation for power model output calculation and environmental impact similarity analysis during clustering.

[0044] 4. Power Model: Combining the hardware parameters of the structural model and the environmental parameters of the meteorological model, the core operating characteristics such as the maximum output of photovoltaic power and the charging and discharging power of energy storage are output through the component model (equivalent circuit), the battery model (Thevenin et al.), and the comprehensive efficiency model (multiple loss factor calculation), which become the key quantitative indicators for clustering.

[0045] 5. Regulation Model: Clarify the upper and lower limits of response capability at each level (such as the power over-limit threshold of distribution transformers and the power flow limit of the line) and response cost constraints. This not only supplements the five-dimensional feature vector with regulation constraint parameters, but also provides a basis for setting the target of subsequent hierarchical clustering.

[0046] II. Horizontal Federated Learning Clustering Based on Hierarchical Regions: This is the core transformation link connecting the five-dimensional model data with actual regulatory applications. It relies on the standardized feature vectors output by the five-dimensional model in step one to complete data aggregation and decision grouping according to hierarchical regulatory needs, while relying on the subsequent transmission system to complete cross-subject data interaction.

[0047] 1. Determine the hierarchical regions and control requirements: Divide the physical regions according to the hierarchy of distribution area - feeder area - substation supply area - county. Combine the constraints of the control model in the five-dimensional model (such as distribution transformer load rate ≤80% in distribution area and high-voltage line power flow limit in county), clarify the core control objectives of each level, and set the grouping direction and evaluation criteria for clustering.

[0048] 2. Beam Federated Learning Clustering:

[0049] (1) Feature filtering: Based on the feature vector generated by the five-dimensional model, the common features of power companies and photovoltaic / storage operators (such as structural coding, geographical coordinates, and maximum output) are filtered out to eliminate privacy data.

[0050] (2) Data preprocessing: Each participant performs normalization processing on the feature data of the five-dimensional model locally to eliminate the difference in dimensions and provide input data of a uniform scale for clustering.

[0051] (3) Federated clustering training: By coordinating the distributed interaction between nodes (power companies) and participating nodes (photovoltaic / storage operators), and taking the hierarchical regulation needs as the target, the five-dimensional feature vectors (structure, geography, meteorology, power, regulation) are clustered, and the results of photovoltaic / storage clusters at each level are finally output (such as the "high load adaptation cluster" at the substation level), providing a grouping basis for the implementation of regulation strategies. The training process is as follows:

[0052] A. Initialize clustering parameters (anchoring hierarchical requirements and five-dimensional features):

[0053] a. Determine the number of clusters K based on the hierarchical requirements. For example, for the distribution area layer, which focuses on "preventing transformer load rate from exceeding limits", k is set to 2 based on the "power scale (five-dimensional power model) + geographical distribution (five-dimensional geographical model)" of the photovoltaic / storage system at that layer (corresponding to "low load adaptation cluster" and "high load adaptation cluster"). For example, for the county layer, which focuses on "high voltage line power flow safety", k is set to 3 based on the line coverage area and the total output scale of photovoltaic / storage (corresponding to "low power flow pressure cluster", "medium power flow pressure cluster" and "high power flow pressure cluster").

[0054] b. Instead of being randomly generated, typical samples that meet the hierarchical requirements are selected from the five-dimensional feature vectors as the initial cluster centers; for example, the initial center 1 of the transformer substation layer (low load adaptation) selects the feature vector of the photovoltaic / storage system with "transformer load rate ≤ 60% area + small output (100-150kW, power model) + near transformer (0.3-0.5km, geographical model)"; for example, the initial center 2 (high load adaptation) selects the feature vector of "transformer load rate ≥ 70% area + large output (200-300kW) + energy storage discharge power ≥ 80kW (structural model)".

[0055] c. The coordinating node distributes the K value, initial cluster center, and distance calculation rules (such as weighted Euclidean distance) to all participating nodes in the hierarchical region via cloud-edge-device encrypted transmission (5G secure network).

[0056] B. Participate in local node clustering calculations (without leaking original data):

[0057] Each participating node loads its own five-dimensional feature vector locally and completes the clustering classification determination based on the rules issued by the coordinating node. The original data does not flow out of the local machine throughout the entire process.

[0058] a. Feature matching and weighted distance calculation: The weighted Euclidean distance is used to calculate the distance between the local feature vector and each initial center, with the weights tilted towards the "hierarchical control of key dimensions";

[0059] For example, the weight allocation for the transformer substation layer is as follows: power dimension (maximum output, energy storage power, five-dimensional power / structural model) weight 0.3, geographical dimension (distance from transformer, five-dimensional geographical model) weight 0.25, regulation dimension (transformer load rate threshold, five-dimensional regulation model) weight 0.25, and structural / meteorological dimension weight total 0.2. Then the distance formula is:

[0060] ;

[0061] In the formula, P represents power characteristics, D represents geographical characteristics, L represents regulation characteristics, S represents structural characteristics, M represents meteorological characteristics, band ′ represents normalized values, and band 0′ represents central values.

[0062] b. Determine the local cluster affiliation: Assign the local feature vector to the initial center cluster with the smallest weighted distance; for example, if a participating node has a maximum output of 208kW of optical / storage system, is 0.8km away from the distribution transformer, and has a distribution transformer load rate of 75%, and is closer to the initial center 2 (high output, distant distribution transformer, high load rate), then it is assigned to the "high load adaptation cluster".

[0063] C. Participating nodes upload local results (only aggregated data is uploaded to protect privacy):

[0064] After participating nodes complete local clustering, they only transmit three types of non-raw data to the coordinating node via the edge device-cloud platform in encrypted form to avoid privacy leaks.

[0065] a. Local clustering results: e.g., belonging to cluster 2 (high load adaptation);

[0066] b. The average of the five-dimensional features of the local cluster; such as the average structural encoding, average power, and average geographical coordinates of the node's own features within the cluster it belongs to.

[0067] c. Layered and associated labels: For example, the substation layer is labeled "Transformer A Coverage Area" and "July Average Irradiance 1050W / m²". 2 ", associated with a five-dimensional geographic / meteorological model.

[0068] D. Coordinate nodes to update global cluster centers (to meet hierarchical requirements):

[0069] After receiving the uploaded data from all participating nodes, the coordinating node filters the data according to the effectiveness of hierarchical regulation and updates the global center to ensure that the new center is more in line with the regulation objectives.

[0070] a. Data filtering and grouping: First, remove data with "abnormal hierarchical labels" (such as outliers that are labeled "Transformer B Coverage Area" but belong to "Transformer A Center"). Then, group the data according to "clustering results" (such as grouping all local mean values ​​that "belong to cluster 1" according to the standard).

[0071] b. Weighted calculation of global mean: The global feature mean of each group is calculated by weighting the "installed capacity ratio" (core parameter of the five-dimensional power model) of the participating nodes, and used as the new global cluster center (e.g., if there are 3 participating nodes in the transformer area layer cluster 2 (high load adaptation) with an installed capacity ratio of 2:3:5, then the maximum output feature of the new center = (260×0.2+280×0.3+300×0.5) = 286kWA).

[0072] c. Control constraint verification: Check whether the new center meets the hierarchical control constraints (e.g., the "distribution transformer load rate correlation value" of the new center at the transformer substation level must be ≤80%, and the threshold of the five-dimensional control model). If it exceeds the standard, readjust the weights to ensure that the center meets the control safety requirements.

[0073] E. Iterative convergence verification (both data and control targets are met):

[0074] The coordinating node distributes the new global cluster center to all participating nodes via encrypted transmission, repeating step B → step C → step C until the dual convergence condition is met.

[0075] a. Data target convergence: The change in the five-dimensional feature vector between the new global center and the previous center is ≤ a preset threshold (e.g., 0.05), ensuring data feature stability;

[0076] b. Convergence of control targets: The core control indicators of each cluster meet the target rate ≥90% (e.g., the "distribution transformer load rate ≤60%" of cluster 1 at the transformer substation level ≥90%, and the "energy storage discharge power ≥80kW" of cluster 2 ≥90%), ensuring that the clustering results can support the control requirements.

[0077] If both convergence conditions are met, the iteration stops; otherwise, the iteration continues until the condition is met.

[0078] F. Output hierarchical optical / storage cluster results (connecting to control strategies):

[0079] The coordinating node outputs the final clustering result of the hierarchical region. The result clearly defines the five-dimensional features and hierarchical regulation attributes, which directly provides the grouping basis for subsequent regulation strategies.

[0080] a. Example of a transformer substation:

[0081] Cluster 1: Low-load adaptable cluster (Features: maximum output 100-150kW, distance from distribution transformer 0.3-0.5km, distribution transformer load rate ≤60%; Adaptation strategy: prioritize local consumption, no need for energy storage dispatch).

[0082] Cluster 2: High-load adaptable cluster (Features: maximum output 200-300kW, distance from distribution transformer 0.6-1.0km, energy storage discharge power ≥80kW; Adaptation strategy: start energy storage discharge when distribution transformer load rate exceeds 70%).

[0083] b. Example of county-level:

[0084] Cluster 3: High power flow pressure cluster (characteristics: total output 5000-8000kW, close to 110kV line, power flow ratio ≥85%; adaptation strategy: reduce some photovoltaic output when the power flow exceeds the limit).

[0085] The final results are synchronized to the hierarchical control system through the cloud-edge-device system, completing the connection from clustering and grouping to the implementation of control measures.

[0086] Third, establish a cloud-edge-device data aggregation and transmission system: This serves as the technical foundation to ensure the efficient implementation of the aforementioned two steps. It enables real-time acquisition of five-dimensional model data, secure interaction of federated clustering data, and closed-loop distribution of final control strategies through local area networking.

[0087] 1. Terminal equipment deployment: Deploy terminal sensors in each photovoltaic power station area to collect raw data required by the five-dimensional model in real time (such as hardware parameters of structural coding, irradiance of meteorological model, output value of power model, etc.) to complete the initial data collection.

[0088] 2. Edge device aggregation: Edge devices are installed at the center of the photovoltaic power station to receive data collected by the surrounding transformer substations. The feature data of the five-dimensional model is processed locally, and the normalized feature vectors of each participant and the local cluster mean are temporarily stored during the federated clustering process to reduce the transmission pressure on the cloud.

[0089] 3. Cloud-edge collaborative transmission: Through the 5G secure network, edge devices encrypt and upload standardized feature data of the five-dimensional model and interactive data of federated clustering to the cloud platform to support cluster training; at the same time, the cluster control strategies (such as energy storage charging and discharging instructions) generated by the cloud platform are sent back to the end devices to realize the closed loop of "data collection-cluster analysis-control execution" and ensure the efficiency of the implementation of the above two steps and data security.

[0090] 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 the claims of the present invention.

Claims

1. A method for aggregating data and multi-dimensional model of a district photovoltaic, characterized in that, Comprise the following steps: S1, establish a five-dimensional digital twin model: use as the data input core of the whole method, by structuring modeling of the physical characteristics, operating environment, regulation constraints of distributed photovoltaic, generate the standardized feature vector required for subsequent clustering and aggregation, provide a unified analysis dimension for the core link; S2, hierarchical regional horizontal federated learning clustering: used as the core conversion link connecting the five-dimensional model data and the actual regulation application, relying on the standardized feature vector output by the five-dimensional model in step S1, complete data aggregation and decision grouping according to hierarchical regulation requirements, and at the same time rely on the subsequent transmission system to complete cross-subject data interaction; S3, build a cloud-edge-end data aggregation transmission system: used as the technical cornerstone to ensure the efficient landing of steps S1 and S2, through local area networking to realize real-time collection of five-dimensional model data, safe interaction of federated clustering data, and closed-loop issuance of final regulation strategy.

2. The method of claim 1, wherein: The five-dimensional model of step S1 is as follows: S11, structural model: encode the installation method and energy storage configuration of distributed photovoltaic, form a quantifiable structural feature code, and provide a unified feature dimension of system hardware configuration for subsequent clustering; S12, geographic model: obtain the geometric center geographic coordinates of each photovoltaic installation area, combine the hierarchical regional boundaries of the transformer area-county area, generate geographic feature parameters, and provide geographic positioning support for regional correlation analysis of hierarchical clustering; S13, weather model: collect real-time data of the installation area through micro-meteorological devices, construct a weather feature data set, and provide a data basis for power model output calculation and environmental influence similarity analysis during clustering; S14, power model: combine hardware parameters of the structural model and environmental parameters of the weather model, output the core operating characteristics of photovoltaic and energy storage through component model, battery model and comprehensive efficiency model, and become the key quantitative index of clustering; S15, regulation model: clearly define the upper and lower limits of response capacity and response cost constraints at each level, which not only supplements the regulation constraint parameters of the five-dimensional feature vector, but also provides a basis for subsequent hierarchical clustering target setting. 3.The method of claim 1, wherein: The step S2 specifically comprises: S21, determine hierarchical regions and regulation demand: divide the physical region according to the level of transformer area-feeder area-substation supply area-county area, combine the constraint conditions of the regulation model in the five-dimensional model, and clearly define the core regulation target of each level, set the grouping direction and evaluation standard for clustering; S22, horizontal beam federated learning clustering: S221, feature selection: based on the feature vector generated by the five-dimensional model, filter the common features of power companies and photovoltaic / storage operators, and remove privacy data; S222, data preprocessing: based on the feature data of the five-dimensional model, each participant completes normalization processing locally to eliminate dimension differences and provide input data of uniform scale for clustering; S223, federated clustering training: through distributed interaction of coordination nodes and participating nodes, cluster the five-dimensional feature vector according to hierarchical regulation requirements, and finally output photovoltaic / storage cluster results at each level to provide grouping basis for regulation strategy landing.

4. The method of claim 3, wherein: The specific steps of step S223 federated clustering training are as follows: S2231, initialization: determine the number of clusters K value according to the hierarchical demand, select typical samples meeting the hierarchical demand from the five-dimensional feature vector as the initial cluster center, and coordinate the node through cloud-edge-end encryption transmission, and then download the K value, initial cluster center and distance calculation rule to all participating nodes in the hierarchical area; S2232, local clustering calculation of participating nodes: each participating node loads its own five-dimensional feature vector locally, and completes clustering attribution determination based on the rules downloaded by the coordinating node, and the original data does not flow out of the local; S2233, participating nodes upload local results: after completing local clustering, the participating nodes only upload three types of non-original data to the coordinating node through edge device-cloud platform encryption transmission to avoid privacy leakage, the three types of non-original data are local clustering attribution results, local five-dimensional feature mean of the cluster and hierarchical association label; S2234, the coordinating node updates the global clustering center: after receiving the uploaded data of all participating nodes, the coordinating node filters the data according to the hierarchical regulation effectiveness and updates the global center to ensure that the new center is more suitable for the regulation target; S2235, iteration convergence verification: the coordinating node downloads the new global clustering center to all participating nodes through encryption transmission, and repeats steps S2232→S2233→S2234 until the double convergence conditions are met, that is, the data and the regulation target are both up to standard; S2236, output hierarchical light / storage cluster result: the coordinating node outputs the final clustering result of the hierarchical area, which clearly shows the five-dimensional features and hierarchical regulation attributes, and directly provides grouping basis for subsequent regulation strategies.

5. The method of claim 4, wherein: In the step S2235, the data convergence target is that the five-dimensional feature vector change amount of the new global center and the last round center is less than or equal to a preset threshold, ensuring the stability of the data features; The regulation convergence target is that the core regulation index of each cluster meets the standard rate of 90%, ensuring that the clustering result can support the regulation demand. 6.The method of claim 1, wherein: The step S3 specifically includes: S31, end device deployment: deploy terminal sensors at each area photovoltaic station to collect real-time original data required by the five-dimensional model and complete preliminary data collection; S32, edge device collection: install edge devices at the center of the photovoltaic station to receive collected data from the surrounding area end devices, process the feature data of the five-dimensional model nearby, and temporarily store the normalized feature vectors and local clustering means of each participating party in the federal clustering process; S33, cloud-edge collaborative transmission: through the 5G secure network, the edge device uploads the standardized feature data of the five-dimensional model and the interaction data of the federal clustering to the cloud platform in an encrypted manner to support clustering training; at the same time, the cluster regulation strategy generated by the cloud platform is reversely downloaded to the end device.