Urban area photovoltaic cluster dynamic grid division method based on space-time coupling tensor

Abstract

The invention discloses a city area photovoltaic cluster dynamic grid division method based on a space-time coupling tensor. The method comprises the steps that active power time sequences and geographic coordinates of all photovoltaic nodes in a city area are synchronously collected; constructing a heterogeneous space adjacency matrix considering the electrical topology constraint, and calculating the static space correlation degree between any two photovoltaic nodes; constructing a time-shifting cross-correlation matrix for capturing the moving characteristics of the weather system, and quantifying the dynamic time correlation degree between nodes; executing spatio-temporal feature tensor fusion to obtain a global spatio-temporal affinity matrix; dividing photovoltaic nodes by using an improved spectral clustering algorithm, adaptively determining an optimal grid number through a contour coefficient, and finally generating a photovoltaic cluster grid division scheme at the moment; and triggering grid reconstruction according to the grid drift threshold. According to the method, physical topology constraints and meteorological time delay characteristics can be effectively considered, accurate aggregation of photovoltaic clusters is realized, and the acceptance capability and regulation and control flexibility of a power distribution network to distributed energy sources are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of power grids and relates to distributed photovoltaic management and control technology, specifically to a dynamic grid partitioning method for urban photovoltaic clusters based on spatiotemporal coupling tensors. Background Technology

[0002] With the deepening implementation of the "dual-carbon" strategy, the penetration rate of distributed photovoltaic (PV) power in urban power distribution networks continues to rise, exhibiting characteristics of numerous points of interest, wide distribution, and small individual capacity. The massive influx of distributed PV power has transformed the power distribution network from a traditional passive network to an active network. Its inherent intermittency and volatility pose severe challenges to the power quality and safe, stable operation of the grid. To achieve efficient management and control of massive, dispersed PV power, the strategy of cluster partitioning and hierarchical control has become an industry consensus. This involves dividing PV sites with similar geographical locations or electrical characteristics into several grids and managing them in a cluster-based manner.

[0003] However, existing photovoltaic (PV) cluster grid partitioning technologies have significant limitations in practical applications. First, existing spatial partitioning dimensions are relatively singular, mostly relying on simple physical clustering based on geographical Euclidean distance or administrative divisions, ignoring the actual electrical topology constraints of the distribution network. Due to the complex routing of urban distribution network lines, two geographically adjacent PV sites may belong to different feeders or even substations. Simple geographical clustering leads to low electrical coupling within the grid, making it difficult to achieve effective voltage and power flow coordinated control.

[0004] Secondly, and more critically, existing technologies generally employ instantaneous correlation analysis when dealing with the temporal correlation of photovoltaic (PV) output fluctuations. This method assumes that PV fluctuations within the same region occur synchronously. However, in real-world scenarios, drastic fluctuations in PV output are primarily caused by cloud cover, and clouds are mobile under the influence of atmospheric winds. This means that power drops at upstream sites are transmitted to downstream sites only after a certain time delay. Existing static correlation calculation methods cannot capture this spatiotemporal delay characteristic caused by cloud movement. This leads to nodes with strong causal relationships but time differences being incorrectly assigned to different grids, severely weakening the cluster's ability to mitigate fluctuations through mutual complementarity.

[0005] Therefore, there is an urgent need to develop a dynamic grid partitioning method that can simultaneously take into account both electrical and geographical constraints and accurately capture the spatiotemporal delay characteristics caused by cloud movement. Summary of the Invention

[0006] Purpose of the invention: To address the control mismatch problem caused by neglecting cloud movement delay and electrical constraints in existing photovoltaic cluster partitioning, this invention provides a dynamic grid partitioning method for urban photovoltaic clusters based on spatiotemporal coupling tensors. By integrating the geographic-electrical heterogeneous adjacency matrix with the time-shift cross-correlation matrix that captures fluctuation lag, the spatiotemporal drift characteristics of photovoltaic output fluctuations are accurately quantified, enabling adaptive dynamic reconstruction of the grid and significantly improving the flexibility and fluctuation smoothing capability of the distribution network for distributed photovoltaic cluster control.

[0007] Technical Solution: To achieve the above objectives, this invention provides a dynamic grid partitioning method for urban photovoltaic clusters based on spatiotemporal coupling tensors, comprising the following steps:

[0008] S1: Establish a city-wide photovoltaic monitoring database and simultaneously collect the active power time series and corresponding geographical coordinates of all photovoltaic nodes within the city within a preset historical sampling period;

[0009] S2: Based on the data collected in step S1, construct a heterogeneous spatial adjacency matrix that considers electrical topology constraints, and calculate the static spatial correlation degree between any two photovoltaic nodes by fusing geographical Euclidean distance and electrical connection impedance.

[0010] S3: Based on the data collected in step S1, construct a time-shift cross-correlation matrix to capture the movement characteristics of the weather system, calculate the maximum cross-correlation coefficient between photovoltaic power output sequences and their corresponding time lag through a sliding time window, and quantify the dynamic time correlation between nodes.

[0011] S4: Perform spatiotemporal feature tensor fusion, mapping the static spatial correlation and dynamic temporal correlation into a global spatiotemporal affinity matrix through a nonlinear coupling function;

[0012] S5: Based on the global spatiotemporal affinity matrix, the photovoltaic nodes are divided using an improved spectral clustering algorithm, and the optimal number of grids is determined adaptively by the contour coefficient, thus generating the photovoltaic cluster grid division scheme at this moment.

[0013] S6: Set a grid drift threshold. When the grid center offset caused by the spatiotemporal characteristics of newly acquired data exceeds the threshold, grid reconstruction is triggered; otherwise, the current partitioning scheme is maintained.

[0014] Furthermore, the method for constructing the heterogeneous space adjacency matrix in step S2 includes:

[0015] A1: Using the Delaunay triangulation algorithm, based on geographic coordinates Construct the physical adjacency graph of the photovoltaic nodes;

[0016] A2: For physically adjacent node pairs Define its spatial distance metric for:

[0017]

[0018] in, For nodes and The geographical Euclidean distance; This represents the electrical distance between two nodes in the distribution network topology, i.e., the equivalent impedance magnitude. and These are the maximum geographical distance and the maximum electrical distance within the city limits, respectively. The geographical-electrical weighting coefficient has a range of values. ;

[0019] A3: Use the Gaussian kernel function to transform the distance metric into spatial similarity, and obtain the matrix elements. :

[0020]

[0021] in, For spatial scale attenuation parameters; based on matrix elements Obtain the adjacency matrix of the heterogeneous space .

[0022] Furthermore, the method for constructing the time-shift cross-correlation matrix in step S3 includes:

[0023] B1: For any two photovoltaic nodes and Power normalized sequence and Define the cross-correlation function ,in, This is the time lag, and its value range is... ;

[0024] B2: Calculate the maximum cross-correlation coefficient within the allowable lag range. and its corresponding optimal lag time :

[0025]

[0026] B3: Construct a dual time correlation degree that includes fluctuation similarity and response synchronization. :

[0027]

[0028] in, As a time synchronization tolerance constant, this formula ensures that high temporal correlation only exists at nodes with similar fluctuation patterns and similar occurrence times; based on dual time correlation... Obtain the time-shift cross-correlation matrix .

[0029] Furthermore, in step S4, the spatiotemporal feature tensor fusion adopts a weighted coupling strategy based on Hadamard product to obtain the global spatiotemporal affinity matrix. elements The calculation formula is:

[0030]

[0031] in, To adjust the coupling coefficient for spatiotemporal correlation preference; For connectivity indicator functions, only if the node and The value is 1 if there is a physical connection path within the power supply range of substations at each voltage level in the distribution network, and 0 otherwise, to ensure that the divided grid does not physically break the power grid structure.

[0032] Furthermore, the process of dividing photovoltaic nodes using the improved spectral clustering algorithm in step S5 includes:

[0033] C1: Calculate the global spatiotemporal affinity matrix degree matrix ,in It is a diagonal matrix, with diagonal elements ;

[0034] C2: Construct the normalized Laplace matrix :

[0035]

[0036] in, It is the identity matrix;

[0037] C3: Calculate the normalized Laplacian matrix The former The eigenvectors corresponding to the smallest eigenvalues ​​form the feature matrix. and to The row vectors are normalized.

[0038] C4: The feature matrix Each row in the array is considered a sample point. The K-means++ algorithm is used to cluster them, and the resulting cluster label is the grid number to which each photovoltaic node belongs.

[0039] Furthermore, the method for adaptively determining the optimal number of meshes through contour coefficients in step S5 includes:

[0040] Based on a comprehensive evaluation of grid cohesion and inter-grid separation, a grid partitioning effectiveness index is defined. :

[0041]

[0042] The first term is the average profile coefficient. For grid The average dissimilarity of internal nodes For grid The average dissimilarity to the nearest neighbor grid node; the second term is a penalty term. To divide the total power transmitted by the severed tie line, The total transmission power of the system, This is the penalty coefficient;

[0043] Select to make The largest The value is used as the optimal number of grid cells.

[0044] Further, in step S5, a preliminary grid is generated according to the photovoltaic cluster grid partitioning scheme, and the grid boundary smoothing correction is performed on the preliminary grid, specifically including:

[0045] D1: Identify enclave nodes located at grid boundaries, i.e., nodes whose grids are inconsistent with the grids of their main neighbors in physical space.

[0046] D2: Define the boundary correction cost function Calculate the enclave node Spatiotemporal affinity gain after reclassification to its neighboring grid:

[0047]

[0048] like Then the node Ownership transferred to neighboring grid This eliminates geographically fragmented, tiny, isolated grids. and Representing nodes respectively The spatiotemporal affinity with neighboring grid m and the original grid.

[0049] Furthermore, the mesh reconstruction triggering mechanism in step S6 is achieved by calculating the mesh centroid drift rate. This is achieved, specifically as follows:

[0050]

[0051] in, yes Time of the first The centroid vector of each grid in the spatiotemporal feature space; only when greater than the preset drift threshold If a significant change is found in the current weather flow field or power grid operation mode, steps S2 to S5 are re-executed; otherwise, the grid division results and corresponding control strategies from the previous moment are used.

[0052] Beneficial effects: Compared with existing technologies, this invention solves the problems of spatiotemporal output fluctuation mismatch and control difficulties caused by cloud movement in urban wide-area photovoltaic clusters. It can effectively balance physical topology constraints and meteorological time delay characteristics, achieving precise aggregation of photovoltaic clusters and significantly improving the distribution network's capacity to accept distributed energy and its control flexibility. Specifically, this is reflected in the following four aspects:

[0053] 1. This invention overcomes the limitations of traditional static partitioning and significantly improves the ability to mitigate power fluctuations in solar power clusters. By constructing a time-shifted cross-correlation matrix that captures the movement characteristics of weather systems, this invention can accurately quantify the "time lag" characteristic of photovoltaic power output fluctuations caused by cloud movement between different nodes. By grouping nodes with strong causal relationships but temporal misalignments into the same grid, the natural spatiotemporal smoothing effect within the cluster is effectively utilized, resulting in a significant reduction in the overall power fluctuation rate of the cluster and greatly alleviating the peak-shaving and frequency regulation pressure on the distribution network side.

[0054] 2. Balancing geographical and electrical constraints to ensure grid partitioning conforms to the physical laws of power grid operation. Addressing the issue of low electrical coupling caused by existing technologies relying solely on geographical distance for partitioning, this invention constructs a heterogeneous spatial adjacency matrix that integrates geographical Euclidean distance and electrical connection impedance. This ensures that the partitioned grids are not only geographically proximate but also tightly connected in electrical topology, avoiding the forced aggregation of physically adjacent nodes belonging to different substations or feeders. This effectively guarantees the integrity of the power grid structure and facilitates coordinated control of voltage and power flow.

[0055] 3. An adaptive dynamic reconstruction mechanism was established, improving the system's robustness to environmental changes. This invention introduces a reconstruction triggering mechanism based on the grid centroid drift rate. The system can monitor changes in the spatiotemporal characteristic space in real time. When significant changes occur in the weather field or power grid operation mode, causing the drift to exceed a threshold, grid reconstruction is automatically triggered. This dynamic response mechanism ensures that the grid partitioning scheme always maintains optimal matching with the current meteorological environment and power grid status, overcoming the shortcomings of static partitioning models in adapting to complex and variable weather conditions.

[0056] 4. The grid boundary morphology has been optimized, improving the practical feasibility of engineering applications. This invention introduces a grid boundary smoothing correction strategy to identify and process enclave nodes, and uses a boundary correction cost function to eliminate geographically fragmented micro-island grids, thereby optimizing the grid space. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention;

[0058] Figure 2 A spatiotemporal trajectory map of photovoltaic distribution and cloud movement within the city;

[0059] Figure 3 This is a spatiotemporal correlation analysis diagram;

[0060] Figure 4 A comparison chart of the segmentation results;

[0061] Figure 5 This is a comparison chart of the cluster fluctuation mitigation capabilities under two partitioning strategies. Detailed Implementation

[0062] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0063] Example 1:

[0064] like Figure 1 As shown, this embodiment provides a dynamic grid partitioning method for urban photovoltaic clusters based on spatiotemporal coupling tensors, including the following steps:

[0065] S1: Establish a city-wide photovoltaic monitoring database, and conduct historical sampling within a preset period. Within the city, data is collected simultaneously from all photovoltaic nodes. Active power time series and corresponding geographic coordinates ;

[0066] S2: Based on the data collected in step S1, construct a heterogeneous spatial adjacency matrix that considers electrical topology constraints, and calculate the static spatial correlation degree between any two photovoltaic nodes by fusing geographical Euclidean distance and electrical connection impedance.

[0067] Methods for constructing adjacency matrices in heterogeneous spaces include:

[0068] A1: Using the Delaunay triangulation algorithm, based on geographic coordinates Construct the physical adjacency graph of the photovoltaic nodes;

[0069] A2: For physically adjacent node pairs Define its spatial distance metric for:

[0070]

[0071] in, For nodes and The geographical Euclidean distance; This represents the electrical distance between two nodes in the distribution network topology, i.e., the equivalent impedance magnitude. and These are the maximum geographical distance and the maximum electrical distance within the city limits, respectively. The geographical-electrical weighting coefficient has a range of values. ;

[0072] A3: Use the Gaussian kernel function to transform the distance metric into spatial similarity, and obtain the matrix elements. :

[0073]

[0074] in, For spatial scale attenuation parameters; based on matrix elements Obtain the adjacency matrix of the heterogeneous space .

[0075] S3: Based on the data collected in step S1, construct a time-shift cross-correlation matrix to capture the movement characteristics of the weather system, calculate the maximum cross-correlation coefficient between photovoltaic power output sequences and their corresponding time lag through a sliding time window, and quantify the dynamic time correlation between nodes.

[0076] Methods for constructing time-shift cross-correlation matrices include:

[0077] B1: For any two photovoltaic nodes and Power normalized sequence and Define the cross-correlation function ,in, This is the time lag, and its value range is... ;

[0078] B2: Calculate the maximum cross-correlation coefficient within the allowable lag range. and its corresponding optimal lag time :

[0079]

[0080] B3: Construct a dual time correlation degree that includes fluctuation similarity and response synchronization. :

[0081]

[0082] in, As a time synchronization tolerance constant, this formula ensures that high temporal correlation only exists at nodes with similar fluctuation patterns and similar occurrence times; based on dual time correlation... Obtain the time-shift cross-correlation matrix .

[0083] S4: Perform spatiotemporal feature tensor fusion, mapping the static spatial correlation and dynamic temporal correlation into a global spatiotemporal affinity matrix through a nonlinear coupling function;

[0084] The spatiotemporal feature tensor fusion employs a weighted coupling strategy based on Hadamard product to obtain a global spatiotemporal affinity matrix. elements The calculation formula is:

[0085]

[0086] in, To adjust the coupling coefficient for spatiotemporal correlation preference; For connectivity indicator functions, only if the node and The value is 1 if there is a physical connection path within the power supply range of substations at each voltage level in the distribution network, and 0 otherwise, to ensure that the divided grid does not physically break the power grid structure.

[0087] S5: Based on the global spatiotemporal affinity matrix, the photovoltaic nodes are divided using an improved spectral clustering algorithm, and the optimal number of grids is determined adaptively by the contour coefficient, thus generating the photovoltaic cluster grid division scheme at this moment.

[0088] The process of partitioning photovoltaic nodes using an improved spectral clustering algorithm includes:

[0089] C1: Calculate the global spatiotemporal affinity matrix degree matrix ,in It is a diagonal matrix, with diagonal elements ;

[0090] C2: Construct the normalized Laplace matrix :

[0091]

[0092] in, It is the identity matrix;

[0093] C3: Calculate the normalized Laplacian matrix The former The eigenvectors corresponding to the smallest eigenvalues ​​form the feature matrix. and to The row vectors are normalized.

[0094] C4: The feature matrix Each row in the array is considered a sample point. The K-means++ algorithm is used to cluster them, and the resulting cluster label is the grid number to which each photovoltaic node belongs.

[0095] Methods for adaptively determining the optimal mesh number using profile coefficients include:

[0096] Based on a comprehensive evaluation of grid cohesion and inter-grid separation, a grid partitioning effectiveness index is defined. :

[0097]

[0098] The first term is the average profile coefficient. For grid The average dissimilarity of internal nodes For grid The average dissimilarity to the nearest neighbor grid node; the second term is a penalty term. To divide the total power transmitted by the severed tie line, The total transmission power of the system, This is the penalty coefficient;

[0099] Select to make The largest The value is used as the optimal number of grid cells.

[0100] A preliminary grid is generated based on the photovoltaic cluster grid partitioning scheme. The preliminary grid is then subjected to grid boundary smoothing correction, specifically including:

[0101] D1: Identify enclave nodes located at grid boundaries, i.e., nodes whose grids are inconsistent with the grids of their main neighbors in physical space.

[0102] D2: Define the boundary correction cost function Calculate the enclave node Spatiotemporal affinity gain after reclassification to its neighboring grid:

[0103]

[0104] like Then the node Ownership transferred to neighboring grid This eliminates geographically fragmented, tiny, isolated grids. and Representing nodes respectively The spatiotemporal affinity with neighboring grid m and the original grid.

[0105] S6: Set a grid drift threshold. When the grid center offset caused by the spatiotemporal characteristics of newly acquired data exceeds the threshold, grid reconstruction is triggered; otherwise, the current partitioning scheme is maintained.

[0106] The mesh reconstruction is triggered by calculating the mesh centroid drift rate. This is achieved, specifically as follows:

[0107]

[0108] in, yes Time of the first The centroid vector of each grid in the spatiotemporal feature space; only when greater than the preset drift threshold If a significant change is found in the current weather flow field or power grid operation mode, steps S2 to S5 are re-executed; otherwise, the grid division results and corresponding control strategies from the previous moment are used.

[0109] Example 2:

[0110] To verify the effectiveness and superiority of the method of the present invention, this embodiment constructs a municipal power distribution network model containing 80 distributed photovoltaic nodes on the Matlab simulation platform for testing, as detailed below:

[0111] I. Simulation Scenarios and Data Foundation

[0112] The experimental setting area is as follows Eighty photovoltaic (PV) nodes were randomly distributed, and the electrical topology of the distribution network was simulated using the Minimum Spanning Tree (MST) algorithm. To realistically represent the PV output fluctuations caused by cloud movement, the experiment introduced the Taylor frozen turbulence assumption, simulating a cloud with Gaussian distribution characteristics moving at a wind speed of... Scan the area.

[0113] Figure 2 The spatiotemporal trajectory of the simulated scene is displayed. Figure 2 The solid black line represents the electrical connection topology of the distribution network, and the colored contour lines show the topology of the distribution network. The spatial location and movement path of the cloud cluster at three sampling times. It can be seen that the cloud occlusion does not instantly cover the entire area, but exhibits a significant spatiotemporal evolution process, which provides the necessary physical scenario for verifying the spatiotemporal coupling mechanism of the present invention.

[0114] II. Analysis of the Spatiotemporal Correlation Mechanism of Photovoltaic Fluctuations

[0115] Figure 3 It reveals the volatility hysteresis that is ignored by existing technologies.

[0116] Figure 3 (a) in the figure selects the power time-series curves of two typical nodes, node A and node B, located upstream and downstream of the cloud's movement path. It can be clearly observed that when the cloud passes by, the blue curve of upstream node A experiences a power drop first, while the orange curve of downstream node B shows a similar fluctuation pattern only after a certain time delay. The two are physically strongly correlated, but they are misaligned on the time axis.

[0117] Figure 3 (b) further presents the analysis results of the time-shift cross-correlation function of the two nodes. Traditional zero-lag correlation calculations yield low coefficients, failing to identify the association between the two nodes; however, the method of this invention, through a sliding window search, accurately captures the correlation at the optimal lag time. As shown by the red dashed line in the figure, the cross-correlation coefficient between the two reaches its peak. This strongly demonstrates the necessity and scientific validity of introducing the time-shift cross-correlation matrix in this invention.

[0118] III. Comparison of Grid Generation Schemes

[0119] To intuitively evaluate the clustering effect of this invention, the number of grid cells is set. The method of this invention is compared with the traditional K-means clustering method based on geographic Euclidean distance, and the results are as follows: Figure 4 As shown.

[0120] Figure 4 As shown in (a), the grid generated by the traditional geographic partitioning method exhibits regular block clustering characteristics, such as circular distribution. It only considers static geographic proximity. Because it ignores the movement direction of cloud clusters, nodes that are on the same cloud cluster path but are geographically far apart are forcibly separated into different grids.

[0121] Figure 4 (b) shows the partitioning result of the method of the present invention. It can be clearly seen that the generated grid shape exhibits a strip-like feature extending along the direction of cloud movement. This partitioning method successfully aggregates nodes with fluctuating causal relationships and temporal sequential characteristics into the same cluster, realizing deep coupling between physical space and meteorological space, which is the core advantage of the technical solution of the present invention.

[0122] IV. Quantitative Evaluation of Cluster Suppression Performance

[0123] Figure 5 The average volatility of aggregated power within each cluster was compared under two partitioning strategies, measured by the standard deviation of the power difference. Figure 5The bar chart data shows that when using traditional geographical partitioning (gray bars), the asynchronicity of fluctuations among nodes within the cluster is weak, resulting in limited complementary mitigation effects and a still high volatility after aggregation. However, by adopting the spatiotemporal partitioning method of this invention (orange bars), the nodes within the cluster are optimized based on spatiotemporal correlation, effectively utilizing the natural smoothing effect brought about by "spatiotemporal misalignment," thus reducing the overall power volatility of the cluster by approximately 35%.

[0124] In summary, the simulation experiments fully demonstrate that the method of the present invention can accurately capture the spatiotemporal drift characteristics of photovoltaic power output, and significantly improve the fluctuation smoothing capability of photovoltaic clusters through dynamic grid partitioning, providing a better control boundary for the stable operation of the distribution network.

Claims

1. A method for dynamic grid partitioning of a city-territory photovoltaic cluster based on spatiotemporal coupling tensors, characterized in that, The method comprises the following steps: S1: Establishing a city-area photovoltaic monitoring database, synchronously collecting active power time series and corresponding geographic coordinates of all photovoltaic nodes in the city area within a preset historical sampling period; S2: According to the data collected in step S1, constructing a heterogeneous space adjacency matrix considering electrical topology constraints, and calculating the static spatial correlation degree between any two photovoltaic nodes by fusing the geographical Euclidean distance and electrical connection impedance; S3: According to the data collected in step S1, constructing a time-shift cross-correlation matrix capturing the moving characteristics of weather systems, calculating the maximum cross-correlation coefficient and its corresponding time lag between photovoltaic output sequences through a sliding time window, and quantifying the dynamic time correlation degree between nodes; S4: Performing space-time feature fusion, mapping the static spatial correlation degree and the dynamic time correlation degree into a global space-time affinity matrix through a nonlinear coupling function; S5: Based on the global space-time affinity matrix, the photovoltaic nodes are divided by using an improved spectral clustering algorithm, and the optimal grid number is adaptively determined by using a contour coefficient, and finally a photovoltaic cluster grid division scheme at this moment is generated; S6: Setting a grid drift threshold, when the grid center offset caused by the change of the space-time features of the newly collected data exceeds the threshold, triggering grid reconstruction, otherwise keeping the current division scheme.

2. The method of claim 1, wherein, The construction method of the heterogeneous space adjacency matrix in step S2 comprises: A1: Using Delaunay triangulation algorithm, based on geographic coordinates constructing a physical adjacency graph of the photovoltaic nodes; A2: For a pair of physically adjacent nodes , define its spatial distance metric as: ； wherein, is the node with geographical Euclidean distance; is the electrical distance between two nodes in the power distribution network topology, i.e., the equivalent impedance modulus value; and are the maximum geographical distance and the maximum electrical distance in the city area, respectively; is the geographical-electrical weight coefficient, which has a value range of ; A3: Transform the distance metric into spatial similarity using a Gaussian kernel function, resulting in matrix elements : ； wherein is a spatial scale decay parameter; according to the matrix element obtaining a heterogeneous spatial adjacency matrix .

3. The method of claim 2, wherein, The construction method of the time-shift cross-correlation matrix in step S3 comprises: B1: Power normalized sequence of any two photovoltaic nodes and B2: Cross-correlation function of any two photovoltaic nodes and B3: Time lag between any two photovoltaic nodes where, is the time lag quantity; B2: calculating the maximum cross-correlation coefficient within the allowed lag range and its corresponding optimal lag time : ； B3: Constructing a dual temporal correlation degree comprising wave fluctuation similarity and response synchronicity : ； wherein, is a time synchronization tolerance constant; according to the double time correlation degree obtaining the time-shift cross-correlation matrix .

4. The method of claim 3, wherein, The spatiotemporal feature tensor fusion in the step S4 adopts a weighted coupling strategy based on Hadamard product, and a global spatiotemporal affinity matrix is obtained whose elements are calculated by the following formula: ； wherein, is the coupling coefficient to adjust the preference of spatiotemporal correlation; is the connectivity indicator function, which is only 1 when node and is 1 if there exists a physical tie-line path between the substations in each voltage level of the distribution network, otherwise 0, to ensure that the divided grid does not physically split the network structure.

5. The method of claim 4, wherein, The process of dividing the photovoltaic nodes by using the improved spectral clustering algorithm in step S5 comprises: C1 : Compute the global spatio-temporal affinity matrix of the degree matrix where is a diagonal matrix with diagonal elements ; C2: Constructing the normalized Laplacian matrix : ； wherein is the identity matrix; C3: compute the normalized Laplacian matrix The eigenvectors corresponding to the first minimum eigenvalues form the feature matrix , and normalize the row vectors of . C4: Each row in the feature matrix is regarded as a sample point, and the K-means++ algorithm is used to cluster it, and the cluster label obtained is the grid number to which each photovoltaic node belongs.

6. The method of claim 5, wherein, The method of adaptively determining the optimal grid number by using the contour coefficient in step S5 comprises: Based on the comprehensive evaluation of the grid cohesion and the grid separation, a grid division effectiveness index is defined : ； wherein the first term is the average profile coefficient, is the grid the average dissimilarity of the inner nodes, is the grid the average dissimilarity to the nearest neighbor grid node; the second term is the penalty term, is the total power transmitted by the partitioned tie lines, is the total power transmitted by the system, is the penalty coefficient; Select to make The largest The value is used as the optimal number of grid cells.

7. The method of claim 6, wherein, The method of generating a preliminary grid according to the photovoltaic cluster grid division scheme in step S5, performing grid boundary smoothing correction on the preliminary grid, specifically comprises: D1: Identifying the enclave nodes on the grid boundary, that is, the nodes whose grid is inconsistent with the grid of the main neighbor nodes in the physical space; D2: define a boundary correction cost function , compute the spatiotemporal affinity gain after reclassifying enclave nodes to their neighbor grids: ； If , the ownership of node is transferred to the neighboring mesh , thus eliminating the geographically fragmented tiny island mesh; and represent the spatiotemporal affinity of node to the neighboring mesh m and the original mesh, respectively.

8. The method of claim 7, wherein, The trigger mechanism of the grid reconstruction in the step S6 is realized by calculating the grid centroid drift rate , specifically: ； in, yes Time of the first The centroid vector of each grid in the spatiotemporal feature space; only when greater than the preset drift threshold If a significant change is found in the current weather flow field or power grid operation mode, steps S2 to S5 are re-executed; otherwise, the grid division results and corresponding control strategies from the previous moment are used.

Images