Photovoltaic support group form autonomous optimization and cooperative control method oriented to complex terrains

Abstract

The invention provides a photovoltaic support group form autonomous optimization and cooperative control method for a complex terrain, and the method comprises the steps: achieving the precise modeling of a coverage region irradiation field through high-resolution irradiation distribution diagram obtaining, multi-source data fusion and spatial interpolation; the gradient sensitivity and action entropy of each support node are combined to realize judgment and cooperative processing of high-conductivity-low-response risk nodes; according to the method, a lightweight graph attention model and a consistency voting mechanism are utilized to generate locally consistent action correction suggestions, node-level action instruction conflicts are automatically eliminated, the action certainty and response robustness of a photovoltaic support group under the complex illumination condition are effectively improved, and the safety and reliability of system cooperative control are enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent collaborative control technology for photovoltaic power plants, and in particular to a method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain. Background Technology

[0002] Currently, cluster collaborative control technology for photovoltaic power plants in complex terrain and non-uniform cloud shadow environments continues to receive widespread attention. Mainstream photovoltaic support control systems in the industry generally employ centralized command distribution and action scheduling mechanisms based on prior strategies or priority tables to address the collaborative light tracking of distributed support clusters. However, these traditional technical approaches have significant limitations under dynamic and asymmetric irradiance conditions. Constrained by factors such as spatial shading, rapid movement of local cloud shadows, and transient changes in irradiance intensity, the action response requirements of individual support nodes change frequently, and the sensitivity of each node to irradiance gradient exhibits strong spatial heterogeneity. Existing technologies struggle to guarantee the consistency and robustness of the overall response. Currently, some industry efforts have attempted to introduce multi-agent control, cloud shadow prediction-based global scheduling, and reinforcement learning-based distributed strategies. However, mainstream technologies generally rely on complex dynamic cloud shadow modeling or wide-area environmental perception, require high-computing-power central servers for decision-making, or employ static priority allocation and centralized anomaly detection mechanisms. These solutions have the following typical application scenarios and applicable scopes in practical deployments: suitable for small and medium-sized photovoltaic arrays with relatively uniform terrain and stable light distribution. For extremely complex terrain and rapidly changing asymmetric irradiance scenarios, the system still cannot meet the engineering requirements of high availability and high robustness in terms of high-frequency action conflicts, local priority mismatches, and excessive global re-optimization latency. Especially in large-scale distributed support system collaboration, when multiple batches of action commands issued by the central controller lack dynamic quantitative constraints on the actual gradient sensitivity and response determinism of nodes, the ambiguity or misjudgment of priority decisions by support nodes will lead to frequent action command conflicts, which in severe cases can cause local collaboration failures and a decrease in overall power generation efficiency. In addition, existing methods such as abnormal clustering, static weight mapping, and rolling optimization cannot achieve real-time response to the dynamic characteristics of physical coupling between nodes and local conflict resolution. They are easily limited by measurement and control bandwidth, data latency, and computing resources, making it difficult to support the intelligent operation and maintenance needs of large-scale photovoltaic power plants in heterogeneous and complex terrain in the future. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, this invention provides a method for autonomous optimization and collaborative control of photovoltaic support group morphology for complex terrain.

[0004] The technical solution of this invention is implemented as follows: a method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain, comprising: S1: Obtain a real-time irradiance distribution map of the area covered by the photovoltaic support group. The spatial resolution of the irradiance distribution map is 10 cm / pixel, which serves as the input data source for asymmetric irradiance field modeling. S2: Based on the real-time irradiance distribution map, construct a spatial adjacency graph with the physical coordinates of each photovoltaic support as the graph node, calculate the irradiance intensity difference vector of each node's 8-neighborhood, and generate a three-dimensional irradiance gradient tensor as the basic data for gradient field topology encoding. S3: Perform gradient propagation weighted aggregation on the three-dimensional irradiation gradient tensor and output a node-level gradient sensitivity index, which characterizes the propagation position attribute of the node in the irradiation gradient field. S4: Based on the action command type, amplitude range and execution result feedback of each node in the most recent N control cycles, calculate the node action entropy, which serves as a quantitative measure of the node response determinism. S5: Determine whether each node simultaneously satisfies the condition that the gradient sensitivity index is greater than the gradient sensitivity threshold and the node action entropy is less than the action entropy threshold. If the condition is met, the node is determined to be in a high transmission-low response risk state. S6: For nodes in a high transmission-low response risk state, trigger the local caching mode and generate a neighborhood negotiation request, which includes the node's gradient sensitivity index and action entropy; S7: In response to the neighborhood negotiation request, based on the gradient sensitivity index and action entropy of the k nearest neighbor nodes, generate a locally consistent action correction suggestion set through a lightweight graph attention model; S8: Perform a consensus vote on the action correction suggestion set, output a unique correction instruction and overwrite the original control instruction to complete the local resolution of node action instruction conflicts.

[0005] The method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain provided by this invention has the following beneficial effects: (1) This invention achieves accurate identification and pre-intervention of the "high conduction-low response" risk state in the photovoltaic support control system by introducing a collaborative mechanism between the irradiation gradient field encoding submodule and the action entropy constraint submodule. Based on the real-time irradiation distribution map, a spatial adjacency graph is constructed. The node-level three-dimensional gradient tensor is extracted using differential vectors, and the gradient sensitivity index is generated by weighted aggregation of adjacent edges, effectively characterizing the energy conduction trend of each support under non-uniform illumination. At the same time, the conditional probability distribution of historical action commands is statistically analyzed by sliding window and the action entropy is calculated to quantify the diversity and stability of node response behavior. When a node simultaneously satisfies the gradient sensitivity index being greater than the gradient sensitivity threshold and the node action entropy being less than the action entropy threshold, it is determined to be in a risk area that is prone to system oscillation, triggering the local caching and neighborhood negotiation mechanism, thereby achieving dynamic isolation and strategy adjustment before the conflict occurs, significantly improving the system's robustness in strong interference environments. (2) This invention decentralizes the conflict resolution process to the edge controller level. After detecting a risk state, the local node actively enters the cache mode and broadcasts a negotiation request to its k nearest neighbors. The neighboring nodes call the lightweight graph attention model to fuse their respective gradient tensors and action entropy states, generate an action correction proposal set, and output the final correction instruction through consensus voting. The entire process does not require the central controller to intervene in the local negotiation process. This mechanism realizes the elastic migration of control from the center to the edge, greatly reduces the instruction transmission load and response latency, and at the same time enables the system to have stronger local autonomy and topology adaptability, which is especially suitable for complex photovoltaic power station scenarios with large terrain undulations and frequent cloud shadow dynamics. (3) This invention constructs a novel collaborative control path that is completely independent of mainstream intelligent optimization paradigms. It abandons high-complexity algorithm frameworks such as reinforcement learning, multi-agent game theory, blockchain consensus, and rolling optimization, thus avoiding engineering implementation bottlenecks such as model training dependence, difficulty in parameter tuning, and high computational resource consumption. All functional modules are deployed on the edge controller, relying only on local sensing data and neighborhood communication to complete closed-loop control. There is no need for cloud backhaul, model updates, or large-scale computing power support, truly achieving "plug-and-play" deployment. This method forms a complete and patentable technology chain—from irradiation gradient tensor encoding, action entropy modeling, dual threshold criterion setting to distributed negotiation execution, each step has clear logic and well-defined boundaries, and possesses high reproducibility and verifiability. Attached Figure Description

[0006] Figure 1 The flowchart shows the method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain according to the present invention. Figure 2 This is a sub-flowchart of the method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain according to the present invention. Figure 3 This is another sub-flowchart of the present invention's method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. like Figure 1 As shown, this invention provides a method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain, specifically including: S1: Obtain a real-time irradiance distribution map of the area covered by the photovoltaic support group. The spatial resolution of the irradiance distribution map is 10 cm / pixel, which serves as the input data source for asymmetric irradiance field modeling. S2: Based on the real-time irradiance distribution map, construct a spatial adjacency graph with the physical coordinates of each photovoltaic support as the graph node, calculate the irradiance intensity difference vector of each node's 8-neighborhood, and generate a three-dimensional irradiance gradient tensor as the basic data for gradient field topology encoding. S3: Perform gradient propagation weighted aggregation on the three-dimensional irradiation gradient tensor and output a node-level gradient sensitivity index, which characterizes the propagation position attribute of the node in the irradiation gradient field. S4: Based on the action command type, amplitude range and execution result feedback of each node in the most recent N control cycles, calculate the node action entropy, which serves as a quantitative measure of the node response determinism. S5: Determine whether each node simultaneously satisfies the condition that the gradient sensitivity index is greater than the gradient sensitivity threshold and the node action entropy is less than the action entropy threshold. If the condition is met, the node is determined to be in a high transmission-low response risk state. S6: For nodes in a high transmission-low response risk state, trigger the local caching mode and generate a neighborhood negotiation request, which includes the node's gradient sensitivity index and action entropy; S7: In response to the neighborhood negotiation request, based on the gradient sensitivity index and action entropy of the k nearest neighbor nodes, generate a locally consistent action correction suggestion set through a lightweight graph attention model; S8: Perform a consensus vote on the action correction suggestion set, output a unique correction instruction and overwrite the original control instruction to complete the local resolution of node action instruction conflicts.

[0009] Step S1: Obtain a real-time irradiance distribution map of the area covered by the photovoltaic support group. The spatial resolution of the irradiance distribution map is 10 cm / pixel, which serves as the input data source for asymmetric irradiance field modeling. Specifically, this includes: S1.1: Real-time acquisition and processing of raw visible light image sequences collected by sky imagers deployed above photovoltaic power plants, as the basic data source for irradiance intensity sensing; For the raw visible light image sequence signals collected by the sky imager deployed above the photovoltaic power station, a multi-channel high frame rate image acquisition method (parameters: acquisition frame rate ≥ 30fps, channel bandwidth ≥ 100Mbps) is adopted to realize the function of continuous visible light data acquisition covering the entire photovoltaic support group. Furthermore, by using a time-domain synchronization calibration method (parameter: synchronization error ≤ 5ms), the consistency of the acquisition timestamps of each channel is corrected, and a time-aligned set of multi-channel original image frames is obtained. Furthermore, a lens distortion correction method (parameters: radial distortion coefficients k1, k2, tangential distortion coefficients p1, p2) is used to correct the optical distortion of the original visible light image and generate a set of image frames with geometric correction. Furthermore, by using a pixel-level linearization method for illumination intensity (parameters: grayscale value range [0,255], Gamma correction coefficient γ), a linear mapping conversion between pixel brightness values ​​and actual irradiance intensity is achieved, and linearized irradiance image data is generated. By using the spatial geometric resampling processing method (parameter: projection matrix accuracy error ≤ 0.01), the linearized image frames from the previous step are resampled into a unified planar projection coordinate system to form a raw visible light image sequence without pose deviation, which serves as the basic data source for irradiance sensing and provides high-precision data support for subsequent multi-source fusion steps. For example, four sky imagers were deployed at a mountain photovoltaic power station, with a frame rate set to 40fps and a channel bandwidth configured to 120Mbps to ensure coverage of the entire support structure. After processing the acquired multi-channel images using a time-domain synchronization calibration algorithm, the synchronization error was 3ms, meeting high consistency requirements. Lens distortion parameters (radial k1=-0.032, k2=0.014, tangential p1=0.001, p2=-0.0005) were obtained using a calibration board, and the image edge linearity was significantly improved after distortion correction. A Gamma correction coefficient γ=2.2 was used to perform a linearization transformation on the image grayscale values, using the following formula to convert the grayscale values... Mapped to irradiance E: in, The original pixel grayscale value. Gamma coefficient The corresponding irradiance values ​​are used. The mapped data undergoes spatial geometric resampling and, using a projection matrix with a precision error of 0.005, all images are stitched together into a panoramic sequence in a unified coordinate system. The sequences sampled at different time segments provide a stable and high-resolution basic input for subsequent cloud image fusion and gradient field construction, effectively improving the system's real-time response capability under complex lighting conditions. S1.2: Based on real-time cloud image data from meteorological satellite downlink, perform geographic coordinate registration processing to accurately map the satellite cloud image to the geographic coordinate system of the photovoltaic support group coverage area to ensure spatial consistency of multi-source data; For real-time cloud image data downlinked from meteorological satellites, a geographic coordinate registration method (parameters: reference coordinate system type WGS-84, number of calibration control points M≥5) is used to transform the cloud image pixel coordinate system to a geographic coordinate system consistent with the coverage area of ​​the photovoltaic support group. Based on the timestamps and orbital parameters in the cloud image metadata, perform time-space synchronization processing (parameter: satellite velocity vector). Imaging time This enables time-domain frame matching with visible light images from the sky imager, ensuring that data from different sources correspond to the same geographical location at the same time. Furthermore, a projection transformation model (parameters: four-parameter affine or seven-parameter Helmert model) is used to map the pixel coordinates of the initially registered cloud map to the geographic coordinate plane through an affine matrix, and the transformed cloud map geographic location matrix is ​​obtained. Furthermore, perform least-squares adjustment of the control point residuals (parameter: geographic coordinate vector of the control points). Control point pixel coordinate vector This optimizes registration accuracy and generates the sum of squared residuals of control points as output. Furthermore, boundary clipping processing is performed on the registered cloud map (parameter: polygon of the boundary of the photovoltaic support group coverage area) to retain only the corresponding cloud map information within the coverage area and eliminate irrelevant images outside the area; Through the above geographic coordinate registration process, the satellite cloud image is converted into geographic coordinate system data consistent with the photovoltaic support group, realizing the spatial consistency technical effect of multi-source irradiation sensing data; For example, in a photovoltaic power plant deployment scenario, the real-time cloud image received from the meteorological satellite has a resolution of 500m / pixel and uses WGS. Using the 84-baseline coordinate system, ten ground control points were selected, evenly distributed within the power station's coverage area, with latitude and longitude accuracy all less than 0.00001°. The imaging time from the cloud image metadata was matched with the sky imager frame sequence in the time domain, ensuring a synchronization window length of less than 200ms. A seven-parameter Helmert projection model was used to batch convert satellite cloud image pixel coordinates into regional geographic coordinates, obtaining preliminary mapping results. Residual calculations were performed on the ten control points, resulting in: in, For the geographic coordinates of the control point target, To register the geographic coordinates of the results, Given the total number of control points, output the average squared residuals. =2.35m. Using this residual as the optimization objective, the local registration parameters were adjusted, ultimately reducing R to 1.12m. The cloud map was cropped to the boundary polygon of the photovoltaic support group, resulting in cloud map data containing only the power plant area. This registered cloud map can be accurately overlaid with the sky imager image, significantly improving spatial consistency and providing a high-precision foundation for subsequent multi-source data fusion. S1.3: Perform multi-source data fusion processing on the original visible light image sequence of the sky imager and the registered satellite cloud image data to generate a preliminary irradiance distribution map, so as to integrate multi-view irradiance information and eliminate the bias of a single data source; Using the raw visible light image sequence acquired by the sky imager and the satellite cloud image data registered with geographic coordinates as the input dataset, a multi-source image irradiance fusion method (parameters: multi-view weight matrix, spatial matching tolerance 0.05m) is adopted to establish the spatial correspondence of irradiance information from different viewpoints and to initially align the irradiance values. Furthermore, by using the irradiance intensity normalization processing method (parameter: minimum-maximum normalization threshold [0,1]), the dimensionality of irradiance values ​​from different source data is unified, and the normalized irradiance intensity matrix is ​​obtained as the basis for fusion calculation. Furthermore, a weighted overlay fusion algorithm (with weight parameters based on the fitting results of the imaging platform height ratio and sensor response curve) is employed to achieve weighted fusion of the sky imager irradiance matrix and the satellite cloud image irradiance matrix, generating a fused irradiance matrix. The weighted calculation formula is as follows: in, This represents the irradiance value after fusion. and These are weighted terms for the irradiance data from the sky imager and satellite cloud images, with the weights set based on the sensor's signal-to-noise ratio and spatial resolution performance; Furthermore, by using a spatial consistency correction method (parameters: block size 2m×2m, gradient smoothing coefficient 0.8), spatial smoothing of local differences within the fused irradiance matrix is ​​achieved, and bias regions from a single data source are eliminated, generating irradiance distribution data with enhanced spatial consistency. By combining computation and spatial consistency correction, the independent irradiance observations from the sky imager and satellite cloud images in the previous step are transformed into a preliminary irradiance intensity distribution map containing multi-source information and with significantly reduced bias, thus realizing the high-precision input data required for asymmetric irradiance field modeling. For example, in a mountainous photovoltaic power station scenario, the sky imager's imaging height is 150m, the original image resolution is 0.12m / pixel, and the satellite cloud image resolution is 10m / pixel. The fusion algorithm sets the multi-view weights as follows: =0.65、 =0.35, normalization mapping the sky imager irradiance range [280, 950] W / m² and the satellite cloud image irradiance range [300, 900] W / m² to [0, 1]. A weighted superposition formula is used: The corresponding grid cells are fused and calculated, and then a spatial consistency correction with a smoothing coefficient of 0.8 is applied to the 2m×2m blocks. The spatial consistency of the preliminary irradiance distribution map output by this embodiment is significantly improved, and the intensity abrupt changes caused by cloud shadows in local areas are weakened, which meets the accuracy requirement of subsequent bilinear interpolation to improve to 10cm / pixel, and provides a stable and high-quality data source for the gradient field topology encoding process; S1.4: The preliminary irradiance intensity distribution map is processed by bilinear spatial interpolation algorithm to improve the spatial resolution to 10 cm / pixel to meet the accuracy requirements of asymmetric irradiance field modeling; S1.5: Perform accuracy verification processing on the upgraded irradiance distribution map to confirm that its spatial resolution reaches 10 cm / pixel, and output it as the input data source for asymmetric irradiance field topology coding.

[0010] Step S2: Based on the real-time irradiance distribution map, a spatial adjacency graph is constructed using the physical coordinates of each photovoltaic support as graph nodes. The irradiance intensity difference vector of each node's 8-neighborhood is calculated, and a three-dimensional irradiance gradient tensor is generated as the basic data for gradient field topological encoding. Specifically, this includes: S2.1: Perform spatial adjacency relationship construction on the physical coordinate set of the photovoltaic support, and determine the connection relationship between nodes based on the Euclidean distance being less than a preset threshold to generate a spatial adjacency graph structure; For the input data of the physical coordinate set of photovoltaic support, a spatial adjacency relationship construction algorithm based on Euclidean distance is adopted (parameter: threshold). This allows for the preliminary determination of spatial connectivity relationships between nodes; Furthermore, by using the normalized Euclidean distance calculation method (parameters: coordinate system units, boundary conditions), the distance between node pairs is quantitatively evaluated, and a dataset of adjacent pairs that meet the threshold conditions is obtained. Furthermore, an adjacency matrix generation method (parameters: matrix size, sparsity control) is adopted to convert the adjacency pair dataset into matrix form and generate a spatial adjacency graph matrix structure that can be used for subsequent gradient field topological encoding. Furthermore, by using an index mapping optimization method (parameters: node unique identifier mapping table, storage optimization strategy), a bidirectional reference relationship is established between the adjacency matrix and the physical coordinate set of photovoltaic support nodes, and an efficient adjacency access index set is generated; By constructing a spatial adjacency relationship, the physical coordinate set results from the previous step are transformed into structured spatial adjacency graph data, realizing a standardized expression of the connection relationship between nodes and meeting the input requirements for subsequent gradient calculation and topology coding. For example, in a photovoltaic power station deployment scenario in a mountainous area, the total number of nodes in the photovoltaic support group is 256, and the physical coordinates of the nodes are stored in the form of a planar projection in the WGS-84 coordinate system. The distance is set to 1.5 meters. The Euclidean distance calculation formula is used. in, , Let these be the planar coordinates of the first node. , Given the planar coordinates of the second node, the distance between any two nodes is calculated. By traversing all node pairs and comparing distances to 1.5 meters, 1024 adjacency pairs are obtained. The adjacency pair dataset is input into the adjacency matrix generation module, resulting in a 256×256 sparse matrix with a sparsity of 93.7%. Through index mapping optimization, the mapping between matrix row and column indices and the original node's unique identifier is stored in a hash table, reducing the adjacency query time to 0.36 milliseconds. In this scenario, the output spatial adjacency graph structure significantly improves the computational efficiency of subsequent gradient differencing and topological coding without reducing connection accuracy. S2.2: Based on the generated spatial adjacency graph structure and real-time irradiance distribution map, extract the irradiance value at the corresponding location of each photovoltaic support node to obtain node irradiance input data; Based on the joint relationship input of the spatial adjacency graph structure and the real-time irradiance distribution map, a node position index matching algorithm (parameters: node physical coordinate set, image coordinate system transformation matrix) is adopted to map the physical coordinates of each photovoltaic support node to the corresponding pixel position of the irradiance distribution map. Furthermore, by using the bilinear interpolation method (parameters: irradiance distribution image matrix, mapped pixel coordinates), the irradiance intensity of the four neighboring pixels around the mapped position is weighted and fused to obtain the interpolated irradiance intensity value of a single node, ensuring that the spatial resolution is maintained at the sampling accuracy of 10 cm / pixel. Furthermore, by using the irradiance intensity numerical reading sub-algorithm (parameters: interpolation coordinates, irradiance distribution map matrix index), the irradiance intensity benchmark value of the mapped location is extracted from the real-time irradiance distribution map, and the node irradiance intensity input raw data is generated. Furthermore, a normalization algorithm (parameters: irradiance baseline value, global maximum irradiance, global minimum irradiance) is adopted to achieve linear normalization transformation of irradiance and obtain standardized node irradiance values ​​suitable for gradient calculation. By combining a node position index matching, bilinear interpolation, benchmark extraction and normalization, the physical coordinates of nodes in the spatial adjacency graph are transformed into standardized node irradiance intensity data that can reflect their position irradiance characteristics, thereby achieving accurate acquisition of the basic input required for gradient field topology coding. For example, in a complex hilly photovoltaic power station containing 200 photovoltaic support nodes, the physical coordinates of the nodes are obtained by differential GPS measurement with an accuracy of ±2 cm. These coordinates are mapped onto a real-time irradiance distribution map with a resolution of 10 cm / pixel. The transformation matrix is ​​obtained by fitting ground calibration points, with an RMSE of 0.8 pixels. The interpolated irradiance values ​​for each node are calculated using a bilinear interpolation method with a four-neighbor weight of 0.25. For instance, the physical coordinates of node A (102.3 m, 58.7 m) are mapped to irradiance image pixels (523, 614), resulting in an interpolation of 812 W / m². After reading the baseline values, a normalization formula is used: in, This is the normalized nodal irradiance baseline value. The reference value for node irradiance. and These are the global minimum and maximum irradiance values, respectively. In this example... =812 W / m², =745 W / m², =955 W / m², and the normalized node irradiance baseline value was calculated to be 0.319. This value was then input into the 8-neighborhood irradiance intensity data generation module in S2.3 to ensure the stability and accuracy of gradient calculation and maintain the physical consistency of gradient encoding even in scenarios where the illumination conditions of different nodes vary significantly. S2.3: For each node in the spatial adjacency graph, determine its 8-neighborhood node set and obtain the irradiance value of each neighboring node to form node-level 8-neighborhood irradiance data; For each photovoltaic support node in the spatial adjacency graph, based on the Euclidean distance threshold determination result and the node mapping relationship in the real-time irradiance distribution map, a neighborhood search algorithm (parameter: neighborhood radius) is adopted. Maximum number of neighboring nodes This allows for the limitation of the spatial elements surrounding the target node and the locking of a set of candidate neighborhoods that satisfy the connection relationship; Furthermore, an 8-directional neighborhood determination method is used (parameters: direction set D = {East, Northeast, North, Northwest, West, Southwest, South, Southeast}, minimum direction angle). This allows for directional classification of candidate neighborhood sets, ensuring that each node's neighborhood covers all eight directions, and obtaining a list of directionally classified neighboring nodes. Furthermore, a node irradiance data mapping method is used (parameter: mapping matrix between the pixel coordinate system and the physical coordinate system of the real-time irradiance distribution map). This allows the physical coordinates of each directional neighboring node to be converted into pixel coordinates of the irradiance distribution map, and the irradiance intensity value to be extracted from the specified pixel position, thus obtaining the original dataset of irradiance intensity of the 8-directional neighboring nodes. Furthermore, by using the missing value interpolation method (parameters: interpolation type = bilinear interpolation, missing value identifier NaN), missing or outlier values ​​in the original dataset of irradiance intensity of 8-direction neighborhood nodes are interpolated and repaired, and a complete list of 8-neighborhood irradiance values ​​is generated. By structuring the neighborhood intensity data, the list of 8-neighborhood irradiance values ​​repaired in the previous step is integrated into a node-level 8-neighborhood irradiance data matrix, which provides complete and accurate input basic data for subsequent horizontal differential calculation and curvature approximation calculation. For example, in a mountainous photovoltaic power station deployment scenario, the Euclidean distance threshold between node physical coordinates is set to 5.0 meters, and the neighborhood radius is... Set to 5.5 meters, maximum number of neighboring points Set to 8. The direction set D adopts an 8-direction standard, with the minimum direction angle... The degree is set to 22.5 degrees. After the neighborhood search algorithm is executed, a node obtains a set of neighboring points that accurately cover all eight directions. Mapping matrix. Composed of pixel coordinate resolution (0.10 m / pixel) and mapping coordinate system transformation factor, physical coordinates are transformed into image coordinates through matrix multiplication. Irradiance intensity values ​​at 8 pixel locations are extracted from the real-time irradiance distribution map, with units of W / m². For missing intensity values ​​of neighboring points in the northeast direction, a bilinear interpolation algorithm is used to calculate interpolation results based on the adjacent north and east direction intensity values, generating a repaired numerical set. The irradiance intensity of all 8 neighboring points reaches valid values, forming a node-level 8-neighborhood irradiance intensity data matrix for subsequent difference and curvature calculations, effectively improving the accuracy and robustness of gradient field topology coding. S2.4: Perform horizontal differential calculation on the 8-neighborhood irradiance data at the node level, including calculating the irradiance difference value in the x-direction and the irradiance difference value in the y-direction to generate a horizontal differential vector; S2.5: Calculate the approximate value of curvature in the z-direction based on the node-level 8-neighborhood irradiance intensity data, and combine the horizontal difference vector with the approximate value of curvature in the z-direction to form a three-dimensional irradiance gradient tensor, which serves as the basic data for gradient field topological encoding.

[0011] like Figure 2 As shown, step S3 involves performing gradient propagation weighted aggregation on the three-dimensional irradiation gradient tensor to output a node-level gradient sensitivity index. This index characterizes the propagation location attribute of the node in the irradiation gradient field. Specifically, this includes: S3.1: Calculate the inverse distance weight coefficients of the Euclidean distances of each adjacent edge in the spatial adjacency graph to generate the weight basis for gradient propagation weighted aggregation. Specifically, based on the spatial adjacency relationship of the physical coordinate set of photovoltaic support nodes, the inverse Euclidean distance algorithm is used to perform a normalized inverse transformation operation on the distances of adjacent edges to obtain a standardized set of distance weight coefficients as input parameters for gradient propagation weighted aggregation. S3.2: Based on the spatial adjacency graph, determine the set of 8 neighboring nodes for each photovoltaic support node, and extract the three-dimensional irradiance gradient tensor of the node and its 8 neighboring nodes as the input data for weighted aggregation. Specifically, the spatial adjacency relation matrix is ​​used to perform boundary constraint processing on the node's neighborhood range to obtain the local subgraph data structure of the three-dimensional irradiance gradient tensor, so as to form the original data source for gradient propagation weighted aggregation. S3.3: Apply the inverse distance weight coefficient to perform a weighted average aggregation operation on the 3D irradiance gradient tensor of the 8-neighborhood of node to generate a node-level aggregated gradient vector. Specifically, perform linear weighted summation on the neighborhood gradient tensor data based on the standardized distance weight coefficient set, and output the aggregated gradient vector that characterizes the local propagation properties of the gradient field as the input for Euclidean norm calculation. S3.4: Perform Euclidean norm calculation on the node-level aggregated gradient vector to obtain the preliminary value of the node gradient sensitivity. Specifically, use the vector magnitude calculation algorithm to perform magnitude quantization on the aggregated gradient vector to generate the preliminary value of the gradient sensitivity as the input parameter for linear normalization. S3.5: Map the initial value of the node gradient sensitivity to the [0,1] interval through linear normalization, and output the node-level gradient sensitivity index. Specifically, perform linear scaling on the initial value of gradient sensitivity based on the preset normalization threshold to generate a standardized node-level gradient sensitivity index as a quantitative basis for determining the high transmission-low response risk state.

[0012] like Figure 3As shown, step S4 involves calculating the node action entropy based on the action command type, amplitude range, and execution result feedback of each node over the most recent N control cycles. This action entropy serves as a quantitative measure of the node's response determinism. Specifically, this includes: S4.1: Perform sliding window time series analysis on the set of action command types, amplitude range division criteria and execution result feedback categories of photovoltaic support nodes to construct a conditional probability distribution model of node actions and define the calculation framework of node action entropy; For the set of action command types, amplitude range division criteria and execution result feedback categories of photovoltaic support nodes, a sliding window time series analysis method (parameters: window length = N control cycles, step size = 1 cycle) is used to construct a state evolution dataset of nodes on a time series scale. Furthermore, by using the instruction type mapping encoding method (parameters: lift = encoding 1, translation = encoding 2, rotation = encoding 3, etc.), discrete action instruction types are transformed into computable category variable vectors, and an instruction type vector matrix is ​​obtained. Furthermore, an amplitude interval normalization method is adopted (parameters: upper limit of amplitude = maximum drive stroke, lower limit = 0) to map the amplitude data of each action command to the [0,1] interval and generate amplitude feature vectors; Furthermore, by performing result feedback classification and coding processing (parameters: success = code 1, delay = code 0.5, conflict = code 0), the qualitative feedback is converted into a quantitatively computable response state vector; By combining triplets of action command type vector, amplitude feature vector, and response state vector, a conditional probability estimation algorithm is used to construct a conditional probability distribution model of node actions. The calculation formula is as follows: in, For conditional probability, For the j-th action category, This represents the current state of the i-th node. This represents the number of times an action category and state co-occur within the sliding window. The number of times the state occurs; Based on the above conditional probability distribution model, a calculation framework for node action entropy is defined, and the information entropy formula is used for measurement: The degree of determinism in node action responses is quantified using information entropy; By using sliding window time series analysis and conditional probability model construction, the instruction type, amplitude range and feedback category data from the previous step are transformed into conditional probability distributions of node actions, thus providing a unified probability framework for subsequent action entropy calculation and risk assessment. For example, in a photovoltaic power station located in hilly terrain, the sliding window length is configured to 20 control cycles, the window step size is 1 cycle, and the three types of actions—lifting, translation, and rotation—are coded as 1, 2, and 3, respectively. The amplitude range is normalized to the range [0,1] based on the maximum driveable stroke of 1.5 meters. The execution feedback category is coded as success 1, delay 0.5, and conflict 0. In the time series analysis of node A, the lifting action occurs 6 times in state s1, translation occurs 4 times in state s1, and rotation occurs 0 times in state s1. The total number of occurrences of state s1 is 12. Therefore, the conditional probability of lifting = =0.5, translation conditional probability = =0.333, rotation conditional probability = =0, substituting into the information entropy formula, we can calculate the action entropy of node A in state s1. The calculation results show that the response uncertainty of this node is low, and it can be used as a subsequent high-transmission node. The input criteria for low response risk assessment significantly improve the accuracy of risk detection and support the effective triggering of node local negotiation mechanisms; S4.2: Based on the node action conditional probability distribution model, perform edge controller local storage unit data acquisition operation on the specified photovoltaic support node identifier to extract the action command type, amplitude range and execution result feedback data of the most recent N control cycles, forming a node historical action record sequence; S4.3: Apply a sliding window statistical mechanism to the node's historical action record sequence to calculate the conditional probability distribution of each type of action instruction in a given state, and quantify the node's action response mode; S4.4: Perform a negative logarithmic summation operation on the conditional probability distribution of node actions to calculate the entropy value of node actions and quantify the degree of determinism of node response; For the conditional probability distribution data of node actions, the information entropy calculation algorithm is used (parameter: probability distribution vector). This enables the measurement of uncertainty in various action response modes; Furthermore, through mathematical formula calculations, the formula for calculating information entropy is defined as follows: This involves performing a negative logarithmic weighted summation of the conditional probabilities for each action category to obtain the original action entropy value. For the j-th type of action, For node status, For statistical probability; Furthermore, by using the probability vector traversal processing method (parameter: probability vector length m), the natural logarithm operation is performed on the conditional probability of each type of action, and the product is performed term by term with the corresponding probability value to form a probability-weighted logarithmic vector; Furthermore, by using the vector summation method (parameter: set of elements of the weighted logarithmic vector), a full summation operation is performed on the probability weighted logarithmic vector, and the sign is reversed to generate the node action entropy value, thereby quantifying the degree of determinism of the node action response; By using the information entropy solution method, the conditional probability distribution of the previous step is transformed into the node action entropy value, thereby realizing a quantitative characterization of the consistency and uncertainty of the node's action response in multiple control cycles. For example, in the edge controller of a photovoltaic support group, the statistical results of the action commands of node i in the most recent 5 control cycles are as follows: the probability of the lifting command is 0.4, the probability of the translation command is 0.3, and the probability of the rotation command is 0.3. According to the information entropy calculation formula, the following is calculated: After inverting the sign, the node action entropy is approximately 1.088 bits. This value is below the system's preset action entropy threshold of 1.2, indicating that the node's response mode has high stability. It can be used as a basis for low response uncertainty in subsequent risk status determination, thereby improving the early identification accuracy of control command conflicts. S4.5: Perform numerical range verification and standardization on the calculated node action entropy values ​​to output standardized node action entropy as a quantitative measure of node response determinism.

[0013] Step S5: Determine whether each node simultaneously satisfies the condition that the gradient sensitivity index is greater than the gradient sensitivity threshold and the node action entropy is less than the action entropy threshold. If satisfied, the node is determined to be in a high transmission-low response risk state. Specifically, this includes: S5.1: Perform real-time data acquisition processing on the node-level gradient sensitivity index stored in the edge controller to obtain its current value as an input parameter for risk status determination; The edge controller performs node-level gradient sensitivity index data call interface initialization processing on the storage unit (parameters: node unique identifier, data access permission verification token) to realize the real-time query function of gradient sensitivity for the target node; Furthermore, a high-precision timestamp index retrieval method is used (parameter: sampling time range). The maximum depth of the data cache (M) enables precise positioning of the latest frame gradient sensitivity index record of the node and obtains the raw numerical data with the sampling time identifier. Furthermore, by employing numerical decoding and unit standardization methods (parameters: original encoding format, target unit system), the gradient sensitivity index is transformed from a stored encoding form into a unified physical quantity expression form, generating numerical results with dimensional consistency. ; Furthermore, outlier filtering and data consistency testing methods (parameter: maximum permissible deviation) are used. Historical average ), to achieve Perform anomaly removal and validity verification, and obtain the current value of the valid gradient sensitivity that has passed the verification. ; Through the above data acquisition link, the original records of node-level gradient sensitivity stored in the edge controller are transformed into accurate, quantitative, and standardized current values, thereby providing accurate and timely input technical indicators for subsequent risk status determination. For example, in the edge controller of a photovoltaic support group, the node ID is set to 1024, the data access permission verification token length is 128 bits, and the sampling time range is... The interval is set to 0.5 seconds, and the maximum data cache depth M is set to 512 records. After initializing the interface, the latest gradient sensitivity index record with a sampling time of 2024-06-11 10:35:26 is retrieved using the timestamp index. The original encoding format is a 16-bit fixed-point number, and the target unit system is dimensionless fractional. A unit normalization algorithm is used to convert the original 16-bit fixed-point number into a floating-point physical quantity. The decoded result is... During the outlier filtering phase, the maximum permissible deviation is... Set as Historical average for The test results show that the test results meet the requirements. The condition is used to determine the current value of the effective gradient sensitivity. The final output As an input parameter for risk status determination, it is subsequently used in conjunction with the gradient-sensitive threshold. The accuracy and stability of the indicators were ensured during the comparison process; S5.2: Based on the current value of the obtained node-level gradient sensitivity index, perform a comparison operation with the gradient sensitivity threshold to generate a Boolean judgment result that the gradient sensitivity index is greater than the gradient sensitivity threshold. Based on the current value input of the node-level gradient sensitivity index, a threshold comparison algorithm is used (parameter: gradient sensitivity threshold). Based on the pre-survey data of photovoltaic power plants (such as topographic factors such as slope and aspect), the statistical results of historical irradiance distribution (such as the spatial variation coefficient of irradiance intensity and the frequency of extreme gradient events), and the pre-calibrated parameters of the support array (such as spacing, height, and installation tilt angle), the function of quantitatively judging gradient conduction intensity is realized. Furthermore, through numerical analysis methods (parameter: floating-point precision) This enables high-precision analysis of node gradient sensitivity indices and yields real-valued data results that can be directly used in logical decisions. Furthermore, floating-point comparison operations are employed (parameter: threshold). This function determines the relationship between the gradient sensitivity index and a preset threshold, and generates a Boolean value for the comparison result. Furthermore, the Boolean value is encapsulated through a logical decision function, and a data-structured decision flag is generated for use as the input interface for downstream risk status determination; A Boolean decision result caching mechanism is adopted to store the result of the previous step in the real-time decision buffer of the edge controller, so as to realize synchronous calling with the action entropy decision branch; The above algorithm transforms the results of the previous step into Boolean decision data with gradient sensitivity greater than the threshold, enabling accurate identification of high-conduction nodes and providing a technical basis for the joint determination of high-conduction-low-response risk states. For example, in a mountain photovoltaic power station, the node-level gradient sensitivity index After normalization, it is 0.823, which is the gradient sensitivity threshold. Topographical data, including slopes ranging from 15° to 30° and southeast-facing aspect, were obtained through UAV mapping. Historical radiation field distribution results, including a spatial variation coefficient of 0.23 for radiation intensity and an average of 12 extreme gradient events per year (daily radiation variation rate >20%), were statistically analyzed over the past three years. Combined with array arrangement parameters of 3.5m support spacing, 2.2m height, and 25° installation tilt angle, a parameter coupling model was constructed. Simulation tests were conducted, and the motion entropy threshold was set to 0.80. Numerical analytical methods were used to... Parse into double-precision floating-point format, then apply floating-point comparison operations: in As a node gradient sensitivity index, The threshold is set to the gradient sensitivity threshold. If the comparison result is True, a Boolean flag of 1 is generated through a logical decision function and stored in the real-time decision buffer. This node is then combined with the Boolean flag in the action entropy decision branch and ultimately determined as a high-transmission node, triggering the local cache mode. In this embodiment, because the node sensitivity is significantly higher than the threshold, the system can complete the decision process in milliseconds, significantly improving the real-time performance and accuracy of risk detection. S5.3: Perform real-time data acquisition processing on the standardized node action entropy stored in the edge controller to obtain its current value as another input parameter for risk status determination; For the standardized node action entropy dataset in the local storage unit of the edge controller, a real-time data access interface protocol (parameters: node identifier and timestamp) is used to implement the function of reading the current value of the action entropy of a specified node. Furthermore, by optimizing the data retrieval algorithm (parameters: bidirectional index table structure and LRU caching strategy), the algorithm can quickly locate the action entropy records of high-frequency access nodes and obtain the latest standardized action entropy value of the target node. Furthermore, a numerical consistency verification method (parameter: CRC16 check code and timestamp comparison rule) is adopted to detect whether the read action entropy value has been changed during transmission and storage, and to generate a current action entropy value that has passed the verification. Furthermore, by using a time series integrity verification method (parameters: sliding window length N and continuous period determination rule), the temporal continuity of action entropy data is confirmed, and a valid current value that can be used for risk status determination is output. By calling the data interface within the edge controller, the current value of the qualified action entropy obtained and verified in the previous step is transformed into a risk judgment input parameter, realizing another key quantitative basis required for the judgment of high transmission-low response risk status; For example, in the coordinated control cycle of a photovoltaic power station in complex terrain, the edge controller needs to determine the risk status of support node A. The standardized action entropy data of node A within the most recent N=10 control cycles has been stored in the local storage unit. The system calls the data access interface protocol, inputting the node identifier A and the current cycle timestamp. The bidirectional index table is retrieved to locate data storage block #05. The LRU caching strategy is used to directly hit this storage block and read the current value of the action entropy. =0.287. The data integrity was verified using a CRC16 checksum, and the result met expectations. Subsequently, the timestamp sequence of this value was compared, confirming coverage of the most recent 10 consecutive periods. This value serves as another input parameter for risk status determination and will be compared with the gradient sensitivity index obtained in S5.2. =0.842 used in conjunction. In the subsequent S5.4, the system will compare With action entropy threshold =0.3, determining whether node A meets the low response condition. Verification showed that the local data read latency was controlled within 3ms, ensuring the real-time performance and high reliability of risk assessment. S5.4: Based on the obtained standardized node action entropy current value, perform a value comparison operation with the action entropy threshold to generate a Boolean judgment result that the node action entropy is less than the action entropy threshold; Based on the standardized node action entropy values ​​acquired in real time from the edge controller, a numerical comparison algorithm is used (parameter: action entropy threshold). Based on the historical action execution data (such as action command type, amplitude range, and execution result feedback) of the photovoltaic support drive mechanism, the control cycle response characteristics (such as response delay and execution success rate), and the historical command conflict records (such as conflict frequency and conflict type) are pre-calibrated to realize the threshold determination function of the current response determinism of the node. Furthermore, using the double-precision floating-point comparison method (parameters: IEEE 754 standard, comparison precision) This achieves standardized motion entropy and preset thresholds. The precise numerical comparison was obtained, and the judgment difference was determined. ; Furthermore, a symbolic function processing method is adopted (parameters: Input), to achieve The symbols are logically parsed, and a Boolean decision flag is generated to indicate whether the node action entropy is lower than 1. ; Furthermore, by using a Boolean value mapping method (parameter: Boolean decision flag value, mapping table {true:1,false:0}), the decision result is encoded as a logical quantity output so that it can be used for subsequent logical AND operations with the gradient sensitivity decision result; Through the above numerical comparison and symbol mapping processing, the current value of the standardized node action entropy is transformed into a Boolean judgment result that the node action entropy is lower than the threshold, thereby realizing the function of generating the second input parameter in the high transmission-low response risk state judgment chain. For example, in the edge controller of a photovoltaic power station, the current value of the standardized action entropy of node i is 0.345, and the action entropy threshold is... By collecting data on the support drive mechanism over the past 12 months, including execution records of three types of commands: lifting, rotation, and translation (82% successful execution, 15% delayed response, and 3% conflict events), the average response delay of the control cycle was statistically determined to be 0.8 seconds. Historical command conflicts were mainly concentrated during periods of sudden irradiation changes (an average of 2.3 times per day). A mapping model between motion entropy and response determinism was established, and the motion entropy threshold was set to 0.40 through simulation testing. The judgment difference was calculated using a double-precision floating-point comparison method. =0.345-0.40. The formula is as follows: Where Δ is the decision difference, 0.345 is the current standardized action entropy value of the node, and 0.40 is the action entropy threshold. This is obtained by performing symbolic function processing. The sign is negative, corresponding to a Boolean decision flag of true. The Boolean value mapping method maps true to a logical value of 1, outputting the decision result that the node's action entropy is below a threshold. Under different conditions, for example, if a node's action entropy is 0.42, then the calculation... =0.02, the sign is positive, the corresponding Boolean decision flag is false, and the logical value is 0. The above process can be repeatedly executed in batch node determination. The verification results show that the logical value output is completely consistent with the manual data comparison results, ensuring the reliability of the high transmission-low response risk state determination; S5.5: Perform logical AND operation processing on the Boolean decision results where the gradient sensitivity index is greater than the gradient sensitivity threshold and the Boolean decision results where the node action entropy is less than the action entropy threshold, so as to output a high transmission-low response risk state determination signal.

[0014] Step S6: For nodes in a high propagation-low response risk state, trigger the local caching mode and generate a neighborhood negotiation request. This request includes the node's gradient sensitivity index and action entropy. Specifically, it includes: S6.1: Parse and process the risk status determination results to identify nodes in a high-transmission-low-response risk state; initialize the local cache data structure based on the identification results to allocate storage space for instruction temporary storage; configure a conflict detection flag in the cache to monitor the instruction execution status in real time; enable the cache locking mechanism when the node risk status is determined to be true to prevent the original control instructions from being executed directly; output a local cache mode activation signal as a prerequisite for generating a neighborhood negotiation request. The high-conduction-low-response risk status judgment signal is parsed and processed using a logical parsing algorithm (parameters: Boolean judgment matrix, node identifier index table) to achieve accurate identification of the risk node set; Furthermore, through memory allocation management methods (parameters: cache block size = 64KB, allocation strategy = contiguous page locking), the local cache data structure based on the risk node set is initialized, and the cache address mapping table is obtained; Furthermore, a conflict detection flag configuration method (parameters: detection period = 5ms, flag type = binary trigger bit) is adopted to realize real-time monitoring of instruction execution status in the buffer area and generate a conflict detection initialization state set; Furthermore, by using a cache locking mechanism to initiate the method (parameters: locking mode = write protection, locking trigger condition = risk status is true), the entry point for executing the original control instruction is blocked, and a cache locking control signal is generated. By using a signal routing control method (parameters: routing delay ≤ 1ms, signal type = activation signal), the cache locking control signal is converted into a local cache mode activation signal, thereby setting the preconditions for generating neighborhood negotiation requests. For example, in the edge controller of a mountain photovoltaic power station, when the support unit with node ID 23 is determined by S5.5 to meet the requirements... and When the condition is met, the logical parsing algorithm parses the Boolean value "true" at the corresponding position in the risk status judgment matrix and adds node 23 to the risk node set. The memory allocation management module allocates a continuously locked 64KB cache area to this node and registers its cache address range in the address mapping table. The conflict detection flag is configured as a binary trigger bit that is updated periodically every 5ms, so that the instruction status in the cache can be monitored in each execution cycle. When the locking mechanism is activated, write protection is implemented to block all action instruction entries issued by the central controller, and a locking control signal is generated. The signal routing module outputs a local cache mode activation signal with a delay of ≤1ms, which becomes a necessary condition for S6.2 to execute interception and cache storage. In this embodiment, the link delay from cache activation to conflict detection activation is kept within 2ms, ensuring timely freezing and subsequent negotiation preparation for instruction conflicts under rapidly changing irradiance conditions, significantly improving system response consistency and control stability. S6.2: Trigger local cache mode based on local cache mode activation signal; intercept and process raw control instructions issued by the central controller and store them in the local cache area; monitor and analyze the execution status of cached instructions to detect potential instruction conflict events; maintain cache lock state when a conflict is detected to ensure instruction temporary storage safety; output cache ready signal as input basis for generating neighborhood negotiation request; Based on the local cache mode activation signal, the original control commands issued by the central controller are intercepted by an instruction interception algorithm (parameters: cache mode status bit = true, instruction priority identifier = original) to realize the interrupted transmission of the instruction stream and transfer it to the local cache unit in the edge controller. Furthermore, by using the memory-mapped write method (parameters: cache address offset = 0x1000, data block size = dynamically allocated according to instruction length), the binary data sequence of the original control instructions is stored in an ordered manner in the local cache, and the cache index table record is obtained as input for subsequent status monitoring. Furthermore, an instruction execution status monitoring method (parameters: detection period = 5ms, monitoring flag = conflict detection enabled) is adopted to realize virtual simulation of the execution status of instructions stored in the buffer and generate a status vector, which includes conflict risk bits, delay risk bits, and execution reliability bits. Furthermore, the potential conflict risk value is calculated using the conflict event determination formula. : in This is the Boolean value for the conflict risk bit. For the delay risk bit, a Boolean value is given when... Greater than At this time, maintain the cache locked state and prevent the direct issuance and execution of cache instructions; The state vector from the previous step is converted into a cache lock signal through the state lock control logic (parameters: lock delay = 10ms, relock threshold = the conflict risk value does not decrease within the interval between two detections), thereby ensuring the security of cache temporary storage. For example, in a mountain photovoltaic power station scenario, the edge controller detects that the support node numbered Node_37 is in a high conduction-low response risk state, triggering a local cache mode activation signal. The original lift instruction issued by the central controller is 128 bytes long. After being moved to the cache area at memory address 0x1000 by the instruction interception algorithm, a cache index ID of 37-20240315 is generated. The instruction execution status monitoring algorithm runs at a detection cycle of 5ms. When it detects a conflict risk bit of 1 and a delay risk bit of 0, it calculates the conflict risk value. = If the risk value reaches the locking threshold, the cache locking mechanism is immediately activated and a locking delay of 10ms is recorded. During this delay, the central controller will not send new action instructions to the node, ensuring that the cached data remains intact and secure before the neighborhood negotiation request is generated, improving the online conflict resolution effect and significantly improving control stability. S6.3: Obtain gradient sensitivity index and action entropy data as negotiation request parameters; normalize the gradient sensitivity index to adapt to communication protocol requirements; quantize and encode the action entropy value to compress data transmission volume; generate a negotiation request data packet based on the normalized index and quantized entropy value; output a neighborhood negotiation request message containing gradient sensitivity index and action entropy. S6.4: Perform integrity verification on the neighborhood negotiation request message to ensure data accuracy; calculate the k-nearest neighbor list based on the node's physical coordinates to determine the broadcast range; sort the k-nearest neighbor list to prioritize nodes with high connection reliability; generate broadcast control commands to configure the wireless communication module; output broadcast commands to send negotiation requests. S6.5: Start the wireless communication module based on the broadcast command; perform modulation and coding processing on the negotiation request message to adapt to the characteristics of the wireless channel; perform message broadcast operation to send the request to the k nearest neighbor nodes; monitor the broadcast confirmation signal to verify that the message was successfully sent; output the broadcast completion signal as the basis for maintaining the local buffer mode; The input parameters of the broadcast control command are parsed, and the wireless communication module start-up control method (parameters: module operating frequency band, transmit power, modulation and coding scheme) is adopted to realize the initialization and activation function of the module's radio frequency front-end and baseband processing unit. Furthermore, through modulation and coding processing methods (parameters: QPSK modulation scheme, forward error correction coding rate 3 / 4, symbol rate setting), the data field of the neighborhood negotiation request message is mapped to the modulation symbol stream, and the channel coding mechanism is used to enhance the transmission robustness in complex terrain scattering environments, generating a coded symbol sequence adapted to the characteristics of the wireless channel. Furthermore, by using a wireless channel frame encapsulation method (parameters: frame synchronization sequence length, frame header control field format), the encoded symbol stream is embedded into a frame structure that conforms to the communication protocol, and the channel identifier and target node address information specified by the broadcast control command are injected to obtain a complete broadcast frame data packet; Furthermore, the radio frequency transmission process of the broadcast frame data packet is realized through the message broadcast operation execution method (parameters: k nearest neighbor node list, antenna directivity control parameters), and the antenna pointing is adaptively adjusted according to the physical coordinates of the node to improve the received signal strength within the target coverage area; Furthermore, by using a broadcast confirmation signal monitoring method (parameters: confirmation timeout time, retransmission number threshold), the confirmation signal returned by the target node is compared and verified. After successful reception, a broadcast completion signal is output as a reliable basis for maintaining the local cache mode. Through the modulation and coding processing and broadcast control command execution of the wireless communication module, the neighborhood negotiation request generated in the previous step is transformed into wireless broadcast data that can be stably distributed within a multi-node range, thereby achieving low-latency information synchronization between nodes with high conduction-low response risk and their neighbors. For example, in a mountain photovoltaic power station scenario, when the edge controller receives a broadcast control command, it configures the wireless communication module to operate at a frequency of 2.4 GHz, a transmit power of 100 mW, a modulation scheme of QPSK, a forward error correction coding rate of 3 / 4, a symbol rate of 250 kSym / s, and encapsulates the neighborhood negotiation request data packet using a frame structure with a synchronization sequence length of 64 bits. During the modulation and coding process, the gradient sensitivity index is normalized. and quantized action entropy Each data segment is encoded into a fixed-length binary field and embedded within the data segment specified by the frame header control field. A directional adjustable antenna is used to limit the coverage area to a Euclidean distance of [missing information]. There are k=5 neighboring nodes within a meter. During the broadcast, the channel acknowledgment timeout is set to... milliseconds, retransmission threshold is Finally, all five neighboring nodes returned acknowledgment signals, and the broadcast completion signal was received. Generate within milliseconds to ensure the local caching mode remains effective and the neighborhood negotiation mechanism is triggered in a timely manner, providing accurate information without packet loss for the graph attention model input in the subsequent S7 steps.

[0015] Step S7: In response to the neighborhood negotiation request, based on the gradient sensitivity index and action entropy of the k nearest neighbor nodes, a locally consistent action correction proposal set is generated using a lightweight graph attention model. Specifically, this includes: S7.1: A lightweight graph attention model architecture is defined and processed for the 3D irradiance gradient tensor and node action entropy of photovoltaic support nodes. The lightweight graph attention model is an efficient graph neural network for edge computing scenarios. The input consists of a 5-dimensional feature vector composed of three components: gradient sensitivity index, action entropy, and 3D irradiance gradient tensor. The model contains two attention calculation layers. Each layer first performs a low-rank projection transformation on the node features, and then calculates the attention weights through dot product and Softmax, which are used for weighted aggregation of neighborhood features. The output layer generates action correction vectors for lifting angle, horizontal rotation angle, and translation distance through a fixed-coefficient linear transformation. This model requires offline training: a transfer learning strategy is adopted, first pre-training using historical irradiance-action datasets, and then fine-tuning online using local power plant data. The low-rank decomposition matrix and output layer coefficients are optimized using MSE as the loss function. After training, the parameters are fixed and deployed to the edge controller. During inference, only forward computation is performed, without the need for online training, thereby achieving low-latency real-time inference at the edge to generate lightweight graph attention model instances suitable for edge controllers. S7.2: Based on the neighborhood negotiation request triggered in S6, the gradient sensitivity index and action entropy data contained in the request are parsed and processed to extract the gradient sensitivity index set and action entropy set of the k nearest neighbor nodes, so as to form the input dataset for local collaborative decision-making. S7.3: Using the constructed lightweight graph attention model instance, the node feature vector dot product calculation is performed on the set of gradient sensitivity indices and action entropy of the k nearest neighbor nodes extracted in S7.2. The attention weight distribution between nodes is generated by softmax normalization to quantify the contribution of each neighbor node in the local correction. S7.4: Based on the generated attention weight distribution among nodes, perform weighted aggregation processing on the action instruction suggestions of the k nearest neighbor nodes, calculate the weighted average action instruction vector, and generate a locally consistent action correction suggestion set; S7.5: Perform consistency verification on the generated locally consistent action correction proposal set. By calculating the Euclidean distance of each action instruction in the proposal set and comparing it with the preset threshold, confirm that the proposal set meets the local consistency condition, and output the final action correction proposal set. Based on the locally consistent action correction proposal set generated by S7.4, the Euclidean distance calculation method (parameters: action command vector dimension d, element value of each proposal vector) is used to realize the distance measurement between each action command in the proposal set, so as to quantify the consistency level between them. Furthermore, using the Euclidean distance formula Implement the distance calculation between two action instruction suggestion vectors, where and Let p and q represent the magnitude parameters of the i-th dimension, respectively. Furthermore, a distance matrix is ​​generated using the distance calculation results between all proposals, and a threshold comparison algorithm is used (parameter: preset consistency threshold). To achieve consistency determination of the suggestion set, the distance between any pair of action suggestions is compared with... Compare the results to obtain a Boolean decision matrix, and mark whether the consistency condition is satisfied. Furthermore, by applying row-by-row logic and aggregation processing methods of Boolean matrices, overall consistency verification of the proposal set is achieved, i.e., only when all proposal pairs satisfy the condition that the distance is less than 1 / 2... When the condition is met, the overall consistency is determined and the suggestion set is marked as the final valid set; Through consistency verification, the local consistent action correction proposal set from the previous step is transformed into the final action correction proposal set, thereby ensuring the tightness of the action schemes of each node after local negotiation in the vector space, so as to improve the response stability of collaborative control. For example, in a collaborative light-tracking scenario for a mountain photovoltaic power station, the number of k-nearest neighbor nodes is set to 5, the dimension d of the action command vector is 3 (corresponding to the lifting angle, horizontal rotation angle, and translation distance, respectively), the range of the correction suggestion vector elements for each node is 0 to 10, and a preset consistency threshold is set. Set to 1.5. For the five action correction suggestions obtained in S7.4, the distance between each pair of suggestions is calculated using the aforementioned Euclidean distance formula, resulting in a 5×5 distance matrix. The maximum distance is 1.32, and the minimum distance is 0.54, all less than the threshold. The overall consistency is determined by Boolean matrix aggregation. The average value of the final output suggestion set in the three dimensions of lifting angle, horizontal rotation angle, and translation distance is used as the final action correction suggestion vector. After confirmation by the S8 consistency voting process, it overrides the original control command and drives the support to complete the synchronous light-tracking action. Field tests show that the response deviation between nodes is significantly reduced, and the cooperative control maintains stable output power under rapid cloud and shadow changes.

[0016] Step S8: Perform a consensus vote on the action correction suggestion set, output a unique correction instruction that overrides the original control instruction, and complete the local resolution of node action instruction conflicts. Specifically, this includes: S8.1: Based on the confidence weight of each correction suggestion in the action correction suggestion set, the confidence weight is calculated by the gradient sensitivity index of the k nearest neighbor node and the action entropy through a preset weight function. The confidence weight calculation formula is: weight = gradient sensitivity index × (1 Action entropy), the confidence of action correction suggestions is positively correlated with gradient sensitivity index and negatively correlated with action entropy. Consistency voting rule construction processing is performed to generate voting judgment threshold and voting logic. The voting judgment threshold is set to a weight ratio of more than 50%, and the voting logic is weighted majority voting to quantify the reliability of each correction suggestion and use it as the basis for decision-making in subsequent voting operations. S8.2: Apply the consensus voting rule to the action correction suggestion set and perform weighted majority voting algorithm processing. First, calculate the total weight of each action correction suggestion and select the suggestion with the highest total weight as the candidate result. If there are multiple suggestions with the same weight, select the suggestion with the smallest Euclidean distance between the vectors to filter out the unique correction instruction that meets the local consensus. The unique correction instruction represents the optimal action scheme after negotiation among the neighboring nodes and eliminates the risk of instruction conflict. S8.3: Perform instruction difference comparison processing between the unique correction instruction and the original control instruction. When an inconsistency is detected in the action type or amplitude range, perform an instruction overwrite operation to update the node control instruction to the corrected action instruction and ensure the continuity of the instruction execution sequence. S8.4: Based on the updated control instructions, trigger the node drive unit to execute the corrected action instructions and receive the execution result feedback signal to confirm whether the instruction execution status and action response delay meet the system robustness requirements; S8.5: Perform conflict resolution verification on the execution result feedback signal, confirm that the node action command conflict has been resolved locally and the system response consistency has been improved, output a command resolution completion flag and end the current collaborative control cycle.

[0017] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0018] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain, characterized in that, Includes the following steps: S1: Obtain a real-time irradiance distribution map of the area covered by the photovoltaic support group; S2: Based on the real-time irradiance distribution map, construct a spatial adjacency graph with the physical coordinates of each photovoltaic support as the graph node, calculate the irradiance intensity difference vector of each node in the 8-neighborhood, and generate a three-dimensional irradiance gradient tensor. S3: Perform gradient propagation weighted aggregation on the three-dimensional irradiation gradient tensor and output a node-level gradient sensitivity index; S4: Calculate the node action entropy based on the action command type, amplitude range, and execution result feedback of each node in the most recent N control cycles; S5: Determine whether each node simultaneously satisfies the condition that the gradient sensitivity index is greater than the gradient sensitivity threshold and the node action entropy is less than the action entropy threshold. If the condition is met, the node is determined to be in a high transmission-low response risk state. S6: For nodes in the high transmission-low response risk state, trigger the local caching mode and generate a neighborhood negotiation request; S7: In response to the neighborhood negotiation request, based on the gradient sensitivity index and action entropy of the k nearest neighbor nodes, generate a locally consistent action correction suggestion set through a graph attention model; S8: Perform a consensus vote on the action correction suggestion set, output a unique correction instruction and overwrite the original control instruction to complete the local resolution of node action instruction conflicts.

2. The method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain as described in claim 1, characterized in that, The acquisition of the real-time irradiance distribution map of the photovoltaic support group coverage area specifically includes: The raw visible light image sequence acquired by the sky imager and the downlink real-time cloud image data received from meteorological satellites are calibrated, linearized, registered, multi-source fused, and spatially interpolated, respectively, and finally output a real-time irradiance distribution map with a spatial resolution of 10 cm / pixel.

3. The method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain as described in claim 1, characterized in that, The three-dimensional irradiation gradient tensor is composed of the horizontal irradiation difference in the x and y directions of the node and the curvature approximation in the z direction.

4. The method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain as described in claim 1, characterized in that, Step S3 specifically includes: The Euclidean distance of each adjacent edge in the spatial adjacency graph is calculated using the inverse distance weight coefficient to generate the weight basis for gradient propagation weighted aggregation. Based on the spatial adjacency graph, the set of 8 neighboring nodes for each photovoltaic support node is determined, and the three-dimensional irradiance gradient tensor of the node and its 8 neighboring nodes is extracted. The reciprocal distance weighting coefficient is applied to the three-dimensional irradiance gradient tensor to perform a weighted average aggregation operation, generating a node-level aggregated gradient vector; Perform Euclidean norm calculation on the node-level aggregated gradient vector to obtain a preliminary value of the node gradient sensitivity. The initial values ​​of the node gradient sensitivity are mapped to a unified interval through linear normalization, and the node-level gradient sensitivity index is output.

5. The method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain as described in claim 4, characterized in that, The node-level gradient sensitivity index characterizes the transmission location attribute of the node in the irradiation gradient field.

6. The method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain as described in claim 1, characterized in that, Step S4 specifically includes: A sliding window time series analysis was performed on the set of action command types, amplitude range division criteria, and execution result feedback categories of photovoltaic support nodes to construct a conditional probability distribution model of node actions. Based on the node action conditional probability distribution model, the edge controller local storage unit data acquisition operation is performed on the specified photovoltaic support node identifier to extract the action command type, amplitude range and execution result feedback data of the most recent N control cycles, forming a node historical action record sequence. The sliding window statistical mechanism is applied to the node historical action record sequence to calculate the node action conditional probability distribution for each type of action instruction in a given state; Perform a negative logarithmic summation operation on the conditional probability distribution of the node actions to calculate the node action entropy value; The node action entropy value is verified and standardized to output the standardized node action entropy.

7. The method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain as described in claim 6, characterized in that, The standardized node action entropy serves as a quantitative measure of the determinism of node response.

8. The method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain as described in claim 1, characterized in that, Step S5 specifically includes: Perform real-time data acquisition processing on the node-level gradient sensitivity index stored in the edge controller to obtain the current value of the node-level gradient sensitivity index; Based on the current value of the node-level gradient sensitivity index, a comparison operation is performed with the gradient sensitivity threshold to generate a Boolean judgment result that the gradient sensitivity index is greater than the gradient sensitivity threshold. Perform real-time data acquisition processing on the standardized node action entropy stored in the edge controller to obtain the current value of the standardized node action entropy; Based on the current value of the standardized node action entropy, a comparison operation is performed with the action entropy threshold to generate a Boolean judgment result that the node action entropy is less than the action entropy threshold. Perform a logical AND operation on the Boolean decision result where the gradient sensitivity index is greater than the gradient sensitivity threshold and the Boolean decision result where the node action entropy is less than the action entropy threshold, and output a high conduction-low response risk state determination signal.

9. The method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain as described in claim 8, characterized in that, The gradient sensitivity threshold is based on the early topographic survey data of the photovoltaic power station, the statistical results of historical irradiance field distribution, and the preset calibration of the support array arrangement parameters. The action entropy threshold is based on the historical action execution data of the photovoltaic support drive mechanism, the control cycle response characteristics, and the preset calibration of historical command conflict records.

10. The method for autonomous optimization and collaborative control of photovoltaic support group morphology in complex terrain according to claim 1, characterized in that, The graph attention model is constructed as follows: the three-dimensional irradiance gradient tensor and node action entropy of the photovoltaic support node are processed by defining the graph attention model architecture, and the model input layer and attention calculation layer are constructed based on the action sensitivity features on the gradient propagation path and the deterministic measure of node response, thereby generating a graph attention model instance.

Images