Equipment anomaly detection method and photovoltaic power generation equipment detection system

Abstract

The invention relates to the technical field of photovoltaic power generation fault identification, in particular to an equipment anomaly detection method and a photovoltaic power generation equipment detection system.The equipment anomaly detection method comprises the steps that convolution processing is carried out on the power offset ratio of each node in a path through a graph convolution neural network, and adjacent conduction is carried out on difference values in combination with the upstream and downstream connection sequence of the nodes; a structure with a continuous power increase trend in a path has a cross-node continuity expression capability, a multi-channel label classification structure is constructed by taking nodes as granularity in kick synchronization judgment through a random forest, and synchronous high-variable unit screening is completed through high-frequency channel category combination and variation amplitude sorting. According to the method, sudden change synchronism judgment breaks through a traditional three-channel single-variable difference threshold mode, and fault separability, response granularity and type description definition in a complex scene are improved through progressive increase aggregation of an offset trend between serial number structures, superposition of a recovery period and offset rate statistics of a reference shielding curve.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation fault identification, and in particular to an equipment anomaly detection method and a photovoltaic power generation equipment detection system. Background Art

[0002] The field of photovoltaic power generation fault identification technology aims to timely identify and locate abnormal power generation units by analyzing data on the electrical parameters of the operating status of various equipment in the photovoltaic system, thereby improving system operation reliability, power generation efficiency and operation and maintenance response speed, and achieving early detection and accurate positioning of typical faults such as component attenuation, shading, electrical faults, and connection abnormalities.

[0003] A photovoltaic power generation equipment detection system is designed to automatically analyze the operating status of photovoltaic power generation equipment and identify abnormal units, identify possible faults in components or sub-arrays, including shading, attenuation, connection failure, etc., and output clear positioning information. The purpose is to improve the system's power generation efficiency, reduce maintenance costs, and ensure operational safety. Through quantitative efficiency evaluation or comparative analysis, it provides executable diagnostic results to support operation and maintenance decisions.

[0004] The existing anomaly identification mode is centered on the data of a single node or component, ignoring the connectivity relationship between nodes in the topological structure and the path trend of electrical parameter changes. It is unable to establish a power disturbance conduction chain between multiple nodes, resulting in the lack of recognition of path-level anomaly diffusion characteristics and the inability to evaluate the joint emergence of multi-channel jumps within the time window, resulting in misjudgments and missed detections. In addition, the occlusion judgment process relies on power drop characteristics or efficiency evaluation indicators, and lacks modeling of the occlusion offset caused by the angle between the bracket orientation angle and the sunlight incidence angle. It is difficult to achieve geometric restoration of regional occlusion causes, which in turn limits the structural classification capability of occlusion types. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a device abnormality detection method and a photovoltaic power generation equipment detection system.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solution: A photovoltaic power generation equipment detection system includes: Electrical topology construction module: Based on the established PV module number table, combiner box connection records, and inverter wiring files, it performs number matching, power difference calculation, and voltage difference judgment, constructs directed edge connection pairs and edge weight sequences, and generates an electrical directed connection graph; Abnormal response aggregation module: Based on the power mutation trend of nodes in the electrical directed connection graph, it performs jump point synchronization detection, adjacent path identification and synchronization concentration screening to obtain mutation starting point combinations with high distribution frequency and establish a candidate set of abnormal propagation starting points; Abnormal path identification module: Based on the candidate set of abnormal propagation starting points, a graph convolutional neural network is used to extract the power offset value ratio of nodes in the path, count the continuously increasing paths, determine the path with the maximum total offset, form a hop path mapping, and generate a power offset propagation chain group; Perturbation trigger judgment module: Based on the power offset propagation chain group, it extracts the multi-channel sudden jump values ​​of voltage, current and temperature rise, uses random forest to determine the combination of multi-channel synchronous jump nodes in the same period, screens the points with high index change rate, and establishes a multi-channel abnormal sudden jump identification map; Occlusion morphology deconstruction module: Based on the multi-channel abnormal jump recognition map, the power recovery period and the occlusion reference curve are called, the occlusion offset ratio is calculated in combination with the bracket orientation angle and the incident angle, and an occlusion category area distribution map is established.

[0007] As a further solution of the present invention, the electrical topology construction module includes: Number verification submodule: Based on the established PV module number table, combiner box connection records, and inverter wiring files, the module numbers are matched item by item with the combiner box records, and unmatched numbers are screened out. The module numbers are cross-checked item by item with the inverter files, and the corresponding positions of the numbers are marked. The power output difference between the module and the combiner box is calculated, and a difference matrix is ​​generated. The power output difference between the combiner box and the inverter is calculated, and the difference variation range is recorded. The numerical difference between the module output voltage and the inverter access voltage is measured, and the voltage offset range is confirmed to obtain the electrical parameter difference mapping matrix. Connection generation submodule: Based on the electrical parameter difference mapping matrix, the power difference between the component and the combiner box is compared with the set interval, and the connection combinations with abnormal differences are eliminated. The voltage difference between the combiner box and the inverter is compared with the boundary of the allowed interval, and the connection pairs with qualified voltage differences are screened. The triplet number combination of the photovoltaic component, combiner box and inverter node is generated, and the connection edges with consistent directions are constructed. The power difference, voltage difference and number position sequence are extracted, and the numerical value set corresponding to the edge is generated to obtain the edge connection attribute set; Graph construction submodule: Based on the edge connection attribute set, a directed chain of numbering order from component nodes to inverter nodes is constructed, and the node numbering path is connected and constructed, and the edge connection weights are marked in sequence and associated with the corresponding node numbers in the graph. The outgoing and incoming edge structures of all nodes are organized, and the directionality and node type are distinguished and marked to generate an electrical directed connection graph.

[0008] As a further solution of the present invention, the abnormal response aggregation module includes: Power mutation identification submodule: Based on the electrical directed connection graph, the power output value of each node at each moment is extracted time-period by time period, and a time series power array is constructed. The adjacent differences of the node power output values ​​are calculated, and the positions of sudden increases and decreases in power jumps are located. The difference of the jump time of multiple nodes is cross-searched, and the set of synchronous nodes whose change time difference is less than the set window value is marked to obtain the set of synchronous mutation nodes. Path aggregation and extraction submodule: Based on the synchronous mutation node set, it traces the path of each synchronous node in the electrical directed connection graph and extracts the path sequence from the start point to the end point. It performs position index statistics of each jump node number in the path sequence, records the hop count difference and connection level between the numbers, performs statistical counting of the starting point numbers in all paths, and marks the path combination with the highest number of repetitions to obtain a centralized mutation path mapping table. Starting point set generation submodule: Based on the centralized mutation path mapping table, the total frequency of occurrence of the starting numbers of all paths is sorted, and a mapping set of numbers and corresponding frequencies is generated. The top numbers after frequency sorting are screened, and the number combinations with the highest frequencies are intercepted. A candidate node number set is constructed and output in the original order to establish a candidate set of abnormal propagation starting points.

[0009] As a further solution of the present invention, the abnormal path identification module includes: Offset extraction submodule: Based on the candidate set of abnormal propagation starting points, a graph convolutional neural network is used to sequentially extract the numbers of all nodes in the starting point number path, and the power data of each node in the path is read node by node. The difference in power values ​​of the previous and next nodes is calculated and the ratio is recorded. A list of power offset ratios between nodes is constructed and a continuous increasing sequence is screened. The set of paths that meet the monotonicity requirement is extracted to generate a power increasing path sequence. Path construction submodule: Based on the power-increasing path sequence, the node and its offset ratio in each path within the path number are jointly identified and constructed, the paths are sorted from high to low according to the cumulative offset value of the path, and the corresponding path numbers are recorded. The path set with the largest offset value is extracted and the hop count is counted. A path sequence index and hop count matching table is generated, and a hop count path mapping set is generated. Chain group generation submodule: Based on the hop path mapping set, the node numbers in each path are connected in sequence to form a directed node chain group, and a connection structure with consistent direction between nodes is constructed chain by chain. A path topology set with power increasing as the connection condition is generated, and the data organization and numbering of the chain group set are completed to generate a power offset propagation chain group.

[0010] As a further solution of the present invention, the graph convolutional neural network is based on the formula: in: Represents graph convolution Nodes in the layer The power offset characteristics, represents a nonlinear activation function, Representation and Node Adjacent nodes , Representation node With node The power difference slope change rate of the path segment is Representation node To Node The frequency ratio of the path segment in the historical anomaly propagation, Representation node The power adjustment value of the degree Representation node The power adjustment value of the degree, Represents graph convolution The weight matrix of the layer, Represents graph convolution Nodes in the layer The power ratio characteristic input, Representation node The power fluctuation factor, Representation node The frequency ratio of multiple power increasing paths.

[0011] As a further solution of the present invention, the perturbation trigger determination module includes: The outburst extraction submodule extracts the voltage, current, and temperature channel data of each node in each chain based on the power offset propagation chain group and aligns them by time. It compares the difference between the channel values ​​at the same node and the same time point and the values ​​at adjacent moments, marks the time points that are greater than the threshold, and collects statistics on channel types and outburst amplitudes to generate a multi-channel outburst dataset. Synchronous judgment submodule: Based on the multi-channel sudden jump data set, random forest is used to count the types of sudden jump channels in each time period and compare whether the three types of channels have common mutations. The node numbers of simultaneous sudden jumps are screened, the channel jump amplitude is extracted and the change ratio per unit time is calculated. The nodes with a change rate lower than the median are screened out, the numbers are retained and a set is generated to generate a set of synchronous high-change nodes; Graph generation submodule: Based on the synchronous high-variable node set, channel mutation type association connections are generated between node pairs, edge-connected node pairs are extracted and the channel names and change rate values ​​are marked, connection edges with strong channel type consistency are organized to form a subgraph, a multi-node sudden jump synchronization relationship graph structure is constructed, and a multi-channel abnormal sudden jump identification graph is established.

[0012] As a further solution of the present invention, the random forest is according to the formula: in: represents the final classification result, represents the total number of decision trees built in the random forest, Indicates the The weight coefficient of each tree, Indicates the The classification output function of the tree, Indicates the The original jump feature vector received by the tree, Represents the adjustment weight coefficient of the channel frequency variation factor, Indicates the channel jump frequency change value, represents the adjustment weight coefficient of the neighborhood synergy factor, represents the sudden jump coordination factor of the neighborhood nodes, represents the adjustment weight coefficient of the energy ratio factor, represents the sudden energy ratio factor, Represents the adjustment weight coefficient of the density factor, represents the outburst event density factor.

[0013] As a further solution of the present invention, the perturbation trigger judgment module is specifically an identification output structure for calibrating multi-channel jump behavior, the jump identification map includes a synchronous jump channel index set, a jump amplitude threshold distribution layer and a jump node period grouping label, the high value point of the abnormal indicator is specifically a combination of voltage nodes, current nodes and temperature rise nodes whose jump amplitude exceeds a preset change rate threshold, and the power offset propagation chain group specifically refers to a channel sequence set that forms a jump response path.

[0014] As a further solution of the present invention, the occlusion form deconstruction module includes: The occlusion response extraction submodule is based on the multi-channel abnormal jump identification map. The numbers of all nodes in the map are read one by one and the path sequence is indexed. The power recovery time series of the corresponding nodes is extracted and the recovery segments are divided into time intervals. The point-by-point difference between each recovery period and the occlusion reference curve is calculated and the relative offset rate is recorded. The cycle length of each node in the path is counted and the paths are classified and numbered. The numbered sets are aggregated according to the path labels to generate the occlusion cycle response set. Angle offset calculation submodule: Based on the shading period response set, the module performs mapping extraction of each path number to the bracket number and obtains the orientation angle parameter. The module extracts the sunlight incident angle sequence of the period corresponding to the node jump time and matches the path number. The module calculates the angle difference between the orientation angle and the incident angle of each path bracket and converts it into an angle ratio value. The module generates a list of path number and offset angle ratio combinations and generates an shading offset ratio table. Regional structure generation submodule: Based on the occlusion offset ratio table, the node position coordinates corresponding to each path number are extracted and mapped to two-dimensional space, the grading range is set according to the offset ratio and the nodes are divided into corresponding grade intervals, the coordinates of nodes of the same grade are aggregated to form regional blocks and the regional boundary numbers are recorded, the numbering structure of each regional block is organized and the spatially adjacent blocks are merged to establish an occlusion category regional distribution map.

[0015] A device anomaly detection method is implemented based on the photovoltaic power generation equipment detection system, comprising the following steps: S1: Based on the PV module number table, combiner box connection records, and inverter wiring files, extract the correspondence between each node number and wiring, match the module number with the wiring number, calculate the power value difference between the number pairs, determine the degree of difference in the corresponding voltage values, establish a directed edge set and edge weight sequence with the number pairs as node connections, and obtain the electrical connection map; S2: Based on the electrical connection map, identify the power change sequence of all nodes within the observation period, select the jump points in the power sequence whose intervals are no more than two sampling periods and whose increase or decrease exceeds a set difference, identify the adjacent path nodes of each jump point, count the frequency of occurrence of the same node in the jump path in multiple paths, select the node combination with an occurrence frequency greater than a specified value, and obtain the abnormal propagation starting point combination set; S3: Based on the abnormal propagation starting point combination set, a graph convolutional neural network is used to construct a number sequence pair for each pair of upstream and downstream nodes in the path, and the ratio of the power value of each node on the path to the power value of its upstream node is calculated to form a ratio matrix indexed by the number pair. The comparison value matrix is ​​weighted superimposed and normalized along the connection direction under the entire graph structure. All number combinations with continuously increasing ratios in the connection path are extracted, the cumulative sum of the ratios in the combination is calculated, and the number jump values ​​are marked to obtain the offset propagation path mapping group; S4: Based on the offset propagation path mapping group, a random forest is used to extract a continuous value sequence of voltage, current, and temperature for each node in the path within the same period. The mutation amplitude values ​​of the three types of sequences in adjacent periods are calculated respectively, and a multidimensional vector set of the three types of mutation amplitudes is established. The multidimensional vectors of all nodes in the same path are clustered according to the path structure. The numbered combinations that simultaneously have the three types of mutation amplitude peak points in multiple paths are identified. The occurrence frequency of each numbered combination and its corresponding maximum amplitude are recorded to obtain a multi-channel synchronous sudden jump node set. S5: Based on the multi-channel synchronous jump node set, the number of cycles required for each node to recover to the previous state is extracted, the power change reference curve of the preset shading type is called, and the offset value is compared point by point with the recovery process curve. Combined with the orientation angle of the component bracket where the numbered node is located and the solar incidence angle in the time period, the ratio of the total offset value of each node to the total offset value of the reference curve is calculated, and a matching table of numbers and shading types is generated to obtain a shading morphology distribution map.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are: 1. This invention uses a graph convolutional neural network to convolve the power offset ratios of each node in a path. This difference is then transferred to adjacent nodes based on the node's upstream and downstream connection order. This allows structures with sustained power growth trends in the path to be expressed as continuous across nodes, improving the path recognition's ability to perceive global relationships in the graph structure. 2. This invention uses random forests to construct a multi-channel label classification structure at the node level for sudden jump synchronization judgment. Synchronous high-variable units are screened by combining high-frequency channel categories and sorting by change amplitude, enabling sudden jump synchronization judgment to break through the traditional three-channel single-variable difference thresholding model. 3. This method uses the incremental aggregation of offset trends between numbering structures to create a sequence-driven hop count mapping of path propagation directions. By superimposing the recovery period and offset rate statistics of the reference occlusion curve, and using the conversion ratio of the angle between the heading angle and the incident angle as a factor in calculating the occlusion shape, local occlusion events are mapped to regional offset groups, improving fault separability, response granularity, and type characterization clarity in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a system flow chart of the present invention; Figure 2 Schematic diagram of the method steps of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0020] See also Figure 1 The present invention provides a technical solution: a photovoltaic power generation equipment detection system comprising: Electrical topology construction module: Based on the established PV module number table, combiner box connection records, and inverter wiring files, it performs number matching, power difference calculation, and voltage difference judgment, constructs directed edge connection pairs and edge weight sequences, and generates an electrical directed connection graph; Abnormal response aggregation module: Based on the node power mutation trend in the electrical directed connection graph, it performs jump point synchronization detection, adjacent path identification, and synchronization concentration screening to obtain mutation starting point combinations with high distribution frequency and establish a candidate set of abnormal propagation starting points; Abnormal path identification module: Based on the candidate set of abnormal propagation starting points, a graph convolutional neural network is used to extract the power offset value ratio of nodes in the path, count the continuously increasing paths, determine the path with the maximum total offset, form a hop path mapping, and generate a power offset propagation chain group; Perturbation trigger judgment module: Based on the power offset propagation chain group, it extracts the multi-channel sudden jump values ​​of voltage, current and temperature rise. It uses random forest to determine the combination of multi-channel synchronous jump nodes in the same period, screens the points with high index change rate, and establishes a multi-channel abnormal sudden jump identification map; Occlusion morphology deconstruction module: Based on the multi-channel abnormal jump recognition map, the power recovery period and occlusion reference curve are called, the occlusion offset ratio is calculated in combination with the bracket orientation angle and the incident angle, and an occlusion category area distribution map is established.

[0021] The electrical topology building blocks include: Number verification submodule: Based on the established PV module number table, combiner box connection records, and inverter wiring files, the module numbers are matched item by item with the combiner box records, and unmatched numbers are screened out. The module numbers are cross-checked item by item with the inverter files, and the corresponding positions of the numbers are marked. The power output difference between the module and the combiner box is calculated, and a difference matrix is ​​generated. The power output difference between the combiner box and the inverter is calculated, and the difference variation range is recorded. The numerical difference between the module output voltage and the inverter access voltage is measured, and the voltage offset range is confirmed to obtain the electrical parameter difference mapping matrix. Connection generation submodule: Based on the electrical parameter difference mapping matrix, the power difference between the module and the combiner box is compared with the set interval, and connection combinations with abnormal differences are eliminated. The voltage difference between the combiner box and the inverter is compared with the boundary of the allowed interval, and connection pairs with voltage differences that meet the conditions are screened. The triplet number combination of the PV module, combiner box and inverter node is generated, and connection edges with consistent directions are constructed. The power difference, voltage difference and number position sequence are extracted, and the numerical value set corresponding to the edge is generated to obtain the edge connection attribute set; Graph construction submodule: Based on the edge connection attribute set, a directed chain is constructed from the component node to the inverter node in the numbering order, and the node number path is connected. The edge connection weights are sequentially labeled and associated with the corresponding node numbers in the graph. The outgoing and incoming edge structures of all nodes are organized, and the directionality and node type are distinguished and labeled to generate an electrical directed connection graph. Number verification submodule: Based on the established PV module number table, combiner box connection record and inverter wiring file, a bidirectional hash mapping method is used to map the PV module number to the number field in the combiner box record one by one. The key-value pair in the hash table with the key being the module number and the value being the combiner box number is used for initial index matching. The combiner box number is used as the reverse key to establish a reverse table structure, perform bidirectional verification and comparison, and filter out unmatched numbers. The Boolean intersection filtering method is used to filter the number consistency of the module number and the access number in the inverter file. A Boolean matrix is ​​generated according to the index position, the intersection Boolean matrix is ​​calculated, and the number corresponding to all matching index bits is marked. For the wiring position, a numerical difference matrix generation method is used to subtract the component output power value from the corresponding combiner box power input value point by point and construct a difference array. The difference array is filled into a two-dimensional matrix according to the number index dimension to generate a power difference matrix. The same operation is used to construct the difference between the combiner box output power value and the inverter input power value. All elements of the matrix are traversed and the difference change range of the corresponding number pairs is recorded. The voltage vector difference comparison method is called. The component terminal voltage measurement sequence and the inverter input voltage sequence are used as the input array. A one-to-one sequence subtraction operation is performed on each node number position. The voltage offset difference vector is output and the maximum and minimum values ​​are counted to obtain the voltage offset range and generate the electrical parameter difference mapping matrix. Connection generation submodule: Based on the electrical parameter difference mapping matrix, the partition interval comparison method is adopted, the set power tolerance interval array is called, and the power difference matrix between the component and the combiner box is compared at the element level. The matrix cells that exceed the tolerance boundary are set as invalid connection flag values ​​and the corresponding connection pairs are eliminated. The boundary condition logic judgment method is used to perform a double-boundary logic comparison operation on the voltage difference sequence between the combiner box and the inverter, where the upper and lower limits are the preset allowable voltage offset value lower limit 8.5 and upper limit 13.2 respectively. A logical masking operation is performed on the connection number combination that does not meet the boundary conditions. The ternary combination generation function module is used to take the photovoltaic component number, combiner box number, and inverter number as triples number one, number two, and number three respectively, and the numbering structure is spliced ​​in tuple format to form a connection number triple that conforms to the direction annotation. The sequence extraction function group is used to extract the power difference, voltage difference, and physical location index number between the component and the inverter from the above triple connection edges in turn, and the three indicators are generated into a floating-point value combination to form an edge attribute index vector group to obtain the edge connection attribute set; Graph construction submodule: Based on the edge connection attribute set, a directed graph structure construction method is adopted. The number triplet is used as the path basis. The starting node is set as the component number and the ending node is set as the inverter number. The number path is defined in the graph structure according to the connection direction. The number mapping chain generation operation is called to establish an array index sequence of the node connection order in the path, and the number path is constructed in sequence. The weight mapping annotation strategy is adopted to annotate the weight values ​​in the edge connection attributes to the connection edges in sequence according to the edge connection sequence, and the corresponding node numbers are recorded for indexing in the graph. The node connection table organization logic is adopted to organize the outgoing node numbers and incoming node numbers of all nodes into two list structures respectively, generate a node direction structure mapping table, and finally traverse according to the node number, set the directional mark value to 0 as input and 1 as output, and splice it with the node type code 0, 1, 2 to generate an electrical directed connection graph.

[0022] The exception response aggregation module includes: Power mutation identification submodule: Based on the electrical directed connection graph, the power output value of each node at each moment is extracted time-period by time period, and a time series power array is constructed. The adjacent differences of the node power output values ​​are calculated, and the positions of sudden increases and decreases in power jumps are located. The difference of the jump time of multiple nodes is cross-searched, and the set of synchronous nodes with change time differences less than the set window value is marked to obtain the set of synchronous mutation nodes. Path aggregation and extraction submodule: Based on the synchronous mutation node set, it traces the path of each synchronous node in the electrical directed connection graph and extracts the path sequence from the start point to the end point. It then performs positional index statistics on the number of each jump node in the path sequence, records the hop count difference and connection level between the numbers, performs statistical counting of the starting point numbers in all paths, and marks the path combination with the highest number of repetitions to obtain a centralized mutation path mapping table. Starting point set generation submodule: Based on the centralized mutation path mapping table, the total frequency of occurrence of the starting numbers of all paths is sorted, and a mapping set of numbers and corresponding frequencies is generated. The top numbers after frequency sorting are screened, and the number combinations with the highest frequencies are intercepted to construct a candidate node number set and output them in the original order to establish a candidate set of abnormal propagation starting points; Power mutation identification submodule: Based on the electrical directed connection graph, the local extreme difference mutation detection algorithm is used to extract the power output value of each node at each moment in the sampling time sequence. A two-dimensional time series matrix is ​​constructed with the node number as the main index and the power value of each sampling point as the horizontal data. The power values ​​are subtracted at adjacent sampling moments, and the absolute value of all the result values ​​is taken. The positions of the elements with a difference greater than the set threshold of 9.7 are screened and recorded as the jump point index set. The jump moments of different nodes in the set are combined and crossed one by one. The synchronous time window matching algorithm is used, and the synchronous judgment time difference threshold is set to 3 sampling periods. The jump time difference between nodes is calculated, and the number combination with a difference less than or equal to 3 is screened. The node group that mutates at the same time is selected as the synchronous mutation node set element. Finally, the synchronous mutation node set is aggregated according to the number combination; Path aggregation extraction submodule: Based on the synchronous mutation node set, a graph path backtracking algorithm is used. Starting from each number in the synchronous mutation node, all reachable paths are traced step by step along the connection direction in the electrical directed connection graph. The maximum path tracing depth is set to 12 layers. The numbers of each layer of connection are recorded in sequence as a path number sequence. The number sequence difference is calculated according to the occurrence position of each pair of jump node numbers in the path. All number sequence differences are uniformly converted to positive numbers to represent the hop count difference. The nodes in the path are assigned values ​​of 0, 1, and 2 according to the three categories of sequence numbers: components, combiner boxes, and inverters. The connection level sequence of the corresponding numbers is recorded. The starting number of each path is subjected to number frequency statistics. The frequency values ​​are compared and sorted according to the starting number table. The starting number combination with the highest frequency is used as the path aggregation key identifier. The corresponding number path combination structure is recorded to generate a centralized mutation path mapping table. Starting point set generation submodule: Based on the centralized mutation path mapping table, the number frequency ranking algorithm is used to extract the frequency value of the starting node number in all paths, count the total number of times each number appears in all paths, sort them from high to low by frequency value, and set the selection threshold to the top 15% number groups in frequency ranking. The position number of the top-ranked node number combinations in the sorting results is summarized and recombined according to the number arrangement order in the path mapping table to construct a number set in a continuous output format. Number pairs are generated for all combination numbers to form a non-repeated candidate number combination table, and finally a candidate set of abnormal propagation starting points is obtained.

[0023] The abnormal path identification module includes: The offset extraction submodule uses a graph convolutional neural network (GCNN) to sequentially extract all node numbers within the path with the starting point number based on the candidate set of anomaly propagation starting points. It also reads the power data of each node within the path node by node, calculates the power difference between the previous and next nodes, and records the ratio. It then constructs a list of power offset ratios between nodes and filters for continuously increasing sequences. It then extracts a set of paths that meet monotonicity requirements and generates a sequence of power-increasing paths. Path construction submodule: Based on the power-increasing path sequence, the node and its offset ratio in each path within the path number are jointly identified and constructed. Paths are sorted from high to low by the cumulative offset value and the corresponding path numbers are recorded. The path set with the largest offset value is extracted and the hop count is counted. A path sequence index and hop count matching table is generated, and a hop count path mapping set is generated. Chain group generation submodule: Based on the hop-count path mapping set, the nodes in each path are sequentially connected to form a directed node chain group. The connection structure with consistent direction between nodes is constructed chain by chain, and a path topology set with increasing power as the connection condition is generated. The data of the chain group set is organized and numbered, and the power offset propagation chain group is generated. Offset extraction submodule: Based on the candidate set of abnormal propagation starting points, a graph convolutional neural network is used to construct a graph structure node index with the path number as the graph structure node index and the edge weight between nodes as the graph convolution input feature. The starting number in the path number sequence is called the first node, and the node number extraction operation is performed in sequence on all paths starting from the starting number in the graph structure. Each number is regarded as a row in the network input feature matrix. The feature dimension is set to power value, the convolution operation step size is 1, and the adjacency matrix is ​​constructed using a Boolean structure in the edge connection attribute matrix where the connection relationship is 1 and the no connection relationship is 0. Graph convolution is performed on all path numbers, and at each node, the difference between the power value of the node and the power value of the upstream node is extracted. All differences are constructed into a one-dimensional difference vector according to the node order. Ratio operation is performed on the power difference of any three consecutive nodes in the difference vector. All ratios are recorded and a ratio increasing sequence is screened out. A path number set is established that satisfies the positive growth relationship between any two ratios in the sequence. A path sequence set whose numbering order is not reversed is extracted, and a path number structure with continuous and consistent power offset direction is generated, generating a power increasing path sequence. Path construction submodule: Based on the power-increasing path sequence, the path incremental accumulation algorithm is used to add the corresponding power ratios of all nodes in each path number sequence node by node, and a list of the total cumulative offset values ​​of the path is constructed. The path number and the total offset value are combined into a two-dimensional index set. The index set is sorted in descending order of the offset value, and the path number corresponding to each sorted position is recorded. The path number group with the highest cumulative offset value is extracted as the high-offset path structure. For each number sequence in the path number structure, the number of node hops between the first and last node numbers is calculated, and the hop value is recorded as the path span indicator. All path number indexes and corresponding hop numbers are formed into a key-value corresponding set. A path extension structure with one-to-one correspondence between path numbers and hop number indexes is established to generate a hop path mapping set. Chain group generation submodule: Based on the hop path mapping set, the node chain topology generation method is adopted. The node numbers in each path number are connected in sequence into directed number pairs, and all directed number pairs are spliced ​​into a complete path chain structure. The chain connection condition is set as the upstream node power value is less than the downstream node power value, the node number increases continuously, and any node is only connected once. The directed edge structure between nodes is constructed for all path chain sets that meet the connection conditions, and three structural attributes are assigned to each directed connection in the graph structure: node position difference, path number index and cumulative offset value. The structure is registered and numbered for each chain, and finally all number chain groups are output in the order of path numbers to generate a numbered chain path set with unidirectional consistent power offset direction, and a power offset propagation chain group is generated.

[0024] Graph convolutional neural network, according to the formula: in: Represents graph convolution Nodes in the layer The power offset characteristics, represents a nonlinear activation function, Representation and Node Adjacent nodes , Representation node With node The power difference slope change rate of the path segment is Representation node To Node The frequency ratio of the path segment in the historical anomaly propagation, Representation node The power adjustment value of the degree Representation node The power adjustment value of the degree, Represents graph convolution The weight matrix of the layer, Represents graph convolution Nodes in the layer The power ratio characteristic input, Representation node The power fluctuation factor, Representation node The frequency ratio of multiple power increasing paths; Execution process: First, collect the historical power data of each node in the path from the photovoltaic equipment node, build a graph structure model through the topological relationship between nodes, and extract each node in turn. Power ratio characteristics , and then calculate its power fluctuation factor , through the node The power value standard deviation on the path is normalized to measure the power stability of the node, and then the node is counted. The number of times it appears in all power-increasing paths is compared with the total number of paths to obtain its path coverage factor. , and then for the current target node Connected path segments, computing node pairs The derivative of the power difference with respect to the time difference gives the power change rate of the edge The path stability factor is obtained by statistically analyzing the frequency of occurrence of path segments in abnormal propagation through the historical fault path database and normalizing it. , then extract the nodes and Connectivity and , and bring in the experience adjustment coefficient , normalized to control the influence bias of high-connectivity nodes in propagation, and finally calculated nodes exist Power offset characteristics in layers , as the structural basis for judging the power increasing path, the candidate path sequence that meets the monotonicity rule is screened out to achieve accurate extraction of abnormal power paths.

[0025] The perturbation trigger determination module includes: The outburst extraction submodule extracts voltage, current, and temperature channel data from each node in each chain based on the power offset propagation chain group and aligns them by time. It compares the difference between the channel values ​​at the same node and time point and the values ​​at adjacent moments, marks the time points where the difference exceeds the threshold, and collects statistics on channel types and outburst amplitudes to generate a multi-channel outburst dataset. Synchronous judgment submodule: Based on a multi-channel sudden jump dataset, a random forest is used to count the types of sudden jump channels in each time period and compare whether the three types of channels have common mutations. The node numbers of simultaneous sudden jumps are screened, the channel jump amplitude is extracted, and the change ratio per unit time is calculated. Nodes with a change rate below the median are screened out, and the numbers are retained and generated into a set to generate a set of synchronous high-change nodes. Graph generation submodule: Based on the synchronous high-variability node set, it generates channel mutation type association connections between node pairs, extracts edge-connected node pairs and labels the channel names and change rate values, organizes connected edges with strong channel type consistency into a subgraph, constructs a multi-node sudden jump synchronization relationship graph structure, and establishes a multi-channel abnormal sudden jump identification graph; Jump extraction submodule: Based on the power offset propagation chain group, the channel timing difference extraction method is adopted to extract the original sequence data of the three channels of voltage, current and temperature values ​​corresponding to the nodes in each chain in the order of number. Each type of channel data is aligned according to the node number and the time axis structure is established with the sampling period as the horizontal axis. The length of each channel data is set to a fixed value of 120 and the sampling frequency is 5 seconds. The three types of channel values ​​of each node at the same time point are calculated with the difference between the two adjacent sampling points in the channel to form the difference between the current time point and the previous value, the current time The difference between the point value and the subsequent value is two difference vectors. The three types of channel differences are stored in a matrix according to the node number. Each value in the matrix is ​​compared with the channel jump setting threshold. The voltage setting threshold is 3.4, the current is 0.85, and the temperature rise is 1.3. The time point position greater than the threshold is marked, and the field group consisting of the marked channel type, power value, and time index position is stored in the multi-channel difference identification table. All jump records for each node number are counted, and a multidimensional array consisting of channel type, jump amplitude, and jump position is output to generate a multi-channel jump data set; Synchronous judgment submodule: Based on the multi-channel outburst dataset, random forest is used to construct a sample set of node channel outburst conditions corresponding to each time period. Each sample contains channel number, outburst amplitude, outburst time index, node number and channel type label. The training feature items are set as channel type and amplitude value. 100 subtrees are used to build the model and the maximum tree depth is set to 6. A Boolean judgment is performed on whether the three channel types of each node are simultaneously marked as outbursts at each sampling time point. The node numbers of all simultaneous outbursts are recorded and the types of outburst channels are marked. The value corresponding to the outburst amplitude is extracted, and the change ratio is calculated by dividing the outburst amplitude by the time interval between the two sampling points. The ratio values ​​of the three types of channels are stored in the channel ratio vector table respectively. The median value of the ratio is calculated according to the channel type. The records with the channel ratio of each node number less than the corresponding median value are set to invalid and the numbers are eliminated. Only the node numbers that meet the conditions of all three channels outburst and the change rate is higher than the median value are retained, and finally a synchronous high-variable node set is generated; Graph generation submodule: Based on a synchronous high-variable node set, the channel type mapping group construction method is adopted to extract number pairs from the node set in sequence, and the connection structure between nodes is generated pair by pair. The connection generation condition is set as the existence of a directed connection relationship between the numbers in the chain group structure, and the channel jump occurrence time index interval does not exceed two sampling cycles. The channel type intersection extraction operation is performed on the number pairs that meet the conditions, and the average jump change rate of each channel type in the intersection is calculated. The consistency judgment threshold is set to a jump ratio deviation of no more than 15%. All node connection structures that meet the conditions are formed into an edge set, and the edge attribute field is constructed with the channel name and the average change rate. A subgraph is established according to the connection direction. The consistency of all edge channel types in the subgraph is stronger than the threshold value. Finally, the subgraph number index, channel type group, and change rate field are organized into a structural graph to generate a multi-channel abnormal jump identification graph.

[0026] Random forest, according to the formula: in: represents the final classification result, represents the total number of decision trees built in the random forest, Indicates the The weight coefficient of each tree, Indicates the The classification output function of the tree, Indicates the The original jump feature vector received by the tree, Represents the adjustment weight coefficient of the channel frequency variation factor, Indicates the channel jump frequency change value, represents the adjustment weight coefficient of the neighborhood synergy factor, represents the sudden jump coordination factor of the neighborhood nodes, represents the adjustment weight coefficient of the energy ratio factor, represents the sudden energy ratio factor, Represents the adjustment weight coefficient of the density factor, represents the outburst event density factor; Execution process: First, collect multi-channel sudden jump data of photovoltaic equipment and divide it into time windows to form the basic input feature vector , including the jump amplitude, occurrence time, channel type code and node number of each channel during the sudden jump, and then calculate the extended characteristic parameters, extract the change value of the sudden jump number per unit time to construct the frequency change index , calculate the proportion of synchronous sudden jump nodes within three hops of the sudden jump node to generate the neighborhood coordination factor , the jump energy ratio is obtained by integrating the jump power curve and dividing it by the total energy of the node period , calculate the density of similar sudden jump events in the current time period to obtain the density factor , and then multiply the above extended features by the adjustment weight coefficient With the original feature vector Make linear combinations to form enhanced feature inputs and input them into the forest separately Each tree is weighted according to its classification accuracy in the independent validation set. , output the result based on the weighted voting average of all trees , judge whether there is synchronous sudden jump behavior of the three types of channels within the time period. If so, retain the corresponding node number, and further screen out the nodes with a jump ratio higher than the median value per unit time, and finally establish a synchronous high-variable point set.

[0027] The perturbation trigger judgment module is specifically an identification output structure used to calibrate multi-channel jump behavior. The jump identification map includes a set of synchronous jump channel indexes, a jump amplitude threshold distribution layer, and a jump node period grouping label. The high-value points of the abnormal indicators are specifically the combination of voltage nodes, current nodes, and temperature rise nodes whose jump amplitudes exceed the preset change rate threshold. The power offset propagation chain group specifically refers to the set of channel sequences that form the jump response path.

[0028] The occlusion morphology deconstruction module includes: The occlusion response extraction submodule uses a multi-channel abnormal jump recognition map to read the numbers of all nodes in the map one by one and index the path sequence. The power recovery time series of the corresponding nodes is extracted and the recovery segments are divided into time intervals. The point-by-point difference between each recovery period and the occlusion reference curve is calculated and the relative offset rate is recorded. The cycle length of each node in the path is counted and the paths are classified and numbered. The numbered sets are aggregated by path labels to generate the occlusion cycle response set. Angle offset calculation submodule: Based on the shading cycle response set, the mapping of each path number to the bracket number is extracted and the orientation angle parameters are obtained. The sunlight incident angle sequence of the period corresponding to the node jump time is extracted and matched with the path number. The angle difference between the orientation angle and the incident angle of each path bracket is calculated and converted into an angle ratio value. A list of path number and offset angle ratio combinations is generated, and a shading offset ratio table is generated. The regional structure generation submodule extracts the node position coordinates corresponding to each path number based on the occlusion offset ratio table and maps them to two-dimensional space. The grading range is set according to the offset ratio and the nodes are divided into corresponding grade intervals. The coordinates of nodes of the same grade are aggregated to form regional blocks and the regional boundary numbers are recorded. The number structure of each regional block is organized and spatially adjacent blocks are merged to establish the occlusion category regional distribution map. Occlusion response extraction submodule: Based on the multi-channel abnormal jump recognition map, the power recovery difference matching method is adopted to read each node number in the map structure one by one, and a numbered path list is established in sequence according to the path number index order. The power recovery time series data of each numbered node is extracted, and the sampling interval is set to 5 seconds. The starting point of the recovery period is set to the first rising point after the jump point, and the end point is set to the last sampling point where the power fluctuation in the stable segment is not greater than 0.5. The power recovery segment is divided according to the above time period, and a point-by-point difference operation is performed on the data of each recovery segment and the preset occlusion reference curve. The actual power value corresponding to each time point is subtracted from the reference curve value, divided by the reference value and converted into a percentage form, recorded as a relative offset rate. The length of all recovery periods in the path is stored as an array according to the node number. The node numbers in the path are grouped according to the recovery segment length, aggregated in number order and a label field is added to each path. A data structure set consisting of each path number and the number list is output to generate an occlusion period response set. Angle offset calculation submodule: Based on the occlusion period response set, the angle calculation method is adopted to extract the bracket direction parameters for each bracket number corresponding to each number in the path number index set. The direction angle data uses true north as the reference, the angle unit is degree, and the range is 0 to 360 degrees. A direction angle vector list is constructed in the order of the number. For the time period of each numbered jump time point, the sunshine incident angle time series is extracted. The time series range is set to 30 minutes before and after the jump point, and sampling is performed every 5 minutes. For each time point, the angle difference between the orientation angle corresponding to all numbers in the path number and the sunshine incident angle is extracted. The difference is the difference between the absolute values ​​of the two angles. If it exceeds 180 degrees, 360 minus the difference is taken. For each node, the angle offset ratio is constructed by dividing the difference by 90, retaining four decimal places of precision, and establishing a two-dimensional list of one-to-one mapping between path numbers and offset ratios. The number index table and the offset ratio value comparison structure are output to generate an occlusion offset ratio table. Regional structure generation submodule: Based on the occlusion offset ratio table, a spatial hierarchical clustering method is used to obtain the geographical coordinates of the components of each numbered node in the path numbering structure. The coordinate format is a two-dimensional plane rectangular coordinate with a unit of meter. A position coordinate list is constructed in the order of numbering. A four-level classification interval range is set for each offset ratio, and the interval boundary values ​​are set to 0.00 to 0.25, 0.26 to 0.50, 0.51 to 0.75, and 0.76 to 1.00. Each node is assigned to the corresponding level interval according to the comparison result between the node number and the corresponding ratio. Number aggregation is performed on the node coordinate set in the same interval, and the aggregated number list is recorded to construct a regional boundary number set. The regional boundary number is the boundary area identifier formed by the minimum and maximum numbers in each group of nodes. An index table is established for the numbering structure in all level areas, and spatial adjacent nodes are merged. The merging condition is that the difference between the horizontal and vertical coordinates of the two node coordinates does not exceed 1.8 meters. The merged result is output as a number block index to generate an occlusion category regional distribution map.

[0029] See also Figure 2 A device abnormality detection method is provided. The device abnormality detection method is performed based on the photovoltaic power generation equipment detection system, and includes the following steps: S1: Based on the PV module number table, combiner box connection records, and inverter wiring files, extract the correspondence between each node number and wiring, match the module number with the wiring number, calculate the power value difference between the number pairs, determine the degree of difference in the corresponding voltage values, establish a directed edge set and edge weight sequence with the number pairs as node connections, and obtain the electrical connection map; S2: Based on the electrical connection graph, identify the power change sequence of all nodes within the observation period, select the jump points in the power sequence whose interval is no more than two sampling periods and whose increase or decrease exceeds the set difference, identify the adjacent path nodes of each jump point, count the frequency of occurrence of the same node in multiple paths in the jump path, select the node combination with an occurrence frequency greater than the specified value, and obtain the combination set of abnormal propagation starting points; S3: Based on the set of abnormal propagation starting point combinations, a graph convolutional neural network is used to construct the number sequence pairs of each pair of upstream and downstream nodes in the path. The ratio of the power value of each node on the path to the power value of its upstream node is calculated to form a ratio matrix indexed by the number pair. The comparison value matrix is ​​weighted and normalized along the connection direction under the entire graph structure. All number combinations with continuously increasing ratios in the connection path are extracted, the cumulative sum of the ratios within the combination is calculated, and the number jump values ​​are marked to obtain the offset propagation path mapping group; S4: Based on the offset propagation path mapping group, a random forest is used to extract the continuous value sequence of voltage, current, and temperature for each node in the path within the same period. The mutation amplitude values ​​of the three types of sequences in adjacent periods are calculated respectively, and a multidimensional vector set of the three types of mutation amplitudes is established. The multidimensional vectors of all nodes in the same path are clustered according to the path structure. The numbered combinations with the three types of mutation amplitude peak points in multiple paths are identified. The occurrence frequency of each numbered combination and its corresponding maximum amplitude are recorded to obtain the multi-channel synchronous jump node set; S5: Based on the multi-channel synchronous jump node set, the number of cycles required for each node to recover to the previous state is extracted, and the power change reference curve of the preset shading type is called. The offset value is compared point by point with the recovery process curve. Combined with the orientation angle of the component bracket where the numbered node is located and the solar incidence angle in the time period, the ratio of the total offset value of each node to the total offset value of the reference curve is calculated, and a matching table of numbers and shading types is generated to obtain a shading morphology distribution map.

[0030] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A photovoltaic power generation equipment detection system, characterized by: The system comprises: Electrical topology construction module: Based on the established PV module number table, combiner box connection records, and inverter wiring files, it performs number matching, power difference calculation, and voltage difference judgment, constructs directed edge connection pairs and edge weight sequences, and generates an electrical directed connection graph; Abnormal response aggregation module: Based on the power mutation trend of nodes in the electrical directed connection graph, it performs jump point synchronization detection, adjacent path identification and synchronization concentration screening to obtain mutation starting point combinations with high distribution frequency and establish a candidate set of abnormal propagation starting points; Abnormal path identification module: Based on the candidate set of abnormal propagation starting points, a graph convolutional neural network is used to extract the power offset value ratio of nodes in the path, count the continuously increasing paths, determine the path with the maximum total offset, form a hop path mapping, and generate a power offset propagation chain group; Perturbation trigger judgment module: Based on the power offset propagation chain group, it extracts the multi-channel sudden jump values ​​of voltage, current and temperature rise, uses random forest to determine the combination of multi-channel synchronous jump nodes in the same period, screens the points with high index change rate, and establishes a multi-channel abnormal sudden jump identification map; Occlusion morphology deconstruction module: Based on the multi-channel abnormal jump recognition map, the power recovery period and the occlusion reference curve are called, the occlusion offset ratio is calculated in combination with the bracket orientation angle and the incident angle, and an occlusion category area distribution map is established.

2. The photovoltaic power generation equipment detection system according to claim 1, characterized in that: The electrical topology construction module includes: Number verification submodule: Based on the established PV module number table, combiner box connection records, and inverter wiring files, the module numbers are matched item by item with the combiner box records, and unmatched numbers are screened out. The module numbers are cross-checked item by item with the inverter files, and the corresponding positions of the numbers are marked. The power output difference between the module and the combiner box is calculated, and a difference matrix is ​​generated. The power output difference between the combiner box and the inverter is calculated, and the difference variation range is recorded. The numerical difference between the module output voltage and the inverter access voltage is measured, and the voltage offset range is confirmed to obtain the electrical parameter difference mapping matrix. Connection generation submodule: Based on the electrical parameter difference mapping matrix, the power difference between the component and the combiner box is compared with the set interval, and the connection combinations with abnormal differences are eliminated. The voltage difference between the combiner box and the inverter is compared with the boundary of the allowed interval, and the connection pairs with qualified voltage differences are screened. The triplet number combination of the photovoltaic component, combiner box and inverter node is generated, and the connection edges with consistent directions are constructed. The power difference, voltage difference and number position sequence are extracted, and the numerical value set corresponding to the edge is generated to obtain the edge connection attribute set; Graph construction submodule: Based on the edge connection attribute set, a directed chain of numbering order from component nodes to inverter nodes is constructed, and the node numbering path is connected and constructed, and the edge connection weights are marked in sequence and associated with the corresponding node numbers in the graph. The outgoing and incoming edge structures of all nodes are organized, and the directionality and node type are distinguished and marked to generate an electrical directed connection graph.

3. The photovoltaic power generation equipment detection system according to claim 1, characterized in that: The abnormal response aggregation module includes: Power mutation identification submodule: Based on the electrical directed connection graph, the power output value of each node at each moment is extracted time-period by time period, and a time series power array is constructed. The adjacent differences of the node power output values ​​are calculated, and the positions of sudden increases and decreases in power jumps are located. The difference of the jump time of multiple nodes is cross-searched, and the set of synchronous nodes whose change time difference is less than the set window value is marked to obtain the set of synchronous mutation nodes. Path aggregation and extraction submodule: Based on the synchronous mutation node set, it traces the path of each synchronous node in the electrical directed connection graph and extracts the path sequence from the start point to the end point. It performs position index statistics of each jump node number in the path sequence, records the hop count difference and connection level between the numbers, performs statistical counting of the starting point numbers in all paths, and marks the path combination with the highest number of repetitions to obtain a centralized mutation path mapping table. Starting point set generation submodule: Based on the centralized mutation path mapping table, the total frequency of occurrence of the starting numbers of all paths is sorted, and a mapping set of numbers and corresponding frequencies is generated. The top numbers after frequency sorting are screened, and the number combinations with the highest frequencies are intercepted. A candidate node number set is constructed and output in the original order to establish a candidate set of abnormal propagation starting points.

4. The photovoltaic power generation equipment detection system according to claim 1, characterized in that: The abnormal path identification module includes: Offset extraction submodule: Based on the candidate set of abnormal propagation starting points, a graph convolutional neural network is used to sequentially extract the numbers of all nodes in the starting point number path, and the power data of each node in the path is read node by node. The difference in power values ​​of the previous and next nodes is calculated and the ratio is recorded. A list of power offset ratios between nodes is constructed and a continuous increasing sequence is screened. The set of paths that meet the monotonicity requirement is extracted to generate a power increasing path sequence. Path construction submodule: Based on the power-increasing path sequence, the node and its offset ratio in each path within the path number are jointly identified and constructed, the paths are sorted from high to low according to the cumulative offset value of the path, and the corresponding path numbers are recorded. The path set with the largest offset value is extracted and the hop count is counted. A path sequence index and hop count matching table is generated, and a hop count path mapping set is generated. Chain group generation submodule: Based on the hop path mapping set, the node numbers in each path are connected in sequence to form a directed node chain group, and a connection structure with consistent direction between nodes is constructed chain by chain. A path topology set with power increasing as the connection condition is generated, and the data organization and numbering of the chain group set are completed to generate a power offset propagation chain group.

5. The photovoltaic power generation equipment detection system according to claim 4, characterized in that: The graph convolutional neural network is constructed according to the formula: in: Represents graph convolution Nodes in the layer The power offset characteristics, represents a nonlinear activation function, Representation and Node Adjacent nodes , Representation node With node The power difference slope change rate of the path segment is Representation node To Node The frequency ratio of the path segment in the historical anomaly propagation, Representation node The power adjustment value of the degree, Representation node The power adjustment value of the degree, represents the weight matrix of the graph convolution layer, Represents graph convolution Nodes in the layer The power ratio characteristic input, Representation node The power fluctuation factor, Representation node The frequency ratio of multiple power increasing paths.

6. The photovoltaic power generation equipment detection system according to claim 1, characterized in that: The perturbation trigger determination module includes: The outburst extraction submodule extracts the voltage, current, and temperature channel data of each node in each chain based on the power offset propagation chain group and aligns them by time. It compares the difference between the channel values ​​at the same node and the same time point and the values ​​at adjacent moments, marks the time points that are greater than the threshold, and collects statistics on channel types and outburst amplitudes to generate a multi-channel outburst dataset. Synchronous judgment submodule: Based on the multi-channel sudden jump data set, random forest is used to count the types of sudden jump channels in each time period and compare whether the three types of channels have common mutations. The node numbers of simultaneous sudden jumps are screened, the channel jump amplitude is extracted and the change ratio per unit time is calculated. The nodes with a change rate lower than the median are screened out, the numbers are retained and a set is generated to generate a set of synchronous high-change nodes; Graph generation submodule: Based on the synchronous high-variable node set, channel mutation type association connections are generated between node pairs, edge-connected node pairs are extracted and the channel names and change rate values ​​are marked, connection edges with strong channel type consistency are organized to form a subgraph, a multi-node sudden jump synchronization relationship graph structure is constructed, and a multi-channel abnormal sudden jump identification graph is established.

7. The photovoltaic power generation equipment detection system according to claim 6, characterized in that: The random forest is based on the formula: in: represents the final classification result, represents the total number of decision trees built in the random forest, Indicates the The weight coefficient of each tree, Indicates the The classification output function of the tree, Indicates the The original jump feature vector received by the tree, Represents the adjustment weight coefficient of the channel frequency variation factor, Indicates the channel jump frequency change value, represents the adjustment weight coefficient of the neighborhood synergy factor, represents the sudden jump coordination factor of the neighborhood nodes, Represents the adjustment weight coefficient of the energy ratio factor, represents the sudden energy ratio factor, Represents the adjustment weight coefficient of the density factor, represents the outburst event density factor.

8. The photovoltaic power generation equipment detection system according to claim 6, characterized in that: The perturbation trigger judgment module is specifically an identification output structure for calibrating multi-channel jump behavior. The jump identification map includes a set of synchronous jump channel indexes, a jump amplitude threshold distribution layer and a jump node period grouping label. The high-value point of the abnormal indicator is specifically a combination of voltage nodes, current nodes and temperature rise nodes whose jump amplitude exceeds a preset change rate threshold. The power offset propagation chain group specifically refers to a set of channel sequences that form a jump response path.

9. The photovoltaic power generation equipment detection system according to claim 1, characterized in that: The occlusion morphology deconstruction module includes: The occlusion response extraction submodule is based on the multi-channel abnormal jump identification map. The numbers of all nodes in the map are read one by one and the path sequence is indexed. The power recovery time series of the corresponding nodes is extracted and the recovery segments are divided into time intervals. The point-by-point difference between each recovery period and the occlusion reference curve is calculated and the relative offset rate is recorded. The cycle length of each node in the path is counted and the paths are classified and numbered. The numbered sets are aggregated according to the path labels to generate the occlusion cycle response set. Angle offset calculation submodule: Based on the shading period response set, the module performs mapping extraction of each path number to the bracket number and obtains the orientation angle parameter. The module extracts the sunlight incident angle sequence of the period corresponding to the node jump time and matches the path number. The module calculates the angle difference between the orientation angle and the incident angle of each path bracket and converts it into an angle ratio value. The module generates a list of path number and offset angle ratio combinations and generates an shading offset ratio table. Regional structure generation submodule: Based on the occlusion offset ratio table, the node position coordinates corresponding to each path number are extracted and mapped to two-dimensional space, the grading range is set according to the offset ratio and the nodes are divided into corresponding grade intervals, the coordinates of nodes of the same grade are aggregated to form regional blocks and the regional boundary numbers are recorded, the numbering structure of each regional block is organized and the spatially adjacent blocks are merged to establish an occlusion category regional distribution map.

10. A device anomaly detection method, characterized in that: The photovoltaic power generation equipment detection system according to any one of claims 1 to 9 is executed, The following steps are involved: S1: Based on the PV module number table, combiner box connection records, and inverter wiring files, extract the correspondence between each node number and wiring, match the module number with the wiring number, calculate the power value difference between the number pairs, determine the degree of difference in the corresponding voltage values, establish a directed edge set and edge weight sequence with the number pairs as node connections, and obtain the electrical connection map; S2: Based on the electrical connection map, identify the power change sequence of all nodes within the observation period, select the jump points in the power sequence whose intervals are no more than two sampling periods and whose increase or decrease exceeds a set difference, identify the adjacent path nodes of each jump point, count the frequency of occurrence of the same node in the jump path in multiple paths, select the node combination with an occurrence frequency greater than a specified value, and obtain the abnormal propagation starting point combination set; S3: Based on the abnormal propagation starting point combination set, a graph convolutional neural network is used to construct a number sequence pair for each pair of upstream and downstream nodes in the path, and the ratio of the power value of each node on the path to the power value of its upstream node is calculated to form a ratio matrix indexed by the number pair. The comparison value matrix is ​​weighted superimposed and normalized along the connection direction under the entire graph structure. All number combinations with continuously increasing ratios in the connection path are extracted, the cumulative sum of the ratios in the combination is calculated, and the number jump values ​​are marked to obtain the offset propagation path mapping group; S4: Based on the offset propagation path mapping group, a random forest is used to extract a continuous value sequence of voltage, current, and temperature for each node in the path within the same period. The mutation amplitude values ​​of the three types of sequences in adjacent periods are calculated respectively, and a multidimensional vector set of the three types of mutation amplitudes is established. The multidimensional vectors of all nodes in the same path are clustered according to the path structure. The numbered combinations that simultaneously have the three types of mutation amplitude peak points in multiple paths are identified. The occurrence frequency of each numbered combination and its corresponding maximum amplitude are recorded to obtain a multi-channel synchronous sudden jump node set. S5: Based on the multi-channel synchronous jump node set, the number of cycles required for each node to recover to the previous state is extracted, the power change reference curve of the preset shading type is called, and the offset value is compared point by point with the recovery process curve. Combined with the orientation angle of the component bracket where the numbered node is located and the solar incidence angle in the time period, the ratio of the total offset value of each node to the total offset value of the reference curve is calculated, and a matching table of numbers and shading types is generated to obtain a shading morphology distribution map.

Images