Photovoltaic power station anomaly detection system based on multi-mode perception

Abstract

The invention discloses a photovoltaic power station anomaly detection system based on multi-modal perception, and the system comprises the following steps: a topology collection module which is used for constructing an equipment topological graph and a data synchronization graph, collecting node signal data, and generating a multi-modal signal data set; the resonance identification module is used for constructing an initial propagation relation graph according to correlation values among the nodes; the atlas construction module is used for inputting the initial propagation relation atlas into the improved AGCN model and outputting a causal propagation atlas; the anomaly elimination module is used for constructing an environment data sequence and generating a corrected propagation map; the path control module is used for updating and correcting the propagation atlas and outputting a stable control atlas; and the weight updating module is used for generating an event label set and outputting a final result graph. The method has the advantages of high identification accuracy, strong interference elimination capability, high node positioning precision and the like, and can effectively improve the operation safety and fault response efficiency of the photovoltaic power station.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant operation monitoring and intelligent diagnosis technology, and in particular to a photovoltaic power plant anomaly detection system based on multimodal perception. Background Technology

[0002] With the rapid increase in the demand for large-scale construction and centralized operation and maintenance management of photovoltaic power plants, anomaly identification and risk warning based on equipment operating status have become key means to ensure power generation efficiency and equipment safety. Existing photovoltaic power plant anomaly detection technologies are mostly based on independent node modeling or static rule analysis, primarily relying on single-point voltage, current, or power signal thresholds for judgment. These technologies are significantly insufficient in identifying anomalies associated with different devices and complex propagation paths. In photovoltaic systems with multiple devices operating in a network, local device failures or communication anomalies may propagate along node links, inducing cascading performance degradation. However, traditional methods struggle to effectively trace anomaly propagation paths or identify resonant coupling relationships.

[0003] Existing anomaly identification schemes generally ignore the potential dynamic relationship structure between devices and lack the ability to analyze the causal relationship of faults from the perspective of propagation mechanism, making it difficult to distinguish between the source anomaly and the affected nodes. At the same time, most technologies do not consider the interference of environmental conditions such as light intensity and ambient temperature on the propagation path, resulting in a high false identification rate and low positioning accuracy. In terms of graph structure construction, existing models often adopt static adjacency methods or fixed topology update logic, which cannot reflect the dynamic changes of the connection relationship between nodes in actual operation. The discrimination strategies for closed structures or frequently occurring nodes in complex paths are still inadequate, and it is impossible to implement dynamic suppression or edge weight correction for high-frequency trigger points of anomalies, which limits the accurate expression of the propagation characteristics of the whole graph and the ability to continuously optimize them.

[0004] Therefore, how to provide a photovoltaic power plant anomaly detection system based on multimodal perception is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an anomaly detection system for photovoltaic power plants based on multimodal perception. This invention fully utilizes multimodal signal acquisition, resonance identification and analysis, graph structure modeling, and causal propagation path analysis techniques. It details the complete detection process of constructing equipment topology graphs and data synchronization graphs, calculating the fusion correlation values ​​between nodes, outputting causal propagation graphs based on the improved AGCN model, removing environmental driving edges to construct corrected propagation graphs, adding attenuation factors based on propagation path loop detection to generate stable control graphs, and further adjusting the causal propagation weights based on node trigger frequencies to generate the final result graph. This system has the advantages of high fault identification accuracy, clear anomaly source location, and strong environmental interference removal capabilities.

[0006] A photovoltaic power plant anomaly detection system based on multimodal sensing according to an embodiment of the present invention includes:

[0007] The topology acquisition module is used to construct device topology diagrams and data synchronization diagrams; it collects signal data from nodes and generates multimodal signal datasets.

[0008] The resonance identification module is used to calculate the correlation value between each node in the multimodal signal dataset; if the correlation value is greater than a first preset threshold in multiple consecutive sampling periods, it is determined that there is a resonance relationship between the nodes, and the node pairs with resonance relationship are marked as propagation edges to construct an initial propagation relationship graph.

[0009] The graph construction module is used to input the initial propagation relationship graph into the improved AGCN model and output the causal propagation graph; the improved AGCN model includes a multimodal embedding unit, a dynamic adjacency generation structure, a graph convolutional interaction structure, a causal path attention structure, and a temporal cumulative coding structure;

[0010] The anomaly removal module is used to collect environmental status data of photovoltaic power plants and construct environmental data sequences; calculate the correlation value between each propagation edge in the causal propagation graph and the environmental data sequence; if the correlation value is greater than a second preset threshold and the causal strength value of the corresponding propagation edge is less than a third preset threshold, then the propagation edge is marked as an environmental driving edge and a path removal operation is performed to generate a corrected propagation graph.

[0011] The path control module is used to perform loop detection based on the modified propagation graph to determine whether there is a closed path structure in the propagation path; if there is a closed structure in the propagation path and the cumulative propagation intensity is greater than the fourth preset threshold, then an attenuation factor is added to each propagation edge in the propagation path, the modified propagation graph is updated and a stable control graph is output.

[0012] The weight update module is used to generate an event label set based on the stable control graph; count the cumulative number of triggers of each node in the propagation path; if the cumulative number of triggers is greater than a fourth preset threshold, then perform an adjustment operation on the causal propagation weight of the node in the stable control graph, and output the final result graph.

[0013] Preferably, the steps between modules include the following:

[0014] S1. Construct a device topology diagram and data synchronization diagram; collect signal data from nodes and generate a multimodal signal dataset;

[0015] S2. Calculate the correlation value between each node in the multimodal signal dataset; if the correlation value is greater than the first preset threshold in multiple consecutive sampling periods, it is determined that there is a resonance relationship between the nodes, and the node pairs with resonance relationship are marked as propagation edges to construct an initial propagation relationship graph.

[0016] S3. Input the initial propagation relationship graph into the improved AGCN model and output the causal propagation graph; the improved AGCN model includes a multimodal embedding unit, a dynamic adjacency generation structure, a graph convolutional interaction structure, a causal path attention structure, and a temporal cumulative coding structure;

[0017] S4. Collect environmental status data of photovoltaic power station and construct environmental data sequence; calculate the correlation value between each propagation edge in the causal propagation graph and the environmental data sequence; if the correlation value is greater than the second preset threshold and the causal strength value of the corresponding propagation edge is less than the third preset threshold, then mark the propagation edge as an environment-driven edge and perform path elimination operation to generate a corrected propagation graph.

[0018] S5. Perform loop detection based on the modified propagation graph to determine whether there is a closed path structure in the propagation path; if there is a closed structure in the propagation path and the cumulative propagation intensity is greater than the fourth preset threshold, then add an attenuation factor to each propagation edge in the propagation path, update the modified propagation graph and output a stable control graph.

[0019] S6. Generate an event label set based on the stable control graph; count the cumulative number of triggers for each node in the propagation path; if the cumulative number of triggers is greater than a fourth preset threshold, perform an adjustment operation on the causal propagation weight of the node in the stable control graph, and output the final result graph.

[0020] Preferably, the construction of the device topology diagram and data synchronization diagram specifically includes:

[0021] Construct a node set, which includes photovoltaic modules, combiner boxes, inverters, and monitoring nodes;

[0022] Based on the electrical connection relationship and physical layout structure, the connection relationship between nodes is generated. The connection relationship includes the logical connection edge and the physical connection edge between nodes. The logical connection edge records the signal transmission channel, and the physical connection edge records the component wiring relationship and installation location information.

[0023] A device topology graph is constructed based on the node set and connection relationship. The device topology graph is graph structure data, which includes a node identifier set, an edge set, and a node type label.

[0024] Collect the sampling configuration parameters of each node, including sampling interval, sampling frequency, signal dimension, and data transmission delay; measure the physical distance between nodes, which is the Euclidean distance or path hop distance in the relative spatial coordinate system;

[0025] Synchronization factors are constructed based on the sampling interval difference, data transmission delay difference, and physical distance between each pair of nodes; a synchronization weight matrix is ​​generated based on the synchronization factors, and the synchronization weight matrix records the degree of synchronization matching of each pair of nodes during the multimodal signal acquisition process.

[0026] Preferably, the signal data from the acquisition node is used to generate a multimodal signal dataset, specifically as follows:

[0027] Collect signal data from photovoltaic modules, combiner boxes, inverters and monitoring nodes. The signal data includes voltage signals, current signals, temperature signals, power signals and communication status signals.

[0028] The acquired signal data is processed for time synchronization. The multi-source signals are time-aligned according to a uniform sampling period to generate a time-aligned sequence.

[0029] Noise suppression processing is performed on each type of signal in the time-aligned sequence. High-frequency noise is removed by bandpass filtering for voltage and current signals, and smoothing is performed by moving average processing for temperature and power signals. Abnormal states are smoothed by moving window statistics for communication status signals.

[0030] Normalization is performed on the noise-suppressed signal data. The maximum and minimum value normalization method is used to map the signal values ​​to a unified dimension range, generating a normalized signal set.

[0031] Assign a time index identifier to each type of normalized signal. The time index identifier records the sampling time and sampling period number. Calculate the sampling confidence weight for each type of signal. The sampling confidence weight is generated based on sampling stability, signal integrity, and data delay. Combine the time index, the normalized signal set, and the sampling confidence weight to form a multimodal signal dataset, and output the multimodal signal dataset.

[0032] Preferably, S2 specifically includes:

[0033] Based on the device topology diagram and the synchronization weight matrix, select node pairs with related structures as the candidate node pair set;

[0034] Time-aligned signal sequences of candidate node pairs are extracted from the multimodal signal dataset, and multimodal feature vectors are constructed respectively. The multimodal feature vectors include voltage feature vector, current feature vector, temperature feature vector, power feature vector and communication state feature vector.

[0035] For each type of feature vector, a corresponding modal similarity value is calculated. The modal similarity value is constructed based on the signal change trend, amplitude difference, and spectral characteristics. A weighted fusion operation is performed on each modal similarity value to generate a fused correlation value. The weighting weights are jointly determined based on the modal confidence weights and the synchronization weight matrix.

[0036] The fusion correlation value is tracked over multiple consecutive sampling periods. If the fusion correlation value is continuously greater than the first preset threshold for a preset number of sampling periods, it is determined that the corresponding node pair has a resonance relationship. The node pair with resonance relationship is marked as a propagation edge. The propagation edge records the corresponding node pair identifier, start and end time index and correlation strength. All propagation edges are combined to generate an initial propagation relationship graph. The initial propagation relationship graph contains a set of propagation nodes, a set of propagation edges and correlation labeling information.

[0037] Preferably, the improved AGCN model includes a multimodal embedding unit, a dynamic adjacency generation structure, a graph convolutional interaction structure, a causal path attention structure, and a temporal cumulative coding structure, specifically:

[0038] The multimodal embedding unit constructs independent modal embedding channels for voltage, current, temperature, power and communication status signals in the multimodal signal dataset. Each embedding channel extracts modal feature vectors based on a one-dimensional convolutional structure and outputs a set of modal embedding tensors.

[0039] The dynamic adjacency generation structure constructs an initial adjacency matrix based on the device topology graph and the synchronization weight matrix. It integrates modal confidence weights and inter-node synchronization factors to calculate the dynamic connection probability between each pair of nodes and generate a dynamic adjacency matrix set. The dynamic adjacency matrix set contains edge connection weight information under different time indices.

[0040] The graph convolutional interactive structure performs graph convolution operations on the modality embedding tensor set respectively. The graph convolution operations are based on the dynamic adjacency matrix to generate the propagation feature representation of each modality. The propagation feature representation is input into the interactive fusion path, which calculates the modality fusion weight coefficients based on the modality confidence weights and outputs the modality fusion propagation feature vector.

[0041] The causal path attention structure constructs a set of candidate propagation paths based on the initial propagation relationship graph, extracts the position information of the propagation edges, propagation intensity and modality consistency index in each propagation path, and calculates the path attention score by combining the path structure features. The path attention score is used to weight the path context information to generate a path-level causal representation tensor.

[0042] The temporal cumulative coding structure constructs a propagation intensity time series based on the path-level causal representation tensor. The propagation intensity time series is input into the gated cyclic unit structure for sequence modeling, and outputs the causal propagation intensity value of each propagation edge at different time indices, which constitutes the edge weight information of the causal propagation graph.

[0043] Preferably, S4 specifically comprises:

[0044] Collect environmental status data of photovoltaic power plants, including light intensity, ambient temperature, wind speed, humidity and rainfall; perform time alignment processing on the environmental status data, and construct an environmental data sequence according to a unified sampling period and time index. The environmental data sequence records the numerical changes of each environmental parameter at different sampling times.

[0045] The correlation degree between the propagation edge and the environmental data sequence is calculated based on the modified propagation map and the environmental data sequence. The correlation degree is calculated based on the time synchronization factor, amplitude response factor and change trend factor. Under each time index, the causal strength value of the propagation edge and the change gradient of the corresponding environmental parameter are extracted to form a correlation calculation sample set. The correlation calculation operation is performed on the correlation calculation sample set to generate a correlation degree matrix. The correlation degree matrix records the correlation index between each propagation edge and each environmental parameter.

[0046] The maximum correlation value and the corresponding causal strength value of each propagation edge in the correlation matrix are jointly determined. If the correlation value is greater than the second preset threshold and the causal strength value is less than the third preset threshold, the corresponding propagation edge is marked as an environment-driven edge. All environment-driven edges are removed from the modified propagation graph, the propagation edge set is updated, and the modified propagation graph is generated. The modified propagation graph includes a propagation node set, a propagation edge set, correlation labeling information, and time index records.

[0047] Preferably, S5 specifically includes:

[0048] Extract the propagation edge set and propagation node set from the modified propagation graph to construct the propagation path set; perform closed path detection operation on the propagation path set. A closed path is a path structure in the propagation path where the start node identifier and the end node identifier are the same and the path length is not less than three.

[0049] For each closed path in the propagation path set, calculate the cumulative propagation intensity value. The cumulative propagation intensity value is the sum of the causal intensity values ​​of all propagation edges in the path. If the cumulative propagation intensity value is greater than the fourth preset threshold, then assign an attenuation factor to all propagation edges in the closed path. The attenuation factor is calculated and generated based on the cumulative propagation intensity value and the propagation path length.

[0050] The attenuation factor is applied to the causal strength value of the corresponding propagation edge to generate an updated propagation edge set; the updated propagation edge set is combined with the propagation node set to construct a stable control graph; the stable control graph includes the propagation node set, the propagation edge set, the path structure set, and causal strength correction labeling information.

[0051] Preferably, the step of generating an event label set based on the stable control map specifically involves:

[0052] Extract the propagation node set, propagation edge set, and path structure set from the stable control graph, and establish a node index table and a path index table; determine the trigger time range of the propagation path based on the time index record, and arrange the propagation paths in chronological order; calculate the average propagation intensity and the rate of change of propagation intensity for each propagation path, and generate a causal intensity index by combining the causal intensity values ​​recorded in the propagation edge.

[0053] Extract node identifiers and propagation direction information from each propagation path to form a set of node pairs; map the set of node pairs to propagation path identifiers to construct an event mapping table; record node identifiers, propagation path identifiers, trigger time indexes, and causality strength indicators in the event mapping table;

[0054] Extract environmental state information corresponding to the trigger time index from the environmental data sequence and calculate the environmental state index value; merge the environmental state index value with the record items in the event mapping table to generate an event tag set; the event tag set includes propagation path identifier, node identifier, trigger time index, causal strength value and environmental state index.

[0055] Preferably, the cumulative number of triggers for each node in the propagation path is counted; if the cumulative number of triggers is greater than a fourth preset threshold, the causal propagation weight of the node in the stable control graph is adjusted, and the final result graph is output, specifically:

[0056] Extract the node identifier field and trigger time index field from the event tag set to construct a node trigger record table; perform aggregation operation on the node trigger record table according to the node identifier, count the cumulative number of triggers of each node in all propagation paths, and generate a node trigger frequency list.

[0057] The node trigger frequency list is compared with the fourth preset threshold, and the target node set with a cumulative trigger number greater than the fourth preset threshold is selected; for each node in the target node set, all propagation edges connected to the target node in the stable control graph are extracted;

[0058] An adjustment factor is added to the causal strength value of each propagation edge. The adjustment factor is calculated and generated based on the node trigger frequency and the number of propagation paths. During the adjustment process, the propagation direction remains unchanged, and the causal propagation weight value in the propagation edge set is updated to form a propagation edge correction record.

[0059] The propagation edge correction record is combined with the original propagation node set and path structure set to construct the final result graph; the final result graph includes the propagation node set, propagation edge set, path structure set, causal propagation weight update identifier and node trigger frequency mapping table.

[0060] The beneficial effects of this invention are:

[0061] This invention constructs a device topology graph and a data synchronization graph, collects signal data from nodes, and generates a multimodal signal dataset. Addressing the problems of low accuracy in identifying anomaly propagation relationships in photovoltaic power plants, susceptibility to misjudgments due to environmental interference, and difficulty in controlling the dynamic evolution of propagation paths, this invention employs a resonance identification module to calculate the correlation values ​​between nodes in the multimodal signal dataset, constructing an initial propagation relationship graph. This initial propagation relationship graph is then input into an improved AGCN model, which includes multimodal embedding units, a dynamic adjacency generation structure, a graph convolutional interaction structure, a causal path attention structure, and a temporal cumulative coding structure, outputting a causal propagation graph. An anomaly removal module collects environmental state data from the photovoltaic power plant, constructs an environmental data sequence, calculates the correlation between propagation edges in the causal propagation graph and the environmental data sequence, identifies environmental driving edges, and performs path removal operations to generate a corrected propagation graph. A path control module performs closed path detection and introduces an attenuation factor, outputting a stable control graph. Finally, a weight update module counts the cumulative trigger counts of nodes in the propagation path, adjusts the causal propagation weights, and outputs the final result graph. This invention has the advantages of high propagation path accuracy, strong ability to eliminate environmentally driven interference, and high robustness in identifying abnormal evolution. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a block diagram of a photovoltaic power station anomaly detection system based on multimodal sensing proposed in this invention;

[0064] Figure 2 This is a flowchart of a photovoltaic power station anomaly detection system based on multimodal sensing proposed in this invention;

[0065] Figure 3 This is a data flow diagram of a photovoltaic power plant anomaly detection system based on multimodal perception proposed in this invention. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0067] refer to Figure 1-3 A photovoltaic power plant anomaly detection system based on multimodal sensing includes:

[0068] The topology acquisition module is used to construct device topology diagrams and data synchronization diagrams; it collects signal data from nodes and generates multimodal signal datasets.

[0069] The resonance identification module is used to calculate the correlation value between each node in the multimodal signal dataset; if the correlation value is greater than a first preset threshold in multiple consecutive sampling periods, it is determined that there is a resonance relationship between the nodes, and the node pairs with resonance relationship are marked as propagation edges to construct an initial propagation relationship graph.

[0070] The graph construction module is used to input the initial propagation relationship graph into the improved AGCN model and output the causal propagation graph; the improved AGCN model includes a multimodal embedding unit, a dynamic adjacency generation structure, a graph convolutional interaction structure, a causal path attention structure, and a temporal cumulative coding structure;

[0071] The anomaly removal module is used to collect environmental status data of photovoltaic power plants and construct environmental data sequences; calculate the correlation value between each propagation edge in the causal propagation graph and the environmental data sequence; if the correlation value is greater than a second preset threshold and the causal strength value of the corresponding propagation edge is less than a third preset threshold, then the propagation edge is marked as an environmental driving edge and a path removal operation is performed to generate a corrected propagation graph.

[0072] The path control module is used to perform loop detection based on the modified propagation graph to determine whether there is a closed path structure in the propagation path; if there is a closed structure in the propagation path and the cumulative propagation intensity is greater than the fourth preset threshold, then an attenuation factor is added to each propagation edge in the propagation path, the modified propagation graph is updated and a stable control graph is output.

[0073] The weight update module is used to generate an event label set based on the stable control graph; count the cumulative number of triggers of each node in the propagation path; if the cumulative number of triggers is greater than a fourth preset threshold, then perform an adjustment operation on the causal propagation weight of the node in the stable control graph, and output the final result graph.

[0074] This implementation method, by setting up a topology acquisition module, achieves unified modeling of the structural connection information and data acquisition information of photovoltaic modules, combiner boxes, inverters, and monitoring nodes, generating equipment topology diagrams and data synchronization diagrams, and outputting structured multimodal signal datasets, effectively improving the accuracy of expressing equipment hierarchical relationships and data time alignment. Through a resonance identification module, multimodal correlation values ​​between nodes are calculated, and resonant node pairs are identified within continuous periods, constructing an initial propagation relationship graph, thereby improving the modeling capability for abnormal linkage phenomena between multiple devices. Through a graph construction module, an improved AGCN model is introduced, integrating multimodal embedding information, dynamic graph structure, and path attention mechanism to output a causal propagation graph, effectively enhancing the spatiotemporal recognition accuracy of abnormal causal relationships. Through an anomaly removal module, environmental state data is introduced as a screening criterion, performing removal operations on environment-driven propagation edges in the causal propagation graph, further eliminating environmental interference factors and improving the reliability and stability of the propagation graph. The path control module performs cumulative evaluation of the propagation intensity of closed paths in the corrected propagation graph and adds propagation edge attenuation factors to further suppress the amplified impact of abnormal chain propagation paths on the graph evolution, thereby improving the overall model's stability and anomaly localization accuracy. The weight update module constructs an event label set based on node trigger frequency, identifies high-frequency trigger nodes, and dynamically adjusts their propagation weights. This enhances the system's response capability to latent fault propagation sources and improves the sensitivity and targeting of photovoltaic power plant anomaly detection.

[0075] In this embodiment, the steps between modules include the following:

[0076] S1. Construct a device topology diagram and data synchronization diagram; collect signal data from nodes and generate a multimodal signal dataset;

[0077] S2. Calculate the correlation value between each node in the multimodal signal dataset; if the correlation value is greater than the first preset threshold in multiple consecutive sampling periods, it is determined that there is a resonance relationship between the nodes, and the node pairs with resonance relationship are marked as propagation edges to construct an initial propagation relationship graph.

[0078] S3. Input the initial propagation relationship graph into the improved AGCN model and output the causal propagation graph; the improved AGCN model includes a multimodal embedding unit, a dynamic adjacency generation structure, a graph convolutional interaction structure, a causal path attention structure, and a temporal cumulative coding structure;

[0079] S4. Collect environmental status data of photovoltaic power station and construct environmental data sequence; calculate the correlation value between each propagation edge in the causal propagation graph and the environmental data sequence; if the correlation value is greater than the second preset threshold and the causal strength value of the corresponding propagation edge is less than the third preset threshold, then mark the propagation edge as an environment-driven edge and perform path elimination operation to generate a corrected propagation graph.

[0080] S5. Perform loop detection based on the modified propagation graph to determine whether there is a closed path structure in the propagation path; if there is a closed structure in the propagation path and the cumulative propagation intensity is greater than the fourth preset threshold, then add an attenuation factor to each propagation edge in the propagation path, update the modified propagation graph and output a stable control graph.

[0081] S6. Generate an event label set based on the stable control graph; count the cumulative number of triggers for each node in the propagation path; if the cumulative number of triggers is greater than a fourth preset threshold, perform an adjustment operation on the causal propagation weight of the node in the stable control graph, and output the final result graph.

[0082] In this embodiment, the construction of the device topology diagram and data synchronization diagram specifically includes:

[0083] Construct a node set, which includes photovoltaic modules, combiner boxes, inverters, and monitoring nodes;

[0084] Based on the electrical connection relationship and physical layout structure, the connection relationship between nodes is generated. The connection relationship includes the logical connection edge and the physical connection edge between nodes. The logical connection edge records the signal transmission channel, and the physical connection edge records the component wiring relationship and installation location information.

[0085] A device topology graph is constructed based on the node set and connection relationship. The device topology graph is graph structure data, which includes a node identifier set, an edge set, and a node type label.

[0086] Collect the sampling configuration parameters of each node, including sampling interval, sampling frequency, signal dimension, and data transmission delay; measure the physical distance between nodes, which is the Euclidean distance or path hop distance in the relative spatial coordinate system;

[0087] Synchronization factors are constructed based on the sampling interval difference, data transmission delay difference, and physical distance between each pair of nodes; a synchronization weight matrix is ​​generated based on the synchronization factors, and the synchronization weight matrix records the degree of synchronization matching of each pair of nodes during the multimodal signal acquisition process.

[0088] In this embodiment, the signal data from the acquisition node is used to generate a multimodal signal dataset, specifically as follows:

[0089] Collect signal data from photovoltaic modules, combiner boxes, inverters and monitoring nodes. The signal data includes voltage signals, current signals, temperature signals, power signals and communication status signals.

[0090] The acquired signal data is processed for time synchronization. The multi-source signals are time-aligned according to a uniform sampling period to generate a time-aligned sequence.

[0091] Noise suppression processing is performed on each type of signal in the time-aligned sequence. High-frequency noise is removed by bandpass filtering for voltage and current signals, and smoothing is performed by moving average processing for temperature and power signals. Abnormal states are smoothed by moving window statistics for communication status signals.

[0092] Normalization is performed on the noise-suppressed signal data. The maximum and minimum value normalization method is used to map the signal values ​​to a unified dimension range, generating a normalized signal set.

[0093] Assign a time index identifier to each type of normalized signal. The time index identifier records the sampling time and sampling period number. Calculate the sampling confidence weight for each type of signal. The sampling confidence weight is generated based on sampling stability, signal integrity, and data delay. Combine the time index, the normalized signal set, and the sampling confidence weight to form a multimodal signal dataset, and output the multimodal signal dataset.

[0094] In this embodiment, S2 specifically refers to:

[0095] Based on the node connection relationships recorded in the device topology diagram and the synchronization matching degree values ​​in the synchronization weight matrix, node pairs with associated structures are selected to form a candidate node pair set; the candidate node pair set includes all node pairs with logical or physical connection edges and synchronization weight values ​​greater than the preset synchronization threshold.

[0096] Extract the time-aligned signal sequence of each pair of nodes from the candidate node pairs in the multimodal signal dataset; extract time segments for each type of signal according to a preset sliding window, and construct the corresponding modal feature vector; the modal feature vector includes: a voltage feature vector constructed based on the amplitude change rate and main frequency component of the voltage signal, a current feature vector constructed based on the zero crossing rate and peak frequency of the current signal, a temperature feature vector constructed based on the sliding trend slope of the temperature signal, a power feature vector constructed based on the average power change rate of the power signal, and a communication status feature vector constructed based on the interruption frequency and recovery interval of the communication status signal;

[0097] For each type of modal feature vector, the cosine similarity function is used to calculate the corresponding modal similarity value; before the similarity calculation, all feature vectors are subjected to L2 normalization; the modal similarity value represents the degree of consistency of the candidate node pair's response under a single modality;

[0098] The fusion weight coefficients are generated by jointly using the modal confidence weights and the synchronization weight matrix; the similarity values ​​of each modality are weighted and accumulated with the corresponding fusion weight coefficients to generate the fusion correlation value; the fusion correlation value represents the consistency level of the candidate node pair's comprehensive response under multimodal and synchronization constraints.

[0099] Within a preset sliding time window, the fusion correlation value is periodically tracked; the number of high correlation periods for each pair of candidate nodes within a continuous sampling period is counted; if the number of consecutive high correlation periods is greater than a preset period threshold and the corresponding fusion correlation value is continuously greater than a first preset threshold, then it is determined that the corresponding node pair has a resonance relationship.

[0100] Node pairs with resonance relationships are marked as propagation edges; the propagation edges record the node pair identifier, resonance start time index, resonance end time index, and average correlation strength value; the average correlation strength value is obtained by fitting the fusion correlation value within the high correlation period, and the fitting method is a moving weighted average.

[0101] All propagation edges are integrated and deduplicated to generate an initial propagation relationship graph. The initial propagation relationship graph includes a set of propagation nodes, a set of propagation edges, and a set of correlation annotation information. The set of propagation nodes includes the node identifiers of all participating propagation edges, the set of propagation edges records resonance relationship information, and the set of correlation annotation information records the fusion correlation statistics, signal category matching degree, and signal phase deviation index.

[0102] This implementation improves the accuracy and robustness of identifying abnormal resonance relationships between multiple nodes by introducing multimodal feature vector construction, joint weighted fusion of confidence weights, and continuous periodic tracking mechanism in the correlation identification process, thereby enhancing the reliability and structural stability of the initial propagation map.

[0103] In this embodiment, the improved AGCN model includes a multimodal embedding unit, a dynamic adjacency generation structure, a graph convolutional interaction structure, a causal path attention structure, and a temporal cumulative coding structure, specifically:

[0104] The multimodal embedding unit constructs independent modal embedding channels for voltage, current, temperature, power, and communication status signals in the multimodal signal dataset. Each modal embedding channel includes an input normalization layer, a one-dimensional convolutional layer, a nonlinear activation layer, and a pooling layer. The input normalization layer standardizes the input signal. The one-dimensional convolutional layer extracts local temporal features with a fixed kernel width. The nonlinear activation layer uses a modified linear unit function to enhance the nonlinear expression of the feature distribution. The pooling layer downsamples the features in the time dimension to extract the main temporal variation trends. The output modal embedding tensor set includes voltage mode tensor, current mode tensor, temperature mode tensor, power mode tensor, and communication status mode tensor.

[0105] The dynamic adjacency generation structure generates an initial adjacency matrix based on the logical connection relationships between nodes in the device topology graph and the synchronization matching degree recorded in the synchronization weight matrix. A connection probability model is constructed based on modal confidence weights and synchronization factors between nodes. The connection probability model uses a Gaussian kernel function to fit the time difference and spatial distance between nodes, and generates dynamic connection probability values ​​through confidence weight normalization. Under each time index, the corresponding edge weights in the adjacency matrix are updated according to the connection probability values. Finally, a dynamic adjacency matrix set is generated, which records the edge connection strength sequence under different time indices.

[0106] The graph convolutional interactive structure receives a set of modal embedding tensors and a set of dynamic adjacency matrices. It performs feature aggregation operations based on graph convolution operators on each modal channel. The graph convolution operator weights the neighbor node features through the adjacency matrix and performs a linear transformation to generate a propagation feature representation. This propagation feature representation is input to the interactive fusion path, which obtains modal fusion weight coefficients based on modal confidence weights and node feature variances. The propagation features of different modalities are weighted and fused to output a modal fusion propagation feature vector. This modal fusion propagation feature vector records the collaborative change patterns between multiple modalities.

[0107] The causal path attention structure constructs a set of candidate propagation paths based on the propagation edge information marked in the initial propagation relationship graph. Each propagation path consists of continuous propagation edges. The position information, propagation intensity value, and modality consistency index of the propagation edges in the path are extracted. The modality consistency index is calculated based on the similarity function of the modality embedding tensor. The path attention score is obtained by comprehensively fitting the mean propagation intensity, mean modality consistency, and path length within the path. The path attention score is weighted on the path context tensor to generate a path-level causal representation tensor. The path-level causal representation tensor describes the structural dependency and modal interaction features of the propagation relationship.

[0108] The temporal cumulative encoding structure receives path-level causal representation tensors and constructs a propagation intensity time series. The propagation intensity time series is arranged in time index order for causal propagation intensity samples. It is input into a gated recurrent unit structure to perform temporal dependency modeling. The gated recurrent unit contains update gates and reset gates. The update gate controls the proportion of historical state retention, and the reset gate controls the proportion of short-term feature forgetting. The model performs weighted fitting on the propagation intensity at different time indices through a time weighting function to generate a causal propagation intensity sequence. The output is the causal propagation intensity value of each propagation edge at different time indices. The temporal intensity values ​​of all propagation edges are integrated to form the edge weight information of the causal propagation graph.

[0109] This implementation further enhances the spatiotemporal feature interaction capability and causal relationship modeling capability between multimodal signals by introducing a multimodal embedding channel, dynamic adjacency generation mechanism and path-level attention structure into the improved AGCN model; and further improves the accuracy and stability of causal propagation map in abnormal propagation direction identification and intensity estimation by dynamically tracking the propagation intensity and fitting the time weight through a time-series cumulative coding structure.

[0110] In this embodiment, S4 specifically refers to:

[0111] Environmental status data of the photovoltaic power station is collected, including irradiance, ambient temperature, wind speed, humidity, and rainfall. Irradiance data is collected through an irradiance sensor, ambient temperature and humidity are collected through a meteorological monitoring module, and wind speed and rainfall are collected through a wind speed sensor and a rain gauge, respectively. All environmental status data are synchronously sampled according to a unified time base. Time alignment processing is performed on the collected environmental status data to construct a time index sequence. Time alignment adopts an interpolation resampling method, and at missing sampling points, cubic spline interpolation is used to fit and generate equally spaced sequences to form an environmental data sequence. The environmental data sequence records the numerical changes of each environmental parameter at different sampling times and retains the sampling time index, measurement point number, and confidence weight.

[0112] The correlation between propagation edges and environmental data sequences is calculated based on the modified propagation map and environmental data sequences. Propagation intensity sequences are extracted for each propagation edge in the modified propagation map. The propagation intensity sequences and environmental data sequences are time-synchronized and paired. The correlation between the rate of change of the causal intensity value of the propagation edge and the rate of change of each environmental parameter is calculated. The correlation calculation includes three parts: a time synchronization factor, an amplitude response factor, and a trend factor. The time synchronization factor is obtained by fitting a phase difference function of the propagation edge intensity change. The amplitude response factor is calculated based on the ratio of the propagation intensity amplitude difference to the fluctuation amplitude of the environmental parameters. The trend factor is calculated using the cosine of the gradient direction of the propagation intensity and the environmental parameters. The three factors are fused using a weighted summation method, and the fusion weights are obtained by fitting a confidence regression model, outputting a single correlation value.

[0113] Under each time index, the causal strength value of the propagation edge and the gradient of the corresponding environmental parameter change are extracted to construct a correlation calculation sample set. The sample set includes the propagation edge identifier, environmental parameter category, time index, propagation strength change value, and environmental gradient value. The correlation calculation operation is performed on the sample set, and the correlation degree matrix is ​​generated by the joint calculation method of Pearson correlation coefficient and Spearman rank correlation coefficient. The correlation degree matrix records the correlation index between each propagation edge and each environmental parameter, and a confidence interval estimate is attached to each index.

[0114] A joint determination is performed on the maximum correlation value and the corresponding causal strength value of each propagation edge in the correlation matrix. When the correlation value is greater than the second preset threshold and the causal strength value is less than the third preset threshold, the corresponding propagation edge is marked as an environment-driven edge. The environment-driven edge records the propagation edge identifier, correlation value, corresponding environmental parameter category, and time index range. All environment-driven edges are removed from the modified propagation graph. An edge set reconstruction operation is performed on the graph after removal to update the propagation edge set. The updated propagation edge set, together with the original propagation node set, the correlation annotation information set, and the time index record, constitute the modified propagation graph. The status of each propagation edge in the modified propagation graph is marked as a retained edge or a removed edge, and the latest update time is recorded.

[0115] This implementation method further enhances the ability of the propagation map to distinguish environmentally driven anomalies by introducing environmental multi-parameter correlation calculation, time synchronization pairing, and dynamic elimination mechanisms during the propagation map correction stage. By fusing a multi-factor fitting model to generate a correlation matrix, the accuracy and stability of the propagation path elimination process are further improved, thereby significantly enhancing the false judgment suppression capability and judgment reliability of the photovoltaic power station anomaly detection system in complex environments.

[0116] In this embodiment, S5 specifically refers to:

[0117] Extract the propagation edge set and propagation node set from the modified propagation graph; construct the propagation path set based on the propagation edge connection relationship; construct the propagation path set by graph traversal with constraints on path depth and hop count, and record the path number, path length, node sequence and edge sequence;

[0118] Perform closed path detection operations on the propagation path set; a closed path is defined as a path structure in which the starting node identifier and the ending node identifier are the same and the path length is not less than three; the detection results form a closed path set, which includes path number, node sequence, propagation edge list and path length parameter;

[0119] For each path in the set of closed paths, calculate the cumulative propagation intensity value; the cumulative propagation intensity value is the sum of the causal propagation intensity values ​​of all propagation edges in the propagation path; the causal propagation intensity value is the time mean of the edge weight sequence in the stable control graph.

[0120] A path risk index is generated based on the cumulative value of propagation intensity and the path length; the path risk index represents the degree of interference of a closed path on the stability of abnormal information propagation; when the path risk index is greater than the fourth preset threshold, the path is marked as a high-risk propagation path.

[0121] Attenuation factors are assigned to all propagation edges in high-risk propagation paths. The attenuation factors are obtained by constructing a mapping relationship based on path risk indicators and fitting the cumulative value of propagation intensity with path structural features. The attenuation factors record the propagation edge identifier, original intensity value, corrected intensity value, and adjustment label.

[0122] The decay factor is applied to the causal propagation strength value of the propagation edge to construct an updated propagation edge set. The update process retains the original time index and node connection attribute of the propagation edge, and only replaces the edge weight information. The updated propagation edge set and the propagation node set are merged to construct a stable control graph.

[0123] The stable control graph includes a set of propagation nodes, a set of propagation edges, a set of path structures, and causal strength correction annotation information. The set of path structures includes the structure type, node sequence, and propagation direction of all propagation paths. The causal strength correction annotation information includes edge identifiers, time indexes, original values, correction values, and sources of attenuation factors.

[0124] This implementation method identifies high-intensity closed-loop structures in the propagation map by introducing path risk indicators. It further dynamically generates attenuation factors based on the structural features and intensity distribution in the propagation path, and then further adjusts the causal intensity values ​​of the propagation edges to weaken the dominant position of the loop links in information propagation, thereby improving the accuracy of identifying multi-source abnormal linkage paths and the stability of the control map.

[0125] In this embodiment, generating an event label set based on the stable control map specifically involves:

[0126] Extract the propagation node set, propagation edge set, and path structure set from the stable control graph; construct a node index table based on the propagation node set, which includes node identifier, node type, component number to which the node belongs, and adjacency information in the graph; construct a path index table based on the path structure set, which records path number, start and end node pairs, path length, and a list of propagation edges.

[0127] A time mapping matrix is ​​constructed based on the time indices recorded in the propagation edges. The time mapping matrix records the activation state and propagation intensity changes of each propagation path under different time indices. The start time index and end time index of each propagation path are statistically analyzed based on the time mapping matrix to form a trigger time range record table. The trigger time range record table is merged with the path index table to generate a time-sorted propagation path sequence.

[0128] In each propagation path, the causal propagation strength values ​​of all propagation edges are counted, and the average propagation strength and the rate of change of propagation strength are calculated. The rate of change of propagation strength represents the speed at which the propagation strength changes over time. A causal strength index is generated based on the average strength and the rate of change. The causal strength index represents the overall causal propagation capability of the path.

[0129] Extract the node connection direction from the propagation path to generate a node pair set; the node pair set includes the start node identifier, end node identifier, and propagation direction information; connect the node pair set with the path index table to generate a node path mapping set; the node path mapping set records the propagation path, propagation direction, path number, and trigger time index of each node pair;

[0130] Construct an event mapping table, which combines each record item in the node path mapping set with its corresponding causal intensity index to form a four-tuple record structure of node-path-time-causal intensity; extract environmental state values ​​that match the trigger time index from the environmental data sequence, construct an environmental state index value set, and record the combined feature values ​​of each environmental factor at the corresponding time.

[0131] The four-tuple records in the event mapping table are associated and matched with the set of environmental state index values ​​to generate an event tag set. The event tag set has a five-tuple structure, which includes the propagation path identifier, node identifier, trigger time index, causal strength value, and environmental state index. The event tag set serves as a structured description of the propagation events in the stable control graph and is used for subsequent behavior frequency statistics and propagation weight adjustment operations.

[0132] In this embodiment, the cumulative number of triggers for each node in the propagation path is counted; if the cumulative number of triggers is greater than a fourth preset threshold, the causal propagation weight of the node in the stable control graph is adjusted, and the final result graph is output, specifically:

[0133] Extract the node identifier field and trigger time index field from the event tag set to construct a node trigger record table; perform aggregation operation on the node trigger record table according to the node identifier, count the total number of triggers for each node in all propagation paths, and generate a node trigger frequency list; the node trigger frequency list records each node identifier and its corresponding number of triggers;

[0134] The node trigger frequency list is compared with the fourth preset threshold to construct a threshold judgment table. The threshold judgment table marks all nodes whose trigger frequency exceeds the fourth preset threshold, forming a target node set. In the target node set, all propagation edges associated with the target nodes in the stable control graph are extracted to form a target propagation edge set. The target propagation edge set records the propagation start point identifier, end point identifier, causal propagation weight value, and time index.

[0135] For each propagation edge in the target propagation edge set, the propagation edge adjustment factor is calculated based on the value of its associated node in the node trigger frequency list and the number of propagation paths. The propagation edge adjustment factor is obtained by fitting an offset function, with the node trigger frequency and the number of connected propagation paths as inputs and the dynamic adjustment amplitude value as outputs.

[0136] The propagation edge adjustment factor is applied to the causal propagation weight value of the corresponding propagation edge, and a weighted correction operation is performed. The correction method is to make a linear adjustment based on the original propagation weight, and output the corrected propagation weight value. The correction operation keeps the original propagation direction and path structure unchanged and generates a propagation edge correction record table. The propagation edge correction record table records the weight value of the corrected propagation edge, the start and end node identifiers, and the source of the adjustment factor.

[0137] The propagation edge correction record table is merged with the propagation node set and path structure set in the original stable control graph to construct the final result graph. The final result graph includes the propagation node set, propagation edge set, path structure set, causal propagation weight update identifier, and node trigger frequency mapping table. The node trigger frequency mapping table marks the total number of triggers and whether a weight correction operation is triggered for each propagation node during the statistical period.

[0138] This implementation method constructs a statistical mechanism linking node trigger frequency and propagation path, further fitting the propagation adjustment factor with frequency as input, and further correcting the causal propagation edge weights, dynamically updating the graph weight structure, effectively improving the robustness and directionality of the causal graph in multi-time period abnormal recurrence scenarios.

[0139] Example 1:

[0140] To verify the feasibility of this invention in practice, it was applied to a 100MW centralized photovoltaic power station in a coastal area for a three-month operational test. This power station consists of over 2600 photovoltaic modules, connected to 32 combiner boxes, 8 inverters, and an environmental monitoring system, possessing complete signal acquisition channels and data interfaces. The test period spanned the late summer and early autumn climate fluctuation period, exhibiting typical high temperature and humidity, high wind and sand interference, and strong thunderstorms, making it highly representative.

[0141] During the operation of this photovoltaic power station, the original fault detection system mainly relied on threshold alarms and equipment self-test mechanisms, which could only detect inverter cascade faults and module failures, lacking the ability to identify hidden resonance propagation phenomena between nodes. In particular, when changes in module signals caused abnormal voltage in the combiner box, which in turn affected the temperature rise characteristics and power feedback of multiple branches, the traditional system could not trace such complex cross-node causal relationships, resulting in delayed alarm response, unclear diagnostic paths, and inefficient maintenance processes that relied on manual node-by-node troubleshooting.

[0142] After deployment, the system first collects voltage, current, temperature, power, and communication status signals from each component, combiner box, and inverter node, and constructs a synchronization map based on the equipment topology. The system uses a resonance identification module to capture frequently changing node pairs and constructs propagation paths based on an improved AGCN model, accurately identifying multiple high-frequency anomalous resonance clusters during the map construction phase. An anomaly removal module further filters conduction paths caused by non-equipment factors by incorporating environmental parameters (such as sudden temperature changes and wind speed interference), retaining only critical propagation edges driven solely by equipment structural coupling. Subsequently, the system performs loop detection on the propagation closed-loop structure and adds attenuation factors to weaken the interference of pseudo-enhanced paths on the overall causal propagation map, improving the system's ability to focus on genuine anomaly sources.

[0143] During the three-month operation, the system of this invention was compared and evaluated with traditional operation and maintenance testing platforms, and the results are shown in the table below:

[0144] Table 1. Comparison of Accuracy Rates for Multi-Source Anomaly Identification in Photovoltaic Power Plants

[0145] Exception types Accuracy of traditional system (%) Accuracy rate (%) of the system of this invention Increase (%) Combiner box voltage anomaly propagation identification 72.5 93.8 21.3 Component-level resonance anomaly identification 61.2 90.4 29.2 Identification of communication status fluctuations 75.3 92.1 16.8 Multi-node chain anomaly identification 58.9 87.6 28.7

[0146] As shown in Table 1, the system of this invention significantly outperforms traditional systems in recognition accuracy across various anomaly scenarios, particularly in component-level resonance and multi-node chain anomalies, where the improvement is nearly 30%. This is because the present invention, through resonance recognition and causal graph modeling, can understand the propagation relationship of abnormal signals in the network, no longer limited to single-point feature fluctuations. Simultaneously, multimodal perception and dynamic adjacency mechanisms significantly enhance the expressive power of abnormal paths, maintaining stable judgment capabilities even under environmental disturbances.

[0147] In addition, to verify the system's improvement in fault early warning and repair efficiency, the system's response speed, accurate fault location capability, and repair time during this period were statistically analyzed. The results are shown in the table below:

[0148] Table 2 Comparison of Fault Handling Efficiency of Photovoltaic Power Plants

[0149] Evaluation indicators Traditional system mean The system mean of this invention Efficiency improvement rate Fault response time (minutes) 18.6 5.2 71.0% Fault location error (node) 3.4 0.8 76.5% Average repair time (hours) 6.9 3.2 53.6%

[0150] Table 2 reflects the comprehensive improvement in fault handling efficiency brought about by the system of this invention. Regarding response time, traditional systems rely on periodic polling and manual judgment, resulting in an average response time of nearly 20 minutes. In contrast, this invention can trigger graph analysis immediately during signal interaction, outputting a diagnostic report within just 5 minutes. In terms of positioning accuracy, this system iteratively adjusts node weights through propagation path strength, ultimately achieving accurate identification of anomaly sources with node errors controlled to within one. Simultaneously, rapid positioning and path focusing significantly shorten the subsequent troubleshooting scope and operation time for maintenance personnel, reducing the overall average repair cycle by nearly half.

[0151] This embodiment not only overcomes the problems of propagation ambiguity, complex paths, and interference-induced misjudgments in traditional photovoltaic power plant anomaly detection, but also achieves high-confidence fault tracing based on true causal mechanisms, demonstrating significant advantages in ensuring power plant power generation stability and reducing operation and maintenance costs. The system is highly versatile and scalable, suitable for photovoltaic scenarios with various string sizes and heterogeneous topologies, and also shows promising prospects for widespread adoption in future energy architectures such as distributed photovoltaics and smart microgrids.

[0152] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A photovoltaic power station anomaly detection system based on multimodal sensing, characterized in that, include: The topology acquisition module is used to construct device topology diagrams and data synchronization diagrams; Collect signal data from the nodes and generate a multimodal signal dataset; The resonance identification module is used to calculate the correlation values ​​between nodes in the multimodal signal dataset; If the correlation value is greater than the first preset threshold in multiple consecutive sampling periods, it is determined that there is a resonance relationship between the nodes, and the node pairs with resonance relationship are marked as propagation edges to construct an initial propagation relationship graph. The graph construction module is used to input the initial propagation relationship graph into the improved AGCN model and output the causal propagation graph; the improved AGCN model includes a multimodal embedding unit, a dynamic adjacency generation structure, a graph convolutional interaction structure, a causal path attention structure, and a temporal cumulative coding structure; The anomaly removal module is used to collect environmental status data of photovoltaic power plants, construct environmental data sequences, and calculate the correlation value between each propagation edge in the causal propagation graph and the environmental data sequence. If the correlation value is greater than the second preset threshold and the causal strength value of the corresponding propagation edge is less than the third preset threshold, then the propagation edge is marked as an environment-driven edge and a path elimination operation is performed to generate a corrected propagation graph. The path control module is used to perform loop detection operations based on the modified propagation map to determine whether there is a closed path structure in the propagation path; If a closed structure exists in the propagation path and the cumulative propagation intensity is greater than the fourth preset threshold, then an attenuation factor is added to each propagation edge in the propagation path, the propagation spectrum is updated and corrected, and a stable control spectrum is output. The weight update module is used to generate an event label set based on the stable control graph; Count the cumulative number of triggers for each node in the propagation path; If the cumulative number of triggers is greater than the fourth preset threshold, then the causal propagation weight of the node in the stable control graph is adjusted, and the final result graph is output.

2. The photovoltaic power station anomaly detection system based on multimodal sensing according to claim 1, characterized in that, The steps between modules are as follows: S1. Construct a device topology diagram and data synchronization diagram; collect signal data from nodes and generate a multimodal signal dataset; S2. Calculate the correlation values ​​between nodes in the multimodal signal dataset; If the correlation value is greater than the first preset threshold in multiple consecutive sampling periods, it is determined that there is a resonance relationship between the nodes, and the node pairs with resonance relationship are marked as propagation edges to construct an initial propagation relationship graph. S3. Input the initial propagation relationship graph into the improved AGCN model and output the causal propagation graph; the improved AGCN model includes a multimodal embedding unit, a dynamic adjacency generation structure, a graph convolutional interaction structure, a causal path attention structure, and a temporal cumulative coding structure; S4. Collect environmental status data of photovoltaic power plants and construct environmental data sequences; Calculate the correlation value between each propagation edge in the causal propagation graph and the environmental data sequence; If the correlation value is greater than the second preset threshold and the causal strength value of the corresponding propagation edge is less than the third preset threshold, then the propagation edge is marked as an environment-driven edge and a path elimination operation is performed to generate a corrected propagation graph. S5. Perform loop detection based on the modified propagation graph to determine whether there is a closed path structure in the propagation path; if there is a closed structure in the propagation path and the cumulative propagation intensity is greater than the fourth preset threshold, then add an attenuation factor to each propagation edge in the propagation path, update the modified propagation graph and output a stable control graph. S6. Generate an event label set based on the stable control map; Count the cumulative number of triggers for each node in the propagation path; If the cumulative number of triggers is greater than the fourth preset threshold, then the causal propagation weight of the node in the stable control graph is adjusted, and the final result graph is output.

3. The photovoltaic power station anomaly detection system based on multimodal sensing according to claim 2, characterized in that, The construction of the device topology diagram and data synchronization diagram is specifically as follows: Construct a node set, which includes photovoltaic modules, combiner boxes, inverters, and monitoring nodes; Based on the electrical connection relationship and physical layout structure, the connection relationship between nodes is generated. The connection relationship includes the logical connection edge and the physical connection edge between nodes. The logical connection edge records the signal transmission channel, and the physical connection edge records the component wiring relationship and installation location information. A device topology graph is constructed based on the node set and connection relationship. The device topology graph is graph structure data, which includes a node identifier set, an edge set, and a node type label. Collect the sampling configuration parameters of each node, including sampling interval, sampling frequency, signal dimension, and data transmission delay; measure the physical distance between nodes, which is the Euclidean distance or path hop distance in the relative spatial coordinate system; A synchronization factor is constructed based on the difference in sampling interval, the difference in data transmission delay, and the physical distance between each pair of nodes; A synchronization weight matrix is ​​generated based on the synchronization factor. The synchronization weight matrix records the degree of synchronization matching of each pair of nodes during the multimodal signal acquisition process.

4. The photovoltaic power station anomaly detection system based on multimodal sensing according to claim 3, characterized in that, The signal data from the acquisition nodes is used to generate a multimodal signal dataset, specifically: Collect signal data from photovoltaic modules, combiner boxes, inverters and monitoring nodes. The signal data includes voltage signals, current signals, temperature signals, power signals and communication status signals. The acquired signal data is processed for time synchronization. The multi-source signals are time-aligned according to a uniform sampling period to generate a time-aligned sequence. Noise suppression processing is performed on each type of signal in the time-aligned sequence. High-frequency noise is removed by bandpass filtering for voltage and current signals, and smoothing is performed by moving average processing for temperature and power signals. Abnormal states are smoothed by moving window statistics for communication status signals. Normalization is performed on the noise-suppressed signal data. The maximum and minimum value normalization method is used to map the signal values ​​to a unified dimension range, generating a normalized signal set. A time index identifier is assigned to each type of normalized signal. The time index identifier records the sampling time and the sampling period number. Calculate the sampling confidence weight for each type of signal. The sampling confidence weight is generated based on sampling stability, signal integrity, and data delay. Combine the time index, normalized signal set, and sampling confidence weight to form a multimodal signal dataset, and output the multimodal signal dataset.

5. The photovoltaic power station anomaly detection system based on multimodal sensing according to claim 4, characterized in that, Specifically, S2 is: Based on the device topology diagram and the synchronization weight matrix, select node pairs with related structures as the candidate node pair set; Time-aligned signal sequences of candidate node pairs are extracted from the multimodal signal dataset, and multimodal feature vectors are constructed respectively. The multimodal feature vectors include voltage feature vector, current feature vector, temperature feature vector, power feature vector and communication state feature vector. For each type of feature vector, a corresponding modal similarity value is calculated. The modal similarity value is constructed based on the signal change trend, amplitude difference, and spectral characteristics. A weighted fusion operation is performed on each modal similarity value to generate a fused correlation value. The weighting weights are jointly determined based on the modal confidence weights and the synchronization weight matrix. The fusion correlation value is tracked over multiple consecutive sampling periods. If the fusion correlation value is continuously greater than the first preset threshold for a preset number of sampling periods, it is determined that the corresponding node pair has a resonance relationship. The node pair with resonance relationship is marked as a propagation edge. The propagation edge records the corresponding node pair identifier, start and end time index and correlation strength. All propagation edges are combined to generate an initial propagation relationship graph. The initial propagation relationship graph contains a set of propagation nodes, a set of propagation edges and correlation labeling information.

6. The photovoltaic power station anomaly detection system based on multimodal sensing according to claim 5, characterized in that, The improved AGCN model includes a multimodal embedding unit, a dynamic adjacency generation structure, a graph convolutional interaction structure, a causal path attention structure, and a temporal cumulative coding structure, specifically: The multimodal embedding unit constructs independent modal embedding channels for voltage, current, temperature, power and communication status signals in the multimodal signal dataset. Each embedding channel extracts modal feature vectors based on a one-dimensional convolutional structure and outputs a set of modal embedding tensors. The dynamic adjacency generation structure constructs an initial adjacency matrix based on the device topology graph and the synchronization weight matrix. It integrates modal confidence weights and inter-node synchronization factors to calculate the dynamic connection probability between each pair of nodes and generate a dynamic adjacency matrix set. The dynamic adjacency matrix set contains edge connection weight information under different time indices. The graph convolutional interactive structure performs graph convolution operations on the modality embedding tensor set respectively. The graph convolution operations are based on the dynamic adjacency matrix to generate the propagation feature representation of each modality. The propagation feature representation is input into the interactive fusion path, which calculates the modality fusion weight coefficients based on the modality confidence weights and outputs the modality fusion propagation feature vector. The causal path attention structure constructs a set of candidate propagation paths based on the initial propagation relationship graph, extracts the position information of the propagation edges, propagation intensity and modality consistency index in each propagation path, and calculates the path attention score by combining the path structure features. The path attention score is used to weight the path context information to generate a path-level causal representation tensor. The temporal cumulative coding structure constructs a propagation intensity time series based on the path-level causal representation tensor. The propagation intensity time series is input into the gated cyclic unit structure for sequence modeling, and outputs the causal propagation intensity value of each propagation edge at different time indices, which constitutes the edge weight information of the causal propagation graph.

7. The photovoltaic power station anomaly detection system based on multimodal sensing according to claim 6, characterized in that, Specifically, S4 is: Collect environmental status data of photovoltaic power plants, including light intensity, ambient temperature, wind speed, humidity and rainfall; perform time alignment processing on the environmental status data, and construct an environmental data sequence according to a unified sampling period and time index. The environmental data sequence records the numerical changes of each environmental parameter at different sampling times. The correlation value between the propagation edge and the environmental data is calculated based on the modified propagation map and the environmental data sequence. The correlation value is calculated based on the time synchronization factor, amplitude response factor and change trend factor. At each time index, the causal strength value of the propagation edge and the gradient of the corresponding environmental parameter change are extracted to form a correlation calculation sample set; Perform correlation calculation on the correlation calculation sample set to generate a correlation degree matrix, which records the correlation index between each propagation edge and each environmental parameter; The maximum correlation value and the corresponding causal strength value of each propagation edge in the correlation matrix are jointly determined. If the correlation value is greater than the second preset threshold and the causal strength value is less than the third preset threshold, the corresponding propagation edge is marked as an environment-driven edge. All environment-driven edges are removed from the modified propagation graph, the propagation edge set is updated, and the modified propagation graph is generated. The modified propagation graph includes a propagation node set, a propagation edge set, correlation labeling information, and time index records.

8. The photovoltaic power station anomaly detection system based on multimodal sensing according to claim 7, characterized in that, Specifically, S5 is: Extract the propagation edge set and propagation node set from the modified propagation graph to construct the propagation path set; perform closed path detection operation on the propagation path set. A closed path is a path structure in the propagation path where the start node identifier and the end node identifier are the same and the path length is not less than three. For each closed path in the propagation path set, calculate the cumulative propagation intensity value, which is the sum of the causal intensity values ​​of all propagation edges in the path; If the cumulative propagation intensity value is greater than the fourth preset threshold, then an attenuation factor is assigned to all propagation edges in the closed path. The attenuation factor is calculated and generated based on the cumulative propagation intensity value and the propagation path length. The attenuation factor is applied to the causal strength value of the corresponding propagation edge to generate an updated propagation edge set; the updated propagation edge set is combined with the propagation node set to construct a stable control graph; the stable control graph includes the propagation node set, the propagation edge set, the path structure set, and causal strength correction labeling information.

9. A photovoltaic power station anomaly detection system based on multimodal sensing according to claim 8, characterized in that, The generation of the event label set based on the stable control map specifically includes: Extract the propagation node set, propagation edge set, and path structure set from the stable control graph, and establish a node index table and a path index table; determine the trigger time range of the propagation path based on the time index record, and arrange the propagation paths in chronological order; calculate the average propagation intensity and the rate of change of propagation intensity for each propagation path, and generate a causal intensity index by combining the causal intensity values ​​recorded in the propagation edge. Extract node identifiers and propagation direction information from each propagation path to form a set of node pairs; map the set of node pairs to the propagation path identifiers to construct an event mapping table; Record node identifiers, propagation path identifiers, trigger time indexes, and causality strength indicators in the event mapping table; Extract environmental state information corresponding to the trigger time index from the environmental data sequence and calculate the environmental state index value; merge the environmental state index value with the record items in the event mapping table to generate an event tag set; the event tag set includes propagation path identifier, node identifier, trigger time index, causal strength value and environmental state index.

10. A photovoltaic power station anomaly detection system based on multimodal sensing according to claim 9, characterized in that, The cumulative number of triggers for each node in the propagation path is counted; if the cumulative number of triggers is greater than a fourth preset threshold, the causal propagation weight of the node in the stable control graph is adjusted, and the final result graph is output, specifically: Extract the node identifier field and trigger time index field from the event tag set to construct a node trigger record table; perform aggregation operation on the node trigger record table according to the node identifier, count the cumulative number of triggers of each node in all propagation paths, and generate a node trigger frequency list. Compare the list of node trigger frequencies with the fourth preset threshold, and filter out the set of target nodes whose cumulative trigger count is greater than the fourth preset threshold; For each node in the target node set, extract all propagation edges connecting the target nodes in the stable control graph; An adjustment factor is added to the causal strength value of each propagation edge. The adjustment factor is calculated and generated based on the node trigger frequency and the number of propagation paths. During the adjustment process, the propagation direction remains unchanged, and the causal propagation weight value in the propagation edge set is updated to form a propagation edge correction record. The propagation edge correction record is combined with the original propagation node set and path structure set to construct the final result graph; the final result graph includes the propagation node set, propagation edge set, path structure set, causal propagation weight update identifier and node trigger frequency mapping table.

Images