Photovoltaic array intelligent operation and maintenance decision system based on large language model

Abstract

The application discloses a photovoltaic array intelligent operation and maintenance decision system based on a large language model, relates to the technical field of fault evolution modeling, collects thermal imaging images, electrical parameters, meteorological data and operation and maintenance text records of photovoltaic components, and constructs multi-modal semantic embedding vectors; an operation and maintenance semantic graph is constructed based on component space topology and electrical connection relationship, and a hidden fault candidate set is identified through a graph attention mechanism; in combination with a large language model and a fault variation factor prompt template, semantic completion is performed on missing or unobservable data, and a semantic feature residual tensor is generated; through comparison with a historical abnormal event path, a causality similarity index curve is calculated, and a fault evolution graph is constructed based on trend changes; finally, the large language model driven instruction generation engine is input, and a multi-target optimized structured operation and maintenance suggestion is output; the application realizes early warning, evolution prediction and intelligent scheduling of photovoltaic faults, and improves operation and maintenance efficiency and decision intelligent level.

Description

Technical Field

[0001] This invention relates to the field of fault evolution modeling technology, specifically to a photovoltaic array intelligent operation and maintenance decision-making system based on a large language model. Background Technology

[0002] Currently, photovoltaic power plants are widely deployed globally. However, photovoltaic arrays face various complex environmental conditions during actual operation, such as dust pollution, module shading, wiring faults, and hot spot effects. These factors can significantly reduce the power generation efficiency of photovoltaic modules and even cause safety hazards such as fires. Therefore, the operation and maintenance management of photovoltaic arrays is gradually shifting from "manual periodic inspections" to "intelligent sensing + predictive decision-making".

[0003] Traditional photovoltaic array fault detection and operation and maintenance decision-making methods are mostly based on image recognition, electrical parameter analysis, or rule-driven models, which have the following main technical problems: Data silos and severe heterogeneity: Data sources are diverse, including infrared images, voltage and current sensors, meteorological data, operation and maintenance records, etc., with large differences in data formats, making it difficult to unify modeling; Limitations of knowledge representation: Existing methods rely on fixed rules or shallow neural networks, making it difficult to effectively learn complex causal relationships and semantic context; Minor faults are difficult to identify: such as slight shading or component performance drift, which do not significantly affect the overall power generation in the early stages and are often overlooked, leading to maintenance delays. Lack of global decision-making optimization capability: Current decision-making strategies are mostly based on the optimization of local indicators, making it difficult to take into account the global risk evolution trend and resource allocation efficiency.

[0004] Breakthroughs in large language models (LLM) in cross-modal semantic understanding, knowledge extraction, and reasoning generation have provided new technical pathways for solving the aforementioned problems. By introducing large language models with multimodal fusion and causal reasoning capabilities, deep semantic perception and high-precision intelligent operation and maintenance decisions regarding the operating status of photovoltaic arrays can be achieved. Summary of the Invention

[0005] The purpose of this invention is to provide a photovoltaic array intelligent operation and maintenance decision-making system based on a large language model to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a photovoltaic array intelligent operation and maintenance decision-making system based on a large language model, comprising: Data acquisition and fusion module: acquires the original heterogeneous operation dataset of the target photovoltaic power station, including component-level thermal imaging sequence, historical change data of AC and DC side electrical parameters, regional meteorological time series data, and historical maintenance records text of operation and maintenance engineers; Feature alignment module: The original heterogeneous operating dataset is decoupled and normalized in terms of mode, and a time-series vector set, a heat map set and a text semantic stream are constructed respectively, and a group of photovoltaic operating status semantic embedding vectors with semantic alignment capability is generated. Operation and maintenance semantic graph construction module: Based on the predefined component space topology and electrical path dependency graph, construct the operation and maintenance semantic embedding graph of the photovoltaic array, and mine the potential semantic coupling relationship between weakly connected nodes through the graph attention mechanism to generate an initial set of hidden fault candidates; Semantic reasoning and completion module: For the initial set of hidden fault candidates, combined with the fault variation factor reverse reasoning mechanism introduced in the large language model, high-confidence semantic completion is performed on the missing or unobservable data regions to generate component-level semantic feature residual tensors. Evolutionary modeling module: Based on the multi-layer nested comparison between the semantic feature residual tensor and the evolution path of historical anomalous events, calculate the causal similarity index curve of potential failure events, and construct the failure evolution map of the photovoltaic array based on its trend change. Operation and maintenance instruction generation module: Input the fault evolution diagram into the operation and maintenance instruction generation engine driven by the large language model, and output natural language structured operation and maintenance suggestions containing multi-objective optimization strategies.

[0007] Preferably, the step of constructing the operation and maintenance semantic embedding graph of the photovoltaic array based on the predefined component spatial topology and electrical path dependency graph includes: The semantic embedding vector of each photovoltaic module is bound to its physical spatial coordinates in the array to generate a set of module-level nodes; Based on the electrical connection relationship of components and the actual wiring structure, a set of electrical path dependency edges between nodes is established, and edge weights are assigned to represent connection strength or fault propagation probability. The graph attention mechanism in graph neural networks is used to perform feature aggregation on the above set of nodes and edge sets; The aggregated graph structure is optimized by structured embedding to generate an operation and maintenance semantic embedding graph of the photovoltaic array.

[0008] Preferably, the step of generating the initial set of latent fault candidates includes: The graph attention mechanism is used to perform multi-level adjacency node feature aggregation on each component node in the operation and maintenance semantic embedding graph; The semantic similarity index and the rate of change of connection weights between node pairs are calculated to identify weak connection node pairs with low structural strength but prominent semantics. Based on weakly connected nodes, potential fault impact areas are constructed, and local clustering analysis is performed by integrating historical operational status change trends within the areas. The subset of nodes with anomalous semantic distribution and deviation from historical evolutionary trajectories in the clustering results are identified as the initial latent fault candidate set, and their positions and temporal labels in the graph are associated with them.

[0009] Preferably, the step of performing high-confidence semantic completion of missing or unobservable data regions by combining the fault variation factor reverse reasoning mechanism introduced in the large language model includes: Anomaly scores are calculated for the multimodal embedding vectors of each component node in the initial set of latent fault candidates, and nodes with scores higher than a set threshold are selected as inference targets. Based on the structured knowledge corpus and historical operation and maintenance semantic embedding samples built into the large language model, a causal relationship prompt template containing fault category, evolution path and mutation mode is constructed. The semantic context information of the reasoning target node and the prompt template are jointly input into the large language model, and the semantic feature vector of the running state of the missing or unobservable data region is output. The semantic feature vector of the running state is compared with the original embedding vector by dimension-wise difference calculation to obtain the component-level semantic feature residual tensor that reflects the degree of fault change.

[0010] Preferably, the step of jointly inputting the semantic context information of the reasoning target node and the prompt template into the large language model includes: The semantic embedding vector of the reasoning target node is expanded by time window, its state change features in continuous running cycle are extracted, and a context description sequence containing component number, spatial location and historical running trend is generated. The context description sequence is embedded into a pre-built fault causal relationship prompt template to form a joint semantic input sequence containing the operating background, potential anomaly types, and inference target constraints; The joint semantic input sequence is fed into a large language model, and the runtime semantic feature vector corresponding to the missing or unobservable data region is predicted through conditional generation.

[0011] Preferably, the step of calculating the causal similarity index curve of potential failure events includes: An evolution path sample library containing residual tensor sequences with semantic features of various types of abnormal events is constructed based on historical operation and maintenance data, and a set of staged evolution trajectories is formed in chronological order. The semantic feature residual tensor sequence of the current component node is matched and calculated stage by stage with the residual features of each stage in the evolution path sample library to obtain the feature offset similarity value of each time stage. Construct time series similarity curves based on the similarity change trends over continuous time periods; The smoothed similarity curve is used as the causal similarity index curve of potential failure events.

[0012] Preferably, the steps for constructing a fault evolution graph of a photovoltaic array based on the trend changes of the causal similarity index curve include: extracting the slope changes of the continuously rising intervals in the causal similarity index curve, identifying time segments in the early, middle, or accelerated phases of fault evolution, and marking their positions on the time axis; performing cluster analysis on multiple component nodes with similar time trends, constructing a semantic propagation relationship graph between node subsets, and establishing cross-node evolution coupling paths; introducing component spatial location and electrical connection topology information, mapping the causal propagation paths between nodes to the actual array structure, and constructing a fault propagation chain with spatial continuity; dynamically updating the graph structure edge weights according to the directionality of the fault propagation chain and the node activity, and finally forming a fault evolution graph characterizing the local or global fault propagation trend of the photovoltaic array.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a fault reasoning mechanism and causal similarity index curve driven by semantic feature residual tensors by combining a large language model with multimodal semantic modeling. Compared with traditional photovoltaic fault detection methods that rely on fixed thresholds or single-modal features, this invention can effectively identify the evolution process of hidden faults in complex backgrounds, and has stronger fault trend capture capabilities, semantic completion capabilities, and early warning performance, realizing a deep understanding and prediction of the operating status of photovoltaic arrays.

[0014] 2. The fault evolution graph constructed in this invention integrates spatial topology, electrical dependencies, and semantic propagation paths, enabling it to dynamically describe the fault propagation chain from local to global scope. This provides accurate and structured input for the subsequent generation of operation and maintenance instructions driven by a large language model. This solution not only improves the accuracy and response efficiency of fault identification but also enhances the interpretability, scalability, and intelligent resource scheduling level of the intelligent operation and maintenance system, demonstrating significant engineering application value. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] For examples, please refer to Figure 1 As shown in this embodiment, the photovoltaic array intelligent operation and maintenance decision-making system based on a large language model includes: Data Acquisition and Fusion Module: Acquires the original heterogeneous operation dataset of the target photovoltaic power station, including component-level thermal imaging sequence, historical change data of AC and DC side electrical parameters, regional meteorological time series data, and historical maintenance records of operation and maintenance engineers.

[0019] In this invention, the data acquisition and fusion module serves as the data entry point for the entire system, undertaking the crucial tasks of acquiring, preprocessing, and unifying the format of multi-source heterogeneous operating data. Its accuracy and completeness directly affect the effectiveness of subsequent multimodal feature fusion and semantic reasoning.

[0020] Specifically, the data acquisition and fusion module is used to acquire the original heterogeneous operation dataset of the target photovoltaic power station, which includes the following four main types of data: Module-level thermal imaging sequence: Thermal infrared image data of the photovoltaic module surface is periodically acquired using a drone infrared inspection system or a fixed thermal imaging camera, with image time intervals ranging from minutes to hours. These images reflect the surface temperature distribution of the module and can be used to identify potential hazards such as hot spot effects, localized shading, and connection anomalies. During acquisition, image timestamps can be linked to the module number to ensure spatiotemporal accuracy and traceability.

[0021] Historical changes in AC and DC side electrical parameters: Real-time acquisition of electrical parameters of the photovoltaic array, including but not limited to module voltage, current, power, power factor, frequency, etc., divided into two channels: DC side (module output) and AC side (inverter output). The data sampling frequency can be 1Hz~10Hz, and is buffered and uploaded through local edge devices. This type of data reflects the module's operating status and electrical stability, and is an important basis for identifying abnormal power drift, mismatch faults, etc.

[0022] Regional meteorological time-series data: Environmental parameters of the geographical location are collected using on-site meteorological monitoring equipment or by connecting to a third-party meteorological platform, including solar radiation intensity, temperature, humidity, wind speed, wind direction, rainfall, and PM2.5 concentration. Meteorological data serves as an environmental disturbance factor, used in the model to eliminate the impact of operational anomalies caused by non-equipment failures, thereby enhancing the robustness and interpretability of predictions.

[0023] Maintenance engineers' historical repair records: This category collects textual records from the on-site maintenance management system, including fault descriptions, repair suggestions, work logs, component numbers, and annotations on manual inspection photos, typically in natural language format. This data, after text mining and semantic extraction, can be used to construct a semantic prior knowledge base, assisting large language models in understanding equipment state evolution paths and typical fault semantic labels.

[0024] To achieve efficient fusion of the above data, the data acquisition and fusion module further includes the following sub-modules: Unified Timeline Alignment Submodule: Performs timestamp calibration on multi-source data, unifying it to a unified UTC timeline, thus resolving alignment error issues caused by asynchronous data acquisition; Data Identification and Association Submodule: Based on metadata such as component ID, inverter number, and physical location information, establish association index relationships between data to ensure the accuracy of subsequent feature mapping; Data integrity detection and missing value marking submodule: Performs quality detection on uploaded data, identifies missing segments, abnormal jumps or invalid values, and fills in missing values ​​or retains marks for subsequent inference and completion processing; Format Conversion and Caching Submodule: Converts raw data (such as .csv, JSON, image sequences, etc.) from different acquisition systems into a unified data structure format supported by the system, and caches and uploads them locally and periodically by component.

[0025] In this embodiment, the design of the data acquisition and fusion module fully considers the high-dimensionality, multimodal, and time-varying characteristics of photovoltaic power plant operation data, providing a structured and standardized data foundation for subsequent multimodal collaborative embedding and semantic graph construction. Simultaneously, by introducing natural language maintenance records as an important data source, the limitation of traditional operation and maintenance systems relying solely on structured numerical data is broken, significantly improving the system's ability to understand "non-obvious faults" and "semantically ambiguous events."

[0026] Feature alignment module: The original heterogeneous operating dataset is decoupled and normalized in terms of mode, and a time-series vector set, a heat map set and a text semantic stream are constructed respectively, and a group of semantic embedding vectors of photovoltaic operating status with semantic alignment capability is generated.

[0027] In this invention, the feature alignment module, as a key link in the data processing flow, is responsible for modal decoupling and standardization of the heterogeneous raw data from the "data acquisition and fusion module," and for constructing a unified semantic expression format, laying the embedded foundation for subsequent semantic graph modeling and inference. The goal of this module is to solve the following core technical problems: There are significant inconsistencies in physical dimensions, semantic structure, and data granularity among different modalities of data (images, time series, text). Traditional methods lack the ability to semantically align multimodal information, resulting in semantic drift and distortion even after feature fusion. Some key state features are only manifested in a certain modality and are difficult to explicitly align to component-level state units.

[0028] Therefore, this module introduces a three-stage strategy of modal decoupling, standardization, and semantic alignment. The specific steps include: Modal decoupling processing: The original heterogeneous data is divided into modalities according to the data type, and the core expression form within the modality is extracted: thermal imaging image data → decoupled into [time frame sequence + thermal image pixel matrix], and component-level temperature difference distribution map is extracted for each frame; Electrical parameter timing data → decoupled into [component number + voltage / current / power time series], retaining sampling frequency information; Meteorological data → decoupled into [time point + multi-factor vector (such as irradiance, wind speed, temperature, etc.)], and bound to station identifiers; Text inspection records → Segmentation and annotation of key entities (such as component number, event type, and exception description) are performed to construct structured semantic tuples.

[0029] To eliminate differences in scale, dimensions, and distribution among different modalities, normalization transformation is performed on various modal data: z-score normalization was applied to the time series data; The image data is processed by centering and standardization, and local heatmaps are constructed using sliding window cropping. Pre-trained language models such as BERT are used to extract contextual semantic embedding vectors from text data, preserving inter-sentence dependencies; Align data segments from different modalities to the same time period (e.g., using a sliding window based on hours).

[0030] Based on modal processing, the following are constructed respectively: Time-series vector set: Represents the electrical and meteorological change characteristics at the component level, with the structure [component ID, feature dimension, time window]; Thermal map set: a two-dimensional spatial thermal feature map representing the surface state of a component, with the structure [component ID, number of thermal image frames, pixel matrix]; Text semantic stream: Represents the fault semantic stream from inspection and maintenance records, with the structure of [event entity embedding sequence + timestamp + operation type].

[0031] By employing multimodal attention mechanisms, cross-modal contrastive learning, and a shared embedding space mapping function, the three types of representation objects are projected into a unified semantic embedding space. This ultimately generates a set of semantic embedding vectors for the photovoltaic operating status of each component within a given time window, which are then used as input for subsequent semantic graph construction and inference engine.

[0032] Embedded vector groups have the following characteristics: Each vector is strongly bound to a component number, ensuring structural traceability; The vector dimension is unified, which facilitates batch processing and downstream model integration; The vector internally encodes temporal evolution features, spatial heat distribution features, and fault semantic context features, possessing high-dimensional information fusion capabilities.

[0033] Operation and maintenance semantic graph construction module: Based on the predefined component space topology and electrical path dependency graph, construct the operation and maintenance semantic embedding graph of the photovoltaic array, and mine the potential semantic coupling relationship between weakly connected nodes through the graph attention mechanism to generate an initial set of hidden fault candidates.

[0034] The semantic embedding vector of each photovoltaic module is bound to its physical spatial coordinates in the array to generate a set of module-level nodes.

[0035] Before constructing the operation and maintenance semantic embedding graph, the nodes in the graph structure must first be defined. First, for each photovoltaic module, based on the photovoltaic operation status semantic embedding vector obtained in the previous processing step, it is bound one-to-one with the module's physical spatial location information in the array. This spatial location information is represented in two-dimensional coordinates, including the module's row and column indices in the array. Through this binding relationship, each photovoltaic module is defined as a graph node, and the node's attributes include its corresponding semantic embedding vector (representing the operation status), physical coordinates (representing the spatial location), and module number. Finally, all component nodes constitute a component-level node set for graph structure initialization.

[0036] Based on the electrical connection relationship of components and the actual wiring structure, a set of electrical path dependency edges between nodes is established, and edge weights are assigned to represent connection strength or fault propagation probability.

[0037] Based on the actual electrical design drawings of the photovoltaic array, determine the series and parallel connections between the modules. Then, construct the edge set in the graph based on the cable lengths, current carrying capacities, and inverter access topology corresponding to the module connections. Each edge connects two module nodes, indicating an electrical path between them. To enable the graph structure to be learnable, each edge needs to be assigned a numerical weight. This edge weight can be set as follows: The edge weight is defined as an index of the electrical coupling strength between components. The calculation method is as follows: based on the DC current correlation coefficient between components, Pearson correlation analysis is performed on the current changes of the two components within the sampling period; if the absolute value of the correlation coefficient is higher than 0.8, the connection strength is considered to be high. Alternatively, the edge weight can be defined as the probability of fault propagation between components. Its value can be estimated based on historical fault propagation statistics or the inverse function of the physical path length of the cabling. If the cabling distance is short, the weight is set to 0.9, and if the distance is long, it is reduced to below 0.4.

[0038] Finally, an edge set containing the electrical dependencies between all component nodes is generated and assigned weighted edge attributes, forming a preliminary graph structure.

[0039] The graph attention mechanism in graph neural networks is used to perform feature aggregation on the above set of nodes and edge sets.

[0040] To learn the latent semantic relationships between nodes in a graph structure, after the graph is constructed, a graph attention mechanism from graph neural networks is introduced for feature propagation and aggregation. Specifically, a multi-head graph attention network is used, which learns the importance weights between each pair of adjacent nodes to achieve adaptive weighted feature fusion. The implementation steps are as follows: For each component node, iterate through its set of adjacent nodes; Calculate the attention weights between the node and each of its neighboring nodes. The attention weights are obtained by weighting the embedding vectors of the two nodes using a dot product function and then inputting them into the LeakyReLU activation function. All attention weights are normalized using Softmax to ensure that the total weight sum is 1; The normalized weights are used to perform a weighted summation of the embedding features of adjacent nodes, which is then used as the updated feature representation of the current node. Repeat the above process for 3 to 5 iterations to allow the semantic influence between distant nodes in the graph to propagate.

[0041] In the final output graph structure, each component node has an updated embedding representation that has been fused with adjacent feature information, which is used for subsequent anomaly detection tasks.

[0042] The aggregated graph structure is optimized by structured embedding to generate an operation and maintenance semantic embedding graph of the photovoltaic array.

[0043] After graph attention feature aggregation is completed, a structured embedding optimization step is introduced to further compress information dimensionality and enhance structural expressiveness. Specifically, a graph autoencoder network is used to reconstruct the representation of all graph nodes. The process includes: During the encoding stage, graph convolution operations are used to compress the embedding vectors of all nodes into a low-dimensional representation, reducing the dimension from the original 128 dimensions to 32 dimensions. In the decoding phase, the inner product decoding function is used to predict the connection relationship between any two nodes; The graph embedding vector is optimized by training the cross-entropy loss between the predicted values ​​and the real graph connection structure. The final output is the embedded representation of each node after structural optimization, forming a full-graph form of the semantic embedding graph of photovoltaic array operation and maintenance.

[0044] This map will be used in the subsequent anomaly clustering and fault reasoning stages, and it features clear representation, stable structure, and significant coupling relationships.

[0045] The graph attention mechanism is used to perform multi-level adjacency node feature aggregation on each component node in the operation and maintenance semantic embedding graph.

[0046] To enhance the ability to detect potential latent faults, the constructed operational semantic embedding graph undergoes further semantic enhancement. A multi-layer graph attention network (e.g., three layers) is used to propagate multi-hop adjacency information to each node, enabling each node to not only perceive its direct neighbors but also integrate the semantic information of second- or third-order neighbors.

[0047] In each layer, attention weights are recalculated, and a residual connection mechanism is used to preserve the original state of nodes to avoid over-smoothing. The final output representation is used for anomaly relationship identification.

[0048] The semantic similarity index and the rate of change of connection weights between node pairs are calculated to identify weakly connected node pairs with low structural strength but prominent semantics.

[0049] Next, we analyze any pair of nodes in the graph, calculating their semantic similarity and the trend of changes in their connection relationships. The semantic similarity is measured by the cosine similarity between the embedding vectors of the two nodes, ranging from 0 to 1; the closer to 1, the closer the semantics. The rate of change of connection weights is calculated as follows: Calculate the average change in edge weights over the past 7 days, and then calculate the ratio of the current weight change to that average. If the rate of change is greater than 30%, the connection is considered to be state unstable. When two nodes have a weak structural connection (edge ​​weight less than 0.3), but a semantic similarity greater than 0.9 and a connection weight change rate greater than 30%, they are identified as a weak connection node pair, which may hide early signs of fault propagation.

[0050] Based on weakly connected nodes, potential fault impact areas are constructed, and local clustering analysis is performed by integrating historical operational status change trends within these areas.

[0051] Using all identified weakly connected node pairs as centers, a local subgraph containing potential anomalous relationships is constructed by expanding outwards to their first-order neighbors. Cluster analysis is then performed on the embedding vectors of all nodes in this subgraph using a density-based spatial clustering method (DBSCAN), with a minimum sample size of 5 and a semantic vector Euclidean distance threshold of 0.5.

[0052] Meanwhile, the temperature fluctuation range and current instability index of each node over the past 14 days are superimposed as weighting factors during the clustering process to improve the sensitivity of the clustering results to abnormal states.

[0053] The subset of nodes with anomalous semantic distribution and deviation from historical evolutionary trajectories in the clustering results are identified as the initial latent fault candidate set, and their positions and temporal labels in the graph are associated with them.

[0054] Ultimately, the set of nodes identified as outliers or high-density anomalous regions in the cluster analysis will serve as the initial candidate set of latent faults. Each component node in this set needs to be bound to its spectral spatial coordinates (i.e., physical location) and its current time window label for use in subsequent mutation factor inference and fault evolution modeling stages.

[0055] This process ensures early detection of potential latent faults caused by minute feature drift and weak structural coupling, providing a source of risk input for the entire operation and maintenance decision-making process.

[0056] Semantic Reasoning and Completion Module: For the initial set of hidden fault candidates, combined with the fault variation factor reverse reasoning mechanism introduced in the large language model, high-confidence semantic completion is performed on the missing or unobservable data regions to generate component-level semantic feature residual tensors.

[0057] In this invention, to achieve accurate semantic completion of missing or unobservable data regions and further capture anomalous offset features of component-level operational states, a fault variation factor reverse inference method combining a large language model is proposed. This method generates semantic feature residual tensors for component nodes with high anomalous behavior in the initial set of latent fault candidates, serving as input for subsequent fault evolution modeling. The complete implementation steps are as follows: First, for each component node in the initial set of latent fault candidates, state volatility analysis is performed based on its multimodal semantic embedding vector time series. The specific method is as follows: The time window length is set to 7 consecutive days, and the semantic embedding vector of each day is extracted for each node (each vector is 128-dimensional). Calculate the Euclidean distance between the embedding vectors of each pair of adjacent dates, accumulate them, and take the average as the "mean of state perturbation" for that node; The perturbation mean of all candidate nodes is standardized and normalized to between 0 and 1; Set an anomaly score threshold, for example, 0.75, indicating that the node state change significantly exceeds the upper limit of normal fluctuation; Nodes with abnormal scores higher than the threshold are included in the "inference target node set".

[0058] This scoring method eliminates nodes whose state fluctuations are insufficient to constitute anomalies, retaining only nodes with high uncertainty and high risk of missing data for the next step of completion reasoning.

[0059] To improve the accuracy of semantic completion and the consistency of causal relationships within the context, a "fault causal relationship prompt template" was developed, incorporating professional knowledge from the photovoltaic equipment operation and maintenance field. The template consists of the following three parts: Fault type structure: including six types of faults: obstruction, loose wiring, voltage deviation, output attenuation, local overheating, and power interruption; Evolutionary path structure: Records the historical change sequence of each fault in the semantic embedding space, such as "voltage stabilizes → slight decrease → power drops suddenly → temperature rise fluctuation"; Variation factor structure: Considering the influence of factors such as season, regional location, array density, and component aging degree, the expression changes of the same type of fault under different scenarios are defined.

[0060] The prompt template is expressed in text form, with reserved semantic padding space for subsequent embedding of target node context information.

[0061] For each inference target node, extract its observable historical state and operating environment to construct context information. The operation steps are as follows: Extract the node number (e.g., A108), two-dimensional spatial location (e.g., row 6, column 12), and the number of the series circuit to which it belongs (e.g., column 4). Obtain the multimodal embedding vector sequence for the past 7 days, analyze trend changes, and extract event descriptions such as "temperature continues to rise", "voltage drops and then stabilizes", and "power anomaly occurs in adjacent component B107". Fill the above context information into the prompt template to form a natural language sequence with operational background, fault clues and time constraints, which serves as the conditional input for the language model.

[0062] Example input sequence is as follows: "Component A108 is located in row 6 and column 12 of the photovoltaic array. Over the past 7 days, its temperature has been continuously rising, and its voltage has been decreasing since day 3. The neighboring component B107 has reported a loose wiring fault. Current data is missing. Based on the historical evolution trajectories of shading and wiring anomaly categories, infer that the current state of this component should possess the following semantic feature vector:" The constructed joint semantic input sequence is then fed into a large language model that has been fine-tuned using a multimodal semantic alignment task, and a conditional generation operation is performed: The model architecture is based on the Transformer encoding and decoding structure, which supports the generation of semantic vectors from natural language; The model used has been trained using photovoltaic equipment corpus (including image annotations, descriptions of electrical parameter anomalies, expert maintenance statements, etc.); The model output is a set of real number vectors of length 128, corresponding to the current running state of the component in the semantic embedding space; The semantic dimension corresponding to each dimension has been bound during the training phase. For example, the 12th dimension represents "temperature stability" and the 57th dimension represents "voltage noise". The generated results achieve semantic completion based on the scenario and historical data by matching the historical trend of the current input with the fault mode prompts through attention.

[0063] To quantify the deviation between the completion result and the original state, a "semantic feature residual tensor" is constructed using dimension-wise interpolation calculation, as follows: Extract the embedding vector V{obs} (128-dimensional) of the last observable time point of the inference target node; Compare the semantic completion vector V{pred} (128-dimensional) generated by the language model. Perform vector interpolation: For each dimension i, calculate the residual. ; Concatenate all dimensional residuals into a residual tensor R=[R1,R2,...,R{128}], which serves as the semantic offset representation of the current component; If the number of dimensions in the residual tensor that exceed a specified sensitivity threshold (e.g., a residual of a certain dimension is greater than 0.2) exceeds a set threshold (e.g., 10 dimensions), then the component is considered to have a significant deviation from its historical state and is at risk of failure evolution.

[0064] This tensor is not only used for fault evolution diagram construction, but can also be used as input for downstream operation and maintenance instruction generation modules to support interpretive operation and maintenance judgments.

[0065] Evolutionary modeling module: Based on the multi-layer nested comparison of the semantic feature residual tensor with the evolution path of historical anomalous events, calculate the causal similarity index curve of potential failure events, and construct the failure evolution map of the photovoltaic array based on its trend change.

[0066] In this invention, to identify and model the potential fault evolution trend of photovoltaic modules, a similarity comparison method based on semantic feature residual tensors and historical anomalous evolution paths is proposed to construct a causal similarity index curve with a temporal causal structure. This curve measures the causal proximity between the current module state change and historical fault paths, providing a temporal input basis for subsequently constructing a fault evolution graph. Specifically, the following steps are included: An evolution path sample library containing residual tensor sequences with semantic features of various types of abnormal events is constructed based on historical operation and maintenance data, and a set of staged evolution trajectories is formed in chronological order.

[0067] First, using the labeled historical photovoltaic operation and maintenance data, we extract the component operation records of those that have experienced abnormal events and construct a semantic evolution path sample library.

[0068] The specific steps are as follows: For each known abnormal event, backtrack its operational status records for the 7 days prior to the event; The multimodal semantic embedding vector sequence within this time period is processed, and the semantic feature residual tensor is calculated on a daily basis, that is, the semantic offset between adjacent time points, to form a "residual evolution sequence". The sequence is divided into three stages according to time sequence: early minor change period, middle disturbance period, and late significant abnormality period; For each type of abnormal event (such as temperature rise, voltage drop, etc.), a corresponding set of staged evolution trajectories is established to form a structured sample library for subsequent matching and comparison.

[0069] The structure of each trajectory sample includes: anomaly type label, stage number, time index, and residual tensor representation for each stage.

[0070] The semantic feature residual tensor sequence of the current component node is matched and calculated stage by stage with the residual features of each stage in the evolution path sample library to obtain the feature offset similarity value of each time stage.

[0071] For the component node to be analyzed, extract its semantic feature residual tensor sequence from the past 7 days. The structure is consistent with the sample database and includes a 128-dimensional residual vector generated daily. The matching calculation steps are as follows: Align the current residual sequence with each outlier trajectory in the sample library by performing stage division; Within each stage, the cosine similarity function is used to calculate the similarity between the current residual vector and the historical trajectory residuals. The cosine similarity is defined as the dot product of two vectors divided by the product of their respective norms, and the result ranges from -1 to 1. The average of all similarity scores within a stage is taken to obtain the feature shift similarity value for that stage. The above calculations were performed for multiple anomaly types, and the phased similarity index corresponding to each anomaly type was recorded.

[0072] The output of this step is a structured table containing a set of similarity scores between the current component and different failure types at each time stage.

[0073] Construct time series similarity curves based on the similarity change trends over continuous time periods.

[0074] To capture the temporal evolution characteristics of abnormal states, the similarity values ​​of each stage are organized along the time dimension to form a time series similarity curve. The specific construction steps are as follows: Arrange the similarity values ​​of each stage in chronological order to form a discrete curve with the horizontal axis representing the time index and the vertical axis representing the similarity value. If a sliding window method is used for comparison, the window length can be set to 3 days and the step size to 1 day. The residual sequence can be segmented by sliding to obtain a more fine-grained similarity change curve. To reduce the interference of occasional anomalies or data noise on the curve shape, a weighted moving average smoothing method is introduced. The similarity values ​​of three adjacent time points are summed in a weighted manner, and the weights can be set to [0.2, 0.5, 0.3] to preserve the current trend of change while suppressing the interference of sudden changes.

[0075] The resulting curve is a dynamic trend graph representing the degree of proximity between the component and the historical failure evolution path over time.

[0076] The smoothed similarity curve is used as the causal similarity index curve of potential failure events.

[0077] Finally, the smoothed similarity time series is used as a causal similarity index curve to quantify the strength of the causal association between the current component and one or more types of historical anomalous events. The key properties of this curve are as follows: The peak position indicates the point in time when a potential fault is closest to a historical anomaly; The slope of the curve reflects the speed of fault evolution; A multi-peak structure may indicate a complex failure trend or intermittent fluctuation behavior.

[0078] When the overall trend of the curve continues to rise and the peak value exceeds the similarity threshold (such as 0.85), a potential fault evolution risk warning can be triggered, and the curve can be used as input for fault evolution diagram construction and operation and maintenance instruction generation.

[0079] To accurately capture the potential propagation relationships of faults among photovoltaic modules, this invention proposes a method for constructing a "fault evolution graph" with spatial coherence and evolutionary dependence based on the dynamic trend of causal similarity index curves. This method not only reveals the temporal characteristics of fault occurrence but also identifies semantic propagation paths and structural coupling pathways between modules, thereby achieving visualized modeling of local and global fault trends. Specifically, it includes the following steps: Extract the slope changes of the continuously rising intervals in the causal similarity index curve, identify time segments in the early, middle or accelerated phase of fault evolution, and mark their positions on the time axis.

[0080] For each component node that has generated a causal similarity index curve, perform the following operations: The sliding window method was used to analyze the continuous rising intervals in the curve, with the window size set to 3 days and the step size to 1 day. In each window, the linear slope of the similarity value is calculated. If the slope is between 0.01 and 0.05, it is considered an early, gradual growth stage; if the slope is between 0.05 and 0.15, it is considered a mid-term, significant evolution stage; if the slope exceeds 0.15, it is considered an accelerated evolution stage. The time segments of the different stages mentioned above are labeled on the timeline and recorded as a quadruple of (component number, stage type, start time, end time). This step is used to define the temporal boundary of the development of fault risk and to lay the foundation for a time coordinate reference for spatial propagation modeling.

[0081] Cluster analysis is performed on multiple component nodes with similar time trends to construct a semantic propagation relationship graph between node subsets and establish an evolutionary coupling path across nodes.

[0082] After obtaining the evolution stage labels for all components, perform the following operations: For all component nodes within the same time period, extract features such as the slope of the rising segment of their causal similarity curve, the mean residual amplitude, and the spatial location number to form a node feature vector; Cluster the nodes using a density clustering algorithm (such as DBSCAN), with the following parameters: minimum number of samples of 3 and similarity threshold of 0.8. Each cluster represents a subset of components with similar semantic evolution trends; Within each cluster, based on the order of node time tags and residual similarity, directed edges are constructed to represent possible paths for fault semantics to propagate from node A to node B, forming a semantic propagation relationship graph.

[0083] In the propagation graph, the edges represent the presumed causal evolution path, and the nodes represent the affected component individuals. This graph structure will serve as the initial skeleton of the evolution graph.

[0084] By introducing component spatial location and electrical connection topology information, the causal propagation path between nodes is mapped to the actual array structure, thus constructing a fault propagation chain with spatial continuity.

[0085] To ensure that the semantic propagation path remains consistent with the actual physical structure, the following operations are performed: Obtain the physical two-dimensional coordinates (row number, column number) of each component and the connection path number of its electrical circuit; Traverse each pair of directed connections in the semantic propagation graph and verify their spatial distance and electrical connectivity: If the spatial Euclidean distance is less than the spacing between 3 components and the electrical connections are in series or parallel, it is determined to be a valid propagation edge; Otherwise, remove the propagation edge to avoid introducing cross-regional connections without a physical basis; All valid propagation paths are sorted in chronological order to construct a spatially continuous and topologically reasonable fault propagation chain. Each chain is recorded as: starting node → intermediate node sequence → ending node.

[0086] This chain represents a dynamic path in which a fault originates in a component and gradually spreads to neighboring areas.

[0087] The graph structure edge weights are dynamically updated based on the directionality of the fault propagation chain and the activity of the nodes, ultimately forming a fault evolution graph that characterizes the local or global fault propagation trend of the photovoltaic array.

[0088] Finally, all the completed propagation chains are merged into a unified graph structure, and the edge weights are dynamically adjusted to enhance the expressiveness of the propagation trend: The initial weight of each propagation edge is defined as the average similarity value between node pairs; A node activity factor is further introduced, which is defined as the mean of the absolute values ​​of the changes in the similarity slope of nodes in the last 3 days, and is used to adjust the edge weights. If the activity level of upstream nodes on an edge is significantly higher than that of downstream nodes, the weight of that propagation edge is increased to reflect causal driving force. Repeat the above steps to update all edge weights, eventually forming a fault evolution graph. In the graph, the path represents the propagation direction, the edge weight represents the propagation strength, and the node attributes mark the state level.

[0089] This fault evolution diagram serves as a key input for subsequent operation and maintenance instruction generation and resource scheduling, possessing clear temporal evolution logic, spatial structure mapping, and multi-node coupling characteristics.

[0090] To verify the effectiveness of the evolution modeling module in this invention, the operating data of an actual photovoltaic power station was used for verification. The power station is located in a plateau region, has a capacity of 50 megawatts, contains 22,000 modules, and is equipped with a module-level infrared imaging system, electrical parameter acquisition device, and environmental meteorological monitoring equipment.

[0091] The experimental data sources include: Data collection period: six months; Daily sampling frequency: once per hour; Anomaly samples: A total of 325 historical component failure events were manually marked, covering types such as obstruction, loose wiring, hot spots, and output drift. Distribution of anomalous events: Approximately 52% are located at the array edge, and 48% are located in the array center block. The validation methods for the evolutionary modeling module include: Residual tensor generation and verification: The difference between historical running data and semantic completion results is calculated to construct a component-level residual tensor to characterize the failure trend; Similarity curve calculation: The residual tensor is compared with the historical evolution path sample library to generate a causal similarity index curve; Stage classification accuracy: The failure evolution stages are classified into early, middle and acceleration stages based on the changes in the slope of the curve, and compared with the actual occurrence time of the event; Propagation path identification rate: Based on the propagation chain structure, the complete path is identified in multi-node propagation events, and its consistency with the posterior fault area is evaluated.

[0092] The key indicators and results are as follows:

[0093] The experimental results above show that: The causal similarity index curve construction method proposed in this invention significantly improves the accuracy of identifying photovoltaic module failure trends, especially showing good predictive ability in the early stage of failure. The fault evolution graph construction method can effectively reconstruct the fault propagation path between components and has both spatial and temporal continuity. Compared with traditional diagnostic methods based on static thresholds, this invention shows significant advantages in multiple evaluation metrics and has stronger adaptability, robustness and scalability.

[0094] This module provides accurate timing and structure inputs for the subsequent intelligent operation and maintenance instruction generation module driven by the large language model, significantly improving the predictability, optimization and execution efficiency of the overall system.

[0095] Operation and maintenance instruction generation module: Input the fault evolution diagram into the operation and maintenance instruction generation engine driven by the large language model, and output natural language structured operation and maintenance suggestions containing multi-objective optimization strategies.

[0096] In this invention, the operation and maintenance instruction generation module is located at the end of the entire operation and maintenance reasoning process. It is responsible for converting the fault evolution diagram obtained based on historical data, residual analysis, and fault propagation modeling into structured natural language operation and maintenance suggestions with multi-objective optimization characteristics. This module is implemented by an instruction generation engine driven by a large language model and has core functions such as task sequencing, path optimization, resource allocation suggestions, and interpretable operation output. The specific implementation steps of this module are as follows: First, the fault evolution diagram constructed in the previous stage is structured and encoded, including the following key information: Component node attributes include component number, physical location coordinates, current status level, and similarity trend slope; Propagation path information: including the direction of the propagation chain, the start and end node numbers, the propagation delay, and the spatial distance; Edge weight: Represents the intensity of fault propagation and is used to measure priority areas for handling. Node activity sequence: used to assess potential diffusion capacity and urgency.

[0097] The above structural information is embedded into a graph structure sequence or a tabular natural language description, such as: "Component B105 (position 3,9) is in the middle of the fault propagation chain, the propagation path is A104→B105→B106, the edge weight is 0.85, and the current state is 'moderate anomaly'." This description serves as the input context to guide the large language model in completing semantic understanding.

[0098] In the process of generating operation and maintenance instructions, this invention introduces the following multi-objective decision-making elements: Prioritization objective: Construct a comprehensive priority score based on fault level, propagation intensity, and node activity score, and prioritize high-risk nodes; Path planning objective: Based on the spatial coordinates of components and the distance to the maintenance operation path, generate the minimum operation travel path; Resource scheduling objective: Incorporate boundary conditions such as the current number of maintenance personnel, available time windows, and spare parts inventory to dynamically adapt job allocation; Task combination objective: Automatically identify task nodes that can be executed in parallel, reduce redundant scheduling, and improve response efficiency.

[0099] The aforementioned objective function is embedded in the prompts using logical templates or mathematical constraints, guiding the model to comprehensively consider maintainability, executability, and resource efficiency during the generation phase.

[0100] Combine the above graph structure context with the operation and maintenance target prompts to form a complete instruction generation input statement, such as: "Based on the following component propagation paths and fault level assessment results, please generate high-priority operation and maintenance task instructions, requiring the shortest total task time and optimal on-site scheduling." The input is fed into a large language model fine-tuned for the photovoltaic scenario (e.g., based on T5 or GPT architecture), and the model output includes: The instruction number and task description (e.g., "Task 1: Check if the B105 wiring is loose, high priority"); recommended personnel or team configuration; suggested maintenance time and replacement parts; maintenance route order and parallel task marking; confirmation content to be provided after the operation is completed. The output format is structured natural language, which is convenient for direct import into the digital operation and maintenance platform or for quick manual interpretation.

[0101] The generated natural language operation and maintenance suggestions are formatted, validated, and distributed by the post-processing module, and can be output in the following form: Text report (with component location map, priority labels, and suggested actions); Interactive task cards (integrated into the operation and maintenance workflow system); A visual dashboard based on risk levels.

[0102] If the output of the large language model contains ambiguous instructions or conflicting tasks, the system will automatically highlight them and prompt for manual confirmation to ensure that the generated content is interpretable and actually executable.

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A photovoltaic array intelligent operation and maintenance decision-making system based on a large language model, characterized in that: include: Data acquisition and fusion module: acquires the original heterogeneous operation dataset of the target photovoltaic power station, including component-level thermal imaging sequence, historical change data of AC and DC side electrical parameters, regional meteorological time series data, and historical maintenance records text of operation and maintenance engineers; Feature alignment module: The original heterogeneous operating dataset is decoupled and normalized in terms of mode, and a time-series vector set, a heat map set and a text semantic stream are constructed respectively, and a group of photovoltaic operating status semantic embedding vectors with semantic alignment capability is generated. Operation and maintenance semantic graph construction module: Based on the predefined component space topology and electrical path dependency graph, construct the operation and maintenance semantic embedding graph of the photovoltaic array, and mine the potential semantic coupling relationship between weakly connected nodes through the graph attention mechanism to generate an initial set of hidden fault candidates; Semantic reasoning and completion module: For the initial set of hidden fault candidates, combined with the fault variation factor reverse reasoning mechanism introduced in the large language model, high-confidence semantic completion is performed on the missing or unobservable data regions to generate component-level semantic feature residual tensors. Evolutionary modeling module: Based on the multi-layer nested comparison between the semantic feature residual tensor and the evolution path of historical anomalous events, calculate the causal similarity index curve of potential failure events, and construct the failure evolution map of the photovoltaic array based on its trend change. Operation and maintenance instruction generation module: Input the fault evolution diagram into the operation and maintenance instruction generation engine driven by the large language model, and output natural language structured operation and maintenance suggestions containing multi-objective optimization strategies.

2. The photovoltaic array intelligent operation and maintenance decision-making system based on a large language model according to claim 1, characterized in that: The steps for constructing the operation and maintenance semantic embedding graph of the photovoltaic array based on the predefined component spatial topology and electrical path dependency graph include: The semantic embedding vector of each photovoltaic module is bound to its physical spatial coordinates in the array to generate a set of module-level nodes; Based on the electrical connection relationship of components and the actual wiring structure, a set of electrical path dependency edges between nodes is established, and edge weights are assigned to represent connection strength or fault propagation probability. The graph attention mechanism in graph neural networks is used to perform feature aggregation on the above set of nodes and edge sets; The aggregated graph structure is optimized by structured embedding to generate an operation and maintenance semantic embedding graph of the photovoltaic array.

3. The photovoltaic array intelligent operation and maintenance decision-making system based on a large language model according to claim 2, characterized in that: The step of generating the initial set of latent fault candidates includes: The graph attention mechanism is used to perform multi-level adjacency node feature aggregation on each component node in the operation and maintenance semantic embedding graph; The semantic similarity index and the rate of change of connection weights between node pairs are calculated to identify weak connection node pairs with low structural strength but prominent semantics. Based on weakly connected nodes, potential fault impact areas are constructed, and local clustering analysis is performed by integrating historical operational status change trends within the areas. The subset of nodes with anomalous semantic distribution and deviation from historical evolutionary trajectories in the clustering results are identified as the initial latent fault candidate set, and their positions and temporal labels in the graph are associated with them.

4. The photovoltaic array intelligent operation and maintenance decision-making system based on a large language model according to claim 1, characterized in that: The steps for performing high-confidence semantic completion of missing or unobservable data regions by combining the fault variation factor back reasoning mechanism introduced in the large language model include: Anomaly scores are calculated for the multimodal embedding vectors of each component node in the initial set of latent fault candidates, and nodes with scores higher than a set threshold are selected as inference targets. Based on the structured knowledge corpus and historical operation and maintenance semantic embedding samples built into the large language model, a causal relationship prompt template containing fault category, evolution path and mutation mode is constructed. The semantic context information of the reasoning target node and the prompt template are jointly input into the large language model, and the semantic feature vector of the running state of the missing or unobservable data region is output. The semantic feature vector of the running state is compared with the original embedding vector by dimension-wise difference calculation to obtain the component-level semantic feature residual tensor that reflects the degree of fault change.

5. The photovoltaic array intelligent operation and maintenance decision-making system based on a large language model according to claim 4, characterized in that: The step of jointly inputting the semantic context information of the reasoning target node and the prompt template into the large language model includes: The semantic embedding vector of the reasoning target node is expanded by time window, its state change features in continuous running cycle are extracted, and a context description sequence containing component number, spatial location and historical running trend is generated. The context description sequence is embedded into a pre-built fault causal relationship prompt template to form a joint semantic input sequence containing the operating background, potential anomaly types, and inference target constraints; The joint semantic input sequence is fed into a large language model, and the runtime semantic feature vector corresponding to the missing or unobservable data region is predicted through conditional generation.

6. The photovoltaic array intelligent operation and maintenance decision-making system based on a large language model according to claim 1, characterized in that: The step of calculating the causal similarity index curve of potential failure events includes: An evolution path sample library containing residual tensor sequences with semantic features of various types of abnormal events is constructed based on historical operation and maintenance data, and a set of staged evolution trajectories is formed in chronological order. The semantic feature residual tensor sequence of the current component node is matched and calculated stage by stage with the residual features of each stage in the evolution path sample library to obtain the feature offset similarity value of each time stage. Construct time series similarity curves based on the similarity change trends over continuous time periods; The smoothed similarity curve is used as the causal similarity index curve of potential failure events.

7. The photovoltaic array intelligent operation and maintenance decision-making system based on a large language model according to claim 6, characterized in that: The steps for constructing a fault evolution map of a photovoltaic array based on the trend changes of the causal similarity index curve include: The slope changes of continuously rising intervals in the causal similarity index curve are extracted to identify time segments in the early, middle, or accelerated stages of fault evolution, and their positions on the time axis are marked. Multiple component nodes with similar time trends are clustered to construct a semantic propagation relationship graph between node subsets and establish cross-node evolutionary coupling paths. Component spatial location and electrical connection topology information are introduced to map the causal propagation paths between nodes to the actual array structure, constructing a fault propagation chain with spatial continuity. The graph structure edge weights are dynamically updated according to the directionality of the fault propagation chain and the node activity, ultimately forming a fault evolution graph that characterizes the local or global fault propagation trend of the photovoltaic array.

Images