A photovoltaic power station fault identification method and system based on hybrid reasoning

Abstract

The application discloses a photovoltaic power station fault identification method and system based on mixed reasoning, belongs to the technical field of photovoltaic power station fault identification, and through a double-channel cooperative screening mechanism of atlas reasoning screening and model reasoning screening, the interpretability of fault identification is ensured, the diagnosis precision is improved through data law mining, the problem that precision and interpretability cannot be compatible in a single technical route is completely solved, global features are added and verification is performed through a comprehensive evaluation mechanism, isolated faults and regional faults can be accurately distinguished in addition to eliminating instantaneous false alarms, the problem that traditional multi-focusing is focused on single equipment or single fault point identification, lacks a system global perspective, and cannot identify common cause faults, chain faults and fault propagation paths is solved, and under complex working conditions, the accuracy of fault identification can be ensured, and the identification reliability is relatively high.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant fault identification technology, specifically to a photovoltaic power plant fault identification method and system based on hybrid reasoning. Background Technology

[0002] Fault identification in photovoltaic power plants is a core technological link in ensuring the safe and stable operation of photovoltaic power generation systems and improving power generation efficiency. Its development has mainly progressed from manual periodic inspections to data-driven diagnosis, and now to the application of intelligent algorithms. The manual-dependent stage relied primarily on maintenance personnel conducting regular on-site inspections or issuing alarms based on simple monitoring data thresholds. This approach was slow to respond, insensitive to latent and early-stage faults, and highly dependent on personnel experience, making it difficult to scale up. The automated diagnosis stage saw the emergence of fault diagnosis methods based on electrical characteristic analysis (such as IV curves, current and voltage signal analysis) and mathematical models (such as equivalent circuit models) with the popularization of sensor technology and SCADA systems. These methods achieved automation to a certain extent, but had poor adaptability to complex operating conditions and limited diagnostic accuracy. The intelligent algorithm exploration stage has seen the introduction of machine learning and deep learning algorithms in recent years, aiming to quickly locate fault information by training machine learning models based on historical fault data or training knowledge-driven models using knowledge graphs, thereby achieving automated fault diagnosis. However, existing intelligent solutions still have significant shortcomings in terms of the depth of technological integration and adaptability to engineering practices.

[0003] In summary, traditional methods and early automation solutions have limited diagnostic accuracy and are susceptible to interference from environmental noise and equipment aging, leading to untimely fault handling and significant power generation losses. Secondly, both comparative methods based on historical data models and reasoning methods based on knowledge graphs suffer from a lack of diversity in their technical approaches: data-driven models lack domain knowledge guidance, have poor interpretability, and function like a "black box"; knowledge-driven models are rigid, lagging, and have weak adaptive capabilities. Furthermore, existing technologies generally lack effective fusion and collaborative analysis of multimodal data, resulting in incomplete fault feature mining. Simultaneously, current technologies primarily focus on identifying individual devices or single fault points, lacking cross-device correlation analysis and fault propagation path early warning capabilities from a system-wide perspective.

[0004] Chinese patent, publication number CN119382616A, discloses a method for diagnosing distributed photovoltaic power station equipment based on big data analysis. This method collects data transmitted from sensors on various devices within the distributed photovoltaic power station, and then extracts characteristic parameters reflecting the equipment's operating status from the preprocessed data through data analysis and data mining. Finally, based on historical fault data and corresponding characteristic parameters, a fault diagnosis model is constructed for fault diagnosis. However, the model construction relies on learning and training from massive amounts of fault data samples. In actual engineering, many fault data samples are scarce and difficult to constitute qualified training data samples, leading to an imbalance in fault scenarios during model training and resulting in low detection accuracy of the model when facing these faults. Summary of the Invention

[0005] This invention addresses the problem that most existing photovoltaic power plant fault identification methods employ a single technical approach, which struggles to guarantee accuracy and reliability in complex operating conditions. It provides a photovoltaic power plant fault identification method and system based on hybrid reasoning. By employing a dual-channel collaborative screening mechanism combining graph-based reasoning and model-based reasoning, it ensures interpretability of fault identification while improving diagnostic accuracy through data pattern mining. This completely solves the dilemma of the trade-off between accuracy and interpretability in a single technical approach. Furthermore, it adds global features and verifies the results through a comprehensive evaluation mechanism, eliminating instantaneous false alarms and accurately distinguishing between isolated and regional faults. This solves the problem that traditional methods, which focus on individual equipment or single fault points, lack a system-wide perspective, resulting in the inability to identify common-cause faults, cascading faults, and fault propagation paths.

[0006] In a first aspect, one technical solution provided in this embodiment of the invention is: a photovoltaic power plant fault identification method based on hybrid reasoning, comprising the following steps: S1. Collect multi-dimensional operational data of photovoltaic power stations and preprocess the data to obtain raw multi-source data; S2. Construct a fault indication map and perform map inference and filtering on the original multi-source data to obtain the first candidate fault set; construct a fault screening model and perform model inference and filtering on the original multi-source data to obtain the second candidate fault set; S3. Based on the dual-channel collaborative mechanism, the first candidate fault set and the second candidate fault set are collaboratively analyzed to obtain the fused candidate fault set. Global features are added to the fused candidate fault set to obtain the fused feature vector. S4. Based on the comprehensive evaluation mechanism, the fusion feature vectors are comprehensively screened to obtain the final fault list of the photovoltaic power station.

[0007] This solution addresses the shortcomings of traditional single-data feature mining by analyzing multi-dimensional operational data, thus achieving comprehensive coverage of both electrical and non-electrical faults. Through a dual-channel design of fault indication maps and fault screening models, it ensures interpretability of reasoning through knowledge-driven approaches and uncovers hidden faults through data-driven approaches, overcoming the pain points of "black box" decision-making and rigid, lagging systems, thereby improving identification accuracy. By employing a dual-channel collaborative mechanism and global feature enhancement, it integrates dual candidate sets and supplements system-level information such as regional fault density and topology correlation, thereby upgrading the assessment from isolated faults to common-cause and cascading faults. Finally, a comprehensive evaluation mechanism outputs an accurate fault list, automating the entire process to adapt to equipment aging and topology change scenarios, reducing reliance on maintenance experience, improving response efficiency, and minimizing ineffective maintenance costs and power generation losses, demonstrating both engineering practicality and technological scalability.

[0008] Optionally, in S1, the multi-dimensional operational data includes power plant operational data, equipment static data, environmental data, and topology data; The power plant operation data includes the current, voltage, output power, and equipment temperature of each device in the photovoltaic power plant; The static data of the equipment includes equipment ID, type, historical fault records, and equipment images; The environmental data includes ambient temperature, solar irradiance, and wind speed; The topology data includes the electrical connections, geographical locations, and regions of various devices within the photovoltaic power plant.

[0009] This solution comprehensively captures full-scene features such as current, images, and connectivity by collecting multi-dimensional data covering power plant operation, equipment static data, environmental data, and topology data, thus overcoming the shortcomings of traditional single-data mining. It supports full-coverage identification of electrical and non-electrical faults and provides a high-quality data foundation for subsequent dual-channel inference and global feature analysis. Through the integrity and diversity of the data, the comprehensiveness and accuracy of fault identification are improved, false alarms and missed alarms are reduced, and reliable data support is provided for operation and maintenance decisions.

[0010] Optionally, in S2, the process of constructing the fault indication map includes: Device nodes are constructed based on the geographical location and region of the devices in the topology data, with fault type as fault type node and fault manifestation as symptom node; The original atlas is obtained by connecting various device nodes based on the connection relationship between devices, connecting device nodes with fault type nodes based on the causal relationship between fault type and faulty device, and connecting symptom nodes with fault type nodes based on the correlation between fault type and fault manifestation. The original map is bound to preset evaluation rules and updated in real time based on the actual operating status of the photovoltaic power station to obtain the fault indication map.

[0011] In this solution, by constructing device nodes according to geographical location and region, and associating fault type and symptom nodes, the semantic and topological relationships between devices, faults, and symptoms are clarified, making the reasoning logic traceable and thus solving the data-driven "black box" problem. By binding preset evaluation rules and updating them in real time, it can adapt to changes in equipment and operating status, thereby avoiding rigid and lagging graphs. At the same time, it provides structured knowledge support for subsequent accurate location of potential faults and discovery of fault associations, improving the efficiency and reliability of fault screening and laying a solid knowledge foundation for dual-channel collaborative reasoning.

[0012] Optionally, in S2, the first candidate fault set is obtained by performing graph inference filtering on the original multi-source data, including the following steps: The original multi-source data is mapped to the symptom nodes in the fault indication map as mapping nodes. Starting from the mapping nodes, the possible fault type nodes and equipment nodes are traced based on causal and correlation relationships to obtain the corresponding tracing chain. Each traceability chain is evaluated based on preset evaluation rules to obtain inference confidence. Traceability chains with inference confidence exceeding the confidence threshold are selected and arranged in descending order of confidence to generate a structured first candidate fault set. The first candidate fault set consists of entries composed of candidate fault device ID, fault type, associated symptoms and inference confidence.

[0013] In this solution, the original multi-source data is mapped to symptom nodes of the fault indication map. Based on causal and correlation relationships, the fault type and equipment node are traced, making the reasoning logic traceable and solving the data-driven "black box" problem. By calculating the reasoning confidence through preset evaluation rules, a structured candidate set is generated after threshold screening and sorting. Invalid clues are accurately filtered out, improving the efficiency of fault location. At the same time, it provides high-quality, deteriorating, and interpretable fault clues for dual-channel collaborative analysis, thereby reducing the redundancy of subsequent screening and laying a solid foundation for the final accurate diagnosis, taking into account both the reliability and efficiency of fault identification.

[0014] Optionally, in S2, the specific process of constructing the fault screening model includes: The fault screening model is trained by using historical fault records of photovoltaic power plants as positive samples, normal operation data of photovoltaic power plants as negative samples, and the probability of identifying various fault types as the training objective.

[0015] In this solution, a fault screening model is trained using historical fault records as positive samples and normal operation data as negative samples, with fault type probability as the training objective. This approach can fully uncover hidden fault patterns in the data, making up for the shortcomings of knowledge graphs in covering complex and novel faults. At the same time, it adapts to the characteristics and patterns of multimodal data, which can improve the accuracy of fault identification and provide high-quality candidate clues driven by data for dual-channel collaborative screening, thus facilitating efficient and accurate diagnosis in the future.

[0016] Optionally, in S2, the second candidate fault set is obtained by model inference filtering of the original multi-source data, including the following steps: Based on convolutional neural networks, feature extraction is performed on the static data of each device to obtain static features of the device; time series analysis is performed on the power plant operation data to obtain power plant operation features; and normalization processing is performed on the environmental data to obtain environmental features. The static characteristics of the equipment, the operating characteristics of the power plant, and the environmental characteristics are spliced ​​together and weighted to obtain the multi-source fusion characteristics of each equipment. The multi-source fusion features are used as input to the fault screening model to obtain the fault probability of each fault type. Fault types with fault probabilities higher than a set threshold are selected and sorted in descending order based on the fault probability to obtain a second candidate fault set. The second candidate fault set consists of entries composed of candidate fault device ID, fault type and corresponding fault probability.

[0017] In this solution, convolutional neural networks are used to extract static features of equipment, time-series analysis is used to mine power plant operation features, and environmental features are normalized to accurately capture core information of various types of data. Multi-source fusion is achieved through feature splicing and weighting to integrate comprehensive fault clues. By outputting fault probabilities through the model, a structured second candidate fault set is generated through threshold filtering and descending sorting. This can efficiently uncover hidden and complex faults, thereby making up for the limitations of knowledge graphs and providing high-quality data-driven clues for subsequent dual-channel collaboration, thus improving the accuracy and efficiency of fault identification and screening.

[0018] Optionally, in S3, a fused candidate fault set is obtained by performing collaborative analysis on the first candidate fault set and the second candidate fault set based on a dual-channel collaborative mechanism, including the following steps: Using the device ID as a unique identifier, the fault information of the same device in the first candidate fault set and the second candidate fault set are associated and the union is taken; If the same device appears in both the first candidate fault set and the second candidate fault set, the fault type in the first candidate fault set and the fault probability in the second candidate fault set are retained to obtain the initial candidate set. The fusion candidate fault set is obtained by removing duplicate device IDs and corresponding fault types from the initial candidate set.

[0019] In this solution, by using the device ID as a unique identifier and associating the first and second candidate fault sets with a union, comprehensive coverage of fault clues is ensured without omission. By prioritizing the retention of knowledge-driven fault types in the first candidate set to guarantee the interpretability of reasoning, and retaining data-driven fault probabilities in the second candidate set to strengthen quantitative support, the advantages of the two channels are complemented. By eliminating duplicate entries, redundant information can be simplified and subsequent processing efficiency can be improved, ultimately forming a high-quality fused candidate fault set, laying a solid foundation for subsequent global feature enhancement and comprehensive evaluation, and taking into account the comprehensiveness, reliability, and efficiency of fault identification.

[0020] Optionally, in S3, global features are added to the fused candidate fault set to obtain a fused feature vector, including the following steps: The regional fault density is obtained by calculating the fault density of the area where the faulty equipment is located. The distribution pattern of fault types in the time dimension is statistically analyzed to obtain the fault time clustering degree, which includes time synchronization degree, duration and recurrence frequency. The topological correlation degree is obtained by analyzing the electrical connection relationships between each candidate faulty device; Environmental consistency is obtained by analyzing the environmental data of candidate faulty devices within the same area; The regional fault density, fault time clustering, topological correlation, and environmental consistency are used as global features and associated with the fusion candidate fault set to obtain a fusion feature vector.

[0021] In this solution, regional common-cause faults can be accurately identified by calculating regional fault density; instantaneous false alarms can be effectively distinguished from real faults by statistically analyzing fault time clustering; cascading fault relationships between devices can be uncovered by analyzing topological correlation; and false faults caused by environmental factors can be eliminated by assessing environmental consistency. By associating the above four types of global features with the fused candidate fault set, fault analysis is upgraded from an isolated device perspective to a system-wide perspective, supplementing the global constraint information missing from single fault clues, providing comprehensive feature support for subsequent comprehensive evaluation, significantly improving the accuracy of fault identification, reducing false alarms and missed alarms, and helping maintenance personnel trace fault propagation paths and accurately locate root causes.

[0022] Optionally, in S4, the comprehensive evaluation mechanism is as follows: Retain candidate fault device IDs whose regional fault density and time synchronization exceed the density threshold and synchronization threshold respectively, and whose environmental consistency is less than the consistency threshold. Candidate faulty device IDs whose regional fault density and topological correlation are less than the density threshold and correlation threshold, respectively, and whose duration is less than the duration threshold are removed. The neighboring device IDs of candidate faulty devices that have a topological correlation degree exceeding the correlation degree threshold and a recurrence frequency exceeding the recurrence frequency threshold are included in the candidate faulty devices.

[0023] This solution accurately identifies regional common-cause faults by retaining devices that meet the standards for regional fault density, time synchronization, and environmental consistency. It effectively filters out transient false alarms by removing devices with low fault density, low topological correlation, and short duration. Furthermore, it captures cascading fault propagation paths by including adjacent devices of devices with high topological correlation and high recurrence frequency. This mechanism achieves refined screening based on global features, avoiding fault omissions and suppressing invalid clues. It upgrades fault identification from single-device judgment to system-level analysis, significantly improving diagnostic accuracy. This helps maintenance personnel quickly trace root causes and coordinate fault handling, reducing maintenance costs and power generation losses.

[0024] Secondly, one technical solution provided in this embodiment of the invention is: a photovoltaic power station fault identification system based on hybrid reasoning, including a data acquisition module, a dual-channel reasoning module, a global collaboration module, and a comprehensive evaluation module; The data acquisition module collects multi-dimensional operational data of the photovoltaic power station and preprocesses it to obtain raw multi-source data. The dual-channel inference module is equipped with a fault indication map and a fault screening model. Based on the fault indication map, it performs map inference screening on the original multi-source data to obtain a first candidate fault set; based on the fault screening model, it performs model inference screening on the original multi-source data to obtain a second candidate fault set. The global collaboration module performs collaborative analysis on the first candidate fault set and the second candidate fault set based on the dual-channel collaboration mechanism to obtain a fused candidate fault set, and adds global features to the fused candidate fault set to obtain a fused feature vector. The comprehensive evaluation module performs a comprehensive evaluation of the fusion feature vector based on the comprehensive evaluation mechanism, and obtains the final fault list of the photovoltaic power station based on the evaluation results.

[0025] In this solution, a corresponding system is built to integrate the photovoltaic power station fault identification method, thereby realizing human-computer interaction and improving the user experience.

[0026] The beneficial effects of this invention are as follows: By adopting a dual-channel collaborative screening mechanism of graph reasoning screening and model reasoning screening, this invention not only ensures the interpretability of fault identification, but also improves the diagnostic accuracy through data pattern mining. It completely solves the problem that accuracy and interpretability cannot be achieved simultaneously by a single technical approach. At the same time, it adds global features and verifies them through a comprehensive evaluation mechanism, which not only eliminates instantaneous false alarms, but also accurately distinguishes between isolated faults and regional faults. This solves the problem that traditional methods focus on identifying individual devices or single fault points and lack a system-wide perspective, resulting in the inability to identify common-cause faults, cascading faults, and fault propagation paths.

[0027] The above description of the invention is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0028] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0029] Figure 1 This is a flowchart of a photovoltaic power plant fault identification method based on hybrid reasoning according to the present invention; Figure 2 This is a schematic diagram of a photovoltaic power plant fault identification system based on hybrid reasoning according to the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only one preferred embodiment of this invention and are only used to explain this invention. They do not limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0031] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures; the process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0032] Example 1: To address the problem that most existing photovoltaic power plant fault identification methods employ a single technical approach, resulting in low reliability due to difficulty in ensuring accuracy under complex operating conditions, this example provides a photovoltaic power plant fault identification method based on hybrid reasoning, such as... Figure 1 As shown, it includes the following steps: S1: Collect multi-dimensional operational data of photovoltaic power plants and preprocess them to obtain raw multi-source data.

[0033] In this embodiment, the multi-dimensional operational data includes power plant operational data, equipment static data, environmental data, and topology data; The power plant operation data includes the current, voltage, output power, and equipment temperature of each device in the photovoltaic power plant; The static data of the equipment includes equipment ID, type, historical fault records, and equipment images; The environmental data includes ambient temperature, solar irradiance, and wind speed; The topology data includes the electrical connections, geographical locations, and regions of various devices within the photovoltaic power plant.

[0034] This embodiment comprehensively captures full-scene features such as current, images, and connection relationships by collecting multi-dimensional data covering power plant operation data, equipment static data, environmental data, and topology data, thus overcoming the shortcomings of traditional single data mining. It not only supports the full-coverage identification of electrical and non-electrical faults, but also provides a high-quality data foundation for subsequent dual-channel inference and global feature analysis. Through the integrity and diversity of the data, the comprehensiveness and accuracy of fault identification are improved, false alarms and missed alarms are reduced, and reliable data support is provided for operation and maintenance decisions.

[0035] S2: Construct a fault indication map and perform map inference and screening on the original multi-source data to obtain the first candidate fault set; construct a fault screening model and perform model inference and screening on the original multi-source data to obtain the second candidate fault set.

[0036] In this embodiment, the process of constructing the fault indication map includes: Device nodes are constructed based on the geographical location and region of the devices in the topology data, with fault type as fault type node and fault manifestation as symptom node; The original atlas is obtained by connecting various device nodes based on the connection relationship between devices, connecting device nodes with fault type nodes based on the causal relationship between fault type and faulty device, and connecting symptom nodes with fault type nodes based on the correlation between fault type and fault manifestation. The original map is bound to preset evaluation rules and updated in real time based on the actual operating status of the photovoltaic power station to obtain the fault indication map.

[0037] Specifically, the connection relationship between devices is represented by, for example, PV-01-05 being connected to combiner box CB-02, representing the electrical connection between devices; the causal relationship indicates that it may happen, specifically represented by, for example, the inverter may trip, representing the correspondence between devices and potential faults; the correlation relationship indicates that, specifically represented by, the tripping fault manifests as zero power and frequent self-starting, representing the mapping between faults and symptoms.

[0038] The evaluation rules are specifically formulated by experts, transforming the experience of operations and maintenance experts into graph reasoning rules, for example: Rule 1: If a single component has an abnormal current while the upstream and downstream components are normal, then the combiner box is ruled out as the faulty component. Rule 2: If a component has been installed for ≥8 years, then aging-related faults are prioritized. Rule 3: If the power of multiple components in the same combiner box decreases, then the probability of combiner box fault is higher than that of single component fault.

[0039] The graph is updated in real time based on the actual operating status of the photovoltaic power station. Specifically, when the power station undergoes equipment replacement, new fault types are added, topology adjustments are made, or expert rules are iterated, the graph is updated. The update operations include node updates, relationship updates, and rule updates. Node updates include adding / deleting equipment nodes and supplementing new fault nodes. Relationship updates include adjusting equipment connection relationships, optimizing the causal relationship between faults and equipment, and the correlation between fault types and symptoms. Rule updates include iterating the rule base through expert review to avoid graph failure due to equipment aging or technology upgrades.

[0040] This embodiment constructs device nodes based on geographical location and region, and associates fault type and symptom nodes, clarifying the semantic and topological relationships between devices, faults, and symptoms, making the reasoning logic traceable, thereby solving the data-driven "black box" problem. By binding preset evaluation rules and updating them in real time, it can adapt to changes in equipment and operating status, thereby avoiding rigid and lagging graphs. At the same time, it provides structured knowledge support for subsequent accurate location of potential faults and discovery of fault associations, improving the efficiency and reliability of fault screening, and laying a solid knowledge foundation for dual-channel collaborative reasoning.

[0041] In this embodiment, the first candidate fault set is obtained by performing graph inference filtering on the original multi-source data, including the following steps: The original multi-source data is mapped to the symptom nodes in the fault indication map as mapping nodes. Starting from the mapping nodes, the possible fault type nodes and equipment nodes are traced based on causal and correlation relationships to obtain the corresponding tracing chain. Each traceability chain is evaluated based on preset evaluation rules to obtain inference confidence. Traceability chains with inference confidence exceeding the confidence threshold are selected and arranged in descending order of confidence to generate a structured first candidate fault set. The first candidate fault set consists of entries composed of candidate fault device ID, fault type, associated symptoms and inference confidence.

[0042] Specifically, the format of a certain entry in the first candidate fault set can be: {Faulty device ID: PV-01-05, Fault type: Trip fault, Associated symptoms: [Power is 0, Self-starts three times within 10 minutes], Inference confidence: 89 points}.

[0043] This embodiment maps raw multi-source data into symptom nodes of a fault indication map, and traces fault types and equipment nodes based on causal and correlational relationships, making the reasoning logic traceable and solving the data-driven "black box" problem. By calculating the reasoning confidence through preset evaluation rules, and generating a structured candidate set after threshold screening and sorting, invalid clues are accurately filtered out, improving fault location efficiency. At the same time, it provides high-quality, deteriorating, and interpretable fault clues for dual-channel collaborative analysis, thereby reducing subsequent screening redundancy and laying a solid foundation for the final accurate diagnosis, taking into account both the reliability and efficiency of fault identification.

[0044] In this embodiment, the specific process of constructing the fault screening model includes: The fault screening model is trained by using historical fault records of photovoltaic power plants as positive samples, normal operation data of photovoltaic power plants as negative samples, and the probability of identifying various fault types as the training objective.

[0045] Specifically, the fault screening model in this embodiment can use XGBOOST (gradient boosting tree) as the core model.

[0046] This embodiment trains a fault screening model using historical fault records as positive samples and normal operation data as negative samples, with fault type probability as the training objective. This fully uncovers hidden fault patterns in the data, making up for the shortcomings of knowledge graphs in covering complex and novel faults. At the same time, it adapts to the characteristics and patterns of multimodal data, which can improve the accuracy of fault identification and provide high-quality candidate clues driven by data for dual-channel collaborative screening, thus facilitating efficient and accurate diagnosis in the future.

[0047] In this embodiment, the second candidate fault set is obtained by model inference filtering of the original multi-source data, including the following steps: Based on convolutional neural networks, feature extraction is performed on the static data of each device to obtain static features of the device; time series analysis is performed on the power plant operation data to obtain power plant operation features; and normalization processing is performed on the environmental data to obtain environmental features. The static characteristics of the equipment, the operating characteristics of the power plant, and the environmental characteristics are spliced ​​together and weighted to obtain the multi-source fusion characteristics of each equipment. The multi-source fusion features are used as input to the fault screening model to obtain the fault probability of each fault type. Fault types with fault probabilities higher than a set threshold are selected and sorted in descending order based on the fault probability to obtain a second candidate fault set. The second candidate fault set consists of entries composed of candidate fault device ID, fault type and corresponding fault probability.

[0048] Specifically, the format of a certain entry in the second candidate fault set in this embodiment can be: {fault device ID: device identifier, candidate fault list: [(fault type 1, probability value 1), (fault type 2, probability value 2)],}, for example: {fault device ID: PV-01-06, candidate fault list: [(tripping fault, 0.92), (poor line contact, 0.65)]}.

[0049] This embodiment extracts static features of equipment through convolutional neural networks, mines power plant operation features through time-series analysis, and normalizes environmental features to accurately capture core information of various types of data. Multi-source fusion is achieved through feature splicing and weighting to integrate comprehensive fault clues. By outputting fault probabilities through the model, a structured second candidate fault set is generated through threshold filtering and descending sorting. This can efficiently uncover hidden and complex faults, thereby making up for the limitations of knowledge graphs and providing high-quality data-driven clues for subsequent dual-channel collaboration, improving the accuracy of fault identification and the efficiency of screening.

[0050] S3: Based on the dual-channel collaborative mechanism, the first candidate fault set and the second candidate fault set are collaboratively analyzed to obtain the fused candidate fault set. Global features are added to the fused candidate fault set to obtain the fused feature vector.

[0051] In this embodiment, a fused candidate fault set is obtained by performing collaborative analysis on the first candidate fault set and the second candidate fault set based on a dual-channel collaborative mechanism, including the following steps: Using the device ID as a unique identifier, the fault information of the same device in the first candidate fault set and the second candidate fault set are associated and the union is taken; If the same device appears in both the first candidate fault set and the second candidate fault set, the fault type in the first candidate fault set and the fault probability in the second candidate fault set are retained to obtain the initial candidate set. The fusion candidate fault set is obtained by removing duplicate device IDs and corresponding fault types from the initial candidate set.

[0052] This embodiment ensures comprehensive and complete coverage of fault clues by using the device ID as a unique identifier to associate and merge the first and second candidate fault sets. It prioritizes retaining the knowledge-driven fault types from the first candidate set to guarantee the interpretability of reasoning, while retaining the data-driven fault probabilities from the second candidate set to strengthen quantitative support, thus achieving complementary advantages from both channels. By eliminating duplicate entries, redundant information can be simplified and subsequent processing efficiency improved, ultimately forming a high-quality fused candidate fault set. This lays a solid foundation for subsequent global feature enhancement and comprehensive evaluation, balancing the comprehensiveness, reliability, and efficiency of fault identification.

[0053] In this embodiment, adding global features to the fused candidate fault set to obtain the fused feature vector includes the following steps: The regional fault density is obtained by calculating the fault density of the area where the faulty equipment is located. The distribution pattern of fault types in the time dimension is statistically analyzed to obtain the fault time clustering degree, which includes time synchronization degree, duration and recurrence frequency. The topological correlation degree is obtained by analyzing the electrical connection relationships between each candidate faulty device; Environmental consistency is obtained by analyzing the environmental data of candidate faulty devices within the same area; The regional fault density, fault time clustering, topological correlation, and environmental consistency are used as global features and associated with the fusion candidate fault set to obtain a fusion feature vector.

[0054] Specifically, in this embodiment, the regional fault density is the proportion of candidate faulty devices in the region where the faulty device is located within a preset time window. It is used to identify regional common-cause faults. For example, if the regional fault density is ≥10%, it may be caused by common factors such as regional power grid fluctuations or centralized blockages.

[0055] In the fault time clustering degree, the time synchronization degree can be the proportion of devices with a start time difference of ≤10 minutes for candidate faults in the same area, the duration is the continuous duration of the abnormal signal corresponding to the fault (such as power of 0), and the recurrence frequency is the number of times the same abnormal signal of the device is triggered in the past 24 hours. The judgment criteria set can be adjusted according to the actual situation.

[0056] Topology correlation includes the number of directly associated device failures and the topology path length. The number of directly associated device failures is the number of upstream and downstream devices that the device is directly connected to through the connection relationship in the fault indication map, and that belong to the candidate fault set. The topology path length is the shortest topology connection step between the device and the core fault device in the region.

[0057] Environmental consistency refers to the degree of consistency of environmental parameters among candidate faulty devices within the same area. Specifically, analysis of variance (ANOVA) can be used to calculate the variance of environmental parameters of candidate faulty devices within the area. The smaller the variance, the higher the environmental consistency.

[0058] This embodiment can accurately identify regional common-cause faults by calculating regional fault density, effectively distinguish between instantaneous false alarms and real faults by statistically analyzing fault time clustering, uncover cascading fault relationships between devices by analyzing topological correlation, and eliminate false faults caused by environmental factors by assessing environmental consistency. By associating the above four types of global features with the fused candidate fault set, fault analysis is upgraded from an isolated device perspective to a system-wide perspective, supplementing the global constraint information missing from single fault clues, providing comprehensive feature support for subsequent comprehensive evaluation, significantly improving the accuracy of fault identification, reducing false alarms and missed alarms, and helping maintenance personnel trace fault propagation paths and accurately locate root causes.

[0059] S4: Based on the comprehensive evaluation mechanism, the final fault list of the photovoltaic power station is obtained by comprehensively screening the fusion feature vectors.

[0060] In this embodiment, the comprehensive evaluation mechanism is as follows: Retain candidate fault device IDs whose regional fault density and time synchronization exceed the density threshold and synchronization threshold respectively, and whose environmental consistency is less than the consistency threshold. Candidate faulty device IDs whose regional fault density and topological correlation are less than the density threshold and correlation threshold, respectively, and whose duration is less than the duration threshold are removed. The neighboring device IDs of candidate faulty devices that have a topological correlation degree exceeding the correlation degree threshold and a recurrence frequency exceeding the recurrence frequency threshold are included in the candidate faulty devices.

[0061] This embodiment can accurately pinpoint regional common-cause faults by retaining devices that meet the standards for regional fault density, time synchronization, and environmental consistency. By eliminating devices with low fault density, low topological correlation, and short duration, it effectively filters out transient false alarms. By including adjacent devices of devices with high topological correlation and high recurrence frequency, it can capture the propagation path of cascading faults. This mechanism achieves refined screening based on global features, which not only avoids fault omissions but also suppresses invalid clues, upgrading fault identification from single-device judgment to system-level analysis. This significantly improves diagnostic accuracy, helps maintenance personnel quickly trace the root cause and coordinate the handling of faults, and reduces maintenance costs and power generation losses.

[0062] Example 2: This example also provides a photovoltaic power plant fault identification system based on hybrid reasoning, such as... Figure 2As shown, it includes a data acquisition module, a dual-channel inference module, a global collaboration module, and a comprehensive evaluation module; The data acquisition module collects multi-dimensional operational data of the photovoltaic power station and preprocesses it to obtain raw multi-source data. The dual-channel inference module is equipped with a fault indication map and a fault screening model. Based on the fault indication map, it performs map inference screening on the original multi-source data to obtain a first candidate fault set; based on the fault screening model, it performs model inference screening on the original multi-source data to obtain a second candidate fault set. The global collaboration module performs collaborative analysis on the first candidate fault set and the second candidate fault set based on the dual-channel collaboration mechanism to obtain a fused candidate fault set, and adds global features to the fused candidate fault set to obtain a fused feature vector. The comprehensive evaluation module performs a comprehensive evaluation of the fusion feature vector based on the comprehensive evaluation mechanism, and obtains the final fault list of the photovoltaic power station based on the evaluation results.

[0063] This embodiment integrates the photovoltaic power station fault identification method in this solution by constructing a corresponding system, realizing human-computer interaction and improving the user experience.

[0064] As can be seen from the above embodiments, it has at least the following substantial effects: (1) This invention makes up for the shortcomings of traditional single data feature mining by analyzing multi-dimensional operation data, thereby achieving comprehensive coverage of electrical and non-electrical faults; (2) This invention uses a dual-channel design of fault indication map and fault screening model to ensure the interpretability of reasoning through knowledge-driven approach and to uncover hidden faults through data-driven approach, thereby breaking the pain points of "black box" decision-making and rigid lag, and thus improving the identification accuracy. (3) This invention achieves an upgrade in judgment from isolated faults to common causes and cascading faults by integrating dual candidate sets and supplementing system-level information such as regional fault density and topological correlation through a dual-channel collaborative mechanism and global feature enhancement. (4) This invention outputs an accurate fault list through a comprehensive evaluation mechanism, and automatically adapts to equipment aging and topology change scenarios throughout the entire process, thereby reducing the dependence on operation and maintenance experience, improving response efficiency, reducing ineffective operation and maintenance costs and power generation loss, and combining engineering practicality and technical scalability.

[0065] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A photovoltaic power plant fault identification method based on hybrid reasoning, characterized in that: Includes the following steps: S1. Collect multi-dimensional operational data of photovoltaic power stations and preprocess the data to obtain raw multi-source data; S2. Construct a fault indication map and perform map inference and filtering on the original multi-source data to obtain the first candidate fault set; construct a fault screening model and perform model inference and filtering on the original multi-source data to obtain the second candidate fault set; S3. Based on the dual-channel collaborative mechanism, the first candidate fault set and the second candidate fault set are collaboratively analyzed to obtain the fused candidate fault set. Global features are added to the fused candidate fault set to obtain the fused feature vector. S4. Based on the comprehensive evaluation mechanism, the fusion feature vectors are comprehensively screened to obtain the final fault list of the photovoltaic power station.

2. The photovoltaic power plant fault identification method based on hybrid reasoning according to claim 1, characterized in that: In S1, the multi-dimensional operational data includes power plant operational data, equipment static data, environmental data, and topology data; The power plant operation data includes the current, voltage, output power, and equipment temperature of each device in the photovoltaic power plant; The static data of the equipment includes equipment ID, type, historical fault records, and equipment images; The environmental data includes ambient temperature, solar irradiance, and wind speed; The topology data includes the electrical connections, geographical locations, and regions of various devices within the photovoltaic power plant.

3. The photovoltaic power plant fault identification method based on hybrid reasoning according to claim 2, characterized in that: In S2, the process of constructing the fault indication map includes: Device nodes are constructed based on the geographical location and region of the devices in the topology data, with fault type as fault type node and fault manifestation as symptom node; The original atlas is obtained by connecting various device nodes based on the connection relationship between devices, connecting device nodes with fault type nodes based on the causal relationship between fault type and faulty device, and connecting symptom nodes with fault type nodes based on the correlation between fault type and fault manifestation. The original map is bound to preset evaluation rules and updated in real time based on the actual operating status of the photovoltaic power station to obtain the fault indication map.

4. The photovoltaic power plant fault identification method based on hybrid reasoning according to claim 3, characterized in that: In S2, the first candidate fault set is obtained by performing graph inference filtering on the original multi-source data, including the following steps: The original multi-source data is mapped to the symptom nodes in the fault indication map as mapping nodes. Starting from the mapping nodes, the possible fault type nodes and equipment nodes are traced based on causal and correlation relationships to obtain the corresponding tracing chain. Each traceability chain is evaluated based on preset evaluation rules to obtain inference confidence. Traceability chains with inference confidence exceeding the confidence threshold are selected and arranged in descending order of confidence to generate a structured first candidate fault set. The first candidate fault set consists of entries composed of candidate fault device ID, fault type, associated symptoms and inference confidence.

5. The photovoltaic power plant fault identification method based on hybrid reasoning according to claim 2, characterized in that: In S2, the specific process of constructing the fault screening model includes: The fault screening model is trained by using historical fault records of photovoltaic power plants as positive samples, normal operation data of photovoltaic power plants as negative samples, and the probability of identifying various fault types as the training objective.

6. The photovoltaic power plant fault identification method based on hybrid reasoning according to claim 5, characterized in that: In S2, the second candidate fault set is obtained by model inference filtering of the original multi-source data, including the following steps: Based on convolutional neural networks, feature extraction is performed on the static data of each device to obtain static features of the device; time series analysis is performed on the power plant operation data to obtain power plant operation features; and normalization processing is performed on the environmental data to obtain environmental features. The static characteristics of the equipment, the operating characteristics of the power plant, and the environmental characteristics are spliced ​​together and weighted to obtain the multi-source fusion characteristics of each equipment. The multi-source fusion features are used as input to the fault screening model to obtain the fault probability of each fault type. Fault types with fault probabilities higher than a set threshold are selected and sorted in descending order based on the fault probability to obtain a second candidate fault set. The second candidate fault set consists of entries composed of candidate fault device ID, fault type and corresponding fault probability.

7. The photovoltaic power plant fault identification method based on hybrid reasoning according to claim 5, characterized in that: In S3, a fused candidate fault set is obtained by performing collaborative analysis on the first candidate fault set and the second candidate fault set based on a dual-channel collaborative mechanism, including the following steps: Using the device ID as a unique identifier, the fault information of the same device in the first candidate fault set and the second candidate fault set are associated and the union is taken; If the same device appears in both the first candidate fault set and the second candidate fault set, the fault type in the first candidate fault set and the fault probability in the second candidate fault set are retained to obtain the initial candidate set. The fusion candidate fault set is obtained by removing duplicate device IDs and corresponding fault types from the initial candidate set.

8. The photovoltaic power plant fault identification method based on hybrid reasoning according to claim 7, characterized in that: In S3, global features are added to the fused candidate fault set to obtain the fused feature vector, including the following steps: The regional fault density is obtained by calculating the fault density of the area where the faulty equipment is located. The distribution pattern of fault types in the time dimension is statistically analyzed to obtain the fault time clustering degree, which includes time synchronization degree, duration and recurrence frequency. The topological correlation degree is obtained by analyzing the electrical connection relationships between each candidate faulty device; Environmental consistency is obtained by analyzing the environmental data of candidate faulty devices within the same area; The regional fault density, fault time clustering, topological correlation, and environmental consistency are used as global features and associated with the fusion candidate fault set to obtain a fusion feature vector.

9. The photovoltaic power plant fault identification method based on hybrid reasoning according to claim 8, characterized in that: In S4, the comprehensive evaluation mechanism is as follows: Retain candidate fault device IDs whose regional fault density and time synchronization exceed the density threshold and synchronization threshold respectively, and whose environmental consistency is less than the consistency threshold. Candidate faulty device IDs whose regional fault density and topological correlation are less than the density threshold and correlation threshold, respectively, and whose duration is less than the duration threshold are removed. The neighboring device IDs of candidate faulty devices that have a topological correlation degree exceeding the correlation degree threshold and a recurrence frequency exceeding the recurrence frequency threshold are included in the candidate faulty devices.

10. A photovoltaic power plant fault identification system based on hybrid reasoning, applicable to the photovoltaic power plant fault identification method based on hybrid reasoning as described in any one of claims 1-9, characterized in that: It includes a data acquisition module, a dual-channel inference module, a global collaboration module, and a comprehensive evaluation module; The data acquisition module collects multi-dimensional operational data of the photovoltaic power station and preprocesses it to obtain raw multi-source data. The dual-channel inference module is equipped with a fault indication map and a fault screening model. Based on the fault indication map, the module performs map inference screening on the original multi-source data to obtain the first candidate fault set. Based on the fault screening model, a second candidate fault set is obtained by performing model inference screening on the original multi-source data; The global collaboration module performs collaborative analysis on the first candidate fault set and the second candidate fault set based on the dual-channel collaboration mechanism to obtain a fused candidate fault set, and adds global features to the fused candidate fault set to obtain a fused feature vector. The comprehensive evaluation module performs a comprehensive evaluation of the fusion feature vector based on the comprehensive evaluation mechanism, and obtains the final fault list of the photovoltaic power station based on the evaluation results.

Images