A data-driven photovoltaic power plant operation and maintenance fault diagnosis method and system

Abstract

The application relates to the technical field of photovoltaic power station fault detection, and discloses a data-driven photovoltaic power station operation and maintenance fault diagnosis method and system. The method collects inverter operation and environment sequence data, generates a multi-dimensional time sequence data set through timestamp alignment; extracts principal components to build low-dimensional feature vectors by intercepting overlapping time sequence segments through a sliding window; connects adjacent vectors to form an evolution trajectory sequence, calculates vector distance to obtain a change intensity sequence, and marks a potential abnormal window in combination with distance change amplitude and trajectory bending degree; inputs the window feature vector into a pre-trained classification model to determine a fault category label, locks key parameters and deviation degree in combination with feature vector contribution weight, and locates a fault range in combination with an inverter and component string mapping relationship; matches power station layout coordinates to output a spatial fault positioning result containing a fault time, a category and a component string number. The method improves the accuracy and timeliness of fault diagnosis and adapts to the large-scale operation and maintenance requirements of photovoltaic power stations.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant fault detection technology, and in particular to a data-driven method and system for diagnosing operation and maintenance faults in photovoltaic power plants. Background Technology

[0002] Photovoltaic power generation is a core component of the clean energy system and occupies a key position in the global energy transition. With the continuous expansion of photovoltaic power plant installed capacity and the increasingly widespread distribution of equipment, the difficulty and pressure of power plant operation and maintenance management have increased significantly. Efficient fault prediction and health management are crucial for the stable operation of photovoltaic power plants. Accurate fault diagnosis and location not only directly affect the power plant's power generation revenue and equipment lifespan, but also become a core link in ensuring the high-quality and sustainable development of the photovoltaic industry.

[0003] Current fault diagnosis methods for photovoltaic power plants still rely primarily on manual inspections and alarms based on fixed threshold parameters. These methods are ill-suited to the complex actual operating conditions of power plants, revealing significant shortcomings in fault prediction and health management. The operating status of photovoltaic inverters and modules is affected by a combination of environmental factors, including sunlight, temperature, and dust obstruction. The data characteristics of the same fault vary significantly under different scenarios, and fixed threshold detection lacks dynamic adaptability. Furthermore, the sheer number of devices in a power plant and the tendency for various faults to intertwine make it impossible to capture the overall characteristics of a fault through single-indicator monitoring. This leads to frequent false alarms and missed alarms, making it difficult for maintenance personnel to quickly pinpoint the root cause of the fault and severely reducing the efficiency of fault prediction and health management.

[0004] Crucially, the complex coupling relationships among various operating parameters within a photovoltaic system pose a significant challenge to fault prediction and health management. Parameters such as component current, voltage, and power influence each other, and the characteristics of different fault types are superimposed across various physical quantities. Existing technologies cannot effectively integrate and extract features from multi-source coupled operating data, making it difficult to distinguish between dominant faults and interfering factors from complex parameter relationships, easily leading to misjudgments. Therefore, how to extract feature expressions that comprehensively reflect equipment operating status and effectively distinguish different fault modes from the massive, multi-source, and strongly coupled operating data of photovoltaic power plants, achieving accurate fault identification and location, and improving the intelligence level of fault prediction and health management, has become a core technical problem urgently needing to be solved in the field of photovoltaic power plant operation and maintenance. Summary of the Invention

[0005] This invention provides a data-driven method and system for diagnosing operation and maintenance faults in photovoltaic power plants, thereby improving the accuracy and timeliness of fault diagnosis and adapting to the large-scale operation and maintenance needs of photovoltaic power plants.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a data-driven method for diagnosing operation and maintenance faults in photovoltaic power plants, comprising: The inverter's DC-side voltage sequence, current sequence, AC-side power and internal temperature sequence, as well as the corresponding ambient irradiance sequence and component surface temperature sequence, are obtained and aligned by timestamp to generate a multi-dimensional time-series dataset. Overlapping time series segments are obtained by using a preset fixed-length sliding window to extract the multidimensional time series dataset, and principal components with a cumulative variance contribution rate reaching a preset ratio are extracted from each of the overlapping time series segments to form low-dimensional feature vectors for each of the overlapping time series segments. The adjacent low-dimensional feature vectors are connected to form an evolution trajectory sequence, and the vector distance value is calculated to obtain the change intensity sequence; Based on the distance change magnitude of the change intensity sequence and the curvature of the evolution trajectory sequence, abnormal turning points are identified and potential abnormal windows are marked. The low-dimensional feature vector of the potential anomaly window is input into a pre-trained classification model to determine the current fault category label; By combining the fault category labels and the contribution weights of the corresponding low-dimensional feature vectors, the key operating parameter combinations and deviations of the dominant faults are identified, and the location range of the module strings is located based on the mapping relationship between the inverter and the photovoltaic module strings. The location range of the component strings is matched with the layout coordinates of the photovoltaic power station, and the spatial fault location result, including the time of fault occurrence, fault type, and affected component string number, is output.

[0007] In one optional implementation, the acquisition of the inverter's DC-side voltage sequence, current sequence, AC-side power sequence, and internal temperature sequence, as well as the corresponding ambient irradiance sequence and component surface temperature sequence, and the generation of a multi-dimensional time-series dataset by aligning them according to timestamps, includes: The inverter's data acquisition module collects DC-side voltage sequence, current sequence, AC-side power, and internal temperature sequence in real time to obtain the first set of acquired data. By using environmental monitoring equipment deployed in photovoltaic power plants, the environmental irradiance sequence and the component surface temperature sequence at corresponding times are collected simultaneously to obtain the second collection data; Based on the timestamp of the acquisition module, a time matching mapping relationship of multi-source sequence data is established. The first and second acquired data are time axis interpolated and calibrated, and integrated according to the parameter type dimension to form a time-aligned multi-dimensional time series dataset.

[0008] In one optional implementation, the step of using a preset fixed-length sliding window to extract overlapping time series segments from the multidimensional time series dataset, and extracting principal components from each overlapping time series segment whose cumulative variance contribution rate reaches a preset proportion to form a low-dimensional feature vector for each overlapping time series segment, includes: A preset fixed-length sliding window is controlled to segment the multidimensional time series dataset segment by segment according to a set step size to obtain overlapping time series segments with overlapping data. Data normalization processing is performed on each of the overlapping time series segments to construct a multidimensional runtime parameter matrix corresponding to the dimension and runtime parameter type; The multidimensional operating parameter matrix is ​​dimensionality reduced, and the variance contribution rate of each principal component of the multidimensional operating parameter matrix is ​​calculated and accumulated in descending order. Based on the cumulative results, principal components whose cumulative variance contribution rate reaches a preset ratio are selected, and the feature information and dimensional features of the principal components are retained. The selected principal components are combined sequentially in descending order of variance contribution rate to form a low-dimensional feature vector representing the running state of the corresponding overlapping time segments.

[0009] In one optional implementation, the step of concatenating adjacent low-dimensional feature vectors to form an evolutionary trajectory sequence and calculating the vector distance value to obtain a change intensity sequence includes: In chronological order, adjacent low-dimensional feature vectors are sequentially connected to form an evolution trajectory sequence that characterizes the inverter's operating state. The vector distance calculation method is used to calculate the distance between each pair of adjacent low-dimensional feature vectors in the evolution trajectory sequence to obtain the distance value between each adjacent vector. The distance values ​​are arranged in chronological order to form a sequence of change intensity that characterizes the intensity of changes in the inverter's operating state.

[0010] In one optional implementation, identifying anomalous turning points and marking potential anomalous windows based on the distance change magnitude of the change intensity sequence and the curvature of the evolution trajectory sequence includes: Set thresholds for distance change magnitude and trajectory curvature as criteria for determining abnormal inverter operating status; The difference between adjacent distance values ​​in the change intensity sequence is calculated to obtain the distance change amplitude. At the same time, the trajectory curvature value of the evolution trajectory sequence in low-dimensional space is calculated to characterize the degree of trajectory curvature. The distance change amplitude is compared with the distance change amplitude threshold, and the trajectory curvature value is compared with the trajectory curvature threshold. If either the distance change amplitude or the trajectory curvature value exceeds the corresponding threshold, the corresponding position is determined to be an abnormal turning point. The overlapping time segments corresponding to the abnormal turning points are marked as potential abnormal windows, and the corresponding time information and low-dimensional feature vectors are recorded synchronously.

[0011] In one optional implementation, the step of inputting the low-dimensional feature vector of the potential anomaly window into a pre-trained classification model to determine the current fault category label includes: The low-dimensional feature vectors of the potential abnormal windows are matched for feature dimensions and then input into a classification model that has been pre-trained based on photovoltaic power plant fault samples. The classification model is used to identify and match fault features, and the matching probability of each fault category is output. The fault category with the highest matching probability is selected as the judgment result, and the corresponding fault category label is generated.

[0012] In one optional implementation, the step of combining the fault category label and the contribution weight of the corresponding low-dimensional feature vector to lock down the key operating parameter combination and deviation degree of the dominant fault, and locating the position range of the module string based on the mapping relationship between the inverter and the photovoltaic module string, includes: Based on the fault category label, the range of fault-related parameters is determined, and combined with the contribution weight of each principal component in the corresponding low-dimensional feature vector to the original operating parameters, the original operating parameters that meet the preset weight standard are selected to form the key operating parameter combination of the dominant fault. Calculate the deviation between the actual values ​​of the key operating parameter combination and the normal operating threshold, and quantify the degree of deviation from the normal state; Based on the degree of deviation, the corresponding mapping relationship between the inverter and the photovoltaic module string is retrieved. According to the inverter number to which the fault belongs, the photovoltaic module string connected to the inverter is located, and the location range of the module string is determined.

[0013] In one optional implementation, the step of matching the location range of the component strings with the layout coordinates of the photovoltaic power station and outputting a spatial fault location result including the fault occurrence time, fault type, and affected component string number includes: Retrieve the physical layout coordinate data of the photovoltaic power station, match the position range of the component string with the physical layout coordinate data, and determine the actual spatial coordinates corresponding to the component string; Based on the actual spatial coordinates, extract the time information corresponding to the potential anomaly window as the fault occurrence time, and integrate the fault occurrence time, fault category label and affected component string number; The integrated fault occurrence time, fault category label, and affected component string number are used to generate spatial fault location results according to a preset format.

[0014] Secondly, the present invention also provides a data-driven photovoltaic power plant operation and maintenance fault diagnosis system, comprising: Data acquisition and alignment module: acquires the DC side voltage sequence, current sequence, AC side power and internal temperature sequence of the inverter, as well as the corresponding ambient irradiance sequence and component surface temperature sequence, and aligns them according to timestamps to generate a multi-dimensional time series dataset; Feature dimensionality reduction and extraction module: The multidimensional time series dataset is truncated using a preset fixed-length sliding window to obtain overlapping time series segments, and principal components with a cumulative variance contribution rate reaching a preset ratio are extracted from each overlapping time series segment to form a low-dimensional feature vector for each overlapping time series segment; Trajectory sequence construction module: Connects adjacent low-dimensional feature vectors to form an evolutionary trajectory sequence, and calculates the vector distance value to obtain the change intensity sequence; Anomaly window marking module: Based on the distance change amplitude of the change intensity sequence and the curvature of the evolution trajectory sequence, identify abnormal turning points and mark potential anomaly windows; Fault category determination module: Input the low-dimensional feature vector of the potential abnormal window into the pre-trained classification model to determine the current fault category label; Fault location analysis module: Combining the fault category labels and the contribution weights of the corresponding low-dimensional feature vectors, it identifies the key operating parameter combinations and deviations that dominate the fault, and locates the position range of the module string based on the mapping relationship between the inverter and the photovoltaic module string. The result output matching module matches the location range of the component string with the layout coordinates of the photovoltaic power station, and outputs spatial fault location results including the time of fault occurrence, fault type, and affected component string number.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) Achieve fusion processing of multi-source time-series data and break through the limitations of single parameter monitoring. By synchronously collecting inverter operating parameters and power plant environmental parameter sequences, and accurately aligning them with timestamps to construct a multi-dimensional time-series dataset, the system effectively integrates strongly coupled multi-source operating data within the photovoltaic system, fully captures the full-dimensional characteristics of equipment operation, and lays a comprehensive and reliable data foundation for photovoltaic power plant fault prediction and health management.

[0016] (2) Accurately complete feature dimensionality reduction and extraction to meet the data processing needs of large-scale power plants. A fixed-length sliding window is used to extract overlapping time series segments. The core features are extracted by combining principal component analysis and a low-dimensional feature vector is constructed. While retaining the core operating features, the complexity of data processing is greatly reduced and the computational efficiency is improved. It can efficiently process massive time series data of photovoltaic power plants and meet the real-time monitoring and fault diagnosis needs of large-scale power plants.

[0017] (3) Construct a two-dimensional anomaly detection mechanism to improve the accuracy of fault identification. By constructing an evolution trajectory sequence through low-dimensional feature vectors, and combining the distance change amplitude and trajectory curvature of the change intensity sequence, the abnormal operation of the equipment is determined in two dimensions. The abnormal turning point is accurately identified and potential abnormal windows are marked to achieve early warning of faults, greatly reduce the probability of fault misjudgment and missed judgment, and optimize the early warning effect of fault prediction and health management of photovoltaic power plants.

[0018] (4) Achieve integrated fault diagnosis and location, and improve the efficiency of operation and maintenance. Input the feature vector of potential abnormal window into the pre-trained classification model to complete the accurate determination of fault category. Combine the contribution weight of feature vector to lock the key parameter combination and deviation degree of the dominant fault. Based on the mapping relationship between inverter and photovoltaic module string, match the power station layout coordinates to achieve integrated and accurate location of fault category, core cause and physical location, help operation and maintenance personnel quickly lock the root cause of the problem and improve the pertinence and efficiency of fault handling.

[0019] (5) Full-process data-driven automation, adapting to the intelligent operation and maintenance needs of power plants. The present invention achieves automated processing of the entire process from multi-source data acquisition, feature extraction and dimensionality reduction, anomaly detection and identification, to fault category determination and spatial location positioning based on data-driven methods. No manual intervention is required. It can adapt to the large-scale operation and maintenance characteristics of photovoltaic power plants with a large number of equipment and wide distribution, effectively reduce the cost of manual inspection, and greatly improve the intelligence and efficiency of fault prediction and health management of photovoltaic power plants, ensuring the stable and efficient operation of power plants. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a data-driven photovoltaic power plant operation and maintenance fault diagnosis method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a data-driven photovoltaic power plant operation and maintenance fault diagnosis system provided in an embodiment of the present invention. Detailed Implementation

[0021] 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, and 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.

[0022] Reference Figure 1 This invention provides a data-driven method for diagnosing operation and maintenance faults in photovoltaic power plants, comprising the following steps: S11, obtain the DC side voltage sequence, current sequence, AC side power and internal temperature sequence of the inverter, as well as the ambient irradiance sequence and component surface temperature sequence at the corresponding time, and generate a multi-dimensional time series dataset by aligning the timestamps. S12, the multidimensional time series dataset is truncated using a preset fixed-length sliding window to obtain overlapping time series segments, and principal components with a cumulative variance contribution rate reaching a preset ratio are extracted from each overlapping time series segment to form a low-dimensional feature vector of each overlapping time series segment. S13, connect adjacent low-dimensional feature vectors to form an evolution trajectory sequence, and calculate the vector distance value to obtain a change intensity sequence; S14, Based on the distance change amplitude of the change intensity sequence and the curvature of the evolution trajectory sequence, identify abnormal turning points and mark potential abnormal windows; S15, input the low-dimensional feature vector of the potential anomaly window into the pre-trained classification model to determine the current fault category label; S16, Combining the fault category label and the contribution weight of the corresponding low-dimensional feature vector, the key operating parameter combination and deviation degree of the dominant fault are locked, and the position range of the module string is located based on the mapping relationship between the inverter and the photovoltaic module string. S17, Match the location range of the component string with the layout coordinates of the photovoltaic power station, and output the spatial fault location result including the time of fault occurrence, fault type and affected component string number.

[0023] In step S11, the DC-side voltage sequence, current sequence, AC-side power and internal temperature sequence of the inverter are obtained, as well as the ambient irradiance sequence and component surface temperature sequence at the corresponding time, and a multi-dimensional time series dataset is generated by aligning the timestamps.

[0024] In one implementation, the step of acquiring the DC-side voltage sequence, current sequence, AC-side power sequence, and internal temperature sequence of the inverter, as well as the corresponding ambient irradiance sequence and module surface temperature sequence, and generating a multi-dimensional time-series dataset by aligning them according to timestamps, is specifically implemented using a 60MW centralized photovoltaic power station as an application scenario. This power station is equipped with 24 500kW string inverters, each connected to 18 photovoltaic module strings. The power station has established a high-precision multi-source data acquisition system for fault prediction and health management. The inverter data acquisition module's acquisition frequency is set to 2Hz, and the environmental monitoring equipment and acquisition module maintain the same acquisition frequency, with the time synchronization error controlled within ±10ms. This enables accurate synchronous acquisition of inverter operating parameters and environmental parameters. The specific implementation process is as follows: In this embodiment, the inverter's built-in industrial-grade high-precision data acquisition module first collects the inverter's DC-side voltage sequence, DC-side current sequence, AC-side power, and internal temperature sequence in real time at a sampling frequency of 2Hz. The acquisition module has a built-in crystal oscillator clock that generates 10ms-level timestamps. The collected DC-side voltage range is 600~800V, DC-side current range is 0~80A, AC-side power range is 0~500kW, and internal temperature range is -20~60℃. All collected parameters are stored as standardized digital sequences after 16-bit analog-to-digital conversion. The acquisition time for a single set of data is ≤5ms, resulting in the first set of collected data. This data contains four types of operating parameters, forming 1200 sets of time-series data every 10 minutes, completely preserving the time-series dynamic change characteristics of the inverter's core operating parameters. Secondly, dedicated environmental monitoring equipment deployed in the inverter array area and photovoltaic array area of ​​the photovoltaic power station synchronously collects the environmental irradiance sequence and photovoltaic module surface temperature sequence at a sampling frequency of 2Hz. The environmental monitoring equipment is equipped with a silicon-based irradiance sensor and a non-contact infrared temperature sensor. The irradiance detection accuracy is ±5W / ㎡, and the detection range is 0~2000W / ㎡. The module surface temperature detection accuracy is ±0.5℃, and the detection range is -40~85℃. The collected environmental parameters are stored as digital sequences after data formatting processing to obtain the second collection data. This data contains two types of environmental parameters, which, together with the first collection data, form six types of synchronous time-series parameters, effectively capturing the coupling relationship between the photovoltaic equipment operating status and environmental factors. Finally, using the 10ms-level timestamp generated by the inverter data acquisition module as a unified benchmark, a time matching mapping relationship for multi-source sequence data is established. For the minute time deviation of ≤10ms caused by the hardware response delay of the first and second acquired data, linear interpolation is used to perform time axis interpolation calibration on the two sets of acquired data. Missing parameter values ​​on the time axis are supplemented at 10ms time intervals to ensure that the first and second acquired data have unique corresponding parameter values ​​at the same timestamp, completely eliminating the time synchronization problem of multi-source acquisition devices. After calibration, parameters such as DC side voltage, DC side current, AC side power, internal temperature, and environmental conditions are then used for calibration. The two sets of collected data are integrated by combining the parameter types of irradiance and component surface temperature to construct a two-dimensional matrix with 10ms-level timestamps in the row dimension and 6 types of monitoring parameters in the column dimension. A 1200×6 multi-dimensional time-series dataset is formed for a single 10-minute period. This dataset achieves effective fusion of inverter operating parameters and environmental parameters, and completely and accurately captures the changes in the operating status of photovoltaic equipment over time and environmental factors. The processing time for a single dataset is ≤2s. It provides a unified, complete and accurate time-series data foundation for subsequent fault diagnosis, feature extraction, anomaly detection and other links, and adapts to the real-time data requirements of the entire process of fault prediction and health management of photovoltaic power plants.

[0025] In step S12, a preset fixed-length sliding window is used to extract overlapping time series segments from the multidimensional time series dataset, and principal components with a cumulative variance contribution rate reaching a preset ratio are extracted from each overlapping time series segment to form a low-dimensional feature vector for each overlapping time series segment.

[0026] In one implementation, this embodiment follows up on the time-aligned multidimensional time-series dataset obtained from the preceding steps for further processing. This dataset is a two-dimensional matrix of 10ms-level timestamps and six types of monitoring parameters. A single 10-minute time period forms 1200×6 time-series data, fully preserving the coupled time-series characteristics of inverter operating parameters and environmental parameters. Based on this, through sliding window truncation and principal component dimensionality reduction, feature extraction and dimensionality compression of massive time-series data are achieved. This improves the data analysis efficiency for subsequent fault prediction and health management while preserving core operating characteristics. The specific implementation process is as follows: First, a preset fixed-length sliding window is used to segment the multidimensional time-series dataset along the time axis at a set step size. In this embodiment, the fixed length of the sliding window is set to 60 timestamps, and the sliding step size is 30 timestamps. The 1200×6 multidimensional time-series dataset is continuously segmented, resulting in multiple overlapping time-series segments with 50% data overlap. Each overlapping time-series segment is a 60×6 two-dimensional matrix, covering 600ms of continuous running data. This data overlap design effectively avoids the loss of local running features due to window truncation, ensuring the feature integrity of each time-series data segment. Next, data normalization processing is performed on each overlapping time-series segment. First, invalid null values ​​caused by instantaneous acquisition errors of the equipment are removed from the matrix. Then, the maximum-minimum normalization formula is used to normalize all parameter values, eliminating dimensional differences between different parameters such as DC-side voltage, current, and irradiance. The formula is as follows:

[0027] in, These are the normalized parameter values. For the original value of the parameter, This is the maximum value of the parameter in the overlapping time segments. The minimum value of this parameter in the overlapping time series segments is normalized, and all parameter values ​​are mapped to the [0,1] interval. Then, a standardized multidimensional operational parameter matrix is ​​constructed according to the rules of timestamp rows and parameter type columns, ensuring that the dimensions of each matrix are consistent and the data format is uniform, laying the data foundation for subsequent dimensionality reduction processing. Principal component analysis (PCA) is then performed on the standardized multidimensional operational parameter matrix for dimensionality reduction. By solving the covariance matrix of the matrix, the eigenvalues ​​and eigenvectors of each principal component are calculated. The variance contribution rate of each principal component is calculated based on the proportion of eigenvalues. All principal components are then accumulated sequentially in descending order of variance contribution rate. The magnitude of the variance contribution rate directly characterizes the explanatory power of each principal component for the original multidimensional operational parameter features. The accumulation process can intuitively reflect the comprehensive feature explanatory power of the first N principal components. Next, based on the accumulated results, principal components whose cumulative variance contribution rate reaches a preset proportion are selected. In this embodiment, considering the characteristic requirements of photovoltaic power plant fault diagnosis, the preset proportion of cumulative variance contribution rate is set to 95%. After calculation, the cumulative variance contribution rate of the first three principal components of the 6-dimensional original parameters reaches 96.2%, which meets the preset proportion requirement. Based on this, these three principal components are selected, while retaining the characteristic information and dimensional features of each principal component completely, and eliminating redundant principal components with low contribution rates, thus achieving efficient dimensionality reduction of the data. Finally, the three selected principal components are combined sequentially in descending order of variance contribution rate to form a 3-dimensional low-dimensional feature vector. Each low-dimensional feature vector corresponds one-to-one with the corresponding overlapping time series segment. Under the premise of significantly compressing the data dimensionality, it can completely represent the core operating status characteristics of the inverter and photovoltaic modules in the corresponding time period. In this embodiment, the 6-dimensional original data is compressed into a 3-dimensional low-dimensional feature vector, reducing the data processing volume by 50% and retaining 96.2% of the original feature information. The formed low-dimensional feature vector can be directly used for subsequent evolution trajectory sequence construction, adapting to the real-time data processing requirements of photovoltaic power plant fault prediction and health management.

[0028] In step S13, adjacent low-dimensional feature vectors are connected to form an evolution trajectory sequence, and the vector distance value is calculated to obtain a change intensity sequence.

[0029] In one implementation, this embodiment follows up on the low-dimensional feature vector obtained in the aforementioned steps with subsequent processing. This low-dimensional feature vector is a 3-dimensional vector, formed by combining the top three principal components with a cumulative variance contribution rate of 96.2% selected from the original 6-dimensional parameters through principal component analysis, arranged in descending order of variance contribution rate. Each 3-dimensional low-dimensional feature vector corresponds to a 60×6 overlapping time series segment, fully characterizing the core operating state of the inverter and photovoltaic modules during that period. Based on this, an evolution trajectory sequence is constructed and a change intensity sequence is calculated, transforming the time series changes in the inverter's operating state into quantifiable feature space trajectories and intensity indicators. This provides accurate judgment criteria for subsequent anomaly detection, adapting to the status monitoring needs of photovoltaic power plant fault prediction and health management. The specific implementation process is as follows: First, according to the timestamp order of the overlapping time series segments corresponding to each low-dimensional feature vector, adjacent 3D low-dimensional feature vectors are sequentially connected in the 3D feature space. Each low-dimensional feature vector serves as a node in the evolution trajectory sequence. The connections between nodes reflect the continuous change trend of the inverter's operating state from one time period to the next, ultimately forming an evolution trajectory sequence that can completely characterize the temporal changes in the inverter's operating state. In this embodiment, the 10-minute multi-dimensional time series dataset is truncated by a sliding window to generate 39 low-dimensional feature vectors, corresponding to a 3D evolution trajectory sequence containing 39 nodes. This sequence can intuitively present the changing pattern of the inverter's operating state in the feature space, providing a foundation for the visual identification of state anomalies. Second, the Euclidean distance formula is used as a vector distance calculation method adapted to 3D low-dimensional feature vectors. This method has high calculation accuracy and fast computation efficiency, adapting to the computing power requirements of real-time monitoring of photovoltaic power plants. The formula is as follows:

[0030] in, The Euclidean distance between two adjacent low-dimensional feature vectors is given by x=(x1,x2,x3) and y=(y1,y2,y3), which are two adjacent 3-dimensional low-dimensional feature vectors in the evolution trajectory sequence, respectively. The distance between each adjacent 3-dimensional low-dimensional feature vector in the evolution trajectory sequence is calculated according to this formula, yielding the Euclidean distance value between each adjacent vector. The magnitude of this distance value is positively correlated with the change in the inverter's operating state in adjacent time periods. A larger distance value indicates a more significant change in the inverter's operating state, while a smaller distance value indicates a more stable operating state. In this embodiment, the calculated Euclidean distance between adjacent vectors ranges from 0.02 to 0.85, fully quantifying the degree of change in the inverter's operating state in each time period. Finally, all calculated Euclidean distance values ​​are arranged sequentially according to the time order of their corresponding overlapping time segments to form a change intensity sequence that corresponds one-to-one with the nodes of the evolution trajectory sequence. In this embodiment, the change intensity sequence contains 38 distance value data, which can accurately reflect the fluctuation and change intensity of the inverter's operating status within a 10-minute monitoring period. Abrupt changes in the values ​​in the sequence can directly correspond to abnormal fluctuations in the inverter's operating status, providing quantified and orderly intensity index data for subsequent dual-dimensional anomaly detection combining distance change amplitude and trajectory curvature. This change intensity sequence and the evolution trajectory sequence work together to form a dual judgment basis for anomaly detection of inverter operating status, greatly improving the accuracy of anomaly identification in fault prediction and health management.

[0031] In step S14, based on the distance change magnitude of the change intensity sequence and the curvature of the evolution trajectory sequence, abnormal turning points are identified and potential abnormal windows are marked.

[0032] In one embodiment, this example connects the evolution trajectory sequence and the change intensity sequence obtained in the aforementioned steps for subsequent processing. The evolution trajectory sequence is a 3D feature space trajectory containing 39 nodes, which fully represents the continuous change trend of the inverter's operating status within a 10-minute time period. The change intensity sequence consists of 38 Euclidean distance values, with a distance value range of 0.02 to 0.85, which accurately quantifies the change amplitude of the inverter's operating status in adjacent time periods. Based on this, the accurate identification of inverter operating anomalies is achieved through dual-index threshold judgment, thus locking in abnormal data segments for photovoltaic power plant fault prediction and health management. The specific implementation process is as follows: First, by combining the statistical characteristics of the inverter's historical normal operation data and the characteristics of fault precursor data, and after multiple training and verifications, a distance change threshold of 0.6 and a trajectory curvature threshold of 1.5 were set as the core criteria for judging abnormal inverter operation. These thresholds are adapted to the actual operating conditions of the power station and can effectively distinguish between normal state fluctuations and abnormal state abrupt changes of the inverter, taking into account both the accuracy of fault prediction and the control of the missed detection rate, and providing a quantitative standard for the identification of abnormal turning points. Subsequently, the difference between adjacent distance values ​​in the change intensity sequence is calculated to obtain the distance change amplitude, which characterizes the degree of fluctuation in the state change. This value intuitively reflects the abruptness of the inverter's operating state change. The larger the difference, the more abrupt the change in the operating state in adjacent time periods. At the same time, the 3D spatial curvature calculation formula is used to calculate the curvature value of each node in the evolution trajectory sequence point by point. The curvature value characterizes the degree of trajectory curvature. The larger the curvature value, the more abnormal the trend of the inverter's operating state change at that node, and the more likely it is to have fault precursor characteristics. In this embodiment, the calculated distance change amplitude range is 0.01~0.72, and the trajectory curvature value range is 0.1~1.8, which fully covers the fluctuation and abrupt change characteristics of the inverter's operating state. Next, the calculated distance change amplitude is compared one by one with the preset distance change amplitude threshold of 0.6, and the trajectory curvature value is compared one by one with the preset trajectory bending degree threshold of 1.5. According to the rule that any index exceeding the corresponding threshold is judged as abnormal, the state of each node in the evolution trajectory sequence is judged. In this embodiment, it is detected that the distance change amplitude of two nodes reaches 0.68 and 0.72 respectively, both exceeding the distance change amplitude threshold, and the trajectory curvature value of one node reaches 1.65, exceeding the trajectory bending degree threshold. Based on this, these three nodes are accurately judged as abnormal turning points in the inverter's operating state, that is, the key nodes for the inverter to change from the normal operating state to the abnormal state. Finally, all overlapping time segments corresponding to the above-mentioned abnormal turning points are marked as potential abnormal windows. The time information corresponding to each potential abnormal window is recorded synchronously and accurately, including the 10ms-level timestamps of the start and end of the window, as well as the 3-dimensional low-dimensional feature vector corresponding to the window. In this embodiment, a total of 3 potential abnormal windows are marked. The recorded time information can accurately trace the specific time period of the inverter's abnormal operation. The low-dimensional feature vector completely retains the core operating characteristics of the abnormal period, providing a targeted feature data foundation for the subsequent accurate determination of fault categories. This realizes the accurate positioning and feature retention of abnormal states in photovoltaic power plant fault prediction and health management.

[0033] In step S15, the low-dimensional feature vector of the potential anomaly window is input into a pre-trained classification model to determine the current fault category label.

[0034] In one embodiment, this example follows the previous steps to process the potential anomaly windows and corresponding low-dimensional feature vectors obtained in subsequent steps. A total of three potential anomaly windows were marked, each corresponding to a 3D low-dimensional feature vector. This vector is composed of the top three principal components with a cumulative variance contribution rate of 96.2% extracted from inverter operation and environmental parameters through principal component analysis. This fully preserves the core characteristics of photovoltaic equipment operation during abnormal periods. Simultaneously, 10ms-level time information for each window is recorded. Based on this, the pre-trained classification model in the photovoltaic power plant fault prediction and health management system is used to accurately identify fault features and determine their categories. The specific implementation process is as follows: First, feature dimension matching is performed on the 3-dimensional low-dimensional feature vectors corresponding to the 3 potential anomaly windows. The consistency between the vector dimension and the input layer dimension of the classification model is checked to ensure that there is no missing or redundant dimensions. In this embodiment, the input layer dimension of the pre-trained classification model is 3-dimensional, which is completely matched with the dimension of the low-dimensional feature vector. No additional dimension conversion is required. The matched low-dimensional feature vectors are directly input into the classification model that has been pre-trained based on photovoltaic power station fault samples. The pre-trained classification model is described in detail below: Model Architecture: The classification model is built on a random forest algorithm and consists of 100 CART decision trees. The maximum depth of each decision tree is set to 8 layers, the minimum number of splits is 5, and the minimum number of leaf nodes is 2. Ensemble learning is used to reduce the overfitting risk of a single decision tree and improve the model's generalization ability. Training Sample Set: The training samples cover 10 typical conditions common in photovoltaic power plants, including hot spots on modules, module aging, abnormal DC voltage of inverters, abnormal temperature rise of inverter power tubes, sudden irradiance interference, module shading, loose wiring, capacitor aging, heat dissipation failure, and normal operation. The number of fault samples is ≥8000 sets (≥800 sets of each type of fault) and the number of normal operation samples is ≥15000 sets. All samples are from the measured data of this 60MW photovoltaic power plant and three similar power plants. The samples cover the equipment operation status under different seasons, different light intensities, and different temperature environments to ensure that the model is adapted to complex operating conditions. Training Process: The model is trained using a 5-fold cross-validation method. The sample set is divided into 8 layers and 10 layers. The model is divided into training and testing sets in a 2:1 ratio. The training set is used for iterative optimization of model parameters, while the testing set is used to verify model performance. Hyperparameter optimization is performed using a grid search method, with the optimization objective being to minimize the fault classification error rate on the testing set, ultimately determining the optimal hyperparameter combination. An early stopping mechanism is employed during training: training stops when the accuracy on the testing set does not improve after 10 consecutive iterations to avoid overfitting. Model performance metrics: After training, the model achieves a fault identification accuracy of over 98.5% on the testing set, with core faults such as hot spots on modules and abnormal DC-side voltage in inverters achieving an accuracy of ≥99%. The confusion matrix for various faults shows a cross-misclassification rate of ≤1.2% for different fault categories. The model's recall rate is ≥98.3%, and its precision rate is ≥98.7%, meeting the high-precision judgment requirements for fault prediction and health management in photovoltaic power plants. Model deployment: After training, the model is deployed in a lightweight format (ONNX format) on the edge computing nodes of the photovoltaic power plant, supporting real-time inference. The inference time for a single 3D low-dimensional feature vector is ≤10ms, adapting to the computing power requirements for real-time monitoring of the power plant.

[0035] This classification model identifies and matches fault features from the input low-dimensional feature vectors. Internally, the model uses multiple decision trees to extract and compare core fault features from the feature vectors layer by layer. The extracted features are then precisely matched with feature templates for various fault types in the model. Simultaneously, reverse verification is performed using normal operation feature templates. Finally, the matching probability of each fault category corresponding to the low-dimensional feature vector is output, and the probability value is presented as a percentage. The sum of the matching probabilities of all categories is 100%. In this embodiment, after the model matches the three sets of input low-dimensional feature vectors, the matching probabilities of two sets of vectors corresponding to hot spot faults in the components are 92.3% and 94.7%, respectively. The matching probability of one set of vectors corresponding to DC-side voltage anomalies in the inverter is 91.5%. The matching probabilities of the remaining fault categories are all below 10%, and there is no high probability confusion among multiple categories. Finally, based on the principle of highest matching probability, the fault category with the highest matching probability corresponding to each set of feature vectors is selected as the final judgment result. According to the power plant fault coding rules, standardized fault category labels are generated for the judgment results. The label corresponding to the hot spot fault of the module is "RZ-RB01", and the label corresponding to the DC side voltage abnormality fault of the inverter is "NB-ZL02". The label contains key information such as fault type and equipment type. It can be directly connected to the operation and maintenance management system of photovoltaic power plant fault prediction and health management, providing a clear fault type basis for subsequent fault key parameter locking and location positioning.

[0036] In step S16, by combining the fault category label and the contribution weight of the corresponding low-dimensional feature vector, the key operating parameter combination and deviation degree of the dominant fault are locked, and the location range of the module string is located based on the mapping relationship between the inverter and the photovoltaic module string.

[0037] In one embodiment, this embodiment follows up on the fault category labels and corresponding 3D low-dimensional feature vectors obtained in the aforementioned steps for subsequent processing. Two groups of module hot spot faults (labeled "RZ-RB01") and one group of inverter DC-side voltage abnormality faults (labeled "NB-ZL02") have been identified. Each fault category retains 96.2% of the original features in its corresponding 3D low-dimensional feature vector, which is obtained through principal component analysis. The contribution weights of each principal component to six original operating parameters—DC-side voltage, current, AC-side power, internal temperature, ambient irradiance, and module surface temperature—have been quantified and statistically analyzed. Based on this, and combined with the fault tracing requirements for photovoltaic power plant fault prediction and health management, the key parameters of the dominant fault are identified and the degree of deviation is quantified. Simultaneously, the module string position is initially located based on the equipment mapping relationship. The specific implementation process is as follows: First, the range of fault-related parameters is determined based on the generated fault category labels. The core related parameters for the module hot spot fault (“RZ-RB01”) are DC-side current, module surface temperature, and ambient irradiance. The core related parameters for the inverter DC-side voltage anomaly fault (“NB-ZL02”) are DC-side voltage, internal temperature, and AC-side power. Then, the contribution weights of each principal component in the corresponding low-dimensional feature vector to the six original operating parameters are extracted. In this embodiment, principal component analysis shows that the contribution weights of the principal component corresponding to the module hot spot fault to the module surface temperature, DC-side current, and ambient irradiance are 42.5%, 31.2%, and 1%, respectively. 8.7% of the inverter's DC-side voltage anomaly faults contribute 48.3%, 29.6%, and 15.4% to the DC-side voltage, internal temperature, and AC-side power, respectively. A contribution weight ≥15% is set as a high-weight criterion. The original operating parameters under this criterion are selected to form the dominant fault key operating parameter combinations for two types of faults: module hotspot faults are defined as "module surface temperature + DC-side current + ambient irradiance," and inverter DC-side voltage anomaly faults are defined as "DC-side voltage + internal temperature + AC-side power." This combination accurately points to the core influencing parameters of the fault, providing a clear data direction for fault tracing. The high-weight criterion is obtained based on historical operating data statistics or pre-configured in the system by maintenance personnel. Subsequently, the normal operating parameter thresholds of the 500kW string inverter in the photovoltaic power station were retrieved. These included a DC-side voltage threshold of 600-800V, a DC-side current threshold of 0-80A, a module surface temperature threshold of -20-60℃, an internal temperature threshold of -20-60℃, an ambient irradiance threshold of 0-2000W / ㎡, and an AC-side power threshold of 0-500kW. These normal operating parameter thresholds were determined jointly by the inverter manufacturer's rated parameters and historical stable operating data. The actual values ​​of each critical fault operating parameter combination were extracted and calculated using the deviation rate formula:

[0038] The deviation between the actual value and the normal operating threshold is calculated, and the degree of deviation from the normal state is quantified. The quantified degree of deviation intuitively reflects the changes and severity of the core parameters of the fault, providing a quantitative basis for fault level determination. Finally, based on the aforementioned quantified deviation, it was determined that the fault had reached the maintenance level requiring precise equipment location. The pre-set mapping table between inverters and photovoltaic module strings in the photovoltaic power station was then retrieved. This table fully records the unique device number of each 500kW string inverter in the power station, along with the numbers of the 18 connected photovoltaic module strings, array affiliation, and other information. Based on the inverter number to which the fault belongs (in this embodiment, all faults belong to the inverter numbered INV-06), the 18 connected photovoltaic module strings of that inverter were precisely located. Combining the parameter variation characteristics of the deviation, module strings without parameter fluctuations were excluded. Ultimately, the 03-08 photovoltaic module strings under the INV-06 inverter were determined to be the range of module strings affected by the fault. This range clearly defines the equipment interval for subsequent precise fault location, significantly reducing the on-site investigation scope for maintenance personnel and improving the efficiency of photovoltaic power station fault prediction and health management.

[0039] In step S17, the location range of the component string is matched with the layout coordinates of the photovoltaic power station, and a spatial fault location result including the time of fault occurrence, fault type and affected component string number is output.

[0040] In one embodiment, this example follows up on the location range of the affected component strings, fault category labels, and potential anomaly window time information obtained in the aforementioned steps for subsequent processing. The inverter number associated with the fault has been identified as INV-06, and the affected component strings are channels 03-08 under this inverter. Two groups of component hot spot faults (labeled "RZ-RB01") and one group of inverter DC-side voltage anomaly faults (labeled "NB-ZL02") have been identified. Simultaneously, 10ms-level timestamps corresponding to each potential anomaly window are recorded. Based on this, and combined with the operational requirements of photovoltaic power plant fault prediction and health management, precise matching of component string locations with physical layout coordinates is achieved. All fault information is integrated, and standardized spatial fault location results are generated, providing accurate and intuitive fault guidance for on-site operational and maintenance troubleshooting. The specific implementation process is as follows: First, the pre-set physical layout coordinate data of the entire photovoltaic power station is retrieved. This data is a spatial coordinate library of power station equipment built based on a GIS system. It completely records the actual geographical coordinates (latitude and longitude), array area location numbers, and equipment installation point information of all inverters and photovoltaic module strings in the power station. Each module string corresponds to a unique latitude and longitude coordinate and array area grid number with an accuracy of ±0.5m, which can accurately match the physical location of the equipment. In this embodiment, the associated coordinate data of module strings 03-08 under INV-06 inverter is directly retrieved from the coordinate library. The previously determined location range of the module strings affected by the fault is then precisely matched with the physical layout coordinates in the library for each module string. The actual spatial coordinates of each of the six component strings from INV-06-03 to INV-06-08 were determined one by one. Among them, the hot spot fault of the component corresponds to the component strings INV-06-05 and 06, with coordinates of 118.XX.XX.68°E, 32.XX.XX.25°N and 118.XX.XX.69°E, 32.XX.XX.26°N, respectively. The DC side voltage abnormality fault of the inverter corresponds to the component string INV-06-04, with coordinates of 118.XX.XX.67°E, 32.XX.XX.24°N. This achieves the accurate conversion of the location of the faulty equipment from logical number to physical spatial coordinates. Subsequently, based on the matched actual spatial coordinates, the potential anomaly window time information corresponding to the fault is associated, and the starting 10ms timestamp of each anomaly window is extracted as the fault occurrence time. The occurrence times of the two groups of component hotspot faults are XXXX-XX-XX 14:25:06.030 and XXXX-XX-XX 14:32:18.050, respectively, and the occurrence time of the inverter DC side voltage anomaly fault is XXXX-XX-XX 14:28:45.020. The fault occurrence time, standardized fault category label, and unique device number of the affected component string are then integrated one-to-one to form a complete information set for each fault. The data in each dimension of the information set are interconnected and non-redundant, accurately reflecting the time, type, device, and spatial core characteristics of the fault. The fault level is determined based on the degree of deviation and its correspondence with a preset level threshold range.Finally, the integrated full set of fault information is packaged according to the preset standardized format of the photovoltaic power plant fault prediction and health management system. This format is adapted to the display and push requirements of the power plant operation and maintenance monitoring terminal and includes six core dimensions: fault code, occurrence time, fault category, affected equipment, spatial coordinates, and fault level. The fault level is automatically determined based on the deviation of the aforementioned parameters. The fault level can be a level 1 fault, a level 2 fault, etc. For example, a hot spot fault in a module can be a level 2 fault, while an abnormal DC voltage fault in an inverter can be a level 1 fault. The final result is a structured spatial fault location result, which can be directly pushed to the power plant operation and maintenance monitoring platform and the on-site operation and maintenance handheld terminal. It is presented intuitively in the form of "Fault Code: GF-202602-016, Occurrence Time: XXXX-XX-XX14:25:06.030, Fault Category: Hot Spot in Module (RZ-RB01), Affected Module String: INV-06-05, Spatial Coordinates: 118.XX.XX.68°E32.XX.XX.25°N, Fault Level: Level 2", realizing the standardized and visualized output of fault information.

[0041] This step, through precise matching of spatial coordinates and standardized integration of fault information, completes a closed loop from fault diagnosis at the data level to fault location at the physical level. The generated spatial fault location results provide direct and accurate operation and maintenance basis for photovoltaic power plant fault prediction and health management. Operation and maintenance personnel can directly locate the fault site based on the results, significantly shortening the time for fault investigation and handling, effectively improving the operation and maintenance efficiency of photovoltaic power plants, and reducing power generation losses caused by fault downtime.

[0042] In summary, this embodiment uses a 60MW centralized photovoltaic power plant as an application scenario. Addressing the actual needs of photovoltaic power plant fault prediction and health management, it fully implements the entire process of a data-driven photovoltaic power plant operation and maintenance fault diagnosis method. This includes accurate collection and timestamp alignment of multi-source time-series data, feature dimensionality reduction and low-dimensional feature vector construction through sliding window extraction and principal component analysis, construction of evolution trajectory sequences and change intensity sequences, identification of abnormal turning points and marking of potential abnormal windows based on distance change amplitude and trajectory curvature, accurate fault category determination through a pre-trained classification model, locking the key operating parameter combinations of the dominant fault and quantifying the degree of deviation based on principal component contribution weights, preliminary fault range localization based on the mapping relationship between inverters and photovoltaic module strings, and finally, matching the physical layout coordinates of the power plant to generate standardized fault location results including the fault occurrence time, fault category, affected module string number, and spatial coordinates.

[0043] The entire implementation process is data-driven and fully automated, with no manual intervention. The methods employed, such as maximum-minimum normalization and Euclidean distance calculation, effectively solve technical problems related to multi-source parameter coupling, dimensional differences, and excessively high feature dimensions. Data is seamlessly integrated across steps, and quantitative indicators are reasonable. The fault identification accuracy exceeds 98.5%, and fault location accuracy reaches down to a single module string, significantly shortening fault investigation and handling time. This embodiment verifies the feasibility and effectiveness of this fault diagnosis method in the operation and maintenance of large-scale photovoltaic power plants. It overcomes the technical limitations of traditional manual inspections and fixed threshold alarms, achieving integrated fault diagnosis and spatial location. This significantly improves the intelligence, accuracy, and efficiency of photovoltaic power plant fault prediction and health management, effectively reducing downtime losses, ensuring the stable and continuous operation of photovoltaic power plants, and adapting to the operation and maintenance needs of photovoltaic power plants of different sizes. It possesses good practical application value and promising prospects for promotion.

[0044] refer to Figure 2 The second embodiment of the invention provides a data-driven photovoltaic power plant operation and maintenance fault diagnosis system, including: Data acquisition and alignment module: acquires the DC side voltage sequence, current sequence, AC side power and internal temperature sequence of the inverter, as well as the corresponding ambient irradiance sequence and component surface temperature sequence, and aligns them according to timestamps to generate a multi-dimensional time series dataset; Feature dimensionality reduction and extraction module: The multidimensional time series dataset is truncated using a preset fixed-length sliding window to obtain overlapping time series segments, and principal components with a cumulative variance contribution rate reaching a preset ratio are extracted from each overlapping time series segment to form a low-dimensional feature vector for each overlapping time series segment; Trajectory sequence construction module: Connects adjacent low-dimensional feature vectors to form an evolutionary trajectory sequence, and calculates the vector distance value to obtain the change intensity sequence; Anomaly window marking module: Based on the distance change amplitude of the change intensity sequence and the curvature of the evolution trajectory sequence, identify abnormal turning points and mark potential anomaly windows; Fault category determination module: Input the low-dimensional feature vector of the potential abnormal window into the pre-trained classification model to determine the current fault category label; Fault location analysis module: Combining the fault category labels and the contribution weights of the corresponding low-dimensional feature vectors, it identifies the key operating parameter combinations and deviations that dominate the fault, and locates the position range of the module string based on the mapping relationship between the inverter and the photovoltaic module string. The result output matching module matches the location range of the component string with the layout coordinates of the photovoltaic power station, and outputs spatial fault location results including the time of fault occurrence, fault type, and affected component string number.

[0045] It should be noted that the data-driven photovoltaic power plant operation and maintenance fault diagnosis system provided in this embodiment of the invention is used to execute all the process steps of the data-driven photovoltaic power plant operation and maintenance fault diagnosis method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0046] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A data-driven method for diagnosing operation and maintenance faults in photovoltaic power plants, characterized in that, include: The inverter's DC-side voltage sequence, current sequence, AC-side power and internal temperature sequence, as well as the corresponding ambient irradiance sequence and component surface temperature sequence, are obtained and aligned by timestamp to generate a multi-dimensional time-series dataset. Overlapping time series segments are obtained by using a preset fixed-length sliding window to extract the multidimensional time series dataset, and principal components with a cumulative variance contribution rate reaching a preset ratio are extracted from each of the overlapping time series segments to form low-dimensional feature vectors for each of the overlapping time series segments. The adjacent low-dimensional feature vectors are connected to form an evolution trajectory sequence, and the vector distance value is calculated to obtain the change intensity sequence; Based on the distance change magnitude of the change intensity sequence and the curvature of the evolution trajectory sequence, abnormal turning points are identified and potential abnormal windows are marked. The low-dimensional feature vector of the potential anomaly window is input into a pre-trained classification model to determine the current fault category label; By combining the fault category labels and the contribution weights of the corresponding low-dimensional feature vectors, the key operating parameter combinations and deviations of the dominant faults are identified, and the location range of the module strings is located based on the mapping relationship between the inverter and the photovoltaic module strings. The location range of the component strings is matched with the layout coordinates of the photovoltaic power station, and the spatial fault location result, including the time of fault occurrence, fault type, and affected component string number, is output.

2. The data-driven photovoltaic power plant operation and maintenance fault diagnosis method according to claim 1, characterized in that, The process involves acquiring the DC-side voltage sequence, current sequence, AC-side power sequence, and internal temperature sequence of the inverter, as well as the corresponding ambient irradiance sequence and component surface temperature sequence, and aligning them by timestamp to generate a multi-dimensional time-series dataset, including: The inverter's data acquisition module collects DC-side voltage sequence, current sequence, AC-side power, and internal temperature sequence in real time to obtain the first acquisition data. By using environmental monitoring equipment deployed in photovoltaic power plants, the environmental irradiance sequence and the component surface temperature sequence at corresponding times are collected simultaneously to obtain the second collection data; Based on the timestamp of the acquisition module, a time matching mapping relationship of multi-source sequence data is established. The first and second acquired data are time axis interpolated and calibrated, and integrated according to the parameter type dimension to form a time-aligned multi-dimensional time series dataset.

3. The data-driven photovoltaic power plant operation and maintenance fault diagnosis method according to claim 1, characterized in that, The method involves using a preset fixed-length sliding window to extract overlapping time series segments from the multidimensional time series dataset, and extracting principal components whose cumulative variance contribution rate reaches a preset proportion from each overlapping time series segment to form a low-dimensional feature vector for each overlapping time series segment, including: A preset fixed-length sliding window is controlled to segment the multidimensional time series dataset segment by segment according to a set step size to obtain overlapping time series segments with overlapping data. Data normalization processing is performed on each of the overlapping time series segments to construct a multidimensional runtime parameter matrix corresponding to the dimension and runtime parameter type; The multidimensional operating parameter matrix is ​​dimensionality reduced, and the variance contribution rate of each principal component of the multidimensional operating parameter matrix is ​​calculated and accumulated in descending order. Based on the cumulative results, principal components whose cumulative variance contribution rate reaches a preset ratio are selected, and the feature information and dimensional features of the principal components are retained. The selected principal components are combined sequentially in descending order of variance contribution rate to form a low-dimensional feature vector representing the running state of the corresponding overlapping time segments.

4. The data-driven photovoltaic power plant operation and maintenance fault diagnosis method according to claim 1, characterized in that, The step of connecting adjacent low-dimensional feature vectors to form an evolutionary trajectory sequence and calculating the vector distance value to obtain a change intensity sequence includes: In chronological order, adjacent low-dimensional feature vectors are sequentially connected to form an evolution trajectory sequence that characterizes the inverter's operating state. The vector distance calculation method is used to calculate the distance between each pair of adjacent low-dimensional feature vectors in the evolution trajectory sequence to obtain the distance value between each adjacent vector. The distance values ​​are arranged in chronological order to form a sequence of change intensity that characterizes the intensity of changes in the inverter's operating state.

5. The data-driven photovoltaic power plant operation and maintenance fault diagnosis method according to claim 1, characterized in that, The step of identifying anomalous turning points and marking potential anomalous windows based on the distance change magnitude of the change intensity sequence and the curvature of the evolution trajectory sequence includes: Set thresholds for distance variation and trajectory curvature as criteria for determining abnormal inverter operation. The difference between adjacent distance values ​​in the change intensity sequence is calculated to obtain the distance change amplitude. At the same time, the trajectory curvature value of the evolution trajectory sequence in low-dimensional space is calculated to characterize the degree of trajectory curvature. The distance change amplitude is compared with the distance change amplitude threshold, and the trajectory curvature value is compared with the trajectory curvature threshold. If either the distance change amplitude or the trajectory curvature value exceeds the corresponding threshold, the corresponding position is determined to be an abnormal turning point. The overlapping time segments corresponding to the abnormal turning points are marked as potential abnormal windows, and the corresponding time information and low-dimensional feature vectors are recorded synchronously.

6. The data-driven photovoltaic power plant operation and maintenance fault diagnosis method according to claim 1, characterized in that, The step of inputting the low-dimensional feature vector of the potential anomaly window into a pre-trained classification model to determine the current fault category label includes: The low-dimensional feature vectors of the potential anomaly windows are matched for feature dimensions and then input into a classification model that has been pre-trained based on photovoltaic power plant fault samples. The classification model is used to identify and match fault features, and the matching probability of each fault category is output. The fault category with the highest matching probability is selected as the judgment result, and the corresponding fault category label is generated.

7. The data-driven photovoltaic power plant operation and maintenance fault diagnosis method according to claim 1, characterized in that, The method combines the contribution weights of the fault category labels and corresponding low-dimensional feature vectors to identify the key operating parameter combinations and deviations that dominate the fault, and locates the position range of the module strings based on the mapping relationship between the inverter and the photovoltaic module strings, including: Based on the fault category label, the range of fault-related parameters is determined, and combined with the contribution weight of each principal component in the corresponding low-dimensional feature vector to the original operating parameters, the original operating parameters that meet the preset weight standard are selected to form the key operating parameter combination of the dominant fault. Calculate the deviation between the actual values ​​of the key operating parameter combination and the normal operating threshold, and quantify the degree of deviation from the normal state; Based on the degree of deviation, the corresponding mapping relationship between the inverter and the photovoltaic module string is retrieved. According to the inverter number to which the fault belongs, the photovoltaic module string connected to the inverter is located, and the location range of the module string is determined.

8. The data-driven photovoltaic power plant operation and maintenance fault diagnosis method according to claim 1, characterized in that, The process of matching the location range of the component strings with the layout coordinates of the photovoltaic power station, and outputting spatial fault location results including the time of fault occurrence, fault type, and affected component string number, includes: Retrieve the physical layout coordinate data of the photovoltaic power station, match the position range of the component string with the physical layout coordinate data, and determine the actual spatial coordinates corresponding to the component string; Based on the actual spatial coordinates, extract the time information corresponding to the potential anomaly window as the time of fault occurrence, and integrate the time of fault occurrence, fault category label and affected component string number; The integrated fault occurrence time, fault category label, and affected component string number are used to generate spatial fault location results according to a preset format.

9. A data-driven photovoltaic power plant operation and maintenance fault diagnosis system, characterized in that, include: Data acquisition and alignment module: acquires the DC side voltage sequence, current sequence, AC side power and internal temperature sequence of the inverter, as well as the corresponding ambient irradiance sequence and component surface temperature sequence, and aligns them according to timestamps to generate a multi-dimensional time series dataset; Feature dimensionality reduction and extraction module: The multidimensional time series dataset is truncated using a preset fixed-length sliding window to obtain overlapping time series segments, and principal components with a cumulative variance contribution rate reaching a preset ratio are extracted from each overlapping time series segment to form a low-dimensional feature vector for each overlapping time series segment; Trajectory sequence construction module: Connects adjacent low-dimensional feature vectors to form an evolutionary trajectory sequence, and calculates the vector distance value to obtain the change intensity sequence; Anomaly window marking module: Based on the distance change amplitude of the change intensity sequence and the curvature of the evolution trajectory sequence, it identifies anomalous turning points and marks potential anomalous windows; Fault category determination module: Input the low-dimensional feature vector of the potential abnormal window into the pre-trained classification model to determine the current fault category label; Fault location analysis module: Combining the fault category labels and the contribution weights of the corresponding low-dimensional feature vectors, it identifies the key operating parameter combinations and deviations that dominate the fault, and locates the position range of the module string based on the mapping relationship between the inverter and the photovoltaic module string. The result output matching module matches the location range of the component string with the layout coordinates of the photovoltaic power station, and outputs spatial fault location results including the time of fault occurrence, fault type, and affected component string number.

Images