A distributed photovoltaic power station fault diagnosis method based on big data analysis

Abstract

This invention discloses a fault diagnosis method for distributed photovoltaic (PV) power plants based on big data analytics, comprising: collecting multi-source operation monitoring data of the PV power plant and preprocessing it to obtain a unified time-series dataset; dividing the data into groups to form a sequence of grouped variable inputs; constructing an improved AutoFormer model and outputting a diagnostic result set; generating anomaly scores to form a set of candidate driving variables; obtaining sorted and located data group results based on cross-convergence mapping; fusing the diagnostic result set to generate fault type and fault location results; and generating operation and maintenance decisions based on the fault type and fault location. This invention achieves executable operation and maintenance decisions for distributed PV power plant fault diagnosis by combining the improved AutoFormer model with slot-level causal localization using cross-convergence mapping.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a method for fault diagnosis of distributed photovoltaic power plants based on big data analytics. Background Technology

[0002] Distributed photovoltaic (PV) power plants typically consist of multiple inverters, strings, combiners, and grid connection points. They are widely distributed, have a large number of devices, and experience frequent changes in operating status. During operation, they continuously generate multi-source monitoring data, including inverter power, string current, DC voltage and current, grid connection point voltage and frequency, irradiance, and module temperature. Existing operation and maintenance platforms often rely on built-in device alarms, fixed threshold rules, single curve comparisons, or simple statistical analysis to identify anomalies, frequently using an electrical quantity exceeding limits or power deviation as a trigger condition. Due to issues such as inconsistent sampling periods, timestamp offsets, and communication jitter leading to missing values ​​and outliers, aligning different data sources on the same timeline is difficult, resulting in inconsistent representations of the same event across different data sources. PV output is strongly affected by irradiance, temperature, and short-term shading, causing significant differences in output under different weather conditions, seasons, and operating conditions. Diagnostic features constructed directly using raw power or electrical quantities fluctuate with operating conditions, exhibiting feature drift, difficulty in unifying thresholds, and poor cross-site portability, thus affecting the stability and consistency of fault diagnosis.

[0003] In large-scale distributed scenarios, the number of sites is large and dispersed, with significant differences in equipment models, capacity configurations, grid connection conditions, and maintenance levels. Fault patterns include not only string mismatch, wiring abnormalities, inverter derating and shutdown, but also various types of problems such as grid-side disturbances, sensor and communication anomalies. Existing fault diagnosis methods generally suffer from high false alarm and false negative rates, difficulty in distinguishing fault types, and coarse-grained fault location. Threshold- or rule-based methods cannot simultaneously consider different sites and different operating conditions, and are prone to misjudging environmental changes as faults or missing progressive performance degradation. Although some data-driven methods use prediction or anomaly detection models to output anomaly scores, they often treat the model output as a single result, lacking a structured separation of irradiation-driven factors, temperature thermal effects, and equipment health deviations, and also lacking a verifiable localization process for the driving relationships between multiple source variables, making it difficult to accurately attribute anomalies to the DC side, inverter side, or grid-connected side.

[0004] Therefore, how to provide a fault diagnosis method for distributed photovoltaic power plants based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a fault diagnosis method for distributed photovoltaic power plants based on big data analysis. This invention comprehensively utilizes unified time-series processing and grouping fusion of multi-source operation monitoring data, innovative improved AutoFormer model structure, and a causal localization method based on cross-convergence mapping. It details the process from acquiring and preprocessing multi-source operation monitoring data, constructing input sequences grouped by physical source, extracting periodic dependencies and anomaly alignment representations based on a physically consistent three-part decomposition module and a fault-sensitive autocorrelation block module, implementing slotted encoding through a structured bottleneck module, forming a set of candidate driving group variables based on anomaly scores, and using cross-mapping of group variables and topological neighborhood mapping to obtain... This invention provides a complete process from obtaining causal strength ranking and location data sets to generating fault type discrimination, fault location, and operation and maintenance decision output. In terms of model structure, it innovatively achieves a three-part decomposition of irradiation-driven components, temperature thermal effect components, and equipment health offset components, as well as fault-oriented anomaly alignment representation extraction and slot-level traceable location. Furthermore, in causal analysis, it implements group-level cross-mapping and adjacency graph connected neighborhood mapping calculations, thereby reducing false alarms and missed alarms under strong environmental disturbances, improving fault type discrimination and location granularity, and outputting executable operation and maintenance decisions such as risk level, confidence level, and verification suggestions. It possesses advantages such as strong stability, adaptability to large-scale distributed scenarios, and ease of engineering implementation.

[0006] A method for fault diagnosis of a distributed photovoltaic power station based on big data analysis according to an embodiment of the present invention includes: Collect multi-source operation monitoring data of distributed photovoltaic power stations, preprocess the multi-source operation monitoring data, and obtain a unified time series dataset; The unified time series dataset is divided into data groups according to its physical source. The units of measurement are unified and features are constructed for each data group to form a grouped variable input sequence. An improved AutoFormer model is constructed, which decomposes the input sequence of grouped variables based on the physical consistency three-decomposition module, introduces the fault-sensitive autocorrelation block module for period dependency extraction and anomaly alignment characterization extraction, and uses the structured bottleneck module for slotted encoding to output a diagnostic result set. Anomaly scores are generated based on the diagnostic result set, and bottleneck slot characterization sequences are extracted within the anomaly time window to form a set of candidate driving group variables. Based on cross-convergence mapping, group-level cross-mapping is performed on the target sequences of the candidate driver group variable set and the diagnostic result set through group variable cross-mapping. A topological neighborhood graph is constructed to build an adjacency graph and cross-mapping calculation is performed in the connected neighborhood of the adjacency graph to obtain the causal strength ranking results and the location data group results. By integrating the diagnostic results set with the causal strength ranking results and the location data set results, fault type discrimination results and fault location results are generated. Based on the fault type identification results and fault location results, an operation and maintenance decision is generated.

[0007] Optionally, the multi-source operation monitoring data includes inverter power data, string current data, DC voltage and current data, grid connection point voltage and frequency data, irradiance and component temperature data.

[0008] Optionally, obtaining the unified time-series dataset includes: Collect multi-source operation monitoring data of distributed photovoltaic power stations, and record the corresponding data source identifier, equipment identifier and collection timestamp for each data collection to form the original multi-source operation monitoring dataset; The original multi-source operation monitoring dataset is unified in terms of timestamps and sampling period. The collection timestamps of each data source are mapped to discrete time points on a unified time axis. The data from each data source is resampled according to a preset sampling period and aligned to the unified time axis. The system performs missing data filling, abnormal data removal, and operating condition marking on the aligned data. It also performs interpolation to fill in missing time points and removes data that exceeds the preset physical range threshold. The system generates operating condition labels that identify nighttime shutdown, limited power generation, maintenance shutdown, and communication interruption, and associates them with the corresponding time point data.

[0009] Optionally, forming the grouped variable input sequence includes: The data fields in the unified time series dataset are grouped according to the physical source rules, and are divided into environmental data group, DC side data group, inverter data group, and grid-connected side data group respectively. Perform dimensional uniformity processing on each data group, perform unit conversion on each data field according to the unit conversion relationship corresponding to the field, and perform numerical standardization processing on each data field by subtracting the mean from each sample value of the field and then dividing it by the standard deviation, using the mean and standard deviation of the current field in the statistical window as parameters. Features are constructed for each data group after dimensional unification processing to form a grouped variable input sequence. For each data field, the moving mean, moving variance, moving maximum and moving minimum are calculated within the time window. The string current consistency characteristics are calculated for the DC side data group, the power change rate characteristics are calculated for the inverter data group, and the voltage fluctuation amplitude characteristics are calculated for the grid-connected data group. This forms an environmental variable input sequence, a DC side variable input sequence, an inverter variable input sequence, and a grid-connected variable input sequence arranged along a unified time axis.

[0010] Optionally, the output diagnostic result set includes: An improved AutoFormer model is constructed, which includes a physically consistent three-part decomposition module, a fault-sensitive autocorrelation block module, and a structured bottleneck module. The physical consistency three-decomposition module performs a layer-by-layer three-decomposition process on the input sequence of grouped variables. The moving average sequence of the input sequence of grouped variables is calculated as the trend base sequence. The difference between the input sequence of grouped variables and the trend base sequence is used as the seasonal base sequence. The trend base sequence is mapped to the irradiation-driven component sequence and the temperature thermal effect component sequence. The remaining component after deducting the irradiation-driven component sequence and the temperature thermal effect component sequence from the input sequence of grouped variables is determined as the equipment health offset component sequence. Based on the fault-sensitive autocorrelation block module, the seasonal base sequence is subjected to periodic dependency representation extraction processing. The autocorrelation sequence of the seasonal base sequence is calculated, the set of delay quantities is selected from the autocorrelation sequence, and each delay quantity is aligned with the time delay of the seasonal base sequence and aggregated to obtain the periodic dependency representation sequence. Anomaly alignment characterization extraction processing is performed on the equipment health offset component sequence, the autocorrelation sequence of the equipment health offset component sequence is calculated, a set of delay quantities is selected from the autocorrelation sequence, and time delay alignment is performed on each delay quantity of the equipment health offset component sequence. The result is obtained by aggregation. A structured bottleneck module is introduced to perform slotting encoding processing on the environmental variable input sequence, DC side multivariate input sequence, inverter multivariate input sequence and grid-connected multivariate input sequence respectively, to obtain the environmental slotting characterization sequence, DC slotting characterization sequence, inverter slotting characterization sequence and grid-connected slotting characterization sequence, and generate a diagnostic result set, which includes the target sequence prediction result, health offset sequence, fault sensitivity characterization sequence and structured bottleneck slotting characterization sequence.

[0011] Optionally, forming the candidate driver group variable set includes: The health offset sequence and the fault-sensitive characterization sequence are obtained from the diagnostic result set. The sequence segments are obtained by sliding segmentation according to the window length on a unified time axis. An anomaly score is calculated for each sequence segment. The anomaly score includes the amplitude characteristics, persistence characteristics and fluctuation characteristics of the health offset sequence segment and the amplitude characteristics, persistence characteristics and fluctuation characteristics of the fault-sensitive characterization sequence segment. The anomaly score is compared with the anomaly threshold, and the anomaly trigger is determined by combining the working condition mark. The start time point and the end time point corresponding to the sequence segment that meets the anomaly trigger are determined as the anomaly time window. Based on the abnormal time window, the abnormal object level and abnormal object identifier are determined. The structured bottleneck slot characterization sequence is extracted within the abnormal time window. Within the data range corresponding to the abnormal object identifier, the environmental data group sequence, DC side data group sequence, inverter data group sequence and grid-connected side data group sequence, which are time-aligned with the structured bottleneck slot characterization sequence, are extracted to form a candidate drive group variable set.

[0012] Optionally, obtaining the causal intensity ranking result and the location data group result includes: Determine the target sequence and candidate driver group variable set for the cross-convergence mapping. The target sequence is selected from the health offset sequence, and the candidate driver group variable set includes the environmental data group sequence, the DC side data group sequence, the inverter data group sequence, and the grid-connected side data group sequence. Based on topological neighborhood graph construction, adjacency graphs are constructed on the reconstructed manifolds of each data group sequence in the target sequence and the candidate driving group variable set. The reconstructed manifold construction includes constructing each sequence into a delayed vector sequence according to the embedding dimension and the delay step size, and using each vector point of the delayed vector sequence to form the reconstructed manifold. The adjacency graph construction includes using each vector point in the reconstructed manifold as a graph node and establishing edge connections based on the distance between nodes to form adjacency relationships. Cross-mapping of group variables is performed in the connected neighborhood of the adjacency graph obtained by topological neighborhood construction. The set of connected neighborhood nodes corresponding to each node in the manifold reconstructed by the target sequence is used to determine the sample set of driving group variables. Based on the sample set of driving group variables, estimated sequences are generated for each variable sequence of the driving group variables. The cross-mapping calculation includes calculating the correlation coefficient between the estimated sequence and the corresponding real sequence as the cross-mapping capability index. The cross-mapping capability index of each variable sequence in the same data group is summarized to obtain the cross-mapping capability index of the data group. The cross-mapping capability index of each data group is sorted to generate a causal strength ranking result for the data groups. Based on the ranking result, the target localization data group with the highest causal strength is determined, and the localization data group result is output.

[0013] Optionally, the generation of fault type discrimination results and fault location results includes: Obtain the diagnostic result set and the causal strength ranking result of the data group, locate the data group result, and extract the health offset sequence, fault sensitivity characterization sequence and structured bottleneck groove characterization sequence according to the abnormal time window to obtain the diagnostic feature set within the window; Based on the location data set results, the location slot type is determined. The statistical features within the slot are extracted from the structured bottleneck slot characterization sequence corresponding to the location slot type. The key variable sequence aligned with the abnormal time window is extracted from the data set sequence corresponding to the location slot type. The system integrates diagnostic feature sets within the fusion window, location slot type, in-slot statistical features, and key variable sequences to generate fault type discrimination results, and generates fault location results based on the abnormal object level and location slot type.

[0014] Optionally, the generation of operation and maintenance decisions includes: Receive fault type identification results and fault location results, and determine the operation and maintenance object identifier and operation and maintenance object level based on the fault location results. The operation and maintenance object level includes the power plant level, inverter level and string level. Within the abnormal time window, offset amplitude and offset persistence indicators are generated based on the healthy offset sequence. The risk level and confidence level are generated by integrating the causal strength ranking results of the data group and the anomaly score. Based on the fault type identification results, location slot type, and maintenance object hierarchy, suggested inspection objects and suggested inspection data items are generated. Based on the risk level and confidence level, the handling priority is generated. The output includes the maintenance decision output containing the risk level, confidence level, suggested inspection objects, suggested inspection data items, and handling priority.

[0015] The beneficial effects of this invention are: This invention improves the stability and usability of fault diagnosis features under complex operating conditions by performing unified time-series preprocessing, grouping and modeling by physical source, and feature construction on multi-source operation monitoring data of distributed photovoltaic power plants. Compared with existing methods that rely on single variable thresholds or alarm rules, this invention adopts an improved AutoFormer model that introduces a physically consistent three-part decomposition module, a fault-sensitive autocorrelation block module, and a structured bottleneck module in its structure. This achieves the separation of irradiation-driven components, temperature thermal effect components, and equipment health offset components, extracts periodic dependence and anomaly alignment representations, and forms slotted coding output. As a result, it can reduce false alarms and missed alarms and improve the consistency and robustness of anomaly detection even under strong irradiation and temperature disturbances, data loss, and noise.

[0016] This invention further employs cross-convergence mapping to perform group-level cross-mapping between the candidate driver group variable set and the target sequence. Through topological neighborhood construction, cross-mapping calculations are performed within the connected neighborhoods of the adjacency graph to obtain causal strength ranking and location data group results. This enables group-level attribution and location of data for the environment, DC side, inverter, and grid-connected side. Compared to existing solutions that only output anomaly scores, have coarse location granularity, and struggle to form a closed-loop operation and maintenance system, this invention can integrate diagnostic result sets and causal evidence to generate fault type discrimination and fault location results. It outputs operation and maintenance decisions including risk level, confidence level, suggested inspection objects, suggested inspection data items, and handling priorities, improving fault type differentiation capabilities, location accuracy, and engineering feasibility. This makes it suitable for large-scale, refined, and intelligent operation and maintenance of large-scale distributed power plants. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a fault diagnosis method for distributed photovoltaic power plants based on big data analysis proposed in this invention; Figure 2 This is a structural block diagram of the improved AutoFormer model for a fault diagnosis method for distributed photovoltaic power plants based on big data analysis proposed in this invention. Figure 3 This is a functional diagram illustrating the cross-convergence mapping of a fault diagnosis method for distributed photovoltaic power plants based on big data analysis proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A fault diagnosis method for distributed photovoltaic power plants based on big data analysis includes: Collect multi-source operation monitoring data of distributed photovoltaic power stations, preprocess the multi-source operation monitoring data, and obtain a unified time series dataset; The unified time series dataset is divided into data groups according to its physical source. The units of measurement are unified and features are constructed for each data group to form a grouped variable input sequence. An improved AutoFormer model is constructed, which decomposes the input sequence of grouped variables based on the physical consistency three-decomposition module, introduces the fault-sensitive autocorrelation block module for period dependency extraction and anomaly alignment characterization extraction, and uses the structured bottleneck module for slotted encoding to output a diagnostic result set. Anomaly scores are generated based on the diagnostic result set, and bottleneck slot characterization sequences are extracted within the anomaly time window to form a set of candidate driving group variables. Based on cross-convergence mapping, group-level cross-mapping is performed on the target sequences of the candidate driver group variable set and the diagnostic result set through group variable cross-mapping. A topological neighborhood graph is constructed to build an adjacency graph and cross-mapping calculation is performed in the connected neighborhood of the adjacency graph to obtain the causal strength ranking results and the location data group results. By integrating the diagnostic results set with the causal strength ranking results and the location data set results, fault type discrimination results and fault location results are generated. Based on the fault type identification results and fault location results, an operation and maintenance decision is generated.

[0020] In this embodiment, the multi-source operation monitoring data includes inverter power data, string current data, DC voltage and current data, grid connection point voltage and frequency data, irradiance and component temperature data.

[0021] In this embodiment, obtaining the unified time-series dataset includes: Collect multi-source operation monitoring data of distributed photovoltaic power stations, and record the corresponding data source identifier, equipment identifier and collection timestamp for each data collection to form the original multi-source operation monitoring dataset; The original multi-source operation monitoring dataset is unified in terms of timestamps and sampling periods. The collection timestamps of each data source are mapped to discrete time points on a unified time axis. The data from each data source is resampled according to a preset sampling period and aligned to the unified time axis. Specifically, the resampling and alignment of the data from each data source to the unified time axis is performed as follows: A unified time axis is established with a preset sampling period of 300 seconds. Starting from the start of the diagnostic period, a discrete time point is generated every 300 seconds. Each data source is first sorted in ascending order by the collection timestamp. Then, for each discrete time point on the unified time axis, samples whose timestamps fall between 150 seconds before and 150 seconds after the discrete time point are selected from the current data source as the candidate sample set for the current time point. If the candidate sample set is not empty, the arithmetic mean of all values ​​in the candidate sample set is taken as the resampled value of the data source at the discrete time point. If the candidate sample set is empty, the discrete time point is marked as missing. When there are multiple records in the same discrete time point window, they are aggregated by arithmetic mean. When the time interval between two adjacent original records is greater than 900 seconds, the corresponding discrete time point in the time period is kept marked as missing and no interpolation is performed. After resampling all data sources, the resampled values ​​of each data source are merged into the same row of records using the discrete time point of the unified time axis as the index, so as to realize resampling according to the preset sampling period and alignment to the unified time axis. The aligned data undergoes missing data imputation, outlier data removal, and operational condition labeling. Data with missing time points is imputed via interpolation. Data exceeding a preset physical range threshold is removed. Operational condition labels are generated to identify nighttime shutdown, power restriction, maintenance shutdown, and communication interruption, and these labels are associated with the corresponding time point data. Specifically, the removal of data exceeding the preset physical range threshold involves: For each time point after alignment, a physical range check is performed on each variable. The following conditions are considered out of range: irradiance less than 0 or greater than 1400 watts per square meter; component temperature less than -40 degrees Celsius or greater than 90 degrees Celsius; DC voltage less than 0 or greater than 1500 volts; DC current less than 0 or greater than 1.5 times the rated current; string current less than 0 or greater than 20 amps; grid connection point voltage less than 0.8 times the rated voltage or greater than 1.2 times the rated voltage; grid connection point frequency less than 49 Hz or greater than 51 Hz; and inverter power less than 0 or greater than 1.1 times the rated power. Out-of-range samples are removed by setting the sample value to null and writing an out-of-range flag. If the number of null variables at the same time point for the same device exceeds 50% of the total number of variables for the device, all variables for the device at that time point are set to null and marked as a candidate time point for communication interruption.

[0022] In this embodiment, forming the grouped variable input sequence includes: The data fields in the unified time-series dataset are grouped according to the physical source rule, into environmental data group, DC side data group, inverter data group, and grid-connected side data group, respectively. The physical source rule is as follows: Data fields are grouped based on the component to which the corresponding acquisition point belongs. A field mapping table is established for each data field. The mapping table includes two items: acquisition point type and field type. When grouping according to the mapping table, acquisition point types of weather stations or component temperature probes and field types of irradiance or temperature are assigned to the environmental data group. Acquisition point types of string acquisition terminals, combiner boxes or DC input terminals and field types of DC voltage, DC current or string current are assigned to the DC side data group. Acquisition point types of inverter bodies and field types of inverter power, efficiency, internal temperature or alarm statistics are assigned to the inverter data group. Acquisition point types of grid-connected meters or grid-connected side measuring devices and field types of grid-connected voltage or grid-connected frequency are assigned to the grid-connected side data group. When the same field meets two types of grouping conditions at the same time, a unique group is determined according to the priority of grid-connected side, inverter, DC side, and environment, resulting in environmental data group, DC side data group, inverter data group and grid-connected side data group. Perform dimensional uniformity processing on each data group, perform unit conversion on each data field according to the unit conversion relationship corresponding to the field, and perform numerical standardization processing on each data field by subtracting the mean from each sample value of the field and then dividing it by the standard deviation, using the mean and standard deviation of the current field in the statistical window as parameters. Features are constructed for each data group after dimensional unification processing to form a grouped variable input sequence. For each data field, the moving mean, moving variance, moving maximum and moving minimum are calculated within the time window. The string current consistency characteristics are calculated for the DC side data group, the power change rate characteristics are calculated for the inverter data group, and the voltage fluctuation amplitude characteristics are calculated for the grid-connected data group. This forms an environmental variable input sequence, a DC side variable input sequence, an inverter variable input sequence, and a grid-connected variable input sequence arranged along a unified time axis.

[0023] In this embodiment, the output diagnostic result set includes: An improved AutoFormer model is constructed, comprising a physically consistent three-part decomposition module, a fault-sensitive autocorrelation block module, and a structured bottleneck module. Specifically, the construction of the improved AutoFormer model involves: The improved AutoFormer model is built according to the encoder-decoder structure. In each layer, the physical consistency tri-decomposition module and the fault-sensitive autocorrelation block module are connected in series. The structured bottleneck module is connected at the encoder output. The physical consistency tri-decomposition module uses a moving average window of length 25 to calculate the trend base sequence. The seasonal base sequence is obtained by subtracting the trend base sequence from the input sequence. The trend base sequence is decomposed into the irradiation-driven component sequence and the temperature thermal effect component sequence. The remaining components are used as the equipment health offset component sequence. The fault-sensitive autocorrelation block module calculates the autocorrelation sequence for the seasonal base sequence and the equipment health offset component sequence, selects the three delay values ​​with the largest autocorrelation values ​​for time delay alignment and aggregation, and obtains the period-dependent representation sequence and the anomaly-aligned representation sequence. The structured bottleneck module divides the encoder implicit representation into four slots and outputs the representation sequences of the environmental slot, DC slot, inverter slot and grid-connected slot respectively. At the same time, it outputs the fused representation for the decoder to generate the diagnostic result set. The physical consistency three-part decomposition module performs a layer-by-layer three-part decomposition process on the grouped variable input sequence. A moving average sequence is calculated for the grouped variable input sequence as the trend base sequence. The difference between the grouped variable input sequence and the trend base sequence is used as the seasonal base sequence. The trend base sequence is mapped to an irradiation-driven component sequence and a temperature thermal effect component sequence. The remaining components after subtracting the irradiation-driven component sequence and the temperature thermal effect component sequence from the grouped variable input sequence are determined as the equipment health offset component sequence. Where: The process involves performing a layer-by-layer three-stage decomposition: In each layer of the improved AutoFormer model, the same decomposition process is repeated on the current layer's input feature sequence. A moving average is calculated using a sliding window with a length of 25 sampling points to obtain the trend base sequence. The sampling period is 300 seconds, and the window center is aligned with the current time. The trend base sequence is subtracted point by point from the current layer's input feature sequence to obtain the seasonal base sequence. After obtaining the irradiation-driven component sequence and the temperature thermal effect component sequence, the irradiation-driven component sequence and the temperature thermal effect component sequence are subtracted point by point from the layer's input feature sequence to obtain the equipment health offset component sequence. The seasonal base sequence and the equipment health offset component sequence are used as the current layer's output and input into the next layer to continue the decomposition process until the last layer obtains the final decomposition result. The trend base sequence is mapped to an irradiation-driven component sequence and a temperature thermal effect component sequence. Specifically, the most recent 288 sampling points are taken at each moment to form a regression window. The regression window corresponds to 24 hours and the sampling period is 300 seconds. Within the regression window, a linear regression is established with the trend base sequence as the dependent variable and the irradiance sequence and component temperature sequence as independent variables. The regression coefficients are solved using the least squares method. The irradiation-driven component sequence is taken as the product of the irradiance sequence and the irradiance regression coefficient. The temperature thermal effect component sequence is taken as the product of the component temperature sequence and the temperature regression coefficient. The remaining terms after deducting the irradiation-driven component sequence and the temperature thermal effect component sequence from the trend base sequence are incorporated into the equipment health offset component sequence. The seasonal base sequence is processed by extracting periodic dependency representations based on the fault-sensitive autocorrelation block module. The autocorrelation sequence of the seasonal base sequence is calculated, and a set of delay values ​​is selected from the autocorrelation sequence. Each delay value is aligned with the time delay of the seasonal base sequence and then aggregated to obtain the periodic dependency representation sequence. Specifically, the calculation of the autocorrelation sequence of the seasonal base sequence is as follows: The seasonal base sequence is processed to remove the mean and the sample mean and sample variance are recorded. The autocorrelation value is calculated sequentially according to the delay amount from 1 sampling point to 288 sampling points. The delay amount of 288 corresponds to 24 hours and the sampling period is 300 seconds. For any delay amount, all paired samples in the interval that satisfy the condition that there are valid values ​​at both the current time and the delayed time are taken. The product of the paired samples is calculated and divided by the number of paired samples to obtain the covariance estimate. The covariance estimate is divided by the sample variance to obtain the autocorrelation value of the delay amount. Anomaly alignment characterization extraction processing is performed on the equipment health offset component sequence, the autocorrelation sequence of the equipment health offset component sequence is calculated, a set of delay quantities is selected from the autocorrelation sequence, and time delay alignment is performed on each delay quantity of the equipment health offset component sequence. The result is obtained by aggregation. A structured bottleneck module is introduced to perform slotting encoding processing on the environmental variable input sequence, DC-side multivariate input sequence, inverter multivariate input sequence, and grid-connected multivariate input sequence, respectively, to obtain environmental slotting characterization sequences, DC slotting characterization sequences, inverter slotting characterization sequences, and grid-connected slotting characterization sequences. A diagnostic result set is generated, which includes the target sequence prediction result, health offset sequence, fault-sensitive characterization sequence, and structured bottleneck slotting characterization sequence. Specifically, the slotting encoding processing is performed as follows: At each time point, the four sets of input sequences are mapped to slot vectors of fixed dimensions while maintaining time alignment. The environmental variable input sequence, DC side multivariate input sequence, inverter multivariate input sequence and grid-connected side multivariate input sequence are respectively input into their respective linear mapping layers. The multivariate feature vectors of each set at time point are multiplied by the corresponding weight matrix and a bias is added to obtain the slot vector. The four slot vectors have the same dimension and are all set to one-quarter of the model's hidden dimension. For each set of input feature vectors at a given time point, amplitude clipping is first performed. The clipping threshold is the mean of the current feature within the training statistical window plus or minus three standard deviations. Vectors exceeding the upper limit are set to the upper limit, and those below the lower limit are set to the lower limit. After clipping, linear mapping is performed, and the four slot vectors are concatenated in a fixed order to obtain a fused representation, which is then input into the decoder to generate the target sequence prediction result. The equipment health offset component sequence is output as a health offset sequence along the time axis, and the anomaly alignment representation sequence is output as a fault-sensitive representation sequence along the time axis. The four slot vectors are combined along the time axis to form an environmental slot representation sequence, a DC slot representation sequence, an inverter slot representation sequence, and a grid-connected slot representation sequence, which together constitute the diagnostic result set.

[0024] In this embodiment, forming the candidate driver group variable set includes: Health-biased sequences and fault-sensitive characterization sequences are obtained from the diagnostic results set. These sequences are then segmented along a unified time axis using a sliding window length to obtain sequence fragments. An anomaly score is calculated for each sequence fragment. The anomaly score includes the amplitude, persistence, and fluctuation characteristics of the health-biased sequence fragment, as well as the amplitude, persistence, and fluctuation characteristics of the fault-sensitive characterization sequence fragment. The calculation of the anomaly score specifically involves: A sliding segmentation with a window length of 24 sampling points is performed on a unified time axis. The sampling period is 300 seconds and the sliding step is 1 sampling point. For each window, amplitude features, persistence features, and fluctuation features are calculated for the healthy offset sequence segment and the fault sensitive characterization sequence segment respectively. The amplitude feature is the maximum absolute value of each sample value in the segment. The persistence feature is the number of time points in the segment where the absolute value of the sample value is continuously greater than 0.2 and then divided by 24 to obtain the persistence ratio. The fluctuation feature is the maximum absolute value of the difference between adjacent sample values ​​in the segment. The three types of features of the healthy offset sequence segment and the three types of features of the fault sensitive characterization sequence segment are summed to obtain the sum of six features. The sum of the six features is used as the anomaly score of the window. Anomaly scores are compared with anomaly thresholds, and anomaly triggering is determined by combining operating condition markers. The start and end times of the sequence segments that meet the anomaly trigger criteria are defined as the anomaly time window. The anomaly threshold is specifically defined as follows: A threshold statistical set is selected for a continuous 30-day period where the operating condition marker does not belong to nighttime shutdown, limited power generation, maintenance shutdown, or communication interruption. An abnormal score for each window is calculated within the statistical set to form an abnormal score sample set. The mean and standard deviation of the abnormal score sample set are calculated, and the abnormal threshold is set to the mean plus three times the standard deviation. At the same time, the lower limit of the abnormal threshold is set to 1. When the mean plus three times the standard deviation is less than 1, 1 is taken as the abnormal threshold. During operation, when the abnormal score of a window is greater than or equal to the abnormal threshold and the operating condition marker at the corresponding time point does not belong to nighttime shutdown, maintenance shutdown, or communication interruption, it is determined to be an abnormal trigger. Based on the abnormal time window, the abnormal object level and abnormal object identifier are determined. The structured bottleneck slot characterization sequence is extracted within the abnormal time window. Within the data range corresponding to the abnormal object identifier, the environmental data group sequence, DC side data group sequence, inverter data group sequence and grid-connected side data group sequence, which are time-aligned with the structured bottleneck slot characterization sequence, are extracted to form a candidate drive group variable set.

[0025] In this embodiment, obtaining the causal strength ranking result and the location data group result includes: Determine the target sequence and candidate driver group variable set for the cross-convergence mapping. The target sequence is selected from the health offset sequence, and the candidate driver group variable set includes the environmental data group sequence, the DC side data group sequence, the inverter data group sequence, and the grid-connected side data group sequence. Based on topological neighborhood graph construction, adjacency graphs are constructed on the reconstructed manifolds of each data set sequence in the target sequence and the candidate driving group variable set. The reconstructed manifold construction includes constructing delayed vector sequences for each sequence according to the embedding dimension and delay step size, and using the vector points of the delayed vector sequences to form the reconstructed manifold. The adjacency graph construction includes using each vector point in the reconstructed manifold as a graph node and establishing edge connections based on the distance between nodes to form adjacency relationships. Specifically, the topological neighborhood graph construction is as follows: The topological neighborhood graph is formed through manifold nodeization, distance calculation, nearest neighbor edge connection, and threshold truncation. Each delayed vector point in the reconstructed manifold is taken as a graph node. The Euclidean distance between any two nodes is calculated. For each node, the 10 nodes with the smallest distance from the remaining nodes are selected as nearest neighbors and undirected edges are established. At the same time, the distance threshold is set to 2. Nearest neighbors with a distance greater than 2 are not connected by edges. After the nearest neighbor selection and edge connection are completed for all nodes, the topological neighborhood graph is obtained. Cross-mapping of group variables is performed in the connected neighborhood of the adjacency graph obtained by topological neighborhood construction. The set of connected neighborhood nodes corresponding to each node in the manifold reconstructed by the target sequence is used to determine the sample set of driving group variables. Based on the sample set of driving group variables, estimated sequences are generated for each variable sequence of the driving group variables. The cross-mapping calculation includes calculating the correlation coefficient between the estimated sequence and the corresponding real sequence as the cross-mapping capability index. The cross-mapping capability index of each variable sequence in the same data group is summarized to obtain the cross-mapping capability index of the data group. The cross-mapping capability index of each data group is sorted to generate a causal strength ranking result for the data groups. Based on the ranking result, the target localization data group with the highest causal strength is determined, and the localization data group result is output.

[0026] In this embodiment, the generation of fault type discrimination results and fault location results includes: Obtain the diagnostic result set and the causal strength ranking result of the data group, locate the data group result, and extract the health offset sequence, fault sensitivity characterization sequence and structured bottleneck groove characterization sequence according to the abnormal time window to obtain the diagnostic feature set within the window; Based on the location data set results, the location slot type is determined. The statistical features within the slot are extracted from the structured bottleneck slot characterization sequence corresponding to the location slot type. The key variable sequence aligned with the abnormal time window is extracted from the data set sequence corresponding to the location slot type. The system integrates diagnostic feature sets within the fusion window, location slot type, in-slot statistical features, and key variable sequences to generate fault type discrimination results. Furthermore, it generates fault location results based on the anomaly object hierarchy and location slot type. The fault type discrimination results are generated as follows: within the abnormal time window, according to the location slot type branch, the health offset amplitude, health offset duration ratio and fault sensitivity fluctuation are calculated. The amplitude is the maximum absolute value, the duration ratio is the number of time points with an absolute value greater than 0.2 divided by the window length, and the fluctuation is the maximum absolute value of adjacent differences. DC slots with a duration ratio greater than 0.5 and string current consistency less than 0.8 are judged as DC side faults. Inverter slots with an amplitude greater than 0.2 and power change rate greater than 0.3 are judged as inverter faults. Grid-connected slots with voltage fluctuation amplitude greater than 0.05 of rated voltage are judged as grid-connected side faults. Environmental slots with an amplitude less than 0.1 and fluctuation less than 0.10 are judged as environmental changes. The fault location result is generated as follows: output the fault object identifier at the abnormal object level, the power station identifier at the power station level, the inverter identifier at the inverter level, and the string identifier at the string level. Determine the set of location variables according to the location slot type and calculate the deviation. The deviation is the maximum absolute deviation of the variable within the abnormal time window. Select the variable identifier with the largest deviation as the location variable output. Combine the fault object identifier, the location slot type, and the location variable identifier to form the fault location result.

[0027] In this embodiment, the generation of operation and maintenance decisions includes: Receive fault type identification results and fault location results, and determine the operation and maintenance object identifier and operation and maintenance object level based on the fault location results. The operation and maintenance object level includes the power plant level, inverter level and string level. Within the abnormal time window, offset amplitude and offset persistence indicators are generated based on the healthy offset sequence. The risk level and confidence level are generated by integrating the causal strength ranking results of the data group and the anomaly score. Based on the fault type identification results, location slot type, and maintenance object hierarchy, suggested inspection objects and suggested inspection data items are generated. Based on the risk level and confidence level, the handling priority is generated. The output includes the maintenance decision output containing the risk level, confidence level, suggested inspection objects, suggested inspection data items, and handling priority.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a distributed photovoltaic large-scale operation and maintenance scenario. The platform connects to multiple power stations, hundreds of inverters, and thousands of string channels. Monitoring data simultaneously comes from inverters, DC side acquisition, grid connection point metering, and meteorological temperature probes. Due to inconsistencies in sampling periods and communication links between different equipment manufacturers, some devices experience minute-level timestamp drift, and continuous data loss and sudden abnormal values ​​occur when the link jitter or the acquisition end restarts. At the same time, rapid changes in irradiance, lag in module temperature, and local shading cause strong disturbances and fluctuations in electrical quantities. Traditional threshold alarms tend to have concentrated false alarms when cloud shadows are frequent, and slow degradation problems are easily masked, resulting in difficulty in distinguishing types, coarse-grained location, and high operation and maintenance troubleshooting costs.

[0029] Inverter power, string current, DC voltage and current, grid connection point voltage and frequency, irradiance, and module temperature are incorporated into the diagnostic process. First, resampling alignment, missing data marking, out-of-limit removal, and operating condition marking are completed in a fixed 300-second time step. Then, the data is divided into environmental, DC, inverter, and grid connection groups according to their physical origin, forming a grouped input sequence. This grouped input sequence enters the improved AutoFormer model, which outputs irradiance-driven components, temperature-thermal effect components, and health offset components from a physically consistent tri-decomposition. Periodic dependence and anomaly alignment representations are extracted from fault-sensitive autocorrelation blocks, and four types of slot representations are output from structured bottlenecks to form a diagnostic result set. Anomaly scores are calculated for the window and triggered for localization. The health offset target sequence is cross-converged with the candidate group variable inputs. Through cross-mapping of group variables and topological neighborhood construction, causal strength ranking and localization data sets are output, generating fault type, object-level localization, and operation and maintenance decisions.

[0030] In a verification operation covering multiple sites, the system accessed approximately 480 inverter data channels and approximately 3600 string current channels. The data loss rate fluctuated between 0% and 18%. During this period, 187 abnormal events were manually reviewed and confirmed in a closed-loop manner, including string mismatch and obstruction, DC side wiring and acquisition anomalies, inverter derating and intermittent shutdowns, and grid-connected voltage fluctuations limiting power generation. This invention can still stably output healthy offsets on days with strong cloud shadow fluctuations and attribute abnormal drives to the environmental group or grid-connected group. Moreover, it can continuously provide offset accumulation and verification suggestions in degraded states, enabling operation and maintenance personnel to complete the handling with fewer curve verifications and fewer on-site confirmations, and achieving early detection of anomalies and reduction of power generation losses.

[0031] Table 1. Multidimensional Performance Comparison of Distributed Photovoltaic Fault Diagnosis Methods

[0032] As shown in Table 1, this invention achieves optimal performance in both the false positive rate and the false negative rate, with a false positive rate of 3.8% and a false negative rate of 4.9%. Compared to the rule-based threshold method, the false positive rate is reduced from 17.5% to 3.8%, and the false negative rate is reduced from 13.2% to 4.9%. It also maintains lower false positive and false negative rates compared to XGBoost, LSTM, Transformer, and the original AutoFormer, indicating more stable alarm triggering under conditions of strong environmental disturbances and data quality fluctuations.

[0033] In terms of diagnostic usability, this invention demonstrates superior advantages in type accuracy and positioning capability. The type accuracy reaches 88.4%, higher than the original AutoFormer's 82.1%, and also higher than XGBoost's 74.6%, LSTM's 77.3%, and Transformer's 79.8%. Regarding positioning, the inverter positioning accuracy of this invention is 91.2%, and the string positioning accuracy is 84.7%, both the highest among the compared methods. In particular, the string positioning accuracy relative to the rule threshold method is improved from 41.8% to 84.7%, demonstrating a significant improvement in fine-grained positioning.

[0034] From the perspective of engineering implementation metrics, this invention balances response speed and data robustness, with an alarm latency of 6.5 minutes, which is better than the 15 minutes of the rule-based threshold method and remains at a low level in the comparison model; the missing tolerance reaches 18%, which is higher than the comparison method, indicating that it is more tolerant to missing data and short-term interruptions; the operation and maintenance executability rate is 86.0%, which is significantly higher than the 52.0% of the rule-based threshold method and the 76.0% of the original AutoFormer, indicating that the output is more conducive to forming clear verification items and handling priorities, facilitating closed-loop operation and maintenance.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for distributed photovoltaic power plants based on big data analysis, characterized in that, include: Collect multi-source operation monitoring data of distributed photovoltaic power stations, preprocess the multi-source operation monitoring data, and obtain a unified time series dataset; The unified time series dataset is divided into data groups according to its physical source. The units of measurement are unified and features are constructed for each data group to form a grouped variable input sequence. An improved AutoFormer model is constructed, which decomposes the input sequence of grouped variables based on the physical consistency three-decomposition module, introduces the fault-sensitive autocorrelation block module for period dependency extraction and anomaly alignment characterization extraction, and uses the structured bottleneck module for slotted encoding to output a diagnostic result set. Anomaly scores are generated based on the diagnostic result set, and bottleneck slot characterization sequences are extracted within the anomaly time window to form a set of candidate driving group variables. Based on cross-convergence mapping, group-level cross-mapping is performed on the target sequences of the candidate driver group variable set and the diagnostic result set through group variable cross-mapping. A topological neighborhood graph is constructed to build an adjacency graph and cross-mapping calculation is performed in the connected neighborhood of the adjacency graph to obtain the causal strength ranking results and the location data group results. By integrating the diagnostic results set with the causal strength ranking results and the location data set results, fault type discrimination results and fault location results are generated. Based on the fault type identification results and fault location results, an operation and maintenance decision is generated.

2. The method for fault diagnosis of distributed photovoltaic power stations based on big data analysis according to claim 1, characterized in that, The multi-source operation monitoring data includes inverter power data, string current data, DC voltage and current data, grid connection point voltage and frequency data, irradiance and component temperature data.

3. The method for fault diagnosis of distributed photovoltaic power stations based on big data analysis according to claim 1, characterized in that, The obtained unified time series dataset includes: Collect multi-source operation monitoring data of distributed photovoltaic power stations, and record the corresponding data source identifier, equipment identifier and collection timestamp for each data collection to form the original multi-source operation monitoring dataset; The original multi-source operation monitoring dataset is unified in terms of timestamps and sampling period. The collection timestamps of each data source are mapped to discrete time points on a unified time axis. The data from each data source is resampled according to a preset sampling period and aligned to the unified time axis. The system performs missing data filling, abnormal data removal, and operating condition marking on the aligned data. It also performs interpolation to fill in missing time points and removes data that exceeds the preset physical range threshold. The system generates operating condition labels that identify nighttime shutdown, limited power generation, maintenance shutdown, and communication interruption, and associates them with the corresponding time point data.

4. The method for fault diagnosis of distributed photovoltaic power stations based on big data analysis according to claim 1, characterized in that, The process of forming the grouped variable input sequence includes: The data fields in the unified time series dataset are grouped according to the physical source rules, and are divided into environmental data group, DC side data group, inverter data group, and grid-connected side data group respectively. Perform dimensional uniformity processing on each data group, perform unit conversion on each data field according to the unit conversion relationship corresponding to the field, and perform numerical standardization processing on each data field by subtracting the mean from each sample value of the field and then dividing it by the standard deviation, using the mean and standard deviation of the current field in the statistical window as parameters. Features are constructed for each data group after dimensional unification processing to form a grouped variable input sequence. For each data field, the moving mean, moving variance, moving maximum and moving minimum are calculated within the time window. The string current consistency characteristics are calculated for the DC side data group, the power change rate characteristics are calculated for the inverter data group, and the voltage fluctuation amplitude characteristics are calculated for the grid-connected data group. This forms an environmental variable input sequence, a DC side variable input sequence, an inverter variable input sequence, and a grid-connected variable input sequence arranged along a unified time axis.

5. The method for fault diagnosis of distributed photovoltaic power stations based on big data analysis according to claim 1, characterized in that, The output diagnostic result set includes: An improved AutoFormer model is constructed, which includes a physically consistent three-part decomposition module, a fault-sensitive autocorrelation block module, and a structured bottleneck module. The physical consistency three-decomposition module performs a layer-by-layer three-decomposition process on the input sequence of grouped variables. The moving average sequence of the input sequence of grouped variables is calculated as the trend base sequence. The difference between the input sequence of grouped variables and the trend base sequence is used as the seasonal base sequence. The trend base sequence is mapped to the irradiation-driven component sequence and the temperature thermal effect component sequence. The remaining component after deducting the irradiation-driven component sequence and the temperature thermal effect component sequence from the input sequence of grouped variables is determined as the equipment health offset component sequence. Based on the fault-sensitive autocorrelation block module, the seasonal base sequence is subjected to periodic dependency representation extraction processing. The autocorrelation sequence of the seasonal base sequence is calculated, the set of delay quantities is selected from the autocorrelation sequence, and each delay quantity is aligned with the time delay of the seasonal base sequence and aggregated to obtain the periodic dependency representation sequence. Anomaly alignment characterization extraction processing is performed on the equipment health offset component sequence, the autocorrelation sequence of the equipment health offset component sequence is calculated, a set of delay quantities is selected from the autocorrelation sequence, and time delay alignment is performed on each delay quantity of the equipment health offset component sequence. The result is obtained by aggregation. A structured bottleneck module is introduced to perform slotting encoding processing on the environmental variable input sequence, DC side multivariate input sequence, inverter multivariate input sequence and grid-connected multivariate input sequence respectively, to obtain the environmental slotting characterization sequence, DC slotting characterization sequence, inverter slotting characterization sequence and grid-connected slotting characterization sequence, and generate a diagnostic result set, which includes the target sequence prediction result, health offset sequence, fault sensitivity characterization sequence and structured bottleneck slotting characterization sequence.

6. The method for fault diagnosis of distributed photovoltaic power stations based on big data analysis according to claim 1, characterized in that, The formation of the candidate driver group variable set includes: The health offset sequence and the fault-sensitive characterization sequence are obtained from the diagnostic result set. The sequence segments are obtained by sliding segmentation according to the window length on a unified time axis. An anomaly score is calculated for each sequence segment. The anomaly score includes the amplitude characteristics, persistence characteristics and fluctuation characteristics of the health offset sequence segment and the amplitude characteristics, persistence characteristics and fluctuation characteristics of the fault-sensitive characterization sequence segment. The anomaly score is compared with the anomaly threshold, and the anomaly trigger is determined by combining the working condition mark. The start time point and the end time point corresponding to the sequence segment that meets the anomaly trigger are determined as the anomaly time window. Based on the abnormal time window, the abnormal object level and abnormal object identifier are determined. The structured bottleneck slot characterization sequence is extracted within the abnormal time window. Within the data range corresponding to the abnormal object identifier, the environmental data group sequence, DC side data group sequence, inverter data group sequence and grid-connected side data group sequence, which are time-aligned with the structured bottleneck slot characterization sequence, are extracted to form a candidate drive group variable set.

7. The method for fault diagnosis of distributed photovoltaic power stations based on big data analysis according to claim 1, characterized in that, The obtained causal strength ranking results and location data group results include: Determine the target sequence and candidate driver group variable set for the cross-convergence mapping. The target sequence is selected from the health offset sequence, and the candidate driver group variable set includes the environmental data group sequence, the DC side data group sequence, the inverter data group sequence, and the grid-connected side data group sequence. Based on topological neighborhood graph construction, adjacency graphs are constructed on the reconstructed manifolds of each data group sequence in the target sequence and the candidate driving group variable set. The reconstructed manifold construction includes constructing each sequence into a delayed vector sequence according to the embedding dimension and the delay step size, and using each vector point of the delayed vector sequence to form the reconstructed manifold. The adjacency graph construction includes using each vector point in the reconstructed manifold as a graph node and establishing edge connections based on the distance between nodes to form adjacency relationships. Cross-mapping of group variables is performed in the connected neighborhood of the adjacency graph obtained by topological neighborhood construction. The set of connected neighborhood nodes corresponding to each node in the manifold reconstructed by the target sequence is used to determine the sample set of driving group variables. Based on the sample set of driving group variables, estimated sequences are generated for each variable sequence of the driving group variables. The cross-mapping calculation includes calculating the correlation coefficient between the estimated sequence and the corresponding real sequence as the cross-mapping capability index. The cross-mapping capability index of each variable sequence in the same data group is summarized to obtain the cross-mapping capability index of the data group. The cross-mapping capability index of each data group is sorted to generate a causal strength ranking result for the data groups. Based on the ranking result, the target localization data group with the highest causal strength is determined, and the localization data group result is output.

8. The method for fault diagnosis of distributed photovoltaic power stations based on big data analysis according to claim 1, characterized in that, The generated fault type discrimination results and fault location results include: Obtain the diagnostic result set and the causal strength ranking result of the data group, locate the data group result, and extract the health offset sequence, fault sensitivity characterization sequence and structured bottleneck groove characterization sequence according to the abnormal time window to obtain the diagnostic feature set within the window; Based on the location data set results, the location slot type is determined. The statistical features within the slot are extracted from the structured bottleneck slot characterization sequence corresponding to the location slot type. The key variable sequence aligned with the abnormal time window is extracted from the data set sequence corresponding to the location slot type. The system integrates diagnostic feature sets within the fusion window, location slot type, in-slot statistical features, and key variable sequences to generate fault type discrimination results, and generates fault location results based on the abnormal object level and location slot type.

9. The method for fault diagnosis of distributed photovoltaic power stations based on big data analysis according to claim 1, characterized in that, The generation of operation and maintenance decisions includes: Receive fault type identification results and fault location results, and determine the operation and maintenance object identifier and operation and maintenance object level based on the fault location results. The operation and maintenance object level includes the power plant level, inverter level and string level. Within the abnormal time window, offset amplitude and offset persistence indicators are generated based on the healthy offset sequence. The risk level and confidence level are generated by integrating the causal strength ranking results of the data group and the anomaly score. Based on the fault type identification results, location slot type, and maintenance object hierarchy, suggested inspection objects and suggested inspection data items are generated. Based on the risk level and confidence level, the handling priority is generated. The output includes the maintenance decision output containing the risk level, confidence level, suggested inspection objects, suggested inspection data items, and handling priority.

Images