Photovoltaic string fault diagnosis method based on improved rotating forest algorithm

Abstract

The invention discloses a photovoltaic string fault diagnosis method based on an improved rotating forest algorithm, and relates to the technical field of fault diagnosis. The method comprises the steps that string operation basic data is acquired and preprocessed, and a first feature data set is constructed; marking based on the state characteristics of the inverter and the combiner box to generate a first label; performing gradient analysis on the string time sequence characteristics to generate a second label; performing benchmarking analysis on the same type of group strings in the same upper layer, and generating a third label based on the deviation; acquiring defect features based on the component edge AI to generate a fourth label; and constructing a gating loop optimization rotating forest model, performing aggregation diagnosis on the data sets corresponding to the four types of labels, and outputting a string fault result. The technical problems that an existing photovoltaic power station equipment fault diagnosis model is low in robustness and has many invalid alarms are solved, and the technical effects that aggregation diagnosis is carried out by introducing the multi-source feature tags and fusing the gating loop optimization rotating forest model, and the accuracy and robustness of photovoltaic string fault recognition are improved are achieved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and specifically to a photovoltaic string fault diagnosis method based on an improved rotating forest algorithm. Background Technology

[0002] Photovoltaic power generation systems are playing an increasingly important role in the new energy structure, and photovoltaic strings, as the core unit on the DC side, directly affect the overall power generation efficiency and safety of the power station. However, photovoltaic power stations are widely distributed in complex and variable outdoor environments, making them susceptible to various factors such as dust accumulation, hot spot effects, shading, module aging, poor cable contact, and inverter malfunctions, leading to distortion of string operating parameters, performance degradation, and even potential faults. Traditional string fault diagnosis often relies on single features or static threshold judgments, which are difficult to fully adapt to the complex environment of photovoltaic scenarios with multi-source heterogeneous characteristics, multi-condition disturbances, and diverse fault modes, especially with low diagnostic accuracy in scenarios involving weak faults, small deviations, and cross-device comparisons. Summary of the Invention

[0003] This application provides a photovoltaic string fault diagnosis method based on an improved rotating forest algorithm, which solves the technical problems of low robustness and numerous invalid alarms in existing photovoltaic power plant equipment fault diagnosis models.

[0004] This application provides a photovoltaic string fault diagnosis method based on an improved rotating forest algorithm, the method comprising: The process involves acquiring and preprocessing basic string operation data, constructing a first feature dataset based on the string's rated operating parameters, labeling the first feature dataset with the operating status characteristics of the inverter and combiner box to generate a first label, performing gradient analysis on the string's time-series operating characteristics, and generating a second label for the first feature dataset based on gradient jumps, performing benchmark analysis on strings of the same type within the same upper-level device, and generating a third label for the first feature dataset based on deviation characteristics, obtaining module defect characteristics based on photovoltaic module edge AI analysis, and generating a fourth label for the first feature dataset, constructing a gated cyclic optimization rotating forest model, and aggregating and diagnosing the datasets corresponding to the first, second, third, and fourth labels to output photovoltaic string fault diagnosis results.

[0005] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, basic data on the operation of the photovoltaic strings is collected and preprocessed. Then, a first feature dataset is constructed based on the rated parameters, serving as the foundation for subsequent analysis. Next, labels are added to this dataset from different dimensions: firstly, labels are generated based on the operating status of the inverter and combiner box; secondly, second labels are generated by analyzing the gradient transitions of the string's time-series operating characteristics; thirdly, third labels are generated by benchmarking similar strings within the same upper-level equipment; and fourthly, fourth labels are generated by using edge AI analysis of photovoltaic modules to obtain defect features. Finally, a gated cyclic optimization rotating forest model is constructed to integrate the datasets corresponding to the four labels for comprehensive diagnosis, thereby outputting fault diagnosis results for the photovoltaic strings and achieving accurate and comprehensive fault judgment. Attached Figure Description

[0006] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0007] Figure 1 This is a schematic flowchart of a photovoltaic string fault diagnosis method based on an improved rotating forest algorithm, provided in an embodiment of this application.

[0008] Figure 2 This is a schematic diagram illustrating the process of constructing the first feature dataset for the photovoltaic string fault diagnosis method based on the improved rotating forest algorithm provided in the embodiments of this application. Detailed Implementation

[0009] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0010] Examples, such as Figure 1 As shown, this application provides a photovoltaic string fault diagnosis method based on an improved rotating forest algorithm, the method including: Obtain the basic data of the string operation and preprocess it, and construct the first feature dataset by combining the string's rated operating parameters.

[0011] In this embodiment, acquiring and preprocessing the basic operating data of the photovoltaic (PV) strings is a fundamental step in constructing a fault diagnosis model. Specifically, the operating data of the PV strings is first acquired in real time through a power plant monitoring system, inverter acquisition terminal, or data acquisition device. The acquired basic operating data includes key parameters reflecting the operating status of the PV modules, such as string current and maximum power current of the modules. Since the timestamps of data from different acquisition terminals may differ, the acquired raw time-series data needs to be time-aligned to ensure that all parameters correspond on a unified time axis. Subsequently, the aligned data is aggregated according to a preset time step to eliminate interference from instantaneous fluctuations, resulting in a smooth and continuous sequence of string operating characteristics. Based on this, the aggregated features are standardized and normalized to improve the stability of subsequent model training and feature analysis. Finally, based on the rated operating parameters of the photovoltaic string, such as rated current, rated power, and temperature correction coefficient, the processed feature sequences are screened and extracted to meet the requirements of the string's normal operating range. This constructs the first feature dataset for subsequent labeling and model training, realizing the structuring and standardization of photovoltaic string operating data and providing a reliable data foundation for fault diagnosis.

[0012] Furthermore, such as Figure 2 As shown, the basic data for string operation is obtained and preprocessed, and the first feature dataset is constructed by combining the string's rated operating parameters, including: Collect basic string operation data, which includes at least the string real-time current, component maximum power current, temperature coefficient, irradiance, and backplane temperature; perform time-series alignment on the basic string operation data and aggregate it according to a preset time step to obtain an aggregated string operation feature sequence; perform standardization and normalization processing on the aggregated string operation feature sequence to obtain a standard aggregated string operation feature sequence; filter the standard aggregated string operation feature sequence based on the string's rated operating parameters to obtain a first feature dataset.

[0013] Preferably, firstly, the basic operating data of the strings is acquired through the power plant monitoring system, inverter acquisition terminal, or data acquisition device. This basic operating data includes at least parameters such as the real-time current of the strings, the maximum power current of the modules, the temperature coefficient of the modules, the real-time irradiance of the area, and the temperature of the module backplane. These parameters can comprehensively reflect the electrical characteristics of the strings and their operating status affected by the environment. Since the acquisition period and timestamp of different parameters may differ, in order to ensure the accuracy of the correspondence between data, all basic operating data needs to be time-series aligned. That is, the parameter data is interpolated, padded, or truncated based on a unified time axis so that the data at the same point in time can be synchronously matched. Then, in order to reduce the impact of instantaneous noise and extract more representative operating features, the aligned data is aggregated according to a preset time step (e.g., 5 minutes). That is, by calculating the mean, maximum value, or other statistical characteristics within this time step, a more stable aggregated string operating feature sequence is obtained. After obtaining the operating feature sequence of the photovoltaic string, standardization and normalization are required to remove the dimensional differences between different parameters and improve the stability of subsequent feature analysis and model training. Standardization can be achieved using the z-score method, which transforms the features into a distribution with a mean of zero and a variance of one. Normalization can be achieved using the min-max transformation, which maps each feature to a fixed interval (e.g., 0 to 1). The standard operating feature sequence of the photovoltaic string obtained after this dual processing has better comparability and stability, and is more suitable for model input. Finally, the standard operating feature sequence of the photovoltaic string is screened based on the rated operating parameters of the photovoltaic string, such as rated current, rated module power, and performance parameters under stable operating conditions (irradiance fluctuation ≤ ±5% within 5 minutes, backsheet temperature fluctuation ≤ ±2℃ within 5 minutes, etc.). The screening process includes removing data that exceeds the rated operating range, does not meet the normal performance of the equipment, or has significant abnormal noise, ensuring that the retained data can truly reflect the normal or discernible state of the string. After the filtering is completed, a structured, clean, and physically meaningful first feature dataset is obtained, providing a reliable data foundation for subsequent labeling and model analysis.

[0014] Furthermore, the standardization process employs the z-score method, and the normalization process employs the min-max method.

[0015] Optionally, when processing the feature sequences of aggregated strings, standardization and normalization are required sequentially to eliminate dimensional differences between different features and enhance the usability of the data in the model analysis process. Specifically, standardization employs the z-score method, which calculates the mean and standard deviation of each feature across the entire sample sequence and transforms each value of that feature according to the formula "(current value - mean) ÷ standard deviation," resulting in a standard normal distribution with a mean of 0 and a standard deviation of 1. This method effectively eliminates scale differences between features, ensuring a fair representation of the weights of different features in the model. After standardization, to further map all features to a uniform numerical range, the min-max method is used for normalization. This involves linearly compressing all features to the range of 0 to 1 using the formula "(current value - minimum value) ÷ (maximum value - minimum value)." This method not only facilitates model convergence for different features but also improves the model's stability against outliers. By applying z-score standardization and min-max normalization sequentially, the original physical quantity data can be transformed into standardized features with a uniform scale and stable distribution, providing a higher quality data foundation for subsequent fault diagnosis and reasoning.

[0016] The first feature dataset is labeled with the status characteristics of the inverter and combiner box to generate the first label.

[0017] In one embodiment, when labeling the first feature dataset with its status, it is first necessary to obtain operating status features reflecting the health status of the photovoltaic string from the inverter and combiner box. These features include key parameters such as the string inverter's operating signals, power-limiting signals, and DC-side bus current. Subsequently, pre-set normal thresholds corresponding to various status features are obtained. When the actual operating status features collected by the equipment are compared with the corresponding thresholds, if the parameters fall within the corresponding thresholds, the string is considered to be in normal operating condition; if they exceed the corresponding thresholds, for example, if there are significant abnormal fluctuations in current or voltage, or if power-limiting signals or fault signals appear, it indicates that the string may have an operational fault or performance degradation. Based on the above comparison results, the feature data at the corresponding time in the first feature dataset are classified and labeled with their corresponding first label as "normal" or "fault". Through this labeling process, the first feature dataset obtains accurate information identification based on the equipment's operating conditions, providing a reliable basic status basis for subsequent multi-label fusion diagnosis of the model.

[0018] Furthermore, based on the operating status characteristics of the inverter and combiner box, the first feature dataset is labeled with status information to generate a first label, including: The inverter and combiner box operating status characteristics include at least string inverter operating signals, power limiting signals, fault signals, and DC bus current data, voltage data, and power data; a preset normal threshold is determined based on the equipment's factory parameters; the inverter and combiner box operating status characteristics are compared with the preset normal threshold to generate a first label for the first feature dataset, and the first label is divided into a fault label and a normal label.

[0019] Preferably, when labeling the first feature dataset, the operating status features of the inverter and combiner box are first collected, and then these operating status features are subjected to the same preprocessing as described above. These operating status features include at least the operating signals, power-limiting signals, and fault signals of the string inverter, as well as key data that can directly reflect the electrical operating status of the string, such as the current, voltage, and power of the DC bus. Since different devices have clearly defined rated operating parameters and corresponding safety protection ranges at the factory, preset normal thresholds for various operating status features can be determined based on the equipment's factory technical standards, rated electrical parameter ranges, and actual operating experience. When the above operating status features are compared item by item with the preset normal thresholds, if the collected values ​​of current, voltage, power, etc. are all within a reasonable range, and the inverter does not show a power-limiting or fault signal, then the string operating status corresponding to that feature data is determined to be normal; conversely, if any status feature exceeds the normal threshold range, for example, the DC current deviates significantly from the rated range, or a power-limiting control or equipment fault signal appears, then the string corresponding to that feature data is determined to have a potential fault. Based on the comparison results, the corresponding data in the first feature dataset are assigned "normal label" or "fault label" respectively, thereby generating the first label sequence for subsequent multi-label fusion diagnosis, realizing automated status labeling based on the actual operating conditions of the equipment.

[0020] Gradient analysis is performed on the time-series characteristics of the string, and a second label is generated for the first feature dataset based on the gradient transition.

[0021] In one embodiment, when generating the second label, a continuity analysis of the string's time-series operating characteristics is first performed to identify potential abnormal trends during operation. Specifically, the time-related current, voltage, power, and other key parameters concentrated in the first feature data are constructed into a time-series feature sequence in chronological order. Gradient analysis is applied to this sequence to calculate the rate of change between adjacent time points, reflecting the dynamic changes in the string's operating characteristics. When a sudden anomaly occurs in the operating state, such as component hotspots, shading, poor contact, or other faults that cause a sharp drop in performance, the feature sequence usually exhibits a significant gradient jump, meaning the gradient value deviates significantly from the normal range within a short period. By pre-setting a gradient threshold to distinguish between normal fluctuations and suspected anomalies, when the gradient change exceeds the preset threshold, it indicates an abnormal abrupt change in the string's operating characteristics, and the second label for the corresponding sample is marked as a fault warning label; otherwise, the second label is marked as a normal warning label. This process fully utilizes the sensitivity of time-series data to early fault characteristics, transforming dynamic change signals into label information that can be recognized by the model, providing an effective trend reference for subsequent multi-dimensional feature fusion diagnosis.

[0022] Furthermore, gradient analysis is performed on the temporal characteristics of the sequences, and a second label is generated for the first feature dataset based on the gradient transitions, including: Gradient descent analysis is performed on the timing characteristics of the string to generate a gradient transition signal; when the gradient transition signal exceeds a preset gradient threshold, the second label of the first feature dataset is marked as a fault indication label; when the gradient transition signal is within the preset gradient threshold range, the second label of the first feature dataset is marked as a normal indication label.

[0023] Preferably, in the process of generating the second tag, operating parameters such as current, voltage, and power that change over time are first constructed into string time-series operating features according to the sampling time sequence, and the numerical differences between adjacent time points are calculated to obtain the gradient value of the feature at each moment, thereby reflecting the rate of change of the corresponding feature in different time periods. During gradient descent analysis, by continuously comparing the positive and negative directions and the magnitude of the rate of change, the location of sudden changes in the feature's trend can be identified. When the rate of change of a feature shows a significant decrease or abrupt change in a short period of time, its corresponding gradient difference value will undergo a significant jump, so the absolute change in the gradient difference can be used as the gradient jump signal. To determine whether the gradient jump indicates a potential fault, the gradient jump signal needs to be compared with a preset gradient threshold. This preset gradient threshold is obtained based on the analysis of a large amount of historical string operating data and can distinguish between normal operating fluctuations and sudden changes caused by faults. When the gradient jump signal of a certain feature exceeds a preset threshold, indicating a significantly abnormal change, the second label corresponding to that feature is marked as a fault warning label, signifying a sudden abnormal fluctuation in the string's operating status at that moment, potentially indicating a fault. Conversely, when the gradient jump signal is within the preset gradient threshold range, it indicates that the feature change is within the range of normal operation fluctuations, and its corresponding second label is marked as a normal warning label. Through the above processing, the original time-series dynamic change features are transformed into label information based on change trends, enabling the model to identify potential early fault features of abrupt changes and enriching the data dimensions for fault diagnosis.

[0024] A comparative analysis is performed on the same type of strings within the same upper-layer device, and a third label is generated for the first feature dataset based on the deviation characteristics.

[0025] In one embodiment, during the generation of the third tag, a benchmarking analysis is first performed on photovoltaic strings of the same type within the same upper-level device (such as the same combiner box or the same inverter). Specifically, the operating parameters of each string are extracted within the same time period, and strings with stable operating status and excellent historical performance are selected as benchmark reference strings, with their characteristics serving as the standard baseline. For the target string to be analyzed, the difference between its operating parameters and the corresponding characteristics of the benchmark string is calculated, and a deviation coefficient is obtained based on the difference value to quantitatively reflect the degree of operational deviation between the target string and normal strings. When the absolute value of the deviation coefficient exceeds a preset deviation threshold, it indicates that the output characteristics of the target string have significantly deviated from the normal level of similar strings, and there may be power attenuation, poor wiring, or hidden faults. In this case, its corresponding third tag is marked as an abnormal tag. Conversely, when the deviation coefficient is within the preset deviation threshold range, it indicates that the string's operating performance is consistent with similar equipment and is in a normal state, and the third tag is marked as a normal tag. This benchmarking analysis process can fully utilize the relative performance differences between similar strings to achieve sensitive identification of minor anomalies and potential faults, providing complementary feature support for the comprehensive diagnosis of subsequent models.

[0026] Furthermore, a comparative analysis is performed on the same type of strings within the same upper-layer device, and a third label is generated for the first feature dataset based on the deviation characteristics, including: Using the same type of string within the same upper-level device as the benchmark, calculate the deviation coefficient; when the absolute value of the deviation coefficient exceeds a preset deviation threshold, mark the third label of the first feature dataset as an abnormal label; when the absolute value of the deviation coefficient is within the preset deviation threshold range, mark the third label of the first feature dataset as a normal label.

[0027] Preferably, in the process of generating the third label, the target string is first selected as the benchmark object, which is located in the same upper-level device (e.g., the same combiner box or the same inverter) and of the same type. This ensures that the comparison benchmark is consistent in terms of operating environment, light level, and equipment structure. Subsequently, the operating parameters of each string within the same time period, including current, power, and voltage, are extracted from the first feature dataset. For both the target string and the benchmark string, the difference between each operating parameter of the target string and the mean value of the corresponding operating parameters of the benchmark string is calculated. Based on the ratio of the difference to the mean value, a deviation coefficient is obtained to reflect the degree of relative performance deviation of the target string. After obtaining the deviation coefficient, the operating status of the string is determined by comparing it with a preset deviation threshold. This preset deviation threshold is determined based on a large amount of historical operating data from power plants and is used to distinguish between normal operating fluctuations and abnormal performance degradation. When the absolute value of the deviation coefficient exceeds the preset deviation threshold, it indicates that the output performance of the target string deviates significantly from that of similar strings, potentially indicating component damage, loose wiring, obstruction, or other potential faults. In this case, the third label corresponding to this data in the first feature dataset is marked as an abnormal label. Conversely, when the absolute value of the deviation coefficient is within the deviation threshold range, it indicates that the performance of the target string is consistent with that of similar strings and is in normal operation; therefore, its corresponding third label is marked as a normal label. Through the above benchmarking analysis process, slight performance degradation that is difficult to detect using traditional thresholding methods can be effectively identified, thereby enhancing the accuracy and sensitivity of fault diagnosis and providing a reliable basis for relative performance characteristics for subsequent multi-source label fusion.

[0028] Based on edge AI analysis of photovoltaic modules, defect features of the modules are obtained, and a fourth label is generated for the first feature dataset.

[0029] In one embodiment, during the generation of the fourth label, an edge AI analysis model deployed at the photovoltaic module is first used to acquire the module's defect characteristics. This edge AI model is typically integrated into the inverter or edge computing device and can automatically identify potential anomalies in the module based on the module's IV curve scanning results or infrared thermal imaging data, such as hot spots, microcracks, shading, solder ribbon detachment, and uneven attenuation. During system operation, the edge AI analysis model is periodically or on demand triggered to extract features and perform pattern recognition on the electrical characteristics or infrared images collected in real time from the module, outputting corresponding defect detection results. Subsequently, the output of the edge AI is correlated and matched with the first feature dataset to label its corresponding fourth label. Through this process, structural defect identification at the module level can be transformed into label information usable by the model, so that fault diagnosis no longer relies solely on electrical parameters, but introduces auxiliary judgments from the physical and visual levels, providing the model with more comprehensive input information, thereby significantly improving the accuracy and reliability of diagnostic results.

[0030] Furthermore, based on edge AI analysis of photovoltaic modules to obtain module defect features, a fourth label is generated for the first feature dataset, including: The output information of the edge AI analysis model is collected. The edge AI analysis model is either the inverter's built-in IV curve scanning model or the infrared camera scanning model. If the output information is empty, the fourth label of the first feature dataset is marked as a normal label. If the output information is not empty, the fourth label of the first feature dataset is marked as a fault type label according to the fault information.

[0031] Preferably, in the process of generating the fourth label, the output information of the edge AI analysis model at the photovoltaic module end is first obtained. This edge AI model can be deployed inside the inverter, using its built-in IV curve scanning model to periodically test the electrical characteristics of the module, or it can be an infrared camera scanning model to identify physical defects such as abnormal hot spots, microcracks, or shading distribution on the module surface through infrared image analysis. When the system collects the module operating status for the corresponding time period of the first feature dataset, it simultaneously retrieves the detection results of the edge AI model and uses them as the source of module defect features. After obtaining the edge AI output information, its content is judged. When the output information is empty, it means that no module abnormalities were detected in the corresponding scanning period, and the appearance or electrical characteristics of the module are within the normal range. At this time, the fourth label of the first feature dataset corresponding to that time point is marked as a normal label. Conversely, when the output information is not empty, it means that the edge AI model has identified potential defects in the module, such as hot spots, weld detachment, microcracks, abnormal shading, PID effect areas, etc. The system performs corresponding fault classification labeling on the fourth label according to the defect type in the edge AI output report, so that each data record can be matched with a clear module fault type label. Through the above process, the physical defect detection results at the photovoltaic module level can be reliably converted into structured label information that can be used for machine learning models. This allows the fourth label to reflect the health status of the module itself, providing a key module-level judgment basis for subsequent multi-source label fusion diagnosis.

[0032] A gated cyclic optimization rotating forest model is constructed, and the datasets corresponding to the first label, the second label, the third label, and the fourth label are aggregated for diagnosis, outputting the photovoltaic string fault diagnosis results.

[0033] In one embodiment, the first, second, third, and fourth labels generated above are first integrated with their corresponding first feature datasets to form a comprehensive dataset containing multi-dimensional features and multi-label annotations. This comprehensive dataset is then input into a gated recurrent unit (GRU) optimized rotation forest model for analysis. Specifically, the model first utilizes the feature subspace partitioning mechanism of rotation forest, performing principal component analysis (PCA) to rotate the input features, generating multiple subsets with differentiated feature distributions to enhance the model's independent learning ability for different feature dimensions. Next, each subset is analyzed using integrated multiple GRU sub-models to capture the dynamic changes in the sequence running features and output the prediction results of the sub-models. To further improve the accuracy and stability of the overall diagnosis, a weighted voting method is used to integrate and optimize the results of multiple GRU sub-models. The weights are automatically adjusted based on the validation performance of each sub-model during the training phase; that is, the ratio of the sub-model's accuracy on the validation set to the sum of the sub-models' accuracy on the validation set. Finally, the prediction result with the highest weighted probability is used as the final fault diagnosis result, achieving intelligent and refined fault identification driven by multi-dimensional information fusion.

[0034] Furthermore, constructing a gated loop-optimized rotating forest model includes: Using the first label, the second label, the third label, and the fourth label as search criteria, multiple subsets are collected from historical fault records. Principal component analysis is performed on the multiple subsets to obtain feature rotation matrices, and multiple training subsets are generated based on the feature rotation matrices. Multiple GRU sub-models are constructed, and the multiple training subsets are respectively input into the multiple GRU sub-models for training to generate multiple optimized GRU sub-models. A weighted voting mechanism is used to integrate the multiple optimized GRU sub-models to construct a gated cyclic optimization rotating forest model.

[0035] Preferably, when constructing the gated recurrent optimization rotating forest model, the first step is to use the first, second, third, and fourth labels as search criteria to filter operational data corresponding to each label from the historical fault record database. This historical data is then divided into multiple subsets, each containing data samples reflecting a specific fault mode or operational state, enhancing the model's ability to learn from different features. Subsequently, principal component analysis (PCA) is performed on the original feature matrices of each subset to extract the principal feature directions with a cumulative variance contribution rate ≥95%, generating a feature rotation matrix. Transforming the feature space of each subset using this rotation matrix effectively improves the discriminative power between different feature dimensions and enhances the model's ability to express complex feature structures. Based on this feature rotation matrix, the multiple subsets are transformed into multiple training subsets, each with an independent feature subspace representation, thus improving the overall model's diversity. Finally, for each of the generated training subsets, a corresponding GRU sub-model is constructed. The GRU sub-model includes an input layer, hidden layer, output layer, update gate, and reset gate. Subsequently, using Adam as the optimizer and cross-entropy loss as the loss function, the GRU sub-models utilize their gating structure to process the correlation of time-series data. By memorizing key features and suppressing redundant features, they learn different time-series behavioral patterns from each training subset. After sufficient training, each sub-model develops an independent expressive ability for a specific feature rotation space, resulting in multiple optimized GRU sub-models. To improve the stability and accuracy of the overall diagnosis, a weighted voting mechanism is used to integrate the multiple optimized GRU sub-models. The weights of each sub-model are dynamically allocated based on its performance on the validation set, with higher-performing sub-models having a higher proportion in the final integration result. This constructs a robust and highly generalizable gated recurrent optimized rotating forest model, enabling it to aggregate and diagnose input multi-label data and output reliable photovoltaic string fault identification results.

[0036] In summary, the embodiments of this application have at least the following technical effects: First, basic data on the operation of the photovoltaic (PV) modules is acquired and preprocessed, and a first feature dataset is constructed based on the rated operating parameters of the modules. Then, the first feature dataset is labeled with the operating status characteristics of the inverter and combiner box, generating a first label. Next, gradient analysis is performed on the time-series operating characteristics of the modules, and a second label is generated for the first feature dataset based on gradient jumps. Then, a comparative analysis is performed on similar modules within the same upper-level equipment, and a third label is generated for the first feature dataset based on deviation characteristics. Then, module defect characteristics are obtained based on edge AI analysis of the PV modules, generating a fourth label for the first feature dataset. Finally, a gated cyclic optimization rotating forest model is constructed, and the datasets corresponding to the first, second, third, and fourth labels are aggregated for diagnosis, outputting the PV module fault diagnosis results. This solves the technical problems of low robustness and numerous invalid alarms in existing PV power plant equipment fault diagnosis models, achieving improved accuracy and robustness of PV module fault identification by introducing multi-source feature labels and fusing them with a gated cyclic optimization rotating forest model for aggregated diagnosis.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A photovoltaic string fault diagnosis method based on an improved rotating forest algorithm, characterized in that, The method includes: Acquire the basic data of the string operation and preprocess it, and construct the first feature dataset by combining the string's rated operation parameters; The first feature dataset is labeled with the status characteristics of the inverter and combiner box to generate the first label. Gradient analysis is performed on the temporal characteristics of the string, and a second label is generated for the first feature dataset based on the gradient transition. A comparative analysis is performed on the same type of strings within the same upper-layer device, and a third label is generated for the first feature dataset based on the deviation characteristics. Based on edge AI analysis of photovoltaic modules, defect features of the modules are obtained, and a fourth label is generated for the first feature dataset. A gated cyclic optimization rotating forest model is constructed, and the datasets corresponding to the first label, the second label, the third label, and the fourth label are aggregated for diagnosis, outputting the photovoltaic string fault diagnosis results.

2. The photovoltaic string fault diagnosis method based on the improved rotating forest algorithm as described in claim 1, characterized in that, Obtain the basic data for string operation and preprocess it. Combine this data with the string's rated operating parameters to construct the first feature dataset, including: Collect basic data on string operation, including at least the string real-time current, the module maximum power current, the temperature coefficient, the irradiance, and the backplane temperature. The basic data of the string operation is time-series aligned and aggregated according to a preset time step to obtain the aggregated string operation feature sequence. The aggregated string running feature sequence is standardized and normalized to obtain a standard aggregated string running feature sequence. The standard aggregated string operation feature sequence is filtered based on the string's rated operating parameters to obtain the first feature dataset.

3. The photovoltaic string fault diagnosis method based on the improved rotating forest algorithm as described in claim 2, characterized in that, The standardization process uses the z-score method, and the normalization process uses the min-max method.

4. The photovoltaic string fault diagnosis method based on the improved rotating forest algorithm as described in claim 1, characterized in that, Based on the operating status characteristics of the inverter and combiner box, the first feature dataset is labeled with status information to generate a first label, including: The inverter and combiner box operating status characteristics include at least string inverter operating signals, power limiting signals, fault signals, and DC bus current data, voltage data, and power data; Determine the preset normal threshold based on the equipment's factory parameters; The inverter and combiner box operating status characteristics are compared with the preset normal threshold to generate the first label of the first feature dataset. The first label is divided into fault label and normal label.

5. The photovoltaic string fault diagnosis method based on the improved rotating forest algorithm as described in claim 1, characterized in that, Gradient analysis is performed on the temporal characteristics of the sequences, and a second label is generated for the first feature dataset based on the gradient transitions, including: Gradient descent analysis is performed on the timing characteristics of the string to generate gradient transition signals; When the gradient jump signal exceeds a preset gradient threshold, the second label of the first feature dataset is marked as a fault indication label; When the gradient jump signal is within the preset gradient threshold range, the second label of the first feature dataset is marked as a normal prompt label.

6. The photovoltaic string fault diagnosis method based on the improved rotating forest algorithm as described in claim 1, characterized in that, Comparative analysis is performed on similar strings within the same upper-layer device. Based on the deviation features, a third label is generated for the first feature dataset, including: Using the same type of string within the same upper-level device group as the benchmark, calculate the deviation coefficient; When the absolute value of the deviation coefficient exceeds the preset deviation threshold, the third label of the first feature dataset is marked as an abnormal label; When the absolute value of the deviation coefficient is within the preset deviation threshold range, the third label of the first feature dataset is marked as a normal label.

7. The photovoltaic string fault diagnosis method based on the improved rotating forest algorithm as described in claim 1, characterized in that, Based on edge AI analysis of photovoltaic modules, defect features of the modules are obtained, and a fourth label is generated for the first feature dataset, including: Collect the output information of the edge AI analysis model, which is the inverter's built-in IV curve scanning model or the infrared camera scanning model; If the output information is empty, mark the fourth label of the first feature dataset as a normal label; If the output information is not empty, the fourth label of the first feature dataset is marked as the fault type label according to the fault information.

8. The photovoltaic string fault diagnosis method based on the improved rotating forest algorithm as described in claim 1, characterized in that, Construct a gated loop-optimized rotating forest model, including: Using the first tag, the second tag, the third tag, and the fourth tag as search criteria, multiple subsets are collected from historical fault records; Principal component analysis is performed on the multiple subset datasets to obtain feature rotation matrices, and multiple training subsets are generated based on the feature rotation matrices; Construct multiple GRU sub-models, and input the multiple training subsets into the multiple GRU sub-models for training to generate multiple optimized GRU sub-models; A weighted voting mechanism is used to integrate the multiple optimized GRU sub-models to construct a gated cyclic optimization rotating forest model.

Images