Photovoltaic long tail fault identification method and device, electronic equipment and storage medium

Abstract

The application discloses a photovoltaic long tail fault recognition method and device, electronic equipment and a storage medium. First, the monitoring data is unsupervised cleaned to eliminate abnormal noise samples, and feature screening is performed, which can effectively remove interference and redundant information and ensure the quality of model input. At the same time, the original samples are adaptively enhanced according to the current operating condition, which can be targeted to expand the rare fault samples in the long-tail distribution, and avoid that the minority class fault features are submerged by normal samples. Further, the dynamic residual encoder is used to extract deep fault discrimination features, and the reliability of the classification probability is corrected, so that the diagnosis result is more robust and reliable. The rare fault recognition accuracy and reliability of the component hot spot, string mismatch, inverter temperature rise anomaly and the like are significantly improved, which provides effective support for the precise operation and maintenance and repair decision of the photovoltaic power station.

Description

Technical Field

[0001] This disclosure relates to the field of photovoltaic power generation technology, and in particular to a photovoltaic long-tail fault identification method, device, electronic device and storage medium. Background Technology

[0002] As a core infrastructure for new energy power generation, the safe and stable operation of photovoltaic power plants directly affects the reliability of power grid supply and economic benefits. With the continuous growth of photovoltaic installed capacity, intelligent operation and maintenance technology based on monitoring systems has become a key means to ensure the efficient operation of power plants. Among related technologies, an online monitoring system covering key equipment such as strings, combiner boxes, inverters, and transformer substations has been constructed through the collaborative operation of data acquisition, feature extraction, and fault classification models.

[0003] Existing photovoltaic fault diagnosis methods directly train classification models using full monitoring data, without effectively addressing the extreme class imbalance and long-tail distribution characteristics between normal samples and high-risk fault samples. This results in minority class fault features being submerged by majority class samples, affecting the model's ability to identify rare faults such as module hot spots, string mismatch, and abnormal inverter temperature rise. Consequently, the diagnostic results are insufficient to effectively support inspection scheduling and maintenance decisions. Summary of the Invention

[0004] This disclosure provides a photovoltaic long-tail fault identification method, apparatus, electronic device, and storage medium.

[0005] According to a first aspect of this disclosure, a photovoltaic long-tail fault identification method is provided, comprising: Acquire monitoring data of photovoltaic power plants, perform unsupervised cleaning on the monitoring data to remove abnormal noise samples, and perform feature filtering on the cleaned data to obtain the model input feature set; Based on the current operating condition information, adaptive enhancement processing is performed on the original samples in the input feature set of the model to generate enhanced samples; Fault discrimination features are obtained by extracting features from the enhanced samples using a dynamic residual encoder. Based on the fault discrimination features, the system performs classification and identification, outputs the fault category and the corresponding category probability, and performs reliability correction on the category probability to generate the fault identification result.

[0006] Optionally, the unsupervised cleaning of the monitoring data to remove abnormal noise samples includes: The monitoring data is clustered using a density-based noisy spatial clustering algorithm. Data samples with connected densities are clustered into the same cluster, and isolated samples that cannot be assigned to any cluster are identified as abnormal noise samples and removed. The abnormal noise samples include null samples caused by communication interruption, invalid power generation samples at night, sensor mutation samples, abnormal jumps in irradiance samples, inverter offline samples, and non-fault power-limited samples.

[0007] Optionally, the feature filtering of the cleaned data includes: Calculate the correlation metric between each candidate feature and the fault category, as well as the redundancy metric between any two candidate features; The feature relevance quotient of each candidate feature is calculated based on the relevance metric and the redundancy metric. Features are then selected in descending order of the feature relevance quotient to form the model input feature set.

[0008] Optionally, the adaptive enhancement processing of the original sample based on the current operating condition information includes: The irradiance fluctuation range and power fluctuation range of the photovoltaic power station are obtained, and the degree of operating condition fluctuation is calculated and mapped to the enhancement intensity coefficient. Based on the enhancement intensity coefficient, a perturbation matching the current operating condition fluctuation is injected into the original sample, while maintaining the basic physical relationship of the photovoltaic operation data.

[0009] Optionally, the basic physical relationships for maintaining photovoltaic operating data include: Maintain a positive correlation between irradiance and available power, a negative correlation between component temperature and output efficiency, a consistent relationship between currents in adjacent strings within the same combiner box, and an energy conversion relationship between inverter input power and output power.

[0010] Optionally, the dynamic residual encoder adaptively adjusts the convolutional response based on the features of the input samples, including: The enhanced samples are first convolutionally mapped to generate an input feature map, and then two pooling description vectors are generated by global average pooling and global standard deviation pooling, respectively. The two pooling description vectors are concatenated and then passed through a fully connected layer to generate a dynamic convolution kernel weight vector. The preset static convolution kernels are then weighted and combined to generate a dynamic convolution kernel that matches the features of the current input sample. The dynamic convolution kernel is convolved with the input feature map, and then an identity mapping is performed through a residual connection structure. Fault discrimination features are extracted through multi-layer feature transformation.

[0011] Optionally, the classification and identification based on the fault discrimination features includes: A dynamic hard sample mining strategy is adopted to determine hard samples based on the distance relationship between the feature representation of the training samples and the center of each category, and to impose stronger feature consistency constraints on the hard samples. A classification loss term with angular margin and a contrastive learning loss term based on the feature consistency constraint between enhanced samples and original samples are constructed. The two are weighted and summed by setting learnable coefficients, and the learnable coefficients are automatically updated during training to achieve dynamic loss balance.

[0012] Optionally, the reliability correction of the category probability includes: The model output class probability is divided into multiple probability intervals, and the difference between the average predicted probability in each interval and the actual failure frequency is calculated to obtain the expected calibration error value. When the expected calibration error value exceeds a preset threshold, a temperature calibration algorithm is used to correct the category probability. Based on the corrected category probabilities, combined with the importance of the equipment, historical alarm records, and the impact of power generation loss, maintenance priority suggestions are generated.

[0013] According to a second aspect of this disclosure, a photovoltaic long-tail fault identification device is provided, comprising: The acquisition unit is used to acquire monitoring data of photovoltaic power plants, perform unsupervised cleaning on the monitoring data to remove abnormal noise samples, and perform feature filtering on the cleaned data to obtain the model input feature set. The enhancement unit is used to perform adaptive enhancement processing on the original samples in the model input feature set according to the current operating condition information, and generate enhanced samples; The extraction unit is used to extract features from the enhanced sample through a dynamic residual encoder to obtain fault discrimination features; The identification unit is used to classify and identify faults based on the fault discrimination features, output fault categories and corresponding category probabilities, perform reliability correction on the category probabilities, and generate fault identification results.

[0014] Optionally, the acquisition unit is further configured to: The monitoring data is clustered using a density-based noisy spatial clustering algorithm. Data samples with connected densities are clustered into the same cluster, and isolated samples that cannot be assigned to any cluster are identified as abnormal noise samples and removed. The abnormal noise samples include null samples caused by communication interruption, invalid power generation samples at night, sensor mutation samples, abnormal jumps in irradiance samples, inverter offline samples, and non-fault power curtailment samples.

[0015] Optionally, the acquisition unit is further configured to: Calculate the correlation metric between each candidate feature and the fault category, as well as the redundancy metric between any two candidate features; The feature relevance quotient of each candidate feature is calculated based on the relevance metric and the redundancy metric. Features are then selected in descending order of the feature relevance quotient to form the model input feature set.

[0016] Optionally, the enhancement unit is further configured to: The irradiance fluctuation range and power fluctuation range of the photovoltaic power station are obtained, and the degree of operating condition fluctuation is calculated and mapped to the enhancement intensity coefficient. Based on the enhancement intensity coefficient, a perturbation matching the current operating condition fluctuation is injected into the original sample, while maintaining the basic physical relationship of the photovoltaic operation data.

[0017] Optionally, the enhancement unit is further configured to: Maintain a positive correlation between irradiance and available power, a negative correlation between component temperature and output efficiency, a consistent relationship between currents in adjacent strings within the same combiner box, and an energy conversion relationship between inverter input power and output power.

[0018] Optionally, the extraction unit is further configured to: The enhanced samples are first convolutionally mapped to generate an input feature map, and then two pooling description vectors are generated by global average pooling and global standard deviation pooling, respectively. The two pooling description vectors are concatenated and then passed through a fully connected layer to generate a dynamic convolution kernel weight vector. The preset static convolution kernels are then weighted and combined to generate a dynamic convolution kernel that matches the features of the current input sample. The dynamic convolution kernel is convolved with the input feature map, and then an identity mapping is performed through a residual connection structure. Fault discrimination features are extracted through multi-layer feature transformation.

[0019] Optionally, the identification unit is further configured to: A dynamic hard sample mining strategy is adopted to determine hard samples based on the distance relationship between the feature representation of the training samples and the center of each category, and to impose stronger feature consistency constraints on the hard samples. A classification loss term with angular margin and a contrastive learning loss term based on the feature consistency constraint between enhanced samples and original samples are constructed. The two are weighted and summed by setting learnable coefficients, and the learnable coefficients are automatically updated during training to achieve dynamic loss balance.

[0020] Optionally, the identification unit is further configured to: The model output class probability is divided into multiple probability intervals, and the difference between the average predicted probability in each interval and the actual failure frequency is calculated to obtain the expected calibration error value. When the expected calibration error value exceeds a preset threshold, a temperature calibration algorithm is used to correct the category probability. Based on the corrected category probabilities, combined with the importance of the equipment, historical alarm records, and the impact of power generation loss, maintenance priority suggestions are generated.

[0021] According to a third aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect above.

[0022] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are configured to cause the computer to perform the method described in the first aspect above.

[0023] According to a fifth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described in the first aspect above.

[0024] The photovoltaic long-tail fault identification method, device, electronic equipment, and storage medium disclosed herein first perform unsupervised cleaning of monitoring data to remove abnormal noise samples and then perform feature screening, which can effectively remove interference and redundant information and ensure the quality of model input. Simultaneously, it adaptively enhances the original samples according to the current operating conditions, enabling targeted expansion of rare fault samples in the long-tail distribution, preventing minority fault features from being overwhelmed by normal samples. Furthermore, it utilizes a dynamic residual encoder to extract deep fault discrimination features and performs reliability correction on the classification probability, making the diagnostic results more robust and reliable. Therefore, it can solve the technical problems of existing photovoltaic fault diagnosis methods, such as the overwhelming of minority fault features due to extremely unbalanced categories and long-tail distribution characteristics, insufficient rare fault identification capabilities, and difficulty in effectively supporting inspection scheduling and maintenance decisions. It achieves a significant improvement in the accuracy and reliability of rare fault identification, such as module hot spots, string mismatch, and abnormal inverter temperature rise, providing effective support for precise operation and maintenance decisions in photovoltaic power plants.

[0025] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0026] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a flowchart illustrating a photovoltaic long-tail fault identification method provided in an embodiment of this disclosure; Figure 2 This is a schematic diagram of the structure of a photovoltaic long-tail fault identification device provided in an embodiment of this disclosure; Figure 3 A schematic block diagram of an example electronic device provided for embodiments of this disclosure. Detailed Implementation

[0027] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0028] The following description, with reference to the accompanying drawings, outlines a photovoltaic long-tail fault identification method, apparatus, electronic device, and storage medium according to embodiments of the present disclosure.

[0029] Figure 1 This is a flowchart illustrating a photovoltaic long-tail fault identification method provided in an embodiment of this disclosure.

[0030] like Figure 1 As shown, the method includes the following steps: Step 101: Obtain monitoring data of photovoltaic power station, perform unsupervised cleaning on the monitoring data to remove abnormal noise samples, and perform feature screening on the cleaned data to obtain the model input feature set; The monitoring system of a photovoltaic power station continuously collects multi-dimensional operational data, including irradiance, ambient temperature, module temperature, DC voltage, DC current, string current, combiner box branch current, inverter AC / DC power, grid-connected voltage and current, transformer temperature, and alarm status. During long-term collection, this raw data inevitably becomes mixed with various noise samples, such as null records caused by communication interruptions, invalid power generation data during periods of zero irradiance at night, abnormal sampling points caused by instantaneous sensor fluctuations, abnormal data during inverter offline periods, and power curtailment data not caused by faults. If these noise samples are directly used for model training, they will severely interfere with the model's learning of real fault characteristics; therefore, effective data cleaning is necessary before modeling.

[0031] In one implementation, a density-based clustering method is used to perform unsupervised cleaning of the original monitoring data. This method aggregates densely distributed and interconnected samples into clusters based on the density reachability of data samples in the feature space, while isolated samples that cannot be assigned to any valid cluster are identified as anomalous noise points deviating from the main operating condition distribution and are thus removed. After cleaning, the remaining data samples can more accurately reflect the true operating status of the photovoltaic power station under different operating conditions.

[0032] After data cleaning, a subset of features with high predictive value for fault identification needs to be selected from the multidimensional monitoring variables to reduce the redundancy of the model input and improve computational efficiency. The feature selection process needs to consider both the correlation between features and fault categories and the degree of redundancy between different features. For example, an information-theoretic feature selection strategy can be adopted, calculating the mutual information between each candidate feature and the fault category label as a correlation measure, and simultaneously calculating the mutual information between any two candidate features as a redundancy measure. Then, the evaluation index of each feature is calculated by combining both, and features are selected sequentially from high to low evaluation index until a preset feature quantity condition is met, forming the model input feature set. The selected feature subset typically covers key operating variables such as irradiance, ambient temperature, component temperature, DC voltage, DC current, string current, inverter power, transformer temperature, grid connection parameters, and alarm status, as well as composite indicators derived from the original variables, such as performance ratio deviation, temperature-corrected power deviation, and string current imbalance.

[0033] Step 102: Based on the current operating condition information, perform adaptive enhancement processing on the original samples in the model input feature set to generate enhanced samples; The actual operating conditions of photovoltaic power plants exhibit significant dynamic fluctuations. Irradiance fluctuates frequently due to factors such as cloud cover and weather changes, while module temperature changes in real time with ambient temperature and light intensity. These fluctuations can cause shifts in the distribution characteristics of monitoring data. In scenarios with a long-tailed distribution where fault samples are already scarce, if the training data lacks effective coverage of the fluctuation characteristics of different operating conditions, the model will struggle to achieve robust fault identification in real-world operating environments. Therefore, it is necessary to adaptively augment the training samples based on the current operating status of the power plant to expand sample diversity and improve the model's generalization ability.

[0034] The core of adaptive enhancement lies in dynamically adjusting the enhancement intensity based on the fluctuation level of the current operating conditions, so that the generated enhanced samples match the actual operating conditions. Specifically, it first acquires operating condition information such as the current irradiance, component temperature, ambient temperature, wind speed, power fluctuation amplitude, inverter operating status, power curtailment status, and alarm status of the power plant, and then assesses the degree of fluctuation of the operating conditions. The greater the fluctuation of the operating conditions, the higher the uncertainty of the data distribution, and the stronger the enhancement perturbation needs to be applied to cover a wider range of operating scenarios; conversely, when the operating conditions are relatively stable, a lighter enhancement intensity is used to avoid introducing too much noise.

[0035] In one implementation, the degree of operating condition fluctuation is mapped to an enhancement intensity coefficient. Based on this coefficient, random perturbations of corresponding magnitude are injected into the original sample, including operations such as random scaling, local smoothing, and short-term perturbations. During the enhancement process, it is necessary to maintain the basic physical coupling relationships of the photovoltaic operating data. For example, the usable power should increase synchronously when irradiance rises, the output efficiency should decrease accordingly when the module temperature rises, the current of adjacent strings under the same combiner box should maintain numerical consistency, and a reasonable energy conversion ratio should be maintained between the inverter's input power and output power. By introducing variable consistency constraints during the enhancement process, the physical rationality of the synthesized sample can be ensured, avoiding the generation of false data that violates the equipment's operating mechanism.

[0036] Step 103: Extract features from the enhanced samples using a dynamic residual encoder to obtain fault discrimination features; After adaptive enhancement, discriminative features that can effectively distinguish different fault types need to be extracted from the enhanced samples. Traditional convolutional neural networks apply the same feature transformation to all input samples using fixed convolution kernel parameters, which is difficult to adapt to the distribution differences in photovoltaic monitoring data due to fluctuations in operating conditions, and may lead to insufficient discriminative power in feature extraction. To address this, a dynamic residual encoder is used as the feature extraction network. This encoder introduces a dynamic convolution mechanism into the basic architecture of the residual network, adaptively adjusting the convolution response according to the features of each input sample, thereby improving the targeting and effectiveness of feature extraction.

[0037] The fundamental principle of dynamic convolution is to dynamically generate convolution kernel parameters based on the features of the input samples, rather than using a fixed convolution kernel shared by all samples. In one implementation, an initial convolution mapping is first performed on the input samples to obtain an input feature map. Then, global average pooling and global standard deviation pooling operations are performed on this feature map to generate two pooling vectors describing the global statistical characteristics of the input samples. These two vectors are concatenated and passed through a fully connected layer and activation function to generate a dynamic convolution kernel weight vector. This weight vector is then used to weight and combine a set of preset static convolution kernels to obtain a dynamic convolution kernel that matches the features of the current input sample. Performing a convolution operation between this dynamic convolution kernel and the input feature map yields a convolution response optimized for the characteristics of the current sample.

[0038] The dynamic residual encoder retains the residual connection structure based on the dynamic convolution mentioned above, and superimposes the output and input of the dynamic convolution through an identity mapping to alleviate the gradient vanishing problem in deep networks and promote the effective transfer of features. Through layer-by-layer feature transformation of multi-layer dynamic residual modules, the network can extract fault discrimination features from enhanced samples from low to high level and from local to global. These features can be used to distinguish different types of fault modes such as abnormal string current, abnormal component temperature rise, inverter efficiency decline, abnormal combiner box branch, and abnormal transformer temperature.

[0039] Step 104: Classify and identify faults based on the fault discrimination features, output fault categories and corresponding category probabilities, and perform reliability correction on the category probabilities to generate fault identification results.

[0040] The fault discrimination features extracted by the dynamic residual encoder need to be mapped to specific fault categories and their corresponding probabilities by a classifier to support subsequent operation and maintenance decisions. In the long-tail distribution scenario of photovoltaic faults, the number of normal samples far exceeds the number of fault samples, and there are also significant differences in the number of different types of faults. Directly performing classification training can easily lead to the model being biased towards the majority class and neglecting the feature learning of minority class faults. Therefore, targeted strategies are needed to optimize the class boundaries in the feature space and improve the model's ability to identify minority class faults.

[0041] In one implementation, the distribution structure of the feature space is optimized by unifying contrastive learning constraints. The basic idea of ​​contrastive learning is to bring similar samples closer together in the feature space while pushing away the feature representations of dissimilar samples. For the problem of difficult samples in long-tailed distributions, a dynamic difficult sample mining strategy can be adopted. This involves calculating the distance between the feature representation of each training sample and its class center, as well as the minimum distance to other class centers, thereby assessing the difficulty of the sample and imposing stronger feature consistency constraints on difficult samples. Simultaneously, the classifier's weight vector can be used as a dynamic feature class center to guide sample features toward their class centers, reducing feature drift of minority class samples during training.

[0042] The classification and contrastive learning tasks are trained collaboratively using a multi-task loss function. Learnable coefficients are set in the loss function to automatically balance the contribution weights of different tasks, and are dynamically adjusted based on gradient backpropagation during training. The classification loss term can employ a cross-entropy loss function with angular margin to enhance the discriminative power of class boundaries; the contrastive learning loss term is constructed based on the feature consistency constraint between the enhanced sample and the original sample.

[0043] The class probabilities output by the model need to be reliably corrected before they can be used for operation and maintenance decisions. The reliability of the class probabilities can be assessed using the expected calibration error. When the calibration error exceeds a preset threshold, the class probabilities are corrected using a temperature calibration algorithm to ensure that the corrected probability values ​​are consistent with the actual frequency of fault occurrence. Finally, based on the corrected class probabilities, combined with the equipment importance, historical alarm records, and the impact of power generation losses, operation and maintenance priority recommendations are generated to support operation and maintenance activities such as inspection scheduling, equipment maintenance, and shutdown decisions.

[0044] In some embodiments, the unsupervised cleaning of the monitoring data to remove abnormal noise samples includes: The monitoring data is clustered using a density-based noisy spatial clustering algorithm. Data samples with connected densities are clustered into the same cluster, and isolated samples that cannot be assigned to any cluster are identified as abnormal noise samples and removed. The abnormal noise samples include null samples caused by communication interruption, invalid power generation samples at night, sensor mutation samples, abnormal jumps in irradiance samples, inverter offline samples, and non-fault power curtailment samples.

[0045] Density-based noisy spatial clustering is a typical unsupervised learning method. Its core idea is to identify cluster structures based on the local density distribution of data points in the feature space, rather than relying on a pre-defined number of clusters. This algorithm defines two key parameters: neighborhood radius and minimum density threshold. For any sample point in the dataset, a neighborhood region is defined with that point as the center and the neighborhood radius as the radius. If the number of samples contained within this neighborhood is not less than the minimum density threshold, then that point is marked as a core point. From a core point, all sample points within its neighborhood are density-reachable from that core point. If a sample point is located within the neighborhood of multiple core points, then that point and these core points are density-reachable. Through the transitivity of density reachability, all mutually density-reachable sample points are clustered into the same cluster, while sample points that cannot establish a density-reachability relationship with any core point are identified as noise points.

[0046] In the scenario of cleaning photovoltaic monitoring data, the aforementioned density clustering algorithm can effectively identify abnormal samples that deviate from the distribution of the main operating conditions. Null value samples caused by communication interruptions are characterized by multiple monitoring variables being simultaneously missing or zero, forming isolated points in the feature space far removed from normal data clusters. Invalid nighttime power generation samples still record current or power data during periods of zero or extremely low irradiance, showing a significant deviation from normal nighttime shutdown conditions. Sensor mutation samples are characterized by a step jump in a monitoring variable at adjacent sampling times that does not conform to physical laws, such as a sudden increase in temperature of tens of degrees within seconds. Irradiance anomaly jump samples are characterized by drastic fluctuations in irradiance values ​​under unclouded conditions, inconsistent with concurrent meteorological conditions. Inverter offline samples appear during inverter shutdowns due to faults or maintenance, but current data is still recorded on the string side. Non-fault power curtailment samples are characterized by a step-like decrease in power output, but equipment operating parameters are within the normal range, corresponding to proactive power curtailment behavior caused by dispatch instructions. All of the above-mentioned abnormal samples were identified and removed during the density clustering process because they could not be classified into the main cluster composed of normal data.

[0047] In some embodiments, the feature filtering of the cleaned data includes: Calculate the correlation metric between each candidate feature and the fault category, as well as the redundancy metric between any two candidate features; The feature relevance quotient of each candidate feature is calculated based on the relevance metric and the redundancy metric. Features are then selected in descending order of the feature relevance quotient to form the model input feature set.

[0048] The goal of feature selection is to choose a subset of features from multidimensional monitoring variables that have high discriminative power for fault categories and low redundancy among them, thereby reducing the dimensionality of the model input and improving classification performance. This implementation adopts an information theory-based feature evaluation method, which quantifies the contribution value of each candidate feature by comprehensively considering the correlation between features and fault categories as well as the degree of redundancy among features.

[0049] The correlation between features and fault categories is measured using mutual information. Mutual information captures non-linear statistical dependencies between variables; a higher value indicates that the feature carries more information about the fault category. For each candidate feature, the mutual information value between it and the fault category label is calculated as a measure of its correlation. Redundancy between features is also measured using mutual information. The mutual information value between any two candidate features is calculated. If the mutual information between two features is high, it indicates a significant overlap in the information they carry, and selecting both features would result in redundancy in the input space.

[0050] Based on the obtained relevance and redundancy metrics, the feature relevance quotient (MRQ) for each candidate feature is calculated. The calculation logic for the MRQ is as follows: the correlation between the feature and the fault category is used as the numerator, and the redundancy of the feature relative to the average redundancy of all features in the selected feature set is used as the denominator; the ratio of these two is the MRQ. The physical meaning of this metric is that a high-quality feature should be highly correlated with the fault category while maintaining low information overlap with the selected features. Features are selected sequentially in descending order of MRQ. After each selection, the selected feature set is updated, and the redundancy component of the remaining candidate features is recalculated until the number of selected features meets a preset condition, ultimately forming the model input feature set.

[0051] In some embodiments, performing adaptive enhancement processing on the original sample based on the current operating condition information includes: The irradiance fluctuation range and power fluctuation range of the photovoltaic power station are obtained, and the degree of operating condition fluctuation is calculated and mapped to the enhancement intensity coefficient. Based on the enhancement intensity coefficient, a perturbation matching the current operating condition fluctuation is injected into the original sample, while maintaining the basic physical relationship of the photovoltaic operation data.

[0052] The core mechanism of adaptive enhancement processing lies in dynamically adjusting the intensity of data enhancement based on the fluctuation level of the photovoltaic power plant's current operating conditions, ensuring that the generated enhancement samples can cover the range of operating condition changes that may be encountered in actual operation. Quantitative assessment of operating condition fluctuation relies on real-time monitoring of key operating parameters of the power plant, among which irradiance fluctuation amplitude and power fluctuation amplitude are two core indicators reflecting the dynamic characteristics of the operating conditions. Irradiance fluctuation amplitude is obtained by calculating the standard deviation or coefficient of variation of irradiance sampling values ​​within a certain time window, while power fluctuation amplitude is calculated from the corresponding statistics of the inverter's output power. By weighted and fused, a comprehensive measure of the degree of operating condition fluctuation can be obtained.

[0053] The mapping relationship between the degree of operating condition fluctuation and the enhancement intensity coefficient can be expressed as a monotonically increasing function, such as a linear mapping or a piecewise linear mapping. When the degree of operating condition fluctuation is low, the enhancement intensity coefficient takes a small value, applying a slight perturbation to the original sample; when the degree of operating condition fluctuation is high, the enhancement intensity coefficient increases accordingly, injecting a larger perturbation to simulate data changes under extreme operating conditions. Specific forms of perturbation include random scaling, local smoothing, and short-time segment perturbation. Random scaling multiplies each feature value in the sample by a random factor following a specific distribution; local smoothing applies a moving average or Gaussian filter to the sample sequence to eliminate high-frequency noise; short-time segment perturbation injects a random offset into a local time period of the sample to simulate instantaneous sensor anomalies.

[0054] Maintaining the fundamental physical relationships of photovoltaic (PV) operating data during the enhancement process is a key constraint to ensure the practical significance of the synthesized samples. This means that the enhancement operation must not disrupt the inherent coupling relationships between variables, and the generated samples must conform to the basic operating mechanism of the PV power generation system, thereby ensuring that the enhanced training data is physically reliable.

[0055] In some embodiments, the basic physical relationship for maintaining photovoltaic operating data includes: Maintain a positive correlation between irradiance and available power, a negative correlation between component temperature and output efficiency, a consistent relationship between currents in adjacent strings within the same combiner box, and an energy conversion relationship between inverter input power and output power.

[0056] The physical relationships in photovoltaic operation data originate from the energy conversion mechanism and equipment operation characteristics of photovoltaic power generation systems. These relationships must be maintained during data augmentation to ensure the physical rationality of the synthesized samples.

[0057] The positive correlation between irradiance and available power is the most fundamental physical constraint of photovoltaic power generation systems. Photovoltaic modules convert solar irradiance into direct current (DC) electricity, and their output power is approximately proportional to the received irradiance. During data augmentation, if a positive perturbation is applied to the irradiance characteristics, a corresponding positive perturbation of the power characteristics should also be applied, and vice versa; the direction and magnitude of the perturbations must be synchronized.

[0058] The negative correlation between module temperature and output efficiency stems from the temperature characteristics of semiconductor materials. The photoelectric conversion efficiency of photovoltaic modules decreases as temperature increases, with a typical temperature coefficient of approximately -0.3% to 0.5% per degree Celsius. During enhancement, if the module temperature characteristics are increased, the output power or efficiency characteristics should be correspondingly decreased, and the degree of decrease must comply with the physical constraints of the temperature coefficient.

[0059] The consistency of current between adjacent strings within the same combiner box is based on the circuit characteristics of parallel string operation. Under normal operating conditions, since each string in the same combiner box receives similar irradiation conditions and uses components of the same specifications, its output current should remain similar. During enhancement, when a disturbance is applied to the current of one string, the currents of other strings in the same combiner box should be disturbed with a similar amplitude to maintain the current balance between strings.

[0060] The energy conversion relationship between inverter input power and output power is constrained by inverter conversion efficiency. The inverter converts DC-side input power to AC-side output power; the ratio of the two is the inverter efficiency, typically between 95% and 99%. During the enhancement process, disturbances to DC-side power should maintain a reasonable proportional relationship with disturbances to AC-side power, ensuring that their ratio remains within the normal range of inverter efficiency.

[0061] In some embodiments, the dynamic residual encoder adaptively adjusts the convolution response based on the features of the input samples, including: The enhanced samples are first convolutionally mapped to generate an input feature map, and then two pooling description vectors are generated by global average pooling and global standard deviation pooling, respectively. The two pooling description vectors are concatenated and then passed through a fully connected layer to generate a dynamic convolution kernel weight vector. The preset static convolution kernels are then weighted and combined to generate a dynamic convolution kernel that matches the features of the current input sample. The dynamic convolution kernel is convolved with the input feature map, and then an identity mapping is performed through a residual connection structure. Fault discrimination features are extracted through multi-layer feature transformation.

[0062] The feature extraction process of the dynamic residual encoder begins with the initial convolutional mapping of the augmented samples. After the augmented samples are input into the network as one-dimensional or multi-dimensional temporal data, they first undergo preliminary feature transformation through a standard convolutional layer. This convolutional layer uses a fixed-size convolutional kernel and a preset stride to perform sliding convolution operations on the input data, generating an input feature map containing local temporal patterns. The purpose of the initial convolutional mapping is to transform the original monitoring variables into a feature representation suitable for subsequent dynamic convolutional processing.

[0063] Two different global pooling operations are performed on the input feature map to extract the global statistical features of the samples. Global average pooling calculates the mean of the feature map across all channels, outputting a vector describing the average level of the response across each channel. Global standard deviation pooling calculates the standard deviation of the feature map across all channels, outputting a vector describing the dispersion of the response across each channel. These two pooling operations characterize the feature distribution of the input samples from the perspectives of central tendency and dispersion, respectively. These two pooling description vectors are concatenated along the channel dimension to form a joint description vector twice the length of the original vector.

[0064] The joint description vector is passed sequentially through a fully connected layer and an activation function layer, outputting a dynamic convolutional kernel weight vector. The output dimension of the fully connected layer is consistent with the number of convolutional kernels in the pre-defined static convolutional kernel group. The activation function normalizes each component of the weight vector into non-negative weight coefficients that sum to one. The pre-defined static convolutional kernel group contains multiple fixed convolutional kernels of different sizes or receptive fields, each responsible for capturing a specific type of local pattern. By weighting and summing these static convolutional kernels using the dynamic convolutional kernel weight vector, a dynamic convolutional kernel matching the feature distribution of the current input sample can be generated.

[0065] The generated dynamic convolutional kernels are convolved with the input feature map to obtain the dynamic convolutional response. The residual connection structure adds the output of the dynamic convolution to the original input feature map element-wise, achieving identity mapping superposition, allowing the network to retain the original information while learning the residual mapping. Multiple layers of dynamic residual modules are stacked sequentially, with the output of each layer serving as the input of the next layer, extracting fault discrimination features from low-level texture to high-level semantics layer by layer.

[0066] In some embodiments, the classification and identification based on the fault discrimination features includes: A dynamic hard sample mining strategy is adopted to determine hard samples based on the distance relationship between the feature representation of the training samples and the center of each category, and to impose stronger feature consistency constraints on the hard samples. A classification loss term with angular margin and a contrastive learning loss term based on the feature consistency constraint between enhanced samples and original samples are constructed. The two are weighted and summed by setting learnable coefficients, and the learnable coefficients are automatically updated during training to achieve dynamic loss balance.

[0067] The dynamic hard sample mining strategy aims to identify samples that are difficult to classify correctly near the class boundaries in the feature space and impose stronger training constraints on them to improve the classifier's discriminative ability. This strategy first requires determining the feature centers for each class. Feature centers can be obtained by calculating the mean of the feature representations of samples of the same class, or by using the classifier weight vector as the dynamic feature centers for each class. For each training sample, the Euclidean distance between its feature representation and its class center is calculated; this distance reflects the degree of aggregation between the sample and samples of the same class. Simultaneously, the minimum distance between the feature representation and all other class centers is calculated; this distance reflects the distance between the sample and the nearest out-of-class center. The difficulty score is defined as the ratio of the distance to its class center to the minimum out-of-class center distance. The closer the ratio is to 1, the closer the sample is to the class boundary, and the greater the classification difficulty. Samples with difficulty scores exceeding a preset threshold are identified as hard samples. In subsequent training, stronger feature consistency constraints are imposed on these samples, causing their feature representations to move closer to their class center while moving away from other class centers.

[0068] The loss function for classification tasks uses cross-entropy with angular margin. Traditional cross-entropy loss only requires that the angle between the sample feature and the center of the correct class be less than the angle between the sample feature and the center of the other classes. However, angular margin requires that the angle between the sample feature and the center of the correct class be less than a preset angle threshold, thereby forming a more compact intra-class distribution and a clearer inter-class boundary in the feature space.

[0069] The contrastive learning loss term is constructed based on the feature consistency constraint between the augmented samples and the original samples. For augmented sample pairs generated from the same sample, the feature representations of the two samples should be as close as possible in the feature space; for augmented sample pairs generated from different samples, the feature representations of the two samples should be far apart. The learnable coefficients serve as weight parameters for each loss term and are automatically updated during training via gradient descent, so that the contribution ratios of classification loss and contrastive learning loss to the total loss are dynamically adjusted as the training progresses.

[0070] In some embodiments, the reliability correction of the category probability includes: The model output class probability is divided into multiple probability intervals, and the difference between the average predicted probability in each interval and the actual failure frequency is calculated to obtain the expected calibration error value. When the expected calibration error value exceeds a preset threshold, a temperature calibration algorithm is used to correct the category probability. Based on the corrected category probabilities, combined with the importance of the equipment, historical alarm records, and the impact of power generation loss, maintenance priority suggestions are generated.

[0071] Expected calibration error (ECO) is an important metric for measuring the reliability of the output probabilities of a classification model. Its calculation is based on a statistical test of the consistency between the model's predicted probabilities and the actual observation frequencies. Specifically, the model's output class probability range is divided into several equally wide probability intervals, such as 0 to 0.1, 0.1 to 0.2, etc. For all predicted samples falling within the same probability interval, the average of the model's predicted probabilities within that interval is calculated as the expected accuracy. Simultaneously, the proportion of correctly predicted samples is calculated as the observed accuracy. The expected calibration error is defined as the absolute value of the difference between the expected accuracy and the observed accuracy within each probability interval, weighted by the number of samples within each interval. A smaller error value indicates that the model's output probability value is closer to the true confidence level.

[0072] When the expected calibration error exceeds a preset threshold, it indicates a systematic bias in the model's output probability values, requiring calibration correction. Temperature calibration is a commonly used post-processing correction method. Its basic principle is to introduce a learnable temperature parameter into the logistic value output by the last layer of the model. The corrected probability distribution is obtained by dividing the logistic value by the temperature parameter and then applying it through an activation function. The temperature parameter is optimized by minimizing the negative log-likelihood loss on the validation set, achieving an optimal fit between the corrected probability values ​​and the actual observation frequencies.

[0073] The corrected category probabilities reflect the model's confidence in the reliability of each sample belonging to each fault category and can be directly used for operation and maintenance decision support. By combining the equipment importance rating, statistical analysis of historical alarm records, and assessment of the impact of different faults on power generation loss, the priority of handling various faults can be comprehensively ranked, forming operation and maintenance priority recommendations covering inspection arrangements, equipment maintenance, spare parts preparation, and other aspects.

[0074] Corresponding to the photovoltaic long-tail fault identification method described above, this invention also proposes a photovoltaic long-tail fault identification device. Since the device embodiments of this invention correspond to the method embodiments described above, details not disclosed in the device embodiments can be referred to in the method embodiments described above, and will not be repeated here.

[0075] Figure 2 This is a schematic diagram of the structure of a photovoltaic long-tail fault identification device provided in an embodiment of this disclosure, as shown below. Figure 2 As shown, it includes: The acquisition unit 21 is used to acquire monitoring data of the photovoltaic power station, perform unsupervised cleaning on the monitoring data to remove abnormal noise samples, and perform feature filtering on the cleaned data to obtain the model input feature set. Enhancement unit 22 is used to perform adaptive enhancement processing on the original samples in the model input feature set according to the current operating condition information to generate enhanced samples; Extraction unit 23 is used to extract features from the enhanced sample through a dynamic residual encoder to obtain fault discrimination features; The identification unit 24 is used to classify and identify faults based on the fault discrimination features, output fault categories and corresponding category probabilities, perform reliability correction on the category probabilities, and generate fault identification results.

[0076] Furthermore, in one possible implementation of this disclosure, the acquisition unit 21 is further configured to: The monitoring data is clustered using a density-based noisy spatial clustering algorithm. Data samples with connected densities are clustered into the same cluster, and isolated samples that cannot be assigned to any cluster are identified as abnormal noise samples and removed. The abnormal noise samples include null samples caused by communication interruption, invalid power generation samples at night, sensor mutation samples, abnormal jumps in irradiance samples, inverter offline samples, and non-fault power curtailment samples.

[0077] Furthermore, in one possible implementation of this disclosure, the acquisition unit 21 is further configured to: Calculate the correlation metric between each candidate feature and the fault category, as well as the redundancy metric between any two candidate features; The feature relevance quotient of each candidate feature is calculated based on the relevance metric and the redundancy metric. Features are then selected in descending order of the feature relevance quotient to form the model input feature set.

[0078] Furthermore, in one possible implementation of this disclosure, the enhancement unit 22 is further configured to: The irradiance fluctuation range and power fluctuation range of the photovoltaic power station are obtained, and the degree of operating condition fluctuation is calculated and mapped to the enhancement intensity coefficient. Based on the enhancement intensity coefficient, a perturbation matching the current operating condition fluctuation is injected into the original sample, while maintaining the basic physical relationship of the photovoltaic operation data.

[0079] Furthermore, in one possible implementation of this disclosure, the enhancement unit 22 is further configured to: Maintain a positive correlation between irradiance and available power, a negative correlation between component temperature and output efficiency, a consistent relationship between currents in adjacent strings within the same combiner box, and an energy conversion relationship between inverter input power and output power.

[0080] Furthermore, in one possible implementation of this disclosure, the extraction unit 23 is further configured to: The enhanced samples are first convolutionally mapped to generate an input feature map, and then two pooling description vectors are generated by global average pooling and global standard deviation pooling, respectively. The two pooling description vectors are concatenated and then passed through a fully connected layer to generate a dynamic convolution kernel weight vector. The preset static convolution kernels are then weighted and combined to generate a dynamic convolution kernel that matches the features of the current input sample. The dynamic convolution kernel is convolved with the input feature map, and then an identity mapping is performed through a residual connection structure. Fault discrimination features are extracted through multi-layer feature transformation.

[0081] Furthermore, in one possible implementation of this disclosure, the identification unit 24 is further configured to: A dynamic hard sample mining strategy is adopted to determine hard samples based on the distance relationship between the feature representation of the training samples and the center of each category, and to impose stronger feature consistency constraints on the hard samples. A classification loss term with angular margin and a contrastive learning loss term based on the feature consistency constraint between enhanced samples and original samples are constructed. The two are weighted and summed by setting learnable coefficients, and the learnable coefficients are automatically updated during training to achieve dynamic loss balance.

[0082] Furthermore, in one possible implementation of this disclosure, the identification unit 24 is further configured to: The model output class probability is divided into multiple probability intervals, and the difference between the average predicted probability in each interval and the actual failure frequency is calculated to obtain the expected calibration error value. When the expected calibration error value exceeds a preset threshold, a temperature calibration algorithm is used to correct the category probability. Based on the corrected category probabilities, combined with the importance of the equipment, historical alarm records, and the impact of power generation loss, maintenance priority suggestions are generated.

[0083] It should be noted that the foregoing explanation of the method embodiments also applies to the apparatus of the embodiments of this disclosure, and the principle is the same. Therefore, the embodiments of this disclosure are not limited thereto.

[0084] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0085] Figure 3 A schematic block diagram of an example electronic device 400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0086] like Figure 3 As shown, device 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in ROM (Read-Only Memory) 402 or a computer program loaded from storage unit 408 into RAM (Random Access Memory) 403. RAM 403 may also store various programs and data required for the operation of device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via bus 404. I / O (Input / Output) interface 405 is also connected to bus 404.

[0087] Multiple components in device 400 are connected to I / O interface 405, including: input unit 406, such as keyboard, mouse, etc.; output unit 407, such as various types of monitors, speakers, etc.; storage unit 408, such as disk, optical disk, etc.; and communication unit 409, such as network card, modem, wireless transceiver, etc. Communication unit 409 allows device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0088] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, CPUs (Central Processing Units), GPUs (Graphics Processing Units), various special-purpose AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, DSPs (Digital Signal Processors), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above, such as the photovoltaic long-tail fault identification method. For example, in some embodiments, the photovoltaic long-tail fault identification method can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps of the methods described above can be performed. Alternatively, in other embodiments, the computing unit 401 may be configured to perform the aforementioned photovoltaic long-tail fault identification method by any other suitable means (e.g., by means of firmware).

[0089] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application-Specific Standard Products), SOCs (System-on-Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0090] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0091] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0092] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0093] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include LANs (Local Area Networks), WANs (Wide Area Networks), the Internet, and blockchain networks.

[0094] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service system that addresses the shortcomings of traditional physical hosts and VPS (Virtual Private Server) services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0095] It's important to note that artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0096] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0097] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for identifying photovoltaic long-tail faults, characterized in that, include: Acquire monitoring data of photovoltaic power plants, perform unsupervised cleaning on the monitoring data to remove abnormal noise samples, and perform feature filtering on the cleaned data to obtain the model input feature set; Based on the current operating condition information, adaptive enhancement processing is performed on the original samples in the input feature set of the model to generate enhanced samples; Fault discrimination features are obtained by extracting features from the enhanced samples using a dynamic residual encoder. Based on the fault discrimination features, the system performs classification and identification, outputs the fault category and the corresponding category probability, and performs reliability correction on the category probability to generate the fault identification result.

2. The method according to claim 1, characterized in that, The unsupervised cleaning of the monitoring data to remove abnormal noise samples includes: The monitoring data is clustered using a density-based noisy spatial clustering algorithm. Data samples with connected densities are clustered into the same cluster, and isolated samples that cannot be assigned to any cluster are identified as abnormal noise samples and removed. The abnormal noise samples include null samples caused by communication interruption, invalid power generation samples at night, sensor mutation samples, abnormal jumps in irradiance samples, inverter offline samples, and non-fault power curtailment samples.

3. The method according to claim 1, characterized in that, The feature filtering of the cleaned data includes: Calculate the correlation metric between each candidate feature and the fault category, as well as the redundancy metric between any two candidate features; The feature relevance quotient of each candidate feature is calculated based on the relevance metric and the redundancy metric. Features are then selected in descending order of the feature relevance quotient to form the model input feature set.

4. The method according to claim 1, characterized in that, The adaptive enhancement processing of the original sample based on the current operating condition information includes: The irradiance fluctuation range and power fluctuation range of the photovoltaic power station are obtained, and the degree of operating condition fluctuation is calculated and mapped to the enhancement intensity coefficient. Based on the enhancement intensity coefficient, a perturbation matching the current operating condition fluctuation is injected into the original sample, while maintaining the basic physical relationship of the photovoltaic operation data.

5. The method according to claim 4, characterized in that, The basic physical relationships for maintaining photovoltaic operating data include: Maintain a positive correlation between irradiance and available power, a negative correlation between component temperature and output efficiency, a consistent relationship between currents in adjacent strings within the same combiner box, and an energy conversion relationship between inverter input power and output power.

6. The method according to claim 1, characterized in that, The dynamic residual encoder adaptively adjusts the convolution response based on the features of the input samples, including: The enhanced samples are first convolutionally mapped to generate an input feature map, and then two pooling description vectors are generated by global average pooling and global standard deviation pooling, respectively. The two pooling description vectors are concatenated and then passed through a fully connected layer to generate a dynamic convolution kernel weight vector. The preset static convolution kernels are then weighted and combined to generate a dynamic convolution kernel that matches the features of the current input sample. The dynamic convolution kernel is convolved with the input feature map, and then an identity mapping is performed through a residual connection structure. Fault discrimination features are extracted through multi-layer feature transformation.

7. The method according to claim 1, characterized in that, The classification and identification based on the fault discrimination features includes: A dynamic hard sample mining strategy is adopted to determine hard samples based on the distance relationship between the feature representation of the training samples and the center of each category, and to impose stronger feature consistency constraints on the hard samples. A classification loss term with angular margin and a contrastive learning loss term based on the feature consistency constraint between enhanced samples and original samples are constructed. The two are weighted and summed by setting learnable coefficients, and the learnable coefficients are automatically updated during training to achieve dynamic loss balance.

8. The method according to claim 1, characterized in that, The reliability correction of the category probability includes: The model output class probability is divided into multiple probability intervals, and the difference between the average predicted probability in each interval and the actual failure frequency is calculated to obtain the expected calibration error value. When the expected calibration error value exceeds a preset threshold, a temperature calibration algorithm is used to correct the category probability. Based on the corrected category probabilities, combined with the importance of the equipment, historical alarm records, and the impact of power generation loss, maintenance priority suggestions are generated.

9. A photovoltaic long-tail fault identification device, characterized in that, include: The acquisition unit is used to acquire monitoring data of photovoltaic power plants, perform unsupervised cleaning on the monitoring data to remove abnormal noise samples, and perform feature filtering on the cleaned data to obtain the model input feature set. The enhancement unit is used to perform adaptive enhancement processing on the original samples in the model input feature set according to the current operating condition information, and generate enhanced samples; The extraction unit is used to extract features from the enhanced sample through a dynamic residual encoder to obtain fault discrimination features; The identification unit is used to classify and identify faults based on the fault discrimination features, output fault categories and corresponding category probabilities, perform reliability correction on the category probabilities, and generate fault identification results.

10. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

Images