Photovoltaic system fault data closed loop enhancement method

Abstract

This invention discloses a closed-loop enhancement method for photovoltaic system fault data. Addressing the poor generalization issues in fault diagnosis under conditions of small sample size and class imbalance, the method first performs STC correction and UMAP feature extraction on the photovoltaic I-V curve, constructs a heterogeneous GAN generator pool, and uses MOPSO to determine the sampling weights of each sub-model. Then, based on the misclassification rate, classification confidence, and sample size of the downstream fault classifier, class difficulty weights are constructed, dynamically updating the generation quotas for various fault types and the GAN sampling ratio. The final enhanced dataset is obtained through a dual-gating screening process using feature space distance criteria and marginal performance gain criteria, along with an early stopping mechanism, to improve the accuracy and generalization ability of photovoltaic fault diagnosis. The enhanced sample set generated by this invention can perfectly simulate the subtle electrical feature fluctuations that are easily overlooked in early faults such as latent degradation and shadow occlusion under different degrees of shading, breaking through the data silo bottleneck in the application of efficient deep learning algorithms such as CNN / GRU in photovoltaic fields.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data augmentation technology, specifically to a closed-loop augmentation method for photovoltaic system fault data based on generative adversarial networks (GANs). Background Technology

[0002] As photovoltaic (PV) power plants increasingly constitute a significant portion of the global energy mix, these devices operate under constantly changing environmental conditions. Low-probability events such as localized shading, short circuits, open circuits, aging, and their combined failures occur frequently. Therefore, fault diagnosis facilities are necessary to ensure the reliability of PV power plants. Traditional fault diagnosis models suffer from severe data silos and sample scarcity due to equipment safety limitations, high costs associated with inducing faults, and the difficulty in obtaining sufficient fault sample data covering all operating conditions. Researchers have proposed using deep learning and machine learning methods for intelligent fault diagnosis of PV systems. However, the training of such intelligent diagnostic models heavily relies on large-scale, class-balanced sample data. Consequently, some scholars have proposed methods such as simple oversampling or introducing a single generative adversarial network (GAN) to generate fault sample data.

[0003] Reference [1] proposes a photovoltaic fault diagnosis method using deep transfer learning under small sample conditions to deal with the problem of sample scarcity; Reference [2] proposes a photovoltaic array fault diagnosis method based on internal parameters to improve classification accuracy; Reference [3] provides a comprehensive review of the latest progress in photovoltaic system fault diagnosis technology and discusses the advantages and limitations of various intelligent diagnosis technologies. In the Chinese patent application number 2025104243570 entitled "A photovoltaic array small sample fault diagnosis method based on neural network", a scheme is proposed to extract features using neural network and complete small sample diagnosis. However, in the data augmentation stage, it mainly relies on traditional data augmentation methods or basic single generation models, and fails to fundamentally solve the problems of diversity and high fidelity of complex fault data distribution.

[0004] In fact, due to the highly nonlinear and multimodal complex characteristics of photovoltaic IV curve fault data, the above-mentioned single generative adversarial networks cannot avoid the problems of training instability and mode collapse caused by their network characteristics when processing such data. The generated fault samples are often not diverse enough, cannot truly cover the feature boundaries of composite faults, and are difficult to support the high-precision lightweight diagnosis requirements.

[0005] [1] Cai Huangyu, Chen Zhicong, Wu Lijun, et al. Photovoltaic fault diagnosis using deep transfer learning under small sample size[J]. Journal of Fuzhou University (Natural Science Edition), 2025, 53(05):552-559. [2] Zheng Xiubin, Chen Jun, Liu Fei. Fault diagnosis of photovoltaic array based on internal parameters [J / OL]. Control Engineering, 1-10 [2026-01-04]. 20240003. [3] Yang B, Zheng R, Han Y, et al. Recent Advances in Fault DiagnosisTechniques for Photovoltaic Systems: A Critical Review[J]. Protection andControl of Modern Power Systems, 2024, 9(3): 36-59. Summary of the Invention

[0006] The purpose of this invention is to address the problems of pattern collapse, insufficient distribution fidelity, and disconnect between enhancement and diagnosis processes in existing single generation models for photovoltaic fault data generation. This invention provides a closed-loop enhancement method for photovoltaic system fault data in scenarios with small samples and class imbalance. This method achieves high-quality generation of fault samples through an iterative update mechanism driven by feedback from the downstream fault classifier and a dual-gating screening strategy.

[0007] The photovoltaic system fault data closed-loop enhancement method specifically includes the following steps: S1. Obtain historical operating data of the photovoltaic array, including normal state data and fault sample data under multiple fault categories; S2. Preprocess the historical operating data, including: calibrating the IV curves under different irradiance and temperature conditions based on the photovoltaic module physical model, and performing dimensionality reduction and feature extraction on the calibrated IV curve data to construct a low-dimensional feature space characterizing the distribution of each fault category. S3. Construct a heterogeneous generator pool based on the low-dimensional feature space. The heterogeneous generator pool includes multiple generative adversarial network sub-models with different network structures or loss functions. Each generative adversarial network sub-model is trained independently using sample data of each fault category in the low-dimensional feature space. S4. For each fault category, the sampling weights of each generative adversarial network sub-model in the heterogeneous generator pool are optimized using a multi-objective particle swarm optimization algorithm. The maximum mean difference between generated samples and real samples, coverage, Wasserstein distance and pattern score are used as multi-objective evaluation indicators to obtain the initial global optimal sampling weights corresponding to each fault category, and an initial augmented sample set is generated accordingly. S5. Merge the initial augmented sample set with the real sample set to form an initial expanded fault diagnosis dataset. Train a downstream fault classifier based on the initial expanded fault diagnosis dataset, and use the trained downstream fault classifier to infer the validation set to obtain the misclassification rate, classification confidence score and category sample size for each fault category, wherein the category sample size is the total number of samples of the fault category in the current expanded fault diagnosis dataset. S6. Construct category difficulty weights based on the misclassification rate, classification confidence score, and category sample size of each fault category obtained in step S5, and dynamically allocate the target generation quota for each fault category in the next round according to the category difficulty weights; at the same time, combine the initial global optimal sampling weights corresponding to each fault category obtained in step S4 with the effective contribution of each generative adversarial network sub-model to the corresponding fault category in the current round, and update the local sampling ratio of each generative adversarial network sub-model for each fault category in the next round. S7. Based on the updated target quota and local sampling ratio, generate the next round of candidate enhancement samples, and perform dual-gated screening on the candidate enhancement samples. The dual-gated screening includes: The first gate is to determine whether the feature space distance criterion between the candidate enhancement sample and the distribution of the corresponding fault category real sample in the low-dimensional feature space meets the distribution tolerance threshold. The second gating process involves incorporating candidate enhancement samples from the first gating process into the current expanded fault diagnosis dataset for incremental evaluation to determine whether the overall macro-average F1 score gain of the validation set meets the minimum performance gain threshold. A candidate augmentation sample is only incorporated into the current augmented fault diagnosis dataset if it passes both the first and second gating. S8. After each round of candidate enhancement samples are incorporated, check whether the early stopping condition is met. If the macro-average F1 score gain for a consecutive preset number of rounds does not reach the minimum performance gain threshold, or the feature center offset distance of the expanded fault diagnosis dataset caused by the new samples is less than the preset convergence threshold, then stop the iteration and output the final enhancement dataset and the corresponding trained downstream fault classifier. Otherwise, return to step S5 to continue iterating, thereby forming a closed-loop enhancement process for photovoltaic system fault data based on downstream fault classification feedback.

[0008] Specifically, in step S2, when performing standard test condition calibration on the IV curve based on the photovoltaic module physical model, the current calibration calculation formula is as follows: The voltage comprehensive compensation correction formula is: In the formula, and These are the currents before and after correction, respectively; and These are the irradiance before and after correction, respectively; and These are the component temperatures before and after calibration, respectively. The relative temperature coefficient of the short-circuit current; and These are the voltages before and after correction, respectively; This is the equivalent series resistance; This is the curve correction factor; This represents the absolute temperature coefficient of the open-circuit voltage of a photovoltaic module.

[0009] Specifically, in the multi-objective particle swarm optimization algorithm described in step S4, the particles... In the Speed ​​at the next iteration and location The update formula is: In the formula, To balance the inertia weights of global search and local search; and These are cognitive learning factors and social learning factors, respectively. and It is a vector of random numbers uniformly distributed in the interval [0,1]. For particles The historical best position; This is the current globally optimal position.

[0010] Specifically, the multiple generative adversarial network sub-models in the heterogeneous generator pool mentioned in step S3 include: Wasserstein generative adversarial network WGAN-GP with gradient penalty, conditional table generative adversarial network CTGAN, deep convolutional generative adversarial network DCGAN, least squares generative adversarial network LSGAN, and mutual information maximization generative adversarial network InfoGAN.

[0011] Specifically, in step S5, the categories Classification confidence score The average information entropy is calculated by predicting the probability distribution of samples belonging to this category in the validation set, and then inverted after normalization. In the formula, To verify that the real labels in the set are categories The sample set, This represents the number of samples in the sample set. This represents the total number of fault categories. Samples output by the downstream fault classifier Category The soft probability.

[0012] Specifically, the fault category in step S6 Category Difficulty Weight Determined according to the following formula: Among them, categories misclassification rate for: In the formula, For category The total number of samples in the current expanded fault diagnosis dataset, This represents the maximum total number of samples across all categories. , , Let be a normalized non-negative adjustment coefficient, and satisfy . ; This is an indicator function.

[0013] Specifically, in step S6, the next iteration will target the fault category. Target generation quota Dynamic allocation as follows: In the formula, The quota amount generated per transaction is based on this. and These are the absolute lower limit and absolute upper limit of the number of items generated in a single run, respectively.

[0014] Specifically, in step S6, the next round of generative adversarial network sub-models targets the fault category. The local sampling ratio is updated as follows: In the formula, The output of step S4 The globally optimal sampling weights of a generative adversarial network sub-model; For the first The generative adversarial network sub-model in the current... Fault categories in round iteration Effective contribution score; For the first In the first iteration Each generative adversarial network sub-model targets the fault category. The local sampling ratio; For the first In the first iteration Each generative adversarial network sub-model targets the fault category. The actual number generated; To generate the total number of adversarial network sub-models.

[0015] Specifically, in step S7, the first gated judgment feature space distance criterion test satisfies: The second gating method incorporates candidate augmentation samples that have passed the feature space distance criterion into the current expanded fault diagnosis dataset to form a temporary dataset. Under the same training configuration, the overall macro-average F1 score of the corresponding validation set is obtained. And calculate: If and only if When the marginal performance gain test is passed, it is determined that the test is passed. In the formula, Fault Category The true sample distribution For the first Fault categories in round iteration The distribution of candidate augmented samples, and The preset distribution tolerance threshold, The minimum performance gain threshold, For the first The expanded fault diagnosis dataset after rounds of iteration.

[0016] Specifically, the early stopping conditions mentioned in step S8 include: The early stop mechanism is automatically triggered when any of the following conditions are met: continuous All iterations satisfy ,in ; Or satisfy ; In the formula, and The first Wheel and First Round-robin expansion of the feature center vector of the fault diagnosis dataset, The preset convergence threshold is used; After triggering the early stopping mechanism, the final augmented dataset and the corresponding trained downstream fault classifier are output.

[0017] The advantages of this invention are: 1. Strong physically interpretable data standardization capability. Irradiance and temperature drift are compensated by the STC correction formula of the IV curve, environmental disturbances are removed a priori, and then UMAP feature mapping is performed to ensure the consistency of the data input to the generator at the physical and statistical layers, reducing the transmission of spurious correlation noise to the subsequent generation and diagnostic links.

[0018] 2. Heterogeneous generation and multi-objective optimization work together to improve the quality of initial enhancements. A heterogeneous generator pool is constructed based on WGAN-GP, CTGAN, DCGAN, LSGAN, and InfoGAN, and MOPSO optimization is performed in combination with multi-dimensional evaluation metrics. The optimal candidate set and corresponding sampling weights are determined by comprehensive scoring to avoid the performance limitations of a single model on specific fault categories.

[0019] 3. A closed-loop enhancement mechanism of classifier feedback + generator update is proposed. For category k, the classification confidence is calculated first, then the category difficulty weight is constructed, and the generation quota for the next round is dynamically allocated; at the same time, the local sampling ratio and actual generation quantity of each GAN sub-model are updated to realize an adaptive generation strategy that prioritizes enhancement based on difficult categories and prioritizes allocation based on effective models.

[0020] 4. A dual-gating and early stopping mechanism ensures usable gain, rather than blindly incrementally increasing the sample size. Early stopping is triggered only when a candidate sample meets the condition of K consecutive rounds of low gain, thus avoiding overfitting and ineffective iterations. Attached Figure Description

[0021] Figure 1 This is a flowchart of the overall method of the present invention.

[0022] Figure 2 This is a schematic diagram of the multidimensional index evaluation distribution of each GAN sub-model under different fault categories, used to illustrate the complementarity of heterogeneous generators.

[0023] Figure 3 yes Figure 1 The figure shows the comparison of the diagnostic performance of the downstream fault classifier before and after enhancement after the closed-loop enhancement process converges. Detailed Implementation

[0024] Based on the initial globally optimal sampling weights, this invention introduces an iterative update mechanism driven by downstream classification feedback. It constructs category difficulty weights by jointly considering classification confidence, misclassification rate, and sample size, and further completes the adaptive adjustment of target generation quota and local sampling ratio, enabling the generation distribution and diagnostic performance to converge synergistically, thereby breaking through the data bottleneck of deep learning fault diagnosis.

[0025] The overall process of this invention is as follows: Figure 1As shown in the accompanying drawings and embodiments, the present invention will now be described in detail.

[0026] S1. Obtain historical operating data of photovoltaic arrays under real-world, limited conditions, including normal state data and fault sample data under different fault types.

[0027] In this embodiment, a high-temporal-resolution operation record of a 10kW photovoltaic power station based on a SCADA communication architecture was collected for one year. In addition to the massive amount of normal data, an extremely small number of samples (50 sets of short-circuit faults, 30 sets of open-circuit faults, 40 sets of local shading faults, 20 sets of aging faults, and fewer than 10 sets of composite fault samples) were selected to construct an extremely imbalanced original training set, with the number of fault samples accounting for no more than 10% of the total data volume.

[0028] S2. Physics-driven data preprocessing and feature extraction.

[0029] Because the solar irradiance and component surface temperature in actual power plant environments fluctuate drastically, multiple IV curves for the same fault type exhibit significant vertical and horizontal shifts, stretching, and overall dispersion, masking the true fault characteristics due to severe environmental noise. All scattered, discrete IV trajectories were precisely and uniformly mapped and flattened to the STC standard test environment (25°C, 1000 W / m²). 2 On the reference plane of ).

[0030] S21. To shield against the influence of variable weather conditions, based on the single-curve equation of the physical model, all collected discrete IV trajectories are precisely mapped to the STC standard test environment. The following derivation and calculations are performed to clean and standardize all the original non-standard data: (1) Current correction calculation: In the formula, and These are the current measured under actual operating conditions and the current corrected to the STC reference, respectively. This represents the actual irradiance at the time of measurement. STC standard irradiance; The actual temperature of the component at the time of measurement. STC standard temperature; This represents the relative temperature coefficient of the short-circuit current of a photovoltaic module.

[0031] (2) Voltage comprehensive compensation correction: Because the output voltage of photovoltaic modules is extremely sensitive to changes in temperature and current, the corrected standard voltage... It cannot be scaled up directly; instead, it needs to be based on the actual measured voltage. The temperature drift compensation term and internal conduction loss are compensated through multi-dimensional nonlinear comprehensive compensation. The specific calculation formula is as follows: In the formula, The actual voltage at the time of measurement. To correct the voltage to the STC reference; For including equivalent series resistance Compensation for internal voltage drop caused by current correction amplitude difference; The curve correction factor compensation term characterizes the nonlinear thermal resistance properties of the curve, where This is the curve correction factor; This is the absolute temperature coefficient of the open-circuit voltage of a photovoltaic module, used to correct for the inherent shift in the built-in potential of the semiconductor PN junction at different temperatures. This represents the overall voltage reference drift caused by the pure temperature difference.

[0032] S22. For the massive, standardized, nonlinear photovoltaic time-point data, directly inputting it into a GAN would result in severe dimensionality curse or even gradient vanishing. This embodiment uses the UMAP algorithm for feature dimensionality reduction. This algorithm profoundly preserves the unique high-dimensional global and local manifold structure of photovoltaic faults.

[0033] First, a fuzzy topological structure representation is constructed in high-dimensional space using Riemannian geometry theory. Then, the two data points are described by solving the high-dimensional conditional probability. and Connection strength between: Introducing the local connectivity distance parameter ρ i Nearest neighbor distance is used to ensure that each data point is connected to at least one neighbor, even in extremely sparse data regions. Then, symmetry is performed to construct the joint probability. .

[0034] Subsequently, low-dimensional coordinates are initialized, and a variant of the Student T-distribution with a flatter tail is used to design two points in the low-dimensional space (target feature space). and The probability distribution: hyperparameters and It controls the visual appearance and density compactness of the mapped data clusters.

[0035] Step S22 not only continues the high-fidelity characteristics of the cleaned data in step S21, but also eliminates the redundancy of data dimensions, and finally constructs a feature space that retains the global and local structure, providing a strong discriminative foundation for subsequent efficient learning and generation of fault samples.

[0036] S3. Construct an integrated generative adversarial network framework, including n independent GAN sub-models, n≥5. Each GAN sub-model has a different network architecture or loss function characteristics. Train each GAN sub-model independently using the feature space data extracted in step S2 to obtain a heterogeneous generator pool.

[0037] In this embodiment, a heterogeneous generator pool consisting of five independent pre-trained generators is constructed, including: Wasserstein Generative Adversarial Network (WGAN-GP) with gradient penalty, Conditional Table Generative Adversarial Network (CTGAN), Deep Convolutional Generative Adversarial Network (DCGAN), Least Squares Generative Adversarial Network (LSGAN), and Mutual Information Maximization Generative Adversarial Network (InfoGAN). These five sub-models are independently pre-trained based on the local real small sample set preprocessed in step S2, enabling them to possess different fitting capabilities for specific fault morphologies, and initial weight coefficients are assigned to them.

[0038] S4. Use the MOPSO algorithm (Multi-Objective Particle Swarm Optimization) to perform weight optimization calculations on the above heterogeneous generator pool.

[0039] With the largest mean difference Coverage Wasserstein distance and mode score A multi-objective evaluation function is constructed by combining the following elements; where and These represent the probability distributions of real fault sample data and generated fault sample data, respectively; the multi-objective particle swarm optimization algorithm iteratively updates the velocity and position of particles, outputting a set of optimal sampling weight vectors in the Pareto optimal solution set. .

[0040] In the embodiment, a population containing multiple particles is initialized, with each particle representing a set of sampling weights for 5 GANs ( ), enter the iteration process: S41 Candidate Data Generation: In the current iteration cycle, the weight distribution of each particle is used to call the above 5 types of pre-trained sub-models to generate a set of candidate datasets in proportion. .

[0041] S42 Multidimensional Evaluation Metrics: For the currently generated candidate augmentation samples, four core distribution quantification metrics are calculated: the maximum mean difference (MMD) between real and generated samples, the range of generated cluster boundaries (coverage C), the transmission cost space (Wasserstein distance WD), and the conditional class adversarial stability (mode score MS). The optimization aims to minimize MMD and WD, and maximize C and MS.

[0042] The multi-objective evaluation function formula includes: Maximum mean difference is used to measure the similarity between the feature distributions of the real and generated data. In the formula, and These represent the probability distributions of real and generated photovoltaic fault data, respectively. This represents the kernel function that measures the similarity of the eigenvectors of IV curves. Represents the mathematical expectation; For real samples, To generate samples.

[0043] Coverage, used to evaluate the extent to which the generated data covers real-world fault feature clusters: In the formula, Represents the first in the feature space of real fault data Cluster center, Total number of fault types; Indicates sample To the cluster center The distance; Distance threshold for a specific fault type. This indicates the number of elements in the set.

[0044] Wasserstein distance, used to measure the transmission cost between two distributions: In the formula, Indicates the search for the supremum. Representation function It satisfies the 1-Lipschitz continuity condition.

[0045] Pattern score, used to simultaneously evaluate the quality of the generative model and the degree of pattern collapse: In the formula, Indicates the Kullback-Leibler divergence; For a given generated sample Conditional class probability; The edge distribution of the true labels; To predict the marginal distribution of labels for generated samples.

[0046] Combined with appendix Figure 2 The diagram shown illustrates the evaluation distribution of each GAN sub-model in this invention under the aforementioned multi-dimensional evaluation indicators when generating data for different photovoltaic fault categories. Figure 2 The comparison shows that no single GAN model can achieve absolute superiority across all metrics and all fault types. For example, DCGAN may have a high pattern score when generating faults with specific two-dimensional features, but it is prone to collapse in terms of coverage; while InfoGAN achieves a very small Wasserstein distance in mining the decoupled features of compound faults, but its... The difference may be too high. Figure 2 The study clearly reveals the inherent "performance barriers" and "complementary advantages" of each heterogeneous sub-model when generating specific photovoltaic fault data, thus fully demonstrating the strong necessity of constructing an integrated generative adversarial network framework and optimizing weights.

[0047] S43 Particle Swarm Optimization Update: Based on the calculated multi-objective evaluation function, the particle swarm is non-dominated and sorted, the velocity and position vectors of all particles are updated, and the contribution configuration of each sub-model is adjusted. S44 Iteration Termination and Weight Output: When the iteration terminates, the multi-objective particle swarm optimization algorithm outputs a Pareto optimal solution set. To determine the unique final weights, multiple candidate datasets are generated using different weight configurations from the Pareto solution set. The optimal generated sample and its weight are determined by finding the maximum value using the extreme value normalization comprehensive scoring formula. In the formula, , , , The set preference weights for each indicator; For candidate datasets The normalization function for each index to reach its optimal extreme value is selected as the maximum or minimum normalization function; Candidate datasets that reach the maximum value As the optimal generation result, the corresponding particle positions are the optimal sampling weights for each GAN sub-model. An initial augmentation sample set is generated based on this initial set of optimal sampling weights.

[0048] S5. Downstream classification evaluation on the current augmented dataset.

[0049] Based on the initial enhanced sample set output in step S4 or the expanded fault diagnosis dataset obtained in the previous iteration. The training set is divided according to a preset ratio. With the validation set Using the training set Training the downstream fault classifier and using the validation set The effectiveness of the current augmented dataset is evaluated.

[0050] In this embodiment, the downstream fault classifier is preferably a CNN / GRU hybrid diagnostic network, in which the convolutional neural network is used to extract local morphological mutation information from the IV trajectory and its low-dimensional embedding features, and the gated recurrent unit network is used to model the temporal correlation and contextual dependence in the fault evolution process. However, it is not limited to this and can also be replaced by other supervised classification models that can output multi-class soft probabilities.

[0051] S51 states that for the true labels in the validation set, the category is... The sample set is defined as follows: Where C represents the total number of fault categories. The classifier applies this value to any sample. Output the soft probability distribution of each category. To further measure the current augmented dataset in categories The difficulty of diagnosis is assessed, and the misclassification rate of this category on the validation set is calculated: In the formula, This is the indicator function. When the misclassification rate of a certain fault category is high, it indicates that the current augmented dataset does not adequately cover the features of that fault category, and the generation priority of this category should be increased in subsequent iterations.

[0052] After completing the validation set inference, S52 further calculates the classification confidence score for each fault category. For each category... Its classification confidence score The average information entropy is calculated by predicting the probability distribution of samples belonging to this category in the validation set, and then inverted after normalization. In the formula, To verify that the real labels in the set are categories The number of samples, This is used to normalize the entropy value. As shown in the above formula, when the classifier classifies categories... When the sample yields a sharper, more concentrated predictive distribution, its average information entropy decreases, corresponding to... If the error rate is high, the error rate will increase; conversely, if the samples of this category are easily confused with other fault categories, the prediction distribution will be more dispersed, and the average information entropy will increase. Decrease.

[0053] S6. Update category difficulty weights, generation quotas, and local sampling ratios.

[0054] In obtaining the misclassification rate Classification confidence score After obtaining information on the category sample size (i.e., the total number of samples of that category in the current expanded fault diagnosis dataset), fault categories are constructed. Category Difficulty Weight : In the formula, To expand the current fault diagnosis dataset Medium category Total sample size This represents the maximum total number of samples across all categories. , , The normalized nonnegative adjustment coefficient satisfies: .

[0055] The above formulas comprehensively characterize the augmentation difficulty of fault categories from three dimensions: the error-proneness of classification results, the uncertainty of classifier decisions, and the relative scarcity of category samples. If a category simultaneously satisfies high misclassification rate, low classification confidence, and significantly smaller sample size than the mainstream categories, then its... It will increase significantly.

[0056] Regarding the first Categories in round iteration Target generation quota The following dynamic allocation formula is adopted: In the formula, Based on the quota generated per transaction, and These represent the absolute lower and upper limits for the number of faults generated in a single run, respectively. This ensures that: when the category difficulty weight of a certain fault type is higher than the average level of all categories, its generation quota in the next round will be amplified; when a certain fault type has been sufficiently learned, its quota will automatically decrease, but will still not be lower than... .

[0057] To align with the initial globally optimal sampling weights obtained in step S4, each GAN sub-model... The next round of local sampling ratio will be updated as follows: In the formula, The output of step S4 The initial globally optimal sampling weights corresponding to each GAN sub-model For this sub-model in the current wheel pair category Effective contribution score For the first Wheel of Life Each GAN sub-model targets the category. The local sampling ratio, For the first Wheel of Life Each GAN sub-model targets the category. The actual number generated.

[0058] S7. Candidate Sample Generation and Dual-Gated Screening.

[0059] Quotas are generated based on the stated objectives. With local sampling ratio The GAN sub-models are invoked to generate the next round of candidate augmentation sample groups. The candidate augmentation samples generated in the next round need to undergo dual-gating screening before being incorporated into the training set.

[0060] Let's define the category The true sample distribution is denoted as The candidate sample group generated in the current round is denoted as The first gate is the feature space distance criterion test, which must simultaneously satisfy: In the formula, and These are the preset distribution tolerance thresholds. Only when a candidate sample is sufficiently close to the real fault distribution in both the kernel distribution difference and transmission distance dimensions is it considered to have basic physical reliability and statistical fidelity.

[0061] After passing the first gating, the dataset proceeds to the second gating, namely the marginal performance gain test. Let's assume that the candidate samples passing the first gating are merged into the current expanded fault diagnosis dataset to obtain a temporary dataset. And under the same training configuration, the corresponding macro-average F1 score of the validation set is obtained. The marginal performance gain test is then defined as follows: If and only if If a candidate sample fails the first gating, it is discarded or regenerated; if it passes the first gating but fails the second gating, it is not merged into the final training set. Figure 1 The paper provides a closed-loop logic diagram for local sampling update, generation, dual-gated filtering, and merging.

[0062] S8. Early stopping mechanism, final output and generalization verification.

[0063] After each round of dual-gated screening and the formation of a new augmented dataset, it is further checked whether the early stopping condition is met. Let the augmented fault diagnosis dataset after merging in round t+1 be... Its feature center vector is denoted as The feature center offset distance caused by the addition of a new sample is expressed as: The early stop mechanism is automatically triggered when any of the following conditions are met: That is, the minimum gain requirement has not been reached after K consecutive iterations; Or satisfy ; In the formula, This is the preset convergence threshold. When the above condition is met, it indicates that continuing to add samples is no longer sufficient to significantly improve the generalization ability of the downstream fault classifier, or that the perturbation of the overall feature distribution by the new samples has stabilized. At this point, the final augmented dataset is output. and the corresponding downstream fault classifier that has been trained. .

[0064] If the early termination condition is not met, then As input for the next iteration, return to step S5 to continue the feedback loop enhancement of classification evaluation, difficulty update, quota update, local sampling update, generation, and dual-gated screening. Figure 1 The looping arrow on the left corresponds to the continued iteration process described above.

[0065] The technical solution of this invention has four key points: The first step involves organically integrating data standardization preprocessing, heterogeneous multi-GAN ensemble modeling, and MOPSO multi-objective collaborative optimization. First, a stable and separable fault feature space is constructed through IV curve STC correction and UMAP feature extraction. Second, a heterogeneous generator pool consisting of WGAN-GP, CTGAN, DCGAN, LSGAN, and InfoGAN is built. Finally, multi-objective optimization is performed based on maximum mean difference, coverage, Wasserstein distance, and pattern score, outputting the initial globally optimal sampling weights and initial augmented sample sets for each fault category. This core approach addresses the issues of insufficient generation quality and significant single-model bias under conditions of small samples and complex distributions.

[0066] Second: In each iteration, the evaluation results of the downstream fault classifier on the current expanded dataset are introduced. Taking into account the misclassification situation, classification confidence and class sample size, the class difficulty ranking is updated, and the target generation quota and generation priority of each fault class in the next round are dynamically adjusted accordingly to realize the enhancement strategy of prioritizing difficult samples and weak class.

[0067] Third: While retaining the initial global optimal sampling weights in step S4, the local sampling ratio and actual generation quantity of each sub-model are iteratively updated by combining the effective contribution of each GAN sub-model to different fault categories in the current round. This allows the generation resources to adaptively balance between the historical optimal configuration and the current diagnostic requirements, thereby improving the coverage quality of composite faults and long-tail categories.

[0068] Fourth: Before merging candidate samples, they are sequentially subjected to distribution consistency gating and performance gain gating, retaining only samples that make a substantial contribution to diagnostic performance; when the gain is insufficient for several consecutive rounds or the feature distribution changes tend to stabilize, an early stop is triggered, and the final augmented dataset and downstream fault classifier are output, thus forming a controllable closed-loop process of evaluation, updating, generation, screening and convergence.

[0069] when Figure 1 After the feedback closed-loop enhancement process shown satisfies the early stopping condition and outputs the final enhanced dataset and downstream fault classifier, it is further generalized to a real power plant for verification. The verification results are as follows: Figure 3 As shown. Figure 3The figure shows a comparison of the accuracy of the downstream CNN / GRU fault classifier before and after data augmentation using the method of this invention in actual power plant deployment verification. When conducting online generalization tests on a SCADA platform for actual grid-connected photovoltaic power plants, if the CNN / GRU model is directly trained using only extremely limited small samples, its diagnostic accuracy for single faults such as short circuits and open circuits is only around 70%, while the classification accuracy for compound faults even drops to 59%, completely failing to meet engineering requirements. However, when the downstream CNN / GRU classifier receives high-fidelity augmented samples output by this invention for mixed training, its overall diagnostic accuracy can be stabilized above 98%, the recall rate for compound faults can reach above 95%, and the false alarm rate remains at a low level, effectively reducing on-site operation and maintenance costs and improving system robustness.

[0070] The aforementioned realistic and significantly different indicator comparisons powerfully demonstrate that the enhanced sample set generated by this invention can perfectly simulate subtle electrical characteristic fluctuations that are easily overlooked in early faults such as latent degradation and shadow occlusion under different degrees of shading, completely breaking through the data silo bottleneck of efficient deep learning algorithms such as CNN / GRU in photovoltaic field applications. Compared to data augmentation methods using a single network, the downstream intelligent diagnostic system trained with the deep enhancement support of this model achieves a revolutionary improvement in the generalization diagnostic success rate under complex climate flow conditions, endowing the entire diagnostic system with extremely high engineering robustness and industrialization value.

[0071] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for closed-loop enhancement of fault data in a photovoltaic system, characterized in that, Includes the following steps: S1. Obtain historical operating data of the photovoltaic array, including normal state data and fault sample data under multiple fault categories; S2. Preprocess the historical operating data, including: calibrating the IV curves under different irradiance and temperature conditions based on the photovoltaic module physical model, and performing dimensionality reduction and feature extraction on the calibrated IV curve data to construct a low-dimensional feature space characterizing the distribution of each fault category. S3. Construct a heterogeneous generator pool based on the low-dimensional feature space. The heterogeneous generator pool includes multiple generative adversarial network sub-models with different network structures or loss functions. Each generative adversarial network sub-model is trained independently using sample data of each fault category in the low-dimensional feature space. S4. For each fault category, the sampling weights of each generative adversarial network sub-model in the heterogeneous generator pool are optimized using a multi-objective particle swarm optimization algorithm. The maximum mean difference between generated samples and real samples, coverage, Wasserstein distance and pattern score are used as multi-objective evaluation indicators to obtain the initial global optimal sampling weights corresponding to each fault category, and an initial augmented sample set is generated accordingly. S5. Merge the initial augmented sample set with the real sample set to form an initial expanded fault diagnosis dataset. Train a downstream fault classifier based on the initial expanded fault diagnosis dataset, and use the trained downstream fault classifier to infer the validation set to obtain the misclassification rate, classification confidence score and category sample size for each fault category, wherein the category sample size is the total number of samples of the fault category in the current expanded fault diagnosis dataset. S6. Construct category difficulty weights based on the misclassification rate, classification confidence score, and category sample size of each fault category obtained in step S5, and dynamically allocate the target generation quota for each fault category in the next round according to the category difficulty weights; at the same time, combine the initial global optimal sampling weights corresponding to each fault category obtained in step S4 with the effective contribution of each generative adversarial network sub-model to the corresponding fault category in the current round, and update the local sampling ratio of each generative adversarial network sub-model for each fault category in the next round. S7. Based on the updated target quota and local sampling ratio, generate the next round of candidate enhancement samples, and perform dual-gated screening on the candidate enhancement samples. The dual-gated screening includes: The first gate is to determine whether the feature space distance criterion between the candidate enhancement sample and the distribution of the corresponding fault category real sample in the low-dimensional feature space meets the distribution tolerance threshold. The second gating process involves incorporating candidate enhancement samples from the first gating process into the current expanded fault diagnosis dataset for incremental evaluation to determine whether the overall macro-average F1 score gain of the validation set meets the minimum performance gain threshold. A candidate augmentation sample is only incorporated into the current augmented fault diagnosis dataset if it passes both the first and second gating. S8. After each round of candidate enhancement samples are incorporated, check whether the early stopping condition is met. If the macro-average F1 score gain for a consecutive preset number of rounds does not reach the minimum performance gain threshold, or the feature center offset distance of the expanded fault diagnosis dataset caused by the new samples is less than the preset convergence threshold, then stop the iteration and output the final enhancement dataset and the corresponding trained downstream fault classifier. Otherwise, return to step S5 to continue iterating, thereby forming a closed-loop enhancement process for photovoltaic system fault data based on downstream fault classification feedback.

2. The photovoltaic system fault data closed-loop enhancement method according to claim 1, characterized in that, In step S2, when performing standard test condition calibration on the IV curve based on the photovoltaic module physical model, the current calibration calculation formula is as follows: The voltage comprehensive compensation correction formula is: In the formula, and These are the currents before and after correction, respectively; and These are the irradiance before and after correction, respectively; and These are the component temperatures before and after calibration, respectively. The relative temperature coefficient of the short-circuit current; and These are the voltages before and after correction, respectively; This is the equivalent series resistance; This is the curve correction factor; This represents the absolute temperature coefficient of the open-circuit voltage of a photovoltaic module.

3. The photovoltaic system fault data closed-loop enhancement method according to claim 1, characterized in that, In the multi-objective particle swarm optimization algorithm described in step S4, the particles In the Speed ​​at the next iteration and location The update formula is: In the formula, To balance the inertia weights of global search and local search; and These are cognitive learning factors and social learning factors, respectively. and It is a vector of random numbers uniformly distributed in the interval [0,1]. For particles The historical best position; This is the current globally optimal position.

4. The photovoltaic system fault data closed-loop enhancement method according to claim 1, characterized in that, The multiple generative adversarial network sub-models in the heterogeneous generator pool mentioned in step S3 include: Wasserstein generative adversarial network WGAN-GP with gradient penalty, conditional table generative adversarial network CTGAN, deep convolutional generative adversarial network DCGAN, least squares generative adversarial network LSGAN, and mutual information maximization generative adversarial network InfoGAN.

5. The photovoltaic system fault data closed-loop enhancement method according to claim 1, characterized in that, Category in step S5 Classification confidence score The average information entropy is calculated by predicting the probability distribution of samples belonging to this category in the validation set, and then inverted after normalization. In the formula, To verify that the real labels in the set are categories The sample set, This represents the number of samples in the sample set. This represents the total number of fault categories. Samples output by the downstream fault classifier Category The soft probability.

6. The photovoltaic system fault data closed-loop enhancement method according to claim 5, characterized in that, Fault category in step S6 Category Difficulty Weight Determined according to the following formula: Among them, categories misclassification rate for: In the formula, For category The total number of samples in the current expanded fault diagnosis dataset, This represents the maximum total number of samples across all categories. , , Let be a normalized non-negative adjustment coefficient, and satisfy . ; This is an indicator function.

7. The photovoltaic system fault data closed-loop enhancement method according to claim 6, characterized in that, In step S6, the next iteration will target the fault category. Target generation quota Dynamic allocation as follows: In the formula, The quota amount generated per transaction is based on this. and These are the absolute lower limit and absolute upper limit of the number of items generated in a single run, respectively.

8. The photovoltaic system fault data closed-loop enhancement method according to claim 7, characterized in that, In step S6, the next round of generative adversarial network sub-models targets the fault category. The local sampling ratio is updated as follows: In the formula, The output of step S4 The globally optimal sampling weights of a generative adversarial network sub-model; For the first The generative adversarial network sub-model in the current... Fault categories in round iteration Effective contribution score; For the first In the first iteration Each generative adversarial network sub-model targets the fault category. The local sampling ratio; For the first In the first iteration Each generative adversarial network sub-model targets the fault category. The actual number generated; To generate the total number of adversarial network sub-models.

9. The photovoltaic system fault data closed-loop enhancement method according to claim 1, characterized in that, In step S7, the first gated judgment feature space distance criterion is satisfied: The second gating method incorporates candidate augmentation samples that have passed the feature space distance criterion into the current expanded fault diagnosis dataset to form a temporary dataset. Under the same training configuration, the overall macro-average F1 score of the corresponding validation set is obtained. And calculate: If and only if When the marginal performance gain test is passed, it is determined that the test is passed. In the formula, Fault Category The true sample distribution For the first Fault categories in round iteration The distribution of candidate augmented samples, and The preset distribution tolerance threshold, The minimum performance gain threshold, For the first The expanded fault diagnosis dataset after rounds of iteration.

10. The photovoltaic system fault data closed-loop enhancement method according to claim 9, characterized in that, The early stopping conditions mentioned in step S8 include: The early stop mechanism is automatically triggered when any of the following conditions are met: continuous All iterations satisfy ,in ; Or satisfy ; In the formula, and The first Wheel and First Round-robin expansion of the feature center vector of the fault diagnosis dataset, The preset convergence threshold is used; After triggering the early stopping mechanism, the final augmented dataset and the corresponding trained downstream fault classifier are output.

Images