Photovoltaic inverter anomaly detection method based on DDPM-DAGMM under small sample
By generating training data and training models using the DDPM-DAGMM method, the fault diagnosis and anomaly detection problems in small sample scenarios of photovoltaic inverters are solved, and real-time and intelligent detection effects are achieved, which is suitable for newly built photovoltaic stations.
Patent Information
- Application Number
- CN202510927979.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-07
- Publication Date
- 2025-10-03
AI Technical Summary
Existing technologies have poor real-time and intelligence in photovoltaic inverter fault diagnosis and anomaly detection, and deep learning methods have high data requirements, making them difficult to apply to newly built or newly commissioned photovoltaic sites.
A DDPM-DAGMM-based method is adopted to generate training data through the denoising diffusion probability model (DDPM). The model is trained in combination with the deep autoencoder mixed Gaussian mixture model (DAGMM) to obtain the anomaly detection threshold and realize photovoltaic inverter anomaly detection under small sample conditions.
It realizes real-time and intelligent anomaly detection of photovoltaic inverters with strong feature extraction capabilities and low data requirements. It is suitable for small sample scenarios and improves the accuracy and reliability of detection.
Smart Images
Figure CN120744459A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of photovoltaic power generation equipment status monitoring and anomaly detection, and specifically discloses a photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small samples. Background Art
[0002] As fossil fuels deplete and carbon emissions intensify, solar energy, due to its clean, unlimited, and widespread nature, has become a key component of national energy strategies. Photovoltaic solar power generation technology utilizes the photovoltaic effect at semiconductor interfaces to directly convert sunlight into electricity. Its core components include solar panels (monocrystalline silicon, polycrystalline silicon, or thin-film cells), controllers, and inverters, packaged in series to form a power generation unit. These units feature no mechanical parts, easy maintenance, and a long lifespan. In 2024, global photovoltaic capacity will increase by 602 GW, a year-on-year increase of 31.15%, accounting for 81% of new renewable energy capacity, bringing the total installed capacity to 2,247 GW. Optimistic forecasts predict that global installed capacity will exceed 750 GW by 2025. From January to May 2025, my country's newly installed photovoltaic capacity reached nearly 200 GW, a year-on-year growth rate of 57%. By the end of May 2025, China's cumulative installed photovoltaic capacity will reach 1.08 billion kilowatts, accounting for 30% of China's total installed power generation capacity and nearly half of the world's total installed photovoltaic capacity.
[0003] A photovoltaic power generation system is a crucial device that generates direct current (DC) electricity through the photoelectric effect of an internal photovoltaic array. A photovoltaic inverter then converts the resulting variable DC voltage into alternating current (AC), ultimately feeding the energy into the grid or directly supplying it to a load. PV power stations are typically constructed in areas with abundant solar resources. However, due to adverse climate and environmental factors, components such as the power transistors, freewheeling diodes, and filter inductors in the grid-connected inverter circuitry may open or short circuit, adversely affecting the operation of the inverter and the energy internet. In severe cases, this can even lead to unsafe and stable grid operation. Therefore, timely fault diagnosis and anomaly detection of PV inverters are crucial.
[0004] Traditional methods for fault diagnosis and anomaly detection in photovoltaic inverters mostly focus on device detection, including signal detection, circuit structure analysis, and infrared imaging. While these methods are widely used in real-world scenarios, they place high demands on detection equipment and suffer from poor real-time and intelligence performance. In recent years, data analysis-based methods, such as statistical models and classical machine learning models, have garnered increasing attention. However, these methods struggle with high-dimensional data and fail to capture data features well. Deep learning-based methods, owing to their superior ability to process large amounts of data and extract high-dimensional features, have achieved promising experimental results. However, deep learning-based methods rely on large amounts of data for model training, making them less suitable for newly built or commissioned photovoltaic power plants. Therefore, research on fault diagnosis and anomaly detection for photovoltaic power generation equipment using small data sets is crucial.
[0005] CN119646648A discloses a photovoltaic fault detection method based on a CNN-LSTM hybrid neural network. First, the photovoltaic power generation and meteorological data within the same time period are collected at the string level; secondly, the collected data is preprocessed, and then a fault detection neural network model is constructed. The fault detection module extracts features from the preprocessed data through a convolutional layer, uses channel rearrangement to enhance feature diversity, and then processes the time series data through an LSTM, and finally outputs the fault classification result through a fully connected layer. During the training process, the focus loss function is used as the loss function of the model to deal with the category imbalance problem, and the model weights with the best performance are saved for subsequent fault detection. The method of the present invention has the advantages of high accuracy, fast convergence, strong robustness, and good generalization ability. It can effectively improve the accuracy of photovoltaic array fault detection and classification and achieve accurate positioning of fault strings. The above-mentioned existing technologies require too many samples. Summary of the Invention
[0006] The present invention proposes a photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions, which is used to solve the fault diagnosis and anomaly detection problems of photovoltaic inverters in newly built photovoltaic stations.
[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions.
[0008] Aiming at the operation, maintenance and status warning of existing photovoltaic power generation equipment, the present invention proposes a photovoltaic inverter anomaly detection method based on DDPM-DAGMM under small sample conditions, which is used to solve the problems of poor real-time and intelligence of classic detection methods, insufficient feature extraction capabilities of machine learning methods, and large data requirements of deep learning methods, and provides guidance for the safe and reliable operation of photovoltaic equipment.
[0009] The present invention proposes a photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions, which is characterized by comprising the following steps:
[0010] (1) PV inverter operation data collection
[0011] Generally speaking, relevant operating data of the inverter can be obtained from the photovoltaic power station data acquisition and monitoring system. The main monitoring parameters include ambient temperature, wind speed, wind direction, solar radiation intensity, inverter DC input voltage, inverter DC input current, inverter three-phase current, etc.
[0012] (2) Historical data processing
[0013] To ensure the validity of the data used for subsequent data generation and model training, it is necessary to process the historical data of photovoltaic inverters. Due to the limited amount of data from newly commissioned photovoltaic sites, traditional machine learning-based methods are not suitable for processing historical data of photovoltaic equipment. Therefore, this patent mainly processes the historical data of photovoltaic inverters from the perspective of data validity.
[0014] (3) Inverter training data generation based on DDPM
[0015] For the historical data of photovoltaic inverters processed in step (2), the denoising diffusion probabilistic model (DDPM) is used to generate data for the subsequent training of anomaly detection models.
[0016] (4) DAGMM model training and anomaly detection threshold acquisition
[0017] The inverter data generated in step (3) is combined with the original data and divided into a training data set and a test data set in a ratio of 70%:30%. The training data set is used to train the DAGMM model to obtain a normal operating state model; the test data set is used to determine the distribution of state monitoring indicators and the anomaly detection threshold under normal conditions.
[0018] (5) Photovoltaic inverter abnormality detection
[0019] The real-time operating data of the photovoltaic inverter is obtained, and the DAGMM model trained in step (4) is input to obtain the real-time status monitoring index. This index is compared with the warning threshold obtained in step (4): if the real-time status monitoring index exceeds the warning threshold, the photovoltaic inverter status is identified as abnormal; if the real-time status monitoring index does not exceed the warning threshold, the photovoltaic inverter status is identified as normal.
[0020] Beneficial technical effects: excellent real-time performance, good intelligence, strong feature extraction capability of machine learning methods, low data requirements of deep learning methods, and can provide effective guidance for the safe and reliable operation of photovoltaic equipment. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 A diagram showing the steps of the method of the present invention.
[0022] Figure 2 This is the DDPM model workflow.
[0023] Figure 3 It is the DAGMM model structure. DETAILED DESCRIPTION
[0024] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
[0025] like Figure 1 As shown, in this embodiment, the photovoltaic inverter anomaly detection method based on DDPM-DAGMM mainly consists of five steps: photovoltaic inverter operation data collection, historical data processing, inverter training data generation based on DDPM, DAGMM model training and anomaly detection threshold acquisition, and photovoltaic inverter anomaly detection.
[0026] (1) PV inverter operation data collection
[0027] Generally speaking, inverter-related operating data can be obtained from the photovoltaic power station data acquisition and monitoring system. In this embodiment, weather station and inverter-related operating parameters are selected for subsequent data generation and model training. These parameters include ambient temperature, wind speed, solar radiation intensity, inverter A-phase voltage, inverter B-phase voltage, inverter C-phase voltage, inverter A-phase current, inverter B-phase current, inverter C-phase current, inverter AB line voltage, inverter BC line voltage, inverter CA line voltage, inverter input power, inverter output power, inverter conversion efficiency, and inverter internal temperature.
[0028] For newly commissioned wind power equipment, collect at least one month's data for subsequent data generation and model training.
[0029] (2) Historical data processing
[0030] To ensure the validity of data used for subsequent data generation and model training, historical PV inverter data must be processed. Due to the limited data volume of newly commissioned PV sites, traditional machine learning-based methods are not suitable for processing historical PV equipment data. Therefore, this patent focuses on processing historical PV inverter data from a data validity perspective. This includes data validity verification, data interpolation, and data normalization.
[0031] Among them, data validity test is to determine whether the collected data has obvious errors or human interference signals, etc., which is mainly tested based on the characteristics of physical quantities, including limit value test, change value test, and relationship test.
[0032] Limit test
[0033] Limit check means that when the inverter data exceeds its maximum possible variation range, the data will be judged as unreliable. In this embodiment, the limit check is performed using the following formula:
[0034]
[0035] Where, are the inverter parameters, is the minimum value of the corresponding parameter, is the maximum value of the parameter.
[0036] Change value test
[0037] The change value test means that when the inverter data changes beyond a certain range and does not conform to the rules within a change cycle, the data can be considered unreliable. In this embodiment, the change value test is performed using the following formula:
[0038]
[0039] Where, is the inverter parameter within the change cycle The minimum value of is the inverter parameter within the change cycle The maximum value of Inverter parameters The change value test threshold.
[0040] Relationship Test
[0041] Relationship verification refers to the mutual verification of related parameters using their relationships. For example, the inverter input power must be greater than the inverter output power, and the inverter phase voltage / phase current / line voltage can be verified against each other.
[0042] (3) Inverter training data generation based on DDPM
[0043] The main working principle of DDPM is to gradually add noise to the original data, and then gradually learn how to remove the noise, thereby generating new data samples. Figure 2 As shown in Figure 3, the DDPM workflow mainly consists of two parts: the forward diffusion process and the backward denoising process.
[0044] In the forward diffusion process, Gaussian noise is gradually added to the original data until it eventually becomes random noise data. is the original data, yes The data after adding noise. The process of adding noise is as follows:
[0045]
[0046]
[0047] Where, is the total number of steps in the diffusion process, represents the probability distribution of data at the time step, Indicates the mean and the variance is Gaussian noise, is the identity matrix with the same shape as the input data, The variance of the Gaussian noise at each step is specified, usually a linear or cosine schedule.
[0048] Among them, the reverse denoising process restores all noise to the original data, and usually uses a neural network to complete the denoising process, as shown below:
[0049]
[0050]
[0051] Where, are the learnable parameters of the neural network, represents random Gaussian noise.
[0052] After multiple iterations, the trained and optimized diffusion model can gradually recover clear data from a series of Gaussian-distributed noise samples and eventually generate samples that are consistent with the original data distribution.
[0053] (4) DAGMM model training and anomaly detection threshold acquisition
[0054] The inverter data generated in step (3) is combined with the original data and divided into a training data set and a test data set in a ratio of 70%:30%. The training data set is used to train the DAGMM model to obtain a normal operating state model; the test data set is used to determine the distribution of state monitoring indicators and the anomaly detection threshold under normal conditions.
[0055] Among them, the DAGMM model structure is as follows Figure 3 As shown in Figure 1, it mainly consists of two parts: compression network and evaluation network. The compression network reduces the dimensionality of the input sample through the autoencoder, extracts low-dimensional representation and reconstructs the error features. The features of the compression network include two parts: low-dimensional features learned by the deep autoencoder and low-dimensional features of the reconstruction error , and then formed , provided to the subsequent evaluation network.
[0056] The evaluation network uses these features to predict the abnormal possibility of the sample under the framework of the Gaussian mixture model. After multiple layers of full connection, the output value of the model is finally obtained. , which contains the category probability after softmax , as shown below:
[0057] , ,
[0058]
[0059] After obtaining the output of the model, the current energy can be obtained according to the multivariate Gaussian probability density related formula and the energy evaluation formula, and high-energy samples can be predicted as abnormal through the pre-selected threshold. The DAGMM model training loss function is:
[0060]
[0061] In order to determine the threshold of the monitoring indicator, it is necessary to perform statistical analysis on the monitoring indicators of the normal sample data. The test data set is input into the trained DAGMM model, and the mean of the DAGMM output results is obtained by statistics. and standard deviation , the inverter abnormality detection threshold based on normal distribution is:
[0062]
[0063] (5) Photovoltaic inverter abnormality detection
[0064] The real-time operating data of the photovoltaic inverter is obtained, and the DAGMM model trained in step (4) is input to obtain the real-time status monitoring index. This index is compared with the warning threshold obtained in step (4): if the real-time status monitoring index exceeds the warning threshold, the photovoltaic inverter status is identified as abnormal; if the real-time status monitoring index does not exceed the warning threshold, the photovoltaic inverter status is identified as normal.
[0065] To further illustrate the implementation process of the present invention, specific data is used for calculation. In this embodiment, the operation and maintenance personnel periodically count the warning results of the anomaly detection method, and analyze the warning effect of the proposed method by counting the number of correct alarms, the number of false alarms, and the number of missed alarms. The present invention regards abnormal samples as positive classes and normal samples as negative classes, and constructs a confusion matrix for the binary classification problem. The present invention uses three indicators based on the confusion matrix: precision, recall, and F1 score to evaluate the performance of the time series data anomaly detection model:
[0066]
[0067]
[0068]
[0069] TP (True Positive) indicates samples that are actually positive and correctly detected as positive; TN (True Negative) indicates samples that are actually negative and correctly detected as negative; FP (False Positive) indicates samples that are actually negative but mistakenly detected as positive; FN (False Negative) indicates samples that are actually positive but mistakenly detected as negative.
[0070] For a certain inverter model at a specific electric field, one month's worth of data was selected and processed for data generation using DDPM and model training using DAGMM. The trained model was then used for real-time data anomaly detection. Statistics on the model's anomaly detection performance over a period of time revealed a precision of 0.9269, a recall of 0.9453, and an F1 score of 0.9360, meeting actual project requirements.
[0071] It will be easily understood by those skilled in the art that 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 in the scope of protection of the present invention.
Claims
1. A photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions, characterized by: It includes: photovoltaic inverter operation data collection steps, historical data processing steps, inverter training data generation steps based on DDPM, DAGMM model training and anomaly detection threshold acquisition steps, and photovoltaic inverter anomaly detection steps.
2. The photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions according to claim 1 is characterized by: In the photovoltaic inverter operation data collection step, the relevant operation data of the inverter can be obtained from the photovoltaic power station data acquisition and monitoring system, and at least the following data are monitored: ambient temperature, wind speed, solar radiation intensity, inverter DC input voltage, inverter DC input current, and inverter three-phase current.
3. The photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions according to claim 2, characterized in that: In the historical data processing step, the historical data of the photovoltaic inverter is processed to ensure the validity of the data used for subsequent data generation and model training.
4. The photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions according to claim 3 is characterized by: In the inverter training data generation step based on DDPM, a denoising diffusion probability model is used to generate data for the processed photovoltaic inverter historical data, which is used for the subsequent training of the anomaly detection model.
5. The photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions according to claim 4 is characterized in that: In the DAGMM model training and anomaly detection threshold acquisition steps, the generated inverter data is merged with the original data and divided into a training data set and a test data set in a ratio of 70%:30%, respectively. The training data set is used to train the DAGMM model to obtain a normal operating status model; the test data set is used to determine the status monitoring indicator distribution and anomaly detection threshold under normal conditions.
6. The photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions according to claim 5, characterized in that: In the PV inverter anomaly detection step, real-time operating data of the PV inverter is obtained and input into the trained DAGMM model to obtain real-time status monitoring indicators. This indicator is then compared with the obtained warning threshold. If the real-time status monitoring indicator exceeds the warning threshold, the PV inverter status is identified as abnormal. If the real-time status monitoring indicator does not exceed the warning threshold, the PV inverter status is identified as normal.
7. The photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions according to claim 2, characterized in that: The weather station and inverter-related operating parameters are used for subsequent data generation and model training, namely, ambient temperature, wind speed, solar radiation intensity, inverter A phase voltage, inverter B phase voltage, inverter C phase voltage, inverter A phase current, inverter B phase current, inverter C phase current, inverter AB line voltage, inverter BC line voltage, inverter CA line voltage, inverter input power, inverter output power, inverter conversion efficiency, and inverter internal temperature; for newly commissioned wind power equipment, at least one month's data is collected for subsequent data generation and model training.
8. The photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions according to claim 7, characterized in that: The historical data of photovoltaic inverters is processed, including data validity verification, data interpolation, and data normalization. Data validity verification is to determine whether the collected data has obvious errors or human interference signals. The verification is based on the characteristics of physical quantities and includes limit value verification, change value verification, and relationship verification. Limit value verification means that when the inverter data exceeds its maximum possible change range, the data will be judged as unreliable and the limit verification is performed using the following formula: , where are the inverter parameters, is the minimum value of the corresponding parameter, is the maximum value of the parameter; the change value test means that within a change cycle, when the inverter data changes beyond a certain range and does not conform to the law, the data can be considered unreliable; the change value test is performed using the following formula: , where is the inverter parameter within the change cycle The minimum value of is the inverter parameter within the change cycle The maximum value of Inverter parameters The change value test threshold is used; relationship test refers to the mutual verification of the relationship between related parameters. The inverter input power must be greater than the inverter output power, and the inverter phase voltage / phase current / line voltage are mutually verified.
9. The photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions according to claim 8, characterized in that: In the inverter training data generation step based on DDPM, DDPM gradually adds noise to the original data, and then gradually learns how to denoise it, thereby generating new data samples; the DDPM workflow consists of two parts: the forward diffusion process and the reverse denoising process; in the forward diffusion process, Gaussian noise is gradually added to the original data until it eventually becomes random noise data. Assuming is the original data, yes The data after adding noise; the process of adding noise is as follows: , , where is the total number of steps in the diffusion process, represents the probability distribution of data at the time step, Indicates the mean and the variance is Gaussian noise, is the identity matrix with the same shape as the input data, is the variance of the specified Gaussian distribution noise at each step, usually linear or cosine; the inverse denoising process restores the complete noise to the original data, usually using a neural network to complete the denoising process, as shown below: , , where are the learnable parameters of the neural network, represents random Gaussian noise; after multiple iterations, the trained and optimized diffusion model gradually recovers clear data from a series of Gaussian distributed noise samples, and finally generates samples consistent with the original data distribution.
10. The photovoltaic inverter anomaly detection method based on DDPM-DAGMM for small sample conditions according to claim 9, characterized in that: In the DAGMM model training and anomaly detection threshold acquisition steps, the generated inverter data is merged with the original data and divided into a training data set and a test data set in a ratio of 70%:30% respectively; the training data set is used to train the DAGMM model to obtain a normal operating state model; the test data set is used to determine the distribution of state monitoring indicators and anomaly detection thresholds under normal conditions; the DAGMM model structure consists of two parts: a compression network and an evaluation network. The compression network reduces the dimensionality of the input samples through an autoencoder, extracts low-dimensional representations and reconstructs error features; the features of the compression network include two parts: low-dimensional features learned by the deep autoencoder and low-dimensional features of the reconstruction error , and then formed , provided to the subsequent evaluation network; the evaluation network predicts the abnormal possibility of the sample under the framework of the Gaussian mixture model; after multiple layers of full connection, the output value of the model is finally obtained , which contains the category probability after softmax , as shown below: , , , After obtaining the output of the model, the current energy is obtained according to the multivariate Gaussian probability density related formula and the energy evaluation formula, and the high energy samples are predicted as abnormal through the pre-selected threshold. The DAGMM model training loss function is: In order to determine the threshold of the monitoring indicator, it is necessary to conduct statistical analysis on the monitoring indicators of the normal sample data; the test data set is input into the trained DAGMM model, and the mean of the DAGMM output results is obtained by statistics. and standard deviation , the inverter abnormality detection threshold based on normal distribution is: .