Arcing detection method and device of inverter, electronic equipment and storage medium

Abstract

The invention is suitable for the technical field of inverter detection, and provides an arc discharge detection method and device of an inverter, electronic equipment and a storage medium. The method comprises the following steps: inputting working current data of a photovoltaic inverter into an initial arc discharge detection model; in the preset continuous duration of the first operation, if the output result is normal operation and the confidence value is greater than the threshold value, taking the corresponding working current data as a sample to form a sample data set; performing online training on the initial arc discharge detection model by using the sample data set to obtain a candidate arc discharge detection model; under the condition that the candidate arc discharge detection model meets a preset requirement, determining the candidate arc discharge detection model as a target arc discharge detection model; and finally, performing arc discharge detection on the real-time current window based on the target arc discharge detection model in online operation. According to the scheme, on-line training is carried out on site, so that model parameters can be self-adapted in real time along with fluctuation of a power grid on site, load disturbance and noise level, and the arc discharge detection precision is improved.

Description

Technical Field

[0001] This application belongs to the field of inverter testing technology, and particularly relates to an inverter arcing detection method, device, electronic equipment and storage medium. Background Technology

[0002] Arcing faults on the DC side of photovoltaic inverters are a major fire hazard, requiring reliable detection capabilities in engineering. However, actual power plant environments are complex and variable, with significant grid fluctuations, load disturbances, and noise interference. Current mainstream approaches are mostly based on fixed thresholds: filtering and transforming the current signal to extract harmonics or spectral energy, and comparing it with a preset threshold to determine if an arc has occurred. This method is simple to implement and has low overhead, but the threshold is extremely sensitive to on-site grid fluctuations, load disturbances, and noise, easily leading to false alarms, missed alarms, and insufficient environmental adaptability.

[0003] Some studies have used machine learning or deep learning for automatic feature extraction and classification, but these generally rely on laboratory datasets for training and still suffer from drawbacks such as poor environmental adaptability and inconvenient model updates in actual photovoltaic power plant applications. Summary of the Invention

[0004] In view of this, embodiments of this application provide an inverter arcing detection method, device, electronic device and storage medium. By collecting field samples within a preset continuous duration during the first run and adaptively fine-tuning at the edge, model updates can be achieved without cloud access, thereby improving environmental adaptability and reducing false alarms and missed alarms.

[0005] A first aspect of this application provides a method for detecting arcing in an inverter, the method comprising: The operating current data of the photovoltaic inverter is input into the initial arcing detection model. The initial arcing detection model is deployed to the photovoltaic inverter after offline training. It is used to infer the operating current data and output the arc judgment result and the corresponding confidence value. The arc judgment result includes normal operation and arc fault. During the preset continuous period of the first operation of the photovoltaic inverter, when the arc judgment result output by the initial arc detection model is normal operation and the confidence value is greater than the preset threshold, the corresponding operating current data is stored as sample data to form a sample data set collected within the preset continuous period. The initial arc detection model is trained online using the sample data set to obtain candidate arc detection models; If the candidate arc detection model meets the preset requirements, the candidate arc detection model is determined as the target arc detection model; Based on the target arc detection model, arc detection is performed on the photovoltaic inverter during operation.

[0006] This application embodiment first inputs the operating current data into an initial arcing detection model deployed after offline training on the inverter side, and outputs the arc judgment result and corresponding confidence level in real time, providing an objective basis for subsequent sample selection based on confidence level. Subsequently, within a preset continuous duration of the first run, only when the model determines that it is operating normally and the confidence level is greater than a preset threshold, the corresponding current data is stored as sample data, forming a sample dataset that fits the real operating conditions. This dataset is then merged to train the initial arcing detection model online on the device side, obtaining candidate arcing detection models. If the candidate arcing detection models meet preset requirements, they are determined as the target arcing detection model. During the online training process, the model parameters are made to adapt in real time to the fluctuations of the power grid, load disturbances, and noise levels, significantly reducing the distribution differences between laboratory and power plant operating conditions. Finally, the target arcing detection model is used to perform online detection. Without relying on the cloud, the model gradually adapts to the field data, enhancing the robustness and consistency of the model to power grid fluctuations, load disturbances, and noise changes, thereby improving environmental adaptability in the actual photovoltaic power plant environment and reducing false alarms and false negatives.

[0007] In one possible implementation, the sample dataset includes a training set and a validation set, wherein each sample in the sample dataset has a true label; after training an initial arc detection model online using the sample dataset to obtain a candidate arc detection model, the method further includes: The sample data in the training set is input into the initial arc detection model to obtain the arc judgment result corresponding to each sample data in the training set; Based on the arc judgment results and true labels corresponding to each sample data in the training set, the first training loss information is determined; The model parameters of the initial arc detection model are updated based on the first training loss information to obtain a candidate arc detection model; The sample data in the validation set is input into the candidate arc detection model to obtain the second training loss information; When the second training loss information meets the preset loss requirement and the performance index of the candidate arc detection model meets the preset index requirement, the candidate arc detection model is confirmed to meet the preset requirements.

[0008] In one possible implementation, the initial arc detection model includes a fully connected classification layer; updating the model parameters of the initial arc detection model based on the first training loss information to obtain a candidate arc detection model includes: Based on the first training loss information, the parameters of the fully connected classification layer in the initial arc detection model are updated to obtain the updated arc detection model. The sample data from each sample in the validation set are input into the updated arc detection model to determine the third training loss information; If the third training loss information is less than the first training loss information, the updated arc detection model is determined as the candidate arc detection model.

[0009] In one possible implementation, the performance metrics of the candidate arc detection model are obtained based on the arc judgment results corresponding to the validation set and the true labels, and the arc judgment results of the validation set are output by the candidate arc detection model; the method further includes: The accuracy of the candidate arc detection model is determined based on the first and second counts; the first count refers to the number of cases in the validation set where the arc judgment result is arc fault and the true label of the validation set is arc fault, and the second count refers to the number of cases in the validation set where the arc judgment result is arc fault and the true label of the validation set is normal operation. The recall rate of the candidate arc detection model is determined based on the first and third counts; the third count refers to the number of cases in the validation set where the arc judgment result is normal operation and the true label of the validation set is arc fault. Based on the precision and recall, the performance metrics of the candidate arc detection model are determined.

[0010] In one possible implementation, during the preset continuous period of the photovoltaic inverter's first operation, when the arc judgment result output by the initial arc detection model indicates normal operation and the confidence value is greater than a preset threshold, the corresponding operating current data is stored as sample data to form a sample data set collected within the preset continuous period, including: During the preset continuous duration of the first operation of the photovoltaic inverter, under the action of maximum power point tracking control, the DC side operating point of the inverter is made to run along the maximum power point tracking trajectory; Using the operating current data on the maximum power point tracking trajectory as a grading reference, the current range is set from the first preset current to the maximum operating current reached on the day, and a grading set is generated according to the second preset current as the step. Traverse the set of grades. In each grade, when the arc judgment result output by the initial arc detection model is normal operation and the confidence value is greater than the preset threshold, store the corresponding working current data as sample data. The sample data set is determined based on the sample data within the preset continuous time period.

[0011] In one possible implementation, the offline training step of the initial arc detection model includes: Collect and label normal current data and arc current data; The normal current data and the arc current data are preprocessed to determine the frequency domain feature vectors corresponding to the normal current data and the arc current data, so as to obtain an initial sample data set. The initial sample dataset is divided into an initial training set, an initial validation set, and an initial test set. The arc detection network is trained using the initial training set as input, the network parameters of the arc detection network are determined based on the initial validation set, and the performance index of the arc detection network under the network parameters is determined based on the initial test set. When the performance indicators of the arc detection network meet the preset requirements, the arc detection network is determined as the initial arc detection model.

[0012] A second aspect of this application provides an arcing detection device for an inverter, the device comprising: The data input module is used to input the operating current data of the photovoltaic inverter into the initial arc detection model. The initial arc detection model is deployed to the photovoltaic inverter after offline training. It is used to infer the operating current data and output the arc judgment result and the corresponding confidence value. The arc judgment result includes normal operation and arc fault. The sample acquisition module is used to store the corresponding operating current data as sample data when the arc judgment result output by the initial arc detection model is normal operation and the confidence value is greater than a preset threshold during the preset continuous operation of the photovoltaic inverter, so as to form a sample data set acquired within the preset continuous operation period. The online training module is used to train the initial arc detection model online using the sample data set to obtain candidate arc detection models; The performance verification module is used to determine the candidate arc detection model as the target arc detection model when the candidate arc detection model meets the preset requirements. The arc detection module is used to perform arc detection on the photovoltaic inverter during operation based on the target arc detection model.

[0013] A third aspect of this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect above.

[0014] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in the first aspect above.

[0015] A fifth aspect of this application provides a computer program product that, when run on an electronic device, causes the electronic device to perform the steps of the method described in the first aspect above.

[0016] The beneficial effects of the second to fifth aspects mentioned above can all be referred to the beneficial effects described in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of an inverter arcing detection method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the residual block structure of the arc detection network in the embodiments of this application; Figure 3 This is a schematic diagram of the overall structure of the arc detection network in the embodiments of this application; Figure 4 This is the overall flowchart of the arcing detection method for inverters; Figure 5 This is a schematic diagram of the edge adaptive fine-tuning mechanism in the overall flowchart; Figure 6 This is a schematic diagram of the structure of an arcing detection device for an inverter provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0020] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0021] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0023] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] It should be understood that the sequence number of each step in this embodiment does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0025] Arcing faults on the DC side of photovoltaic inverters are a major fire hazard, requiring reliable detection capabilities in engineering. However, actual power plant environments are complex and variable, with significant grid fluctuations, load disturbances, and noise interference. Current mainstream approaches are mostly based on fixed thresholds: filtering and transforming the current signal to extract harmonics or spectral energy, and comparing it with a preset threshold to determine if an arc has occurred. This method is simple to implement and has low overhead, but the threshold is extremely sensitive to on-site grid fluctuations, load disturbances, and noise, easily leading to false alarms, missed alarms, and insufficient environmental adaptability.

[0026] Some studies have used machine learning or deep learning for automatic feature extraction and classification, but these generally rely on laboratory datasets for training and still suffer from drawbacks such as poor environmental adaptability and inconvenient model updates in actual photovoltaic power plant applications.

[0027] To address the aforementioned issues, embodiments of this application provide an inverter arcing detection method, apparatus, electronic device, and storage medium. The arcing detection method for inverters in this embodiment first inputs the operating current data into an initial arcing detection model deployed after offline training on the inverter side, and outputs the arc judgment result and corresponding confidence level in real time, providing an objective basis for subsequent sample selection based on confidence level. Then, within a preset continuous duration of the first run, only when the model determines that it is operating normally and the confidence level is greater than a preset threshold, the corresponding current data is stored as sample data, forming a sample dataset that fits the real operating conditions. This dataset is then merged to train the initial arcing detection model online on the device side, obtaining candidate arcing detection models. If the candidate arcing detection models meet the preset requirements, they are determined as the target arcing detection model. During the online training process, the model parameters are made to adapt in real time to the fluctuations of the power grid, load disturbances, and noise levels, significantly reducing the distribution difference between laboratory and power plant operating conditions. Finally, the target arcing detection model is used to perform online detection. Without relying on the cloud, the model gradually adapts to the field data, enhancing the robustness and consistency of the model to the fluctuations of the power grid, load disturbances, and noise changes, thereby improving environmental adaptability in the actual photovoltaic power plant environment and reducing false alarms and false negatives.

[0028] The following is a detailed description of an inverter arcing detection method, apparatus, electronic device, storage medium, and computer program provided in the embodiments of this application, with reference to the accompanying drawings.

[0029] See Figure 1 The diagram illustrates a flowchart of an inverter arcing detection method provided in an embodiment of this application; as shown below. Figure 1 As shown, the method may include the following steps: Step 101: Input the operating current data of the photovoltaic inverter into the initial arc detection model.

[0030] The initial arc detection model is trained offline and then deployed to the photovoltaic inverter to infer the working current data and output the arc judgment result and the corresponding confidence value.

[0031] The arc judgment results include normal operation and arc fault, and the confidence value is used to characterize the credibility of the output arc judgment results.

[0032] In this embodiment of the application, the operating current data of the photovoltaic inverter can refer to the current data obtained by the embedded processing unit of the photovoltaic inverter after the DC side current of the photovoltaic inverter is converted into an analog voltage by a current transformer or a Hall sensor, and then sampled at a fixed sampling rate.

[0033] In one possible implementation, the initial arc detection model is trained offline and quantized and compressed before being deployed to the photovoltaic inverter.

[0034] Specifically, the inverter includes an arcing detection module. This module can preload an initial arcing detection model that has been trained offline and quantized and compressed. Then, it assembles the operating current data into a real-time current window (1024 points) of fixed length as input to the initial arcing detection model. It then performs forward inference and outputs the arcing judgment result of "arc fault / normal operation" and the corresponding confidence value (i.e., arcing probability). To ensure the consistency of subsequent processing, this implementation records and submits the timestamp and current grading information together with the inference of each window to the upper-level logic after completion. This allows subsequent steps to perform sample acquisition and alarm control based on threshold strategies. The above process is executed cyclically at the photovoltaic inverter end, thereby realizing the embedded edge deployment of the quantization model and real-time inference of the 1024-point current window, providing a stable and consistent discrimination and confidence value output for subsequent sample acquisition and fine-tuning.

[0035] In one possible implementation, the offline training steps of the initial arc detection model include: Collect and label normal current data and arc current data; The normal current data and arc current data are preprocessed to determine the frequency domain feature vectors corresponding to the normal current data and the arc current data, so as to obtain the initial sample data set. The initial sample dataset is divided into an initial training set, an initial validation set, and an initial test set. The arc detection network is trained using the initial training set as input, the network parameters of the arc detection network are determined based on the initial validation set, and the performance index of the arc detection network under the network parameters is determined based on the initial test set. When the performance indicators of the arc detection network meet the preset requirements, the arc detection network is determined as the initial arc detection model.

[0036] Among them, normal current data refers to the sequence of DC-side AC current collected under arc-free conditions; arc current data refers to the sequence of DC-side AC current collected under arc fault conditions; frequency domain feature vector refers to the current spectrum sequence obtained by preprocessing and fast Fourier transforming the current data, which is used as the model input feature; the initial sample data set refers to the data set formed by merging the frequency domain feature vectors corresponding to the normal current data and the arc current data. The initial training set, initial validation set, and initial test set refer to three subsets obtained by dividing the initial sample data set according to a preset ratio, which are used for model training, parameter selection, and performance verification, respectively.

[0037] Among them, the arc detection network is a neural network used to extract and classify features from the input frequency domain feature vector; the initial arc detection model refers to the arc detection network whose parameters are fixed after offline training and performance confirmation, and is used for subsequent deployment and inference.

[0038] For example, firstly, in a laboratory environment, an arc-drawing machine can be used to simulate an electric arc fault, with a DC current ranging from 3A to the maximum operating current. I max Within the specified range, normal current data and arc current data were collected in 1A increments; 10 sets of data were collected for each current level, with each set lasting 500 ms and a sampling rate of 250 kHz, to obtain the raw current data. I original [n].

[0039] Secondly, for I original [n] is subjected to exponentially weighted moving average filtering to obtain the filtered and denoised current data. I filter [n] is calculated using the following formula:

[0040] in, I filter [n] represents the filtered current data at the current moment; I filter [n-1] represents the filtered current data from the previous time step; I original [n] represents the raw current data at the current moment; α∈(0,1) is the smoothing coefficient, used to control the weight of the current data.

[0041] Then, to I filter [n] Perform a Fast Fourier Transform, taking the number of sampling points K=1024, to obtain the spectral data F[k] for training:

[0042] In this context, F[k] of each sample is taken as its corresponding frequency domain feature vector, K is the length, k is the index of the current frequency domain component, and m is the index of the current time domain sampling point.

[0043] After obtaining the frequency domain feature vectors corresponding to the normal current data and the arc current data, the two are merged to form an initial sample data set. This initial sample data set is then divided into an initial training set, an initial validation set, and an initial test set in a ratio of 4:3:3.

[0044] The arc detection network is trained using the initial training set as input: local features are extracted first through convolutional layers, then deep features are extracted through three residual blocks, and then the arc detection results are output through global average pooling layers and fully connected classification layers. The network parameters and training hyperparameters are determined based on the initial validation set, and the performance indicators under these parameters are evaluated on the initial test set. When the performance indicators meet the preset requirements, the arc detection network is solidified as the initial arc detection model.

[0045] In one possible implementation, the residual block structure of the arc detection network is as follows: Figure 2 As shown, the input feature x corresponding to the output or input sequence of the previous layer is fed into the residual block. First, it passes through the first convolutional layer, batch normalization, and ReLU to obtain intermediate features. Then, it passes through the second convolutional layer to obtain the residual mapping F(x). At the same time, an identity mapping is performed on x to form a direct connection branch, and it is summed with the residual mapping in the element-wise dimension to obtain H(x) = F(x) + x. Then, it is activated by ReLU to obtain the output feature H(x) and fed into the next network layer.

[0046] To ensure the feasibility of element-by-element addition, the following approach is adopted when the number of channels and the stride are consistent: Figure 2 The identity mapping is shown; when the number of channels or the time dimension resolution changes, a 1×1 one-dimensional convolution (with the same stride as the main branch) can be configured on the direct branch to achieve channel / stride matching, and then added to F(x), and then output after ReLU.

[0047] In this embodiment, the preferred one-dimensional convolution kernel length is 3, the stride is 1, and zero padding is used at both ends to keep the length constant. Batch normalization is used to stabilize training and suppress internal covariate shifts, and ReLU is used as a nonlinear mapping to enhance feature representation capabilities.

[0048] At the network level, multiple residual blocks are stacked sequentially to form a "multi-layer residual block," used for hierarchical deep feature extraction of the current frequency domain feature vector. A global average pooling layer is then used to average the time frequency dimension of each channel to obtain a fixed-length representation vector, which is then connected to a fully connected classification layer to output the arc judgment result ("arc fault or normal operation") and the corresponding confidence value. During training, the same input caliber as the data preprocessing is used, and the loss function used to calculate the loss information can be the binary cross-entropy function. During deployment, the quantized and compressed network is loaded as the initial arc detection model at the photovoltaic inverter end, performing forward inference on a fixed-length (e.g., 1024 points) real-time current window to generate arc probability, which is used for subsequent sample data acquisition and online judgment.

[0049] Based on the above implementation method, the overall structure diagram of the arc detection network is as follows: Figure 3As shown, this belongs to a one-dimensional convolutional neural network (ResNet-1DCNN) based on an improved residual structure. The network receives a preprocessed frequency domain feature vector F[k] (of length K, for example, K=1024) at the input layer, and then sequentially passes it through convolution, batch normalization, and ReLU to extract low-level local features. At the end, a shortcut branch is set up to add to the main branch in an element-wise dimension, forming the first residual fusion, which is used to stabilize deep training and alleviate gradient vanishing.

[0050] Subsequently, the network consists of three sets of cascaded residual blocks (corresponding to...) Figure 3 The three substructures from left to right). Each set of residual blocks uses "convolutional layer → batch normalization → ReLU → convolutional layer → (add with shortcut) → ReLU" as the basic unit: Specifically, when the number of input and output channels and the time step are the same for the residual block, the shortcut branch uses an identity mapping to directly add the output of the main branch; when the number of channels or the step size changes, the shortcut branch uses a 1×1 one-dimensional convolution (with the same stride as the main branch) to achieve dimension matching before adding; the preferred kernel length is 3, the stride is 1, and zero padding is used at both ends to keep the length constant; batch normalization is used to suppress internal covariate shift, and ReLU provides nonlinear mapping.

[0051] After hierarchically refining the frequency domain features through multiple residual blocks, the network enters a global average pooling layer to obtain a fixed-length representation vector by averaging each channel along the time or frequency dimension. This is then followed by a fully connected classification layer (with an output dimension of 2, corresponding to "arc fault and normal operation") to form the output layer. During training, binary cross-entropy is used as the loss function to optimize the network parameters. After deployment to an embedded edge, this network serves as the initial arc detection model, performing forward inference on the frequency domain feature vector corresponding to a real-time current window of length 1024 points, outputting the arc judgment result and its confidence value (arc probability). In subsequent edge adaptation, only the fully connected classification layer is unlocked as the trainable part, while the parameters of the remaining convolutional layers and residual blocks remain unchanged to achieve low-overhead, controllable, and stable online updates.

[0052] Step 102: During the preset continuous time period of the first operation of the photovoltaic inverter, when the arc judgment result output by the initial arc detection model is normal operation and the confidence value is greater than the preset threshold, the corresponding operating current data is stored as sample data to form a sample data set collected within the preset continuous time period.

[0053] The preset continuous duration refers to a continuously set data collection period (e.g., 7 consecutive days) during the initial operation of the photovoltaic inverter, used to acquire on-site sample data. Operating current data refers to real-time current windows (e.g., 1024 points long) collected from the DC side of the inverter. Sample data refers to the operating current data that is retained and stored long-term when the model outputs an arc judgment result indicating normal operation and a confidence value greater than a preset threshold. The sample dataset refers to the collection of all sample data accumulated within the preset continuous duration, used for subsequent division into training and validation sets.

[0054] In this embodiment, within a preset continuous duration, the inverter continuously acquires DC-side operating current data at a fixed sampling rate, and assembles the data into a real-time current window of 1024 points in accordance with the same criteria as in step 101. Each acquired window is fed into the initial arc detection model for forward inference to obtain the arc judgment result and confidence value for that window. When the arc judgment result output by the window indicates normal operation and the confidence value is greater than a preset threshold, the system determines the operating current data corresponding to that window as sample data and writes it into the local sample library. Simultaneously, it records the timestamp associated with the window, device identifier, DC operating point information (such as current value or current level), operating environment information (such as temperature and irradiance), and the confidence value output by the model for traceability and subsequent data management. If any condition is not met, the data is not included in the sample library, and only a minimal operating log is retained. The above process is executed cyclically according to a window period until the preset continuous duration ends. At the end, the system automatically summarizes all sample data accumulated during this period, forming a sample data set collected within the preset continuous duration. This set is then transferred to subsequent steps for dividing the training and validation sets according to a preset ratio and for edge-side fine-tuning. This setup ensures that the acquired sample data comes from real-world scenarios and is highly reliable, providing a stable and consistent data foundation for subsequent model updates.

[0055] In one possible implementation, step 102 above may specifically include: During the preset continuous duration of the first operation of the photovoltaic inverter, under the action of maximum power point tracking control, the DC side operating point of the inverter is made to run along the maximum power point tracking trajectory; Using the operating current data on the maximum power point tracking trajectory as a grading reference, the current range is set from the first preset current to the maximum operating current reached on the day, and a grading set is generated according to the second preset current as the step. Traverse the set of grades. Within each grade, when the arc judgment result output by the initial arc detection model is normal operation and the confidence value is greater than the preset threshold, store the corresponding working current data as sample data. The sample data set is determined based on the sample data within a preset continuous time period.

[0056] Among them, the maximum power point tracking trajectory refers to the operating trajectory formed by the DC side operating point of the inverter under MPPT control as the operating conditions change, which is used to determine the reference range for current grading.

[0057] Wherein, the first preset current refers to the lower limit of the current range; the second preset current refers to the step current of the range. In this embodiment, the first preset current can be 3A and the second preset current can be 1A.

[0058] The confidence score is output by the initial arc detection model and calculated using the Softmax function, which is the linear output of the last fully connected layer.

[0059] For example, the inverter operates under maximum power point tracking (MPPT) control for a preset continuous duration. The system continuously acquires DC-side operating current data at a sampling rate of 250 kHz and assembles real-time current windows with a time length of 500 ms. Each acquired window is fed into the initial arc detection model deployed to the arc detection module for inference, yielding the arc judgment result and confidence value. p normal The arc judgment result is normal operation if and only if the output is normal operation. p normal If the confidence level is >0.9 (preset threshold), the window is considered high-confidence normal operating data; otherwise, it is discarded. The confidence level can be calculated using the following formula:

[0060] in, e Znormal and e Zarc These represent the linear output values ​​of the last fully connected layer for normal operation and arc fault, respectively.

[0061] Regarding current coverage, the operating current on the MPPT track is used as the reference for grading, and the grading range is set from 3A to the maximum operating current of the day. I mppt-max The system iterates through each sub-segment in increments of 1A to ensure that samples of each current level are included in the set.

[0062] For windows that meet the conditions, the system automatically marks them as operating normally, counts them, and writes them into the sample data set (sample count N++). The system runs continuously for 7 consecutive days, accumulating high-confidence samples to form a sample data set covering multiple operating conditions and multiple current levels.

[0063] The above implementation can also record the timestamp, device identifier and classification information of each sample, so that the sample data set can be preprocessed and divided into training or validation sets (see subsequent steps), while ensuring consistency and traceability with the actual working conditions on site.

[0064] Based on the above implementation method, a sample data set is determined according to sample data within a preset continuous time period, including: Preprocess each sample data within a preset continuous time period to determine the frequency domain feature vector corresponding to each sample data; The sample data set is determined based on the frequency domain feature vector corresponding to each sample data.

[0065] In this embodiment, after the preset continuous duration ends, all sample data during the period (i.e., the working current window that satisfies "the output arc judgment result is normal operation and the confidence value is greater than the preset threshold") are read. Each sample data is preprocessed and feature generated according to the same caliber as offline training to obtain the frequency domain feature vector corresponding to each sample data. Then, all frequency domain feature vectors obtained within the preset continuous duration are summarized at the sample level to form the data basis for subsequent fine-tuning, i.e., the sample data set.

[0066] In one possible implementation, the sample dataset is divided into a training set and a validation set.

[0067] Specifically, the frequency domain feature vectors in the sample data set are divided according to a preset ratio to determine the training set and the validation set.

[0068] For example, assuming there are N×100 sets of frequency domain feature vectors, they can be divided according to a preset ratio of 8:2, where N×80 sets are the training set and N×20 sets are the validation set. It should be noted that the test set can be fixed as the original test set.

[0069] Step 103: Use the sample data set to train the initial arc detection model online to obtain candidate arc detection models.

[0070] In this embodiment of the application, the training set obtained in step 102 (composed of the frequency domain feature vector and its label corresponding to each sample data) is used as input, and the initial arc detection model with the same architecture is loaded for online training.

[0071] Specifically, the parameters of the convolutional layers, residual blocks, and global average pooling layers in the model can be frozen, while only the fully connected classification layer can be set as trainable parameters. A method for calculating training loss information is established, preferably using binary cross-entropy as the loss function (for binary classification of arc faults and normal operation). Different weights can be assigned to positive and negative samples to stabilize training when class imbalance occurs. A gradient-based first-order optimization algorithm is used to update the parameters of the fully connected classification layer. The frequency domain feature vectors of the training set are input into the model, and forward inference is performed batch by batch to obtain the arc judgment results and confidence values. Training loss information is calculated, and backpropagation and parameter updates are performed, with gradient updates performed only on the fully connected classification layer parameters, while the parameters of other layers remain unchanged. When the preset termination condition is met (e.g., reaching the maximum number of rounds or the training loss no longer decreases), the parameters of the best-performing fully connected classification layer during the training phase are combined with the remaining frozen parameters of the initial arc detection model to generate and solidify a candidate arc detection model for subsequent performance evaluation on the validation set.

[0072] In one possible implementation, each sample data in the sample dataset has a real label; step 103 above may specifically include: The sample data in the training set is input into the initial arc detection model to obtain the arc detection judgment result corresponding to each sample data in the training set; based on the arc detection result and the true label corresponding to each sample data in the training set, the first training loss information is determined; based on the first training loss information, the model parameters of the initial arc detection model are updated to obtain the candidate arc detection model. Based on this, the candidate arc detection models can also be tested, that is, to check whether the candidate arc detection models meet the preset requirements and can be put into use. Specifically, this can include: The sample data from the validation set is input into the candidate arc detection model to obtain the second training loss information; When the second training loss information meets the preset loss requirement and the performance index of the candidate arc detection model meets the preset index requirement, the candidate arc detection model is confirmed to meet the preset requirements.

[0073] In another possible implementation, the initial arc detection model includes a fully connected classification layer, and step 103 above may specifically include: The parameters of the fully connected classification layer in the initial arc detection model are updated based on the first training loss information to obtain the updated arc detection model. The sample data from each sample in the validation set are input into the updated arc detection model to determine the third training loss information; If the third training loss information is less than the first training loss information, the updated arc detection model will be identified as a candidate arc detection model.

[0074] Among them, the first training loss information, the second training loss information, and the third training loss information are all used to measure the scalar of the model classification error. In this embodiment, the binary cross-entropy loss function can be used as the function to calculate each training loss information.

[0075] Among them, updating the arc detection model refers to the model obtained by only updating the parameters of the fully connected classification layer (while keeping the convolutional layer, residual block and global average pooling layer frozen).

[0076] For example, the training set obtained in step 102 (the frequency domain feature vector and its label corresponding to each sample) is input into the initial arc detection model in mini-batch; only the fully connected classification layer is unlocked as trainable parameters, the binary cross-entropy is used as the first training loss information, and SGD / Adam is used for backpropagation and parameter update to obtain the updated arc detection model after one iteration.

[0077] Training loss information (illustrated formula):

[0078] in, N 1 represents the number of samples in the validation set; y i For the first i The true binary classification labels of each sample; p(y i ) The model predicts that the sample belongs to the label. y i The probability of.

[0079] The validation set is input into the updated arc detection model, and the third training loss information of the updated arc detection model is calculated. Loss new Simultaneously, the first training loss information of the initial arc detection model is retained. Loss old .calculate:

[0080] If ΔL≥0, then the updated arc detection model is determined as a candidate arc detection model; otherwise, this round of update is considered invalid, the fully connected classification layer parameters are rolled back, and the initial arc detection model is maintained as a candidate arc detection model.

[0081] The above process ensures that updates are retained only if the loss on the validation set does not increase, thus avoiding performance degradation caused by overfitting of the training set and laying the foundation for further acceptance based on performance metrics.

[0082] Step 104: If the candidate arc detection model meets the preset requirements, the candidate arc detection model is determined as the target arc detection model.

[0083] In this embodiment, the validation set is input batch by batch into the candidate arc detection model for forward inference to obtain the arc judgment result and confidence value of each sample. Based on the true label of the validation set and the model output, the second training loss information of the candidate arc detection model on the validation set is first calculated and compared with the first training loss information on the same validation set before fine-tuning. At the same time, the validation set results are evaluated according to the performance indicators preset by the system to obtain the corresponding performance indicator values. The above two items are used as joint acceptance conditions: if and only if "the second training loss information of the validation set is not greater than the first training loss information before fine-tuning" and "the performance indicator of the validation set is not lower than the preset indicator requirements and does not decrease relative to before fine-tuning", the candidate arc detection model is confirmed as the target arc detection model and written into the running system; if either condition is not met, the update in this round is not effective and the previously running model remains unchanged. Through this dual-condition acceptance process, performance fluctuations caused by invalid updates can be suppressed, ensuring the stability and reliability of the written model under field working conditions.

[0084] Specifically, after inputting the validation set into the candidate arc detection model, the method also includes: When a candidate arc detection model does not meet the preset requirements, the model parameters in the candidate arc detection model will be rolled back to the model parameters in the initial arc detection model. The initial arc detection model is determined as the target arc detection model.

[0085] In one possible implementation, the performance metrics of the candidate arc detection model are obtained based on the arc judgment results corresponding to the validation set and the true labels, and the arc judgment results of the validation set are output by the candidate arc detection model; the method further includes: Based on the first and second sample counts, determine the accuracy of the candidate arc detection model; The recall rate of the candidate arc detection model is determined based on the number of the first and third cases. Based on precision and recall, the performance metrics of candidate arc detection models are determined.

[0086] The first number of cases (TP) refers to the number of cases in the validation set where the arc judgment result is arc fault and the true label of the validation set is arc fault; the second number of cases (FP) refers to the number of cases in the validation set where the arc judgment result is arc fault and the true label of the validation set is normal operation; and the third number of cases (FN) refers to the number of cases in the validation set where the arc judgment result is normal operation and the true label of the validation set is arc fault.

[0087] In this embodiment, the validation set is input into the candidate arc detection model batch by batch to obtain the arc judgment result for each sample. After aligning with the true label, TP, FP, and FN are calculated. Then, the accuracy is calculated. Precision Recall rate Recall And performance metrics F1:

[0088]

[0089]

[0090] After calculating F1, it can be determined whether to adopt the candidate arc detection model as the target arc detection model based on whether ΔF1 is greater than or equal to zero, where ΔF1 = F1 new -F1 old F1 old F1 is the performance metric for the initial arc detection model. new The performance metrics for candidate arc detection models are denoted as .

[0091] It should be noted that, to verify the effectiveness of the method proposed in this application, the same dataset was used to train both a regular CNN model and the proposed ResNet-1DCNN model. A real-world field was selected to collect operating current signals to construct a test set. Based on this field test, the edge adaptive fine-tuning method proposed in this application was further applied to update the parameters of the original ResNet-1DCNN model using lightweight optimization, resulting in an adaptively fine-tuned ResNet-1DCNN model. The inference results of the three models on the same test set were statistically analyzed to obtain the corresponding confusion matrices.

[0092] The accuracy of a standard CNN model is calculated from the confusion matrix. Precision =1.000, Recall Rate Recall The accuracy of the ResNet-1DCNN model is 0.909 and the F1 score is 0.952. Precision =1.000, Recall Rate Recall =0.985 and F1 score =0.993, while the accuracy of the model after adaptive fine-tuning is... Precision Recall rate Recall Both the F1 index and the F1 index reached 1.000.

[0093] By comparing the performance of the ordinary CNN, ResNet-1DCNN, and the fine-tuned ResNet-1DCNN on the same test set, it can be seen that the ordinary CNN model is insufficient in terms of recall, with only 0.909, resulting in a large number of false negatives. The ResNet-1DCNN model proposed in this application effectively alleviates the feature degradation problem by introducing an improved residual structure, thereby increasing the recall to 0.985 and achieving a near-perfect F1 score, significantly reducing false negatives. Furthermore, after applying the edge adaptive fine-tuning mechanism proposed in this application, all the model's metrics reach 1.000, achieving completely correct classification of arc fault and normal operation samples.

[0094] This verifies that the ResNet-1DCNN model and edge adaptive fine-tuning mechanism of the method in this application can further improve the classification accuracy and stability of the model, enabling arc detection to maintain high accuracy and strong robustness under different working conditions.

[0095] Step 105: Based on the target arc detection model, perform arc detection on the photovoltaic inverter during operation.

[0096] In this embodiment of the application, the photovoltaic inverter continuously collects DC side operating current data at a sampling rate consistent with offline training during normal operation, and assembles it into a real-time current window (e.g., 1024 points, step size 256 points) at a fixed length. Each time a window is obtained, it is sent to the target arc detection model for forward inference to obtain the arc judgment result (arc fault / normal operation) and confidence value (arc probability) of the window.

[0097] Photovoltaic inverters can compare the arcing probability with a detection threshold: when the arcing probability is not less than the detection threshold, the window is recorded as "arc fault"; otherwise, it is recorded as "normal operation".

[0098] To improve the stability of online judgment, the system can use debouncing or persistent criteria for secondary confirmation. For example, in the sliding window sequence, a "3 / 5" strategy can be applied (at least 3 out of the most recent 5 real-time current windows are arc judgments), and an arc fault event is triggered when the shortest duration (e.g., ≥100 ms) is met. After triggering, an alarm message is immediately output and the protection logic is activated to perform disconnection control, recording the event timestamp, current current level, confidence value of the model output, and relevant operating parameters. If the triggering condition is not met, the system maintains normal operation and continues to process subsequent real-time current windows.

[0099] To facilitate post-event traceability and operational analysis, the system can save the original data and model output of several windows before and after the arc event according to a set buffer depth (e.g., 10 windows before and after the event). Simultaneously, the version number of the target arc detection model used in this online detection and the threshold configuration are written into the event log. This process is executed cyclically to achieve online arc detection and protection linkage based on the target arc detection model, providing real-time and reliable arc fault identification capabilities without changing the inverter's main control strategy.

[0100] In the above method embodiment, firstly, the operating current data is input into the initial arc detection model deployed after offline training on the inverter side, and the arc judgment result and corresponding confidence level are output in real time, providing an objective basis for subsequent sample selection based on confidence level. Subsequently, within the preset continuous duration of the first run, only when the model determines that it is operating normally and the confidence level is greater than the preset threshold, the corresponding current data is stored as sample data, forming a sample dataset that fits the real operating conditions. This dataset is then merged to train the initial arc detection model online on the device side, obtaining candidate arc detection models. If the candidate arc detection models meet the preset requirements, they are determined as the target arc detection model. During the online training process, the model parameters are made to adapt in real time to the fluctuations of the power grid, load disturbances, and noise levels, significantly reducing the distribution difference between laboratory and power plant operating conditions. Finally, the target arc detection model is used to perform online detection. Without relying on the cloud, the model gradually adapts to the field data, enhancing the robustness and consistency of the model to the fluctuations of the power grid, load disturbances, and noise changes, thereby improving environmental adaptability in the actual photovoltaic power plant environment and reducing false alarms and false negatives.

[0101] See Figure 4 The overall flowchart of the arcing detection method for inverters is shown.

[0102] First, data acquisition was conducted in a laboratory environment: arc current data and normal current data were acquired using an arc-generating machine; exponentially weighted moving average filtering was applied to the acquired current data to suppress transient noise and power frequency disturbances; the filtered sequence was then subjected to a fast Fourier transform to obtain spectral data, which served as the frequency domain feature vector for subsequent network input. After annotation, the original current dataset was divided into an initial training set, an initial validation set, and an initial test set in a 4:3:3 ratio; a one-dimensional convolutional neural network (ResNet-1DCNN) based on an improved residual structure was trained using the initial training set as input; hyperparameters were determined using the initial validation set; and performance was confirmed on the initial test set, resulting in the initial arc detection model. Subsequently, this initial arc detection model was deployed to the arc detection module of a photovoltaic inverter for preliminary inference.

[0103] During the edge adaptive fine-tuning phase, the inverter operates in the field for an extended period: the system automatically collects high-confidence normal operation data covering different current levels over several consecutive days, forming a sample data set; this set is divided into a training set and a validation set in an 8:2 ratio, while the test set remains fixed as the original test set. During fine-tuning, feature extraction layers such as convolutional layers and residual blocks are frozen, and only the fully connected classification layer is unlocked as trainable parameters; iterative updates are performed using the training set, and a model detection mechanism based on the validation set loss is set up. After each update, the validation set is first used to determine if the loss "does not increase / meets the standard." If it passes, the model is then reviewed on the fixed original test set. If the performance index of the test set does not decrease compared to the baseline (meets the preset performance index), the weights of the fully connected layer are updated and written, resulting in the target arc detection model; otherwise, the model is rolled back, maintaining the original running model (i.e., the initial arc detection model).

[0104] Finally, the inverter enters long-term online monitoring: it infers the arc fault probability (i.e., confidence value) for a fixed-length real-time current window; when the probability is not less than the detection threshold and the continuous criterion is met, an alarm / disconnection is triggered; the system periodically acquires field samples according to the above strategy and triggers fine-tuning and acceptance, thereby achieving adaptive updates and stable arc detection on site without relying on the cloud.

[0105] See Figure 5 The diagram illustrates the specific process flow of the edge adaptive fine-tuning mechanism in the overall flowchart.

[0106] In this embodiment, during the initial field run, the ResNet-1DCNN initial arc detection model, trained and deployed offline, is used to infer the real-time current window and calculate the probability of normal operation. p normal ;when p normal When the current value is greater than 0.9, current segments of 500 ms length are written to the cache in units of 1A and automatically labeled as "normal operation," accumulating to form a sample dataset. After the runtime reaches a preset number of days, the sample data is sequentially subjected to exponentially weighted moving average filtering and fast Fourier transform to obtain N×100 sets of current spectrum feature vectors. Subsequently, the samples are proportionally allocated into a training set (N×80 sets) and a validation set (N×20 sets), with the test set fixed as the original test set. During fine-tuning, convolutional layers and residual blocks are frozen, and only fully connected layers are made trainable. If the change in validation set loss ΔL≥0, the test set is reviewed, and the performance increment ΔF1 is calculated. When ΔF1≥0, the fully connected layer parameters are updated to obtain the target arc detection model and it is put into operation; otherwise, it is rolled back to the initial arc detection model.

[0107] See Figure 6The diagram shows a schematic of the arc detection device for an inverter provided in an embodiment of this application; for ease of explanation, only the parts related to the embodiment of this application are shown.

[0108] The inverter arcing detection device 600 includes: The data input module 601 is used to input the operating current data of the photovoltaic inverter into the initial arc detection model. After offline training, the initial arc detection model is deployed to the photovoltaic inverter to infer the operating current data and output the arc judgment result and the corresponding confidence value. The arc judgment result includes normal operation and arc fault. The sample acquisition module 602 is used to store the corresponding operating current data as sample data when the arc judgment result output by the initial arc detection model is normal operation and the confidence value is greater than the preset threshold during the preset continuous time of the first operation of the photovoltaic inverter, so as to form a sample data set collected within the preset continuous time. The online training module 603 is used to train the initial arc detection model online using the sample data set to obtain candidate arc detection models. The performance verification module 604 is used to determine the candidate arc detection model as the target arc detection model if the candidate arc detection model meets the preset requirements. The arc detection module 605 is used to perform arc detection on photovoltaic inverters during operation based on a target arc detection model.

[0109] In this embodiment, the sample dataset includes a training set and a validation set, and each sample data in the sample dataset has a real label; the online training module 603 includes: The initial detection unit is used to input the sample data in the training set into the initial arc detection model to obtain the arc judgment result corresponding to each sample data in the training set. The first loss determination unit is used to determine the first training loss information based on the arc judgment result and the true label corresponding to each sample data in the training set. The candidate determination unit is used to update the model parameters of the initial arc detection model based on the first training loss information to obtain the candidate arc detection model; Correspondingly, the inverter's arcing detection device 600 also includes: The second loss determination unit is used to input the sample data in the validation set into the candidate arc detection model to obtain the second training loss information; The model detection unit is used to confirm that the candidate arc detection model meets the preset requirements when the second training loss information meets the preset loss requirements and the performance index of the candidate arc detection model meets the preset index requirements.

[0110] In this embodiment, the initial arc detection model includes a fully connected classification layer; the candidate determination unit is used for: The parameters of the fully connected classification layer in the initial arc detection model are updated based on the first training loss information to obtain the updated arc detection model. The sample data from each sample in the validation set are input into the updated arc detection model to determine the third training loss information; If the third training loss information is less than the first training loss information, the updated arc detection model will be identified as a candidate arc detection model.

[0111] In this embodiment of the application, the arcing detection device 600 for the inverter further includes: The parameter rollback judgment module is used to roll back the model parameters in the candidate arc detection model to the model parameters in the initial arc detection model when the candidate arc detection model does not meet the preset requirements. The first target determination module is used to determine the initial arc detection model as the target arc detection model.

[0112] In this embodiment, the performance index of the candidate arc detection model is obtained based on the arc judgment result corresponding to the validation set and the true label, and the arc judgment result of the validation set is output by the candidate arc detection model; the arc detection device 600 of the inverter also includes: The accuracy determination module is used to determine the accuracy of the candidate arc detection model based on the first number of cases and the second number of cases. The first number of cases refers to the number of cases in the validation set where the arc judgment result is arc fault and the true label of the validation set is arc fault. The second number of cases refers to the number of cases in the validation set where the arc judgment result is arc fault and the true label of the validation set is normal operation. The recall determination module is used to determine the recall rate of the candidate arc detection model based on the first and third counts; the third count refers to the number of cases in the validation set where the arc judgment result is normal operation and the true label of the validation set is arc fault. The performance metrics determination module is used to determine the performance metrics of candidate arc detection models based on precision and recall.

[0113] In this embodiment of the application, the sample acquisition module 602 may specifically include: The tracking control unit is used to ensure that the DC-side operating point of the photovoltaic inverter runs along the maximum power point tracking trajectory under the action of maximum power point tracking control during a preset continuous period of time during the first operation of the photovoltaic inverter. The grading unit is used to use the working current data on the maximum power point tracking trajectory as a grading reference, set the current range to the first preset current to the maximum working current reached on the day, and generate a grading set according to the second preset current as the step. The storage unit is used to traverse the graded set. Within each grade, when the arc judgment result output by the initial arc detection model is normal and the confidence value is greater than the preset threshold, the corresponding working current data is stored as sample data. The sample determination unit is used to determine the sample data set based on the sample data within a preset continuous time period.

[0114] In this embodiment of the application, the offline training steps of the initial arc detection model include: Collect and label normal current data and arc current data; The normal current data and arc current data are preprocessed to determine the frequency domain feature vectors corresponding to the normal current data and the arc current data, so as to obtain the initial sample data set. The initial sample dataset is divided into an initial training set, an initial validation set, and an initial test set. The arc detection network is trained using the initial training set as input, the network parameters of the arc detection network are determined based on the initial validation set, and the performance index of the arc detection network under the network parameters is determined based on the initial test set. When the performance indicators of the arc detection network meet the preset requirements, the arc detection network is determined as the initial arc detection model.

[0115] The arcing detection device 600 for inverters provided in this application embodiment can be applied to the arcing detection method for inverters provided in the foregoing embodiments. For details, please refer to the description of the arcing detection method for inverters provided in the foregoing embodiments, which will not be repeated here.

[0116] See Figure 7 The diagram illustrates the structure of an electronic device according to an embodiment of this application. Figure 7 As shown, the electronic device 700 of this embodiment includes: at least one processor 710 ( Figure 7 Only one is shown in the diagram), memory 720, and computer program 721 stored in memory 720 and executable on at least one processor 710. When processor 710 executes computer program 721, it implements the steps in the above-described inverter arc detection method embodiment.

[0117] Electronic device 700 can be a server, physical server, computing device, etc. This electronic device may include, but is not limited to, a processor 710 and a memory 720. Those skilled in the art will understand that... Figure 7 This is merely an example of electronic device 700 and does not constitute a limitation on electronic device 700. It may include more or fewer components than shown, or combine certain components, or different components. For example, it may also include input / output devices, network access devices, etc.

[0118] The processor 710 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0119] In some embodiments, memory 720 may be an internal storage unit of electronic device 700, such as a hard disk or memory of electronic device 700. In other embodiments, memory 720 may be an external storage device of electronic device 700, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on electronic device 700. Furthermore, memory 720 may include both internal and external storage units of electronic device 700. Memory 720 is used to store operating system, application programs, boot loader, data, and other programs, such as program code of computer programs. Memory 720 may also be used to temporarily store data that has been output or will be output.

[0120] In specific implementations, the processor 710, memory 720, and computer program 721 described in the embodiments of this application can execute the embodiments of the arc detection method for the inverter of this application, which will not be repeated here.

[0121] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0122] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0123] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0124] In the embodiments provided in this application, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0125] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0127] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0128] The implementation of all or part of the processes in the methods of the above embodiments can also be accomplished by a computer program product. When the computer program product is run on an electronic device, the electronic device can implement the steps in the various method embodiments described above.

[0129] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for detecting arcing in an inverter, characterized in that, The method includes: The operating current data of the photovoltaic inverter is input into the initial arcing detection model. The initial arcing detection model is deployed to the photovoltaic inverter after offline training. It is used to infer the operating current data and output the arc judgment result and the corresponding confidence value. The arc judgment result includes normal operation and arc fault. During the preset continuous period of the first operation of the photovoltaic inverter, when the arc judgment result output by the initial arc detection model is normal operation and the confidence value is greater than the preset threshold, the corresponding operating current data is stored as sample data to form a sample data set collected within the preset continuous period. The initial arc detection model is trained online using the sample data set to obtain candidate arc detection models; If the candidate arc detection model meets the preset requirements, the candidate arc detection model is determined as the target arc detection model; Based on the target arc detection model, arc detection is performed on the photovoltaic inverter during operation.

2. The method as described in claim 1, characterized in that, The sample dataset includes a training set and a validation set, and each sample in the sample dataset has a true label; The initial arc detection model is trained online using the aforementioned sample dataset to obtain candidate arc detection models, including: The sample data in the training set is input into the initial arc detection model to obtain the arc judgment result corresponding to each sample data in the training set; Based on the arc judgment results and true labels corresponding to each sample data in the training set, the first training loss information is determined; The model parameters of the initial arc detection model are updated based on the first training loss information to obtain a candidate arc detection model; After training the initial arc detection model online using the sample data set to obtain candidate arc detection models, the method further includes: The sample data in the validation set is input into the candidate arc detection model to obtain the second training loss information; When the second training loss information meets the preset loss requirement and the performance index of the candidate arc detection model meets the preset index requirement, the candidate arc detection model is confirmed to meet the preset requirements.

3. The method as described in claim 2, characterized in that, The initial arc detection model includes a fully connected classification layer; updating the model parameters of the initial arc detection model based on the first training loss information to obtain a candidate arc detection model includes: Based on the first training loss information, the parameters of the fully connected classification layer in the initial arc detection model are updated to obtain the updated arc detection model. The sample data from each sample in the validation set are input into the updated arc detection model to determine the third training loss information; If the third training loss information is less than the first training loss information, the updated arc detection model is determined as the candidate arc detection model.

4. The method as described in claim 2, characterized in that, The method further includes: When the candidate arc detection model does not meet the preset requirements, the model parameters in the candidate arc detection model are rolled back to the model parameters in the initial arc detection model; The initial arc detection model is determined to be the target arc detection model.

5. The method as described in claim 2, characterized in that, The performance metrics of the candidate arc detection model are obtained based on the arc judgment results and the true labels corresponding to the validation set, and the arc judgment results of the validation set are output by the candidate arc detection model; the method further includes: The accuracy of the candidate arc detection model is determined based on the first and second counts; the first count refers to the number of cases in the validation set where the arc judgment result is arc fault and the true label of the validation set is arc fault, and the second count refers to the number of cases in the validation set where the arc judgment result is arc fault and the true label of the validation set is normal operation. The recall rate of the candidate arc detection model is determined based on the first and third counts; the third count refers to the number of cases in the validation set where the arc judgment result is normal operation and the true label of the validation set is arc fault. Based on the precision and recall, the performance metrics of the candidate arc detection model are determined.

6. The method as described in claim 1, characterized in that, During the preset continuous operating time of the photovoltaic inverter, when the arc judgment result output by the initial arc detection model indicates normal operation and the confidence value is greater than a preset threshold, the corresponding operating current data is stored as sample data to form a sample data set collected within the preset continuous operating time, including: During the preset continuous duration of the first operation of the photovoltaic inverter, under the action of maximum power point tracking control, the DC side operating point of the inverter is made to run along the maximum power point tracking trajectory; Using the operating current data on the maximum power point tracking trajectory as a grading reference, the current range is set from the first preset current to the maximum operating current reached on the day, and a grading set is generated according to the second preset current as the step. Traverse the set of grades. In each grade, when the arc judgment result output by the initial arc detection model is normal operation and the confidence value is greater than the preset threshold, store the corresponding working current data as sample data. The sample data set is determined based on the sample data within the preset continuous time period.

7. The method as described in claim 1, characterized in that, The offline training steps for the initial arc detection model include: Collect and label normal current data and arc current data; The normal current data and the arc current data are preprocessed to determine the frequency domain feature vectors corresponding to the normal current data and the arc current data, so as to obtain an initial sample data set. The initial sample dataset is divided into an initial training set, an initial validation set, and an initial test set. The arc detection network is trained using the initial training set as input, the network parameters of the arc detection network are determined based on the initial validation set, and the performance index of the arc detection network under the network parameters is determined based on the initial test set. When the performance indicators of the arc detection network meet the preset requirements, the arc detection network is determined as the initial arc detection model.

8. An arcing detection device for an inverter, characterized in that, The device includes: The data input module is used to input the operating current data of the photovoltaic inverter into the initial arc detection model. The initial arc detection model is deployed to the photovoltaic inverter after offline training. It is used to infer the operating current data and output the arc judgment result and the corresponding confidence value. The arc judgment result includes normal operation and arc fault. The sample acquisition module is used to store the corresponding operating current data as sample data when the arc judgment result output by the initial arc detection model is normal operation and the confidence value is greater than a preset threshold during the preset continuous operation of the photovoltaic inverter, so as to form a sample data set acquired within the preset continuous operation period. The online training module is used to train the initial arc detection model online using the sample data set to obtain candidate arc detection models; The performance verification module is used to determine the candidate arc detection model as the target arc detection model when the candidate arc detection model meets the preset requirements. The arc detection module is used to perform arc detection on the photovoltaic inverter during operation based on the target arc detection model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images