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

Abstract

The application is suitable for the technical field of inverter detection, and provides a method and device for arc detection of an inverter, an electronic device and a storage medium. The method comprises: inputting working current data of a photovoltaic inverter into an initial arc detection model; within a preset continuous time length of first operation, if the output result is normal operation and the confidence value is greater than a threshold value, corresponding working current data is taken as a sample to form a sample data set; the initial arc detection model is trained online using the sample data set to obtain a candidate arc detection model; in the case that the candidate arc detection model meets a preset requirement, the candidate arc detection model is determined as a target arc detection model; finally, the real-time current window is detected based on the target arc detection model in online operation. The above scheme can make the model parameters adapt to the field power grid fluctuation, load disturbance and noise level in real time through online training on site, and improve the arc detection precision.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of inverter detection, and particularly relates to an arc detection method and device for an inverter, an electronic device, and a storage medium. BACKGROUND

[0002] Arc fault on the DC side of a photovoltaic inverter is a major hidden danger that causes fire, and engineering requires it to have reliable detection capability. However, the actual power station environment is complex and changeable, and power grid fluctuations, load disturbances, and noise interference are significant. The existing mainstream approach is mostly based on a fixed threshold: filtering and transforming the current signal to extract harmonics or spectral energy, and comparing it with a preset threshold to determine whether there is an arc. This method is simple to implement and has low overhead, but the threshold is extremely sensitive to on-site power grid fluctuations, load disturbances, and noise, and is prone to false positives, false negatives, and insufficient environmental adaptability.

[0003] Some research has also used machine learning or deep learning for automatic feature extraction and classification, but generally relies on laboratory data sets for training, and still has defects such as weak environmental adaptability and inconvenient model updates in actual photovoltaic power station applications. SUMMARY

[0004] Therefore, the embodiments of the present application provide an arc detection method and device for an inverter, an electronic device, and a storage medium, which collect on-site samples within a preset continuous duration of the first operation and adaptively fine-tune on the edge, realize model updates without the cloud, and thus improve environmental adaptability and reduce false positives and false negatives.

[0005] A first aspect of the embodiments of the present application provides an arc detection method for an inverter, which comprises:

[0006] inputting working current data of a photovoltaic inverter into an initial arc detection model, wherein the initial arc detection model is deployed to the photovoltaic inverter after offline training, used for reasoning the working current data and outputting an arc judgment result and a corresponding confidence value; the arc judgment result comprises normal operation and arc fault;

[0007] within a preset continuous duration 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 a preset threshold, storing the corresponding working current data as sample data to form a sample data set collected within the preset continuous duration;

[0008] training the initial arc detection model online using the sample data set to obtain a candidate arc detection model;

[0009] in the case where the candidate arc detection model meets a preset requirement, determining the candidate arc detection model as a target arc detection model;

[0010] Detecting an arc based on the target arc detection model.

[0011] The embodiment of the present application firstly inputs the working current data on the inverter side into the initial arc detection model deployed after offline training, and outputs the arc judgment result and the corresponding confidence in real time, which provides an objective basis for subsequent sample screening based on the confidence; then within the preset continuous time of the first operation, only when the model determines normal operation and the confidence is greater than the preset threshold, the corresponding current data is stored as sample data, forming a sample data set that fits the real working condition, and the initial arc detection model is trained online at the device end based on this, to obtain a candidate arc detection model, and in the case that the candidate arc detection model meets the preset requirements, the candidate arc detection model is determined as the target arc detection model, so that the model parameters adapt to the real-time fluctuations of the field power grid, load disturbance and noise level during the online training process, significantly reducing the distribution difference between the laboratory and the power station working condition; finally, the target arc detection model performs online detection, and the model gradually adapts to the field data without relying on the cloud, thereby enhancing the robustness and consistency of the power grid fluctuations, load disturbance and noise changes, improving the environmental adaptability in the actual photovoltaic power station environment, and reducing false positives and false negatives.

[0012] In a possible implementation, the sample data set includes a training set and a validation set, and each sample data in the sample data set has a true label; after the initial arc detection model is trained online using the sample data set to obtain a candidate arc detection model, the method further includes:

[0013] Inputting 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;

[0014] Based on the arc judgment result corresponding to each sample data in the training set and the true label, determining first training loss information;

[0015] Updating the model parameters of the initial arc detection model based on the first training loss information to obtain a candidate arc detection model;

[0016] Inputting the sample data in the validation set into the candidate arc detection model to obtain second training loss information;

[0017] 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, it is confirmed that the candidate arc detection model meets the preset requirement.

[0018] In a possible implementation, the initial arc-detection model comprises a fully connected classification layer; and updating the model parameters of the initial arc-detection model based on the first training loss information to obtain a candidate arc-detection model comprises:

[0019] updating the parameters of the fully connected classification layer in the initial arc-detection model based on the first training loss information to obtain an updated arc-detection model;

[0020] inputting each sample data in the verification set into the updated arc-detection model to determine third training loss information;

[0021] in a case where the third training loss information is less than the first training loss information, determining the updated arc-detection model as the candidate arc-detection model.

[0022] In a possible implementation, the performance indicator of the candidate arc-detection model is obtained based on the arc judgment result of the verification set and the real label, and the arc judgment result of the verification set is output by the candidate arc-detection model; and the method further comprises:

[0023] determining the precision of the candidate arc-detection model according to a first number and a second number, the first number being a number of cases in which the arc judgment result of the verification set is an arc fault and the real label of the verification set is an arc fault, and the second number being a number of cases in which the arc judgment result of the verification set is an arc fault and the real label of the verification set is normal operation;

[0024] determining the recall rate of the candidate arc-detection model according to the first number and a third number, the third number being a number of cases in which the arc judgment result of the verification set is normal operation and the real label of the verification set is an arc fault;

[0025] determining the performance indicator of the candidate arc-detection model based on the precision and the recall rate.

[0026] In a possible implementation, 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 within the preset continuous duration of the first operation of the photovoltaic inverter, the corresponding working current data is stored as sample data to form a sample data set collected within the preset continuous duration, comprising:

[0027] within the preset continuous duration of the first operation of the photovoltaic inverter, under the maximum power point tracking control, the operating point of the direct current side of the inverter is caused to operate along a maximum power point tracking trajectory;

[0028] The working current data on the maximum power point tracking trajectory is taken as a grading reference, a current range is set as a first preset current to a maximum working current reached on the day, and a grading set is generated according to a second preset current as a step;

[0029] The grading set is traversed, and 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 value in each grading, the corresponding working current data is stored as sample data;

[0030] The sample data set is determined according to the sample data in the preset continuous time length.

[0031] In a possible implementation, the offline training step of the initial arc detection model includes:

[0032] Normal current data and arc current data are collected and labeled;

[0033] The normal current data and the arc current data are preprocessed to determine the frequency domain feature vector corresponding to the normal current data and the frequency domain feature vector corresponding to the arc current data, so as to obtain an initial sample data set;

[0034] The initial sample data set is divided into an initial training set, an initial validation set, and an initial test set;

[0035] The initial training set is taken as input to train the arc detection network, 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;

[0036] When the performance index of the arc detection network meets a preset index requirement, the arc detection network is determined as the initial arc detection model.

[0037] A second aspect of the embodiment of the application provides an arc detection device of an inverter, and the device includes:

[0038] A data input module is configured to input working current data of a photovoltaic inverter to an initial arc detection model, the initial arc detection model is deployed to the photovoltaic inverter after offline training, and is configured to perform inference on the working current data and output an arc judgment result and a corresponding confidence value; the arc judgment result includes normal operation and arc failure;

[0039] The sample collection module is configured to, within a preset continuous time length when the photovoltaic inverter is first operated, store working current data as sample data when the initial arc detection model outputs an arc judgment result of normal operation and the confidence value is greater than a preset threshold, to form a sample data set collected within the preset continuous time length.

[0040] The online training module is configured to perform online training on the initial arc detection model using the sample data set to obtain a candidate arc detection model.

[0041] The performance verification module is configured to, in a case where the candidate arc detection model meets a preset requirement, determine the candidate arc detection model as a target arc detection model.

[0042] The arc detection module is configured to perform arc detection on the photovoltaic inverter in operation based on the target arc detection model.

[0043] A third aspect of the embodiments of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the method of the first aspect described above when executing the computer program.

[0044] A fourth aspect of the embodiments of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method of the first aspect described above.

[0045] A fifth aspect of the embodiments of the present application provides a computer program product, which, when executed on an electronic device, causes the electronic device to perform the steps of the method of the first aspect described above.

[0046] The beneficial effects of the second aspect to the fifth aspect described above can refer to the beneficial effects of the first aspect described above, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0048] Figure 1 is a flowchart of an arc detection method of an inverter provided by the embodiments of the present application;

[0049] Figure 2FIG. 1 is a structural schematic diagram of a residual block of an arc detection network in an embodiment of the present application;

[0050] Figure 3 FIG. 2 is a schematic diagram of the overall structure of an arc detection network in an embodiment of the present application;

[0051] Figure 4 FIG. 3 is a schematic diagram of the overall flow of an arc detection method for an inverter;

[0052] Figure 5 FIG. 4 is a schematic diagram of the specific flow of an edge adaptive fine-tuning mechanism in the overall flow;

[0053] Figure 6 FIG. 5 is a structural schematic diagram of an arc detection device for an inverter provided in an embodiment of the present application;

[0054] Figure 7 FIG. 6 is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0055] In the following description, specific details are set forth, such as particular system configurations, techniques, etc., in order to provide a thorough understanding of the present application. However, persons skilled in the art will understand that the present application can be practiced without these specific details. In other instances, well-known structures, devices, circuits, and methods have not been described in detail in order to avoid obscuring the present application.

[0056] It should be understood that the term “includes” when used in the present application and the appended claims indicates the presence of the described features, integers, steps, operations, elements, and / or components but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0057] It should also be understood that the term “and / or” when used in the present application and the appended claims indicates any combination of the associated listed items as well as all possible combinations of the items, and includes these combinations.

[0058] As used in the present application and the appended claims, the term “if’ can be interpreted as meaning “when” or “once” or “in response to a determination” or “in response to detecting” depending on the context. Similarly, the phrase “if determined” or “if detected [the described condition or event]” can be interpreted as meaning “once determined” or “in response to a determination” or “once detected [the described condition or event]” or “in response to detecting [the described condition or event]” depending on the context.

[0059] In addition, in the description in the specification and the attached claims of the present application, the terms "first", "second", "third" and the like are only used to distinguish descriptions, and cannot be understood as indicating or implying relative importance.

[0060] It should be understood that the magnitude of the serial number of each step in the embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0061] The arc fault on the DC side of the photovoltaic inverter is the main hidden danger of fire, and engineering requires it to have reliable detection capability. However, the actual power station environment is complex and changeable, and the power grid fluctuation, load disturbance and noise interference are significant. The existing mainstream method is mostly based on a fixed threshold: filtering, transforming and then extracting the harmonic or frequency spectrum energy of the current signal, and comparing it with the preset threshold to determine whether there is an arc. This method is simple to implement and low in cost, but the threshold is very sensitive to the power grid fluctuation, load disturbance and noise in the field, and false positives, false negatives and insufficient environmental adaptability are prone to occur.

[0062] There are also studies that use machine learning or deep learning for automatic feature extraction and classification, but they generally rely on laboratory data sets for training, and still have defects such as weak environmental adaptability and inconvenient model updating in actual photovoltaic power station applications.

[0063] In order to solve the above problems, the present application provides an arc detection method and device for an inverter, an electronic device and a storage medium. The arc detection method for the inverter in the present application first inputs the working current data on the inverter side into the initial arc detection model deployed after offline training, and outputs the arc judgment result and the corresponding confidence in real time, providing an objective basis for subsequent sample selection based on confidence; then within a preset continuous time length of the first run, only when the model determines normal operation and the confidence is greater than a preset threshold, the corresponding current data is stored as sample data, forming a sample data set that fits the real working conditions and using it to perform online training of the initial arc detection model on the device side to obtain a candidate arc detection model, and in the case that the candidate arc detection model meets the preset requirements, the candidate arc detection model is determined as the target arc detection model, so that the model parameters adapt to the power grid fluctuation, load disturbance and noise level in real time during the online training process, significantly reducing the distribution difference between the laboratory and the power station working conditions; finally, the target arc detection model performs online detection, the model gradually adapts to the on-site data without relying on the cloud, and the robustness and consistency to the power grid fluctuation, load disturbance and noise change are enhanced, thereby improving the environmental adaptability in the actual photovoltaic power station environment and reducing false positives and false negatives.

[0064] The arc detection method, device, electronic device, storage medium and computer program provided by the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0065] Referring to Figure 1 , a flowchart of a method for detecting arc in an inverter is shown; as Figure 1 shown, the method can include the following steps:

[0066] Step 101, input the working current data of the photovoltaic inverter to the initial arc detection model.

[0067] Wherein, the initial arc detection model is deployed to the photovoltaic inverter after offline training, used for inferring the working current data and outputting the arc judgment result and the corresponding confidence value.

[0068] Wherein, the arc judgment result includes normal operation and arc fault, and the confidence value is used to represent the credibility of the output arc judgment result.

[0069] In the embodiments of the present application, the working current data of the photovoltaic inverter can be the current data obtained by sampling the DC side current of the photovoltaic inverter converted into analog voltage by the transformer or Hall sensor at a fixed sampling rate by the embedded processing unit of the photovoltaic inverter.

[0070] In one possible implementation, the initial arc detection model is deployed to the photovoltaic inverter after offline training and quantization compression.

[0071] Specifically, the inverter includes an arc detection module, which can preload the initial arc detection model trained offline and completed quantization compression, then assemble the working current data into real-time current window (length of 1024 points) according to fixed length, as the input of the initial arc detection model, and then execute forward inference, output the arc judgment result of "arc fault / normal operation" and the corresponding confidence value (i.e. arc probability); in order to ensure the consistency of subsequent processing, this embodiment records and submits the time stamp and current grading information together after completing the inference of each window to the upper layer logic for subsequent steps to acquire samples and control alarm according to threshold strategy; the above process is executed in the photovoltaic inverter, so as to realize the embedded edge deployment of quantization model and the real-time inference of 1024 point current window, and provide stable and consistent discrimination and confidence value output for subsequent sample collection and fine tuning.

[0072] In one possible implementation, the offline training step of the initial arc detection model includes:

[0073] Collect and label normal current data and arc current data;

[0074] Preprocess the normal current data and the arc current data, determine the frequency domain feature vector corresponding to the normal current data and the frequency domain feature vector corresponding to the arc current data, to obtain an initial sample data set;

[0075] Divide the initial sample data set into an initial training set, an initial validation set and an initial test set;

[0076] Train the arc drawing detection network with the initial training set as input, determine the network parameters of the arc drawing detection network based on the initial validation set, and determine the performance index of the arc drawing detection network under the network parameters based on the initial test set;

[0077] When the performance index of the arc drawing detection network meets the preset index requirement, the arc drawing detection network is determined as an initial arc drawing detection model.

[0078] The normal current data refers to the alternating current quantity sequence of the DC side collected under the condition of no arc; the arc current data refers to the alternating current quantity sequence of the DC side collected under the condition of arc fault; the frequency domain feature vector refers to the current spectrum sequence obtained by preprocessing and fast Fourier transform of the current data, which is used as the input feature of the model; 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, the initial validation set and the initial test set refer to three subsets obtained by dividing the initial sample data set according to a preset proportion, which are respectively used for model training, parameter selection and performance confirmation.

[0079] The arc drawing detection network is a neural network for feature extraction and classification of input frequency domain feature vectors; the initial arc drawing detection model refers to the arc drawing detection network whose parameters are fixed after offline training and performance confirmation, which is used for subsequent deployment and reasoning.

[0080] Exemplarily, first, in a laboratory environment, the arc drawing machine can be used to simulate arc fault, and the normal current data and the arc current data can be collected in the range of 3A to the maximum working current I max of the DC current with a step of 1A; 10 groups of current data are collected at each current level, each group has a duration of 500ms and a sampling rate of 250kHz, to obtain the original current data I original [n].

[0081] Secondly, the I original [n] is subjected to exponential weighted moving average filtering to obtain the current data filtered and denoised I filter [n], and the calculation formula is as follows:

[0082]

[0083] wherein, I filter [n] is the filtered current data at the current time; I filter [n-1] is the filtered current data at the previous time; I original [n] is the original current data at the current time; and a e (0, 1) is a smoothing coefficient for controlling the weight of the current data.

[0084] Then, the fast Fourier transform is performed on I filter [n] to obtain the frequency spectrum data F[k] for training, with the sampling point number K = 1024:

[0085]

[0086] wherein F[k] of each sample is taken as the corresponding frequency domain feature vector, K is the length, k is the current frequency domain component index, and m is the current time domain sampling point index.

[0087] After obtaining the frequency domain feature vectors corresponding to the normal current data and the arc current data, the two are combined to form an initial sample data set. The initial sample data set is divided into an initial training set, an initial validation set and an initial test set according to the proportions 4:3:3.

[0088] The initial training set is taken as the input to train the arc detection network: first, local features are extracted through a convolutional layer, then deep layer feature extraction is performed through three residual blocks, and then the arc detection result is output through a global average pooling layer and a fully connected classification layer; the network parameters and training hyperparameters are determined based on the initial validation set, and the performance indicators under the parameters are evaluated on the initial test set; when the performance indicators meet the preset indicator requirements, the arc detection network is solidified as an initial arc detection model.

[0089] In one possible implementation, the residual block structure of the arc detection network is as shown in Figure 2 The input feature x corresponding to the output or input sequence of the previous layer is fed into the residual block, first passing through a first convolutional layer, batch normalization and ReLU in sequence to obtain intermediate features; then a second convolutional layer is used to obtain a residual mapping F(x). At the same time, x is mapped to form a direct branch, and the residual mapping is summed in the element dimension to obtain H(x) = F(x) + x; the output feature H(x) is obtained through ReLU activation and fed to the next network layer.

[0090] To ensure the feasibility of element-wise addition, when the number of channels and the stride are consistent, Figure 2The identity mapping is shown; in the case of changes in the number of channels or time dimension resolution, a 1x1 one-dimensional convolution (stride consistent with the main branch) can be configured on the direct branch to realize channel / stride matching, then added to F(x), and then output through ReLU.

[0091] The one-dimensional convolution kernel length is preferably 3, the stride is 1, and the ends are zero-padded to maintain the length. Batch normalization is used to stabilize training and suppress internal covariant bias. ReLU is used as a nonlinear mapping to enhance feature expression capability.

[0092] At the network level, multiple residual blocks are stacked in sequence to form a "multi-layer residual block" for hierarchical deep feature extraction of the current frequency domain feature vector. A global average pooling layer is then set to average each channel along the time-frequency dimension to obtain a fixed-length representation vector, and then a fully connected classification layer is connected to output the arc fault or normal operation arc judgment result and the corresponding confidence value. In the training stage, the input aperture is consistent with the data preprocessing, and the loss function used to calculate the loss information can use the binary cross-entropy function. When deployed, the quantized and compressed network is loaded as the initial arc detection model on the photovoltaic inverter end to perform forward inference on the fixed-length (e.g., 1024-point) real-time current window and generate an arc probability for subsequent sample data acquisition and online determination.

[0093] Based on the above embodiment, the overall structure of the arc detection network is shown in Figure 3 The network is based on an improved residual structure one-dimensional convolutional neural network (ResNet-1DCNN). The network receives the frequency domain feature vector F[k] (length K, e.g., K=1024) obtained by preprocessing at the input layer, and then extracts low-level local features through convolution, batch normalization, and ReLU in sequence. A shortcut branch is set at the end and added element-wise to the main branch, forming the first residual fusion to stabilize deep training and alleviate gradient vanishing.

[0094] Subsequently, the network includes three groups of residual blocks in cascade (corresponding Figure 3 from left to right). Each group of residual blocks has "convolution layer -> batch normalization -> ReLU -> convolution layer -> (added to the shortcut) -> ReLU" as the basic unit:

[0095] When the number of input and output channels of the residual block and the time step are consistent, the shortcut branch uses an identity mapping to directly add the output of the main branch. When the number of channels or steps changes, the shortcut branch uses a 1x1 one-dimensional convolution (stride consistent with the main branch) to realize dimension matching before adding. The convolution kernel length is preferably 3, the stride is 1, and the ends are zero-padded to maintain the length. Batch normalization is used to suppress internal covariant bias, and ReLU provides nonlinear mapping.

[0096] After the hierarchical deep refinement of the frequency domain features of multiple groups of residual blocks, the network enters the global average pooling layer, and the mean value of each channel is obtained along the time or frequency dimension to obtain a fixed-length feature vector. Then, a fully connected classification layer (output dimension is 2, corresponding to "arc fault and normal operation") is connected to form the output layer. In the training stage, binary cross-entropy is used as the loss function to optimize the network parameters. After being deployed to the embedded edge terminal, the network serves as an initial arc detection model to perform forward inference on the frequency domain feature vector corresponding to the real-time current window of 1024 points, and outputs the arc judgment result and its confidence value (arc probability). In the subsequent edge adaptive process, only the fully connected classification layer is unlocked as the trainable part, and the parameters of the remaining convolutional layers and residual blocks remain unchanged, so as to realize the online update with small overhead and controllable stability.

[0097] Step 102, within a preset continuous time length when the photovoltaic inverter is first operated, 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, the corresponding working current data is stored as sample data to form a sample data set collected within the preset continuous time length.

[0098] Among them, the preset continuous time length refers to the continuous collection period (for example, 7 consecutive days) during the first operation of the photovoltaic inverter, which is used to obtain the field sample data. The working current data refers to the real-time current window (for example, 1024 points in length) collected from the DC side of the inverter and constructed in a fixed length. The sample data refers to the corresponding working current data that is retained and stored for a long time when the arc judgment result output by the model is normal operation and the confidence value is greater than the preset threshold. The sample data set refers to the collection of all sample data accumulated within the preset continuous time length, which is used for subsequent division of training set and validation set.

[0099] In the embodiment of the present application, within a preset continuous time length, the inverter continuously acquires the direct current side working current data at a fixed sampling rate, and assembles the data into a real-time current window with a length of 1024 points according to the criterion consistent with step 101; every time a window is acquired, it is sent to the initial arc striking detection model for forward inference to obtain the arc judgment result and the confidence value of the window. When the output arc judgment result of the window is normal operation and the confidence value is greater than a preset threshold, the system determines the working current data corresponding to the window as sample data and writes it into the local sample library, while recording the time stamp, device identifier, direct current working point information (such as the current value or current sub-range), operating environment information (such as temperature, irradiance) and the confidence value output by the model associated with the window, for tracing and subsequent data management; if any condition is not met, it is not included in the sample library, only the minimum operation log is retained. The above process is executed in a window cycle until the end of the preset continuous time length; at the end, the system automatically summarizes all the sample data accumulated during this period to form a sample data set collected within the preset continuous time length, and transfers it to the subsequent step for dividing the training set and the validation set according to the preset proportion and carrying out the edge side fine tuning. This setting can ensure that the acquired sample data comes from the real scene and has high credibility, providing a stable and consistent data basis for subsequent model updating.

[0100] In a possible implementation, the step 102 can specifically include:

[0101] Within a preset continuous time length when the photovoltaic inverter is first operated, under the maximum power point tracking control, the direct current side working point of the inverter is operated along the maximum power point tracking trajectory;

[0102] The working current data on the maximum power point tracking trajectory is taken as the sub-range reference, the current range is set to the first preset current to the maximum working current reached on the day, and the sub-range set is generated according to the second preset current as the step;

[0103] Traverse the sub-range set, and when the arc judgment result output by the initial arc striking detection model is normal operation and the confidence value is greater than a preset threshold, store the corresponding working current data as sample data in each sub-range;

[0104] According to the sample data within the preset continuous time length, determine the sample data set.

[0105] The maximum power point tracking trajectory refers to the operating trajectory of the direct current side working point of the inverter under MPPT control as the working condition changes, which is used to determine the reference interval of the current sub-range.

[0106] The first preset current refers to the lower limit of the sub-range interval; the second preset current refers to the step current of the sub-range. In the embodiment, the first preset current can be 3A and the second preset current can be 1A.

[0107] wherein the confidence value is output by the initial arc detection model, and is calculated by the Softmax of the linear output of the last full connection layer.

[0108] Exemplarily, within the preset continuous time length, the inverter is operated under the maximum power point tracking (MPPT) control; the system continuously obtains the DC side working current data at a sampling rate of 250 kHz, and assembles a real-time current window with a time length of 500 ms. Each time a window is obtained, it is sent to the initial arc detection model deployed to the arc detection module for inference to obtain an arc judgment result and a confidence value p normal When and only when the output arc judgment result is normal operation, and p normal >0.9 (preset threshold), it is determined that the window is high-confidence normal operation data; otherwise, it is discarded. The confidence value can be calculated using the following formula:

[0109]

[0110] wherein, e Znormal and e Zarc respectively represent the linear output values of the last full connection layer for normal operation and arc failure.

[0111] In terms of current coverage, the working current on the MPPT trajectory is taken as the reference for grading, and the grading interval is set to 3A to the maximum working current of the day I mppt-max , and each grading is traversed with a step of 1A to ensure that each current level has samples entering the set.

[0112] For the windows that meet the conditions, the system automatically labels them as normal operation, counts them and writes them into the sample data set (sample count N++); the system continuously runs, and accumulates high-confidence samples for 7 consecutive days to form a sample data set covering multiple working conditions and multiple current levels.

[0113] The above implementation can also record the timestamp, device identifier and grading information of each sample, so as to subsequently preprocess and divide the training set or validation set based on the sample data set (see the subsequent steps), while ensuring consistency and traceability with the real working conditions.

[0114] On the basis of the above embodiment, according to the sample data within the preset continuous time length, the sample data set is determined, comprising:

[0115] preprocessing each sample data within the preset continuous time length to determine the frequency domain feature vector corresponding to each sample data;

[0116] Based on the frequency domain feature vector corresponding to each piece of sample data, the sample data set is determined.

[0117] In the embodiment of the present application, after the end of the preset continuous duration, all sample data in the period (i.e. the working current window satisfying that the output arc judgment result is normal operation and the confidence value is greater than the preset threshold) are read, each piece of sample data is preprocessed and feature generated according to the same caliber as offline training, the frequency domain feature vector corresponding to each piece of sample data is obtained, and then all the frequency domain feature vectors obtained within the preset continuous duration are aggregated by sample as the data basis for subsequent fine-tuning, i.e. the sample data set.

[0118] In a possible implementation, the sample data set is divided into a training set and a validation set.

[0119] Specifically, the frequency domain feature vectors in the sample data set are divided according to a preset proportion to determine the training set and the validation set.

[0120] For example, assuming there are N×100 groups of frequency domain feature vectors, they can be divided according to a preset proportion of 8:2, where N×80 groups are the training set and N×20 groups are the validation set. It should be noted that the test set can be fixed as the original test set.

[0121] Step 103, using the sample data set to perform online training on the initial arc detection model to obtain a candidate arc detection model.

[0122] In the embodiment of the present application, the training set obtained in step 102 (composed of the frequency domain feature vector corresponding to each piece of sample data and its label) is taken as input, the initial arc detection model of the same architecture is loaded, and online training is performed.

[0123] Specifically, the convolution layer, residual block and global average pooling layer parameters in the model can be set to a frozen state, and only the fully connected classification layer is set to a trainable parameter; a calculation method of training loss information is established, and binary cross entropy is preferably used as the loss function (for arc fault and normal operation two-classification), and different weights can be given to positive and negative samples when the classes are imbalanced to stabilize the training; a first-order optimization algorithm based on gradient is selected to update the parameters of the fully connected classification layer; the frequency domain feature vectors of the training set are input into the model, forward inference is performed batch by batch to obtain the arc judgment result and the confidence value, the training loss information is calculated and back propagation and parameter updating are performed, only the gradient of the fully connected classification layer parameters is updated, and the parameters of the remaining layers remain unchanged; when the preset termination condition (such as reaching the maximum number of rounds or the training loss no longer decreasing) is met, the best fully connected classification layer parameters in 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 determination on the validation set.

[0124] In a possible implementation, each sample data in the sample data set has a true label; and the step 103 can specifically include:

[0125] inputting the sample data in the training set into the initial arc detection model to obtain a detection arc judgment result corresponding to each sample data in the training set; determining first training loss information based on the detection arc judgment result corresponding to each sample data in the training set and the true label; and updating the model parameter of the initial arc detection model based on the first training loss information to obtain a candidate arc detection model;

[0126] On this basis, the candidate arc detection model can also be detected, that is, whether the candidate arc detection model meets the preset requirement and can be put into use, and the detection can specifically include:

[0127] inputting the sample data in the verification set into the candidate arc detection model to obtain second training loss information;

[0128] 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, it is confirmed that the candidate arc detection model meets the preset requirement.

[0129] In another possible implementation, the initial arc detection model includes a fully connected classification layer, and the step 103 can specifically include:

[0130] updating the parameter of the fully connected classification layer in the initial arc detection model based on the first training loss information to obtain an updated arc detection model;

[0131] inputting each sample data in the verification set into the updated arc detection model to determine third training loss information;

[0132] In a case where 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.

[0133] 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, and in this embodiment, a binary cross-entropy loss function can be taken as a function for calculating each training loss information.

[0134] The updated arc detection model refers to a model obtained after only updating the parameter of the fully connected classification layer (the convolution layer, the residual block, and the global average pooling layer are kept frozen).

[0135] 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.

[0136] Training loss information (illustrated formula):

[0137]

[0138] 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.

[0139] 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:

[0140]

[0141] 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.

[0142] 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.

[0143] 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.

[0144] In the embodiment of the present application, the verification set is input into the candidate arc detection model in batches for forward inference to obtain the arc judgment result and the confidence value of each sample. According to the verification set real label and the model output, first, the second training loss information of the candidate arc detection model on the verification set is calculated, and compared with the first training loss information on the same verification set before fine-tuning. At the same time, the performance index of the verification set result is evaluated according to the pre-set performance index of the system to obtain the corresponding performance index value. The above two items are used as joint acceptance conditions: when and only when the verification set second training loss information is not greater than the first training loss information before fine-tuning and the verification set performance index is not lower than the pre-set index requirement and does not decrease compared with before fine-tuning, the candidate arc detection model is confirmed as the target arc detection model and is written into operation; if any condition is not met, the current round of update is invalid, and the previous running model remains unchanged. Through the double-condition acceptance process, the performance fluctuation caused by invalid update can be inhibited to ensure the stability and reliability of the written model under field working conditions.

[0145] Specifically, after the verification set is input into the candidate arc detection model, the method further comprises:

[0146] When the candidate arc detection model does not meet the pre-set requirement, the model parameters in the candidate arc detection model are rolled back to the model parameters in the initial arc detection model;

[0147] The initial arc detection model is determined as the target arc detection model.

[0148] In a possible implementation, the performance index of the candidate arc detection model is obtained based on the arc judgment result of the verification set and the real label, and the arc judgment result of the verification set is output by the candidate arc detection model; the method further comprises:

[0149] According to the first number and the second number, the precision of the candidate arc detection model is determined;

[0150] According to the first number and the third number, the recall of the candidate arc detection model is determined;

[0151] Based on the precision and the recall, the performance index of the candidate arc detection model is determined.

[0152] The first number (TP) refers to the number of cases in which the arc judgment result of the verification set is an arc fault and the real label of the verification set is an arc fault, the second number (FP) refers to the number of cases in which the arc judgment result of the verification set is an arc fault and the real label of the verification set is normal operation, and the third number (FN) refers to the number of cases in which the arc judgment result of the verification set is normal operation and the real label of the verification set is an arc fault.

[0153] In the embodiment of the present application, the verification set is input into the candidate arc detection model in batches, and the arc judgment result of each sample is obtained. After aligning with the true label, TP, FP and FN are counted. Then the precision Precision , recall Recall and performance index F1 are calculated:

[0154]

[0155]

[0156]

[0157] After calculating F1, whether to use the candidate arc detection model as the target arc detection model can be determined according to whether △F1 is greater than or equal to zero, wherein △F1 = F1 new -F1 old , F1 old is the performance index of the initial arc detection model, and F1 new is the performance index of the candidate arc detection model.

[0158] It should be noted that, in order to verify the effectiveness of the method, the same data set is used to train the ordinary CNN model and the ResNet-1D CNN model proposed in the present application, and a test set is constructed by collecting the running current signal in an actual field. On the basis of the field test, the edge adaptive fine-tuning method proposed in the present application is further applied to update the parameters of the original ResNet-1D CNN model to obtain the ResNet-1D CNN model after adaptive fine-tuning. The inference results of the above three types of models on the same test set are counted to obtain the corresponding confusion matrix.

[0159] According to the confusion matrix, the precision Precision =1.000, the recall Recall =0.909 and the F1 index =0.952 of the ordinary CNN model, the precision Precision =1.000, the recall Recall =0.985 and the F1 index =0.993 of the ResNet-1D CNN model, and the precision Precision , recall Recall and F1 index of the model after adaptive fine-tuning all reach 1.000.

[0160] By comparing the performances of the three models, i.e., the ordinary CNN, the ResNet-1D CNN and the fine-tuned ResNet-1D CNN, on the same test set, it can be seen that the ordinary CNN model is insufficient in recall rate, which is only 0.909, and there are many missed reports. The ResNet-1D CNN model proposed in the method introduces an improved residual structure, effectively alleviates the feature degradation problem, and improves the recall rate to 0.985, with the F1 index close to full score, and greatly reduces the missed report situation. After further applying the edge adaptive fine-tuning mechanism proposed in the method, all indicators of the model reach 1.000, and the arc fault and normal operation samples are completely and correctly classified.

[0161] Therefore, it is verified that the ResNet-1D CNN model and the edge adaptive fine-tuning mechanism of the method can further improve the classification accuracy and stability of the model, so that the arc detection can maintain high precision and strong robustness under different working conditions.

[0162] In step 105, the arc detection of the photovoltaic inverter during operation is performed based on the target arc detection model.

[0163] In the embodiment of the application, the photovoltaic inverter continuously collects the direct current side working current data during normal operation at a consistent sampling rate with offline training, and assembles the real-time current window (for example, 1024 points, step length 256 points) with 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 the confidence value (arc probability) of the window.

[0164] The photovoltaic inverter can compare the arc probability with the detection threshold: when the arc probability is not less than the detection threshold, the window is recorded as "arc fault", otherwise it is recorded as "normal operation".

[0165] In order to improve the stability of online judgment, the system can use a debouncing or persistent criterion for secondary confirmation, for example, applying a "3 / 5" strategy (at least 3 of the last 5 real-time current windows are arc judgments) in the sliding window sequence, and meeting the minimum duration (such as ≥100 ms) to trigger the arc fault event; immediately output the alarm information and link the protection logic to perform the disconnect control after triggering, record the event timestamp, current current level, model output confidence value and related operating parameters. If the trigger condition is not met, the normal operation state is maintained and the subsequent real-time current window is processed.

[0166] To facilitate post-tracing and operation and maintenance analysis, the system can save the original data and model output of a plurality of windows before and after the arc event according to a set buffer depth (for example, 10 windows before and after the event), and write the target arc detection model version number and threshold value used for this online detection into the event record. The above process is executed in a loop to realize online arc detection and protection linkage based on the target arc detection model, and provide real-time and reliable arc fault identification capability without changing the main control strategy of the inverter.

[0167] In the above method embodiment, first, the working current data is input into the initial arc detection model deployed after offline training on the inverter side, and the arc judgment result and the corresponding confidence are output in real time, providing an objective basis for subsequent sample screening based on confidence; then, within a preset continuous time length of the first operation, only when the model determines normal operation and the confidence is greater than a preset threshold, the corresponding current data is stored as sample data, forming a sample data set that fits the real working condition, and the initial arc detection model is trained online at the device end based on the sample data set, to obtain a candidate arc detection model, and in the case that the candidate arc detection model meets the preset requirements, the candidate arc detection model is determined as the target arc detection model, so that the model parameters adapt to the field power grid fluctuations, load disturbances and noise levels in real time during the online training process, significantly reducing the distribution difference between the laboratory and the power station working condition; finally, the target arc detection model is used for online detection, and under the premise of not relying on the cloud, the model gradually adapts to the field data, enhances the robustness and consistency to the power grid fluctuations, load disturbances and noise changes, thereby improving the environmental adaptability in the actual photovoltaic power station environment and reducing false positives and false negatives.

[0168] Referring to Figure 4 , a general flowchart of an arc detection method for an inverter is shown.

[0169] First, data collection is performed in a laboratory environment: an arc current data and a normal current data are obtained by using an arc machine; the collected current data is subjected to exponential weighted moving average filtering to suppress transient noise and power frequency disturbances; the filtered sequence is subjected to fast Fourier transform to obtain frequency spectrum data as a frequency domain feature vector for subsequent network input. After labeling, the original current data set is divided into an initial training set, an initial validation set and an initial test set in a ratio of 4:3:3; a one-dimensional convolutional neural network (ResNet-1DCNN) based on an improved residual structure is trained with the initial training set as input, the hyperparameters are determined by the initial validation set, and the performance is confirmed on the initial test set to obtain and solidify the initial arc detection model. Subsequently, the initial arc detection model is deployed to the arc detection module of the photovoltaic inverter for preliminary reasoning.

[0170] In the edge adaptive fine-tuning stage, the inverter runs on site for a long time: the system automatically collects high-confidence normal operation data covering different current levels for continuous days to form a sample data set; the set is divided into a training set and a validation set in a ratio of 8:2, and the test set is fixed as the original test set. During fine-tuning, the convolutional layer and the residual block are frozen, and only the fully connected classification layer is unlocked as a trainable parameter; the training set is iteratively updated, and a model detection mechanism based on the validation set loss is set. After completing an update, first determine whether the loss does not increase or meets the standard on the validation set, and then recheck on the fixed original test set. If the performance indicators of the test set do not decrease compared with the baseline (meet the preset performance indicators), update the weights of the fully connected layer and write them, and obtain the target arc detection model; otherwise, roll back to maintain the original running model (i.e., the initial arc detection model).

[0171] Finally, the inverter enters long-term online detection: the real-time current window of a fixed length is inferred and the arc fault probability (i.e., the confidence value) is output; when the probability is not less than the detection threshold and meets the continuous criterion, an alarm is triggered / disconnected; the system periodically obtains on-site samples according to the above strategy and triggers fine-tuning and acceptance, thereby realizing adaptive updating and stable arc detection on site without relying on the cloud.

[0172] Referring to Figure 5 , a specific flowchart of the edge adaptive fine-tuning mechanism in the overall flowchart is shown.

[0173] In the embodiments of the present application, when the inverter is first run on site, the ResNet-1D CNN initial arc detection model deployed by offline training is used to infer the real-time current window, and the normal operation probability p normal is calculated. p normal When P > 0.9, the current segment of the current 500 ms length is written into the cache in units of 1A and is automatically labeled as "normal operation", and is accumulated as a sample data set. When the running time reaches the preset number of days, the sample data is sequentially subjected to exponential weighted moving average filtering and fast Fourier transform to obtain N x 100 groups of current spectral feature vectors; then the sample is proportionally allocated as a training set (N x 80 groups) and a validation set (N x 20 groups), and the test set is fixed as the original test set. During fine-tuning, the convolutional layer and the residual block are frozen, and only the fully connected layer is opened as a trainable parameter; if the validation set loss change ΔL ≥ 0, the test set is rechecked, the performance increment ΔF1 is calculated, when ΔF1 ≥ 0, the fully connected layer parameters are updated, the target arc detection model is obtained and replaced online; otherwise, roll back to the initial arc detection model.

[0174] Referring to Figure 6Fig. 1 is a structural schematic diagram of an arc detection device of an inverter provided by an embodiment of the present application, which shows only parts related to the embodiments of the present application for the convenience of illustration.

[0175] The arc detection device 600 of the inverter comprises:

[0176] A data input module 601 is configured to input working current data of the photovoltaic inverter to an initial arc detection model, the initial arc detection model is deployed to the photovoltaic inverter after offline training, and is configured to perform inference on the working current data and output an arc judgment result and a corresponding confidence value; the arc judgment result comprises normal operation and arc fault;

[0177] A sample collection module 602 is configured to, within a preset continuous time length when the photovoltaic inverter is first operated, store corresponding working 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, to form a sample data set collected within the preset continuous time length;

[0178] An online training module 603 is configured to perform online training on the initial arc detection model by using the sample data set, to obtain a candidate arc detection model;

[0179] A performance verification module 604 is configured to, in a case where the candidate arc detection model meets a preset requirement, determine the candidate arc detection model as a target arc detection model;

[0180] An arc detection module 605 is configured to perform arc detection on the photovoltaic inverter in operation based on the target arc detection model.

[0181] In the embodiments of the present application, the sample data set comprises a training set and a verification set, and each sample data in the sample data set has a true label; the online training module 603 comprises:

[0182] An initial detection unit is configured to input the sample data in the training set to the initial arc detection model, to obtain an arc judgment result corresponding to each sample data in the training set;

[0183] A first loss determination unit is configured to determine first training loss information based on the arc judgment result corresponding to each sample data in the training set and the true label;

[0184] A candidate determination unit is configured to update model parameters of the initial arc detection model based on the first training loss information, to obtain the candidate arc detection model;

[0185] Correspondingly, the arc detection device 600 of the inverter further comprises:

[0186] a second loss determination unit, configured to input sample data in the verification set into the candidate arc detection model to obtain second training loss information;

[0187] a model detection unit, configured to confirm that the candidate arc detection model meets the preset requirement when the second training loss information meets the preset loss requirement and a performance index of the candidate arc detection model meets a preset index requirement.

[0188] In the embodiment of the present application, the initial arc detection model comprises a fully connected classification layer; the candidate determination unit is configured to:

[0189] update parameters of the fully connected classification layer in the initial arc detection model based on the first training loss information to obtain an updated arc detection model;

[0190] input each sample data in the verification set into the updated arc detection model to determine third training loss information;

[0191] in a case where the third training loss information is less than the first training loss information, determine the updated arc detection model as the candidate arc detection model.

[0192] In the embodiment of the present application, the arc detection device 600 of the inverter further comprises:

[0193] a parameter rollback judgment module, configured to rollback model parameters in the candidate arc detection model to model parameters in the initial arc detection model when the candidate arc detection model does not meet the preset requirement;

[0194] a first target determination module, configured to determine the initial arc detection model as a target arc detection model.

[0195] In the embodiment of the present application, the performance index of the candidate arc detection model is obtained based on an arc judgment result corresponding to the verification set and a real label, and the arc judgment result of the verification set is output by the candidate arc detection model; the arc detection device 600 of the inverter further comprises:

[0196] a precision determination module, configured to determine a precision of the candidate arc detection model according to the first number of cases and the second number of cases; the first number of cases refers to a number of cases that the arc judgment result of the verification set is an arc fault and the real label of the verification set is an arc fault, and the second number of cases refers to a number of cases that the arc judgment result of the verification set is an arc fault and the real label of the verification set is normal operation;

[0197] a recall rate determination module, configured to determine a recall rate of the candidate arc detection model according to the first number of cases and a third number of cases; the third number of cases refers to a number of cases that the arc judgment result of the verification set is normal operation and the real label of the verification set is an arc fault;

[0198] The performance index determination module is configured to determine the performance index of the candidate arc drawing detection model based on the precision and the recall.

[0199] In the embodiment of the present application, the sample collection module 602 can specifically include:

[0200] The tracking control unit is configured to, within a preset continuous time length in which the photovoltaic inverter is first operated, operate the DC side operating point of the inverter along the maximum power point tracking trajectory under the maximum power point tracking control;

[0201] The grading unit is configured to take the working current data on the maximum power point tracking trajectory as a grading reference, set the current range to a first preset current to a maximum working current reached on the day, and generate a grading set according to a second preset current as a step;

[0202] The storage unit is configured to traverse the grading set, and when the arc judgment result output by the initial arc drawing detection model is normal and the confidence value is greater than a preset threshold value in each grading, store the corresponding working current data as sample data;

[0203] The sample determination unit is configured to determine a sample data set according to the sample data within the preset continuous time length.

[0204] In the embodiment of the present application, the offline training step of the initial arc drawing detection model includes:

[0205] Collect and label normal current data and arc current data;

[0206] Preprocess the normal current data and the arc current data, determine the frequency domain feature vector corresponding to the normal current data and the frequency domain feature vector corresponding to the arc current data, and obtain an initial sample data set;

[0207] Divide the initial sample data set into an initial training set, an initial validation set, and an initial test set;

[0208] Train the arc drawing detection network with the initial training set as input, determine the network parameters of the arc drawing detection network based on the initial validation set, and determine the performance index of the arc drawing detection network under the network parameters based on the initial test set;

[0209] When the performance index of the arc drawing detection network meets the preset index requirement, the arc drawing detection network is determined as the initial arc drawing detection model.

[0210] The arc drawing detection device 600 of the inverter provided in the embodiment of the present application can be applied in the arc drawing detection method of the inverter provided in the foregoing embodiments. For details, refer to the description of the arc drawing detection method of the inverter provided in the foregoing embodiments, which will not be described here.

[0211] Referring to Figure 7, shows a structural schematic diagram of an electronic device provided in an embodiment of the present application. As shown in Figure 7 the electronic device 700 of the embodiment includes at least one processor 710 (only one is shown in the figure), Figure 7 a memory 720, and a computer program 721 stored in the memory 720 and executable on the at least one processor 710, and the processor 710 implements the steps in the arc detection method of the inverter in the above embodiment when executing the computer program 721.

[0212] The electronic device 700 can be a server, a physical server, a computing device, etc. The electronic device can include, but is not limited to, the processor 710 and the memory 720. Those skilled in the art can understand that Figure 7 The electronic device 700 is only an example and does not constitute a limitation on the electronic device 700, and can include more or fewer components than shown, or combine certain components, or include different components, for example, can also include an input / output device, a network access device, etc.

[0213] The processor 710 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, etc. The general-purpose processor can be a microprocessor or can also be any conventional processor.

[0214] The memory 720 can be an internal storage unit of the electronic device 700 in some embodiments, for example, a hard disk or a memory of the electronic device 700. The memory 720 can also be an external storage device of the electronic device 700 in other embodiments, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 700. Further, the memory 720 can include both the internal storage unit and the external storage device of the electronic device 700. The memory 720 is used to store an operating system, an application program, a boot loader, data, and other programs, for example, program codes of the computer program, etc. The memory 720 can also be used to temporarily store data that has been output or will be output.

[0215] In particular implementations, the processor 710, the memory 720, and the computer program 721 described in the embodiments of the present application can perform the embodiments of the arc detection method of the inverter of the present application, which will not be described here.

[0216] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiments can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be realized in the form of hardware or software. In addition, the specific names of each functional unit and module are only for easy distinction, and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0217] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in detail in a certain embodiment can be referred to the related description of other embodiments.

[0218] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0219] In the embodiments provided in the present application, it should be understood that the disclosed devices / equipment and methods can be implemented in other ways. For example, the device / equipment embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed each other can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0220] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, i.e. may be located in one place, or may be distributed to multiple network units. Part or all of the units may be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0221] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0222] If the integrated module / unit is realized in the form of 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 above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the processor executes the computer program, the steps of each method embodiment described above can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable file or some intermediate form. The computer readable medium can include any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier wave signal, telecommunication signal and software distribution medium, etc. It should be noted that the content included in the computer readable medium can be appropriately increased or decreased 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 electric carrier wave signal and telecommunication signal.

[0223] The above-mentioned embodiment methods can also be completed by a computer program product, when the computer program product runs on the electronic device, so that the electronic device executes the steps in each method embodiment described above.

[0224] The above examples are only used to illustrate the technical solutions of the present application, but not to limit it. Although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalent ones. The modification or replacement does not make the essence of the corresponding technical solution deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method of arc detection for an inverter, the method comprising: The method comprises: inputting working current data of a photovoltaic inverter into an initial arc detection model, the initial arc detection model being deployed to the photovoltaic inverter after offline training, for reasoning on the working current data and outputting an arc judgment result and a corresponding confidence value; the arc judgment result comprising normal operation and arc fault; within a preset continuous time length of 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 a preset threshold, corresponding working current data is stored as sample data to form a sample data set collected within the preset continuous time length; training the initial arc detection model using the sample data set to obtain a candidate arc detection model; in a case where the candidate arc detection model meets a preset requirement, determining the candidate arc detection model as a target arc detection model; based on the target arc detection model, performing arc detection on the photovoltaic inverter in operation; wherein the sample data set comprises a training set and a validation set, each sample data in the sample data set has a true label; training the initial arc detection model using the sample data set to obtain a candidate arc detection model comprises: inputting sample data in the training set into the initial arc detection model to obtain an arc judgment result corresponding to each sample data in the training set; based on the arc judgment result corresponding to each sample data in the training set and the true label, determining first training loss information; updating model parameters of the initial arc detection model based on the first training loss information to obtain a candidate arc detection model; after the initial arc detection model is trained using the sample data set to obtain a candidate arc detection model, the method further comprises: inputting sample data in the validation set into the candidate arc detection model to obtain second training loss information; when the second training loss information meets a preset loss requirement and a performance index of the candidate arc detection model meets a preset index requirement, confirming that the candidate arc detection model meets the preset requirement.

2. The method of claim 1, wherein, The initial arc detection model comprises a fully connected classification layer; updating model parameters of the initial arc detection model based on the first training loss information to obtain a candidate arc detection model comprises: updating parameters of the fully connected classification layer in the initial arc detection model based on the first training loss information to obtain an updated arc detection model; inputting each sample data in the validation set into the updated arc detection model to determine third training loss information; in a case where the third training loss information is less than the first training loss information, determining the updated arc detection model as the candidate arc detection model.

3. The method of claim 1, wherein, The method further comprises: when the candidate arc detection model does not meet the preset requirement, rolling back model parameters in the candidate arc detection model to model parameters in the initial arc detection model; determining the initial arc detection model as the target arc detection model.

4. The method of claim 1, wherein, The performance index of the candidate arc drawing detection model is obtained based on the arc judgment result corresponding to the verification set and the true label, and the arc judgment result of the verification set is output by the candidate arc drawing detection model; the method further comprises: According to the first and second numbers, the precision of the candidate arc drawing detection model is determined; the first number refers to the number of cases in which the arc judgment result of the verification set is arc fault and the true label of the verification set is arc fault, and the second number refers to the number of cases in which the arc judgment result of the verification set is arc fault and the true label of the verification set is normal operation; According to the first and third numbers, the recall rate of the candidate arc drawing detection model is determined; the third number refers to the number of cases in which the arc judgment result of the verification set is normal operation and the true label of the verification set is arc fault; Based on the precision and the recall rate, the performance index of the candidate arc drawing detection model is determined.

5. The method of claim 1, wherein, Within the preset continuous time length when the photovoltaic inverter is first operated, when the arc judgment result output by the initial arc drawing detection model is normal operation and the confidence value is greater than a preset threshold, the corresponding working current data is stored as sample data to form a sample data set collected within the preset continuous time length, comprising: Within the preset continuous time length when the photovoltaic inverter is first operated, under the action of maximum power point tracking control, the direct current side operating point of the inverter is operated along the maximum power point tracking trajectory; The working current data on the maximum power point tracking trajectory is used as a grading reference, the current range is set to a first preset current to the maximum working current reached on the day, and a grading set is generated according to a second preset current as a step; Traverse the grading set, and when the arc judgment result output by the initial arc drawing detection model is normal operation and the confidence value is greater than a preset threshold, store the corresponding working current data as sample data in each grading; According to the sample data within the preset continuous time length, the sample data set is determined.

6. The method of claim 1, wherein, The offline training step of the initial arc drawing detection model comprises: Collect and label normal current data and arc current data; Preprocess the normal current data and the arc current data, determine the frequency domain feature vector corresponding to the normal current data and the frequency domain feature vector corresponding to the arc current data, to obtain an initial sample data set; Divide the initial sample data set into an initial training set, an initial verification set and an initial test set; Train the arc drawing detection network with the initial training set as input, determine the network parameters of the arc drawing detection network based on the initial verification set, and determine the performance index of the arc drawing detection network under the network parameters based on the initial test set; When the performance index of the arc drawing detection network meets the preset index requirement, the arc drawing detection network is determined as the initial arc drawing detection model.

7. An arc detection device of an inverter, characterized by comprising: The device comprises: The data input module is configured to input working current data of a photovoltaic inverter to an initial arc detection model, the initial arc detection model is deployed to the photovoltaic inverter after offline training, and is configured to perform inference on the working current data and output an arc judgment result and a corresponding confidence value; the arc judgment result includes normal operation and arc failure; The sample collection module is configured to, within a preset continuous time length when the photovoltaic inverter is first operated, store corresponding working 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, to form a sample data set collected within the preset continuous time length; The online training module is configured to perform online training on the initial arc detection model by using the sample data set, to obtain a candidate arc detection model; The performance verification module is configured to determine the candidate arc detection model as a target arc detection model when the candidate arc detection model meets a preset requirement. The arc detection module is configured to perform arc detection on the photovoltaic inverter in operation based on the target arc detection model. The sample data set includes a training set and a verification set, and each sample data in the sample data set has a true label; the online training module includes: The initial detection unit is configured to input sample data in the training set to the initial arc detection model, to obtain an arc judgment result corresponding to each sample data in the training set; The first loss determination unit is configured to determine first training loss information based on the arc judgment result corresponding to each sample data in the training set and the true label; The candidate determination unit is configured to update model parameters of the initial arc detection model based on the first training loss information, to obtain a candidate arc detection model. Correspondingly, the arc detection device of the inverter further includes: The second loss determination unit is configured to input sample data in the verification set to the candidate arc detection model, to obtain second training loss information; The model detection unit is configured to confirm that the candidate arc detection model meets a preset requirement when the second training loss information meets a preset loss requirement and a performance index of the candidate arc detection model meets a preset index requirement.

8. The pre-arc detection device of the inverter according to claim 7, characterized by, The initial arc detection model includes a fully connected classification layer, and the candidate determination unit is configured to: update parameters of the fully connected classification layer in the initial arc detection model based on the first training loss information, to obtain an updated arc detection model; input each sample data in the verification set to the updated arc detection model, to determine third training loss information; determine the updated arc detection model as the candidate arc detection model when the third training loss information is less than the first training loss information.

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, The processor executes the computer program to implement the method of any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program is executed by the processor to implement the method of any one of claims 1 to 6.

Images