Low-voltage electricity utilization safety state monitoring and evaluating method and system

Abstract

The invention discloses a low-voltage power utilization safety state monitoring and evaluating method and system, and the method comprises the steps: S1, obtaining the operation data of a low-voltage customer, constructing a typical feature data set, and carrying out the preprocessing; s2, constructing an identification model, and training the identification model based on the typical feature data set to obtain a final identification model; the identification model comprises an automatic encoder model, a random forest integrated learning model and a BP neural network model; s3, performing monitoring and early warning on a power utilization fault based on the final identification model; the step of monitoring the power utilization faults comprises accurate fault arc identification based on an automatic encoder model, low-voltage user electric leakage judgment based on a random forest and short circuit sensing based on current sudden change characteristics; the early warning of the power utilization fault comprises low-voltage power utilization potential safety hazard early warning based on a BP algorithm and low-voltage power utilization potential safety hazard early warning based on a random forest ensemble learning model. The method has the advantages of high prediction precision and the like.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of low-voltage power grids, and in particular to a method and system for monitoring and evaluating the safety status of low-voltage electricity use. Background Art

[0002] At present, low-voltage power grids generally use residual current protection devices as the main protection measures to prevent human electric shock accidents and electrical insulation failures. However, according to theoretical research and actual operation results, the existing residual current protection device operates according to the total residual current, and its protection characteristics are affected by the phase difference angle between the natural leakage current and the fault leakage current, and there is a protection dead zone; in addition, there are malfunctions such as failure to close normally when extreme weather and large loads are suddenly put into use. Once the residual current protection device malfunctions or refuses to operate, it will cause accidents such as electrical fires or electric shock casualties of organisms, and the detection difficulty of series arc faults. Therefore, the series arc fault in real life scenarios brings great harm to people's lives and the safe operation of the power system. The fault current of the series arc fault is limited by the line load, which is generally close to or even less than the normal working current of the load, so it is difficult to be detected by the existing overcurrent protection; on the other hand, the current waveform of the series arc fault has a certain randomness, and it will also be affected by the load to produce complex changes, which further increases the difficulty of detecting and accurately judging the series fault arc. At present, many scholars have conducted extensive research and exploration in the detection of series fault arcs and have achieved certain results, but there are also shortcomings. The methods used to detect fault arcs and distinguish them from normal conditions are mostly in the experimental stage and still lack a comprehensive and complete solution. Summary of the invention

[0003] In view of the technical problems existing in the prior art, the present invention provides a method and system for monitoring and evaluating the safety status of low-voltage electricity use with high accuracy.

[0004] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0005] A method for monitoring and evaluating the safety status of low-voltage electricity use, comprising the steps of:

[0006] S1. Obtain the operating data of low-voltage customers in normal operation and various power failure scenarios, build a typical feature data set, and perform preprocessing;

[0007] S2, constructing a recognition model, and training the recognition model based on the preprocessed typical feature data set to obtain a final recognition model; the recognition model includes an autoencoder model, a random forest ensemble learning model and a BP neural network model;

[0008] S3. Monitor and warn of power failures based on the final identification model; the monitoring of power failures includes accurate identification of fault arcs based on the autoencoder model, low-voltage user leakage judgment based on random forests, and short-circuit perception based on current mutation characteristics; the warning of power failures includes low-voltage power safety hazard warning based on the BP algorithm and low-voltage power safety hazard warning based on the random forest ensemble learning model.

[0009] Preferably, in step S1, the corresponding operation data includes residual current data Id i And total current data Ia i .

[0010] Preferably, in step S1, preprocessing includes:

[0011] The original operation data is denoised and the noise in the primary and secondary lines is filtered out by wavelet transform;

[0012] Normalize the denoised running data.

[0013] Preferably, in step S3, accurate identification of fault arcs based on the autoencoder model includes: building an autoencoder model, eliminating the decoding stage, retaining the encoding stage model, and nesting the output of the encoding model with three fully connected layers and one softmax output layer to form a softmax classifier model for arc identification.

[0014] Preferably, when performing arc identification, the core features of the fault arc are extracted using Fourier transform and sorted in order of importance, in the following order: fundamental wave, fundamental wave ratio, third harmonic, third harmonic ratio, fifth harmonic, fifth harmonic ratio, integral, variance, kurtosis, zero crossing time, and Shannon entropy.

[0015] Preferably, in step S3, the low-voltage user leakage identification based on random forest includes the following steps:

[0016] Obtaining raw data and preprocessing it, including eliminating the baseline value of the conditioning circuit, denoising the raw residual current data, extracting the sequence through a sliding time window, and processing the data out of order;

[0017] Use the Fourier transform method to extract different features from the original data, including fundamental wave, third harmonic, fifth harmonic, fundamental wave ratio, third harmonic ratio, fifth harmonic ratio, integral, Shannon entropy, kurtosis, and variance; divide the data set after feature extraction into training set, validation set, and test set according to the proportion;

[0018] Use the random forest algorithm to develop modeling, randomly extract multiple training sample subsets from the training set, perform decision tree modeling on each subset, integrate the decision results of multiple trees, and finally determine the detection results through voting, and complete the establishment of the low-voltage leakage detection model;

[0019] The training set is used as the input of the low-voltage leakage detection model to train the low-voltage system leakage detection model. Finally, the test set is used to predict the accuracy of the low-voltage system leakage detection model.

[0020] Preferably, in step S3, short circuit sensing based on current mutation characteristics includes the following steps:

[0021] Extract different features from the short-circuit voltage and current data. The features are calculated from the beginning of the second power frequency cycle of the voltage and current data; including peak-to-peak values, current and voltage amplitudes, and kurtosis factors;

[0022] Construct the characteristic vectors of short-circuit current and short-circuit voltage and determine the short-circuit judgment threshold;

[0023] The peak-to-peak value, current and voltage amplitude, kurtosis factor and short circuit discrimination threshold are compared to detect whether a short circuit occurs.

[0024] Preferably, in step S3, the low-voltage electricity safety hazard warning based on the random forest ensemble learning model includes the following steps:

[0025] Data acquisition: normal and fault arc data with time series are collected through real experiments, and the data are pre-processed by denoising and normalization, and the experimental data are sampled through time windows;

[0026] Feature engineering: By analyzing the experimental data, we extract characteristic indicators with a high degree of correlation with the fault arc, and construct a multi-dimensional fault arc hidden danger prediction feature data set; the selected features are: three frequency domain features of the first harmonic, third harmonic, and fifth harmonic of the current, two time domain features such as zero-crossing index and peak state, and Shannon entropy;

[0027] Model hyperparameter setting: Select the grid search local optimization method to optimize the model hyperparameters;

[0028] Model data fitting: After obtaining the sampled feature data set with time series, divide the fault arc hazard prediction feature data set into a training set and a test set at a ratio of 4:1, and read the training set into the model so that the model can complete the training set data fitting under the optimal hyperparameter environment and obtain a learned fault arc hazard warning model;

[0029] Model performance evaluation: The prediction accuracy and training time of the trained hidden danger warning model are evaluated through the test set, and multi-fold cross-validation is used to complete the model generalization ability test and overfitting assessment.

[0030] The present invention also discloses a low-voltage electricity safety status monitoring and evaluation system, comprising:

[0031] The data acquisition module is used to obtain the operating data of low-voltage customers in normal operation and various power failure scenarios, build typical feature data sets, and perform preprocessing;

[0032] A model building module, used to build a recognition model and train the recognition model based on the preprocessed typical feature data set to obtain a final recognition model; the recognition model includes an autoencoder model, a random forest ensemble learning model and a BP neural network model;

[0033] The monitoring and early warning module is used to monitor and warn of power failures based on the final identification model; the monitoring of power failures includes accurate identification of fault arcs based on the autoencoder model, low-voltage user leakage judgment based on random forests, and short-circuit perception based on current mutation characteristics; the early warning of power failures includes low-voltage power safety hazard early warning based on the BP algorithm and low-voltage power safety hazard early warning based on the random forest ensemble learning model.

[0034] Preferably, in the data acquisition module, the corresponding operation data includes residual current data Id i And total current data Ia i .

[0035] Preferably, in the data acquisition module, preprocessing includes:

[0036] The original operation data is denoised and the noise in the primary and secondary lines is filtered out by wavelet transform;

[0037] Normalize the denoised running data.

[0038] Preferably, in the monitoring and early warning module, accurate identification of fault arcs based on the autoencoder model includes: building an autoencoder model, eliminating the decoding stage, retaining the encoding stage model, and nesting the output of the encoding model into three fully connected layers and one softmax output layer to form a softmax classifier model for arc identification.

[0039] Preferably, when performing arc identification, the core features of the fault arc are extracted using Fourier transform and sorted in order of importance, in the following order: fundamental wave, fundamental wave ratio, third harmonic, third harmonic ratio, fifth harmonic, fifth harmonic ratio, integral, variance, kurtosis, zero crossing time, and Shannon entropy.

[0040] Preferably, in the monitoring and early warning module, the low-voltage user leakage identification based on random forest includes the following steps:

[0041] Obtaining raw data and preprocessing it, including eliminating the baseline value of the conditioning circuit, denoising the raw residual current data, extracting the sequence through a sliding time window, and processing the data out of order;

[0042] Use the Fourier transform method to extract different features from the original data, including fundamental wave, third harmonic, fifth harmonic, fundamental wave ratio, third harmonic ratio, fifth harmonic ratio, integral, Shannon entropy, kurtosis, and variance; divide the data set after feature extraction into training set, validation set, and test set according to the proportion;

[0043] Use the random forest algorithm to develop modeling, randomly extract multiple training sample subsets from the training set, perform decision tree modeling on each subset, integrate the decision results of multiple trees, and finally determine the detection results through voting, and complete the establishment of the low-voltage leakage detection model;

[0044] The training set is used as the input of the low-voltage leakage detection model to train the low-voltage system leakage detection model. Finally, the test set is used to predict the accuracy of the low-voltage system leakage detection model.

[0045] Preferably, in the monitoring and early warning module, short circuit perception based on current mutation characteristics includes the following steps:

[0046] Extract different features from the short-circuit voltage and current data. The features are calculated from the beginning of the second power frequency cycle of the voltage and current data; including peak-to-peak values, current and voltage amplitudes, and kurtosis factors;

[0047] Construct the characteristic vectors of short-circuit current and short-circuit voltage and determine the short-circuit judgment threshold;

[0048] The peak-to-peak value, current and voltage amplitude, kurtosis factor and short circuit discrimination threshold are compared to detect whether a short circuit occurs.

[0049] Preferably, in the monitoring and early warning module, the low-voltage power safety hazard early warning based on the random forest ensemble learning model includes the following steps:

[0050] Data acquisition: normal and fault arc data with time series are collected through real experiments, and the data are pre-processed by denoising and normalization, and the experimental data are sampled through time windows;

[0051] Feature engineering: By analyzing the experimental data, we extract characteristic indicators with a high degree of correlation with the fault arc, and construct a multi-dimensional fault arc hidden danger prediction feature data set; the selected features are: three frequency domain features of the first harmonic, third harmonic, and fifth harmonic of the current, two time domain features such as zero-crossing index and peak state, and Shannon entropy;

[0052] Model hyperparameter setting: Select the grid search local optimization method to optimize the model hyperparameters;

[0053] Model data fitting: After obtaining the sampled feature data set with time series, divide the fault arc hazard prediction feature data set into a training set and a test set at a ratio of 4:1, and read the training set into the model so that the model can complete the training set data fitting under the optimal hyperparameter environment and obtain a learned fault arc hazard warning model;

[0054] Model performance evaluation: The prediction accuracy and training time of the trained hidden danger warning model are evaluated through the test set, and multi-fold cross-validation is used to complete the model generalization ability test and overfitting assessment.

[0055] The present invention also discloses a computer program product, comprising a computer program, wherein the computer program executes the steps of the method described above when executed by a processor.

[0056] The present invention further discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described above are executed.

[0057] Compared with the prior art, the advantages of the present invention are:

[0058] The present invention targets arc faults that are easy to occur in low-voltage electricity, are highly concealed, and are likely to cause electrical fires once they occur. Based on the early warning model of the BP neural network algorithm, early warning information is issued before the fault occurs and corresponding measures are taken to achieve accurate perception of arc faults in advance; by simulating the real low-voltage electricity scene, current data in normal and arc fault states are collected for waveform analysis, and model training, testing and evaluation are performed based on the low-voltage electricity safety hazard perception feature library. The method has strong generalization ability and can predict the occurrence of arc faults in advance, with a prediction accuracy of up to 80%. In addition, in order to further improve the low-voltage hidden danger warning effect, the random forest ensemble learning model is used for technical upgrading; the random forest ensemble learning model is trained, and the hyperparameters are optimized through the local optimization method, and the model parameters are adjusted with the goal of minimizing the node information entropy to complete the construction of the power safety hidden danger warning model; experimental verification shows that the proposed method can achieve accurate prediction of low-voltage power safety hazards. Under a variety of load operating conditions, the prediction accuracy of the method of the present invention is higher than that of the four mainstream classification prediction models such as GRU, RNN, DT, and BP, which provides a new method for improving the safety factor of residents' electricity use and reducing the occurrence of electrical fires.

[0059] In order to effectively ensure the safety of electricity use for low-voltage customers and accurately perceive common faults in low-voltage systems, such as fault arcs, leakage faults and short-circuit faults, the present invention first establishes a feature database of potential safety hazards in electricity use for low-voltage customers and performs data preprocessing operations on it; a low-voltage customer electricity safety perception model is established based on the feature database, which mainly includes: a fault arc precision identification model based on an automatic encoder, a low-voltage user leakage discrimination model based on a random forest, and a short-circuit perception model based on current mutation characteristics. Through in-depth testing and experiments on the model, it can be seen that the perception accuracy of the perception model can reach 80%. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a format diagram of the original operating data in the present invention.

[0061] Figure 2 These are denoising effect diagrams at different time scales in the present invention; (a) is a denoising effect diagram at a time scale of 0.012s; (b) is a denoising effect diagram at a time scale of 12s.

[0062] Figure 3 This is a structural diagram of the automatic encoder training stage in the present invention.

[0063] Figure 4 It is a schematic diagram of the BP neural network structure in the present invention.

[0064] Figure 5 This is a diagram of the automatic encoder training process in the present invention.

[0065] Figure 6It is a correlation diagram between the theoretical model in the present invention and the actual low-voltage power safety device.

[0066] Figure 7 This is the Softmax classifier training model diagram in the present invention.

[0067] Figure 8 This is a diagram of the process of establishing a low-voltage system leakage detection model based on the random forest algorithm of the present invention.

[0068] Fig. 9 It is a schematic diagram of the leakage detection principle in the present invention.

[0069] Fig.10 It is the original residual current denoising processing image; (a) is the image without denoising processing; (b) is the image after denoising processing.

[0070] Fig.11 It is the characteristic distribution diagram under normal operating state and leakage fault state; among them, (a) is the fundamental wave characteristic; (b) is the third harmonic characteristic; (c) is the fifth harmonic characteristic; (d) is the fundamental wave ratio characteristic; (e) is the third harmonic ratio; (f) is the fifth harmonic ratio; (g) is the integral characteristic; (h) is the kurtosis characteristic; (i) is the Shannon entropy; and (j) is the sum and variance.

[0071] Fig.12 This is a diagram for establishing a low-voltage leakage detection model based on a random forest algorithm in the present invention.

[0072] Fig.13 This is a diagram of the transient process of the short-circuit current in the present invention.

[0073] Fig.14 This is a diagram of the short-circuit voltage transient process in the present invention.

[0074] Fig.15 This is a flow chart of the short-circuit sensing method based on the current mutation characteristics in the present invention.

[0075] Fig.16 The figure is a flow chart of the low-voltage electricity safety hazard early warning method based on the BP algorithm in the present invention.

[0076] Fig.17 This is a flow chart of the BP neural network training in the present invention.

[0077] Fig.18 It is a schematic diagram of the training set accuracy curve in the present invention.

[0078] Fig.19 Schematic diagram of the training set loss value in the present invention.

[0079] Fig. 20 A flow chart is constructed for the electricity safety hazard early warning model in the present invention.

[0080] Fig.21 The following is a flow chart of the monitoring and evaluation method of the present invention in an embodiment. DETAILED DESCRIPTION

[0081] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0082] like Fig.21 As shown, the low-voltage power safety status monitoring and evaluation method of the embodiment of the present invention includes the following steps:

[0083] S1. Obtain the operating data of low-voltage customers in normal operation and various power failure scenarios, build a typical feature data set, and perform preprocessing;

[0084] Establishment of typical characteristic data set: Simulate various power failure scenarios and normal operation status of low-voltage customers, and use the recording device to collect residual current data Id from the simulation experiment i And total current data Ia i , the sampling frequency is set to 10kHz, the power frequency is 50Hz, so each power frequency cycle includes 200 sampling points, forming the original data set matrix OriginalData. Figure 1 As shown, the matrix structure is N rows and 5 columns, where the i-th row represents the data at the i-th sampling point, the 1st column represents the sampling point number, the 2nd column represents the residual current data, the 3rd column represents the total current data, the 4th column represents the voltage data between the live wire and the neutral wire, and the 5th column represents the label value (normal operation label: "0", short circuit label: "1", leakage label: "2", fault arc label: "3").

[0085] Furthermore, the original data is denoised (the original data is formed at a sampling rate of 10kHz) and the noise in the primary and secondary circuits is filtered out by wavelet transform to improve the signal-to-noise ratio. Figure 2 As shown in Figure 2, the actual effect of denoising is demonstrated through two time scales, where the red line represents the signal after denoising and the blue line represents the signal before denoising.

[0086] Furthermore, in order to eliminate the impact of different dimensions between different loads and different features and enable the model to converge faster, the data samples are normalized using a unified standard so that the current values ​​of each load are between [-1,1] after linear transformation. The specific formula is:

[0087]

[0088] In the formula, x and x * are the corresponding sample values ​​before and after normalization; x max is the maximum value of the sample data; xmin is the minimum value of the sample data.

[0089] Finally, in order to balance the model training time and training effect and reduce the dimension of the model input data, the model training feature extraction is performed on the sampled experimental data. On the one hand, common features are selected for the diverse data collected during normal and faulty operation under different load conditions; on the other hand, in order to make the model more effectively trained and learned, feature data with stronger correlation with low-voltage power failures are extracted to achieve state classification and hidden danger perception of the original data set. The signal analysis method combining numerical statistical features and frequency domain features is used to extract relevant eigenvalues. Specifically, the peak-to-peak value, amplitude, integral, variance, kurtosis, power information entropy, signal proportion in different frequency bands, waveform factor, peak index, pulse index, and fault arc "zero-crossing retention" probability index of the total current and residual current are extracted to form a feature vector, which is used as the subsequent model input for training.

[0090] S2. After the typical feature data set is established, a suitable recognition model is constructed, and the recognition model is trained based on the preprocessed typical feature data set to obtain a final recognition model; wherein the recognition model includes an autoencoder model, a random forest ensemble learning model and a BP neural network model;

[0091] During the training process, the autoencoder model continuously improves the value of the extracted features by encoding and decoding, which is the basis for achieving high-precision classification of subsequent classifiers. Figure 3 As shown. The original data is compressed through network models such as the fully connected layer, the single batch normalization layer, and the activation function, and the 100-dimensional sample is compressed into 50 dimensions. The 50-dimensional encoded data is then restored and output by the inverse process of encoding, and the difference between the data before and after encoding is compared, so as to continuously improve the scientific nature of data encoding.

[0092] The random forest ensemble learning model is a supervised machine learning algorithm that combines the bagging ensemble learning theory with the random subspace method. The random forest uses the decision tree as the base classifier. On the basis of the decision tree, the ensemble idea is introduced to increase the randomness of the algorithm, which greatly improves the scope of application of the algorithm. The ensemble idea increases the diversity of decision trees. The randomness of the training set extraction and the node candidate segmentation feature set increases the diversity of decision trees. The diversity and diversity greatly improve the generalization ability and diagnostic accuracy of the algorithm.

[0093] like Figure 4As shown in the figure, the BP neural network model can be divided into three levels: input layer, hidden layer and output layer. It is a multi-layer feedforward self-learning network model that uses nonlinear differentiable functions for weight training. The input layer is used to input the warning indicator system, and the article is the characteristic value of the current data. The hidden layer can be one or more layers, which are used for data processing. Increasing the number of hidden layers can optimize the accuracy, but it will also increase the difficulty of network learning and the training time. The output layer is used to output warning information. The output layer of the article is a node, which is used to output the probability of a fault arc in the next cycle.

[0094] The nodes at the same level are independent of each other, and the output of each level node is only affected by the output of the previous level. The network learning process can be divided into two processes: forward propagation and back propagation. The activation function is the Sigmoid function.

[0095] S3. Monitor power failures based on the final identification model. Specifically, it includes accurate identification of fault arcs based on the autoencoder model, a low-voltage user leakage identification method based on random forests, and a short-circuit perception method based on current mutation characteristics;

[0096] Specifically, the precise identification process of fault arc based on the automatic encoder is as follows:

[0097] like Figure 5 As shown in the figure, the activation function used in the training of the autoencoder is the root mean square error. After every 10 trainings, a sample is selected from the test set to monitor the stage training effect. The blue line represents the training curve, and the orange line represents the stage test curve. The model was trained 200 times in total, and the construction of the autoencoder was completed with a training error of 7.8078e-04 and a test error of 7.3725e-04. After the autoencoder model is built, the decoding stage is removed, the encoding stage model is retained, and the output of the encoding model is nested with three fully connected layers and one softmax output layer to form a softmax classifier model.

[0098] The arc characteristics and frequency selection are as follows:

[0099] 1. Multi-feature selection

[0100] The core features of the fault arc are extracted by Fourier transform, mainly including: fundamental wave, fundamental wave ratio, third harmonic, third harmonic ratio, fifth harmonic, fifth harmonic ratio, integral, variance, kurtosis, zero crossing time, and Shannon entropy. However, the operating environment faced by low-voltage power safety devices is relatively complex. Therefore, in addition to extracting targeted feature vector groups that can effectively characterize the characteristics of fault arc hazards, it is also necessary to sort the importance of multiple feature quantities to obtain the more important fault arc composition feature quantities that contribute more to the fault arc detection effect. Through the effective fusion of feature quantities, the low-voltage power safety device can adapt to fault arc detection under different working conditions (different fault locations, considering line impedance and EMI filter parameter interference). The algorithm model of feature extraction and feature importance sorting is burned into the low-voltage power safety device to realize the functions of feature extraction, feature importance sorting, and fault arc detection.

[0101] 2. Arc characteristic frequency selection

[0102] In order to enable low-voltage electrical equipment to overcome the interference of harmonics, carrier communications, EMI noise and other factors that affect fault arc identification during actual use, it is necessary to carefully select the arc characteristic frequency. The present invention mainly collects arc characteristics under low frequency (within 1kHz), medium frequency (10kHz~100kHz frequency band), and high frequency (frequency above MHz) conditions to analyze the waveform morphology mutation characteristics, high-frequency noise characteristics near the arc zero crossing, and high-frequency noise during the arcing process. The correlation between the theoretical model of the present invention and the actual low-voltage electrical safety device is as follows. Figure 6 As shown:

[0103] Softmax classifier training

[0104] The completion of the autoencoder training indicates that the sequence data has been effectively feature extracted. The next step is to train a classification model that can distinguish arcs by encoding the data sequence. The autoencoder model is nested with three fully connected layers and one softmax output layer to form a softmax classifier. The training process is as follows: Figure 7 As shown in the figure, the loss function uses the root mean square error, the training sample set size is 120345, and after 4533 iterations, the training ends with a loss function value of 0.00300902.

[0105] Specifically, the low-voltage user leakage identification method based on random forest includes the following steps:

[0106] like Figure 8 As shown in the figure, the establishment of the low voltage system leakage detection model based on the random forest algorithm includes five parts: leakage detection experiment (acquisition of original residual current data), preprocessing of original residual current data, feature extraction, model establishment and model evaluation:

[0107] 1) Original data acquisition

[0108] The original data samples of the experiment come from the low-voltage laboratory leakage physics experimental platform. The schematic diagram of the leakage detection principle is as follows: Fig. 9 As shown:

[0109] Fig. 9 In the experiment, a programmable power supply provides 220V AC power to the loop; the line-to-ground capacitance is simulated by the ground capacitance simulation module, and the interference factor of the line-to-ground leakage current on the leakage fault detection of the low-voltage user is added; the leakage action current of the residual current circuit breaker is measured by the residual current transformer, and the measured residual current is converted into a digital signal that can be used for data acquisition, control process, and execution calculation display readout through the conditioning circuit; the original current data of the low-voltage system in normal and leakage fault states are collected by NI data acquisition equipment; the model sensing distance is not less than 25 meters, so in this experiment, the distance between the sensor and the leakage fault occurrence point is set to 25 meters; branch 1 is the leakage fault occurrence branch, and the fault occurrence time is controlled by the control switch, and the fault resistance is set to 10000Ω; branch 2, branch 3 and branch 4 are all load branches, representing nonlinear power electronic load and resistive load types respectively; the control switches of branches 2, 3, and 4 play the role of switching load types, and can also operate the control switches when collecting normal and fault data of the line, so as to take into account the interference factors of the control switch action on the leakage fault detection.

[0110] The original data collection process mainly includes:

[0111] Control the programmable power supply to supply 220V AC power - run the data acquisition system - turn on the dimmer load to work normally - add interference (the control switch of the action resistor branch creates interference) - operate the fault leakage branch control switch (create leakage fault) - add interference (the control switch of the action resistor branch creates interference);

[0112] Control the programmable power supply to supply 220V AC power - run the data acquisition system - turn on the computer load to work normally - add interference (the control switch of the action resistance branch creates interference) - actuate the fault leakage branch control switch (create leakage fault) - add interference (the control switch of the action resistance branch creates interference).

[0113] 2) Data preprocessing, including eliminating the baseline value of the conditioning circuit, denoising the original residual current data, extracting the sequence using a sliding time window, and processing the data out of order;

[0114] Specifically, the process of eliminating the baseline value of the conditioning circuit is as follows: adding a DC component to the experimental circuit during the leakage test can prevent the diode from burning and ensure that the conditioning circuit does not reverse, but it also causes the measurement baseline value of the current transformer to increase. Therefore, if you want to ensure the accuracy of the total residual current detection, you must first cut off the load, sample the DC data in the line, and calibrate the current transformer. The sampling time of this experiment is about 5.1s, the number of sampled data is 51,000, and the average value of all the sampled data is taken; subtracting the average value of the sampled data from all current data can eliminate the baseline value of the conditioning circuit, and you can get accurate total residual current data.

[0115] Specifically, the process of denoising the original residual current data is as follows: the original residual current data collected in the leakage experiment contains noise, such as Fig.10 As shown in (a): The original residual current is denoised using the wavelet transform software in the MATLAB software. The residual current waveform after denoising is as follows: Fig.10 As shown in (b), it can be seen that the total residual current waveform is smoother after denoising, and the burrs are less than before denoising.

[0116] Specifically, the process of extracting the sequence of the sliding time window is as follows: In practical applications, the system has a strong correlation with the data near the working point when it is working, but the correlation with the data far from the working point is not strong. Therefore, this paper establishes a sliding time window model and keeps the data interval length at 200 sampling points. When a new data is added, the earliest data slides out from the upper interval accordingly. As the data interval is continuously updated, the model is also continuously updated by the data in the new interval.

[0117] Specifically, the process of data disorder processing is as follows: in order to improve the generalization ability and accuracy of the low-voltage system leakage detection model, the time window data needs to be shuffled; use Excel's Rand() function to assign random numbers to the data features at each moment and sort them, so as to disrupt the time window data set (time series).

[0118] 3) Feature extraction

[0119] In order to increase the efficiency of model detection and reduce the computer's running memory, the original data is processed by dimensionality reduction, that is, feature extraction, so that a small number of more representative feature data can be used to reflect the effect of the overall original data. The Fourier transform method is used to extract different features from the original time window data table, including fundamental wave (fc), third harmonic (3rdc), fifth harmonic (5rdc), fundamental wave ratio (fcr), third harmonic ratio (3rdcr), fifth harmonic ratio (5rdcr), integral (Integral), Shannon entropy (Shannon), kurtosis (kurtosis), variance (var); the data set after feature extraction is divided into training set, validation set and test set in a ratio of 7:1:2. The feature distribution of the above features under normal operating conditions and leakage fault conditions is shown as follows. Fig.11 shown.

[0120] (4) Model establishment

[0121] The random forest algorithm is used to carry out modeling. The modeling process is as follows Fig.12 As shown in the figure, the Bootstrap resampling method is used to randomly extract multiple training sample subsets from the training sample set, and decision tree modeling is performed on each subset. The decision results of multiple trees are combined, and the detection results are finally determined by voting, and the low-voltage system leakage detection model is established.

[0122] The training set is used as the input of the RF model to train the low-voltage system leakage detection model, and finally the test set is used to predict the accuracy of the low-voltage system leakage detection model. It has been verified that the accuracy of the low-voltage system leakage detection model can reach 99.98%.

[0123] When establishing a random forest model, RandomForestRegressor is used to build the model, n_estimators is the number of decision trees; random_state is the random number seed of the randomly selected samples when each tree uses the Bootstrap in the Bagging strategy for resampling (that is, random sampling out of the bag with replacement); fit is used to train the model, and predict is used to predict the model.

[0124] The short-circuit sensing method based on current mutation characteristics specifically includes the following steps:

[0125] 1) Short circuit transient analysis

[0126] Simulate various power short-circuit fault problem scenarios and normal operating conditions of low-voltage customers. Use a recording device to collect short-circuit voltage and current data in normal and short-circuit fault conditions in the simulation experiment. The sampling frequency is set to 10KHz and the power frequency is 50Hz, so each power frequency cycle includes 200 sampling points to form the original data set matrix OriginalData; intercept the short-circuit voltage and current data of one cycle near the short-circuit moment (nearly 200 sampling points), and use the drawing function of matlab software to draw the waveform of the short-circuit transient voltage and current data, such as Figure 13-Figure 14 As shown:

[0127] It can be seen from the transient diagrams of short-circuit current and short-circuit voltage that when a short circuit occurs, the amplitude and amplitude change rate of the short-circuit current gradually increase; the short-circuit voltage suddenly drops.

[0128] 2) Calculation of short-circuit characteristics

[0129] Different features are extracted from the short-circuit voltage and current data after data cleaning by the following method, and the features are calculated from the beginning of the second power frequency cycle of the voltage and current data;

[0130] Peak-to-peak value: The peak-to-peak value of the current sampling point i is the difference between the previous peak value and the previous reverse peak value; the peak-to-peak value of the short-circuit current and short-circuit voltage is recorded as Pva i , Pvv i ;

[0131] Current and voltage amplitude: The short-circuit current (voltage) amplitude at the current sampling point i. The amplitude of the short-circuit current at time i is recorded as Aa i , the amplitude of the short-circuit voltage at time i is recorded as Av i ;

[0132] Kurtosis factor: Kurtosis factor is defined as the normalized 4th-order central moment, which can characterize the smoothness of the waveform. The kurtosis index is very sensitive to the impact component in the signal. The greater the energy of the impact component, the greater the kurtosis value will be, and the more uneven the waveform will be. The calculation formula of the kurtosis index is as follows:

[0133]

[0134] The above three characteristics under normal state and short circuit state are calculated by using MATLAB software. The calculation results are shown in Table 1:

[0135] Table 1 Characteristic calculation values

[0136]

[0137] It can be seen from Table 1 that when a short circuit occurs, the selected features are significantly different from those in the normal state, which verifies the feasibility of the selected feature values.

[0138] 3) Setting of short-circuit threshold and short-circuit identification

[0139] like Fig.15 As shown in FIG, according to the short-circuit characteristic calculation method, the characteristic vectors of short-circuit current and short-circuit voltage are constructed to determine the short-circuit discrimination threshold. Assume that the threshold vector of the short circuit of sample i is T i :

[0140]

[0141] where a i1 is the characteristic value of the amplitude feature; a i2 is the characteristic value of the peak-to-peak characteristic; a i3 is the characteristic value of the impact factor characteristic. Since the actual electrical operating environment is relatively complex, the voltage and current threshold vectors for judging the short-circuit state are selected with a certain margin:

[0142] when a i1 >10.3,a i2 >80.6,a i3 When <2.7, it is judged as short circuit state, otherwise it is normal state.

[0143] when a i1 <80.5, a i2 <200.2, a i3 When <3.2, it is judged as short circuit state, otherwise it is normal state.

[0144] The above-mentioned discrimination mechanism is embedded in the low-voltage power safety device to detect the occurrence of short circuit.

[0145] Furthermore, based on the final identification model, early warning of power failure is given;

[0146] Specifically, the low-voltage power safety hazard early warning model based on the BP algorithm: the fault arc hazard early warning model constructed predicts the probability of the next cycle of faults through the current data information and development trends, and measures the severity of power safety hazards. When the probability value of the fault arc hazard exceeds the early warning threshold, information is sent to the user and the relevant person in charge of the power grid to conduct a safety hazard inspection. The BP neural network with strong self-learning, organization and adaptability and wide application is used as an early warning model for arc faults. At the same time, the threshold judgment module is combined to divide the severity level of the hazard. The threshold can be adjusted according to the actual situation. The threshold setting is shown in the following formula:

[0147]

[0148] Among them, P f is the fault probability value output by the BP neural network early warning model. The overall idea of ​​the model is as follows Fig.16 shown.

[0149] Based on the extracted arc fault identification system indicators, a BP neural network arc fault warning model is constructed. The input layer has 11 nodes, corresponding to 11 characteristic data of the input current; the output layer corresponds to one node, which outputs the probability of arc fault in the next cycle as the warning information of the model; the number of nodes in the output layer and the input layer jointly affect the number of points added in the hidden layer. The model calculates the number of hidden layer nodes to be 23 by adding 1 to the sum of the number of nodes in the input layer and the output layer; the learning rate is set to 0.008; the algorithm flow is as follows Fig.17 shown.

[0150] 1. Set the random seed initialization weight and threshold: w i , b i ;

[0151] 2. Provide input sample X = (X1, X2, ..., X 11 ) and target output Y=Y1;

[0152] 3. Use the sigmoid function to construct the forward propagation graph and calculate the output of each layer: X(i) = f(X(i-1)*w i +b i );

[0153] 4. Use mean square error as loss function to calculate the training error of each layer;

[0154] 5. Update weights and thresholds;

[0155] 6. Output error indicators after each sample is trained 10 times;

[0156] 7. Compare the output error index with the set error value. When the error index is less than the set error, end the training, otherwise repeat the iterative training.

[0157] Experimental verification:

[0158] Analysis of training results: In the experiment, 3000 sets of normal operation and fault arc current data of various loads were collected. The feature data set was randomly shuffled and divided into training set and test set in a ratio of 4:1, of which the first 2500 sets of data were divided into training set and the last 500 sets of data were divided into test set. The feature data were read into the model for training and testing. The model training loss value curve and accuracy curve are shown as follows: Fig.18 and Fig.19 As shown in the figure, the curve shows that when the number of iterations exceeds 120, the accuracy gradually stabilizes and eventually stabilizes at around 97%, with the highest accuracy exceeding 99%. The model loss value decreases with the increase in the number of iterations, and when the number of iterations exceeds 60, it finally converges to around 0.01.

[0159] Test result analysis: The experimental data of several loads were read into the learned model to test the model performance. The accuracy is shown in Table 2.

[0160] Table 2 Experimental load prediction accuracy

[0161]

[0162] From the data in Table 2, it can be seen that the prediction accuracy of the early warning model for the electronic regulator load, which is difficult to identify whether an arc fault occurs in the time domain waveform, can reach more than 99%. The prediction accuracy for the computer load with a smaller current can also exceed 99%. The prediction accuracy of the arc fault of the hair dryer and electric drill load can reach more than 97%, which proves the effectiveness of the designed model. Accurately predicting the occurrence of a fault before the fault occurs, issuing early warning information and taking protective measures can greatly improve the safety of residents' electricity use and reduce the occurrence of electrical fires. The experimental loads include several common household appliances such as resistive, nonlinear and inductive, which are highly representative, indicating that the BP neural network early warning model has a strong generalization ability and can better mine the characteristics of current data.

[0163] In order to further improve the prediction effect of the model, a low-voltage power safety hazard warning model based on the random forest ensemble learning model is used to build the model. The random forest algorithm with fast training speed, ability to run on large data sets and not prone to overfitting is used. The specific steps are as follows: Fig. 20 As shown:

[0164] 1) Data acquisition. Normal and fault arc data with time series are collected through real experiments, and the data are preprocessed by denoising and normalization, and the experimental data are sampled through time windows.

[0165] 2) Feature engineering. By analyzing the experimental data, we extract the characteristic indicators with a high correlation with the fault arc, and construct a multi-dimensional fault arc hidden danger prediction feature data set. The selected features are: three frequency domain features such as the first harmonic, third harmonic, and fifth harmonic of the current, two time domain features such as the zero-crossing index and peak state, and Shannon entropy.

[0166] 3) Model hyperparameter setting. Random forest has relatively few hyperparameters. Based on the advantages of grid search, such as simple operation and the ability to find the optimal hyperparameters in a local range, the grid search local optimization method is selected to optimize the model hyperparameters.

[0167] 4) Model data fitting. After obtaining the sampled feature data set with time series, the fault arc hazard prediction feature data set is divided into a training set and a test set at a ratio of 4:1, and the training set is read into the model so that the model completes the training set data fitting under the optimal hyperparameter environment and obtains a learned fault arc hazard warning model.

[0168] 5) Model performance evaluation. The prediction accuracy and training time of the trained hidden danger warning model are evaluated through the test set, and multi-fold cross-validation is used to complete the model generalization ability test and overfitting evaluation.

[0169] The embodiment of the present invention also discloses a low-voltage electricity safety status monitoring and evaluation system, including:

[0170] The data acquisition module is used to obtain the operating data of low-voltage customers in normal operation and various power failure scenarios, build typical feature data sets, and perform preprocessing;

[0171] A model building module, used to build a recognition model and train the recognition model based on the preprocessed typical feature data set to obtain a final recognition model; the recognition model includes an autoencoder model, a random forest ensemble learning model and a BP neural network model;

[0172] The monitoring and early warning module is used to monitor and warn of power failures based on the final identification model; the monitoring of power failures includes accurate identification of fault arcs based on the autoencoder model, low-voltage user leakage judgment based on random forests, and short-circuit perception based on current mutation characteristics; the early warning of power failures includes low-voltage power safety hazard early warning based on the BP algorithm and low-voltage power safety hazard early warning based on the random forest ensemble learning model.

[0173] Specifically, in the data acquisition module, the corresponding operation data includes the residual current data Id i And total current data Ia i .

[0174] Specifically, in the data acquisition module, preprocessing includes:

[0175] The original operation data is denoised and the noise in the primary and secondary lines is filtered out by wavelet transform;

[0176] Normalize the denoised running data.

[0177] Specifically, in the monitoring and early warning module, accurate identification of fault arcs based on the autoencoder model includes: building an autoencoder model, eliminating the decoding stage, retaining the encoding stage model, and nesting the output of the encoding model into three fully connected layers and one softmax output layer to form a softmax classifier model for arc recognition.

[0178] Specifically, when performing arc identification, Fourier transform is used to extract the core features of the fault arc and sort them in order of importance, namely: fundamental wave, fundamental wave ratio, third harmonic, third harmonic ratio, fifth harmonic, fifth harmonic ratio, integral, variance, kurtosis, zero-crossing time, and Shannon entropy.

[0179] Specifically, in the monitoring and early warning module, the low-voltage user leakage identification based on random forest includes the following steps:

[0180] Obtaining raw data and preprocessing it, including eliminating the baseline value of the conditioning circuit, denoising the raw residual current data, extracting the sequence through a sliding time window, and processing the data out of order;

[0181] Use the Fourier transform method to extract different features from the original data, including fundamental wave, third harmonic, fifth harmonic, fundamental wave ratio, third harmonic ratio, fifth harmonic ratio, integral, Shannon entropy, kurtosis, and variance; divide the data set after feature extraction into training set, validation set, and test set according to the proportion;

[0182] Use the random forest algorithm to develop modeling, randomly extract multiple training sample subsets from the training set, perform decision tree modeling on each subset, integrate the decision results of multiple trees, and finally determine the detection results through voting, and complete the establishment of the low-voltage leakage detection model;

[0183] The training set is used as the input of the low-voltage leakage detection model to train the low-voltage system leakage detection model. Finally, the test set is used to predict the accuracy of the low-voltage system leakage detection model.

[0184] Specifically, in the monitoring and early warning module, short circuit perception based on current mutation characteristics includes the following steps:

[0185] Extract different features from the short-circuit voltage and current data. The features are calculated from the beginning of the second power frequency cycle of the voltage and current data; including peak-to-peak values, current and voltage amplitudes, and kurtosis factors;

[0186] Construct the characteristic vectors of short-circuit current and short-circuit voltage and determine the short-circuit judgment threshold;

[0187] The peak-to-peak value, current and voltage amplitude, kurtosis factor and short circuit discrimination threshold are compared to detect whether a short circuit occurs.

[0188] Specifically, in the monitoring and early warning module, the low-voltage power safety hazard early warning based on the random forest ensemble learning model includes the following steps:

[0189] Data acquisition: normal and fault arc data with time series are collected through real experiments, and the data are pre-processed by denoising and normalization, and the experimental data are sampled through time windows;

[0190] Feature engineering: By analyzing the experimental data, we extract characteristic indicators with a high degree of correlation with the fault arc, and construct a multi-dimensional fault arc hidden danger prediction feature data set; the selected features are: three frequency domain features of the first harmonic, third harmonic, and fifth harmonic of the current, two time domain features such as zero-crossing index and peak state, and Shannon entropy;

[0191] Model hyperparameter setting: Select the grid search local optimization method to optimize the model hyperparameters;

[0192] Model data fitting: After obtaining the sampled feature data set with time series, divide the fault arc hazard prediction feature data set into a training set and a test set at a ratio of 4:1, and read the training set into the model so that the model can complete the training set data fitting under the optimal hyperparameter environment and obtain a learned fault arc hazard warning model;

[0193] Model performance evaluation: The prediction accuracy and training time of the trained hidden danger warning model are evaluated through the test set, and multi-fold cross-validation is used to complete the model generalization ability test and overfitting assessment.

[0194] The present invention targets arc faults that are easy to occur in low-voltage electricity, are highly concealed, and are likely to cause electrical fires once they occur. Based on the early warning model of the BP neural network algorithm, early warning information is issued before the fault occurs and corresponding measures are taken to achieve accurate perception of arc faults in advance; by simulating the real low-voltage electricity scene, current data in normal and arc fault states are collected for waveform analysis, and model training, testing and evaluation are performed based on the low-voltage electricity safety hazard perception feature library. The method has strong generalization ability and can predict the occurrence of arc faults in advance, with a prediction accuracy of up to 80%. In addition, in order to further improve the low-voltage hidden danger warning effect, the random forest ensemble learning model is used for technical upgrading; the random forest ensemble learning model is trained, and the hyperparameters are optimized through the local optimization method, and the model parameters are adjusted with the goal of minimizing the node information entropy to complete the construction of the power safety hidden danger warning model; experimental verification shows that the proposed method can achieve accurate prediction of low-voltage power safety hazards. Under a variety of load operating conditions, the prediction accuracy of the method of the present invention is higher than that of the four mainstream classification prediction models such as GRU, RNN, DT, and BP, which provides a new method for improving the safety factor of residents' electricity use and reducing the occurrence of electrical fires.

[0195] In order to effectively ensure the safety of electricity use for low-voltage customers and accurately perceive common faults in low-voltage systems, such as fault arcs, leakage faults and short-circuit faults, the present invention first establishes a feature database of potential safety hazards in electricity use for low-voltage customers and performs data preprocessing operations on it; a low-voltage customer electricity safety perception model is established based on the feature database, which mainly includes: a fault arc precision identification model based on an automatic encoder, a low-voltage user leakage discrimination model based on a random forest, and a short-circuit perception model based on current mutation characteristics. Through in-depth testing and experiments on the model, it can be seen that the perception accuracy of the perception model can reach 80%.

[0196] The embodiment of the present invention further discloses a computer program product, including a computer program, which executes the steps of the above method when executed by a processor.

[0197] The embodiment of the present invention further discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described above are executed.

[0198] The product and medium of the present invention correspond to the above method and also have the advantages described in the above method.

[0199] The present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned method embodiment when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. Computer-readable storage media include: any entity or device that can carry computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. The memory is used to store computer programs and / or modules, and the processor implements various functions by running or executing computer programs and / or modules stored in the memory, and calling data stored in the memory. The memory may include high-speed random access memory and may also include non-volatile memory, such as a hard disk, an internal memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0200] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A low-voltage electricity safety status monitoring and evaluation method, characterized in that: Includes steps: S1. Obtain the operating data of low-voltage customers in normal operation and various power failure scenarios, build a typical feature data set, and perform preprocessing; S2. constructing a recognition model and training the recognition model based on the preprocessed typical feature data set to obtain a final recognition model; The identification model includes an autoencoder model, a random forest ensemble learning model and a BP neural network model; S3. Monitor and warn of power failures based on the final identification model; The monitoring of power failures includes accurate identification of fault arcs based on the autoencoder model, low-voltage user leakage identification based on random forests, and short-circuit perception based on current mutation characteristics. The early warning for power failure includes low-voltage power safety hazard early warning based on BP algorithm and low-voltage power safety hazard early warning based on random forest ensemble learning model.

2. The low-voltage electricity safety status monitoring and evaluation method according to claim 1 is characterized in that: In step S1, the corresponding operation data includes residual current data Id i And total current data Ia i .

3. The low-voltage electricity safety status monitoring and evaluation method according to claim 1 is characterized in that: In step S1, the preprocessing includes: The original operation data is denoised and the noise in the primary and secondary lines is filtered out by wavelet transform; Normalize the denoised running data.

4. The low-voltage electricity safety status monitoring and evaluation method according to claim 1, 2 or 3, characterized in that: In step S3, accurate identification of fault arcs based on the autoencoder model includes: building an autoencoder model, eliminating the decoding stage, retaining the encoding stage model, and nesting the output of the encoding model with three fully connected layers and one softmax output layer to form a softmax classifier model for arc identification.

5. The low-voltage electricity safety status monitoring and evaluation method according to claim 4 is characterized in that: When performing arc identification, Fourier transform is used to extract the core features of the fault arc and sort them in order of importance, namely: fundamental wave, fundamental wave ratio, third harmonic, third harmonic ratio, fifth harmonic, fifth harmonic ratio, integral, variance, kurtosis, zero-crossing time, and Shannon entropy.

6. The low-voltage electricity safety status monitoring and evaluation method according to claim 1, 2 or 3, characterized in that: In step S3, the low-voltage user leakage identification based on random forest includes the following steps: Obtaining raw data and preprocessing it, including eliminating the baseline value of the conditioning circuit, denoising the raw residual current data, extracting the sequence through a sliding time window, and processing the data out of order; Use the Fourier transform method to extract different features from the original data, including fundamental wave, third harmonic, fifth harmonic, fundamental wave ratio, third harmonic ratio, fifth harmonic ratio, integral, Shannon entropy, kurtosis, and variance; divide the data set after feature extraction into training set, validation set, and test set according to the proportion; Use the random forest algorithm to develop modeling, randomly extract multiple training sample subsets from the training set, perform decision tree modeling on each subset, integrate the decision results of multiple trees, and finally determine the detection results through voting, and complete the establishment of the low-voltage leakage detection model; The training set is used as the input of the low-voltage leakage detection model to train the low-voltage system leakage detection model. Finally, the test set is used to predict the accuracy of the low-voltage system leakage detection model.

7. The low-voltage electricity safety status monitoring and evaluation method according to claim 1, 2 or 3, characterized in that: In step S3, short circuit sensing based on current mutation characteristics includes the following steps: Extract different features from the short-circuit voltage and current data. The features are calculated from the beginning of the second power frequency cycle of the voltage and current data; including peak-to-peak values, current and voltage amplitudes, and kurtosis factors; Construct the characteristic vectors of short-circuit current and short-circuit voltage and determine the short-circuit judgment threshold; The peak-to-peak value, current and voltage amplitude, kurtosis factor and short circuit discrimination threshold are compared to detect whether a short circuit occurs.

8. The low-voltage electricity safety status monitoring and evaluation method according to claim 1, 2 or 3, characterized in that: In step S3, the low-voltage power safety hazard warning based on the random forest ensemble learning model includes the following steps: Data acquisition: normal and fault arc data with time series are collected through real experiments, and the data are pre-processed by denoising and normalization, and the experimental data are sampled through time windows; feature Engineering: By analyzing the characteristics of experimental data, characteristic indicators with a high degree of correlation with fault arcs are extracted, and a multi-dimensional fault arc hidden danger prediction feature data set is constructed; the selected features are: three frequency domain features of the first harmonic, third harmonic, and fifth harmonic of current, two time domain features such as zero-crossing index and peak state, and Shannon entropy; Model hyperparameter setting: Select the grid search local optimization method to optimize the model hyperparameters; Model data fitting: After obtaining the sampled feature data set with time series, divide the fault arc hazard prediction feature data set into a training set and a test set at a ratio of 4:1, and read the training set into the model so that the model can complete the training set data fitting under the optimal hyperparameter environment and obtain a learned fault arc hazard warning model; Model performance evaluation: The prediction accuracy and training time of the trained hidden danger warning model are evaluated through the test set, and multi-fold cross-validation is used to complete the model generalization ability test and overfitting assessment.

9. A low voltage electricity safety status monitoring and evaluation system, characterized in that: include: The data acquisition module is used to obtain the operating data of low-voltage customers in normal operation and various power failure scenarios, build typical feature data sets, and perform preprocessing; A model building module is used to build a recognition model and train the recognition model based on the preprocessed typical feature data set to obtain a final recognition model; The identification model includes an autoencoder model, a random forest ensemble learning model and a BP neural network model; Monitoring and early warning module, used to monitor and warn of power failures based on the final identification model; The monitoring of power failures includes accurate identification of fault arcs based on the autoencoder model, low-voltage user leakage identification based on random forests, and short-circuit perception based on current mutation characteristics. The early warning for power failure includes low-voltage power safety hazard early warning based on BP algorithm and low-voltage power safety hazard early warning based on random forest ensemble learning model.

10. The low-voltage electricity safety status monitoring and evaluation system according to claim 9 is characterized in that: In the data acquisition module, the corresponding operation data includes the residual current data Id i And total current data Ia i .

11. The low-voltage electricity safety status monitoring and evaluation system according to claim 9, characterized in that: In the data acquisition module, preprocessing includes: The original operation data is denoised and the noise in the primary and secondary lines is filtered out by wavelet transform; Normalize the denoised running data.

12. The low voltage electricity safety status monitoring and evaluation system according to claim 9, 10 or 11, characterized in that: In the monitoring and early warning module, accurate identification of fault arcs based on the autoencoder model includes: building an autoencoder model, eliminating the decoding stage, retaining the encoding stage model, and nesting the output of the encoding model into three fully connected layers and one softmax output layer to form a softmax classifier model for arc identification.

13. The low-voltage electricity safety status monitoring and evaluation system according to claim 12, characterized in that: When performing arc identification, Fourier transform is used to extract the core features of the fault arc and sort them in order of importance, namely: fundamental wave, fundamental wave ratio, third harmonic, third harmonic ratio, fifth harmonic, fifth harmonic ratio, integral, variance, kurtosis, zero-crossing time, and Shannon entropy.

14. The low voltage electricity safety status monitoring and evaluation system according to claim 9, 10 or 11, characterized in that: In the monitoring and early warning module, the low-voltage user leakage identification based on random forest includes the following steps: Obtaining raw data and preprocessing it, including eliminating the baseline value of the conditioning circuit, denoising the raw residual current data, extracting the sequence through a sliding time window, and processing the data out of order; Use the Fourier transform method to extract different features from the original data, including fundamental wave, third harmonic, fifth harmonic, fundamental wave ratio, third harmonic ratio, fifth harmonic ratio, integral, Shannon entropy, kurtosis, and variance; divide the data set after feature extraction into training set, validation set, and test set according to the proportion; Use the random forest algorithm to develop modeling, randomly extract multiple training sample subsets from the training set, perform decision tree modeling on each subset, integrate the decision results of multiple trees, and finally determine the detection results through voting, and complete the establishment of the low-voltage leakage detection model; The training set is used as the input of the low-voltage leakage detection model to train the low-voltage system leakage detection model. Finally, the test set is used to predict the accuracy of the low-voltage system leakage detection model.

15. The low voltage electricity safety status monitoring and evaluation system according to claim 9, 10 or 11, characterized in that: In the monitoring and early warning module, short circuit perception based on current mutation characteristics includes the following steps: Extract different features from the short-circuit voltage and current data. The features are calculated from the beginning of the second power frequency cycle of the voltage and current data; including peak-to-peak values, current and voltage amplitudes, and kurtosis factors; Construct the characteristic vectors of short-circuit current and short-circuit voltage and determine the short-circuit judgment threshold; The peak-to-peak value, current and voltage amplitude, kurtosis factor and short circuit discrimination threshold are compared to detect whether a short circuit occurs.

16. The low voltage electricity safety status monitoring and evaluation system according to claim 9, 10 or 11, characterized in that: In the monitoring and early warning module, the low-voltage power safety hazard early warning based on the random forest ensemble learning model includes the following steps: Data acquisition: normal and fault arc data with time series are collected through real experiments, and the data are pre-processed by denoising and normalization, and the experimental data are sampled through time windows; feature Engineering: By analyzing the characteristics of experimental data, characteristic indicators with a high degree of correlation with fault arcs are extracted, and a multi-dimensional fault arc hidden danger prediction feature data set is constructed; the selected features are: three frequency domain features of the first harmonic, third harmonic, and fifth harmonic of current, two time domain features such as zero-crossing index and peak state, and Shannon entropy; Model hyperparameter setting: Select the grid search local optimization method to optimize the model hyperparameters; Model data fitting: After obtaining the sampled feature data set with time series, divide the fault arc hazard prediction feature data set into a training set and a test set at a ratio of 4:1, and read the training set into the model so that the model can complete the training set data fitting under the optimal hyperparameter environment and obtain a learned fault arc hazard warning model; Model performance evaluation: The prediction accuracy and training time of the trained hidden danger warning model are evaluated through the test set, and multi-fold cross-validation is used to complete the model generalization ability test and overfitting assessment.

17. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are performed.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 8.

Images