An adaptive fault protection method considering uncertainty of distribution network equipment access

Abstract

The application discloses a self-adaptive fault protection method considering distribution network equipment access uncertainty, and on the basis of an existing feature-driven intelligent model, proposes a forgetting learning method to improve a continuous learning method to update a fault detection model, adaptively detects and judges a new type of fault caused by new equipment conditions, and after adaptively detecting a fault arc, updates a fault positioning model by using a forgetting learning method to improve a category incremental learning method, and adaptively selects a line to output branch and equipment information where the fault is located. After judging that detection and line selection features are invalid by using a maximum mutual information coefficient, effective features are obtained by using a parameter optimization method or a feature set screening method, so that the problem that detection and line selection are caused by fault characteristic changes caused by distribution network equipment access is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electrical fault detection, and particularly relates to a self-adaptive fault protection method considering the uncertainty of distribution network equipment access. BACKGROUND

[0002] Flexible resources such as photovoltaic and energy storage will be adjusted in real time according to the power load, and thus these source-load uncertainty factors will cause real-time changes in the fault arc characteristics. In addition, the diverse power electronic topology strategies, multi-element power electronic device composition stages and connection modes in the system will also interfere with the conduction process of the fault arc characteristics. The increase in cable length, distributed capacitance parameters and connection heads will cause the phenomenon of attenuation or even capture of interference noise in the fault frequency band of 100 kHz to 500 MHz, thus causing changes in the collected fault electrical characteristics. The original direct current fault detection and positioning technology in the special scene is no longer applicable, and there will be a risk of failure when dealing with these new fault modes. Therefore, it is urgent to research a self-adaptive fault protection method, continuously improve and optimize the existing fault diagnosis algorithm, and improve its adaptability to new fault types and fault modes.

[0003] At present, the fault detection and positioning technology mainly includes the following: (1) fault detection and positioning technology based on current and voltage signals: by monitoring the changes of current and voltage signals, the characteristics of fault arc are extracted to realize detection and positioning, which is mainly applied to low-voltage power systems; (2) fault detection and positioning technology based on optical fiber sensors: the changes of optical fiber temperature, optical fiber bending and other parameters in the power system are monitored by optical fiber sensors to realize fault detection and positioning, which is mainly applied to medium and high voltage power systems; (3) fault detection and positioning technology based on infrared imaging technology: infrared thermal imager is used to realize infrared imaging of fault arc in the power system, and the thermal characteristics of temperature rise of fault-caused equipment or environment are analyzed to realize fault detection and positioning, which is mainly applied to high-voltage power systems; (4) fault detection and positioning technology based on multi-source signal sensing analysis: wireless sensor network is used to monitor the changes of current, voltage and temperature and other parameters in the power system to realize fault detection and positioning, which has the advantages of wireless transmission and real-time monitoring, and is mainly applied to medium and high voltage power systems.

[0004] Currently, fault diagnosis technology mainly focuses on the analysis level of application electrical signal. For example, the patent entitled "A series arc fault detection and positioning method for multiple loads" first collects the alternating current loop current signal, extracts multiple dimensional features including variance, correlation coefficient and peak factor index based on time domain analysis, harmonic amplitude based on discrete Fourier transform, and frequency band energy and wavelet entropy features based on discrete wavelet transform, and inputs these features into a deep neural network for arc detection and positioning. However, this fault protection method can only deal with the load types and arc characteristics learned during neural network training. The features of new working conditions may be mixed due to the aliasing of historical working condition features, and the existing detection and line selection model often performs a series of targeted designs such as frequency band selection and threshold construction to obtain better performance. It can only describe the historical working condition feature form and its recognition situation, and cannot detect the fault arc in unknown new working conditions when new distribution network equipment is connected. The detection performance of the model may be limited, and the detected arc cannot be judged for the location. SUMMARY

[0005] In order to solve the problems in the prior art, the present application provides an adaptive fault protection method considering the uncertainty of distribution network equipment access, which proposes a forgetting learning method to improve continuous learning for updating the fault detection model, and a forgetting learning method to improve class incremental learning for updating the fault positioning model, thereby solving the detection and line selection problems caused by the change of fault arc characteristics due to uncertain source and load.

[0006] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows: an adaptive fault protection method considering the uncertainty of distribution network equipment access, comprising the following steps:

[0007] Step 1, real-time acquisition of system current information under the condition of flexible access and switching of distribution network equipment;

[0008] Step 2, time-frequency domain analysis of the acquired system current information, extraction of detection and line selection features, and obtaining of a feature set;

[0009] Step 3, failure judgment of the detection features, if the detection features are not failed, go to step 4, if the detection features are failed, first optimize the parameters for the current detection feature form, then perform failure judgment, if the improved detection features are not failed, combine the historical features and perform dimensionality reduction processing, then go to step 4, if the improved detection and line selection features are still failed, screen the fault features from the feature set obtained in step 2, combine the obtained fault features with the historical features and perform dimensionality reduction processing, then go to step 4;

[0010] Step 4, the fault intelligent detection model is extended by a method for improving continuous learning based on forgetting learning, and the fault intelligent detection model is used for fault state detection based on the detection features obtained in step 2, the fault adaptive detection is completed, and after the fault is detected, the process goes to step 5, otherwise, the process returns to step 1;

[0011] Step 5, the selectivity of the feature is judged, if the selectivity of the feature is not lost, the process goes to step 6, if the selectivity of the feature is lost, the parameters of the current selectivity feature are optimized, and then the selectivity of the improved feature is judged, if the selectivity of the improved feature is not lost, the feature is combined with the historical feature and processed by dimensionality reduction, and then the process goes to step 7, if the selectivity of the improved feature is still lost, the feature is filtered in the feature set, and then the feature is combined with the historical feature and processed by dimensionality reduction, and then the process goes to step 7;

[0012] Step 6, it is judged whether the selectivity of the feature is new distribution equipment information, if yes, the process goes to step 7, otherwise, the process goes to step 8;

[0013] Step 7, the fault intelligent positioning model is extended by a method for improving category incremental learning based on forgetting learning;

[0014] Step 8, the fault is located by using an artificial intelligence network, and the branch and equipment information where the fault is located are output.

[0015] Further, the fault detection and selectivity feature is a time-frequency domain feature.

[0016] Further, the fault detection and selectivity model is a recurrent neural network; for the fault detection model, the recurrent neural network is trained by using the electromagnetic oven features based on wavelet packet decomposition, and then a fault intelligent detection model is obtained; for the selectivity model, the constructed recurrent neural network is trained by using the steady-state features of the old scene, and then a fault intelligent positioning model is obtained.

[0017] Further, the fault intelligent positioning model is extended by a method for improving category incremental learning based on forgetting learning, including the following steps: first, a node is added to the output layer of the direct current fault arc selectivity model, then the data of the old task is reserved by using the knowledge distillation method and the weight factor, so that the prediction value of the new model for the old task is consistent with the prediction value of the old model for the old task, and the cross-entropy loss and the weight factor are used to remember the new task.

[0018] Further, the detection feature failure judgment is based on:

[0019] Firstly, the maximum mutual information coefficient is used to measure the correlation between the fault feature and the historical normal operation system feature, and the mutual information calculation formula is as follows:

[0020]

[0021] In the formula: I(x, y) is mutual information, p(x, y) is the joint probability distribution function of x and y, p(x) and p(y) are the marginal probability distribution functions of x and y, when the mutual information is greater than a set threshold, it is determined that the detection feature is related and fails;

[0022] The line selection feature failure judgment is based on:

[0023] First, the maximum mutual information coefficient is used to measure the correlation degree between the normal features and the normal operation system features corresponding to different distribution network equipment and branch states in history, when the mutual information is greater than a set threshold, it is determined that the line selection feature is related and fails.

[0024] Further, the feature parameter optimization process is that, without changing the feature framework, the feature parameters are solved by using an optimization algorithm to realize that the feature output is not failed under the current fault state or equipment operation state, and the optimization algorithm adopts a particle swarm optimization algorithm or a genetic algorithm.

[0025] Further, the feature set includes time domain features, statistical features, time-frequency domain features and magneto-optical acoustic physical features, and the selection of the features whose outputs are not failed under the current fault state or the features whose outputs are not failed under the equipment operation state is completed by evaluation.

[0026] Further, first, the maximum mutual information coefficient is used to measure the correlation degree between the non-failed line selection features in step 6 and the normal operation system features corresponding to different distribution network equipment and branch states in history, when the mutual information coefficient reflecting the correlation degree is greater than a set threshold, it is determined that the non-failed line selection feature presents related historical distribution network equipment information, otherwise, it is determined that the non-failed line selection feature presents new distribution network equipment information.

[0027] Further, for a plurality of line selection features, the ternary loss method in the formula is used to narrow the distance between the data of the same type of equipment and to widen the distance between the data of different types of equipment:

[0028] L = max (D (a, p) - D (a, n) + a, 0)

[0029] In the formula: L represents ternary loss, D(a, p) represents the distance between equipment samples a and sample b, D(a, n) represents the distance between equipment samples a and sample n, and a represents a margin function.

[0030] Further, the feature parameter optimization and feature set screening method automatically analyzes the most suitable fault detection feature according to the system architecture or application scenario.

[0031] Compared with the prior art, the present application has at least the following beneficial effects:

[0032] The direct current fault arc adaptive detection based on the improved continuous learning of the forgetting learning is proposed, the existing fault intelligent detection model is updated through the improved continuous learning of the forgetting learning algorithm, the diagnosis rate of the existing fault intelligent detection model facing unknown working conditions can be effectively improved, the problems of efficient and adaptive detection are solved, the catastrophic forgetting phenomenon of the model fine-tuning direct current fault intelligent detection model does not occur, the original fault arc diagnosis rate reduction problem under the condition of new type fault detection does not occur, cross-condition fault arc accurate detection is realized, and environmental adaptability is achieved.

[0033] The direct current fault arc adaptive line selection based on the improved category incremental learning of the forgetting learning is proposed, the existing direct current fault arc line selection model is updated through the improved category incremental learning of the forgetting learning algorithm, the diagnosis rate of the existing direct current fault arc line selection model facing unknown distribution network equipment can be effectively improved, the problems of efficient and adaptive line selection are solved, the catastrophic forgetting phenomenon of the model fine-tuning direct current fault intelligent line selection model does not occur, the original system equipment identification diagnosis rate reduction problem under the condition of new type distribution network system operation does not occur, cross-condition fault arc reliable line selection is realized, and a foundation is laid for convenient operation and maintenance of faults.

[0034] The fault feature efficient dimensionality increasing new idea of parameter optimization first and feature set screening later is proposed, the fault detection range is expanded through feature dimensionality increasing, first, new features suitable for the current working condition are formed through parameter optimization on the existing feature architecture, and then when the current feature structure is invalid, effective feature types are efficiently searched in the feature set according to the current working condition and are integrated, so that the new protection problem caused by the detection and line selection feature invalidation of new fault working conditions is solved, and then the faults under unknown new working condition conditions when new type distribution network equipment is accessed are maximally detected, and the position of the detected arc is pointed out.

[0035] The feature efficient dimensionality increasing new idea can also automatically analyze the most suitable detection new arc feature according to the parameter input of different system architectures, so that the current scene can be applied by using only a single new type feature without integrating the original feature set, so that the feature invalidation risk is reduced, the development cost of the fault arc detection device or the disconnector is reduced, and the development time of the device is shortened. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The flowchart of the adaptive fault protection method considering the uncertainty of distribution network equipment access in the embodiment of the application.

[0037] Figure 2a The current signal waveform diagram of the induction cooker.

[0038] Figure 2b The current signal waveform diagram of the television.

[0039] Figure 3a Wavelet feature waveform diagram of current signal of induction cooker.

[0040] Figure 3b Wavelet feature waveform diagram of current signal of television.

[0041] Figure 4a Adaptive detection result of fault arc in induction cooker and television scene based on gated recurrent unit algorithm.

[0042] Figure 4b Adaptive detection result of fault arc in multiple new scenes based on gated recurrent unit algorithm.

[0043] Figure 5a Adaptive detection result of fault arc in induction cooker and television scene based on model fine-tuning improved gated recurrent unit algorithm.

[0044] Figure 5b Adaptive detection result of fault arc in induction cooker and television scene based on forgetful learning improved continuous learning.

[0045] Figure 5c Adaptive detection result of fault arc in multiple new scenes based on forgetful learning improved continuous learning.

[0046] Figure 6a Adaptive line selection result of fault arc in multiple new scenes based on gated recurrent unit algorithm.

[0047] Figure 6b Adaptive line selection accuracy of fault arc in multiple new scenes based on gated recurrent unit algorithm.

[0048] Figure 7a Adaptive line selection accuracy of fault arc in multiple new scenes based on model fine-tuning improved gated recurrent unit algorithm.

[0049] Figure 7b Adaptive line selection accuracy of fault arc in multiple new scenes based on forgetful learning improved category incremental learning.

[0050] Figure 8 Structure diagram of experimental system model built by the embodiment of the present application. DETAILED DESCRIPTION

[0051] The present application provides a kind of adaptive fault protection method considering distribution network equipment access uncertainty, comprising the following steps:

[0052] Step 1, real-time acquisition system current information under the condition of distribution network equipment flexible access, switching;

[0053] Step 2, time-frequency domain analysis is performed on the obtained system current information to extract detection and line selection features, and a feature set is obtained;

[0054] Step 3, failure judgment is performed on the detection features. If the detection features are not failed, go to Step 4. If the detection features are failed, first perform parameter optimization on the current detection feature form, and then perform failure judgment. If the improved detection features are not failed, combine the historical features for dimensionality reduction processing, and then go to Step 4. If the improved detection and line selection features are still failed, filter the fault features from the feature set obtained in Step 2, combine the obtained fault features with the historical features for dimensionality reduction processing, and then go to Step 4;

[0055] Step 4, the fault intelligent detection model is extended through the method of improving continuous learning based on forgetting learning. The fault state is detected based on the fault intelligent detection model and the detection features obtained in Step 2, the fault adaptive detection is completed, and after the fault is detected, go to Step 5, otherwise return to Step 1;

[0056] Step 5, failure judgment is performed on the line selection features. If the line selection features are not failed, go to Step 6. If the line selection features are failed, first perform parameter optimization on the current line selection feature form, and then perform failure judgment. If the improved line selection features are not failed, combine the historical features for dimensionality reduction processing, and then go to Step 7. If the improved line selection features are still failed, filter the fault features from the feature set, combine the obtained fault features with the historical features for dimensionality reduction processing, and then go to Step 7;

[0057] Step 6, judge whether the unfailed line selection features present new distribution network equipment information. If yes, go to Step 7, otherwise go to Step 8;

[0058] Step 7, the fault intelligent positioning model is extended through the method of improving category incremental learning based on forgetting learning;

[0059] Step 8, the fault is line-selected and positioned by using an artificial intelligence network, and the branch and equipment information where the fault is located are output.

[0060] The method of improving continuous learning based on forgetting learning for extending the fault intelligent detection model includes the following steps:

[0061] a. judge whether the original detection features are failed;

[0062] b. if failed, perform parameter optimization or feature set filtering;

[0063] c. on the basis of the old model, new data is processed through the branch of the old task to obtain the output Y0 of the old task;

[0064] d Then use the new data to train the network, make the old task keep the output Y0 by the method of knowledge distillation, calculate the loss and weight factor for gradient update constraint, so that it will not completely deviate from the old task;

[0065] e If not failed, skip the steps of performing parameter optimization or performing feature set screening, and perform steps c and d.

[0066] The method for improving the class incremental learning based on the forgetting learning is used to extend the fault intelligent positioning model, including the following steps:

[0067] a Determine whether the original selected line feature is failed;

[0068] b If the original selected line feature is failed, perform parameter optimization or perform feature set screening;

[0069] c Add a node to the output layer of the network for realizing the multi-classification task;

[0070] d On the basis of the old model, make the prediction value consistent through the branch of the old task by using the new data, obtain the output Y0 of the old task, and use the new data to train the network, so that the output Y0 of the old task is kept through the method of knowledge distillation;

[0071] e Calculate the loss and weight factor for gradient update constraint, so that it will not completely deviate from the old task;

[0072] f If the original selected line feature is not failed and new distribution network equipment appears, steps c-e are performed; if no new distribution network equipment appears, jump to step 8.

[0073] The application will be described in detail below in combination with the drawings and embodiments. The embodiment of the adaptive fault protection method considering the uncertainty of distribution network equipment access is only used for explaining the application, and does not limit the protection scope of the application.

[0074] Step one, according to the frequency f s The current signal output by the low-voltage direct-current distribution network is sampled to obtain the fault current waveform under different conditions and load conditions. As an example, the current data of the electromagnetic oven, television and other direct-current loads under two conditions of normal operation and occurrence of direct-current fault arc are collected.

[0075] Step two, the current waveform of the load is decomposed into four layers with Rbio3.1 as the fundamental wave, and the frequency band 31.25kHz-62.5kHz is selected to analyze the fault arc current. The current signal of the load is decomposed by wavelet packet to obtain the characteristic signal diagram of the current, and the characteristic signal diagram of the current is combined with the current signal diagram to obtain the direct-current fault arc current and detection characteristic diagram.

[0076] Step three, the fault arc in the detection results into the trained artificial intelligence model network based on the gated recurrent unit, in the process, the features extracted from other loads will be as the unknown working conditions that may occur in the power system, and then use the constructed direct current fault arc detection model to detect the fault arc of the features of other loads, output level 1 when the load in the line is normal, output level 2 when the load in the line has a fault arc, judge whether a fault arc occurs, if no fault arc occurs, return to step one for new signal acquisition, if a fault arc occurs, go to step five, and record the actual detection situation of the experiment, finally compare with the true label of the data, and calculate the diagnosis rate.

[0077] The direct current fault arc detection model constructed in this embodiment is as follows: a multi-source load direct current fault arc experimental simulation platform is built according to the UL1699B standard, and the structure of the experimental system is as shown in Figure 8

[0078] It includes three voltage level uses: 750VDC, 400VDC, and 48VDC. Since fault arcs may not occur during the experiment, in order to obtain current data when a fault arc occurs, the fault arc occurrence device is connected in series to the system loop, and by controlling the opening and closing of the fault arc occurrence device, the current data of the load in normal operation and when a fault arc occurs can be obtained. The arrangement of the detection nodes is the red circles A and B in the figure, and by controlling the opening and closing of the switch, different loads can be connected to the system to obtain current data of different loads. Since there are many types of loads involved in the experiment, four loads, 2.1kW electromagnetic oven, 120W TV, 90W LED (Light Emitting Diode, LED) lamp, and 36W electric fan, are taken as examples.

[0079] Step four, using the continuous learning method improved based on forget learning for self-adaptive detection of direct current fault arc, the means of adaptive updating of the fault arc detection model is also the most critical step, in this stage, first the decision output Y0 of the old direct current fault arc detection model to the old task is obtained, then the new direct current fault arc detection model is made to still maintain the decision output Y0 to the old task through the way of knowledge distillation, because the uncertainty of loss reduction may cause catastrophic forgetting to the old task, and the gradient cannot find a common point between the new task and the old task, so by calculating the loss and the weight factor to constrain the update of the gradient, the decision output will not completely deviate to the old task.

[0080] ​Step five, test the existing DC arc fault line selection model with the old scene of induction cooker and TV and the new scene of electric fan and LED lamp. The current data of different loads is calculated by skewness to obtain the feature extraction result, and the feature quantity based on skewness is upgraded to obtain the effective feature quantity for load identification. The steady-state feature quantity of the new scene after S transformation is obtained, and the current data of different loads has obvious distinguishability. A new artificial intelligence model network based on gated recurrent unit (GRU) is constructed, and the steady-state features of induction cooker and TV are used as old scenes to train the constructed GRU network, and then a fault intelligent positioning model is obtained. The steady-state features of electric fan and LED lamp are used as unknown load labels and new scenes. Determine whether the original feature is invalid: if it is invalid, continue to find or add new feature analysis until the feature is not invalid; if the feature is not invalid and new distribution network equipment information is presented, the fault intelligent positioning model is extended by using the LWF improved category incremental learning method. Specifically, a node is added to the output layer of the DC arc fault line selection model to realize multi-classification task. Then, the knowledge distillation method and weight factor are used to retain the data of the old task, so that the prediction value of the new model for the old task is consistent with the prediction value of the old model for the old task. Then, cross-entropy loss and weight factor are used to remember the new task. In this stage, the prediction probability of the new model for the new task is encouraged to be consistent with the true label of the data.

[0081] From Figure 2a and Figure 2b It can be seen that the current amplitude of induction cooker and TV is close to 4A when they are running normally, and after the occurrence of arc fault, the current amplitude decreases, the high frequency harmonic component increases, and the time domain waveform distorts, etc. When the arc fault ends, the circuit has been damaged, which may cause electrical accidents, so it is necessary to detect the arc fault, and the fluctuation can roughly identify the occurrence of arc fault.

[0082] Figure 3a and Figure 3bThe two loads are taken as Rbio3.1 as the small base wave, and the four-layer decomposition is selected, and the wavelet packet decomposition result in the frequency band of 31.25 kHz-62.5 kHz. The more the decomposition layers, the higher the frequency resolution, and the more complex the calculation. From the analysis of the DC fault arc feature extraction effect diagram, it can be concluded that the best scheme for DC fault arc feature extraction is to select four-layer decomposition and take the frequency band of 31.25 kHz-62.5 kHz for wavelet packet decomposition. At this time, the characteristic quantity difference before and after the fault arc occurs is large. The arc wavelet features of the television set load under the operating condition and the arc wavelet features of the induction cooker under the operating condition although the pulses generated at the fault occurrence time present significant differences, but the feature distribution amplitude range under the fault state is very close, so it is possible to produce misoperation under the fault condition of the television set load interference.

[0083] The old scene is the arc wavelet feature of the induction cooker, which is input into the artificial intelligence model network based on the gated recurrent unit (GRU) for detection model training, and the trained GRU arc detection model is obtained. The fault arc conditions under the operating conditions of the induction cooker and the television set are tested, Figure 4a The fault arc detection results of the induction cooker and the television set are 99.84% for the induction cooker fault arc diagnosis rate, which reaches high-precision detection. The diagnosis rate of the television set fault arc is 48.04%, and the diagnosis effect is not ideal, so the existing DC fault arc detection model faces the single working condition of the new scene, and the diagnosis effect of the DC fault arc is poor.

[0084] The old scene is the arc wavelet feature of the induction cooker, which is input into the artificial intelligence model network based on the gated recurrent unit (GRU) for detection model training, and the trained GRU arc detection model is obtained. The fault arc detection results of the new scene under multiple working conditions are as Figure 4b shown. In order to ensure the accuracy of the experiment, this time the new working condition increases the load of the refrigerator and other loads. As can be seen from the figure, the diagnosis accuracy of the fault arc under the old scene is more than 99%, and the diagnosis effect is good. When the existing fault arc detection model faces multiple new scenes, the diagnosis accuracy of most normal load stages is high, but the fault arc diagnosis accuracy of the fault arc stage and the overall stage is not ideal, so the existing DC fault arc detection model has certain limitations.

[0085] By fine-tuning the existing DC fault arc adaptive detection model, the knowledge and data characteristics learned in the source domain are used to help the target domain training, accelerate model convergence, and perform fault arc adaptive detection. When the old scene is the induction cooker and the new scene is the television set, the DC fault arc detection results are as Figure 5aAs shown in the table, the fault arc diagnosis rates are 43.88% and 99.09% respectively, so it can be seen that the fault diagnosis rate of the current DC fault arc detection model for the new scene has had a qualitative leap compared to before. The fault arc diagnosis rate of the current DC fault arc detection model for the induction cooker is only 43.88%, so it shows that the DC fault arc detection model optimized by the fine-tuning method has "catastrophic forgetting" for the previous data, resulting in a decrease in the fault arc diagnosis rate of the old scene. Although fine-tuning does not need to train the network from scratch for new tasks, saving time and cost, the optimized fault arc detection model still has certain limitations, that is, it has forgotten the old data.

[0086] The model is fine-tuned as follows: on the basis of retaining the existing network architecture and parameter values, the trained model is optimized and adjusted to improve its performance and accuracy. After the initial training of the model, the model parameters, hyperparameters or model structure are adjusted and optimized according to the characteristics of the actual application scene and data set, without involving the pruning and reconstruction of network parameters, using the trained model parameters as the initialization parameters of the new model, and requiring the source model and target model to have the same layer name, type and layer design parameters, etc. Fine-tuning includes the following steps:

[0087] (1) Pre-train a neural network model on the source domain data set, i.e. the source model;

[0088] (2) Create a target model that copies all model designs and parameters of the source model except the output layer;

[0089] (3) Add an output layer with an output size of the number of target data set types to the target model, and randomly initialize the model parameters of the layer;

[0090] (4) Train the target model on the target data set. We will train the output layer from scratch, while the parameters of the remaining layers are frozen based on the parameters of the source model.

[0091] The adaptive detection results of the DC fault arc based on the LWF improved continuous learning method are shown in Table 2. Figure 5b As shown in the table, the adaptive detection results of the DC fault arc for the new scene of the TV set are shown in Table 2, and the fault arc diagnosis rate is 98.36%. From the figure, it can be seen that the diagnosis accuracy of the existing DC fault arc detection model for the new scene has been significantly improved compared to before, and there is no phenomenon of catastrophic forgetting for the old scene. From this, it can be seen that the existing DC fault arc detection can adaptively detect the DC fault arc under the condition of a single working condition for the new scene, so next we will test whether the DC fault arc detection model optimized by the LWF improved continuous learning method can adaptively detect the DC fault arc under the condition of multiple working conditions for the new scene.

[0092] The old scene is an induction cooker, and the fault arc adaptive detection results of the new scene under multiple working conditions are as shown in Figure 5c To ensure the accuracy of the experiment, other loads such as refrigerators are also added in the same sample new working condition. As can be seen from the figure, the diagnostic accuracy of the old load has a slight decline compared with Figure 4b However, the diagnostic accuracy of the improved direct current fault arc adaptive detection model based on the LWF algorithm has significantly improved, and the detection accuracy of the three stages is more than 80%. When new working conditions appear in the line, the existing direct current fault arc detection model can adaptively detect whether a fault arc has occurred in the new working condition.

[0093] A new artificial intelligence model network based on a gated recurrent unit is constructed, and then the steady-state characteristics of the induction cooker and the television are used as the old scene to train the constructed GRU network, and then a fault intelligent positioning model is obtained. The steady-state characteristics of the electric fan and the LED lamp are used as the label of the unknown load and as the new scene. Among them, the induction cooker, television, electric fan and LED lamp are marked as levels 1, 2, 3 and 4 respectively. Then the existing pre-recognition positioning model is used to adaptively recognize and position the steady-state characteristics of the electric fan and the LED lamp, and then the adaptive recognition results of different loads are as shown in Figure 6a As can be seen from the figure, the recognition accuracy of the existing pre-recognition positioning model for the induction cooker and the television is as high as 100%, while a large number of misjudgments occur for the electric fan and the LED lamp, and the recognition accuracy is 0%.

[0094] The existing direct current fault arc selection model is tested with the old scene as the induction cooker and the television, and the new scene as the electric fan and the LED lamp. The fault arc selection results of the new scene under multiple working conditions when the old scene is the induction cooker and the television are as shown in Figure 6b As can be seen from the figure, the existing direct current fault arc selection model can make very good identification for the old scene, while for the new scene, it appears to be ineffective, so the existing direct current fault arc pre-selection model has certain limitations and needs to explore new technologies to solve this problem.

[0095] The adaptive recognition results of the direct current fault arc selection model based on model fine-tuning improvement when the old scene is the induction cooker and the television and the new scene is the LED lamp and the electric fan are as shown in Figure 7a As can be seen from the figure, the recognition accuracy of the fault intelligent positioning model after model fine-tuning extension for the new scene of the electric fan and the LED lamp is 95.5% and 96.68% respectively, and the recognition effect is good. However, the recognition accuracy for the old scene is 0%. Therefore, the fault intelligent positioning model after model fine-tuning extension has catastrophic forgetting for the old task, and the overall recognition effect is poor.

[0096] By training based on class-incremental learning improvement, when the old scene is an induction cooker and a TV set, the adaptive fault arc selection results of multiple working conditions in the new scene are as shown in Figure 7b Fig. 3, from which it can be seen that the fault intelligent positioning model improved by the LWF algorithm and the class-incremental learning method can still maintain good load identification accuracy for multiple new working conditions, and can also maintain good load identification rate for old working conditions. The results show that when a direct current fault arc occurs in a line, the existing direct current fault arc selection model can adaptively locate the line where the fault arc occurs, thereby effectively preventing electrical accidents caused by fault arcs.

Claims

1. An adaptive fault protection method considering the uncertainty of distribution network equipment access, characterized in that, Includes the following steps: Step 1: Real-time acquisition of system current information under conditions of flexible access and switching of distribution network equipment; Step 2: Perform time-frequency domain analysis on the acquired system current information to extract detection and line selection features, and obtain a feature set; Step 3: Determine the failure status of the detection features. If the detection features are not failed, proceed to Step 4. If the detection features are failed, first optimize the parameters for the current detection feature form, then determine the failure status. If the improved detection features are not failed, combine them with historical features for dimensionality enhancement and proceed to Step 4. If the improved detection and line selection features are still failed, then filter the feature set obtained in Step 2 to obtain fault features, combine the obtained fault features with historical features for dimensionality enhancement, and proceed to Step 4. The criteria for determining the failure status of the detection features are as follows: First, the maximum mutual information coefficient is used to measure the correlation between fault characteristics and historical normal operating system characteristics. The mutual information calculation formula is as follows: In the formula: I (x, y) represents mutual information. p ( x , y )yes x and y The joint probability distribution function, p ( x )and p ( y ) are respectively x and y The marginal probability distribution function is used to determine that the detection feature is correlated and fails when the mutual information is greater than a set threshold. The criteria for determining the failure of the line selection feature are as follows: First, the maximum mutual information coefficient is used to measure the correlation between normal features and the normal operating system features corresponding to different distribution network equipment and branch states in history. When the mutual information is greater than the set threshold, it is determined that the line selection feature is related and fails. Step 4: Extend the fault intelligent detection model by improving continuous learning based on the method of learning without forgetting. Based on the fault intelligent detection model and the detection features obtained in step 2, perform fault state detection to complete the adaptive fault detection. If a fault is detected, proceed to step 5; otherwise, return to step 1. Step 5: Determine the failure status of the route selection features. If the route selection features are not failed, proceed to step 6. If the route selection features are failed, first optimize the parameters for the current route selection feature form, and then determine the failure status. If the improved route selection features are not failed, combine them with historical features for dimensionality improvement and proceed to step 7. If the improved route selection features are still failed, filter the feature set, combine the obtained fault features with historical features for dimensionality improvement and proceed to step 7. Step 6: Determine whether the non-failed line selection features present new distribution network equipment information. If so, proceed to Step 7; otherwise, proceed to Step 8. Step 7: Extend the fault intelligent localization model by improving the category incremental learning method based on forgetting learning; specifically, it includes the following steps: First, add a new node to the output layer of the DC fault arc selection model, then use knowledge distillation and weight factors to retain the data of the old task, so that the prediction value of the new model for the old task is consistent with the prediction value of the old model for the old task, and use cross-entropy loss and weight factors to memorize the new task. Step 8: Based on the artificial intelligence network, perform fault location and output the branch and equipment information where the fault is located. First, use the maximum mutual information coefficient to measure the correlation between the non-failed fault location features described in Step 6 and the normal operating system features corresponding to different distribution network equipment and branch states in history. When the mutual information coefficient reflecting the correlation is greater than a set threshold, it is determined that the non-failed fault location features present relevant historical distribution network equipment information; otherwise, it is determined that the non-failed fault location features present new distribution network equipment information. For multiple fault location features, use the ternary loss method in the following formula to narrow the distance between data of similar equipment and widen the distance between data of dissimilar equipment: In the formula: L Indicates a three-element loss. D ( a , p ) represents equipment sample a With sample b The distance between them D ( a , n ) represents equipment sample a With sample n The distance between them α This represents the margin function.

2. The adaptive fault protection method considering the uncertainty of distribution network equipment access according to claim 1, characterized in that, In step 2, the detection and line selection features are time-frequency domain features.

3. The adaptive fault protection method considering the uncertainty of distribution network equipment access according to claim 1, characterized in that, In step 3, a recurrent neural network is used. For the fault detection model, the recurrent neural network is first trained using the features of the induction cooker based on wavelet packet decomposition, and then a fault intelligent detection model is obtained. In step 7, for the line selection model, the recurrent neural network is trained using the steady-state features of the old scene, and then a fault intelligent localization model is obtained.

4. The adaptive fault protection method considering the uncertainty of distribution network equipment access as described in claim 1, characterized in that, The feature parameter optimization process involves using the feature parameters as constraints without changing the feature framework, and then using an optimization algorithm to solve for the feature parameters to ensure that the feature output does not fail under the current fault state or equipment operating state.

5. The adaptive fault protection method considering the uncertainty of distribution network equipment access according to claim 1, characterized in that, The feature set includes time-domain features, statistical features, time-frequency domain features, and magneto-optical-acoustic physical features. The selection is completed by evaluating and selecting features that do not fail in the current fault state or features that do not fail in the equipment operation state.

6. The adaptive fault protection method considering the uncertainty of distribution network equipment access according to claim 1, characterized in that, When fault features are obtained from the feature set obtained in step 2, the most suitable fault detection features are automatically analyzed based on the system architecture or application scenario.

Images