CN113298124B - Parameter matching-based multi-level direct current fault arc detection method - Google Patents
Parameter matching-based multi-level direct current fault arc detection method Download PDFInfo
- Publication number
- CN113298124B CN113298124B CN202110502113.1A CN202110502113A CN113298124B CN 113298124 B CN113298124 B CN 113298124B CN 202110502113 A CN202110502113 A CN 202110502113A CN 113298124 B CN113298124 B CN 113298124B
- Authority
- CN
- China
- Prior art keywords
- fault arc
- detection
- current
- parameters
- direct current
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
- Y04S10/52—Outage or fault management, e.g. fault detection or location
Abstract
The invention discloses a parameter matching-based multi-level direct current fault arc detection method, which comprises the following steps: acquiring system signals in real time, improving the proportion of fault arc components through signal filtering processing, then carrying out fault arc characteristic analysis and state learning, and taking training time as a core evaluation target in the fault arc state identification training process; testing the trained fault arc state identification model, and taking the comprehensive judgment accuracy, response time, detection sensitivity and parameter optimization time for detecting the occurrence of the direct-current fault arc by using the model as a direct-current fault arc detection performance evaluation target; the particle swarm optimization is adopted to carry out multi-objective optimization, so that multi-link detection parameters in the whole process of direct current fault arc detection achieve global optimal matching and the detection effect is cooperatively improved, and the urgent requirements of reliable detection of various complex fault arcs caused by multi-noise source interference and operation environment difference in different direct current power supply scenes are dynamically adapted.
Description
Technical Field
The invention belongs to the technical field of electrical fault detection, and relates to a direct current fault arc detection method, in particular to a method for performing global optimization matching on detection parameters of a multilevel classification model by a constructed multidimensional fault arc detection performance evaluation target group and applying a particle swarm algorithm, so that the requirements of accurate detection of various complex direct current fault arcs under the conditions of multi-noise source interference and operation environment difference of a dynamic adaptive application scene are met.
Background
The fault arc protection device is one of core equipment for ensuring the safety and reliability of direct current power supply, fundamentally ensures the reliable operation and power utilization safety of a system, and is one of bottlenecks which restrict the rapid development of a direct current system at present. The power system has the characteristics of high-proportion renewable energy sources and high-proportion power electronic equipment by means of the double-carbon target of coping with climate change, meanwhile, direct-current power supply scenes represented by the renewable energy sources and generalized direct-current loads are continuously increased, and once the phenomena of line insulation aging, line insulation damage or connector loosening and the like occur in the direct-current systems, fault arc accidents which harm the running safety of the direct-current systems can be caused. At present, the research on the characteristics of the direct-current fault arc is in a starting stage, and a corresponding direct-current fault arc circuit breaker (DC AFCI) product is not mature yet.
Source load equipment, line parameters, power electronic devices and external environment characteristics in different direct current power supply scenes are different, and different arc characteristics are presented after interaction with a fault arc; meanwhile, the DC power supply scene is more open, and the operation characteristics of the system can be changed to different degrees due to the diversity of the connected components or when the connected components are in different electric service life periods, so that different arc characteristics are formed. Due to the factors, the analysis results of the same characteristic form show larger form difference, so that the existing fault arc detection method is difficult to keep compatible and effective, the fault arc cannot be detected in time, the related equipment and devices of the direct current system are seriously damaged, regional power failure and fire accidents are triggered, and the life and property safety of surrounding residents is threatened. Therefore, research on a fault arc accurate detection algorithm and a multi-parameter dynamic self-adaptive control model thereof is urgently needed, and further the requirement of different direct current power supply scenes on arc accurate detection is met.
The particle swarm optimization algorithm is a heuristic algorithm, the randomness in the iteration process enables the optimization result to jump out of a local optimal point, the memorability enables the optimization result to quickly converge on a global optimal point, and the particle swarm optimization algorithm has the characteristics of simplicity and quickness. However, at present, a method for optimizing and matching multi-link parameters in a multilevel direct current arc fault detection process by using a particle swarm algorithm is not seen, and a multidimensional target for comprehensively evaluating the direct current arc fault detection performance is not seen.
Disclosure of Invention
The invention aims to provide a parameter matching-based multi-level direct current fault arc detection method, which improves the accuracy of direct current fault arc detection and adapts to the efficient matching capability of detection parameters required by a direct current power supply scene.
In order to achieve the purpose, the invention adopts the following technical scheme:
1) sampling current signals of a direct current system under arc-like working conditions and fault arc conditions in the operation of the direct current system to obtain a sample data set of current detection signals, and turning to the step 2) or the step 3);
2) Filtering the current detection signals in the sample data set to improve the proportion of fault arc components, and turning to the step 3);
3) Determining available characteristic quantity and a classifier through fault arc characteristic analysis and system state learning, determining an optimized vector of a particle swarm algorithm by combining detection parameters to be optimized, then initializing the particle swarm, and turning to the step 4);
4) Respectively training each fault arc state identification model corresponding to the current weight value of the optimized vector by using the current detection signal, the filtered current detection signal or feature data obtained by analyzing feature quantity according to one of the two signals, taking training time t when the state identification performance y (t) in the training process reaches 90% -100% (e.g., 95%) of steady state performance y (∞) for the first time as a first-class evaluation target, wherein infinity is training time required by the fault arc state identification performance to reach the steady state, simultaneously carrying out multi-cycle (e.g., 50-100 analysis cycles) analysis on the real-time performance state of the trained fault arc state identification model according to a current detection signal time window, and taking comprehensive judgment accuracy, response time, detection sensitivity and parameter optimization time obtained by analysis as a second-class evaluation target, and turning to step 5);
5) In the one-time iteration process of the particle swarm algorithm, updating the individual and the population extreme value after evaluating the particle fitness, and updating the speed and the position of each particle according to the individual and the population extreme value, wherein the particle fitness is determined according to various evaluation targets obtained in the step 4), and turning to the step 6);
6) If the iteration process termination condition is met, outputting the weight value of each optimized vector under the optimal evaluation target, determining the optimal classifier, the feature quantity matching result and the optimal detection parameter, and turning to the step 7), otherwise, turning to the step 4);
7) And determining optimal detection parameters according to the weight values of the optimized vectors, and then identifying the system state according to the real-time sampling result of the current signal, wherein the system state is selected from any one of normal interference and fault arc.
Preferably, in the step 1), the sample data set under the fault arc condition is obtained by sampling the fault arc current under one or more of the following conditions: interference of a power electronic device, pure resistive load operation, remote transmission and two branch interference when a multi-branch power supply direct current system is adopted; the sample data set under the arc-like working condition is obtained by sampling one or more of the following normal interference currents under the source load running condition: normal running state, transient state, and shutdown state.
Preferably, the step 1) further comprises the following steps: and calculating a data occupation ratio of 10-20 ms fault arc and the Rbio3.1 wavelet coefficient of the current signal under the normal operation state condition from [ L, S ] to [ L, E ] less than 1.2 according to a sample data set, wherein L represents the number of layers of wavelet packet decomposition, L = 3-5, S represents a selected starting node number, S =1, 2, E represents a selected termination node number, E =5, 6, and if the occupation ratio is more than 70%, judging that the signal-to-noise ratio of the detection signal of the fault arc current is low, and performing signal filtering processing on all current detection signals in the sample data set at the moment.
Preferably, the detection parameters to be optimized relate to parameters of a plurality of links in the following direct-current fault arc detection process: sampling, filtering, characteristic quantity analysis, system state classification and fault arc judgment; the parameters of the sampling link comprise current sampling frequency, the parameters of the filtering link comprise filtering frequency band parameters, the parameters of the characteristic quantity analysis link comprise time-frequency characteristic resolution and characteristic transformation necessary adjustment parameters, the parameters of the system state classification link comprise classifier necessary adjustment parameters, the parameters of the fault arc judgment link comprise direct current fault arc criteria, and the direct current fault arc criteria are threshold values of the times of the model needing to output the value corresponding to the fault arc when the fault arc is judged to occur by using the trained fault arc state identification model.
Preferably, the position vector is set to X when initializing the particle group i ＝(x i1 ,x i2 ,···,x ikM ，x j1 ，···，x jN ) T Velocity vector set to V i ＝(v i1 ，v i2 ,···,v ikM ，v j1 ,···,v jN ) T The first kM weight in the position vector represents the pairwise combination probability of k classifiers available for use and M characteristic quantities, wherein k is larger than or equal to 1 and M is larger than or equal to 0, the last N weights in the position vector represent the regulation proportion of detection parameters to be optimized, the detection parameters to be optimized are sensitive parameters which influence the fault arc identification performance within a 30% change range in respective parameter value ranges, and the fault arc identification performance is determined according to the fault arc detection accuracy.
Preferably, the step 4) further comprises the following steps: when the detection process has both links of characteristic quantity analysis and system state classification, namely M is not equal to 0, matching characteristic quantities for a classifier by using the position vectors of the initial particle swarm or the particle swarm after each iteration, specifically by setting a screening threshold (for example, 0.5-0.8) and neglecting X i The weight of the preceding kM item is smaller than the threshold value, so that the characteristic quantity matched with the classifier is determined according to the combination relation between the classifier corresponding to the other items and the characteristic quantity, and then a fault arc state identification model is trained; at the same time, according to X i The last N items of the data are set with the parameter values of corresponding links in sampling, filtering, characteristic quantity analysis, system state classification and fault arc judgment in the direct current fault arc detection process.
Preferably, the organization architecture of the classifier in the fault arc state identification model adopts a Stacking model, and the first-layer base classifier is composed of a machine learning classifier and/or a data-driven classifier, wherein the machine learning classifier uses characteristic data as input for training, and the data-driven classifier uses current detection signals or characteristic data as input for training; and when the number of the first-layer classifiers is more than 1, establishing a logistic regression classifier of the second layer, and training by using a state identification result output by the first-layer base classifier as input.
Preferably, in the training process of the arc fault state identification model, the current detection signal or the filtered current detection signal input to the base learner of the first layer or the feature data obtained by performing feature quantity analysis according to one of the two signals is normalized, and the normalization process is to divide the current detection signal or the feature data in a certain time window by the average value of the current detection signal or the feature data of each time window in the normal operation state of the sample data set.
Preferably, when the particle swarm algorithm evaluates the particle fitness, the training time, the comprehensive judgment accuracy, the response time, the detection sensitivity and the parameter optimization time are selected, and the fitness of each particle is obtained by solving according to an analytic hierarchy process.
Preferably, in the particle swarm optimization, the iteration process terminating condition is that the number of iterations reaches 500 to 700 (e.g., 500), or the iteration process terminating condition is that the numerical value variation of the optimal value of the fitness of the particles in the swarm is kept within 2% in the latest 150 to 200 (e.g., 150) iterations.
The invention has the beneficial effects that:
the invention combines the training of the direct current fault arc detection model, the matching of detection parameters, the establishment of multiple evaluation targets of detection performance and the iterative process of the particle swarm algorithm, efficiently matches the detection parameters suitable for the detection algorithm aiming at the characteristics of the direct current power supply scene, and can quickly respond to the differentiated detection requirements of branch access and equipment aging change arc characteristics in the emerging open direct current scene. The evaluation system selected by the invention comprises comprehensive judgment accuracy, response time, detection sensitivity and parameter optimization time, and multiple evaluation targets are beneficial to realizing the overall optimization of multiple parameters of the detection framework, so that the detection parameters under the condition of a given evaluation target scheme are comprehensively considered, and the method is more reasonable and is pertinently suitable for the DC power supply scene to be popularized and applied. The invention has higher commercial value and wide market application.
The invention further obtains the technical effects that:
1) The fault arc sample data set considered by the invention relates to the situations of power electronic device interference, pure resistive load operation, remote transmission interference and multi-power-supply branch interference, and simultaneously considers the algorithm operation condition of an arc detection algorithm under the normal operation condition of source load, thereby being beneficial to improving the fault arc detection reliability.
2) The invention divides the detection process of the direct current fault arc into a plurality of links such as sampling, filtering, characteristic quantity analysis, system state classification and the like, can form a plurality of alternative schemes from the angle of combination of detection parameters, characteristics and classification, can not only reduce the misjudgment rate by utilizing a plurality of characteristics and a plurality of classifiers, but also adaptively match the optimal detection parameters aiming at the characteristics of the considered direct current power supply scene, more effectively form an optimal detection framework, realize the optimal fault arc cooperative detection effect and greatly improve the reliability and rapidity of the fault arc detection.
3) The optimization method adopted by the invention is a particle swarm algorithm, and the particle fitness is calculated by introducing a chromatographic analysis method, so that the optimization result can be rapidly obtained while the consumption of calculation resources in the optimization process of the detection parameters is reduced, and the high efficiency of parameter optimization is ensured.
4) The invention fully considers the multi-party requirements of the realized hardware resource and application scene detection, the characteristic quantity and the number of the classifiers can set the upper limit according to the actual situation, simultaneously, different classifiers are matched with the optimal characteristic quantity group, unnecessary secondary characteristics are abandoned, the time spent by training and mastering the fault arc rule by each classifier is greatly reduced, the characteristic quantity group used by each classifier is in the set output by the characteristic layer, the effective utilization degree of each characteristic by the characteristic layer and the state layer is enhanced, and the optimal adaptation of the existing hardware resource to the arc detection direct current power supply scene is realized.
Drawings
Fig. 1 is a flowchart of a method for accurately detecting a multi-level dc fault arc according to an embodiment of the present invention.
Fig. 2 is a flowchart of a dc fault arc detection parameter efficient matching method in an embodiment of the present invention.
Fig. 3a shows an input current signal for accurate detection of dc fault arcs and efficient matching of parameters.
Fig. 3b shows the output result of the normalization process on the input current signal.
Fig. 4a shows an input power signal (where a voltage signal is used to assist in determining the system state) for accurate detection of dc fault arcs and efficient matching of parameters.
Fig. 4b is an output result of the feature extraction processing performed on the input electric quantity signal.
FIG. 5a is the output result of system state identification under the condition of insufficient training time.
FIG. 5b is the output result of the system state identification under the condition of sufficient training time.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples. The examples are given solely for the purpose of illustration and are not intended to limit the scope of the invention.
(I) accurate detection of multi-level DC fault arc
The process of the accurate detection of the multilayer direct current fault arc adopted by the invention is shown in figure 1, and comprises the following specific steps:
step one, during the operation of a direct current system, sampling current signals of the direct current system under the conditions of normal interference and fault arc by fault arc detection hardware at a frequency f to form a detection signal data set, and turning to step two;
the data set considers current signals under four fault arc conditions, namely power electronic device interference, pure resistive load operation, fault arc current under a long-distance transmission condition and fault arc current under two branch interference condition when a multi-branch power supply direct current system is adopted, and the current signals under a normal interference condition specifically consider normal interference current signals under a source load operation condition.
Step two, processing the data set by applying a signal filtering method, improving the proportion of the fault arc components, and turning to step three;
the signal filtering link is used when the signal-to-noise ratio of the fault arc is low; the method for judging the low signal-to-noise ratio of the fault arc comprises the following steps: the ratio of the fault arc to the Rbio3.1 wavelet coefficient of the normal operation current signal from [4,1] to [4,5] is less than 1.2 (the data is a two-dimensional concept and comprises not only the node but also the influence of the time window, for example, 5 nodes multiplied by 100 time windows, totally 500 data) of more than 70%.
Step three, carrying out characteristic quantity analysis on the obtained current signal directly or carrying out characteristic quantity analysis after time-frequency transformation processing, then carrying out a learning training process of system state identification aiming at a direct current fault arc power supply scene to be applied, and turning to step four;
the classification model of the system state identification is used for carrying out organization architecture design of the classifier according to a Stacking model, and the first layer is composed of a machine learning classifier and a data driving classifier, wherein the machine learning classifier uses characteristic data as input for training, and the data driving classifier uses current signals or characteristic data as input for training; and when the number of the first-layer base classifiers is more than 1, establishing a logistic regression classifier of the second layer, and applying a state identification result output by the first-layer base classifier as the input of the logistic regression classifier and training.
The current signal or the characteristic data input to the first layer of base learner needs to be standardized; the standardization processing method comprises the following steps: the current signal or signature data within an arbitrary time window (analysis period) is divided by the average of the normal operating state current signals or signature data within the data set.
The labeling of the packing classification model in the learning and training process is as follows: when the arc voltage signal in the corresponding period is zero (arc-like interference or normal interference: normal operation, transient state and shutdown state), it is marked as 0, and when the arc voltage in the corresponding period is non-zero (fault arc state), it is marked as 1.
Step four, performing real-time system state identification by using the classification model after learning training, if the analysis period is considered to present a fault arc state, outputting a value of 1, continuing to perform system state identification of a plurality of analysis periods, and sending an action signal to cut off the fault arc when triggering a direct-current fault arc judgment standard; otherwise, judging the arc interference, clearing the counting variable, and analyzing the detection signal in the next time window.
Efficient matching of (II) DC fault arc detection parameters
The process of efficient matching of direct-current fault arc detection parameters adopted by the invention is shown in fig. 2, and comprises the following specific steps:
the detection parameters to be optimized can be selected from current sampling frequency, filtering parameters, time-frequency characteristic resolution, characteristic transformation and classifier key parameters, a combination form of characteristics and a classifier and the like according to a detection link. And determining an optimization vector according to the given characteristic quantity and the upper limit of the number of the classifiers and the detection parameters to be optimized, and randomly initializing the position and the speed of the particle swarm.
Upon initialization of particle swarm parameters, the position vector is set to X i ＝(x i1 ，x i2 ，···，x ikM ，x j1 ，···，x jN ) T Velocity vector is V i ＝(v i1 ，v i2 ,···,v ikM ,v j1 ,···,v jN ) T The front kM item corresponds to M characteristic quantities which can be used by k classifiers, wherein k is more than or equal to 1, M is more than or equal to 0, and the rear N item corresponds to sensitive parameters which influence the detection accuracy rate to reach a 30% change range in the value range of each parameter in detection links such as hardware sampling, filtering, characteristic analysis, system state classification and the like.
In order to obtain the optimal direct current fault arc detection effect, a plurality of evaluation targets are set in the optimization process of the particle swarm optimization. For the fault arc state identification training process of the third step in (a), a training time t of 95% is set as a core estimation target, where the state identification performance y (t) first reaches a steady-state performance y (∞) (∞ indicates a sufficiently long training time required for the fault arc state identification performance to change by not more than 2%). And (3) identifying and judging the fault arc state in the step four in the step (I) in real time, taking comprehensive judgment accuracy, response time, detection sensitivity and parameter optimization time as evaluation targets, and forming the plurality of evaluation targets together with the core evaluation target.
The comprehensive judgment accuracy is as follows: and calculating the identification accuracy of the fault arc according to the sample ratio of the fault arc to the normal interference signal in the data set as 1.
The response time is as follows: calculated as the mean value of the fault arc detection time.
The detection sensitivity is as follows: and (3) obtaining continuous 5-time waveforms of four types of fault arcs in the step (I) by applying simulation and experimental means, and calculating according to the detection rate of the fault arcs (generally, one waveform relates to a plurality of analysis periods).
The parameter optimization time is as follows: the particle swarm algorithm is the time from initialization to the end condition.
When evaluating the particle fitness, the particle swarm algorithm selects training time, comprehensive judgment accuracy, response time, detection sensitivity and parameter optimization time to carry out multi-target evaluation on the detection effect, and the specific mode is as follows: the multiple evaluation targets are subjected to standardization processing according to historical data (variation ranges formed by evaluation target values in the processes from initial iteration to final iteration of a historical scene and in the process of 50 times of current scene iteration), then importance is judged by comparing every two evaluation targets, a multi-target weight matrix is calculated according to an analytic hierarchy process, and the optimal solution (particle fitness) under multiple targets is obtained through the comprehensive evaluation system.
The particle swarm algorithm calculates the fitness of each particle by combining a position vector obtained by each iteration with multiple evaluation targets, evaluates the fitness of each particle obtained by each iteration, updates an individual and a population extreme value, updates the speed and position vector of each particle according to the individual and the population extreme value, checks whether the population meets a termination condition, and outputs an optimal detection parameter under an optimal evaluation target scheme if the population meets the termination condition; otherwise, continuing to carry out iterative solution. The terminating condition of the multi-objective optimization solution of the particle swarm optimization is as follows: the iteration number reaches the upper limit of 500 times, or the optimal point of the population keeps within 2% of the change in the last 150 iterations.
When the detection algorithm has both a feature quantity and a classifier link (M ≠ 0), the position vector of the particle swarm is a classifier matching feature quantity: set the screening threshold and discard X i And obtaining the characteristic quantity combination matched and used by each classifier by the combination with the weight smaller than the threshold value in the front kM item. When optimizing the parameters, according to X i The last N items are arranged in multiple links of hardware sampling, filtering, characteristic quantity analysis, system state classification and the likeThe parameters can be adjusted, and the global optimal matching cooperative detection effect is achieved. Therefore, the provided efficient matching method for the direct-current fault arc detection parameters outputs the optimal detection parameter combination, and the comprehensive evaluation effect of multiple evaluation targets is optimal, so that the direct-current fault arc detection method achieves the overall optimal detection performance and dynamically adapts to the direct-current fault arc characteristic difference caused by multiple noise source interferences and operation environment factors in different direct-current power supply scenes.
(III) accurate detection of direct-current fault arc and current data efficiently matched with parameters
Such as the dc fault arc current signal shown in fig. 3 a. Before 3.25s, the direct current system is in a normal working state, and the current waveform still has a normal electric quantity interference disturbance phenomenon caused by system instability; when a 3.25s system has a fault arc, a current signal has a large amplitude pulse and shows a phenomenon of amplitude reduction, the unstable condition is more obvious, and the current signal is always in a fault arc state; 7.17s, the electric arc is extinguished by disconnecting the loop, the direct current system is in a shutdown state, and the corresponding system state is recovered to be normal.
For data-driven such base classifiers, system state recognition can be performed with direct input of current data. However, because the current levels under different experimental conditions are different, the obtained fault arc current also shows difference, so that the current data needs to be standardized according to the method of the invention before deep learning is carried out. The results obtained after normalization of the data using the min-max normalization of the variants are shown in FIG. 3 b. When a large pulse exists in the current at the moment of occurrence of the fault arc, the maximum value of data appears, and if the standardization processing is directly carried out, the judgment of the fault arc is greatly influenced. Therefore, the current average value of the 2s normal time period before the whole data is selected, so that the fault arc current standardized data in different time periods is limited to be between 0 and 1, and errors caused by accidental pulses are avoided.
(IV) characteristic data of accurate detection and efficient parameter matching of direct-current fault arc
Such as the dc fault arc power signal shown in fig. 4 a. Before 2.56s, the direct current system is in a normal working state, the current waveform still has a normal electric quantity interference disturbance phenomenon caused by system instability, an electric arc does not occur, and the corresponding electric arc voltage is 0;2.56s system fault electric arc, electric quantity signal all appear large amplitude pulse, electric arc current presents the reduction of amplitude, electric arc voltage presents dozens of V numerical value because of existence of the arc process, the unstable situation of the two is more obvious at the same time, after that, always in the fault electric arc state; 6.61s extinguishes the arc by breaking the loop, the direct current system is in an open circuit state, the corresponding arc current drops to 0 suddenly, and the system state returns to normal.
For machine learning such base classifiers, most unsupervised, semi-supervised, and supervised learning classifiers commonly use feature data for system state recognition. And performing time-frequency transformation analysis on the current data by adopting short-time Fourier transformation, and then performing calculation and solving by using the characteristic quantity to obtain a result as shown in FIG. 4 b. It can be seen that the characteristic quantities are valid overall, i.e. at the normal operating state before 2.56s and at the shutdown state after 6.61s, the resulting characteristic amplitudes are at a lower level; in the fault arc stage, the output characteristic amplitude is in a larger level on the whole, although the amplitude fluctuates in a certain range due to the instability of the arc, the whole fault arc characteristic value is still separated from the normal state characteristic value obviously, so that the machine learning classifier can master the difference between the fault arc state and various normal interference states through a training process, the fault arc state and the normal interference states are effectively distinguished, and the accurate detection of the fault arc is realized.
(V) influence of efficient matching of detection parameters on detection effect of direct-current fault arc
By using the current data of fig. 3 or the characteristic data of fig. 4, the accurate detection of the dc fault arc can be realized according to the detection framework of the present invention. The current data of fig. 3b is used for fault arc detection, and the importance of the detection parameters to the detection result is contrasted and analyzed, so that the importance of the efficient parameter matching method provided by the invention is reflected.
Applying 10010 current data samples, respectively marking the direct current fault arc state and the normal interference state as 1 and 0, respectively constructing a training set and a testing set, and inputting the training set into the long-term and short-term memory network for learning. Referring to fig. 5a, compared with the system states (actual values, i.e. the normal operation state is 0, the fault arc state is 1, and the shutdown state is 0) that should be output corresponding to each stage of the current signal of fig. 3, when the method of the present invention is not used for parameter matching and only training is performed for a short time, the state identification result is an identification error in the normal operation state, the normal operation current is identified as the fault arc current, and the state identification results of the other two subsequent stages are matched with the actual values. It can be seen that the ideal fault arc detection accuracy cannot be realized by simply pursuing the optimization of a single target of training time, and the classifier may generate overfitting by increasing the training time, so that the corresponding accuracy cannot be increased any more. Therefore, it is necessary to explore the globally optimal training time and fault arc detection accuracy combination.
As shown in fig. 5b, after the efficient matching method for detection parameters provided by the present invention is applied, the prediction result obtained by the long-term and short-term memory network is significantly improved in the optimal training time, and completely conforms to the actual result, and the current of the corresponding detection algorithm in the normal working state and the shutdown state can give the correct low level indication, and the arc current signal in the fault state can give the correct high level indication.
According to the multi-level direct-current fault arc accurate detection method, the long-term and short-term memory network can input the current data set shown in the figure 3, the constructed fault arc characteristics can be used as classifier input to identify the system state, the analysis frequency of input data can be effectively reduced in the characteristic quantity calculation process, and the state training and learning of the classifier are easier. Meanwhile, when the data-driven classifier adopts data input training analysis, the efficient matching problem of classification parameters also exists, and the application of the efficient matching method for detecting parameters can effectively shorten the optimization time of a plurality of long-short term memory network parameters and obviously improve the system state identification effect.
Claims (9)
1. A multi-level direct current fault arc detection method based on parameter matching is characterized in that: the method comprises the following steps:
1) sampling current signals of a direct current system under arc-like working conditions and fault arc conditions in the operation of the direct current system to obtain a sample data set of current detection signals, and turning to the step 2) or the step 3);
2) Filtering the current detection signal in the sample data set, improving the proportion of fault arc components, and turning to the step 3);
3) Determining available characteristic quantity and a classifier through fault arc characteristic analysis and system state learning, determining an optimized vector of a particle swarm algorithm by combining detection parameters to be optimized, then initializing the particle swarm, and turning to the step 4);
4) Respectively training each fault arc state identification model corresponding to the current weight value of the optimized vector by using the current detection signal, the filtered current detection signal or feature data obtained by analyzing feature quantity according to one of the two signals, taking training time t when state identification performance y (t) in the training process reaches 90% -100% of steady state performance y (∞) for the first time as a first-class evaluation target, wherein infinity is training time required by the fault arc state identification performance reaching the steady state, simultaneously carrying out multi-cycle analysis on the real-time performance state of the trained fault arc state identification model according to a current detection signal time window, and taking comprehensive judgment accuracy, response time, detection sensitivity and parameter optimization time obtained by analysis as a second-class evaluation target, and turning to step 5);
5) In the one-time iteration process of the particle swarm algorithm, updating individual and population extreme values after evaluating the fitness of the particles, updating the speed and the position of each particle according to the individual and the population extreme values, wherein the fitness of the particles is determined according to various evaluation targets obtained in the step 4), and turning to the step 6);
6) If the iteration process termination condition is met, outputting the weight value of each optimized vector under the optimal evaluation target, determining the optimal classifier, the feature quantity matching result and the optimal detection parameter, and turning to the step 7), otherwise, turning to the step 4);
7) Determining optimal detection parameters according to the weight values of the optimized vectors, and then identifying the system state according to the real-time sampling result of the current signal, wherein the system state is selected from any one of normal interference and fault arc;
the step 1) further comprises the following steps: and calculating a data occupation ratio of 10-20 ms fault arc and the Rbio3.1 wavelet coefficient of the current signal under the normal operation state condition from [ L, S ] to [ L, E ] less than 1.2 according to a sample data set, wherein L represents the number of layers of wavelet packet decomposition, L = 3-5, S represents a selected starting node number, S =1, 2, E represents a selected termination node number, E =5, 6, and if the occupation ratio is more than 70%, judging that the signal-to-noise ratio of the detection signal of the fault arc current is low, and performing signal filtering processing on all current detection signals in the sample data set at the moment.
2. The method for detecting the multi-level direct current fault arc based on the parameter matching as claimed in claim 1, wherein: in the step 1), the sample data set under the fault arc condition is obtained by sampling the fault arc current under one or more of the following conditions: interference of a power electronic device, pure resistive load operation, remote transmission and two-branch interference during a multi-branch power supply direct current system; the sample data set under the arc-like working condition is obtained by sampling the following normal interference currents under the source load running condition: normal running state, transient state, and shutdown state.
3. The method for detecting the multi-level direct current fault arc based on the parameter matching as claimed in claim 1, wherein: the detection parameters to be optimized relate to the parameters of a plurality of links in the following direct current fault arc detection process: sampling, filtering, characteristic quantity analysis, system state classification and fault arc judgment; the parameters of the sampling link comprise current sampling frequency, the parameters of the filtering link comprise filtering frequency band parameters, the parameters of the characteristic quantity analysis link comprise time-frequency characteristic resolution and characteristic transformation necessary adjustment parameters, the parameters of the system state classification link comprise classifier necessary adjustment parameters, the parameters of the fault arc judgment link comprise direct current fault arc criteria, and the direct current fault arc criteria are threshold values of the times of the model needing to output the value corresponding to the fault arc when the fault arc is judged to occur by using the trained fault arc state identification model.
4. The method according to claim 3, wherein the method comprises the following steps: when initializing the particle group, the position vector is set to X i ＝(x i1 ,x i2 ,···,x ikM ,x j1 ,···,x jN ) T Velocity vector set to V i ＝(v i1 ,v i2 ,···,v ikM ,v j1 ,···,v jN ) T The first kM weight in the position vector represents the pairwise combination probability of k classifiers available for use and M characteristic quantities, wherein k is larger than or equal to 1 and M is larger than or equal to 0, the last N weights in the position vector represent the regulation proportion of detection parameters to be optimized, the detection parameters to be optimized are sensitive parameters which influence the fault arc identification performance within a 30% change range in respective parameter value ranges, and the fault arc identification performance is determined according to the fault arc detection accuracy.
5. The method according to claim 4, wherein the method comprises the following steps: the step 4) further comprises the following steps: when the detection process has both links of characteristic quantity analysis and system state classification, namely M is not equal to 0, matching characteristic quantities for a classifier by using the position vectors of the initial particle swarm or the particle swarm after each iteration, specifically setting a screening threshold and neglecting X i The weight of the preceding kM item is smaller than the threshold value, so that the characteristic quantity matched with the classifier is determined according to the combination relation between the classifier corresponding to the other items and the characteristic quantity, and then a fault arc state identification model is trained; at the same time, according to X i The last N items set parameters of corresponding links in sampling, filtering, characteristic quantity analysis, system state classification and fault arc judgment of the direct current fault arc detection processThe value is obtained.
6. The method for detecting the multi-level direct current fault arc based on the parameter matching as claimed in claim 1, wherein: the organization structure of the classifier in the fault arc state identification model adopts a Stacking model, and the first-layer base classifier is composed of a machine learning classifier and/or a data-driven classifier, wherein the machine learning classifier uses characteristic data as input for training, and the data-driven classifier uses current detection signals or characteristic data as input for training; and when the number of the first-layer classifiers is more than 1, establishing a logistic regression classifier of the second layer, and training by using a state identification result output by the first-layer base classifier as input.
7. The multi-level direct current fault arc detection method based on parameter matching according to claim 5, characterized in that: in the training process of the fault arc state identification model, the current detection signal input to the base learner of the first layer or the filtered current detection signal or the characteristic data obtained by performing characteristic quantity analysis according to one of the two signals are subjected to standardization processing, wherein the standardization processing refers to dividing the current detection signal or the characteristic data in a certain time window by the average value of the current detection signal or the characteristic data of each time window in the normal operation state of the sample data set.
8. The method for detecting the multi-level direct current fault arc based on the parameter matching as claimed in claim 1, wherein: when the particle swarm algorithm evaluates the particle fitness, the training time, the comprehensive judgment accuracy, the response time, the detection sensitivity and the parameter optimization time are selected, and the fitness of each particle is obtained by solving according to an analytic hierarchy process.
9. The method for detecting the multi-level direct current fault arc based on the parameter matching as claimed in claim 1, wherein: in the particle swarm optimization, the iteration process termination condition is that the iteration frequency reaches 500-700 times, or the iteration process termination condition is that the change of the numerical value of the optimal value of the particle fitness in the population in the latest 150-200 times of iteration process is kept within 2%.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110502113.1A CN113298124B (en) | 2021-05-08 | 2021-05-08 | Parameter matching-based multi-level direct current fault arc detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110502113.1A CN113298124B (en) | 2021-05-08 | 2021-05-08 | Parameter matching-based multi-level direct current fault arc detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113298124A CN113298124A (en) | 2021-08-24 |
| CN113298124B true CN113298124B (en) | 2023-03-24 |
Family
ID=77321166
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110502113.1A Active CN113298124B (en) | 2021-05-08 | 2021-05-08 | Parameter matching-based multi-level direct current fault arc detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113298124B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113866572A (en) * | 2021-09-06 | 2021-12-31 | 西安理工大学 | Direct-current fault arc detection and positioning method under condition of access of multiple power electronic devices |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109560770A (en) * | 2018-10-29 | 2019-04-02 | 西安交通大学 | Photovoltaic system fault arc detection method based on self-adaptive kernel function and instantaneous Frequency Estimation |
| CN110398669A (en) * | 2019-06-11 | 2019-11-01 | 深圳供电局有限公司 | Method for detecting arc |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7242291B2 (en) * | 2005-06-08 | 2007-07-10 | Sensata Technologies, Inc. | Arc fault circuit interrupter test circuit and test board |
| CN112433128B (en) * | 2020-10-13 | 2022-12-02 | 西安理工大学 | Stochastic resonance-based direct-current fault arc diagnosis method and device |
| CN112526243B (en) * | 2020-12-28 | 2021-10-26 | 南京航空航天大学 | Sectional type direct current arc noise model, parameter optimization and identification method |
-
2021
- 2021-05-08 CN CN202110502113.1A patent/CN113298124B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109560770A (en) * | 2018-10-29 | 2019-04-02 | 西安交通大学 | Photovoltaic system fault arc detection method based on self-adaptive kernel function and instantaneous Frequency Estimation |
| CN110398669A (en) * | 2019-06-11 | 2019-11-01 | 深圳供电局有限公司 | Method for detecting arc |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113298124A (en) | 2021-08-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11489490B2 (en) | Arc fault detection method for photovoltaic system based on adaptive kernel function and instantaneous frequency estimation | |
| CN109787979B (en) | Method for detecting electric power network event and invasion | |
| CN110879917A (en) | Electric power system transient stability self-adaptive evaluation method based on transfer learning | |
| CN111585948B (en) | Intelligent network security situation prediction method based on power grid big data | |
| CN110702966B (en) | Fault arc detection method, device and system based on probabilistic neural network | |
| CN111598179B (en) | Power monitoring system user abnormal behavior analysis method, storage medium and equipment | |
| CN112149609A (en) | Black box anti-sample attack method for electric energy quality signal neural network classification model | |
| CN112560596B (en) | Radar interference category identification method and system | |
| CN113298124B (en) | Parameter matching-based multi-level direct current fault arc detection method | |
| CN113935440A (en) | Iterative evaluation method and system for error state of voltage transformer | |
| CN114785573A (en) | Intelligent substation process layer network abnormal flow detection method based on deep learning | |
| CN110852441B (en) | Fire disaster early warning method based on improved naive Bayes algorithm | |
| CN113297786B (en) | Low-voltage fault arc sensing method based on semi-supervised machine learning | |
| Hu et al. | Robust feature extraction and ensemble classification against cyber-physical attacks in the smart grid | |
| CN108508297A (en) | A kind of fault arc detection method based on change coefficient and SVM | |
| CN114386024A (en) | Power intranet terminal equipment abnormal attack detection method based on ensemble learning | |
| CN111062569A (en) | Low-current fault discrimination method based on BP neural network | |
| Jiang et al. | RBF-based real-time hierarchical intrusion detection systems | |
| CN111898446A (en) | Single-phase earth fault studying and judging method based on multi-algorithm normalization analysis | |
| Mei et al. | Random forest based adaptive non-intrusive load identification | |
| Chen et al. | A data preparation method for machine-learning-based power system cyber-attack detection | |
| CN114113868A (en) | Transformer fault diagnosis method | |
| CN113884807A (en) | Power distribution network fault prediction method based on random forest and multi-layer architecture clustering | |
| CN112348159A (en) | Fault arc neural network optimization training method | |
| KR20220025342A (en) | Arc Signal Detection Method Using Logistic Regression |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |