CN116205265A - Power grid fault diagnosis method and device based on deep neural network - Google Patents
Power grid fault diagnosis method and device based on deep neural network Download PDFInfo
- Publication number
- CN116205265A CN116205265A CN202310041168.6A CN202310041168A CN116205265A CN 116205265 A CN116205265 A CN 116205265A CN 202310041168 A CN202310041168 A CN 202310041168A CN 116205265 A CN116205265 A CN 116205265A
- Authority
- CN
- China
- Prior art keywords
- neural network
- sample
- fault
- deep neural
- power grid
- 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.)
- Pending
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
- G01R31/08—Locating faults in cables, transmission lines, or networks
- G01R31/088—Aspects of digital computing
-
- 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
- G01R31/08—Locating faults in cables, transmission lines, or networks
- G01R31/081—Locating faults in cables, transmission lines, or networks according to type of conductors
- G01R31/086—Locating faults in cables, transmission lines, or networks according to type of conductors in power transmission or distribution networks, i.e. with interconnected conductors
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a power grid fault diagnosis method and device based on a deep neural network, wherein the method comprises the steps of obtaining power grid fault alarm information and a network topology structure; constructing a neural network sample according to the power grid fault alarm information and the network topology structure, wherein the neural network sample comprises a neural network learning sample and a neural network testing sample; preprocessing a neural network sample to obtain a fault feature vector; inputting fault feature vectors of a neural network learning sample into a deep neural network model which is constructed in advance and optimized based on a chaotic particle swarm algorithm for learning, and obtaining an optimized and learned deep neural network model; the invention establishes a deep neural network model for the sample and carries out learning training on the deep neural network by using a chaotic particle swarm algorithm, thereby having higher diagnosis accuracy.
Description
Technical Field
The invention relates to a power grid fault diagnosis method and device based on a deep neural network, and belongs to the technical field of intelligent power grid fault diagnosis.
Background
The power grid equipment is affected by aging, natural disasters and the like, faults are inevitably generated in the running process of the power grid, and fault elements are timely and accurately diagnosed, so that the normal running of a power system is facilitated to be recovered. In this case, it is highly necessary to resort to advanced diagnostic techniques. When the power equipment breaks down, the dispatching center can receive a large number of signals at the same time, and how to quickly identify the broken down equipment from the large number of signals, namely the fault diagnosis, has important significance for ensuring and improving the operation safety of the system and becomes standard software of the power dispatching center. To date, the main methods of power system fault diagnosis include Expert Systems (ESs), analytical model methods, artificial neural networks (artificial neural networks, ANNs), petri Networks (PNs), bayesian Networks (BNs), multi-agent systems (MAS), genetic algorithms (genetic algorithms, GA), and optimization algorithm-based methods. The method is the most widely applied method in the fault diagnosis of the actual power system because of the relatively simple principle and stable operation. The expert system, the fuzzy system and the multi-agent technology are combined to be applied to fault diagnosis, various possible faults of the transformer are comprehensively diagnosed by adopting various data sources, and the problem that the method needs to be overcome is how to effectively acquire expert experience knowledge and how to enable an inference mechanism to have expansibility. The analysis model method builds an ideal and actual difference minimization optimization model of various candidate fault devices and corresponding information states in the suspicious fault region, and the most probable fault element is obtained through the optimization method. The method has better theoretical support, but the method can often obtain multiple solutions, and the method still needs to be further overcome and improved. The Petri net graphically represents the logic relationship of a system formed by the protected elements, the switches, the protection and the like, and analyzes the fault occurrence process of the equipment from the viewpoint of the mutual logic relationship of the equipment. Combining the Petri network with fuzzy information, establishing a multi-factor grading model based on the Petri network considering protection and possible false operation rejection of the circuit breaker, and enhancing the fault tolerance of the model by considering time constraint. This type of method requires more sufficient prior probability data of equipment failure, which is often more difficult to obtain, and accordingly limits its application.
The neural network method takes the switch and fault information as the input of the network, and takes the fault information of the corresponding equipment as the output of the network, and has the characteristics of simple and quick online operation and strong fault tolerance. The neural network fault diagnosis method about 2000 and before is limited by the computing capability of a computer and the development of a neural network theory, and a shallow neural network model is generally adopted, so that good effects are obtained when the dimension of an input layer is not high (a single hidden layer is often adopted) and the number of samples is not large. And a diagnosis model based on a shallow neural network, a radial basis neural network and a generalized regression neural network of data mining is established. In this case, the neural network method often has a small training sample size, so that the capability of extrapolation is lacking, and the training time is long. With the rapid development of computer hardware and deep learning theory and algorithm in recent years, the fault diagnosis method based on the deep neural network becomes a method which is most hopeful to improve the existing fault diagnosis problem and improve the fault diagnosis level to a new height. The depth self-coding network is applied to motor fault diagnosis by a learner, fault data of a specific power grid in the last 10 years is used as a training set, and training and learning of the constructed cyclic neural network model are achieved. In fact, the shortages of the sample number and the dispersion of the quality are a great obstacle to the application of the deep learning neural network method to engineering problems, and a learner adopts an unsupervised pre-training mode to improve the problem, but the existing method still does not fundamentally solve the problem.
According to the investigation of a plurality of power company dispatching centers, the above methods are applied to practical engineering in a large number of fault diagnosis methods based on expert systems, and have high practical fault identification rate and low practical fault identification rate
Is not reported. The main problem of the current puzzling power company is that the false alarm rate of the method is high. The artificial neural network with strong fault tolerance and learning ability is widely applied to various fault diagnosis fields, and the power grid fault diagnosis field is no exception. Among many neural networks, deep neural networks are used as deep learning neural networks, and are applied in many fields, and the field of power grid fault diagnosis is not exceptional. And when the deep neural network is used for learning and training, the parameter determination of the deep neural network has great influence on the accuracy of fault diagnosis of the whole power grid.
Disclosure of Invention
The invention aims to overcome the defects in the prior art, and provides a power grid fault diagnosis method and device based on a deep neural network, so as to solve the problems of lower fault diagnosis accuracy, higher false alarm rate and the like in the method based on an expert system and the like in the prior art.
In order to achieve the above purpose, the invention is realized by adopting the following technical scheme:
in a first aspect, the present invention provides a method for diagnosing a power grid fault based on a deep neural network, including:
acquiring power grid fault alarm information and a network topology structure;
constructing a neural network sample according to the power grid fault alarm information and the network topology structure, wherein the neural network sample comprises a neural network learning sample and a neural network testing sample;
preprocessing a neural network sample to obtain fault feature vectors of a neural network learning sample and a neural network testing sample;
inputting fault feature vectors of a neural network learning sample into a deep neural network model which is constructed in advance and optimized based on a chaotic particle swarm algorithm for learning, and obtaining an optimized and learned deep neural network model;
and (3) inputting the fault feature vector of the neural network test sample into the optimized and learned deep neural network model for probability diagnosis test, and obtaining a fault probability result.
Further, the neural network learning sample comprises a sample label and a fault feature vector corresponding to the sample, wherein the sample label is a 0-1 variable, the real fault condition of the equipment is represented, 1 represents a fault, and 0 represents no fault.
Further, the method for constructing the deep neural network model comprises the following steps:
the deep neural network model is obtained by modeling the deep neural network of the neural network sample, and the formula is as follows:
wherein: n is the number of samples; a, a i And y i Respectively representing the network output and the sample label corresponding to the ith sample.
Further, the optimization method of the deep neural network model comprises the following steps:
step 1: weights (w) of deep neural networks i，j ) And bias (b) i，j ) Defining particles, initializing population quantity, maximum iteration times, and positions and speeds of each particle; wherein w is i，j B is the weight between the ith hidden layer neuron and the jth output layer neuron i，j Biasing weights between the ith hidden layer neuron and the jth output layer neuron;
step 2: determining and storing a local best lbest and a global best gbest of particles with initial values;
step 3: updating the speed and the position of each particle, searching the lbest and the gbest of the current particle, comparing the lbest and the gbest of the previous particle, updating the lbest and the gbest, and storing; if the current speed and position are outside the boundary, taking an upper or lower boundary;
step 4: and (3) repeating the step until the termination condition is reached, wherein the value of the global optimal gbest is the bias and output weight of the hidden layer neuron of the optimized deep neural network.
Further, in the step 3, the speed and the position of each particle are updated by the following formula:
in the n-dimensional search space, after the kth iteration, the ith particle position is X i (k)＝{x i1 (k)，x i2 (k)，…，x in (k) Sum of velocity V i (k)＝{v ii (k)，v i2 (k)，…，v in (k) When iterating k+1, the speed and position of the ith particle are updated by the calculation of equations (3) and (4), respectively:
V i (k+1)＝
wV i (k)+c 1 r 1 [lbest i (k)-X i (k)]+c 2 r 2 [gbest i (k)-X i (k)] (3)
X i (k+1)＝X i (k)+V i (k+1) (4)
wherein: w is inertial weight; c 1 And c 2 A constant for controlling the search space between the local best lbest position and the global best gbest position; r is (r) 1 And r 2 Is at [01 ]]Random numbers uniformly distributed in the inner part; lbesti (k) is the local best of the ith particle after the kth iteration; gbest (k) is the global best of the ith particle after the kth iteration; the parameters adjusted by CPSO algorithm are as follows:
w(k+1)＝
4[w min +(w max -w min )w(k)][1-w min +(w max -w min )w(k)] (5)
c z (k+1)＝4[ε min +(c max -c min )c z (k)][1-c min +(c max -c min )c z (k)] (6)
r z (k+1)＝4r(k)[1-r(k)] (7)
wherein: w (k) is the inertia weight of iteration k times; w (k+1) is the inertia weight of iteration k+1 times; c z (k+1) is a constant of the search space iterated k+1 times; r is (r) z (k+1) is a random number iterated k+1 times; w (w) max /w min Is the maximum/minimum inertial weight; c max /c min Z=1, 2, which is a constant of the maximum/minimum search space.
Further, the inputting the fault feature vector of the neural network test sample into the optimized and learned deep neural network model for probability diagnosis test, obtaining a fault probability result, includes:
inputting the fault feature vector of the neural network test sample into the optimized and learned deep neural network model to obtain the network output fault probability;
and comparing the network output fault probability with a preset theoretical fault probability to obtain a difference value, and obtaining the difference between the diagnostic capability of the deep neural network model and the diagnostic capability of the theoretical fault according to the difference value.
Furthermore, according to the number of historical fault elements of the power grid, under the condition that the number of samples is insufficient, a sample expansion extraction and sample simulation generation method is adopted to realize smooth training of the model, and the method specifically comprises the following steps:
firstly, expanding the historical records of other elements which are counted to be the same as or similar to the diagnosed element, and if the obtained records are still insufficient, further expanding the area range until the historical records of the same type of equipment are enough; and then counting error probability of each signal based on the historical record, and obtaining a complete sample of the equipment to be diagnosed through random sampling.
In a second aspect, the present invention provides a deep neural network-based power grid fault diagnosis device, including:
the acquisition module is used for acquiring power grid fault alarm information and a network topological structure;
the system comprises a sample construction module, a network topology module and a network analysis module, wherein the sample construction module is used for constructing a neural network sample according to the power grid fault alarm information and the network topology structure, and the neural network sample comprises a neural network learning sample and a neural network testing sample;
the preprocessing module is used for preprocessing the neural network sample to obtain fault feature vectors of the neural network learning sample and the neural network test sample;
the optimizing and learning module is used for inputting fault feature vectors of the neural network learning sample into a deep neural network model which is constructed in advance and optimized based on a chaotic particle swarm algorithm for learning, and obtaining the optimized and learned deep neural network model;
the probability test module is used for inputting the fault feature vector of the neural network test sample into the optimized and learned deep neural network model for probability diagnosis test, and obtaining a fault probability result.
In a third aspect, the present invention provides an electronic device, comprising a processor and a storage medium;
the storage medium is used for storing instructions;
the processor is operative according to the instructions to perform the steps of the method according to any one of the preceding claims.
In a fourth aspect, the present invention provides a computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of any of the methods described in the preceding claims.
Compared with the prior art, the invention has the beneficial effects that:
the invention provides a power grid fault diagnosis method and device based on a deep neural network, which are characterized in that a deep neural network model is built on a sample, a chaotic particle swarm algorithm is utilized to learn and train the deep neural network, the deep neural network has higher diagnosis accuracy, the model builds a corresponding deep learning neural network for each element, SCADA data is input, the fault probability of the corresponding element is output, and the neural network is trained and learned based on the chaotic particle swarm algorithm; furthermore, aiming at the problems of large number of samples required by deep learning and small actual power system fault history records, the invention further provides a record expansion extraction and sample simulation generation method for realizing smooth training of a model, and the correctness and feasibility of the fault diagnosis model and the sample expansion method built by the invention are demonstrated by a calculation example, and the simulation experiment also shows that the method has strong data characteristic extraction and diagnosis capability and can quickly and accurately obtain element fault probability.
Drawings
FIG. 1 is a schematic diagram of a neuron structure according to one embodiment of the present invention;
FIG. 2 is a schematic diagram of a neural network according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a line fault signature vector according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a transformer fault signature vector according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of sample generation probabilities in accordance with an embodiment of the present invention;
FIG. 6 is a schematic diagram of the number of samples in each sample set simulated in accordance with one embodiment of the invention;
FIG. 7 is a diagram illustrating the change of the loss function values of the training set and the verification set according to an embodiment of the present invention;
FIG. 8 is a schematic diagram of a test sample input feature according to an embodiment of the present invention;
FIG. 9 is a schematic diagram of theoretical failure probability and network output result according to an embodiment of the present invention;
fig. 10 is a flowchart of a deep neural network-based power grid fault diagnosis method according to an embodiment of the present invention.
Detailed Description
The invention is further described below with reference to the accompanying drawings. The following examples are only for more clearly illustrating the technical aspects of the present invention, and are not intended to limit the scope of the present invention.
Example 1
As shown in fig. 10, this embodiment describes a power grid fault diagnosis method based on a deep neural network, including:
acquiring power grid fault alarm information and a network topology structure;
constructing a neural network sample according to the power grid fault alarm information and the network topology structure, wherein the neural network sample comprises a neural network learning sample and a neural network testing sample;
preprocessing a neural network sample to obtain fault feature vectors of a neural network learning sample and a neural network testing sample;
inputting fault feature vectors of a neural network learning sample into a deep neural network model which is constructed in advance and optimized based on a chaotic particle swarm algorithm for learning, and obtaining an optimized and learned deep neural network model;
and (3) inputting the fault feature vector of the neural network test sample into the optimized and learned deep neural network model for probability diagnosis test, and obtaining a fault probability result.
The application process of the deep neural network-based power grid fault diagnosis method provided by the embodiment specifically relates to the following steps:
firstly analyzing a deep neural network structure, constructing a neural network sample through fault alarm information and a network topology structure received by an SCADA (supervisory control and data acquisition) system, expanding the sample under the condition of insufficient historical samples, constructing a deep learning neural network model for each sample element through deep neural network modeling of the sample, inputting element fault characteristic vectors processed into 1/-1, outputting fault probability of the corresponding element, selecting a part of samples to be used as a learning sample set of the neural network through data preprocessing, learning the neural network through the learning sample set, participating in neural network learning by using a chaotic particle swarm algorithm, adjusting the weight of the deep neural network and neuron bias, selecting part of the neural network samples constructed by the SCADA received fault alarm information and the network topology structure, and establishing a test sample set different from the neural network learning sample set through data preprocessing; inputting the processed test sample set into a learned neural network diagnosis module, activating a fault diagnosis module to perform fault diagnosis, and then performing diagnosis analysis on the output result of the fault diagnosis module; and finally obtaining the fault probability of the power grid.
In order to better explain the technical scheme of the invention, the preliminary work needs to pre-process related parameters and data. Wherein:
the invention provides a power grid fault diagnosis technical scheme of a deep neural network based on chaotic particle swarm optimization, which comprises the following steps of:
in a first aspect, preprocessing an input sample data structure to construct a deep neural network model, comprising:
the fault diagnosis problem of the power system equipment is a two-class problem, and y is used for representing a sample observation value, and then positive class y=1 represents equipment faults, and negative class y=0 represents equipment faults. If x is used to represent the input sample feature of the corresponding device, the device failure probability h (θ) x can be expressed as
h θ (x)＝P(y＝1|x)＝1/(1+e -g(θ，x) ) (1)
Wherein: θ is a model parameter; g (θ, x) is a classification boundary, the specific calculation formula of the classification boundary is determined by the functional form to be fitted, and the probability of no fault of the equipment is as follows:
assuming N samples, the observations are y 1 ，y 2 ，…，y N The corresponding sample is characterized by a vector X 1 ，X 2 ，…X N The observation value y can be obtained by combining the formula (1) and the formula (2) i The probability of (2) is
If the samples are independent of each other, the model parameters θ are adjusted by maximum likelihood estimation, then the likelihood function can be obtained from equation (3) as
And carrying out logarithmic operation on the obtained product to obtain:
the optimization method is adopted to obtain the parameter theta so that the log likelihood function lnL (theta) takes the maximum value.
The objective function E is related to all the parameters θ (w i，k，j b i，j ，i＝1，2，3，4，k∈[1，n i-1 ]，j∈[1，n i ])
Fig. 1 shows any neuron structure, where x is the input, w is the weight, b is the bias, f is the activation function, z= Σ i w i x i +b is the input of the activation function and a=f (z) is the activation function strain amount, i.e. the neuron output value.
The overall structure of a triple-hidden-layer-containing fully-connected deep neural network composed of neurons as shown in fig. 1 linked is shown in fig. 2.
In FIG. 2, l 0 For the input layer, the number of neurons in the layer n 0 Equal to the number of bits of the constructed sample feature vector; l (L) 1 -l 3 As hidden layer, l 4 For the output layer, l i Each neuron x of the (i=1, 2,3, 4) layer i，j (j∈[1，n i ]) Are all equal to l i-1 Each neuron x in (a) i-1，k (k∈[1，n i-1 ]) Connected with each other, the connection weight is w i，k，j . Hidden layer l 1 -l 3 The activation function of the neuron is a Relu function, as shown in formula (6); output layer neuron x 4，1 The activation function of (2) is h as shown in formula (1) θ (x) The output a is the network output and represents the event probability corresponding to the sample 4，1 Input of (i) is
I.e., g (θ, x) in formula (1)
The network minimization objective-loss function E is calculated as
Wherein: n is the number of samples; a, a i And y i Respectively representing the network output and the sample label corresponding to the ith sample. Sample structure (data preprocessing): the sample label is 0-1 variable, which represents the real fault condition of the equipment (1 represents fault, 0 represents no fault), if the diagnosis object is a certain line, a set of main protection is providedIf the diagnosis object is a transformer in a generator transformer set, only one end is connected with the circuit breaker, the input feature vector of the fault diagnosis model of the transformer is shown in figure 4, and the input feature vector construction method of other element fault diagnosis models is similar to the method.
For the attributes in fig. 3 and 4, a value of 1 indicates that the regulatory center received the corresponding signal, -1 indicates that no signal was received. The analog information such as line power, transformer power and the like comes from SCADA telemetry data, 4 SCADA power sampling data and 6 SCADA power sampling data at the front 4 SCADA power sampling data and the rear 6 SCADA power sampling data at the time of receiving the 1 st related remote signaling signal are selected, if any two adjacent data differences in the 10 power sampling data exceed 50% of rated values, the attribute value of power change is 1, and otherwise, the attribute value of power change is-1; the 1 st telemetry data is 0, the attribute value of 'initial power is 0' is 1, otherwise, the attribute value is-1; the last telemetry data is 0 with a "final power of 0" attribute value of 1, otherwise it is-1.
When constructing the sample, the protection and switching action signals need to be detected initially to determine whether the action signals are valid. When any one of the following occurs, the corresponding action signal is considered invalid, and is processed as an unreceived signal, and the attribute value is-1:
(1) The switch or the protection signal is marked with a full data judging mark by the SCADA system, and the signal is re-evoked by the master station in a certain time period of the substation and is not real deflection information at the fault moment;
(2) The protection, switch or the station is hung with a special type of signpost, including 'overhaul', 'stop operation', 'standby', etc., in which case the signals of the corresponding devices are all non-fault information;
(3) If the signal action times exceed the maximum set times, namely the control center monitoring system receives the multiple opening and closing action signals of the same switch frequently in a short time, the signals are invalid displacement signals and are not fault information.
According to the principle, corresponding information is obtained from the SCADA database to form an input/output sample.
In a second aspect, learning training is performed on the deep neural network by using a chaotic particle swarm algorithm, including:
the chaotic particle swarm (chaos particle swarm optimization, CPSO) algorithm is an optimization algorithm added with the chaotic algorithm on the basis of a particle swarm (particle swarm optimization, PSO) algorithm. Compared with the PSO algorithm, the method has better searching capability and faster convergence speed.
In the n-dimensional search space, after the kth iteration, the ith particle position is X i (k)＝{x i1 (k)，x i2 (k)，…，x in (k) Sum of velocity V i (k)＝{v i1 (k)，v i2 (k)，…，v in (k) When iterating k+1, the speed and position of the ith particle are updated by the calculation of equations (3) and (4), respectively:
V i (k+1)＝
wV i (k)+c 1 r 1 [lbest i (k)-X i (k)]+c 2 r 2 [gbest i (k)-X i (k)] (3)
X i (k+1)＝X i (k)+V i (k+1) (4)
wherein: w is inertial weight; c 1 (cognitive parameters) and c 2 (social parameters) are constants that control the search space between the local best (lbest) location and the global best (gbest) location; r is (r) 1 And r 2 Is at [01 ]]Random numbers uniformly distributed in the inner part; lbesti (k) is the local best of the ith particle after the kth iteration; the gbesti (k) is the global best of the ith particle after the kth iteration. The parameters adjusted by the CPSO algorithm are as follows.
w(k+1)＝
4[w min +(w max -w min )w(k)][1-w min +(w max -w min )w(k)] (5)
c z (k+1)＝4[ε min +(c max -c min )c z (k)][1-c min +(c max -c min )c z (k)] (6)
r z (k+1)＝4r(k)[1-r(k)] (7)
Wherein: w (k) is the inertia weight of iteration k times; w (k+1) is the inertia weight of iteration k+1 times; c z (k+1) is a constant of the search space iterated k+1 times; r is (r) z (k+1) is a random number iterated k+1 times; w (w) max /w min Is the maximum/minimum inertial weight; c max /c min Z=1, 2, which is a constant of the maximum/minimum search space.
Deep neural network based on chaotic particle swarm optimization algorithm:
when the shallow neural network learns and trains itself, the shallow neural network is slightly insufficient in terms of parameters such as weight and center selection, and is unfavorable for power grid fault diagnosis. Aiming at the problem, the invention utilizes the chaotic particle swarm algorithm to search for the selection of the optimization weight and the bias of the hidden layer neurons during deep nerve training. The method comprises the following specific steps:
(1) Weights (w) of deep neural networks i，j ) And bias (b) i，j ) Defined as particles, initializing population number, maximum number of iterations, position and velocity of each particle, etc.
(2) The lbest and gbest of the particles of the initial value are determined and saved.
(3) Updating the speed and position of each particle through formulas (3) to (7), searching for the lbest and gbest of the current particle and comparing with the lbest and gbest of the previous particle, updating the lbest and gbest and storing. If the current speed and position are outside the boundaries, then either an upper or lower bound is taken.
(4) Repeating the step (3) until the termination condition is reached. In addition, the global optimum (gbest) value is the bias and output weight of the hidden layer neurons of the optimized deep neural network.
In a third aspect, in the case of insufficient data of the historical sample fault, a specific sample expansion method is proposed:
in the invention, in the fault diagnosis of an actual power system, if the historical data is sufficient, the historical data is directly used as the deep neural network training sample provided by the first aspect of the invention to train the deep neural network training sample.
Because the strict equipment network access detection system, the regular maintenance system and the strict personnel operation and training system of the electric power system can consider that the quality and the performance of the same type of equipment are satisfactory as long as the network access detection is passed, the difference is negligible in engineering, the geographical positions of the stations in the same area are very close, the corresponding microenvironments are similar, the faults caused by the human factors of all the stations are similar due to the strict personnel operation and training system, and the factors are main factors influencing the action correctness of the equipment, so that the history record of a certain equipment can be gradually expanded to the history records of the same type of equipment in the same station, the adjacent station and the area as required, and the corresponding obtained sample can reflect the operation characteristics of the specific equipment.
Based on the above consideration, the invention provides a new method for sample expansion, and the basic idea is as follows: firstly, expanding the historical records of other elements which are counted to be the same as or similar to the diagnosed element, and if the obtained records are still insufficient, further expanding the area range until the historical records of the same type of equipment are enough; and then counting error probability of each signal based on the historical record, and obtaining a complete sample of the equipment to be diagnosed through random sampling. The invention takes a certain transformer as an example, and the specific sample expansion method is as follows:
step 1: extracting a fault history record of the transformer;
step 2: if the records are insufficient, continuously extracting the historical fault records of other transformers with the same model, the same configuration and similar installation time in the same factory station; otherwise, the sample expansion is finished;
step 3: if the number of the samples is still insufficient, further expanding the search range to adjacent stations, other stations with the same voltage level in the same area until enough histories are obtained, or obtaining samples as many as possible; otherwise, the sample expansion is finished;
and step 4, if enough records still cannot be obtained as the neural network training data, adopting a probability-based sampling expansion strategy to obtain enough training samples, wherein the method comprises the following steps of: according to the acquired sample size, for a remote signaling signal, calculating the probability of taking 1 of each attribute in the input vector under the conditions of corresponding equipment faults and no faults respectively; otherwise, the sample expansion is finished, the calculation method is that
Wherein: n (y=1), N (y=0) represent the number of equipment failure, failure-free samples in the obtained samples, respectively; p (x) i = 1|y =1) indicates the probability of receiving a corresponding valid telemetry signal at the time of failure; p (x) i = 1|y =0) indicates the probability of receiving a corresponding valid telemetry signal when no fault has occurred.
Similarly, for telemetry data, the processing of the two attributes of "power change" and "power before change 0" is the same as the method described above; and if the final power is 0, respectively counting the probability that the attribute is 1 when the power flow is changed under the fault and fault-free conditions.
Step 5: determining each input x in the sample by adopting a random sampling method according to the obtained value probability of each input i (i∈[1，n 0 ]) The specific value determining method is as follows:
(1) For element failure conditions, sample tag y=1, the ith input x in the sample i The method comprises the following steps:
wherein r is i E (0, 1) is a random number generated by the sample.
(2) Similarly, for a device no fault condition, sample label y=0, sample feature x i Is that
Similarly, two scenarios are discussed for the "final power 0" attribute: (1) when the tide is unchanged, the attribute is equal to the tide before changing which is 0; (2) when the tide changes, the attribute value is determined by a random sampling method.
Step 6: and splicing the sample label and the obtained characteristic value to form a long vector as a complete sample.
Step 7: and 5, repeating the step until at least 2000 samples of faults and faults are obtained, and forming a sample set together with the samples in the SCADA database for network training.
For other types of diagnosed elements, a similar approach may be used to obtain sufficient data to train the network model of the first aspect.
According to the method, in the neural network learning and training process, the chaotic particle swarm algorithm is adopted to participate in optimization of the neural network weight and the neuron bias, and the method can achieve higher power grid fault diagnosis efficiency. Simulation results also verify good performance of the scheme.
The method in the embodiment of the invention is illustrated by simulation verification in the following description with reference to the specific implementation.
In the simulation example, the efficiency of the proposed power grid fault diagnosis scheme of the deep neural network based on chaotic particle swarm optimization is demonstrated through simulation. Considering that the power transmission line fault is one of typical faults in the power system fault, and the number of times of occurrence of the power transmission line fault is more, and the corresponding protection, circuit breaker and other configurations are more complex, the invention takes the power transmission line as an example for two-aspect verification, firstly, sample generation and verification of network training learning correctness based on the sample are carried out; and the fault diagnosis model constructed by the invention is used for fault diagnosis of a certain practical system, and the feasibility and effectiveness of engineering application of the diagnosis model are further verified.
The method in the step 5 and the step 6 in the second aspect are used for generating samples of equipment faults and no faults, and then the simulation samples are used for training the fault diagnosis fully-connected deep neural network constructed by the invention so as to verify the probability learning capability of the fault diagnosis fully-connected deep neural network.
Simulation sample generation: sample setup is the same as in fig. 3, and the sample generation correlation probabilities are shown in fig. 5. The number of failed and non-failed samples of the training set, test set, validation set are shown in fig. 6.
In fig. 5, the fault condition and the no fault condition respectively represent that the diagnosed element actually has a fault or has not a fault, and the corresponding probability represents the probability that the scheduling center receives the corresponding signal in the case.
Network model structure parameter selection: the experimental network structure is determined by a grid search method. For the parameters to be determined of the network, namely the hidden layer number of the network and the number of neurons of each layer, a deep neural network is built by taking a plurality of groups of parameter values within a certain value range, training is carried out by using the generated simulation samples until convergence, the parameters are set to be (30, 20 and 10) (the (30, 20 and 10) represent that the hidden layer number of the network is 3, the number of neurons of the hidden layers of the three layers is 30, 20 and 10 respectively), and subsequent diagnostic test experiments are carried out, wherein the loss function values of a test set and a verification set in the training process are shown in figure 7. It can be seen that the convergence speed of the loss function in the training set and the verification set is ideal, and the performance is good.
Probability diagnostic test:
the theoretical failure probability of the test sample is calculated by the following process: let the input vector be x, according to the bayesian formula, the element failure probability can be obtained as:
P(y＝1|x)＝P(y＝1)P(x|y＝1)/
[P(y＝0)P(x|y＝0)+P(y＝1)P(x|y＝1)] (11)
wherein P (y=1), P (y=0) respectively represent prior probabilities of faults and non-faults of corresponding equipment in the sample set, and fault samples and non-fault sample number N in the sample set 1 、N 0 4100, then:
p (x|y=0) and P (x|y=1) represent the probability of inputting x in the sample in the case of failure/no failure, respectively, i.e
In which x is i Representing the ith feature bit in the signal feature vector x.
The method comprises the steps of testing the capacity of obtaining probability characteristics of a model by using 200 training samples, substituting x and the probability set in fig. 5 into formulas (11) to (13) together to obtain theoretical probability P (y= 1|x) that a diagnosed element breaks down when a characteristic vector is x, selecting other test samples similarly to calculation, defining the error between the model output probability and the theoretical probability as the absolute value of the difference between the model output probability and the calculated probability, and randomly selecting 6 test samples with the average error of the model on the test samples being 2.16%, wherein the input characteristics of the test samples are shown in fig. 8, and the theoretical fault probability and the network output result are shown in fig. 9.
The input features in fig. 9 are signal feature vector X; the theoretical probability is the probability of failure of the element corresponding to the input signal obtained by calculation of the formula; taking the signal characteristics as the input of the deep neural network, and correspondingly outputting the signal characteristics to obtain fault diagnosis probability of the signal characteristics; the error is the absolute value of the difference between the theoretical probability and the network output. As can be seen from the test results, the network model is accurate in predicting the failure probability of the test sample, the error mean value is only 2.16%, and the probability errors of the extracted 6 samples are all less than 6%. Therefore, the deep neural network based on the chaotic particle swarm optimization algorithm constructed by the invention has strong data characteristic extraction and diagnosis capabilities, and can quickly and accurately obtain the element fault probability.
According to the invention, the optimal weight and the neuron bias of the radial basis function neural network are searched by using a chaotic particle swarm algorithm, so that the fault diagnosis accuracy of the power grid is improved. The calculation example shows that the method has higher learning efficiency and higher fault diagnosis accuracy. The model establishes a corresponding deep learning neural network for each element, inputs SCADA data comprising accident total signals, switches and protection signals, outputs fault probability of the corresponding element, and trains and learns the neural network based on a chaotic particle swarm algorithm; furthermore, aiming at the problems of large number of samples required by deep learning and small actual power system fault history records, the invention further provides a record expansion extraction and sample simulation generation method for realizing smooth training of a model. The correctness and feasibility of the fault diagnosis model and the sample expansion method established by the invention are demonstrated by a calculation example, and the simulation experiment also shows that the method has stronger data characteristic extraction and diagnosis capability, and can rapidly and accurately obtain the element fault probability.
Example 2
The embodiment provides a deep neural network-based power grid fault diagnosis device, which comprises:
the acquisition module is used for acquiring power grid fault alarm information and a network topological structure;
the system comprises a sample construction module, a network topology module and a network analysis module, wherein the sample construction module is used for constructing a neural network sample according to the power grid fault alarm information and the network topology structure, and the neural network sample comprises a neural network learning sample and a neural network testing sample;
the preprocessing module is used for preprocessing the neural network sample to obtain fault feature vectors of the neural network learning sample and the neural network test sample;
the optimizing and learning module is used for inputting fault feature vectors of the neural network learning sample into a deep neural network model which is constructed in advance and optimized based on a chaotic particle swarm algorithm for learning, and obtaining the optimized and learned deep neural network model;
the probability test module is used for inputting the fault feature vector of the neural network test sample into the optimized and learned deep neural network model for probability diagnosis test, and obtaining a fault probability result.
Example 3
The embodiment provides an electronic device, which comprises a processor and a storage medium;
the storage medium is used for storing instructions;
the processor is operative according to the instructions to perform the steps of the method according to any one of the preceding claims.
Example 4
The present embodiment provides a computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of any of the methods described in the preceding claims.
The foregoing is merely a preferred embodiment of the present invention, and it should be noted that modifications and variations could be made by those skilled in the art without departing from the technical principles of the present invention, and such modifications and variations should also be regarded as being within the scope of the invention.
Claims (10)
1. The utility model provides a deep neural network-based power grid fault diagnosis method which is characterized by comprising the following steps:
acquiring power grid fault alarm information and a network topology structure;
constructing a neural network sample according to the power grid fault alarm information and the network topology structure, wherein the neural network sample comprises a neural network learning sample and a neural network testing sample;
preprocessing a neural network sample to obtain fault feature vectors of a neural network learning sample and a neural network testing sample;
inputting fault feature vectors of a neural network learning sample into a deep neural network model which is constructed in advance and optimized based on a chaotic particle swarm algorithm for learning, and obtaining an optimized and learned deep neural network model;
and (3) inputting the fault feature vector of the neural network test sample into the optimized and learned deep neural network model for probability diagnosis test, and obtaining a fault probability result.
2. The deep neural network-based power grid fault diagnosis method according to claim 1, wherein the neural network learning samples comprise sample labels and fault feature vectors corresponding to the samples, the sample labels are 0-1 variables, the real fault conditions of equipment are represented, 1 represents a fault, and 0 represents no fault.
3. The deep neural network-based power grid fault diagnosis method according to claim 2, wherein the deep neural network model construction method comprises the following steps:
the deep neural network model is obtained by modeling the deep neural network of the neural network sample, and the formula is as follows:
wherein: n is the number of samples; a, a i And y i Respectively representing the network output and the sample label corresponding to the ith sample.
4. The deep neural network-based power grid fault diagnosis method according to claim 3, wherein the deep neural network model optimization method comprises:
step 1: weights (w) of deep neural networks i,j ) And bias (b) i,j ) Defining particles, initializing population quantity, maximum iteration times, and positions and speeds of each particle; wherein w is i,j B is the weight between the ith hidden layer neuron and the jth output layer neuron i,j Biasing weights between the ith hidden layer neuron and the jth output layer neuron;
step 2: determining and storing a local best lbest and a global best gbest of particles with initial values;
step 3: updating the speed and the position of each particle, searching the lbest and the gbest of the current particle, comparing the lbest and the gbest of the previous particle, updating the lbest and the gbest, and storing; if the current speed and position are outside the boundary, taking an upper or lower boundary;
step 4: and (3) repeating the step until the termination condition is reached, wherein the value of the global optimal gbest is the bias and output weight of the hidden layer neuron of the optimized deep neural network.
5. The deep neural network-based power grid fault diagnosis method according to claim 4, wherein in said step 3, the speed and position of each particle are updated by the following formula:
in the n-dimensional search space, after the kth iteration, the ith particle position is X i (k)＝{x i1 (k),x i2 (k),…,x in (k) Sum of velocity V i (k)＝{v i1 (k),v i2 (k),…,v in (k) When iterating k+1, the speed and position of the ith particle are updated by the calculation of equations (3) and (4), respectively:
wherein w is inertial weight; c 1 And c 2 A constant for controlling the search space between the local best lbest position and the global best gbest position; r is (r) 1 And r 2 Is at [01 ]]Random numbers uniformly distributed in the inner part; lbesti (k) is the local best of the ith particle after the kth iteration; gbest (k) is the global best of the ith particle after the kth iteration; the parameters adjusted by CPSO algorithm are as follows:
w(k+1)＝
4[w min +(w max -w min )w(k)][1-w min +(w max -w min )w(k)] (5)
c z (k+1)＝4[ε min +(c max -c min )c z (k)][1-c min +(c max -c min )c z (k)] (6)
r z (k+1)＝4r(k)[1-r(k)] (7)
wherein w (k) is the inertia weight of iteration k times; w (k+1) is the inertia weight of iteration k+1 times; c z (k+1) is a constant of the search space iterated k+1 times; r is (r) z (k+1) is a random number iterated k+1 times; w (w) max /w min Is the maximum/minimum inertial weight; c max /c min Z=1, 2, which is a constant of the maximum/minimum search space.
6. The deep neural network-based power grid fault diagnosis method according to claim 1, wherein the step of inputting the fault feature vector of the neural network test sample into the optimized and learned deep neural network model for probability diagnosis test to obtain a fault probability result comprises the steps of:
inputting the fault feature vector of the neural network test sample into the optimized and learned deep neural network model to obtain the network output fault probability;
and comparing the network output fault probability with a preset theoretical fault probability to obtain a difference value, and obtaining the difference between the diagnostic capability of the deep neural network model and the diagnostic capability of the theoretical fault according to the difference value.
7. The deep neural network-based power grid fault diagnosis method according to claim 1, wherein under the condition that the number of samples is insufficient, a sample expansion extraction and sample simulation generation method is adopted to realize smooth training of a model according to the number of power grid historical fault elements, specifically comprising the following steps:
firstly, expanding the historical records of other elements which are counted to be the same as or similar to the diagnosed element, and if the obtained records are still insufficient, further expanding the area range until the historical records of the same type of equipment are enough; and then counting error probability of each signal based on the historical record, and obtaining a complete sample of the equipment to be diagnosed through random sampling.
8. A deep neural network-based power grid fault diagnosis device, comprising:
the acquisition module is used for acquiring power grid fault alarm information and a network topological structure;
the system comprises a sample construction module, a network topology module and a network analysis module, wherein the sample construction module is used for constructing a neural network sample according to the power grid fault alarm information and the network topology structure, and the neural network sample comprises a neural network learning sample and a neural network testing sample;
the preprocessing module is used for preprocessing the neural network sample to obtain fault feature vectors of the neural network learning sample and the neural network test sample;
the optimizing and learning module is used for inputting fault feature vectors of the neural network learning sample into a deep neural network model which is constructed in advance and optimized based on a chaotic particle swarm algorithm for learning, and obtaining the optimized and learned deep neural network model;
the probability test module is used for inputting the fault feature vector of the neural network test sample into the optimized and learned deep neural network model for probability diagnosis test, and obtaining a fault probability result.
9. An electronic device, characterized in that: comprises a processor and a storage medium;
the storage medium is used for storing instructions;
the processor being operative according to the instructions to perform the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having stored thereon a computer program, characterized by: the program, when executed by a processor, implements the steps of the method of any of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310041168.6A CN116205265A (en) | 2023-01-12 | 2023-01-12 | Power grid fault diagnosis method and device based on deep neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310041168.6A CN116205265A (en) | 2023-01-12 | 2023-01-12 | Power grid fault diagnosis method and device based on deep neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116205265A true CN116205265A (en) | 2023-06-02 |
Family
ID=86515445
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310041168.6A Pending CN116205265A (en) | 2023-01-12 | 2023-01-12 | Power grid fault diagnosis method and device based on deep neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116205265A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116819345A (en) * | 2023-08-29 | 2023-09-29 | 武汉船用电力推进装置研究所(中国船舶集团有限公司第七一二研究所) | Battery system fault identification method and device, electronic equipment and storage medium |
| CN117034143A (en) * | 2023-10-10 | 2023-11-10 | 南京邮电大学 | Distributed system fault diagnosis method and device based on machine learning |
| CN117111589A (en) * | 2023-10-23 | 2023-11-24 | 中国机械总院集团云南分院有限公司 | Fault diagnosis method for numerical control machine tool control system based on Petri network |
| CN117978612A (en) * | 2024-03-28 | 2024-05-03 | 成都格理特电子技术有限公司 | Network fault detection method, storage medium and electronic equipment |
| CN117978612B (en) * | 2024-03-28 | 2024-06-04 | 成都格理特电子技术有限公司 | Network fault detection method, storage medium and electronic equipment |
-
2023
- 2023-01-12 CN CN202310041168.6A patent/CN116205265A/en active Pending
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116819345A (en) * | 2023-08-29 | 2023-09-29 | 武汉船用电力推进装置研究所(中国船舶集团有限公司第七一二研究所) | Battery system fault identification method and device, electronic equipment and storage medium |
| CN116819345B (en) * | 2023-08-29 | 2023-11-21 | 武汉船用电力推进装置研究所(中国船舶集团有限公司第七一二研究所) | Battery system fault identification method and device, electronic equipment and storage medium |
| CN117034143A (en) * | 2023-10-10 | 2023-11-10 | 南京邮电大学 | Distributed system fault diagnosis method and device based on machine learning |
| CN117034143B (en) * | 2023-10-10 | 2023-12-15 | 南京邮电大学 | Distributed system fault diagnosis method and device based on machine learning |
| CN117111589A (en) * | 2023-10-23 | 2023-11-24 | 中国机械总院集团云南分院有限公司 | Fault diagnosis method for numerical control machine tool control system based on Petri network |
| CN117111589B (en) * | 2023-10-23 | 2024-01-02 | 中国机械总院集团云南分院有限公司 | Fault diagnosis method for numerical control machine tool control system based on Petri network |
| CN117978612A (en) * | 2024-03-28 | 2024-05-03 | 成都格理特电子技术有限公司 | Network fault detection method, storage medium and electronic equipment |
| CN117978612B (en) * | 2024-03-28 | 2024-06-04 | 成都格理特电子技术有限公司 | Network fault detection method, storage medium and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112202736B (en) | Communication network anomaly classification method based on statistical learning and deep learning | |
| CN116205265A (en) | Power grid fault diagnosis method and device based on deep neural network | |
| CN109800875A (en) | Chemical industry fault detection method based on particle group optimizing and noise reduction sparse coding machine | |
| CN112051481B (en) | Alternating current-direct current hybrid power grid fault area diagnosis method and system based on LSTM | |
| US11840998B2 (en) | Hydraulic turbine cavitation acoustic signal identification method based on big data machine learning | |
| CN112838946B (en) | Method for constructing intelligent sensing and early warning model based on communication network faults | |
| CN111339712A (en) | Method for predicting residual life of proton exchange membrane fuel cell | |
| CN109633369B (en) | Power grid fault diagnosis method based on multi-dimensional data similarity matching | |
| CN111413565B (en) | Intelligent power grid fault diagnosis method capable of identifying and measuring tampering attack | |
| CN112069727B (en) | Intelligent transient stability evaluation system and method with high reliability for power system | |
| CN111060779B (en) | Power grid partition fault diagnosis method and system based on probabilistic neural network | |
| CN112821424B (en) | Power system frequency response analysis method based on data-model fusion driving | |
| CN111680875A (en) | Unmanned aerial vehicle state risk fuzzy comprehensive evaluation method based on probability baseline model | |
| CN112215279B (en) | Power grid fault diagnosis method based on immune RBF neural network | |
| CN110188837A (en) | A kind of MVB network fault diagnosis method based on fuzzy neural | |
| CN115481726A (en) | Industrial robot complete machine health assessment method and system | |
| CN113740666B (en) | Method for positioning root fault of storm alarm in power system of data center | |
| CN112085043B (en) | Intelligent monitoring method and system for network security of transformer substation | |
| Kwon et al. | A stochastic approach with hidden Markov model for accident diagnosis in nuclear power plants | |
| CN112327096B (en) | Fault diagnosis information fusion method and device based on self-adaptive strategy | |
| Pérez-Pérez et al. | Fault detection and isolation in wind turbines based on neuro-fuzzy qLPV zonotopic observers | |
| CN114638379A (en) | Edge side multi-agent OPC UA information analysis and decision method | |
| CN113076217A (en) | Disk fault prediction method based on domestic platform | |
| CN113609912A (en) | Power transmission network fault diagnosis method based on multi-source information fusion | |
| Dong et al. | Deep learning based multiple sensors monitoring and abnormal discovery for satellite power system |
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 |