A transformer fault diagnosis method based on improved particle swarm optimization algorithm

Abstract

The application discloses a transformer fault diagnosis method based on an improved particle swarm optimization algorithm, and comprises the following steps: acquiring oil gas chromatographic content and fault types of a transformer as sample data; training a GRNN network which is constructed in advance by using the sample data; optimizing parameters of the trained GRNN network by using an improved particle swarm optimization algorithm to obtain a final fault diagnosis model; and diagnosing real-time data of the transformer by using the final fault diagnosis model, judging whether the transformer has a fault, and outputting a fault type. The application has a fast training convergence speed and can accurately diagnose comprehensive faults of the transformer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer fault diagnosis, and more particularly to a transformer fault diagnosis method based on an improved particle swarm optimization algorithm. BACKGROUND

[0002] With the development of social economy and the improvement of comprehensive national strength, the stable operation of power equipment system is directly related to the future development and construction of the country. The transformer plays an important role in the entire power transmission system and power distribution system, and its operation state is directly related to the stable operation of the entire power equipment system. Once the transformer operation fault occurs, it will not only threaten the safety of the power transmission system and the power distribution system, but also have a certain impact on the efficiency of some production and processing related enterprises, thereby bringing immeasurable economic losses to the enterprises. If a serious transformer fault occurs, it may even threaten the life safety of the workers, so it is extremely important to supervise the relevant personnel to implement the transformer fault diagnosis work, which is of great significance to ensure the safe operation of the entire power system equipment and the safety of people's life and property.

[0003] At present, the methods and decisions for transformer fault diagnosis mainly include the following: (1) electrical quantity sensing technology, through the detection of voltage and current, the relevant values of each component of the transformer can be obtained, so as to judge the insulation state inside the equipment. It includes the method of detecting transformer core grounding current and high-frequency partial discharge method, and the method of mainly detecting short-circuit impedance of transformer winding, etc. (2) Acoustic vibration sensing technology, including methods of detecting partial discharge ultrasonic wave, vibration soundprint and substation noise, etc. (3) Photoelectric sensing technology, including methods of obtaining the temperature of the surface of the equipment by infrared thermal imager and ultraviolet imaging detection method, etc. (4) Chemical sensing technology, mainly for detecting dissolved gas in oil. Because overheat, discharge fault occurs inside the oil-filled and oil-paper insulated equipment, various gases will be generated, and the main gas composition and secondary gas composition detected can be used for fault judgment. (5) Thermal sensing technology, overheat fault has always been one of the important concerns of equipment operation management. The infrared imaging method has a temperature monitoring dead zone, and the infrared camera can only monitor the surface temperature of the outside part of the equipment, and cannot effectively monitor the temperature of the inside equipment and the internal temperature of the equipment. The above methods are more accurate for single fault diagnosis, but cannot accurately diagnose the comprehensive fault of the transformer.

[0004] Therefore, how to provide a transformer fault diagnosis method based on an improved particle swarm optimization algorithm, which can accurately diagnose the comprehensive fault of the transformer, is a technical problem that needs to be solved by those skilled in the art. SUMMARY

[0005] Therefore, the application provides a transformer fault diagnosis method based on an improved particle swarm optimization algorithm.

[0006] In order to achieve the above object, the application adopts the following technical scheme:

[0007] A transformer fault diagnosis method based on an improved particle swarm optimization algorithm comprises the following steps.

[0008] Obtain the oil gas chromatographic content and fault type of the transformer as sample data.

[0009] Train a pre-constructed GRNN network by using the sample data.

[0010] Optimize the parameters of the trained GRNN network by using the improved particle swarm algorithm to obtain a final fault diagnosis model.

[0011] Diagnose the real-time data of the transformer by using the final fault diagnosis model, judge whether the transformer has a fault, and output the fault type.

[0012] Further, the velocity and position update formula of the improved particle swarm algorithm is as follows:

[0013]

[0014]

[0015]

[0016] In the formula, w1 is an initial inertia weight value; w2 is a terminal inertia weight value; t is a current evolution generation; T is a maximum evolution generation; v id (t) is the velocity of the i-th particle in d dimensions at the t-th iteration; η1 and η2 are acceleration coefficients; r1 and r2 are random numbers between 0 and 1; x id (t) is the position of the i-th particle in d dimensions at the t-th iteration; P id is the individual optimal value; P gd is the global optimal value; η 1i is the initial value of η1(t); η 1f is the terminal value of η1(t); η 2i is the initial value of η2(t); η 2f is the terminal value of η2(t).

[0017] Further, the initial value of the acceleration coefficient η1 is set to 2.5, the terminal value is 0.5, the initial value of η2 is set to 0.5, and the terminal value is 2.5.

[0018] Further, the process of optimizing the parameters of the trained GRNN network by using the improved particle swarm algorithm comprises:

[0019] Setting initial values of the improved particle swarm algorithm;

[0020] Taking the root mean square error function of the output value of the GRNN network and the actual label as the fitness function of the particle swarm algorithm;

[0021] For each particle, the fitness function value of its current position is compared with the fitness function value of the best position it has experienced, if better, the corresponding position is taken as the optimal position;

[0022] For each particle, the fitness function value of its current position is compared with the fitness function value of the best position it has experienced, if better, the corresponding position is taken as the optimal position;

[0023] The position and speed of the particle are updated by using the speed and position updating formula of the improved particle swarm algorithm;

[0024] When the number of iterations reaches the maximum value, the optimal position searched is taken as the smoothing factor of the GRNN network.

[0025] Further, before training the pre-constructed GRNN network by using the sample data, further comprising:

[0026] The sample data is normalized.

[0027] Further, the oil and gas chromatographic content in the sample data at least includes monitoring quantities of H2, CH4, C2H2, C2H4 and C2H6 five kinds of gas.

[0028] Further, the fault types in the sample data at least include partial discharge, low-energy discharge, high-energy discharge, low-temperature overheating, medium-temperature overheating, high-temperature overheating, low-energy discharge and overheating, high-energy discharge and overheating, and normal state.

[0029] Further, the structure expression of the GRNN network is:

[0030]

[0031] In the formula, X i A is the training sample; is the predicted coordinate of the network output; and δ is the smoothing factor of the GRNN network.

[0032] Further, the expression of the root mean square error function of the output value of the GRNN network and the actual label is:

[0033]

[0034] In the formula, y^ is the predicted output of the GRNN network when inputting the test sample B; and y is the actual label of the test sample B.

[0035] Compared with the prior art, the transformer fault diagnosis method based on the improved particle swarm optimization algorithm provided by the present application can greatly improve the diagnosis accuracy of the network model by further optimizing the parameters of the trained GRNN network using the improved particle swarm algorithm, and can accelerate the network convergence speed and improve the training speed because the GRNN network has a simple structure and only one free parameter. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of the provided drawings.

[0037] Figure 1 The flowchart of the transformer fault diagnosis method based on the improved particle swarm optimization algorithm provided by the present application;

[0038] Figure 2 The RBF network classification running result provided by the present application;

[0039] Figure 3 The GRNN network classification running result provided by the present application;

[0040] Figure 4 The optimal fitness curve of the particle swarm algorithm optimization provided by the present application;

[0041] Figure 5 The optimal position curve of the particle swarm algorithm particle provided by the present application;

[0042] Figure 6 The diagnosis result diagram of the GRNN network optimized by the improved particle swarm algorithm provided by the present application;

[0043] Figure 7 The diagnosis error diagram of the GRNN network optimized by the improved particle swarm algorithm provided by the present application. DETAILED DESCRIPTION

[0044] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by those skilled in the art without creative efforts belong to the scope of the present application.

[0045] As shown in Figure 1 The embodiment of the present application discloses a transformer fault diagnosis method based on an improved particle swarm optimization algorithm, comprising the following steps:

[0046] S1, obtaining the oil gas chromatography content and fault type of the transformer as sample data;

[0047] S2, training the pre-constructed GRNN network by using the sample data;

[0048] S3, optimizing the parameters of the trained GRNN network by using the improved particle swarm optimization algorithm to obtain the final fault diagnosis model;

[0049] S4, diagnosing the real-time data of the transformer by using the final fault diagnosis model, judging whether the transformer has a fault, and outputting the fault type.

[0050] In one embodiment, in S1, the oil gas chromatography content in the sample data at least includes the monitoring amount of five gases of H2, CH4, C2H2, C2H4 and C2H6. The fault type in the sample data at least includes local discharge, low-energy discharge, high-energy discharge, low-temperature overheating, medium-temperature overheating, high-temperature overheating, low-energy discharge and overheating, high-energy discharge and overheating, and normal state.

[0051] The embodiment of the present application collects a plurality of data samples from the State Grid and published documents, covers various states of the transformer, randomly selects 180 groups of data from them as a training set to train the neural network model, randomly selects another 90 groups of data as a test set to evaluate the final neural network diagnosis performance, and the distribution of various data is shown in Table 1.

[0052] Table 1: Power transformer fault diagnosis sample

[0053]

[0054]

[0055] More advantageously, considering that the discreteness of data is easy to cause output saturation to affect the judgment result, before training the pre-constructed GRNN network by using the sample data, the method further comprises: performing normalization processing on the sample data.

[0056] The specific formula is as follows:

[0057]

[0058] In the formula, x scale is normalized sample data; x is sample data; x min is the minimum value of the index; x max is the maximum value of the index.

[0059] In transformer fault diagnosis, the neural network used is a multiple-input multiple-output model, and the processed data is used as a data set. Radial basis function neural network (RBF) and generalized regression neural network (GRNN) are used for simulation verification. The classification comparison results of the two algorithms are shown in Figure 2 and Figure 3

[0060] It can be found from the identification result of the radial neural network that if the data is not approximated, the identification result is a decimal number. Moreover, for some data, the classification result is between two categories, and it is impossible to determine the category to which it belongs. The radial basis function network can approximate any continuous function, but for discrete data, it cannot be completely approximated in theory.

[0061] It can be seen from the classification result that the classification effect of the generalized neural network is better than that of the radial basis function neural network. The generalized neural network can approximate a nonlinear continuous function with any precision. Its network training is simple, and the network training is completed at the same time when the training sample enters the hidden layer. Its training does not need iteration, and the training efficiency of the GRNN network is much higher than that of the RBF network, and it is more suitable for real-time processing of online data. Because of the simple network structure, there is only one free parameter, which can optimize the generalized neural network by finding the optimal value of the smoothing factor.

[0062] In one embodiment, the classical particle swarm algorithm also uses the concepts of "population" and "evolution", and adjusts according to the fitness of the individual relative to the population. In the particle swarm algorithm, each particle will correct its own position according to the individual optimal position and the group optimal position, and finally the particle will improve and optimize the overall optimal objective function. Each particle has two main parameters of particle motion speed and particle position when moving, and the optimization mode of the algorithm is an iterative process used in continuous space. The implementation of the particle swarm algorithm is to first initialize a random population, and then perform iterative optimization on each particle to find the individual optimal solution and compare the historical optimal position of each search group.

[0063] ​The traditional particle swarm algorithm is simple, fast in convergence speed, high in optimization efficiency and strong in universality, and can be used for solving complex optimization problems of nonlinearity, discontinuity, multi-constraint and multi-variable. However, the traditional particle swarm optimization algorithm has two defects in the optimization process. The first one is that when the particle swarm algorithm calculates some functions of multi-peak value, it is easy to fall into local optimum and difficult to jump out of it. The second one is that when the initial parameters are not set properly, the diversity of particles may disappear, the particles run generally slowly, and concentrate on a local extreme value. In this case, the optimal solution cannot be found. The two defects are interrelated and interacted, which leads to the result that the particle swarm algorithm falls into a local extreme value. In order to find the global optimal solution with good convergence efficiency, the two defects must be solved. The inertia weight concept is introduced in the improved particle swarm algorithm in order to further balance the exploration ability of the particle swarm algorithm. Specifically, in S3, the speed and position updating formula of the improved particle swarm algorithm is:

[0064]

[0065] In the above formula, v id (t) is the speed of the ith particle in d dimensions at the tth iteration; η1 and η2 are acceleration coefficients; r1 and r2 are random numbers between 0 and 1; x id (t) is the position of the ith particle in d dimensions at the tth iteration; P id is the individual optimal value; and P gd is the global optimal value.

[0066] A good strategy always strives to achieve a certain balance between the global exploration ability and the local search ability of the algorithm. The global exploration ability enables the algorithm to constantly search new areas and avoid premature convergence. The local search ability allows the algorithm to perform fine search in the promising area, improves the convergence speed and the accuracy of the solution. w(t) is the inertia weight, which describes the influence of the speed change in the iteration process of the particle, and controlling the size of w can balance the optimization ability of the particle swarm algorithm. Generally, taking a larger w value enables the particle to have a larger flight speed, which can strengthen the global optimization ability of the algorithm. Taking a smaller w value enables the flight speed of the particle to be smaller, which can strengthen the local optimization ability of the algorithm. Therefore, in practical application, w is set to a linearly decreasing value. In the beginning of the algorithm running, the particle performs global search. With the locking of the optimal area, the particle begins to perform fine search around the optimal solution. In the later stage of iteration, the algorithm has strong local search ability, so as to ensure that it can achieve a balance between the global search and the local search ability. The iteration formula of the inertia weight w is:

[0067]

[0068] In the above formula, w1 is an initial inertia weight value; w2 is a final inertia weight value; t is a current evolution number; and T is a maximum evolution number.

[0069] The values of the acceleration constants η1 and η2 are adjusted by the following formula, so that the acceleration constants have the ability to change linearly, and η1 and η2 approach the global optimum as the iteration number increases. The particles can continuously adjust the moving speed according to the size of the acceleration constants η1 and η2.

[0070]

[0071] η 1i is an initial value of η1(t); η 1f is a final value of η1(t); η 2i is an initial value of η2(t); η 2f is a final value of η2(t).

[0072] wherein the initial value of the acceleration constant η1 is set to 2.5, the final value is 0.5, the initial value of η2 is set to 0.5, and the final value is 2.5.

[0073] In one embodiment, in S3, the process of optimizing the parameters of the trained GRNN network by using the improved particle swarm algorithm includes:

[0074] S31, setting the initial value of the improved particle swarm algorithm;

[0075] S32, taking the root mean square error function of the output value of the GRNN network and the actual label as the fitness function of the particle swarm algorithm; wherein the calculation formula of the network structure obtained by training the training sample is as follows:

[0076]

[0077] In the formula, X i A is the training sample; is the predicted coordinate of the network output; and δ is the smoothing factor of the GRNN network.

[0078] Test phase: the oil spectrum content of the test sample is taken as the input of the network, and the output of the GRNN network is the predicted output value. The fitness function here is the root mean square error function of the test output value and the actual label, and the expression is:

[0079]

[0080] In the formula, is the predicted output of the GRNN network when the test sample B is input; and y is the actual label of the test sample B.

[0081] S33, for each particle, the fitness function value of its current position is compared with the fitness function value of its best position experienced, if better, the corresponding position of it is taken as the optimal position;

[0082] S34, for each particle, the fitness function value of its current position is compared with the fitness function value of the best position experienced globally, if better, the particle serial number of the best value is recorded, and the corresponding position of it is taken as the optimal position;

[0083] S35, the position and speed of the particle are updated by using the speed and position updating formula of the improved particle swarm optimization algorithm;

[0084] S36, when the iteration number reaches the maximum value, the optimal position searched is taken as the smoothing factor of the GRNN network.

[0085] The application also simulates and verifies the generalized regression neural network optimized by the particle swarm.

[0086] The acceleration coefficients η1 and η2 of the particle swarm algorithm are respectively the step length for adjusting the flight direction of the particle to the global optimal particle and the individual optimal particle, if too large, the particle may fly over the target region, if too small, the particle may be far away from the target region, at present, the value of η1 and η2 is set to about 2 in most documents, the acceleration constant designed in the experiment changes linearly with the iteration number, in order to improve the fine search ability of the algorithm, the initial value of the acceleration coefficient η1 is set to 2.5, the terminal value is 0.5, the initial value of η2 is set to 0.5, and the terminal value is set to 2.5; the population size of the PSO algorithm is small, and 6 particles can obtain good results for most problems, since the neural network structure in the document is relatively simple, the number of particle swarms is set to 6 in the experiment. The inertia weight coefficients w1 and w2 are used to control the influence of the previous speed on the current speed, a larger inertia weight coefficient can strengthen the global search ability of the PSO, and a smaller one can strengthen the local search ability. The inertia weight w1 and w2 are respectively set to 0.9 and 0.1, so that the PSO can explore a larger region at the beginning, quickly locate the approximate position of the optimal solution, and slow down the particle speed with the decrease of the inertia weight, and start fine local search.

[0087] Figure 4 For the best fitness convergence curve of the PSO, it can be seen that the convergence speed of the best fitness curve is very fast in the first 10 periods, then the fitness curve gradually becomes stable, and finally reaches a stable state. It can be seen from the particle swarm algorithm that all particles are in the optimal position, that is, the parameter optimization is realized; as shown in the figure, the optimal smoothing factor obtained by the particle swarm algorithm is 6.78; and the fault diagnosis result can be seen from Figure 6 and Figure 7 It can be seen that only 3 groups of diagnosis are wrong in 90 groups of training samples.

[0088] Based on the above-adopted feature data, the present application also gives a transformer fault diagnosis example analysis of the unoptimized GRNN method, and the fault diagnosis results of the two algorithms are shown in Table 2.

[0089] Table 2 Fault diagnosis recognition effect before and after optimization

[0090]

[0091] As shown in Table 2, the diagnosis accuracy of the PSO-GRNN algorithm is 95%. It can be seen that the PSO-GRNN algorithm has fast convergence speed and good diagnosis effect.

[0092] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the device disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0093] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A transformer fault diagnosis method based on an improved particle swarm optimization algorithm, characterized in that, The application relates to a transformer fault diagnosis method based on GRNN and improved particle swarm optimization. The method comprises the following steps: Obtaining oil gas chromatographic content and fault types of a transformer as sample data; Training a pre-constructed GRNN network by using the sample data; Optimizing parameters of the trained GRNN network by using an improved particle swarm optimization algorithm to obtain a final fault diagnosis model; Diagnosing real-time data of the transformer by using the final fault diagnosis model to determine whether the transformer has a fault and outputting a fault type; The process of optimizing the parameters of the trained GRNN network by using the improved particle swarm optimization algorithm comprises the following steps: Setting initial values of the improved particle swarm optimization algorithm; Taking a root mean square error function of an output value of the GRNN network and an actual label as an adaptive function of the particle swarm optimization algorithm; For each particle, comparing an adaptive function value of a current position of the particle with an adaptive function value of a best position experienced by the particle, and if the adaptive function value of the current position is better, taking the current position as the best position; For each particle, comparing an adaptive function value of a current position of the particle with an adaptive function value of a best position experienced by the particle, and if the adaptive function value of the current position is better, taking the current position as the best position; Updating the position and speed of the particle by using a speed and position updating formula of the improved particle swarm optimization algorithm; When the iteration number reaches a maximum value, taking the searched optimal position as a smoothing factor of the GRNN network; ； wherein is the predicted output of the GRNN network when inputting the test sample B; is the actual label of the test sample B.

2. The transformer fault diagnosis method based on improved particle swarm optimization algorithm according to claim 1, characterized in that, The expression of the root mean square error function of the output value of the GRNN network and the actual label is as follows: ； ； ； In the formula, w1 is the initial inertia weight value; w2 is the final inertia weight value; t is the current evolution generation; and T is the maximum evolution generation. Let be the velocity of the i-th particle in the d-dimensional dimension at iteration t; , This is the acceleration coefficient; , A random number between 0 and 1; x id (t) represents the position of the i-th particle in dimension d at iteration t; P id For the individual optimal value; P gd This is the globally optimal value; for The initial value; for The final value; for The initial value; for The final value.

3. The transformer fault diagnosis method based on improved particle swarm optimization algorithm according to claim 2, characterized in that, acceleration coefficient The initial value is set to 2.5 and the final value is set to 0.5, The initial value is set to 0.5 and the final value is set to 2.

5.

4. The transformer fault diagnosis method based on improved particle swarm optimization algorithm according to claim 1, characterized in that, The speed and position updating formula of the improved particle swarm optimization algorithm is as follows: Before the step of training the pre-constructed GRNN network by using the sample data, the method further comprises the following step:

5. The transformer fault diagnosis method based on improved particle swarm optimization algorithm according to claim 1, characterized in that, Normalizing the sample data.

6. The transformer fault diagnosis method based on improved particle swarm optimization algorithm according to claim 1, characterized in that, The oil gas chromatographic content in the sample data at least comprises monitoring amounts of H2, CH4, C2H2, C2H4 and C2H6.

7. The transformer fault diagnosis method based on improved particle swarm optimization algorithm according to claim 1, characterized in that, The fault types in the sample data at least comprise local discharge, low-energy discharge, high-energy discharge, low-temperature overheating, medium-temperature overheating, high-temperature overheating, low-energy discharge and overheating, high-energy discharge and overheating and normal state. The structure expression of the GRNN network is as follows: ； wherein is the training sample; predicted coordinates output by the network; and δ is a smoothing factor for the GRNN network.

Images