Transformer Fault Prediction Method Based on Improved Adaptive Fuzzy Neural Network
By improving the parameter optimization method of the adaptive fuzzy neural network, the accuracy and efficiency problems of transformer fault prediction are solved, and higher prediction accuracy and power system stability are achieved.
Patent Information
- Application Number
- CN202310243664.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-14
- Publication Date
- 2025-09-26
- Estimated Expiration
- 2043-03-14
AI Technical Summary
The accuracy of existing transformer fault prediction methods is not ideal, and the particle swarm optimization algorithm is prone to fall into local optimality, resulting in insufficient transformer fault prediction accuracy and efficiency.
Improve the parameter optimization method of adaptive fuzzy neural network. By improving the inertia weight and learning factor function of particle swarm optimization, combining the convergence domain and Euclidean distance judgment, optimize the parameters of TS type adaptive fuzzy neural network and improve the network's optimization ability and diversity.
The accuracy and efficiency of transformer fault prediction are improved, the risk of falling into local optimality is reduced, and the safe and stable operation of the power system is ensured.
Smart Images
Figure CN116298037B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of neural networks, and in particular to a transformer fault prediction method based on an improved adaptive fuzzy neural network. Background Art
[0002] Transformers play an important role in raising and lowering voltage in power systems. Different types and degrees of faults occurring during operation will have an impact on the power system, and the safe operation of the power grid is closely related to them.
[0003] Currently, Liu Hang and his colleagues collected data on multidimensional factors affecting dissolved gases in transformer oil, extracted relevant multidimensional factors using grey correlation analysis, and established a multi-factor prediction model based on a proximity algorithm. Wang Youyuan and his colleagues used a cloud model to mine the association rules between state parameters and different states, constructed a transformer state transition matrix, and established a correction factor system to modify the state transition matrix to achieve transformer fault prediction. However, the accuracy of these methods is not ideal.
[0004] Existing methods improve the particle swarm's velocity update formula and inertia weight to optimize the network's structural parameters, but ignore the problem that the particle swarm is prone to falling into local optimality in the later stage. Summary of the Invention
[0005] Aiming at the shortcomings of existing methods, the present invention improves the traditional particle swarm algorithm and jumps out of the local optimum in the later stage; solves the problems of weak parameter optimization ability and slow convergence speed of TS-type adaptive fuzzy neural network, and improves the accuracy and efficiency of transformer fault prediction.
[0006] The technical solution adopted by the present invention is: a transformer fault prediction method based on an improved adaptive fuzzy neural network includes the following steps:
[0007] Step 1: Collect transformer oil chromatogram data and perform normalization on the oil chromatogram data;
[0008] Furthermore, the input data in the oil chromatography data includes: hydrogen, methane, ethane, ethylene, and acetylene.
[0009] Furthermore, the output data in the oil chromatogram data is the transformer fault type, including: normal, low energy discharge, low temperature overheat, medium and low temperature overheat, medium temperature overheat, high temperature overheat, and high energy discharge.
[0010] Step 2: Establish a TS-type adaptive fuzzy neural network model;
[0011] Further, specifically including:
[0012] Determine the network topology and establish a TS-type adaptive fuzzy neural network model. The TS-type adaptive fuzzy neural network is a feedforward neural network consisting of an input layer, a fuzzification layer, a rule layer, and an output layer.
[0013] The input layer receives oil chromatography data. The fuzzification layer fuzzifies the input data. The rule layer maps the fuzzified data to a set of rules. The output layer generates output based on the rules and input data. The fuzzification layer uses a Gaussian function as the membership function. An improved particle swarm optimization algorithm optimizes the premise parameters of the Gaussian function to form a network learning training layer. The third layer is the rule layer, where each node selects a rule. The number of nodes in this layer is 7. The network model has 5 nodes in the input layer and 1 node in the output layer. The output results are represented by integers 1-7, corresponding to the fault causes. The final network topology for transformer fault diagnosis is 5-35-7-1.
[0014] Step 3: Improve the traditional particle swarm algorithm and optimize the premise parameters of the TS-type adaptive fuzzy neural network;
[0015] Further, specifically including:
[0016] Step 31: Determine the population size, particle dimension, total number of iterations T, and the maximum and minimum values of the inertia weight, set the improved inertia weight function W and the improved learning factor function C, and initialize the particle position X. i =(X i1 , X i2 ,...,X id ,...,X iD ) and speed V i =(V i1 , V i2 ,...V id ,...V iD ), where i represents the number of the current particle, d represents the number of dimensions, and D represents the total dimension of the particle. The premise parameter is obtained by the current particle X i The position vector representation of
[0017] Step 32: Calculate the fitness value of each particle and compare the fitness value of the particle's current position with the optimal position P it has experienced. i =(P i1 , P i2 ,...P id ,...P iD )’s fitness value is compared, if it is better, it will be replaced with the current optimal position P best , otherwise it remains unchanged; the fitness value of the particle's current position is compared with the optimal position G experienced by the group i =(G i1 , Gi2 ,…,G id ,…G iD ) fitness value, if better, then record the particle number and replace it with the group's optimal position G best , otherwise unchanged; among them, P best is the optimal position of the individual, G best is the global optimal position, i represents the number of the current particle, d represents the number of dimensions, and D represents the total dimension of the particle;
[0018] Step 33: Update the current position and velocity of each particle according to the inertia, velocity and convergence domain judgment formula of the particle swarm algorithm; check whether the position and velocity of the particle are out of bounds. If so, eliminate the out-of-bounds and update the position and velocity of the particle; check whether the number of iterations reaches the maximum value or whether it meets the set minimum error requirement. If so, stop the iteration and output the premise parameters finally obtained by training, and use them as the initial parameters of the TS type adaptive fuzzy neural network to complete the optimization of the premise parameters of the TS type adaptive fuzzy neural network; if not, jump to step 32 and iterate again.
[0019] Furthermore, the formula for improving the inertia weight function is:
[0020]
[0021] Among them, W max and W min Represent the maximum and minimum values of W, F i represents the fitness value of particle i, F avg and F min They represent the average fitness value and the minimum fitness value of all particles at present, Rand represents a random number between (0,1), and K represents the ratio of the particle at the tth iteration to the total number of particle iterations T, K = t / T.
[0022] Furthermore, the formula for improving the learning factor function is:
[0023]
[0024]
[0025] Among them, C1 and C2 represent the self-learning factor and the group learning factor respectively, and K represents the ratio of the particle at the tth iteration to the total number of particle iterations T: K = t / T.
[0026] Furthermore, step three also includes: when the When , some particles are judged to have entered the convergence domain of the global optimal solution; the parameters of particles that have not entered the convergence domain and whose fitness function values are in the top 5% of the population are taken as the average of the parameters of the last 20% of the particles in the population; within the convergence domain, if the Euclidean distance between other particles and the optimal particle is less than 1, they are identical particles, and random parameters are assigned to the identical particles to jump out of the convergence domain.
[0027] Beneficial effects of the present invention:
[0028] 1. By improving the inertia weight function and learning factor function, the optimization ability of the particle swarm algorithm is enhanced compared with the traditional particle swarm algorithm;
[0029] 2. The convergence region and Euclidean distance are introduced to distinguish similar particles to increase the diversity of the population and escape from the local optimum in the later stage.
[0030] 3. Through experimental comparison of three situations, the prediction result error of the present invention is small, which improves the accuracy and efficiency of transformer fault prediction and has important practical significance for the safe and stable operation of transformers and power systems. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 This is a flow chart of a transformer fault prediction method based on an improved adaptive fuzzy neural network according to the present invention;
[0032] Figure 2 TS type adaptive fuzzy neural network structure diagram of the present invention;
[0033] Figure 3 This is the flow chart of the improved particle swarm optimization TS-type adaptive fuzzy neural network of the present invention. DETAILED DESCRIPTION
[0034] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore only shows the components related to the present invention.
[0035] like Figure 1 As shown, a transformer fault prediction method based on an improved adaptive fuzzy neural network includes the following steps:
[0036] Step 1: Transformer data collection: select five main characteristic gases from the oil chromatographic data of the oil-immersed transformer, including hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), and acetylene (C2H2), and perform normalization processing on them to obtain characteristic parameters that reflect transformer fault prediction;
[0037] The normalized five characteristic gas data are used as the input variables of the fuzzy neural network, and the Gaussian function is selected as the membership function; the transformer fault types include normal O1, low-energy discharge O2, low-temperature overheating O3, medium-low-temperature overheating O4, medium-temperature overheating O5, high-temperature overheating O6, and high-energy discharge O7, a total of seven transformer failure modes are used as the output variables of the fuzzy neural network.
[0038] Step 2: Establish a TS-type adaptive fuzzy neural network model and determine the number of fuzzy neural network layers, the number of neurons in each layer, and the membership function.
[0039] The TS-type adaptive fuzzy neural network is a feedforward neural network. The antecedent network includes an input layer, a fuzzification layer, and a rule layer, and the succesor network includes an output layer. The input layer accepts input data, the fuzzification layer fuzzifies the input data, the rule layer maps the fuzzified data to a set of rules, and the output layer generates output according to the rules and input data. The fuzzification layer uses a Gaussian function as the membership function, and optimizes the parameters in the Gaussian function through the improved particle swarm algorithm to form network learning training. The third layer is the rule layer. Each node selects a rule. The number of nodes in this layer is 7. The number of nodes in the input layer of the network model is 5, and the number of nodes in the output layer is 1. The output results are represented by integers 1-7, which correspond to the fault causes. The final network topology for transformer fault diagnosis is 5-35-7-1, as shown in the following example. Figure 2 As shown;
[0040] Step 3: Improve the traditional particle swarm algorithm and optimize the premise parameters of the TS-type adaptive fuzzy neural network so that the error of the neural network model reaches the preset accuracy;
[0041] Step 31: Determine the population size, particle dimension, total number of iterations T, and the maximum and minimum values of the inertia weight, set the improved inertia weight function W and the improved learning factor function C, and initialize the particle position X. i =(X i1 , X i2 ,...X id ,...X iD ) and speed V i =(V i1 , V i2 ,...V id ,...V iD ), where i represents the number of the current particle, d represents the number of dimensions, and D represents the total dimension of the particle. The premise parameters of the fuzzy neural network are obtained by the current particle X i The position vector is represented by , and the initialization interval of each dimension parameter is set to [-1,1].
[0042] Step 32: Calculate the fitness value corresponding to each particle using the mean absolute error formula and the transformer fault category output. The fitness value represents the error threshold of the fuzzy neural network. The smaller the error, the better the search performance of the particle. The position parameter when the particle reaches the minimum fitness value is the optimal premise parameter corresponding to the fuzzy neural network.
[0043] The fitness value of the particle's current position is compared with the optimal position P it has experienced. i =(P i1 , P i2 ,...P id ,...P iD )’s fitness value is compared, if it is better, it will be replaced with the current optimal position P best , otherwise it remains unchanged; the fitness value of the particle's current position is compared with the optimal position G experienced by the group i =(G i1 , G i2 ,...G id ,...G iD ) fitness value, if better, then record the particle number and replace it with the group's optimal position G best , otherwise unchanged; among them, P best is the optimal position of the individual, G best is the global optimal position, i represents the number of the current particle, d represents the number of dimensions, and D represents the total dimension of the particle.
[0044] Step 33: Update the current position and velocity of each particle according to the inertia, velocity and convergence domain judgment formula of the improved particle swarm algorithm; check whether the position and velocity of the particle are out of bounds. If so, eliminate the out-of-bounds and update the position and velocity of the particle; check whether the number of iterations reaches the maximum value or whether it meets the set minimum error requirement. If so, stop the iteration and output the premise parameters finally obtained by training as the initial parameters of the TS-type adaptive fuzzy neural network to complete the optimization of the premise parameters of the TS-type adaptive fuzzy neural network; if not, jump to step 32 and iterate again.
[0045] The update formula of the improved inertia weight W is:
[0046]
[0047] Among them, W max and W min Represent the maximum and minimum values of W, F i represents the fitness value of particle i, F avg and F minRepresent the average fitness value and minimum fitness value of all particles at present, Rand represents a random number between (0,1), and K represents the ratio of the particle at the tth iteration to the total number of particle iterations T: k = t / T; by introducing the average fitness value for comparison, we can max and W min Make your selection within.
[0048] The update formula for constructing a new learning factor C is:
[0049]
[0050]
[0051] Among them, C1 and C2 represent the self-learning factor and the group learning factor respectively, and K represents the ratio of the particle at the tth iteration to the total number of particle iterations T: K = t / T.
[0052] In order to ensure the diversity of particle population and avoid falling into local optimization in the later stage, when When a particle enters the convergence region of the global optimal solution, it is determined that some particles have entered the convergence region of the global optimal solution. At this time, the parameters of particles that have not entered the convergence region and whose fitness function values are in the top 5% of the population are calculated by taking the average parameters of the last 20% of the particles in the population (the smaller the fitness value, the better the search performance of the particle). Within the convergence region, if the Euclidean distance between other particles and the optimal particle is less than 1, they are considered to be identical particles. These identical particles are given random parameters to jump out of the convergence region, thereby ensuring the diversity of the population and escaping the local optimum in the later stage.
[0053] Step 4: Input the normalized data into the fuzzy neural network after particle swarm optimization for fault prediction, output the predicted transformer fault type, and evaluate the prediction results;
[0054] The method for evaluating predicted transformer faults is:
[0055]
[0056] Where F is the mean absolute error, k = 1, 2, ... n, n is the number of output values; E k is the expected output value, R k is the actual output value; if E k With R k If the values of are the same, the difference between them is recorded as 0, otherwise it is recorded as 1.
[0057] When the input and output are the same, the error comparison of the results of using the standard TS adaptive fuzzy neural network, the standard particle swarm optimization TS adaptive fuzzy neural network, and the improved particle swarm optimization TS adaptive fuzzy neural network algorithm is shown in Table 1.
[0058] Table 1 Performance comparison of ANFIS, PSO-ANFIS and the present invention
[0059]
[0060] It can be seen that when the present invention sets the same number of iterations, the optimal number of iterations, average error and time consumption are all optimal.
[0061] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.
Claims
1. A transformer fault prediction method based on an improved adaptive fuzzy neural network is characterized by: The following steps are involved: Step 1: Collect transformer oil chromatogram data and perform normalization on the oil chromatogram data; The input data in the oil chromatography data include: hydrogen, methane, ethane, ethylene, acetylene; Step 2: Establish a TS-type adaptive fuzzy neural network model; Step 3: Optimize the premise parameters of TS-type adaptive fuzzy neural network by improving particle swarm optimization algorithm; Step three specifically includes: Step 31: Determine the population size, particle dimension, and total number of iterations T And the maximum and minimum values of the inertia weight, set the improved inertia weight function W and improved learning factor function C , initialize the particle position and speed ;in, Indicates the label of the current particle, d The label indicating the dimension, D Indicates the total dimension of the particle, the premise parameter is passed through the current particle X i The position vector representation of The formula for improving the inertia weight function is: in, and Respectively The maximum and minimum values of Represents particles The fitness value of and Respectively represent the average fitness value and minimum fitness value of all particles at present, Represents a random number between (0,1), Indicates that the particle The total number of iterations of the particle The ratio of ; The formula for improving the learning factor function is: in, and Represent the self-learning factor and group learning factor respectively, Indicates that the particle The total number of iterations of the particle The ratio of: ; Step 32: Calculate the fitness value of each particle and compare the fitness value of the particle's current position with the optimal position it has experienced. If the fitness value is better, it will be replaced with the current optimal position. , otherwise it remains unchanged; the fitness value of the particle's current position is compared with the optimal position experienced by the group If the fitness value is better, the particle number is recorded and replaced with the optimal position of the group. , otherwise unchanged; among them, is the optimal position of an individual, is the global optimal position, Indicates the label of the current particle, d The label indicating the dimension, represents the total dimension of the particle; Step 33: update the current position and velocity of each particle according to the inertia, velocity and convergence domain judgment formula of the particle swarm algorithm; check whether the position and velocity of the particle are out of bounds; if so, eliminate the out-of-bounds and update the position and velocity of the particle; check whether the number of iterations reaches the maximum value or whether it meets the set minimum error requirement; if so, stop the iteration and output the premise parameters finally obtained by training, and use them as the initial parameters of the TS type adaptive fuzzy neural network to complete the optimization of the premise parameters of the TS type adaptive fuzzy neural network; if not, jump to step 32 and iterate again; Step 4: Predict the transformer fault type and evaluate the prediction results.
2. The transformer fault prediction method based on improved adaptive fuzzy neural network according to claim 1 is characterized in that: The output data in the oil chromatogram data is the transformer fault type, including: normal, low energy discharge, low temperature overheat, medium and low temperature overheat, medium temperature overheat, high temperature overheat, and high energy discharge.
3. The transformer fault prediction method based on improved adaptive fuzzy neural network according to claim 1 is characterized in that: The TS-type adaptive fuzzy neural network model consists of an input layer, a fuzzification layer, a rule layer, and an output layer. The number of nodes in the input layer is 5; the number of nodes in the fuzzification layer is 35; each node in the rule layer selects a rule, and the number of nodes is 7; the number of nodes in the output layer is 1.
4. The transformer fault prediction method based on improved adaptive fuzzy neural network according to claim 1 is characterized in that: Step 3 also includes: When the When , some particles are judged to have entered the convergence domain of the global optimal solution; the parameters of particles that have not entered the convergence domain and whose fitness function values are in the top 5% of the population are taken as the average of the parameters of the last 20% of the particles in the population; within the convergence domain, if the Euclidean distance between other particles and the optimal particle is less than 1, they are identical particles, and random parameters are assigned to the identical particles to jump out of the convergence domain.