Transformer internal insulation life prediction method and system based on BP neural network

Abstract

A method and system for predicting the internal insulation life of a transformer based on a BP neural network. The method first obtains historical transformer operating data and preprocesses it using the principal hierarchy analysis method (AHP). The BP neural network model is then trained using the preprocessed historical transformer operating data. Initial network parameters of the BP neural network model are optimized using an ant colony algorithm combined with a genetic algorithm at the beginning of training. Hotspot temperature information of the transformer is predicted based on the trained BP neural network model. The predicted hotspot temperature information is input into an insulation life assessment model to obtain an insulation life assessment value. This insulation life assessment value is then converted into a final insulation life prediction value using the Monte Carlo method. The present invention utilizes a genetic algorithm to improve the ant colony algorithm, thereby increasing search efficiency and avoiding local optimal solutions.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer operation reliability, and in particular relates to a method and system for predicting the internal insulation life of a transformer based on a BP neural network. Background Art

[0002] The safe and reliable operation of transformers directly impacts the normal operation of power systems. A transformer failure not only disrupts power supply but also significantly impacts industrial and agricultural production, the lives of ordinary people, and even threatens personal and equipment safety. Accurately predicting transformer lifespan to improve reliability is a pressing issue and crucial for ensuring reliable power system operation.

[0003] Existing literature indicates that temperature is the primary cause of insulation fatigue failure. High temperatures can cause significant dielectric degradation in a short period of time. Even at lower temperatures, prolonged exposure can lead to irreversible degradation of insulation performance. Higher temperatures accelerate insulation aging and shorten lifespan. Furthermore, localized overheating contributes primarily to transformer oil aging due to the release of various oil-soluble trace gases. In summary, hotspot temperature information can be used to provide an indirect indicator of the transformer's approximate lifespan, enabling prediction of its lifespan.

[0004] Patent application number CN202111458661.5 provides a transformer life prediction method. This method integrates dynamic parameter characteristics with static parameter information of the transformer to be predicted and inputs them into a BP neural network to train a life prediction model. The dynamic parameter information includes at least one of the hotspot temperature of the transformer insulation winding, the transformer load voltage, the transformer load current, and the transformer magnetic flux. The static parameter information includes at least one of the moisture content of the transformer insulation paper, the water-soluble acid content in the transformer oil, and the acidity value of the transformer oil. Although this method can predict the remaining life of the transformer, the BP neural network suffers from the low prediction accuracy.

[0005] Patent application number CN202311701302.7 provides a method for predicting the remaining life of dry-type transformers based on an improved ant colony-optimized BP neural network. This method uses an ant colony algorithm to optimize the BP neural network to train data and obtain a remaining life prediction model. Although this method optimizes the BP neural network through the ant colony algorithm and reasonably adjusts the initial parameters of the BP neural network, thereby compensating for the insufficient prediction accuracy of the neural network and improving the accuracy of the transformer's remaining service life prediction, it suffers from the problem of excessive training time and the tendency to fall into local optimal solutions.

[0006] In summary, the existing technology of using ant colony algorithm to optimize BP neural network has the problem of long training time and easy to fall into local optimal solution. Summary of the Invention

[0007] The purpose of the present invention is to address the above-mentioned problems existing in the prior art and to provide a method, system, device and medium for predicting the internal insulation life of a transformer based on a BP neural network, which can accelerate the algorithm search efficiency and avoid falling into a local optimal solution.

[0008] To achieve the above objectives, the technical solutions of the present invention are as follows:

[0009] In a first aspect, the present invention provides a method for predicting the internal insulation life of a transformer based on a BP neural network, the prediction method comprising the following steps:

[0010] S1. Obtain the historical operating data of the transformer and preprocess it using the principal hierarchy analysis method;

[0011] S2. Using the pre-processed historical operation data of the transformer to train the BP neural network model, during the training, an ant colony algorithm combined with a genetic algorithm is used to optimize the initial network parameters of the BP neural network model, and then the BP neural network model is trained based on the optimal initial network parameters;

[0012] S3, predicting the hot spot temperature information of the transformer based on the trained BP neural network model;

[0013] S4. Inputting the predicted value of the hotspot temperature information into the insulation life assessment model to obtain an insulation life assessment value;

[0014] S5. Convert the insulation life evaluation value into the final insulation life prediction value based on the Monte Carlo method.

[0015] The S2 comprises the following steps:

[0016] S21, inputting pre-processed historical operation data of the transformer;

[0017] S22, parameter initialization: The elements composed of the initial weights and thresholds generated by each node of the BP neural network are recorded as set I wi , let the number of ants in the ant population be N, the ant path coordinates be the weights and thresholds of each node in the BP neural network, let the path i→j be from the starting point e through nodes i and j to the end point s, and the pheromone concentration of the path i→j is obtained by non-uniform distribution processing; the calculation formula for the non-uniform distribution processing is:

[0018]

[0019] In the above formula, v ijis the pheromone concentration from node i to node j; d ij is the Euclidean distance between node i and node j; (x i ,y i )、(x j ,y j )、(x e ,y e )、(x s ,y s ) are the coordinates of node i, node j, starting point e, and end point s respectively; d se is the Euclidean distance from the starting point e to the end point s; d ie is the Euclidean distance between node i and starting point e; d sj is the Euclidean distance from node j to end point s; v0 is the initial pheromone concentration;

[0020] S23. Path selection: Place all ants at the starting point and move them according to the following formula:

[0021]

[0022] In the above formula, represents the probability of the kth ant transferring from node i to node j at time t; α is the inspiration factor; β is the expectation factor; v ij (t) is the pheromone concentration on the path from node i to node j at time t; allowed k is the set of next nodes to which the kth ant can move; is a heuristic function that represents the expected degree of ant transfer from node i to node j;

[0023] S24. When all ants have traversed all the elements in the set, the pheromone concentration on the path is updated according to the following formula:

[0024] v ij (t+n)=(1-ρ)v ij (t)+Δv ij (t,t+n)+(h×(LL i→j )) / L i→j ;

[0025] In the above formula, v ij (t+n) is the pheromone concentration on the path from node i to node j at time t+n; ρ is the pheromone volatility coefficient; Δv ij (t, t+n) is the pheromone left by the ant on the path from node i to node j from time t to time t+n; L is the optimal path length of the previous iteration; L i→j is the path length from the current iteration node i to the node j; h is the reinforcement coefficient;

[0026] S25. Calculate the fitness function value of each ant, form a subpopulation of ants whose fitness values ​​are less than a preset threshold, substitute them into the genetic algorithm for crossover mutation, and form a new ant population with the subpopulation after crossover mutation and the ants whose fitness values ​​are greater than or equal to the preset threshold;

[0027] S26, return to step S23 and iterate until the maximum number of iterations is reached, output the ant with the highest fitness value as the optimal solution, and substitute it into the BP neural network as the optimal initial network parameter;

[0028] S27, calculating the network error of the BP neural network;

[0029] S28. Determine whether the network error of the BP neural network is less than the preset error threshold ε0; if so, the training ends; if not, update the network parameters of the BP neural network and return to step S27 to continue training.

[0030] In S24, the pheromone volatility coefficient ρ is adjusted according to the following formula:

[0031]

[0032] In the above formula, ρ(t) is the pheromone volatility coefficient at time t; ρ min is the minimum value of the pheromone volatility coefficient; η is the global pheromone decrease factor, η∈(0,1).

[0033] In step S25, the specific steps of crossover mutation include:

[0034] A1. Randomly pair the ants in the subpopulations. Calculate the crossover probability by combining Sigmoid and Gaussian distributions. Perform a crossover on each group of individuals to generate a new pair of individuals. The crossover is performed according to the following formula:

[0035]

[0036] In the above formula, a1 and a2 represent a group of paired individuals; b represents the crossover probability; c1 and c2 represent a group of newly generated individuals after crossover;

[0037] The calculation formula of the crossover probability is:

[0038]

[0039] In the above formula, b max 、b min is the maximum and minimum value of the crossover probability; J max represents the maximum fitness value in the subpopulation; k1 is the curve smoothing parameter; J′ represents the larger fitness value between the two individuals undergoing crossover; J avg represents the average fitness value of the subpopulation;

[0040] A2. Use the roulette wheel rule to select the individual to perform mutation. The probability of an individual being selected is:

[0041]

[0042] In the above formula, P i is the probability of individual i being selected in the subpopulation; f i is the fitness function of individual i;

[0043] A3. Randomly generate two natural numbers m1 and m2 that are greater than 1. M is the total number of genes in the gene coding string of the individual; the genes between m1 and m2 in the gene coding string of the individual selected by A2 are arranged in reverse order to obtain a mutant individual.

[0044] The insulation life assessment model is:

[0045]

[0046] In the above formula, L is the remaining life of transformer insulation; m is the mass fraction of the initial moisture content in the oil-paper; DP is the current degree of polymerization of the oil-paper; DP end is the specified value of the degree of polymerization at the end of life; A is a constant; θ H is the hot spot temperature value.

[0047] In a second aspect, the present invention provides a transformer internal insulation life prediction system based on an improved BP neural network, the prediction system comprising a data acquisition module, a model training module, a hotspot temperature prediction module, and an insulation life calculation module;

[0048] The data acquisition module is used to obtain the historical operation data of the transformer and pre-process it using the principal hierarchy analysis method;

[0049] The model training module is used to train the BP neural network model using the preprocessed historical operation data of the transformer. During the training, the initial network parameters of the BP neural network model are optimized by using the ant colony algorithm combined with the genetic algorithm, and then the BP neural network model is trained based on the optimal initial network parameters.

[0050] The hot spot temperature prediction module is used to predict the hot spot temperature information of the transformer based on the trained BP neural network model;

[0051] The insulation life calculation module is used to input the hot spot temperature information prediction value into the insulation life evaluation model to obtain the insulation life evaluation value, and convert the insulation life evaluation value into the final insulation life prediction value based on the Monte Carlo method.

[0052] The model training module is used to perform training according to the following steps:

[0053] S21, inputting pre-processed historical operation data of the transformer;

[0054] S22, parameter initialization: The elements composed of the initial weights and thresholds generated by each node of the BP neural network are recorded as set I wi , let the number of ants in the ant population be N, the ant path coordinates be the weights and thresholds of each node in the BP neural network, let the path i→j be from the starting point e through nodes i and j to the end point s, and the pheromone concentration of the path i→j is obtained by non-uniform distribution processing; the calculation formula for the non-uniform distribution processing is:

[0055]

[0056] In the above formula, v ij is the pheromone concentration from node i to node j; d ij is the Euclidean distance between node i and node j; (x i ,y i )、(x j ,y j )、(x e ,y e )、(x s ,y s ) are the coordinates of node i, node j, starting point e, and end point s respectively; d se is the Euclidean distance from the starting point e to the end point s; d ie is the Euclidean distance between node i and starting point e; d sj is the Euclidean distance from node j to end point s; v0 is the initial pheromone concentration;

[0057] S23. Path selection: Place all ants at the starting point and move them according to the following formula:

[0058]

[0059] In the above formula, represents the probability of the kth ant transferring from node i to node j at time t; α is the inspiration factor; β is the expectation factor; v ij (t) is the pheromone concentration on the path from node i to node j at time t; allowed k is the set of next nodes to which the kth ant can move; is a heuristic function that represents the expected degree of ant transfer from node i to node j;

[0060] S24. When all ants have traversed all the elements in the set, the pheromone concentration on the path is updated according to the following formula:

[0061] v ij (t+n)=(1-ρ)v ij (t)+Δv ij (t,t+n)+(h×(LL i→j )) / L i→j ;

[0062] In the above formula, v ij (t+n) is the pheromone concentration on the path from node i to node j at time t+n; ρ is the pheromone volatility coefficient; Δv ij (t, t+n) is the pheromone left by the ant on the path from node i to node j from time t to time t+n; L is the optimal path length of the previous iteration; L i→j is the path length from the current iteration node i to the node j; h is the reinforcement coefficient;

[0063] S25. Calculate the fitness function value of each ant, form a subpopulation of ants whose fitness values ​​are less than a preset threshold, substitute them into the genetic algorithm for crossover mutation, and form a new ant population with the subpopulation after crossover mutation and the ants whose fitness values ​​are greater than or equal to the preset threshold;

[0064] S26, return to step S23 and iterate until the maximum number of iterations is reached, output the ant with the highest fitness value as the optimal solution, and substitute it into the BP neural network as the optimal initial network parameter;

[0065] S27, calculating the network error of the BP neural network;

[0066] S28. Determine whether the network error of the BP neural network is less than the preset error threshold ε0; if so, the training ends; if not, update the network parameters of the BP neural network and return to step S27 to continue training.

[0067] The model training module adjusts the pheromone volatility coefficient ρ according to the following formula:

[0068]

[0069] In the above formula, ρ(t) is the pheromone volatility coefficient at time t; ρ min is the minimum value of the pheromone volatility coefficient; η is the global pheromone decrease factor, η∈(0,1).

[0070] The model training module performs crossover mutation according to the following steps:

[0071] A1. Randomly pair the ants in the subpopulations. Calculate the crossover probability by combining Sigmoid and Gaussian distributions. Perform a crossover on each group of individuals to generate a new pair of individuals. The crossover is performed according to the following formula:

[0072]

[0073] In the above formula, a1 and a2 represent a group of paired individuals; b represents the crossover probability; c1 and c2 represent a group of newly generated individuals after crossover;

[0074] The calculation formula of the crossover probability is:

[0075]

[0076] In the above formula, b max 、b min is the maximum and minimum value of the crossover probability; J max represents the maximum fitness value in the subpopulation; k1 is the curve smoothing parameter; J′ represents the larger fitness value between the two individuals undergoing crossover; J avg represents the average fitness value of the subpopulation;

[0077] A2. Use the roulette wheel rule to select the individual to perform mutation. The probability of an individual being selected is:

[0078]

[0079] In the above formula, P i is the probability of individual i being selected in the subpopulation; f i is the fitness function of individual i;

[0080] A3. Randomly generate two natural numbers m1 and m2 that are greater than 1. M is the total number of genes in the gene coding string of the individual; the genes between m1 and m2 in the gene coding string of the individual selected by A2 are arranged in reverse order to obtain a mutant individual.

[0081] The insulation life assessment model is:

[0082]

[0083] In the above formula, L is the remaining life of transformer insulation; m is the mass fraction of the initial moisture content in the oil-paper; DP is the current degree of polymerization of the oil-paper; DP end is the specified value of the degree of polymerization at the end of life; A is a constant; θ H is the hot spot temperature value.

[0084] In a third aspect, the present invention provides a transformer internal insulation life prediction device based on a BP neural network, the prediction device comprising a memory and a processor;

[0085] The memory is configured to store computer program code and transmit the computer program code to the processor;

[0086] The processor is configured to execute the aforementioned method according to the instructions in the computer program code.

[0087] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, and the computer program implements the aforementioned method when executed by a processor.

[0088] Compared with the prior art, the present invention has the following beneficial effects:

[0089] 1. The transformer internal insulation life prediction method based on a BP neural network described in the present invention first obtains historical transformer operating data and preprocesses it using the principal hierarchy analysis method. The BP neural network model is then trained using the preprocessed transformer historical operating data. At the beginning of the training, an ant colony algorithm combined with a genetic algorithm is used to optimize the initial network parameters of the BP neural network model. The hotspot temperature information of the transformer is predicted based on the trained BP neural network model. The predicted hotspot temperature information is input into an insulation life assessment model to obtain an insulation life assessment value. The insulation life assessment value is converted into a final insulation life prediction value using the Monte Carlo method. The above design uses a genetic algorithm to improve the ant colony algorithm, thereby accelerating search efficiency and avoiding falling into local optimal solutions. Therefore, the present invention can accelerate algorithm search efficiency and avoid falling into local optimal solutions.

[0090] 2. The transformer internal insulation life prediction method based on a BP neural network, described in this invention, addresses the problem of poor initial search efficiency caused by the uniform initial information concentration of the ant colony algorithm. By applying a non-uniform distribution to the initial information concentration—that is, by differentiating the pheromone concentration of each path—this method enables the ant colony to clearly identify its search objectives in the early stages, thereby improving the algorithm's search efficiency. Therefore, this invention can further improve the algorithm's search efficiency.

[0091] 3. The transformer internal insulation life prediction method based on the BP neural network described in the present invention uses adaptive pheromone volatilization rules to control the pheromone volatilization rate, thereby improving the convergence speed, avoiding falling into local optimality, and thus obtaining a global optimal solution. Therefore, the present invention can obtain a global optimal solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 The present invention is a flowchart of the method.

[0093] Figure 2 Schematic diagram of the structure of the system of the present invention.

[0094] Figure 3 It is a structural schematic diagram of the device described in the present invention.

[0095] Figure 4This is a flow chart of initial network parameter optimization and BP neural network training in step S2 of the method of the present invention. DETAILED DESCRIPTION

[0096] The present invention will be further described in detail below with reference to specific embodiments and the accompanying drawings.

[0097] Example 1:

[0098] See also Figure 1 A transformer internal insulation life prediction method based on BP neural network is carried out in the following steps:

[0099] S1. First, collect the historical operating data of the transformer, such as load rate K, harmonic voltage U, harmonic current I, coil temperature Tc, top oil temperature θ top 、Ambient temperature θ a The principal components are extracted by the principal component analysis method (AHP) and basic information such as ambient humidity H, active power P and reactive power Q. The principal components are extracted by the standard principal component analysis method, which mainly includes four steps, namely data standardization, determination of principal components, calculation of principal component variance contribution rate and cumulative variance contribution rate, and selection of principal components. In order to avoid unnecessary information loss, the cumulative variance contribution rate G(l) ≥ 95% is used as the benchmark when determining the number of principal components, and the first l principal components are extracted. The standard principal component analysis method performs data standardization according to the following formula to make the data basically consistent in terms of discreteness, so that the variance of each indicator of the processed sample data is equal and 1:

[0100]

[0101] In the above formula, x ij represents the sample data without data standardization; x′ ij Represents sample data after data standardization; δ Xj Represent the mean and standard deviation of the sample data respectively;

[0102] However, this method ignores the impact of the difference in the degree of dispersion between data on the selection of principal components, resulting in the sample data after standardization processing, further extraction of principal components cannot well contain most of the information of the original data; therefore, averaging processing is used instead of data standardization. The formula for averaging processing is:

[0103]

[0104] S2. Use the pre-processed historical operating data of the transformer to train the BP neural network model; at the beginning of training, use the ant colony algorithm to optimize the initial network parameters of the BP neural network model, and substitute the optimal initial network parameters into the BP neural network to start training; the BP neural network model includes an input layer, a hidden layer and an output layer. The number of input layers and output layers is determined to be 1 layer, but the number of hidden layers is related to the global error size of the network. When the number of layers is large, the complexity of the network structure will increase, and it is easy to fall into the local minimum, which reduces the network training effect. Therefore, it is necessary to reasonably select the number of layers according to the actual problem. In the hot spot temperature prediction, when the number of hidden layers is selected as one, most of the function information can be processed, and the overall comprehensive evaluation effect of the network is better, so the BP neural network structure is selected as three layers; BP neural network The network training mainly includes two stages, namely the forward transmission of sample data and the reverse transmission of errors. In the process of forward transmission of sample data, the input sample data is converted into the relationship between weights and thresholds and transmitted to the hidden layer nodes. Under the action of the activation function, the processed information is output to the output layer, and finally the output is obtained. If the final output does not match the expected output result and the error between the two is greater than or equal to the preset error threshold ε0, the error reverse transmission stage will begin to execute. The error reverse transmission stage uses the error information as the basis and readjusts the weights and thresholds of each layer according to the principle of gradient descent, and transmits the adjusted error information back to the hidden layer and input layer. When the error between the final output and the expected output result is less than the preset error threshold, the training is terminated and the trained BP neural network model is output.

[0105] See also Figure 4 The initial network parameter optimization and BP neural network training steps are specifically as follows:

[0106] S21, inputting pre-processed historical operation data of the transformer;

[0107] S22, parameter initialization: The elements composed of the initial weights and thresholds generated by each node of the BP neural network are recorded as set I wi , let the number of ants in the ant population be N, the path coordinates of the ants be the weights and thresholds of each node in the BP neural network, and the F of the clustering algorithm effect evaluation Measure Function is used as the fitness function, and the path i→j is set from the starting point e through the node i and node j to the end point s;

[0108] When using the traditional ant colony algorithm to improve the BP neural network, the initial pheromone concentration of each path is the same, which will lead to the problem of excessively long model training time. To solve this problem, the initial information concentration is processed with a non-uniform distribution, that is, the pheromone concentration of each path is differentiated, so that the ant colony can clearly determine the search purpose in the initial stage and improve the search efficiency of the algorithm. The calculation formula for the non-uniform distribution processing is:

[0109]

[0110]

[0111] In the above formula, v ij is the pheromone concentration from node i to node j; d ij is the Euclidean distance between node i and node j; (x i ,y i )、(x j ,y j )、(x e ,y e )、(x s ,y s ) are the coordinates of node i, node j, starting point e, and end point s respectively; d se is the Euclidean distance from the starting point e to the end point s; d ie is the Euclidean distance between node i and starting point e; d sj is the Euclidean distance from node j to end point s; v0 is the initial pheromone concentration; the starting point and end point of the ant path in the ant colony algorithm are usually clearly specified by initializing variables to ensure that the algorithm can optimize according to the established goals. In this embodiment, the starting point e and end point s are determined by the initialization parameters;

[0112] S23. Path selection: Place all ants at the starting point and move them according to the following formula:

[0113]

[0114] In the above formula, represents the probability of the kth ant transferring from node i to node j at time t; α is the inspiration factor; β is the expectation factor; v ij (t) is the pheromone concentration on the path from node i to node j at time t; allowed k is the set of next nodes to which the kth ant can move; is a heuristic function that represents the expected degree of ant transfer from node i to node j;

[0115] S24. After all ants have traversed the elements in the set, the pheromone concentration on the path is updated. Since the pheromone update rule of the traditional ant colony algorithm is to update all pheromones on all paths, it cannot form a significant concentration difference and cannot highlight the shortest path or shorter paths. To solve this problem, the pheromone concentration of the shorter path is strengthened and the shorter path is weakened to speed up the operation of the ant algorithm. The pheromone concentration update formula is as follows:

[0116] v ij (t+n)=(1-ρ)vij (t)+Δv ij (t,t+n)+(h×(LL i→j )) / L i→j ;

[0117] In the above formula, v ij (t+n) is the pheromone concentration on the path from node i to node j at time t+n; ρ is the pheromone volatility coefficient; Δv ij (t, t+n) is the pheromone left by the ant on the path from node i to node j from time t to time t+n; L is the optimal path length of the previous iteration; L i→j is the path length from node i to node j in the current iteration; h is the reinforcement coefficient; after each iteration, when L>L i→j , it means that the path of this iteration is shorter than the optimal path of the previous iteration, save the path and strengthen the pheromone update of this path. On the contrary, when L<L i→j , then the update is weakened;

[0118] If the value of ρ is too small, the pheromone will evaporate slowly, and a large amount of pheromone will accumulate on a better path, causing the search to fall into a local optimum. If the value of ρ is too large, the pheromone will evaporate too quickly, and the amount of information remaining on the path will not attract ants to search, slowing down the convergence rate. To solve this problem, an adaptive pheromone evaporation rule is used to control the pheromone evaporation rate, improve the convergence speed, avoid falling into a local optimum, and obtain the global optimal solution. Specifically, the pheromone evaporation coefficient ρ is adjusted according to the following formula:

[0119]

[0120] In the above formula, ρ(t) is the pheromone volatility coefficient at time t; ρ min is the minimum value of the pheromone volatility coefficient; η is the global pheromone decrease factor, η∈(0,1);

[0121] S25. Calculate the fitness function value of each ant. To prevent premature population growth, ants with fitness values ​​less than a preset threshold are grouped into subpopulations. These subpopulations are substituted into a genetic algorithm for crossover mutation. The subpopulations after crossover mutation are combined with ants with fitness values ​​greater than or equal to the preset threshold to form a new ant population. The specific steps of crossover mutation include:

[0122] A1. Randomly pair the ants in the subpopulations. Calculate the crossover probability by combining Sigmoid and Gaussian distributions. Perform a crossover on each group of individuals to generate a new pair of individuals. The crossover is performed according to the following formula:

[0123]

[0124] In the above formula, a1 and a2 represent a group of paired individuals; b represents the crossover probability; c1 and c2 represent a group of newly generated individuals after crossover;

[0125] In order to make the initial convergence speed faster and prevent premature mutation, the crossover probability calculation formula is improved; the improved crossover probability calculation formula is:

[0126]

[0127] In the above formula, b max 、b min is the maximum and minimum value of the crossover probability; J max represents the maximum fitness value in the subpopulation; k1 is the curve smoothing parameter; J′ represents the larger fitness value between the two individuals undergoing crossover; J avg represents the average fitness value of the subpopulation;

[0128] A2. Use the roulette wheel rule to select the individual to perform mutation. The probability of an individual being selected is:

[0129]

[0130] In the above formula, P i is the probability of individual i being selected in the subpopulation; f i is the fitness function of individual i;

[0131] A3. Randomly generate two natural numbers m1 and m2 that are greater than 1. M is the total number of genes in the individual's genetic code string. The genes between m1 and m2 in the genetic code string of the individual selected by A2 are reversed to obtain a mutant individual. This reverse mutation method can increase population diversity, improve search efficiency, and avoid falling into local optimal solutions. For example, if the original individual's genetic code string is (231459867), two points are randomly selected: (231|4598|67), and the mutated individual is: (231|8954|67).

[0132] S26, return to step S23 and iterate until the maximum number of iterations is reached, output the ant with the highest fitness value as the optimal solution; substitute the optimal solution into the BP neural network, and calculate the network error between the expected result and the predicted result;

[0133] S27. Determine whether the network error is less than a preset error threshold ε0; if so, the training ends; if not, return to step S21 to continue training.

[0134] S3. Predict the hotspot temperature information of the transformer based on the trained BP neural network model, and perform a denormalization operation on the prediction result to obtain the hotspot temperature information prediction value;

[0135] S4. Input the hotspot temperature information prediction value into the insulation life evaluation model to obtain an insulation life evaluation value; the insulation life evaluation model is:

[0136]

[0137] In the above formula, L is the remaining life of transformer insulation; m is the mass fraction of the initial moisture content in the oil-paper; DP is the current degree of polymerization of the oil-paper; DP end is the specified value of the degree of polymerization at the end of life; A is a constant, with a value of 9.34×10 5 θ H is the hot spot temperature value;

[0138] S5. In actual applications, due to changes in manufacturing processes, environmental conditions and operating conditions, the parameters of the device will change, and the time required to reach failure will also vary accordingly. The life of the component should usually be expressed using statistical values; therefore, a Monte Carlo simulation method is used to analyze 10,000 insulation life assessment values ​​obtained based on the insulation life assessment model within a 90% confidence interval, and a reliability assessment model at the transformer system level is established. The confidence level value is selected as the final insulation life prediction value.

[0139] Example 2:

[0140] See also Figure 2 , a transformer internal insulation life prediction system based on improved BP neural network, including data acquisition module, model training module, hot spot temperature prediction module, insulation life calculation module;

[0141] The data acquisition module is used to obtain the historical operation data of the transformer and pre-process it using the principal hierarchy analysis method;

[0142] The model training module is used to train the BP neural network model using the preprocessed historical operation data of the transformer. During the training, the ant colony algorithm combined with the genetic algorithm is used to optimize the initial network parameters of the BP neural network model, and then the BP neural network model is trained based on the optimal initial network parameters. The model training module specifically performs the following steps to optimize the initial network parameters and train the BP neural network model:

[0143] S21, inputting pre-processed historical operation data of the transformer;

[0144] S22, parameter initialization: The elements composed of the initial weights and thresholds generated by each node of the BP neural network are recorded as set I wi, let the number of ants in the ant population be N, the ant path coordinates be the weights and thresholds of each node in the BP neural network, let the path i→j be from the starting point e through nodes i and j to the end point s, and the pheromone concentration of the path i→j is obtained by non-uniform distribution processing; the calculation formula for the non-uniform distribution processing is:

[0145]

[0146] In the above formula, v ij is the pheromone concentration from node i to node j; d ij is the Euclidean distance between node i and node j; (x i ,y i )、(x j ,y j )、(x e ,y e )、(x s ,y s ) are the coordinates of node i, node j, starting point e, and end point s respectively; d se is the Euclidean distance from the starting point e to the end point s; d ie is the Euclidean distance between node i and starting point e; d sj is the Euclidean distance from node j to end point s; v0 is the initial pheromone concentration;

[0147] S23. Path selection: Place all ants at the starting point and move them according to the following formula:

[0148]

[0149] In the above formula, represents the probability of the kth ant transferring from node i to node j at time t; α is the inspiration factor; β is the expectation factor; v ij (t) is the pheromone concentration on the path from node i to node j at time t; allowed k is the set of next nodes to which the kth ant can move; is a heuristic function that represents the expected degree of ant transfer from node i to node j;

[0150] S24. When all ants have traversed all the elements in the set, the pheromone concentration on the path is updated according to the following formula:

[0151] v ij (t+n)=(1-ρ)v ij (t)+Δv ij (t,t+n)+(h×(LL i→j )) / L i→j ;

[0152] In the above formula, v ij (t+n) is the pheromone concentration on the path from node i to node j at time t+n; ρ is the pheromone volatility coefficient; Δv ij (t, t+n) is the pheromone left by the ant on the path from node i to node j from time t to time t+n; L is the optimal path length of the previous iteration; L i→j is the path length from the current iteration node i to the node j; h is the reinforcement coefficient;

[0153] Adjust the pheromone volatility coefficient ρ according to the following formula:

[0154]

[0155] In the above formula, ρ(t) is the pheromone volatility coefficient at time t; ρ min is the minimum value of the pheromone volatility coefficient; η is the global pheromone decrease factor, η∈(0,1);

[0156] S25. Calculate the fitness function value of each ant, form a subpopulation with ants whose fitness value is less than a preset threshold, substitute the subpopulation into the genetic algorithm for crossover mutation, and form a new ant population with the subpopulation after crossover mutation and ants whose fitness value is greater than or equal to the preset threshold. The crossover mutation is performed according to the following steps:

[0157] A1. Randomly pair the ants in the subpopulations. Calculate the crossover probability by combining Sigmoid and Gaussian distributions. Perform a crossover on each group of individuals to generate a new pair of individuals. The crossover is performed according to the following formula:

[0158]

[0159] In the above formula, a1 and a2 represent a group of paired individuals; b represents the crossover probability; c1 and c2 represent a group of newly generated individuals after crossover;

[0160] The calculation formula of the crossover probability is:

[0161]

[0162] In the above formula, b max 、b min is the maximum and minimum value of the crossover probability; J max represents the maximum fitness value in the subpopulation; k1 is the curve smoothing parameter; J′ represents the larger fitness value between the two individuals undergoing crossover; J avg represents the average fitness value of the subpopulation;

[0163] A2. Use the roulette wheel rule to select the individual to perform mutation. The probability of an individual being selected is:

[0164]

[0165] In the above formula, P i is the probability of individual i being selected in the subpopulation; f i is the fitness function of individual i;

[0166] A3. Randomly generate two natural numbers m1 and m2 that are greater than 1. M is the total number of genes in the gene coding string of the individual; the genes between m1 and m2 in the gene coding string of the individual selected by A2 are arranged in reverse order to obtain the mutant individual;

[0167] S26, return to step S23 and iterate until the maximum number of iterations is reached, output the ant with the highest fitness value as the optimal solution, and substitute it into the BP neural network as the optimal initial network parameter;

[0168] S27, calculating the network error of the BP neural network;

[0169] S28, determining whether the network error of the BP neural network is less than a preset error threshold ε0; if so, the training ends; if not, the network parameters of the BP neural network are updated, and the process returns to step S27 to continue training;

[0170] The hotspot temperature prediction module is used to predict the hotspot temperature information of the transformer using the trained BP neural network model; the prediction result output by the BP neural network model is denormalized to obtain the hotspot temperature information prediction value;

[0171] The insulation life calculation module is used to first input the hotspot temperature information prediction value into the insulation life evaluation model to obtain the insulation life evaluation value, and then convert the insulation life evaluation value into the final insulation life prediction value based on the Monte Carlo method; the insulation life evaluation model is:

[0172]

[0173] In the above formula, L is the remaining life of transformer insulation; m is the mass fraction of the initial moisture content in the oil-paper; DP is the current degree of polymerization of the oil-paper; DP end is the specified value of the degree of polymerization at the end of life; A is a constant; θ H is the hot spot temperature value.

[0174] Example 3:

[0175] See also Figure 3, a transformer internal insulation life prediction device based on BP neural network, the prediction device includes a memory and a processor; the memory is used to store computer program code and transmit the computer program code to the processor; the processor is used to execute the method described in Example 1 according to the instructions in the computer program code.

[0176] Example 4:

[0177] A computer-readable storage medium stores a computer program, which implements the method described in Example 1 when executed by a processor.

[0178] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0179] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0180] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0181] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0182] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A method for predicting transformer internal insulation life based on BP neural network, characterized by: The prediction method comprises the following steps: S1. Obtain the historical operating data of the transformer and preprocess it using the principal hierarchy analysis method; S2. Using the pre-processed historical operation data of the transformer to train the BP neural network model. During the training, the ant colony algorithm combined with the genetic algorithm is first used to optimize the initial network parameters of the BP neural network model, and then the BP neural network model is trained based on the optimal initial network parameters. S3, predicting the hot spot temperature information of the transformer based on the trained BP neural network model; S4. Inputting the predicted value of the hotspot temperature information into the insulation life assessment model to obtain an insulation life assessment value; S5. Convert the insulation life evaluation value into the final insulation life prediction value based on the Monte Carlo method; The S2 comprises the following steps: S21, inputting pre-processed historical operation data of the transformer; S22, parameter initialization: The elements composed of the initial weights and thresholds generated by each node of the BP neural network are recorded as set I wi , let the number of ants in the ant population be N, the ant path coordinates be the weights and thresholds of each node in the BP neural network, let the path i→j be from the starting point e through nodes i and j to the end point s, and the pheromone concentration of the path i→j is obtained by non-uniform distribution processing; the calculation formula for the non-uniform distribution processing is: In the above formula, v ij is the pheromone concentration from node i to node j; d ij is the Euclidean distance between node i and node j; (x i ,y i )、(x j ,y j )、(x e ,y e )、(x s ,y s ) are the coordinates of node i, node j, starting point e, and end point s respectively; d se is the Euclidean distance from the starting point e to the end point s; d ie is the Euclidean distance between node i and starting point e; d sj is the Euclidean distance from node j to end point s; v0 is the initial pheromone concentration; S23. Path selection: Place all ants at the starting point and move them according to the following formula: In the above formula, represents the probability of the kth ant transferring from node i to node j at time t; α is the inspiration factor; β is the expectation factor; v ij (t) is the pheromone concentration on the path from node i to node j at time t; allowed k is the next node set to which the kth ant moves; is a heuristic function that represents the expected degree of ant transfer from node i to node j; S24. When all ants have traversed all the elements in the set, the pheromone concentration on the path is updated according to the following formula: v ij (t+n)=(1-ρ)v ij (t)+Δv ij (t,t+n)+(h×(LL i→j )) / L i→j ; In the above formula, v ij (t+n) is the pheromone concentration on the path from node i to node j at time t+n; ρ is the pheromone volatility coefficient; Δv ij (t, t+n) is the pheromone left by the ant on the path from node i to node j from time t to time t+n; L is the optimal path length of the previous iteration; L i→j is the path length from the current iteration node i to the node j; h is the reinforcement coefficient; S25. Calculate the fitness function value of each ant, form a subpopulation of ants whose fitness values ​​are less than a preset threshold, substitute them into the genetic algorithm for crossover mutation, and form a new ant population with the subpopulation after crossover mutation and the ants whose fitness values ​​are greater than or equal to the preset threshold; S26, return to step S23 and iterate until the maximum number of iterations is reached, output the ant with the highest fitness value as the optimal solution, and substitute it into the BP neural network as the optimal initial network parameter; S27, calculating the network error of the BP neural network; S28, determining whether the network error of the BP neural network is less than a preset error threshold ε0; if so, the training ends; if not, the network parameters of the BP neural network are updated, and the process returns to step S27 to continue training; The S5 is specifically: performing Monte Carlo simulation analysis on multiple insulation life assessment values ​​obtained based on the insulation life assessment model within the confidence interval, establishing a transformer system-level reliability assessment model, and selecting a confidence level value as the final insulation life prediction value.

2. The transformer internal insulation life prediction method based on BP neural network according to claim 1 is characterized in that: In S24, the pheromone volatility coefficient ρ is adjusted according to the following formula: In the above formula, ρ(t) is the pheromone volatility coefficient at time t; ρ min is the minimum value of the pheromone volatility coefficient; η is the global pheromone decrease factor, η∈(0,1).

3. The transformer internal insulation life prediction method based on BP neural network according to claim 1 is characterized in that: In step S25, the specific steps of crossover mutation include: A1. Randomly pair the ants in the subpopulations. Calculate the crossover probability by combining Sigmoid and Gaussian distributions. Perform a crossover on each group of individuals to generate a new pair of individuals. The crossover is performed according to the following formula: In the above formula, a1 and a2 represent a group of paired individuals; b represents the crossover probability; c1 and c2 represent a group of newly generated individuals after crossover; The calculation formula of the crossover probability is: In the above formula, b max 、b min is the maximum and minimum value of the crossover probability; J max represents the maximum fitness value in the subpopulation; k1 is the curve smoothing parameter; J′ represents the larger fitness value between the two individuals undergoing crossover; J avg represents the average fitness value of the subpopulation; A2. Use the roulette wheel rule to select the individual to perform mutation. The probability of an individual being selected is: In the above formula, P i is the probability of individual i being selected in the subpopulation; f i is the fitness function of individual i; A3. Randomly generate two natural numbers m1 and m2 that are greater than 1. M is the total number of genes in the gene coding string of the individual; the genes between m1 and m2 in the gene coding string of the individual selected by A2 are arranged in reverse order to obtain a mutant individual.

4. The method for predicting transformer internal insulation life based on BP neural network according to any one of claims 1 to 3, characterized in that: The insulation life assessment model is: In the above formula, L is the remaining life of transformer insulation; m is the mass fraction of the initial moisture content in the oil-paper; DP is the current degree of polymerization of the oil-paper; DP end is the specified value of the degree of polymerization at the end of life; A is a constant; θ H is the hot spot temperature value.

5. Transformer internal insulation life prediction system based on BP neural network, characterized by: The prediction system includes a data acquisition module, a model training module, a hotspot temperature prediction module, and an insulation life calculation module; The data acquisition module is used to obtain the historical operation data of the transformer and pre-process it using the principal hierarchy analysis method; The model training module is used to train the BP neural network model using the preprocessed historical operation data of the transformer. During the training, the initial network parameters of the BP neural network model are optimized by using the ant colony algorithm combined with the genetic algorithm, and then the BP neural network model is trained based on the optimal initial network parameters. The hot spot temperature prediction module is used to predict the hot spot temperature information of the transformer based on the trained BP neural network model; The insulation life calculation module is used to input the hot spot temperature information prediction value into the insulation life evaluation model to obtain the insulation life evaluation value, and convert the insulation life evaluation value into the final insulation life prediction value based on the Monte Carlo method; The model training module is used to perform training according to the following steps: S21, inputting pre-processed historical operation data of the transformer; S22, parameter initialization: Set the elements composed of the initial weights and thresholds generated by each node of the BP neural network as set I wi , let the number of ants in the ant population be N, the ant path coordinates be the weights and thresholds of each node in the BP neural network, let the path i→j be from the starting point e through nodes i and j to the end point s, and the pheromone concentration of the path i→j is obtained by non-uniform distribution processing; the calculation formula for the non-uniform distribution processing is: In the above formula, v ij is the pheromone concentration from node i to node j; d ij is the Euclidean distance between node i and node j; (x i ,y i )、(x j ,y j )、(x e ,y e )、(x s ,y s ) are the coordinates of node i, node j, starting point e, and end point s respectively; d se is the Euclidean distance from the starting point e to the end point s; d ie is the Euclidean distance between node i and starting point e; d sj is the Euclidean distance from node j to end point s; v0 is the initial pheromone concentration; S23. Path selection: Place all ants at the starting point and move them according to the following formula: In the above formula, represents the probability of the kth ant transferring from node i to node j at time t; α is the inspiration factor; β is the expectation factor; v ij (t) is the pheromone concentration of node i and node j; allowed k is the next node set to which the kth ant moves; is a heuristic function that represents the expected degree of ant transfer from node i to node j; S24. When all ants have traversed all the elements in the set, the pheromone concentration on the path is updated according to the following formula: v ij (t+n)=(1-ρ)v ij (t)+Δv ij (t,t+n)+(h×(LL i→j )) / L i→j ; In the above formula, v ij (t) is the pheromone concentration on the path from node i to node j at time t; ρ is the pheromone volatility coefficient; Δv ij (t, t+n) is the pheromone left by the ant on the path from node i to node j from time t to time t+n; L is the optimal path length of the previous iteration; L i→j is the path length from the current iteration node i to the node j; h is the reinforcement coefficient; S25. Calculate the fitness function value of each ant, form a subpopulation of ants whose fitness values ​​are less than a preset threshold, substitute them into the genetic algorithm for crossover mutation, and form a new ant population with the subpopulation after crossover mutation and the ants whose fitness values ​​are greater than or equal to the preset threshold; S26, return to step S23 and iterate until the maximum number of iterations is reached, output the ant with the highest fitness value as the optimal solution, and substitute it into the BP neural network as the optimal initial network parameter; S27, calculating the network error of the BP neural network; S28. Determine whether the network error of the BP neural network is less than the preset error threshold ε0; if so, the training ends; if not, update the network parameters of the BP neural network and return to step S27 to continue training.

6. The transformer internal insulation life prediction system based on BP neural network according to claim 5 is characterized in that: The model training module adjusts the pheromone volatility coefficient ρ according to the following formula: In the above formula, ρ(t) is the pheromone volatility coefficient at time t; ρ min is the minimum value of the pheromone volatility coefficient; η is the global pheromone decrease factor, η∈(0,1).

7. The transformer internal insulation life prediction system based on BP neural network according to claim 5 is characterized in that: The model training module performs crossover mutation according to the following steps: A1. Randomly pair the ants in the subpopulations. Calculate the crossover probability by combining Sigmoid and Gaussian distributions. Perform a crossover on each group of individuals to generate a new pair of individuals. The crossover is performed according to the following formula: In the above formula, a1 and a2 represent a group of paired individuals; b represents the crossover probability; c1 and c2 represent a group of newly generated individuals after crossover; The calculation formula of the crossover probability is: In the above formula, b max 、b min is the maximum and minimum value of the crossover probability; J max represents the maximum fitness value in the subpopulation; k1 is the curve smoothing parameter; J′ represents the larger fitness value between the two individuals undergoing crossover; J avg represents the average fitness value of the subpopulation; A2. Use the roulette wheel rule to select the individual to perform mutation. The probability of an individual being selected is: In the above formula, P i is the probability of individual i being selected in the subpopulation; f i is the fitness function of individual i; A3. Randomly generate two natural numbers m1 and m2 that are greater than 1. M is the total number of genes in the gene coding string of the individual; the genes between m1 and m2 in the gene coding string of the individual selected by A2 are arranged in reverse order to obtain a mutant individual.

8. The transformer internal insulation life prediction system based on BP neural network according to any one of claims 5 to 7, characterized in that: The insulation life assessment model is: In the above formula, L is the remaining life of transformer insulation; m is the mass fraction of the initial moisture content in the oil-paper; DP is the current degree of polymerization of the oil-paper; DP end is the specified value of the degree of polymerization at the end of life; A is a constant; θ H is the hot spot temperature value.

Images