A method and system for predicting healthy operation time of transformer

Abstract

The present invention relates to a method and system for predicting the healthy operation time of a transformer, which belongs to the field of data preprocessing and identification technology, and solves the problem of accuracy of maintenance plan decision-making in existing methods. The method includes collecting index data and status data of multiple transformers and performing noise reduction processing; labeling multiple transformers based on the index data with noise reduction processing to obtain transformers of multiple levels; selecting transformers of the same level from the transformers of multiple levels, building a prediction model based on historical status data, and using the prediction model to predict the trend of real-time status data; judging the future health status of the transformer based on the feature vector and the health status cluster center and determining the healthy operation time of the transformer, extracting features from the indicator data of the transformer based on historical fault data to construct a feature vector and clustering the feature vector to obtain the health status cluster center. Determine the expected operation time of the transformer under healthy operation conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of data preprocessing and identification, and in particular to a method and system for predicting the healthy operation time of a transformer. Background Art

[0002] The distribution transformer is a key device to ensure the normal operation of the entire distribution network. The stability of its performance is crucial to improving the reliability and stability of power supply. Once a transformer fails, it may cause a large-scale power outage, which will not only affect the normal production of the enterprise, but also bring inconvenience to people's daily life, and thus have a negative impact on the economic benefits of the entire society.

[0003] Although many distribution transformers are now equipped with online status monitoring and fault alarm functions, the existing fault management mode still has some hidden dangers, which may pose a potential threat to the safe and stable operation of the distribution network, affect the service life of the distribution transformer, and cause unnecessary economic losses. In order to solve this problem, by predicting the performance change trend of the distribution transformer, the health management of the transformer is transformed from the traditional fault management mode to a more intelligent decay management mode. Through predictive maintenance, the continuous and reliable operation of the distribution transformer can be achieved, the probability of equipment failure can be reduced, and unnecessary planned maintenance can be avoided. Therefore, an effective health management mode plays an important role in promoting the intelligent development of distribution transformers and can significantly improve the operating reliability of distribution transformers. Summary of the invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide a method and system for predicting the healthy operation time of a transformer, so as to solve the problem of accuracy of maintenance plan decision-making in existing methods.

[0005] On the one hand, an embodiment of the present invention provides a method and system for predicting the healthy operation time of a transformer, comprising: collecting index data and status data of multiple transformers and performing noise reduction processing on the index data and the status data; labeling the multiple transformers based on the index data processed with noise reduction to obtain transformers of multiple levels, wherein the different index data correspond to energy efficiency level, load capacity, short-circuit resistance, insulation level and voltage regulation capability; selecting transformers of the same level from the transformers of the multiple levels, constructing a prediction model based on historical status data, and using the prediction model to perform trend prediction on real-time status data to obtain a feature vector; and judging the future health status of the transformer based on the feature vector and the health status cluster center and determining the healthy operation time of the transformer, wherein, based on historical fault data, feature extraction is performed on the index data of the transformer to construct a feature vector and cluster analysis is performed on the feature vector to obtain a health status cluster center.

[0006] The beneficial effects of the above technical solution are as follows: The transformer healthy operation time prediction method can ensure that the information input to the model is relevant and effective by carefully selecting the feature vectors related to the transformer operating state. This helps to reduce the interference of irrelevant information and improve the accuracy of the prediction model. The GLSTM model can effectively capture the time dependence and nonlinear relationship in the transformer operating state feature vector, thereby improving the prediction accuracy. The changes in the transformer operating state feature vector can be adaptively learned. When the operating state of the transformer changes, the model can automatically adjust its parameters to adapt to the new data pattern. By combining historical fault information and real-time feature vector data, a comprehensive and integrated analysis of the transformer health state can be achieved. This method not only takes into account the current operating state of the transformer, but also combines its past behavior and known fault modes, so as to provide more accurate diagnostic results. Further combined with the calculation of the health index, the expected operating time of the transformer under healthy and stable operating conditions can be calculated. This is very valuable for formulating maintenance plans, optimizing resource allocation, and predicting equipment life.

[0007] Based on the further improvement of the above method, the noise reduction processing of the indicator data and the status data further includes: performing M initialization complex data empirical mode decomposition CEMD, adding white noise with an amplitude of k for a certain number of times; n (z) Add the original data sequence x0(z) of the indicator data and the state data to form a complex signal x c (z) = x0(z) + ix n (z), where z is a time series variable; projecting the complex signal onto On, among them, is the projection direction, 1≤k≤N:

[0008]

[0009] Substituting Euler's formula into the above formula:

[0010]

[0011] when When , the direction of the original extreme value selection changes, and the solution is The maximum point of the projection direction is obtained by performing cubic spline interpolation on it; the upper and lower envelopes in each projection direction are obtained; the average value of the boundary envelope m(z) is calculated, and then the component signals h(z)=x are obtained according to the average value of the boundary envelope m(z). c(z)-m(z); obtain the mean of the IMF component and the residual; and calculate the mean of each IMF component obtained after M times of complex data empirical mode decomposition CEMD process as the indicator data and status data after denoising.

[0012] Based on further improvement of the above method, the different indicator data include: the indicator data corresponding to the energy efficiency level include winding insulation resistance, bushing external insulation creepage distance, body oil withstand voltage and bushing main insulation resistance; the indicator data corresponding to the load capacity include: cooler fault and full stop signal, bushing oil color spectrum and core grounding current; the indicator data corresponding to the short-circuit resistance include: low voltage short-circuit impedance, number of short-circuit impacts, hidden dangers of insufficient short-circuit resistance and whether the medium voltage side is half capacity; the indicator data corresponding to the insulation level include: equipment operating years, annual heavy overload time, no-load loss and energy efficiency and real-time oil temperature of the main transformer; and the indicator data corresponding to the voltage regulation capacity include: number of on-load tap changer operations and on-load tap changer over-cycle uninspected.

[0013] Based on the further improvement of the above method, the label levels of the multiple transformers are divided based on different indicator data to obtain transformers of the same level, including: constructing a portrait fact label based on the energy efficiency level, the load capacity, the short-circuit resistance, the insulation level and the voltage regulation capability; identifying and extracting the principal component feature label through an improved particle swarm algorithm, wherein the principal component feature label is used to reflect the performance and status information of the transformer; constructing a model label based on the principal component feature label, wherein the model label includes environmental status attributes, operating status attributes and equipment body attributes; and integrating the portrait fact label and the model label into the portrait system to select the transformers of the same level.

[0014] Based on the further improvement of the above method, the main component feature labels are identified and extracted by improving the particle swarm algorithm, wherein the main component feature labels are used to reflect the performance and status information of the transformer, including: Step 1: Set the following initial parameters: particle swarm size N, sample feature dimension d, cluster number K, maximum number of iterations T, position boundary x min 、x max , and the maximum speed v max; Step 2: Generate a chaotic number through chaotic mapping, then map the chaotic number to a position interval, construct the reverse position of the particle through reverse learning, and compare the fitness of the particle and its reverse position, so that the position corresponding to the better fitness is the initial position of the particle; Step 3: Start searching from the initial position of the particle, update the particle through normal iteration, otherwise use the Cauchy mutation particle position and perform boundary control; Step 4: Divide the denoised index data set according to the nearest neighbor rule, and recalculate the fitness of the particle; Step 5: Compare and update the optimal position and optimal fitness value of the individual and group particles; Step 6: If the iteration stop condition is met, proceed to step 7, otherwise return to step 3 to continue iterating; Step 7: Divide the denoised index data according to the nearest neighbor rule with the optimal solution as the clustering center to obtain K clusters, and complete the clustering of the multiple transformers.

[0015] Based on the further improvement of the above method, the particle speed and position are updated by the following formula:

[0016]

[0017] ω=ω max -(ω max -ω min )×q / T;

[0018] in, represents the d-dimensional velocity of particle i at the t+1th iteration, and the unit is consistent with the data; c1 and c2 represent learning factors, which represent the degree of cognition of the particle to itself and the group; r1 and r2 represent random numbers between (0,1); represents the d-th dimension position of particle i at the t+1th iteration; P id , g id Represent the optimal solutions of the current particle and group respectively; ω represents the inertia weight; q represents the current number of iterations, and T represents the maximum number of iterations; the chaotic number is generated by the chaotic mapping formula:

[0019]

[0020] x′=x max +x min -x;

[0021] Among them, z k is the chaotic number generated for the kth time, and β is taken as (0,1); x represents the indicator data; x′ represents the indicator data x The reverse solution of x max 、x min The indicator data are x The upper and lower limits of the value; the ability of particles to jump out of the local optimum is enhanced through the following Cauchy mutation formula;

[0022] x′=x×(1-tan(π(u-0.5));

[0023] Among them, x′ represents the position after mutation; u Represents a random number in the interval (0,1); the particle fitness function uses the clustering evaluation index error square sum SSE:

[0024]

[0025] Where K represents the number of clusters; μ i Represents cluster C i The cluster center of x Represents cluster C i The smaller the SSE value, the tighter the cluster is and the better the clustering effect is.

[0026] Based on the further improvement of the above method, constructing a prediction model based on the historical status data of transformers of the same level further includes: taking the historical status data of the faulty / defective transformer and the monitoring data during the fault period as samples, determining the correlation between each status data and different fault sets according to the confidence level; selecting state vectors with high correlation according to the principle of the confidence minimum threshold method; forming a time series matrix of the state vectors, environmental meteorological data and power grid operation data of the transformer in chronological order; and constructing a long short-term memory network as the prediction model based on the time series matrix, and then using the real-time status data, real-time environmental meteorological data and real-time power grid operation data as inputs of the prediction model to predict the status data at future moments.

[0027] Based on the further improvement of the above method, judging the future health status of the transformer based on the feature vector and the health status cluster center and determining the stable operation time of the transformer further includes: based on the historical fault data, extracting features of the indicator data of the transformer in the normal state and the fault state to construct a feature vector; using the feature vector as input data to train the k-medoids clustering model to obtain a set of cluster centers; judging the operating status of the transformer based on the feature vector and the health status cluster center and converting the operating status of the transformer into the transformer health confidence; determining the weight of each indicator according to the hierarchical analysis method, and using the indicator weight to perform weighted calculation on the transformer health confidence to obtain the final transformer health confidence; and calculating the stable operation time of the transformer according to the final transformer health confidence.

[0028] On the other hand, an embodiment of the present invention provides a transformer healthy operation time prediction system, including: a data acquisition module, used to collect index data and status data of multiple transformers and a processing module; a noise reduction processing module, used to perform noise reduction processing on the index data and the status data; a grade classification module, used to label the multiple transformers based on the index data processed by the noise reduction to obtain transformers of multiple grades, wherein the different index data correspond to energy efficiency grade, load capacity, short-circuit resistance, insulation level and voltage regulation capacity; a state trend prediction module, used to select transformers of the same grade from the transformers of the multiple grades, build a prediction model based on historical state data, and use the prediction model to perform trend prediction on real-time state data to obtain a feature vector; and a stable operation time determination module, used to judge the future health state of the transformer based on the feature vector and the health state cluster center and determine the healthy operation time of the transformer, wherein, based on the historical fault data, feature extraction is performed on the index data of the transformer to construct a feature vector and cluster analysis is performed on the feature vector to obtain a health state cluster center.

[0029] Based on the further improvement of the above system, the noise reduction processing module is used to: perform M initialization complex data empirical mode decomposition CEMD, add white noise with amplitude k for a certain number of times; convert the white noise x n (z) Add the original data sequence x0(z) of the indicator data and the state data to form a complex signal x c (z) = x0(z) + ix n (z), where z is a time series variable; projecting the complex signal onto On, among them, is the projection direction, 1≤k≤N:

[0030]

[0031] Substituting Euler's formula into the above formula:

[0032]

[0033] when When , the direction of the original extreme value selection changes, and the solution is The maximum point of the projection direction is obtained by performing cubic spline interpolation on it; the upper and lower envelopes in each projection direction are obtained; the average value of the boundary envelope m(z) is calculated, and then the component signals h(z)=x are obtained according to the average value of the boundary envelope m(z). c (z)-m(z); obtain the mean of the IMF component and the residual; and calculate the mean of each IMF component obtained after M times of complex data empirical mode decomposition CEMD process as the indicator data and status data after denoising.

[0034] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0035] 1. Combined with the important parameters of the transformer, a portrait fact label system based on energy efficiency level, load capacity, short-circuit resistance, insulation level, and voltage regulation capability is constructed. The portrait labels are further clustered through the particle swarm algorithm to form model label attributes in three dimensions: environmental state attributes, operating state attributes, and equipment body attributes. Different label attributes are divided into five levels. According to the actual level of the transformer, different types of transformer sets are formed to improve the scientificity and rationality of subsequent fault judgment and state prediction.

[0036] 2. When collecting data, there may be certain data errors. The CEMD method can effectively separate the noise components from the original data, thereby improving the signal-to-noise ratio of the data. The denoised data can better reflect the true characteristics and trends of the data, thereby providing more reliable and accurate information when conducting data mining, pattern recognition or predictive analysis.

[0037] 3. By carefully selecting the feature vectors related to the transformer operating status, it can be ensured that the information input to the model is relevant and valid. This helps to reduce the interference of irrelevant information and improve the accuracy of the prediction model. The GLSTM model is an advanced recurrent neural network (RNN) architecture that is particularly suitable for processing data with time series characteristics. It can effectively capture the time dependency and nonlinear relationship in the transformer operating status feature vector, thereby improving the prediction accuracy. The changes in the transformer operating status feature vector can be adaptively learned. When the operating status of the transformer changes, the model can automatically adjust its parameters to adapt to the new data pattern.

[0038] 4. By combining historical fault information and real-time feature vector data, a comprehensive and integrated analysis of the transformer health status can be achieved. This approach not only takes into account the current operating status of the transformer, but also combines its past behavior and known failure modes, thereby providing more accurate diagnostic results. Further combined with the calculation of the health index, the expected operating time of the transformer under healthy and stable operating conditions can be calculated. This is very valuable for formulating maintenance plans, optimizing resource allocation, and predicting equipment life.

[0039] 5. There is no fixed method to determine the best operating time of the transformer, but it needs to be considered comprehensively according to the specific situation and operating environment of the transformer. In actual operation, a specific operation plan can be formulated according to factors such as the rated capacity, operating environment and operating status of the transformer, and the equipment should be inspected and maintained regularly to ensure its safe and stable operation.

[0040] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be achieved and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0042] Figure 1 is a flow chart of a method for predicting healthy operation time of a transformer according to an embodiment of the present invention;

[0043] Figure 2 A schematic diagram of hierarchical analysis structure division according to an embodiment of the present invention;

[0044] Figure 3 A specific flow chart of a method for predicting healthy operation time of a transformer according to an embodiment of the present invention; and

[0045] Figure 4 FIG. 4 is a block diagram of a system for predicting the healthy operation time of a transformer according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0047] refer to Figure 1 A specific embodiment of the present invention discloses a method for predicting the healthy operation time of a transformer, comprising: in step S101, collecting index data and status data of multiple transformers and performing noise reduction processing on the index data and the status data; in step S102, labeling and classifying multiple transformers based on the index data subjected to noise reduction processing to obtain transformers of multiple levels, wherein different index data correspond to energy efficiency level, load capacity, short-circuit resistance, insulation level and voltage regulation capacity; in step S103, selecting transformers of the same level from transformers of multiple levels, constructing a prediction model based on historical status data, and using the prediction model to perform trend prediction on real-time status data to obtain a feature vector; and in step S104, judging the future health state of the transformer and determining the healthy operation time of the transformer based on the feature vector and the health state cluster center, wherein, based on the historical fault data, feature extraction is performed on the index data of the transformer to construct a feature vector and cluster analysis is performed on the feature vector to obtain the health state cluster center.

[0048] Compared with the prior art, the transformer health operation time prediction method provided in this embodiment can ensure that the information input to the model is relevant and effective by carefully selecting the feature vectors related to the transformer operation state. This helps to reduce the interference of irrelevant information and improve the accuracy of the prediction model. The GLSTM model can effectively capture the time dependence and nonlinear relationship in the transformer operation state feature vector, thereby improving the prediction accuracy. The changes in the transformer operation state feature vector can be adaptively learned. When the operating state of the transformer changes, the model can automatically adjust its parameters to adapt to the new data mode. By combining historical fault information and real-time feature vector data, a comprehensive and integrated analysis of the transformer health state can be achieved. This method not only takes into account the current operating state of the transformer, but also combines its past behavior and known fault modes, so as to provide more accurate diagnostic results. Further combined with the calculation of the health index, the expected operating time of the transformer under healthy and stable operating conditions can be calculated. This is very valuable for formulating maintenance plans, optimizing resource allocation, and predicting equipment life.

[0049] Hereinafter, each step of the method for predicting the healthy operation time of a transformer according to an embodiment of the present invention is described in detail. The method for predicting the healthy operation time of a transformer includes the following steps S101 to S104.

[0050] In step S101, index data and status data of a plurality of transformers are collected and noise reduction processing is performed on the index data and status data.

[0051] Specifically, the noise reduction process for the indicator data and the status data further includes: performing M initialization complex data empirical mode decomposition CEMD, adding white noise with an amplitude of k for each number of times; converting the white noise x n (z) Add the original data sequence x0(z) of the indicator data and status data to form a complex signal x c (z) = x0(z) + ix n (z), where z is a time series variable; project the complex signal onto On, among them, is the projection direction, 1≤k≤N:

[0052]

[0053] Substituting Euler's formula into the above formula:

[0054]

[0055] when When , the original extreme value selection direction changes ("the original extreme value selection direction changes" usually means that when processing signals, the selection of extreme points no longer follows the traditional rising or falling direction, but is determined according to a certain projection direction or transformed signal characteristics). Solve The maximum point of the projection direction is obtained by performing cubic spline interpolation on it; the upper and lower envelopes in each projection direction are obtained; the average value of the boundary envelope m(z) is calculated, and then the component signals h(z)=x are obtained according to the average value of the boundary envelope m(z) c (z)-m(z); get the mean of the IMF component and the residual (the mean of the residual can reflect the average level of the remaining part of the signal. If the residual mean is close to zero, it may mean that most of the signal components have been effectively extracted as IMFs; if the residual mean is large, it may indicate that there are still some components in the original signal that have not been fully explained); and calculate the mean of each IMF component obtained after the M-time complex data empirical mode decomposition CEMD process as the indicator data and status data after denoising.

[0056] In step S102, multiple transformers are labeled and graded based on the index data processed by noise reduction to obtain transformers of multiple grades, wherein different index data correspond to energy efficiency grade, load capacity, short-circuit resistance, insulation level and voltage regulation capability.

[0057] Specifically, different indicator data include: indicator data corresponding to energy efficiency level include winding insulation resistance, bushing external insulation creepage distance, body oil withstand voltage and bushing main insulation resistance; indicator data corresponding to load capacity include: cooler failure and full stop signal, bushing oil color spectrum and core grounding current; indicator data corresponding to short-circuit resistance include: low voltage short-circuit impedance, number of short-circuit impacts, hidden dangers of insufficient short-circuit resistance and whether the medium voltage side is half capacity; indicator data corresponding to insulation level include: equipment operating life, annual heavy overload time, no-load loss and energy efficiency and real-time oil temperature of main transformer; and indicator data corresponding to voltage regulation capacity include: number of on-load tap changer operations and on-load tap changer over-cycle failure to inspect.

[0058] The labeling of multiple transformers based on different indicator data to obtain transformers of the same level includes: constructing portrait fact labels based on energy efficiency level, load capacity, short-circuit resistance, insulation level and voltage regulation capability; identifying and extracting principal component feature labels through improved particle swarm algorithm, wherein the principal component feature labels are used to reflect the performance and status information of the transformer; constructing model labels based on the principal component feature labels, wherein the model labels include environmental status attributes, operating status attributes and equipment body attributes; and integrating portrait fact labels and model labels into the portrait system to select transformers of the same level.

[0059] By improving the particle swarm algorithm, the main component feature labels are identified and extracted. The main component feature labels are used to reflect the performance and status information of the transformer, including:

[0060] Step 1: Set the following initial parameters: particle swarm size N, sample feature dimension d, number of clusters K, maximum number of iterations T, position boundary x min 、x max , and the maximum speed v max .

[0061] Step 2: Generate chaotic numbers through chaos mapping, and then map the chaotic numbers to the position interval. At the same time, construct the reverse position of the particle through reverse learning, and compare the fitness of the particle and its reverse position, so that the position corresponding to the better fitness is the initial position of the particle. Generate chaotic numbers through chaos mapping formula:

[0062]

[0063] x′=x max +x min -x;

[0064] Among them, z k is the chaotic number generated for the kth time (the chaotic number can be used to initialize the position of the particle, making the distribution of the particles in the search space more uniform and extensive, thus helping the algorithm to find the global optimal solution. The position and velocity update of the particle is the core of the PSO algorithm. By continuously adjusting the position and velocity of the particle, the particle can gradually approach the optimal solution of the problem), and β is taken as (0,1); x represents indicator data; x′ represents indicator data x The reverse solution of x max 、x min The indicator data x The upper and lower limits of the value.

[0065] Step 3: Start searching from the initial position of the particle and update the particle through normal iteration, otherwise mutate the particle position through Cauchy and perform boundary control.

[0066] Update particle speed and position by the following formula:

[0067]

[0068] ω=ω max -(ω max -ω min )×q / T;

[0069] in, represents the d-dimensional velocity of particle i at the t+1th iteration, and the unit is consistent with the data; c1 and c2 represent learning factors, which represent the degree of cognition of the particle to itself and the group; r1 and r2 represent random numbers between (0,1); represents the d-th dimension position of particle i at the t+1th iteration; P id , g id They represent the optimal solutions of the current particle and the group respectively; ω represents the inertia weight; q represents the current number of iterations, and T represents the maximum number of iterations.

[0070] The ability of particles to escape from local optimality is enhanced by the following Cauchy mutation formula:

[0071] x′=x×(1-tan(π(u-0.5));

[0072] Among them, x′ represents the position after mutation; u represents a random number in the interval (0,1).

[0073] Step 4: Divide the denoised index data set according to the nearest neighbor rule and recalculate the fitness of the particles.

[0074] Step 5: Compare and update the optimal position and optimal fitness value of individual particles and the group.

[0075] Step 6: If the iteration stop condition is met, proceed to step 7, otherwise return to step 3 to continue iteration.

[0076] Step 7: Use the optimal solution as the cluster center to divide the denoised index data according to the nearest neighbor rule to obtain K clusters, and complete the clustering of multiple transformers. The particle fitness function uses the clustering evaluation index error square sum SSE:

[0077]

[0078] Where K represents the number of clusters; μ i Represents cluster C i The cluster center of i The smaller the SSE value, the tighter the cluster is and the better the clustering effect is.

[0079] In step S103, transformers of the same level are selected from transformers of multiple levels, a prediction model is constructed based on historical status data, and the prediction model is used to perform trend prediction on real-time status data to obtain feature vectors. Through status data prediction, historical fault information is combined with clustering of status data, and then the operating status is judged.

[0080] Specifically, constructing a prediction model based on the historical status data of transformers of the same level further includes: taking the historical status data of the faulty / defective transformer and the monitoring data during the fault period as samples, determining the correlation between each status data and different fault sets according to the confidence level; selecting state vectors with high correlation according to the principle of the confidence minimum threshold method; forming a time series matrix of the transformer's state vectors, environmental meteorological data and power grid operation data in chronological order; and constructing a long short-term memory network as a prediction model based on the time series matrix, and then using the real-time status data, real-time environmental meteorological data and real-time power grid operation data as inputs of the prediction model to predict the status data at future moments.

[0081] In step S104, the future health status of the transformer is judged and the healthy operation time of the transformer is determined based on the feature vector and the health status cluster center, wherein, based on the historical fault data, feature extraction is performed on the indicator data of the transformer to construct a feature vector, and cluster analysis is performed on the feature vector to obtain the health status cluster center.

[0082] Specifically, judging the future health status of the transformer based on the feature vector and the health status cluster center and determining the stable operation time of the transformer further includes: based on the historical fault data, extracting features of the indicator data of the transformer in the normal state and the fault state to construct a feature vector; training the k-medoids clustering model with the feature vector as input data to obtain a set of cluster centers; judging the operating status of the transformer based on the feature vector and the health status cluster center and converting the operating status of the transformer into the transformer health confidence; determining the weight of each indicator according to the hierarchical analysis method, and using the indicator weight to perform weighted calculation on the transformer health confidence to obtain the final transformer health confidence; and calculating the stable operation time of the transformer according to the final transformer health confidence.

[0083] Another specific embodiment of the present invention discloses a healthy operation time prediction system for a transformer, including: a data acquisition module 401, used to collect index data and status data of multiple transformers and a processing module; a noise reduction processing module 402, used to perform noise reduction processing on the index data and status data; a grade classification module 403, used to perform label grade classification on multiple transformers based on the index data processed by noise reduction to obtain transformers of multiple grades, wherein different index data correspond to energy efficiency grade, load capacity, short-circuit resistance, insulation level and voltage regulation capacity; a state trend prediction module 404, used to select transformers of the same grade from transformers of multiple grades, build a prediction model based on historical state data, and use the prediction model to perform trend prediction on real-time state data to obtain a feature vector; and a stable operation time determination module 405, used to judge the future health state of the transformer based on the feature vector and the health state cluster center and determine the healthy operation time of the transformer, wherein, based on the historical fault data, feature extraction is performed on the index data of the transformer to construct a feature vector and cluster analysis is performed on the feature vector to obtain the health state cluster center.

[0084] The noise reduction processing module 402 is used to perform M initialization complex data empirical mode decomposition CEMD, adding white noise with an amplitude of k for each number of times; n (z) Add the original data sequence x0(z) of the indicator data and status data to form a complex signal x c (z) = x0(z) + ix n (z), where z is a time series variable; project the complex signal onto On, among them, is the projection direction, 1≤k≤N: Substituting Euler's formula into the above formula:

[0085]

[0086] when When , the direction of the original extreme value selection changes, and the solution is The maximum point of the projection is obtained by performing cubic spline interpolation on it to obtain the upper and lower envelopes in each projection direction; the average value of the boundary envelope m(z) is calculated, and then the component signals h(z)=x are obtained according to the average value of the boundary envelope m(z) c (z)-m(z); obtain the mean of the IMF component and the residual; and calculate the mean of each IMF component obtained after M times of complex data empirical mode decomposition CEMD process as the indicator data and status data after denoising.

[0087] In the following, reference Figure 2 and Figure 3, the specific details of the method for predicting the healthy operation time of a transformer according to an embodiment of the present invention are described in detail by way of a specific example.

[0088] refer to Figure 2 ,Firstly, a transformer operation profile is constructed and ,combined with the improved particle swarm algorithm, a label portrait system is constructed based on ,transformer energy efficiency grade, load capacity, short circuit resistance capability, insulation level, voltage regulation capability, etc. to identify ,transformers with the same profile for classification and analysis.

[0089] First, we construct a transformer operation portrait, and combine it with the improved particle swarm algorithm to build a portrait system based on labels such as transformer energy efficiency level, load capacity, short-circuit resistance, insulation level, and voltage regulation capability. We identify transformers with the same portrait for classification analysis (the above are five portrait fact labels, including energy efficiency level, load capacity, short-circuit resistance, insulation level, and voltage regulation capability. Further, through the clustering method, we form three model labels: environmental state attributes, operating state attributes, and equipment body attributes. As described below, each label is divided into five levels, which is equivalent to each label having five levels, and transformers with the same level of five labels are classified into one category).

[0090] (1) Construct a portrait label system. The constructed transformer operation status portrait system constructs a transformer operation information label library from five capability dimensions: insulation level, load capacity, short-circuit resistance, energy efficiency level, and voltage regulation capacity. The specific label library is shown in Table 1.

[0091]

[0092] (2) Construction of a portrait system based on an improved particle swarm algorithm. Based on the fact label system (insulation level, load capacity, short-circuit resistance, energy efficiency level and voltage regulation capability), the improved particle swarm algorithm is used to identify and extract the main component feature labels, further mine the multi-attribute features, and construct model labels (environmental state attributes, operating state attributes, and equipment attributes) to achieve the design of the portrait system. The specific steps are to use the improved particle swarm algorithm to identify and extract the main component feature labels from the data. These feature labels can reflect the performance and status information of the transformer to the greatest extent. Construct model labels: Based on the extracted feature labels, construct model labels, including environmental state attributes, operating state attributes, and equipment attributes. These model labels are further abstractions and generalizations of the transformer performance and status. Portrait system design: Integrate fact labels and model labels into the user portrait system. This portrait system should be able to comprehensively and accurately reflect the performance and status information of the transformer and support various application scenarios of operation and maintenance management and optimization decisions.

[0093] 1) Particle Swarm Optimization: PSO is an optimization algorithm that simulates the foraging behavior of bird flocks. In this algorithm, each particle has two attributes: speed and position. The position represents the solution to the problem. Each iteration is based on the optimization fitness function. The particle speed and position are updated according to the following formula:

[0094]

[0095] Where: represents the d-dimensional velocity of particle i at the t+1th iteration, and the unit is consistent with the data; c1 and c2 represent learning factors, which represent the degree of cognition of the particle to itself and the group; r1 and r2 represent random numbers between (0,1); represents the d-th dimension position of particle i at the t+1th iteration; P id , g id Respectively represent the optimal solution of the current particle and the group; ω represents the inertia weight. The larger the value, the stronger the global optimization ability, and vice versa. Dynamic ω can obtain better optimization results than fixed values. The linear decreasing strategy is currently used more. The formula is as follows: q represents the current number of iterations, and T is the maximum number of iterations.

[0096] ω=ω max -(ω max -ω min )×q / T;

[0097] 2) Particle swarm algorithm integrating chaos mapping, reverse learning and Cauchy mutation As the standard particle swarm algorithm iterates, the population diversity will decrease, and it is easy to prematurely mature and fall into the local optimum. This paper improves the standard particle swarm algorithm as follows:

[0098] Chaotic mapping and reverse learning initialization: Chaotic mapping has randomness, ergodicity and regularity. Compared with simple random values, the data distribution is more uniform and extensive. This paper adopts Tent mapping, and the mapping function is as follows: k is the chaotic number generated for the kth time, and β is taken as (0,1).

[0099]

[0100] The reverse learning strategy is based on the current solution, generates new individuals by constructing a reverse solution, and compares and retains better solutions. The construction method of the reverse solution is as follows:

[0101] x′=x max +x min -x;

[0102] Where: x represents the original data (including specific indicators represented by insulation level, load capacity, short-circuit resistance, energy efficiency level and voltage regulation capacity); x′ represents the reverse solution of x; x max、x min are the upper and lower limits of x respectively. Chaotic mapping combined with reverse learning can expand the search space, improve the ability to find the optimal solution, and accelerate the convergence speed of the algorithm and improve the accuracy of the algorithm.

[0103] Cauchy mutation disturbance: Cauchy mutation originates from Cauchy distribution. Its main characteristics are that the peak value at zero is small and the value decreases slowly from the peak value to zero, making the mutation more uniform. If the optimal value of the population does not change for three consecutive iterations, it is considered that the search is stagnant. At this time, Cauchy mutation is introduced to enhance the ability of particles to jump out of the local optimum. The formula of Cauchy mutation is as follows:

[0104] x′=x×(1-tan(π(u-0.5));

[0105] Where: x′ represents the position after mutation; u represents a random number in (0,1).

[0106] The particle fitness function uses the clustering evaluation index error square sum SSE, and the calculation formula is as follows:

[0107]

[0108] Where: K represents the number of clusters; μ i Represents cluster C i The cluster center of i The smaller the SSE value, the tighter the cluster is and the better the clustering effect is. Iteration is performed based on the minimum SSE standard. The velocity and position of the particle are encoded as K×d dimensional vectors. The boundary absorption strategy is adopted at the boundary of the position and velocity, and the maximum velocity of the particle is appropriately reduced in the later stage of the search. The algorithm flow is as follows:

[0109] Step 1: Set the initial parameters, including particle swarm size N, sample feature dimension d, number of clusters K, maximum number of iterations T, position boundary x min 、x max , and the maximum speed v max ; Step 2: Initialize the population. First, generate chaotic numbers through chaotic mapping, then map the chaotic numbers to the position interval, and construct the reverse position of the particle through reverse learning. Compare the fitness of the particle and its reverse position, and take the position corresponding to the better fitness as the initial position of the particle; Step 3: Start searching, iterate and update the particles normally, otherwise mutate the particle position and perform boundary control; Step 4: Divide the data set according to the nearest neighbor rule and recalculate the fitness of the particles; Step 5: Compare and update the optimal position and optimal fitness value of individual particles and groups; Step 6: If the iteration stop condition is met, go to step 7, otherwise return to step 3 to continue iterating; Step 7: Divide the data according to the nearest neighbor rule with the optimal solution as the clustering center to obtain K clusters and complete clustering.

[0110] (3) Label level classification: Different labels are divided into five levels according to the maximum and minimum value ranges, and arranged from small to large. Transformers with the same label level are selected for subsequent fault analysis and health status assessment and prediction.

[0111] Relevant data are collected and sorted, and the complex data empirical mode decomposition method is used to perform noise reduction analysis on the data.

[0112] The use of NACEMD (noise-assisted complex data empirical mode decomposition) method to reduce the noise of real-time captured data (including indicator data in the label and indicator data in the status evaluation) includes: (1) performing M initialization CEMD (complex data empirical mode decomposition) and adding white noise with an amplitude of k. (2) adding the white noise xn(z) to the original data sequence x0(z) of the working state of the smart transformer to form a complex signal xc(z)=x0(z)+ixn(z), where z is a time series variable. (3) projecting the complex signal onto On, among them, is the projection direction, 1≤k≤N:

[0113]

[0114] Substituting Euler's formula into the above formula:

[0115]

[0116] when When , the direction of the original extreme value selection changes, and the solution is The maximum point of the projection is obtained by performing cubic spline interpolation to obtain the upper and lower envelopes in each projection direction. (4) The average value of the boundary envelope m(z) is calculated, and then the component signals h(z)=x are obtained based on the average value of the boundary envelope m(z). c (z)-m(z). (5) Obtain the mean of the IMF components and residuals. (6) Overall average operation. Calculate the mean of each IMF component obtained after M CEMD processes as the output.

[0117] The characteristic vectors that affect the operating status of the transformer are selected and the characteristic vector prediction model is constructed.

[0118] (1) Selection of key state quantities of equipment: Combined with data processing, the historical experimental data of faulty / defective transformer equipment and the monitoring data during the fault period are used as samples to mine the correlation between each state parameter and different fault modes, and then a relationship set corresponding to the key performance of the equipment and the state quantity is constructed. Based on the collected transformer historical data and fault samples, five key performances such as load performance, insulation performance, and mechanical performance are statistically analyzed, and a relationship set is established (see Table 2 below).

[0119] Table 2

[0120] Serial number Itemset Fault type 1 Load performance Short circuit fault, winding fault 2 Insulation performance (overheating) Core failure, current circuit overheating 3 Insulation performance (discharge) Arc discharge, partial discharge 4 Insulation performance (moisture The dielectric loss of the bushing exceeds the standard and the insulating oil deteriorates 5 Mechanical properties Tap changer failure, winding failure

[0121] From the perspective of association rules, define I = {transformer failure mode belongs to the jth category}, X i,j = {the i-th state quantity exceeds the warning value in the j-th fault mode}, Y i = {the i-th state quantity exceeds the warning value in all failure modes}, I represents the data set, X i,j and Y i Represents different item sets.

[0122] X i,j and Y i Simultaneous confidence definition:

[0123]

[0124] Where p(X i,j ) indicates that the data set I contains parameter X i,j The probability of i,j ) indicates X i,j The support count of X i,j The proportion in the transaction database. The confidence level can be used to quantify the correlation between each state quantity and the fault set. The higher the confidence, the stronger the correlation between the corresponding parameter and the transformer fault set. According to the principle of the confidence minimum threshold method, the corresponding state vector screening result is finally obtained (refer to Table 3 below).

[0125] Table 3

[0126] Key Performance State quantity set Load performance Load, winding temperature, ambient temperature Insulation performance (overheating) <![CDATA[Load, ambient temperature, winding temperature, ground current, CH, C2H4, total hydrocarbons, C2H6]]> Insulation performance (discharge) <![CDATA[Load, partial discharge, H2, total hydrocarbons, CO / CO2, C2H2]]> Insulation performance (moisture) <![CDATA[Casing dielectric loss, ambient humidity, H2, micro water in oil]]> Mechanical properties Box vibration

[0127] (2) Constructing prediction model: The single variable sequence of transformer operating state parameters can be expressed as Where X m represents the mth parameter, It represents the measured value of the mth parameter at time t. All state parameters such as transformer equipment state quantity, environmental meteorology, and power grid operation data are arranged in chronological order to form a time series matrix.

[0128]

[0129] Where r represents the number of state parameters, so the prediction of transformer operation state parameters can be regarded as a prediction problem of high-dimensional parameters.

[0130] If x t+τ Related to its previous K data, the prediction task can be described as t+τ =F(x t-k ,...,x t-1 ,θ), where θ is the parameter vector in the F model. is the cost function.

[0131]

[0132] The Grid Long Short-Term Memory Network is a multidimensional spatial network generated by LSTM units, which can be used to process vectors, sequences, or higher-dimensional data. Unlike the traditional LSTM RNN model that organizes LSTM modules into a time chain, the GLSTM model arranges LSTM blocks into a multidimensional grid, so that each grid contains a set of multidimensional LSTM blocks, and each grid is connected between network layers and in the spatiotemporal dimensions of the data. This architecture introduces gated linear dependencies between adjacent unit states in each dimension, reducing the occurrence of problems such as gradient vanishing or exploding. Therefore, GLSTM provides a unified module that can be used for spatial depth and time series calculations.

[0133] In the deep LSTM network, the state and output of each layer are calculated as follows.

[0134]

[0135]

[0136] Where x is the input sequence x=(x1,x2,...x k ), They are respectively the forget gate, input gate, and output gate, representing the state or output of layer l at time t. They are the weight matrices of the forget gate, input gate, and output gate respectively. are the bias terms of the forget gate, input gate, and output gate respectively. σ is the activation function. It represents the output value of the LSTM at time t in layer l And the state of the gate control unit at time t is the unit state of the input at time t in layer l, is the input unit state weight matrix, is the input unit state bias term, and ⊙ represents element-wise multiplication.

[0137] The two-dimensional GLSTM module contains LSTM modules in two dimensions: time length and spatial depth.

[0138]

[0139] Respectively represent the calculation results of the time LSTM sub-block and the space LSTM sub-block, x t,l is the input of the l-layer GLSTM at time t, and is output by the l-1-layer cell in the spatial structure at time t And the output of the l-layer cell at time t-1 in the time dimension Composition. i Represents all parameters of the i_LSTM module.

[0140] The value is uncertain and is usually initially set to 0, which acts as a smoothing function for any input. V represents the linear variation coefficient matrix.

[0141] The transformer state parameter trend prediction model based on the GLSTM network can effectively predict the trend of transformer state parameters. The model uses the transformer online monitoring state, power grid operation status and substation environmental meteorological data as input feature quantities, and uses the GLSTM network to extract features and mine parameter associations. This helps to improve prediction accuracy and reliability, and provides strong support for transformer state monitoring and fault warning. In practical applications, the model parameters can be adjusted and optimized according to specific needs and data conditions to obtain better prediction results.

[0142] 1) The K-order measurement points with strong time dependence are normalized by the deviation standardization method to obtain the data representation X.

[0143] 2) Extract prediction parameters through GLSTM network, where GLSTM uses the time back propagation algorithm to train h t-k is the hidden layer expression, including and The hidden state of the previous layer is used as the input of the next layer of the network, forming a deep mining structure for prediction features.

[0144] 3) The correlation between the feature quantities extracted in the previous step is used as the initial parameter to adjust the weight of the output layer of the feedforward neural network. The feedforward neural network is used as the inference prediction layer to output the prediction result.

[0145] 4) Based on the prediction parameters obtained from the training data, the state quantities at future moments in the test set are tested to verify the prediction accuracy.

[0146] The GLSTM network has three gate switches to protect and control the cell state. The input gate, forget gate, and output gate correspond to the injection, accumulation, and output operations of the transformer-related state quantity. The gate switch is used to realize the temporal memory function to prevent the gradient from disappearing. The unique deep spatial information extraction function in the GLSTM module is to deeply mine each state.

[0147] Finally, combined with historical fault information, cluster analysis is performed on the fault information and feature vectors. Combined with feature vector prediction, the future health status of the transformer is judged. By setting corresponding health index constraints, the time for the transformer to operate healthily and stably is calculated.

[0148] (1) Health status prediction:

[0149] 1) Step 1: Based on historical data, feature extraction is performed on the indicator data of the distribution transformer in normal state and fault state respectively to construct a feature vector. Among them, the indicator data sequence of the normal state and the indicator data sequence of the fault state of the distribution transformer are respectively shown as follows:

[0150]

[0151] Where, X zi represents the i-th indicator data sequence under the normal state of the distribution transformer, represents the jth data object in the i-th normal indicator data sequence; X gi represents the i-th indicator data sequence under the fault state of the distribution transformer, Represents the jth data object in the i-th fault indicator data sequence.

[0152] 2) Step 2: For each relevant indicator, use its different state feature vectors as input to train the k-medoids clustering model and obtain the cluster center set:

[0153] {k zi ,k gi},(i=1,2,...,N);

[0154] Among them, k zi represents the cluster center of the health status of the distribution transformer corresponding to the i-th indicator, k gi Represents the cluster center of the distribution transformer failure state corresponding to the i-th indicator.

[0155] The k-medoids algorithm is a classic clustering method based on partitioning. The PAM (partitioning around medoids) algorithm in k-medoids is used for clustering. The principle is to give a set X containing n elements = {x1, x2, ..., x n}, where xi ∈R s , assign the objects in X to k clusters, each cluster contains a center point. After clustering, the objects in the clusters have similarities, and the objects between clusters have differences. The specific implementation process of the PAM algorithm is as follows:

[0156] Step 1: Initialize the number of clusters k and randomly select k objects from X as the initial center point p of the cluster j (j=1,2,...,k).

[0157] Step 2: Assign the remaining objects in X to the cluster whose center point is closest to it.

[0158] Step 3: For each cluster, take each element x in the cluster i Loop to replace the cluster center point p j , and calculate p j With other elements x i (x i ≠p j ) and select the element x with the smallest distance sum. j Replace p as the new center point j .

[0159] Step 4: Repeat steps 2-3 until the cluster center point does not change or the sum of the distances from other elements in the cluster to the center point does not change. j The distance is defined as follows:

[0160] d(x i ,p j )=||x i -p j || 2 .

[0161] The objective function of k-medoids clustering is as follows:

[0162]

[0163] 3) Step 3: The prediction of the relevant indicators of the distribution transformer is used as the test data set, as shown in the following formula:

[0164] X ti ={x ti}, (i=1,2,...,N; j=1,2,...n);

[0165] Where, X ti represents the forecast data sequence of the i-th indicator, x tiRepresents the jth data object in the i-th indicator prediction data sequence. According to the dissimilarity measure between the calculated test data feature vector and the health status cluster center, it is used as the distribution transformer performance to measure the transformer operation condition. The operation condition TH can be converted into the distribution transformer health confidence CV using the following formula.

[0166]

[0167] CV i =exp(-TH i / c),(i=1,2,...,N);

[0168] Where TH i Evaluate the transformer operation status corresponding to the relevant indicators for the i-th distribution transformer, d ij Represents the dissimilarity measure between the jth data object in the i-th indicator sequence and the health status cluster center; CV i represents the health confidence level corresponding to the evaluation indicators of the ith distribution transformer, and c is the scale factor. i ∈[0,1],CV i The closer it is to 1, the better the health of the distribution transformer. i The closer it is to 0, the worse the performance of the distribution transformer.

[0169] Step 4: Determine the weight w of each indicator according to the hierarchical analysis method i (i=1,2,...,N), weighted calculation is performed on the health confidence of the distribution transformer corresponding to each indicator to obtain the final health confidence of the distribution transformer CV:

[0170]

[0171] Where wi represents the weight corresponding to the evaluation indicators of the i-th distribution transformer.

[0172] First, we build a hierarchical structure, break down the decision-making problem into details, and organize it into a top-down hierarchical structure. Complex problems are broken down into multiple core elements, and the upper-level elements play a decisive role in dominating the lower-level elements.

[0173] When applying AHP to analyze decision problems, we first decompose the problem into multiple elements, then decompose the elements into several levels according to their attributes, and finally construct a hierarchical problem structure model as the basis for the AHP calculation (refer to Figure 2 ).

[0174] 1) Construct a judgment matrix: In order to reflect the weight correspondence between each element, it is necessary to construct a judgment matrix. Generally, the 1-9 scaling method is adopted, and the numbers 1-9 and their reciprocals are used as scales to evaluate the correspondence between elements.

[0175] 2) Single sorting consistency test: The consistency index CI is generally used to test whether each judgment matrix is ​​designed reasonably and whether there are logical errors. It is generally believed that when CI < 0.10, the judgment matrix is ​​reasonable and within an acceptable range, and no further adjustment is required. Otherwise, the judgment matrix needs to be further adjusted and corrected.

[0176] 3) Total sorting consistency check: After the single sorting consistency meets the requirements, the total sorting needs to be checked for consistency. If the check passes, the current weight sorting result can be used as the final decision basis. If the consistency index is greater than 0.1, each indicator layer needs to be rebuilt.

[0177] (2) Transformer healthy operation time prediction: The offline status of the distribution transformer is evaluated using an improved health index calculation formula, as shown in the following formula:

[0178]

[0179] In the formula, the health index HI represents the health index of the distribution transformer corresponding to the current time T, HI0 is the health index corresponding to the initial operation time T0 of the transformer, and B is the aging coefficient. At this time, the value range of HI is [0,1], and the larger the HI value, the better the health of the distribution transformer. It is generally believed that when the health index HI≤0.3, the aging phenomenon of the distribution transformer is serious, the failure rate is significantly increased, and it is difficult to maintain normal and stable operation. Therefore, the present invention selects HI=0.35 as the critical health index to ensure the benign operation of the distribution transformer.

[0180] During the actual operation of the distribution transformer, it is affected by the operating load and the operating environment for a long time. Its actual remaining benign operating time is often equal to or less than the theoretical remaining time. Therefore, the theoretical operating time of the distribution transformer needs to be corrected based on the operating load and operating environment factors, as shown in the following formula:

[0181]

[0182] Where, T act is the expected service life of the distribution transformer after correction, T sj is the theoretical service life of the distribution transformer, i.e., the design life, f L and f E are load factor and environmental factor, respectively, where the environmental factor f E It is determined by the ambient temperature and the location of the distribution transformer. L and fE Determine according to Table 4 and Table 5 below respectively.

[0183] Table 4

[0184] Distribution transformer load factor (%) Load Factor [0,40] 1.0 (40,60) 1.05 (60,70) 1.11 (70,80) 1.25 [80,100] 1.3 >100 1.6

[0185] Table 5

[0186] Distribution transformer environment Environmental Factor Indoor, maximum ambient temperature <35℃ 1 Outdoor, maximum ambient temperature <35℃ 1.02 Indoor, maximum ambient temperature>35℃ 1.05 Outdoor, maximum ambient temperature>35℃ 1.1

[0187] Assuming that the health index of the distribution transformer during initial operation is HI0=0.95, the aging coefficient can be calculated:

[0188]

[0189] Among them, the design life of the distribution transformer is T sj It can be obtained from the ledger information, which is generally 40 years. The present invention conducts a progressive offline evaluation of the performance status of the distribution transformer from three levels. First, the health index HI1 of the distribution transformer based on the natural aging of the service life is calculated according to the equipment health index formula:

[0190] HI1=1-(1-0.95)×e B×T 1-0.05e B×T ;

[0191] Based on the quantitative evaluation level of each indicator corresponding to oil data, dissolved gas, and furfural content, the indicator weight is set to V. The weights of each indicator can be obtained by the hierarchical analysis method. Therefore, the health index of the distribution transformer can be obtained as follows:

[0192]

[0193] The health index of the distribution transformer based on the test information is:

[0194]

[0195] Where HI2 represents the health index of the distribution transformer corresponding to the kth type of test information, Represents the weight coefficient corresponding to the type of test information.

[0196]

[0197] in, The coefficient representing the defect level of the i-th type fault, is the occurrence coefficient corresponding to the i-th fault defect level.

[0198] In summary, the health index of the distribution transformer is:

[0199]

[0200] in, For HI i The weight coefficient of .

[0201] The health status index is calculated through prediction, CV∈[0,1], and the larger the value, the better the health status of the transformer. Therefore, the HI value can be referred to. When CV<0..35 is set, the performance of the distribution transformer is difficult to meet the requirements and it is difficult to maintain safe and stable operation. The reliable operation time of the distribution transformer is calculated according to the eigenvector prediction results as shown in the following formula:

[0202] RUL on =(CV-0.35)·T act ;

[0203] Among them, T act It is the expected service life of the distribution transformer.

[0204] Substituting the critical health index HI = 0.35 of the distribution transformer into the health index formula, we can get:

[0205]

[0206] The transformer stable operation time obtained based on the historical fault data is:

[0207]

[0208] In summary, the healthy operation time of distribution transformers combined with historical data and predicted data is:

[0209]

[0210] in, is the weight coefficient, which can be taken

[0211] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0212] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for predicting the healthy operation time of a transformer, characterized in that: include: Collecting index data and status data of multiple transformers and performing noise reduction processing on the index data and the status data; Based on the index data processed by noise reduction, the multiple transformers are labeled and graded to obtain transformers of multiple grades, wherein the different index data correspond to energy efficiency grade, load capacity, short-circuit resistance, insulation level and voltage regulation capacity, and different types of transformer sets are formed as the transformers of multiple grades according to the actual grades of the transformers. The multiple transformers are labeled and graded based on different index data to obtain transformers of the same grade, including: constructing a portrait fact label based on the energy efficiency grade, the load capacity, the short-circuit resistance, the insulation level and the voltage regulation capacity; identifying and extracting the principal component feature label by improving the particle swarm algorithm, wherein the principal component feature label is used to reflect the performance and status information of the transformer; constructing a model label based on the principal component feature label, wherein the model label includes environmental state attributes, operating state attributes and Equipment body attributes; and integrating the portrait fact label and the model label into the portrait system to select the transformer of the same level, the different indicator data include: the indicator data corresponding to the energy efficiency level include winding insulation resistance, bushing outer insulation creepage, body oil withstand voltage and bushing main insulation resistance; the indicator data corresponding to the load capacity include: cooler fault and full stop signal, bushing oil color spectrum and core grounding current; the indicator data corresponding to the short-circuit resistance include: low voltage short-circuit impedance, number of short-circuit impacts, short-circuit resistance hidden dangers and whether the medium voltage side is half capacity; the indicator data corresponding to the insulation level include: equipment operating years, annual heavy overload time, no-load loss and energy efficiency and main transformer real-time oil temperature; and the indicator data corresponding to the voltage regulation capacity include: number of on-load tap changer operations and on-load tap changer over-cycle unchecked; Transformers of the same grade are selected from the transformers of the plurality of grades, a prediction model is constructed based on historical state data, and the prediction model is used to predict the state data trend of the real-time state data to obtain a first eigenvector. The prediction model is constructed based on the historical state data of the transformers of the same grade, further comprising: taking the historical state data of the faulty / defective transformer and the monitoring data of the fault period as samples, and determining the correlation between each state data and different fault sets according to the confidence level; selecting state data with high correlation as the state vector according to the principle of the confidence minimum threshold method; forming a time series matrix of the state vector of the transformer, the environmental meteorological data and the power grid operation data in chronological order; and constructing a long short-term memory network as the prediction model according to the time series matrix, and then using the real-time state data, the real-time environmental meteorological data and the real-time power grid operation data as the input of the prediction model to predict the state data at the future moment; and Based on the first feature vector and the health status cluster center, the future health status of the transformer is judged and the healthy operation time of the transformer is determined, wherein, based on the historical fault data, feature extraction is performed on the indicator data of the transformer to construct a second feature vector, and cluster analysis is performed on the second feature vector to obtain the health status cluster center.

2. The method for predicting the healthy operation time of a transformer according to claim 1, characterized in that: The noise reduction process for the indicator data and the status data further includes: Performing M initializations of complex data empirical mode decomposition CEMD, counting the number of times white noise with amplitude k is added includes: The white noise x n (z) Add the original data sequence x0(z) of the indicator data and the state data to form a complex signal x c (z) = x0(z) + ix n (z), where z is a time series variable; Project the complex signal onto On, among them, is the projection direction, 1≤k≤N: pp φk (z)=Re(e -iφk (x c (z)); Substituting Euler's formula into the above formula: when When , the direction of the original extreme value selection changes, and the solution is The maximum point of , and perform cubic spline interpolation on it to obtain the upper and lower envelopes in each projection direction; Calculate the average value of the boundary envelope m(z), and then obtain each component signal h(z)=x according to the average value of the boundary envelope m(z). c (z)-m(z); obtain the mean of the IMF component and the residual; and calculate the mean of each IMF component obtained after M times of complex data empirical mode decomposition CEMD process as the indicator data and state data after denoising.

3. The method for predicting the healthy operation time of a transformer according to claim 1, characterized in that: The plurality of transformers are classified into label grades to obtain transformers of the same grade, including: Step 1: Set the following initial parameters: particle swarm size Num, sample feature dimension d, number of clusters KN, maximum number of iterations T, position boundary x min 、x max , and the maximum speed v max ; Step 2: Generate chaotic numbers through chaotic mapping, then map the chaotic numbers to position intervals, construct the reverse position of the particle through reverse learning, and compare the fitness of the particle and its reverse position, so that the position corresponding to the better fitness of the particle and its reverse position is taken as the initial position of the particle; Step 3: Start searching from the initial position of the particle, update the particle through normal iteration, otherwise when the particle falls into the local optimum, perform Cauchy mutation on the particle position and perform boundary control; Step 4: Divide the denoised index data set according to the nearest neighbor rule and recalculate the fitness of the particles; Step 5: Compare and update the optimal position and optimal fitness value of individual particles and the group; Step 6: If the iteration stop condition is met, proceed to step 7, otherwise return to step 3 to continue iterating; Step 7: Taking the optimal solution as the clustering center, the index data of the noise reduction processing is divided according to the nearest neighbor rule to obtain KN clusters, thereby completing the clustering of the multiple transformers.

4. The method for predicting the healthy operation time of a transformer according to claim 3, characterized in that: Update particle speed and position by the following formula: oh = oh max -(oh max -oh min )×q / T; in, represents the d-dimensional velocity of particle i at the t+1th iteration, and the unit is consistent with the data; c1 and c2 represent learning factors, which represent the degree of cognition of the particle to itself and the group; r1 and r2 represent random numbers between (0,1); represents the d-th dimension position of particle i at the t+1th iteration; P id , g id Represent the optimal solutions of the current particle and the group respectively; ω represents the inertia weight; q represents the current number of iterations, and T represents the maximum number of iterations; Generate chaotic numbers through the chaotic mapping formula: x′=x max +x min -x; Among them, z k is the chaotic number generated for the kth time, β is taken as (0,1); x represents the indicator data; x′ represents the reverse solution of the indicator data x; x max 、x min are the upper and lower limits of the value of the indicator data x respectively; The ability of particles to escape from local optimality is enhanced through the following Cauchy mutation formula; x' b =x×(1-tan(π(u-0.5)); Among them, x' b represents the position after mutation; u represents a random number in the interval (0,1); The particle fitness function uses the clustering evaluation index error square sum SSE: Among them, μ i Represents cluster C i The smaller the SSE value, the tighter the cluster is and the better the clustering effect is.

5. The method for predicting the healthy operation time of a transformer according to claim 4, characterized in that: Judging the future health state of the transformer based on the first feature vector and the health state cluster center and determining the time for the transformer to operate stably further includes: Based on the historical fault data, feature extraction is performed on the indicator data of the transformer in a normal state and a fault state to construct a second feature vector; Using the second feature vector as input data to train the k-medoids clustering model to obtain a cluster center set; Based on the first feature vector and the health status cluster center, determine the operation status of the transformer and convert the operation status of the transformer into a transformer health confidence level; Determine the weight of each indicator according to the hierarchical analysis method, and use the indicator weight to perform weighted calculation on the transformer health confidence to obtain the final transformer health confidence; and The transformer stable operation time is calculated according to the final transformer health confidence level.

6. A transformer healthy operation time prediction system, characterized in that: include: Data acquisition module, used to collect index data and status data of multiple transformers and processing module; A noise reduction processing module, used for performing noise reduction processing on the indicator data and the status data; A level classification module is used to classify the multiple transformers into labels based on the indicator data processed by noise reduction to obtain transformers of multiple levels, wherein the different indicator data correspond to energy efficiency level, load capacity, short-circuit resistance, insulation level and voltage regulation capability, and different types of transformer sets are formed as the transformers of multiple levels according to the actual level of the transformer. Classifying the multiple transformers into labels based on different indicator data to obtain transformers of the same level includes: constructing a portrait fact label based on the energy efficiency level, the load capacity, the short-circuit resistance, the insulation level and the voltage regulation capability; identifying and extracting the principal component feature label by improving the particle swarm algorithm, wherein the principal component feature label is used to reflect the performance and status information of the transformer; constructing a model label based on the principal component feature label, wherein the model label includes environmental state attributes, operation State attributes and equipment body attributes; and integrating the portrait fact label and the model label into the portrait system to select the transformer of the same level, the different indicator data include: the indicator data corresponding to the energy efficiency level include winding insulation resistance, bushing outer insulation creepage, body oil withstand voltage and bushing main insulation resistance; the indicator data corresponding to the load capacity include: cooler fault and full stop signal, bushing oil color spectrum and core grounding current; the indicator data corresponding to the short-circuit resistance include: low voltage short-circuit impedance, number of short-circuit impacts, short-circuit resistance hidden dangers and whether the medium voltage side is half capacity; the indicator data corresponding to the insulation level include: equipment operating years, annual heavy overload time, no-load loss and energy efficiency and main transformer real-time oil temperature; and the indicator data corresponding to the voltage regulation capacity include: number of on-load tap changer operations and on-load tap changer over-cycle unchecked; A state trend prediction module is used to select a transformer of the same grade from the transformers of the multiple grades, build a prediction model based on historical state data, and use the prediction model to perform trend prediction on the real-time state data to obtain a first eigenvector, wherein the state trend prediction module is used to use the historical state data of the faulty / defective transformer and the monitoring data of the fault period as samples, and determine the correlation between each state data and different fault sets according to the confidence level; select state data with high correlation as the state vector according to the principle of the confidence minimum threshold method; form a time series matrix of the state vector of the transformer, environmental meteorological data and power grid operation data in chronological order; and build a long short-term memory network as the prediction model according to the time series matrix, and then use the real-time state data, real-time environmental meteorological data and real-time power grid operation data as inputs of the prediction model to predict the state data at future moments; and A stable operation time determination module is used to judge the future health status of the transformer and determine the healthy operation time of the transformer based on the first feature vector and the health status cluster center, wherein, based on historical fault data, feature extraction is performed on the indicator data of the transformer to construct a second feature vector and cluster analysis is performed on the second feature vector to obtain the health status cluster center.

7. The transformer healthy operation time prediction system according to claim 6, characterized in that: The noise reduction processing module is used for: Perform M initializations of the complex data empirical mode decomposition CEMD, adding white noise with an amplitude of k for each time; The white noise x n (z) Add the original data sequence x0(z) of the indicator data and the state data to form a complex signal x c (z) = x0(z) + ix n (z), where z is a time series variable; Project the complex signal onto On, among them, is the projection direction, 1≤k≤N: pp φk (z)=Re(e -iφk (x c (z)) Substituting Euler's formula into the above formula: when When , the direction of the original extreme value selection changes, and the solution is The maximum point of , and perform cubic spline interpolation on it to obtain the upper and lower envelopes in each projection direction; Calculate the average value of the boundary envelope m(z), and then obtain each component signal h(z)=x according to the average value of the boundary envelope m(z). c (z)-m(z); Obtain the means of the IMF components and residuals; and The mean of each IMF component obtained after M times of complex data empirical mode decomposition (CEMD) is calculated as the indicator data and status data after denoising.

Images