Disclosure of Invention
In view of the above analysis, the present invention aims to provide a method and a system for determining maintenance decision of a transformer based on state evaluation and fault rate correction, which are used for solving the problem of how to improve the accuracy of maintenance decision.
On one hand, the embodiment of the invention provides a transformer maintenance decision method based on state evaluation and fault rate correction, which is characterized by comprising the steps of collecting transformer fault data and preprocessing to obtain characteristic quantities, wherein the characteristic quantities correspond to oil chromatographic data, oil test data and electric test data, dividing transformer fault data types into primary division regions by utilizing preprocessed historical transformer fault data, then secondarily dividing the primary division regions into health evaluation regions based on the primary division regions and setting parameters, constructing and training a plurality of neural network models based on the characteristic quantities and the health evaluation regions to obtain a plurality of prediction results of the health condition of a transformer, combining the prediction results to obtain a final prediction region type, generating a fault rate curve of the transformer based on the final prediction region type, correcting the fault rate curve based on individual differences and running state images, and constructing a maintenance policy decision model to obtain optimal maintenance time based on the corrected fault rate curve and minimum cost and reliability constraint.
Based on further improvement of the method, collecting transformer fault data and preprocessing to obtain characteristic quantities comprises the steps of collecting the transformer fault data, wherein the transformer fault data comprise fault data under the condition that a transformer breaks down, corresponding transformer state data and external environment data, preprocessing abnormal data, wherein the preprocessing comprises the steps of filling missing data and removing abnormal values and repeated values, and removing redundant relevant elements from the preprocessed transformer fault data through an entropy weighting method to obtain the characteristic quantities.
Based on further improvement of the method, removing redundant related elements from the preprocessed transformer fault data through an entropy weight method to obtain the characteristic quantity further comprises the steps of utilizing m samples to be evaluated in the transformer fault data and n indexes to form the following original index data matrix:
Wherein xij represents the j index value of the i sample, normalizes each index value in the original index data matrix by using a range normalization method and a standard deviation normalization method, and calculates the specific gravity of each index value by the following formula:
Wherein pij represents the specific gravity of the j-th index value of the i-th sample;
Calculating an index entropy value by the following entropy value calculation formula:
Wherein ej represents the index entropy value of the j-th index value, and calculating redundancy values of the indexes based on the difference degree of the indexes by using the following formula:
dj＝1-ej;
And dj is the difference degree of the j-th index, sequencing the index redundancy values from large to small, and selecting the evaluation index corresponding to the index redundancy value with the maximum index redundancy value and the preset number of index redundancy values as the characteristic quantity.
Based on a further improvement of the method, dividing the transformer fault data type into primary divided sections by the preprocessed transformer fault data, and then secondarily dividing the primary divided sections into health evaluation sections based on the primary divided sections and setting parameters, further comprises dividing the transformer into primary divided sections by the preprocessed transformer fault data based on a Markov chain, the primary divided sections including a no fault section [ a1,b1 ], a slight fault section [ a2,b2 ] and a severe fault section [ a3,b3 ], setting a point between no fault and slight fault as c1 and a point between slight fault and severe fault as c2, setting 2 judgment fuzzy areas of no fault and slight fault, slight fault and severe fault based on the point c1 and the point c2, and calculating the length of the fuzzy area by the following formula:
L1＝min{b1-a1,b2-a2}·k%;
L2＝min{b2-a2,b3-a3}·k%;
The blurred region is determined as c1-L1,c1+L1 based on the length of the blurred region,
And dividing the health status section into health assessment sections based on the length of the blurred region and the blurred region, wherein the health assessment sections include a no-fault section [ a1,c1-L1 ], a slight fault possible section [ c1-L1,c1+L1 ], a slight fault section [ c1+L1,c2-L2 ], a severe fault possible section [ c2-L2,c2+L2 ] and a severe fault possible section [ c2+L2,b3 ].
Based on the further improvement of the method, constructing and training a plurality of neural network models based on the characteristic quantity and the health evaluation interval to obtain a plurality of prediction results of the health condition of the transformer, and combining the plurality of prediction results to obtain a final prediction interval type further comprises the following steps of:
the method comprises the steps of determining a total quantity of characteristic quantities in a transformer, wherein xgy is a normalized value, x is a value of characteristic quantities in a data set, xmin、xmax is a maximum value and a minimum value of the characteristic quantities respectively, respectively establishing three neural network models of the characteristic quantities corresponding to oil chromatographic data, oil test data and electric test data after normalization processing, training the three neural network models, determining the output quantity of each neural network model based on the health evaluation interval, and predicting a final prediction interval of the transformer by utilizing the trained neural network models, wherein the final prediction interval comprises the non-fault interval, the slightly fault possible interval, the severely fault possible interval and the severely fault possible interval.
Based on the further improvement of the method, generating a fault rate curve of the transformer based on the final prediction interval type, and correcting the fault rate curve based on the individual difference and the running state image, the method further comprises the steps of acquiring shape parameters and scale parameters of each stage by using transformer fault data based on the final prediction interval type to establish the fault rate curve of the transformer based on the shape parameters and the scale parameters, carrying out linearization correction on the fault rate curve of the transformer produced by the same manufacturer according to the actual running condition of the transformer to obtain a correction fault rate function considering familial defects, then obtaining a correction fault rate curve considering individual defects based on the correction fault rate curve considering familial defects, constructing a running state image, and correcting the fault rate curve of the transformer based on the running state image to obtain the correction fault rate curve considering the running image.
Based on a further improvement of the above method, the corrected failure rate curve after considering familial defects is expressed as:
the corrected failure rate curve after considering the individual defects is expressed as;
Wherein gt (t) is a correction failure rate function considering individual defects, wu is the average no-failure time of an individual transformer, t2 is the operation period of the transformer, n2 is the number of times the transformer fails during operation, qx (t) is the correction failure rate function considering familial defects, jc (t) is the failure rate function obtained by fitting according to historical data of similar transformers, wf is the average no-failure time of similar transformers, wt is the average no-failure time of transformers produced by the same manufacturer, t1_i is the operation period of the ith transformer in the similar transformers, n1 is the counted total number of times the ith transformer fails during operation, and t0_i is the total number of times the transformer fails during operation.
Based on the further improvement of the method, constructing an operation state image and correcting a fault rate curve of the transformer based on the operation state image to obtain a corrected fault rate curve, wherein the method further comprises the steps of objectively extracting indexes by utilizing a correlation analysis and Laplace scoring method to select a preset number of indexes with different correlation intensities and maximum Laplace scores as main characteristic indexes, generating index labels based on the main characteristic indexes, wherein the index labels comprise indexes respectively corresponding to the regional attribute, the natural attribute and the operation attribute, constructing an operation state image based on the index labels, wherein the operation state image comprises the regional attribute, the natural attribute and the operation attribute, calculating index label weight under each attribute dimension by an entropy weight method, weighting and sequencing the index labels based on the index label weight, carrying out correlation analysis on different operation state images and transformer fault rates to obtain an adjustment coefficient g, generating an image rate curve based on the adjustment coefficient and the transformer fault rate, and a fault rate corrected curve after consideration of familial defects and the fault rate corrected curve after consideration of the fault rate of the transformer, and the fault rate corrected curve after consideration of the fault rate curve is generated:
where g is an adjustment coefficient, HX1 is a failure rate of the corresponding image, and HX2 is an average failure rate.
Based on further improvement of the method, constructing an overhaul strategy decision model based on the corrected fault rate curve and minimum cost and reliability constraint to obtain a final overhaul plan further comprises constructing the overhaul strategy decision model and constraint conditions thereof based on overhaul risks and fault risks, wherein overhaul risks are determined based on random power loss of a power grid and planned loss of load in an n-period overhaul mode m, and fault risks are determined based on random loss of load of the power grid and loss of the power grid in the n-period overhaul mode m, the constraint conditions comprise simultaneous overhaul constraint, mutual exclusion overhaul constraint, overhaul resource constraint and power grid safety constraint, wherein the simultaneous overhaul constraint refers to constraint of overhauling the transformer simultaneously to avoid repeated power failure caused by overhauling the transformer simultaneously, the mutual exclusion overhaul constraint refers to constraint of overhauling the transformer not at the same time to avoid power failure, the overhaul resource constraint refers to constraint that the total number of overhauled transformers cannot exceed overhaul capacity of overhaulers, and the power grid safety constraint refers to constraint of carrying out safety inspection through tide calculation.
On the other hand, the embodiment of the invention provides a transformer overhaul decision system based on state evaluation and fault rate correction, which is characterized by comprising a data acquisition and processing module, a health evaluation interval acquisition module, an interval prediction module, a fault curve generation module, an optimal time determination module and an overhaul time determination module, wherein the data acquisition and processing module is used for acquiring transformer fault data and preprocessing to acquire characteristic quantity, the characteristic quantity corresponds to oil chromatographic data, oil test data and electric test data, the health evaluation interval acquisition module is used for dividing a transformer fault data type into primary division intervals by utilizing preprocessed historical transformer fault data and then dividing the primary division intervals into health evaluation intervals secondarily based on the primary division intervals and set parameters, the interval prediction module is used for constructing and training a plurality of neural network models to acquire a plurality of prediction results based on the characteristic quantity and the health evaluation intervals, the prediction results are combined to acquire a final prediction interval type, the fault rate curve of the transformer is generated based on the final prediction interval type, and the fault rate curve is corrected based on individual difference and running state images, and the optimal time determination module is used for constructing an overhaul constraint model based on the corrected fault rate and minimum cost and reliability.
Compared with the prior art, the invention has at least one of the following beneficial effects:
1. The influence of error data on analysis results can be reduced through preprocessing so as to improve data reliability, detailed fault classification is helpful for more accurately identifying fault modes, accurate information is provided for subsequent steps so as to improve fault identification accuracy, and time sequence analysis can improve prediction accuracy of future fault occurrence.
2. The optimized neural network structure provides more accurate health status, and the multi-source information fusion can provide more comprehensive equipment status information, thereby being beneficial to improving the judgment accuracy. The model adaptation is enhanced so that the model adaptation can keep high judgment accuracy when facing new data.
3. The accuracy of the fault rate curve can be improved by considering more influencing factors, and the dynamic correction can ensure that the fault rate curve always reflects the latest state of the equipment. Uncertainty analysis helps to assess risk and reliability of the predicted outcome.
4. Cost-benefit analysis may ensure that the overhaul strategy is optimized for economy and long-term benefit, and multi-objective optimization may help find the best balance point among multiple objectives. The algorithm optimization can improve the solving efficiency and accuracy of the overhaul strategy decision model.
In the invention, the technical schemes can be mutually combined to realize more preferable combination schemes. Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention may be realized and attained by the structure particularly pointed out in the written description and drawings.
Detailed Description
The following detailed description of preferred embodiments of the application is made in connection with the accompanying drawings, which form a part hereof, and together with the description of the embodiments of the application, are used to explain the principles of the application and are not intended to limit the scope of the application.
In one embodiment of the invention, a transformer overhaul decision method based on state evaluation and fault rate correction is disclosed, as shown in fig. 1, comprising the following specific steps.
In step S101, transformer fault data is collected and preprocessed to obtain feature values, wherein the feature values correspond to oil chromatography data, oil test data, and electrical test data.
Collecting transformer fault data and preprocessing to obtain characteristic quantities comprises the steps of collecting the transformer fault data, wherein the transformer fault data (namely, an original data set) comprises fault data under the condition that a transformer breaks down, corresponding transformer state data and external environment data.
For example, the fault data includes fault types such as winding shorts, insulation aging, cooling system faults, and the like. Fault occurrence time is specifically up to year, month, day, hour and minute, so as to know the occurrence frequency and time period of the fault. The severity of the fault is evaluated, for example, slightly, moderately, severely, or based on outage time, repair difficulty, etc. The equipment state data comprises oil temperature, normal oil temperature range and faults possibly predicted by abnormal rising of the oil temperature, wherein the oil temperature is monitored in real time through an oil temperature sensor. The load condition comprises the load rate of the transformer, the three-phase load balance degree and the like, and reflects the running state of the transformer. Oil level the height of the oil level is directly related to the heat dissipation and insulation performance of the transformer. Winding temperature is monitored by a winding temperature sensor, and excessive temperature can be used for indicating winding faults. Insulation resistance by periodically measuring insulation resistance, the insulation state of the transformer is evaluated. External environment data, namely, temperature, wherein the environment temperature has direct influence on heat dissipation and performance of the transformer. Humidity-high humidity may cause a decrease in the insulation performance of the transformer. The pollution degree, namely dust, salt mist and other pollutants can influence the insulation and heat dissipation of the transformer. Wind speed is the natural cooling effect of the transformer is affected. Lightning activity lightning may cause a lightning strike failure of the transformer. And other related data, namely historical overhaul records, including overhaul time, overhaul contents, overhaul personnel and other information, are helpful for analyzing fault reasons and making overhaul plans. The service life of the transformer is closely related to the operation state of the transformer, and equipment with long service life can be more likely to fail. The operation log records daily operation data of the transformer, such as voltage, current, power and the like, and is helpful for finding potential faults.
Preprocessing the abnormal data, wherein the preprocessing comprises filling in missing data and removing abnormal values and repeated values. Firstly, constructing a data missing value and abnormal value processing model by combining relevant actual requirements, secondly, carrying out reasonable and effective processing on data by means of integration and transformation, and thirdly, generating a complete data set from a data set containing missing values by adopting a missing value processing method based on repeated simulation for missing value processing. The missing data in each dataset is padded with the monte carlo method.
And the missing value processing is to construct a Markov chain by using the variable mean vector and the variance-covariance matrix as prior information, ensure that the distribution of elements can be converged to a stable distribution, repeatedly simulate the Markov chain by sampling, obtain stable posterior distribution and generate the estimation of missing data. The steps can be finished as follows.
A continuous set of data vectors (i.e., original set of data) Yc＝[Y1,Y2,....,Yn is received, where the i-th data vector is Y (i) = [ Yi(1),yi(2),.....,yi (D) ], i=1, 2.
Setting a Gaussian model according to the ith item of data, wherein a parameter space of the Gaussian model is theta, calculating the occurrence probability p (Yqs I Ywz,θg) of missing data according to an estimated value thetag of the space theta, and calculating the occurrence probability of the parameter space theta according to the current complete data and the estimated value of the missing dataAnd updating the estimated value of the parameter space theta of the Gaussian model. Up to the resulting Markov chainAnd estimating the missing data when converging.
The final missing data calculation formula is:
Where Nsample is the total number of samples and NBurn-in is the number of missing samples.
And deleting the abnormal value and the repeated value to obtain the processed data domain set. The fault types are classified empirically (refer to table 1 below).
TABLE 1
Redundant relevant elements are removed from the pre-processed transformer fault data by an entropy weight method to obtain feature quantities (refer to table 2 below). The main purpose of the entropy method is to weight the index system. The larger the entropy, the more chaotic the system, the less the information is carried, the smaller the weight, the smaller the entropy, the more ordered the system, the more the information is carried, and the greater the weight. The entropy method is an objective weighting method, and by means of information entropy thought, the information entropy of the indexes is calculated, and the weights of the indexes are determined according to the influence of the relative change degree of the indexes on the whole system, namely, the weights are weighted according to the difference degree of the mark values of the indexes, so that the corresponding weights of the indexes are obtained, and the indexes with large relative change degree have large weights.
TABLE 2
The step of removing redundant relevant elements from the preprocessed transformer fault data by an entropy weight method to obtain the characteristic quantity further comprises the following steps.
The method comprises the steps of collecting and rectifying original data, wherein m samples to be evaluated and n indexes in transformer fault data are utilized to form the following original index data matrix:
Where xij represents the jth index value of the ith sample. For a certain index Xj, the greater the degree of dispersion of the sample, the greater the role that the index plays in the overall evaluation. If the index flag values are all equal, it is indicated that the index does not play a role in the comprehensive evaluation.
And (3) data processing, namely carrying out normalization processing on each index value in the original index data matrix by using a range normalization method and a standard deviation normalization method. In order to eliminate the influence of different dimensions on the evaluation result, various indexes need to be normalized or standardized. The current normalization standard deviation normalization method and Z-score normalization method.
And (3) performing standard deviation, namely selecting the maximum value and the minimum value of sample data, performing standard treatment on the original data, and finally mapping all the data into intervals [0,1 ].
Wherein X' and X represent the sample data after and before normalization, respectively, and Xmax、Xmin represents the maximum value and minimum value in the sample data, respectively.
Z-score normalization, namely Z-score normalization, which is the most commonly used method for data normalization when the maximum value and the minimum value of a certain index are unknown or outlier values exceeding the value range exist, can be adopted, the data is normalized based on the mean mu and the standard deviation sigma of the original data, and the original value X of A is normalized to X' by using the Z-score method. Specific calculation formula
Calculating the specific gravity of each index value is calculated by the following formula:
Wherein pij represents the specific gravity of the j index value of the i sample, and the ratio matrix based on this can be constructed as follows:
calculating an index entropy value by the following entropy value calculation formula:
wherein ej represents the index entropy value of the j-th index value;
Defining index redundancy values, namely calculating each index redundancy value based on each index difference degree by using the following formula:
wherein dj is the degree of difference of the j-th index;
and sequencing all the index redundancy values from large to small, and selecting the evaluation indexes corresponding to the index redundancy values with the maximum preset number of index redundancy values as the characteristic quantity. And combining the entropy weight method calculation result, and selecting 40% of the maximum calculation result as the subsequent typical characteristic elements.
H2 and the total volume fraction C2H2 are important characteristic quantities for monitoring the occurrence of faults, H2 is typical representative gas of partial discharge faults, the occurrence of partial discharge faults can be reflected most, total hydrocarbon is used as the total sum of hydrocarbon gases related to faults, the occurrence of overheat faults and discharge faults can be reflected, and C2H2 has obvious effects in reflecting the discharge faults, namely CO is one of products of decomposition of insulating paper, and can reflect the insulation ageing condition of a transformer, in particular to the overheat fault condition. The furfural content in the oil-based test data is also one of insulation decomposition products, can reflect the insulation aging condition of the transformer, the micro-water content in the oil is an important parameter for monitoring the insulation wetting condition, has important significance for identifying insulation wetting faults, has similar oil breakdown voltage and aldehyde content and CO gas production rate, can reflect the insulation aging condition to a certain extent, and the oil medium loss can reflect the insulation condition of the oil. The absorption ratio in the electrical test data can reflect the degree of insulation wetting, insulation breakdown or severe overheat aging, namely, the insulation electric anode of the iron core has important significance in judging related faults of the iron core, such as the grounding of the iron core, the direct-current resistance difference of the winding can reflect the thermal faults, and the insulation dielectric loss of the winding can also reflect the insulation condition of the winding.
In step S102, the transformer fault data type is divided into primary divided sections using the preprocessed historical transformer fault data, and then the primary divided sections are secondarily divided into health evaluation sections based on the primary divided sections and the setting parameters.
Dividing the transformer fault data type into primary divided sections using the preprocessed transformer fault data and then secondarily dividing the primary divided sections into health evaluation sections based on the primary divided sections and setting parameters further includes dividing the transformer into primary divided sections including a no fault section [ a1,b1 ], a slight fault section [ a2,b2 ] and a severe fault section [ a3,b3 ] using the preprocessed transformer fault data based on a markov chain, setting a point between the no fault and the slight fault as c1 and a point between the slight fault and the severe fault as c2, setting 2 judgment blur areas of the no fault and the slight fault, the slight fault and the severe fault based on the point c1 and the point c2, and calculating the length of the blur area by the following formula:
L1＝min{b1-a1,b2-a2}·k％;L2＝min{b2-a2,b3-a3}·k％;
determining the blurred region as [ c1-L1,c1+L1],[c2-L2,c2+L2 ] based on the length of the blurred region;
The health status section is secondarily divided into health evaluation sections based on the length of the blurred region and the blurred region, wherein the health evaluation sections include a no-fault section [ a1,c1-L1 ], a slightly faulty section with possible occurrence [ c1-L1,c1+L1 ], a slightly faulty section with possible occurrence [ c1+L1,c2-L2 ], a severely faulty section with possible occurrence [ c2-L2,c2+L2 ], and a severely faulty section with possible occurrence [ c2+L2,b3 ].
The health state interval preliminary division based on Markov chains and Gibbs sampling, markov chains are random processes in probability theory and numerical statistics with Markov properties and existing in discrete index sets and state spaces. Markov chains that are suitable for continuous index sets are known as markov processes, but are sometimes also considered as subsets of markov chains, i.e. continuous time markov chains, correspond to discrete time markov chains, and thus markov chains are a broader concept. Markov chains may be defined by transition matrices and transition diagrams, and in addition to Markov, markov chains may have irreducibility, recurrent, periodic, and traversal properties. An irreducible and normally returnable Markov chain is a strictly smooth Markov chain with a unique smooth distribution. The limit distribution of traversing markov chains converges to its plateau distribution.
Let Xt denote the value of the random variable X at the discrete time t. If the transition probability of the variable over time depends only on its current value, i.e
P(Xt+1＝sj|X0＝s0,X1＝s1,...,Xt＝st)＝P(Xt+1＝sj|Xt＝st);
Representing that the state transition probability depends only on the previous state, this variable is called the Markov variable, where s0,s1,...,si,sj ε Ω is the possible state of the random variable X. This property is called markov property and the random process with markov property is called markov process. Markov chains refer to a sequence of values (X0,X1,...,Xm) of a random variable X over a period of time. The Markov chain is defined by a corresponding transition probability, which refers to the probability of a random variable transitioning from one time instant to the next, from state si to another state sj, namely:
P(i→j)=Pi,j＝P(Xt+1＝sj|Xt＝si);
The probability that the value of the random variable X at time t is sk is represented by the probability that the value of the random variable X at time t+1 is si:
Assuming the number of states is n, there are:
The method comprises the steps of P= (Pi,j)n×n is a transition probability matrix, a Markov chain has periodicity, namely state transition after limited times returns to the Markov chain, and has and irreducibility, namely mutual transition between two states, if one Markov process has no periodicity and is irreducible, the Markov process is called state traversal, the value of the Markov process of each state traversal is taken no matter how an initial value pi(0), and with the increase of transition times, the value distribution of a random variable finally converges to a unique stable distribution pi*, and the condition that the data of a transformer health evaluation system accords with each state traversal is established.
The Monte Carlo method principle is that a system is simulated by sampling a large number of random samples, so that parameters required to be calculated are obtained. Based on the Markov chain method, the Monte Carlo method is combined to realize the sampling of the primary partition of the transformer. The sampled data are multidimensional feature quantities and have certain relevance, and when the probability distribution of the feature quantities is calculated by directly using all the data, the complexity and the time consumption are huge. The multidimensional feature can be sampled using gibbs sampling in the markov chain monte carlo method and the probability of distribution of the feature can be obtained therefrom.
Gibbs sampling is applicable where the joint distribution is not explicitly known or difficult to directly sample, but the conditional distribution of each variable is known and easy to sample. Compared with other MCMC methods, the acceptance rate of the MCMC method is smaller than 1, the acceptance rate of Gibbs sampling is 1, the sampling efficiency is higher, and the sampling problem of multidimensional data can be met. The principle of Gibbs sampling is as follows if the state transition matrix P and probability distribution pi* of the aperiodic Markov chain satisfy for all i, j:
π(i)P(i,j)=π(j)P(j,i);
the probability distribution pi* is called the plateau distribution of the state transition matrix P. Assume that there is a two-dimensional probability distribution pi (i, j) for two points A (x1,y1)、B(x2,y2) with the same x-coordinate:
π(x1,y1)π(y2|x1)＝π(x1)π(y1|x1)π(y2|x1);
π(x1,y2)π(y1|x1)＝π(x1)π(y2|x1)π(y1|x1);
It can be seen that, on the straight line x=x1, the transition between any two points satisfies the fine and smooth condition by using the conditional probability distribution pi (y|x) as the state transition probability of the markov chain, and the same is true for the point C (x2,y1) on the straight line y=y1, so that the state transition probability matrix P between any two points of the plane is constructed:
P(A→B)=π(yB|x1) ifxA＝xB＝x1;
P(A→C)=π(xc|y1) ifyA＝yB＝y1;
P(A→D)=0else;
For the state transition probability matrix P, it can be known that for any two points X, Y on the plane, the fine balance condition is satisfied:
π(X)P(X→Y)=π(Y)P(Y→X);
And carrying out Gibbs sampling on the data of each divided interval according to the characteristic quantity category to obtain the distribution probability of each characteristic quantity. And acquiring sampling results of no fault, slight fault and serious fault of each characteristic quantity as a health evaluation interval divided for the first time.
The normal distribution, also called Gaussian distribution, is a probability distribution which is very important in the fields of mathematics, physics, engineering and the like, and has great influence on a plurality of aspects of statistics. The normal curve is bell-shaped, the two ends are low, the middle is high, and the left and right symmetry is bell-shaped, so people are often called bell-shaped curves.
3 Normal distribution curves, the interval of which is assumed to correspond to [ a1,b1]、[a2,b2]、[a3,b3 ] for no-fault, slight fault, severe fault normal curves, and the points of no-fault and slight fault, and severe fault are c1,c2. The 2 judgment ambiguity zones of no fault and slight fault, slight fault and serious fault are set as new division zones for warning from no fault to slight fault and slight fault to serious fault. The length L1,L2 of the corresponding blur area is:
L1＝min{b1-a1,b2-a2}·k％;L2＝min{b2-a2,b3-a3}·k％;
the corresponding ambiguity interval is:
[c1-L1,c1+L1];[c2-L2,c2+L2];
The normal distribution function of probability distribution can be determined by sampling other parameters in the first divided 3 intervals by the method, then the interval secondary division is realized by setting parameters, and the final interval is divided into 5 intervals. The interval dividing mode can realize a fault early warning function, and the interval dividing boundary can be adaptively adjusted according to an actual data set and requirements, and the processed interval is referred to in the following table 3.
TABLE 3 Table 3
| Dividing section | Numbering device | Interval value |
| No fault | 1 | [a1,c1-L1] |
| With the possibility of minor malfunctions | 2 | [c1-L1,c1+L1] |
| Occurrence of slight failure | 3 | [c1+L1,c2-L2] |
| With the possibility of serious faults | 4 | [c2-L2,c2+L2] |
| Possible serious faults | 5 | [c2+L2,b3] |
The improvement of the step comprises (1) data cleaning and preprocessing, namely introducing more advanced data cleaning technology, removing abnormal values, missing values and repeated values, and improving the data quality. (2) Fault classification refinement-a more detailed division of fault types to more accurately identify fault patterns. (3) Time sequence analysis, namely capturing the trend and periodicity of fault occurrence by using a time sequence analysis technology.
The method has the technical effects that (1) the data reliability is improved, and the influence of error data on an analysis result can be reduced by data cleaning and preprocessing. (2) Fault identification accuracy-detailed fault classification helps to identify fault modes more accurately and provides accurate information for subsequent steps. (3) Prediction accuracy-time series analysis can improve the prediction accuracy of future fault occurrence.
In step S103, a plurality of neural network models are constructed and trained based on the feature quantity and the health evaluation interval to obtain a plurality of prediction results of the transformer health condition, and the plurality of prediction results are combined to obtain a final prediction interval type. Constructing and training a plurality of neural network models based on the feature quantity and the health assessment interval to obtain a plurality of predicted results of the transformer health condition, combining the plurality of predicted results to obtain a final predicted interval type further comprises:
(1) Normalizing the fault related elements, namely normalizing the feature quantity:
Where xgy is a normalized value, x is a value of a feature quantity in the dataset, and xmin、xmax is a maximum value and a minimum value of the feature quantity, respectively.
(2) Establishing three neural network models of the feature quantity corresponding to the oil chromatographic data, the oil test data and the electric test data after normalization processing respectively, and training the three neural network models, wherein the output quantity of each neural network model is determined based on a health evaluation interval;
And combining a plurality of deep neural networks serving as submodels by adopting a feedforward neural network method, and finally coupling the output result by the submodel results. The sensitivity of the characteristic quantity of different types to different faults or health states of the transformer is also different, so that the advantage of classifying the different health states can be highlighted by respectively modeling the characteristic quantity of different types. The combined model can evaluate and screen the result according to the prediction condition of each sub-model, and the final prediction accuracy can be improved by mutually correcting different sub-models.
And obtaining training accuracy. And respectively building a neural network model from training data of oil chromatographic data, oil test data and electric test data to obtain 3 models, namely model 1 to model 3. And then when the optimal training precision is obtained, storing a training model, and outputting 5 divided intervals corresponding to the five processed intervals. Let X1,X2,…,Xn be the BP neural network input vector, Y1,Y2,…,Ym be the output value, wij and wjk be the weights, and the typical BP neural network topology structure diagram is shown in FIG. 2.
When the number of input nodes and the number of output nodes of the BP neural network are n and m respectively, the BP neural network reflects the mapping relation between n independent variables and m dependent variables. The BP neural network modeling prediction comprises three steps of network structural components, training and prediction, and the basic workflow is shown in figure 3.
Let the activation function of each layer node in the network be S-type function, and the input of the first layer i node in the network be neti, the output be oi, the output of the k node in the output layer be yk, then the input of the j node in the middle layer be:
oj＝f(netj);
defining the error of the network as the difference between the expected output and the actual output, there isIf the output layer has i neurons, the square error of the actual output and the expected output is defined as:
Since the BP algorithm corrects the weight according to the negative gradient of the error E, the modification of the weight can be expressed as Wm+1＝wm+Δwm＝wm-λgm;
where lambda is the step size of the learning, m represents the number of iterations,
Because it is the output layer, at this timeIs the actual output value, which is obtained from the definition of ek and the square error:
the definition according to ek can be:
According to the aboveThe method can obtain:
According to the aboveThe method can obtain:
Finally, the method comprises the following steps:
Let the learning error of the output layer be σk＝ekf′(netk now;
hidden layer neural unit weight modifier Δwkj:
According to the aboveAnd oj＝f(netj) available:
because it is the change in weighting of the hidden layer. In this case, the effect of the previous layer on it should be considered, so there are:
According toIt can be seen that
And according toThe method can obtain:
HandleCarry-in typeDeducing:
Let the learning error of hidden layer:
the model is then optimized by genetic algorithm (see fig. 4).
And predicting a final prediction interval of the transformer by using the trained multiple neural network models, wherein the final prediction interval comprises a non-fault interval, a slightly fault possible interval, a slightly fault interval, a severely fault possible interval and a severely fault possible interval.
And (3) establishing a combined model, namely combining all prediction results of the model 1 to the model 3 to obtain the final predicted state type. The rule is that, among the prediction results of model 1 to model 3, the partition section having the largest number of occurrences is the final prediction section. If the occurrence times of a plurality of divided sections are the same, the divided sections are weighted according to the optimal training precision of the model, and the divided section with the highest weight is the final prediction section.
And when the fault diagnosis is fault, the test results of the models 1 to 3 can be removed, so that the prediction accuracy is improved.
The improvement of the step comprises (1) model optimization, namely adopting a more advanced neural network structure such as a deep learning model to improve the accuracy of health state judgment. (2) And multi-source information fusion, namely providing richer input information for the neural network by combining other sensor data (such as temperature, vibration and the like) and environment data. (3) Model adaptivity, namely, the adaptivity of the model to new data and new fault modes is enhanced, and overfitting is avoided.
The method has the technical effects that (1) the judgment accuracy is improved, and the optimized neural network structure can provide more accurate health state judgment. (2) The multi-source information fusion can provide more comprehensive equipment state information, and is beneficial to improving the judgment accuracy. (3) The model adaptation is enhanced, so that the model adaptation can keep higher judgment accuracy when facing new data.
In step S104, a failure rate curve of the transformer is generated based on the final prediction interval type, and the failure rate curve is corrected based on the individual difference and the running state image. Generating a fault rate curve for the transformer based on the final prediction interval type, and correcting the fault rate curve based on the individual differences and the running state image further includes the following steps.
And calculating a conventional fault curve, namely acquiring shape parameters and scale parameters of each stage by using transformer fault data based on Weibull distribution to establish a fault rate curve of the transformer based on the shape parameters and the scale parameters.
The weibull distribution is the theoretical basis for reliability analysis and life testing. The Weibull distribution is widely applied in reliability engineering, and is particularly suitable for the distribution form of wear accumulation failure of electromechanical products. Since it can easily infer its distribution parameters using probability values, it is widely used for data processing of various life tests.
Wherein lambda (t) is the life of the transformer, t is time,β is a scale parameter, and alpha is a shape parameter. By selecting different scale parameters, the change trend of the failure rate in different periods can be described, wherein when beta is less than 1, the failure rate is in a descending trend and can correspond to the early failure period of the bathtub curve, when beta is more than 1, the failure rate is in an ascending trend and can correspond to the loss failure period of the bathtub curve, and when beta is less than 1, the failure rate is constant and can correspond to the accidental failure period of the bathtub curve. Based on the equipment failure history data, the shape parameter alpha and the scale parameterβ of each stage are obtained through fitting the Weibull distribution corresponding to the equipment failure rate in a segmented mode, so that an equipment basic failure rate model is established.
According to the actual running condition of the transformer, the fault rate curves of the transformers produced by the same manufacturer are subjected to linear correction to obtain a correction fault rate function considering the familial defect, and then the correction fault rate curves considering the individual defect are obtained based on the correction fault rate curves considering the familial defect.
The correction of individual differences is considered, namely, the operation condition and the aging failure process of the produced equipment are quite different due to the different manufacturing processes of different factories. In addition, when calculating the equipment failure rate curve, the shape parameter and the scale parameter of the Weibull distribution of the equipment are difficult to accurately obtain due to the limited number of equipment produced by a certain manufacturer. Therefore, the basic fault rate curve needs to be corrected according to the statistical data of the equipment produced by different manufacturers, so that the influence of familial defects on the equipment fault rate is reflected. In engineering application, the actual average failure-free time of the production equipment of the same manufacturer can be compared with the integral average failure-free time of the same type of equipment, and the ratio of the actual average failure-free time to the integral average failure-free time of the production equipment of the same manufacturer is adopted to carry out linearization correction on the basic failure rate of the same type of product.
The corrected failure rate curve after considering familial defects is expressed as:
Similar to familial defects, in order to reflect individual differences among different devices produced by the same manufacturer, the failure rate curves of the devices produced by the same manufacturer need to be corrected linearly according to the actual running conditions of the devices, so that the influence of the individual defects on the failure rate of the devices is reflected.
The corrected failure rate curve after considering the individual defects is expressed as;
Wherein gt (t) is a correction failure rate function considering individual defects, wu is the average no-failure time of an individual transformer, t2 is the operation period of the transformer, n2 is the number of times the transformer fails during operation, qx (t) is the correction failure rate function considering familial defects, jc (t) is the failure rate function obtained by fitting according to historical data of similar transformers, wf is the average no-failure time of similar transformers, wt is the average no-failure time of transformers produced by the same manufacturer, t1_i is the operation period of the ith transformer in the similar transformers, n1 is the counted total number of times the ith transformer fails during operation, and t0_i is the total number of times the transformer fails during operation.
The correction of the operation image is considered, namely, the operation state image is constructed, and the fault rate curve of the transformer is corrected based on the operation state image so as to obtain a corrected fault rate curve after the operation image is considered. Constructing an operating state representation and correcting a fault rate curve of the transformer based on the operating state representation to obtain a corrected fault rate curve, further comprises the steps of.
1. And constructing an operation portrait system, namely performing index objective extraction by utilizing correlation analysis and Laplace scoring to select a preset number of indexes with different correlation strengths with other indexes and maximum Laplace scoring as main characteristic indexes.
Feature extraction is the most critical link in the grid portrait construction process, is the integration of characteristics or commonalities, and extracts main information, reduces dimensionality and simplifies a calculation model by carrying out similarity calculation and modeling on a fact label system under each attribute. The extracted thought is generally to reject or combine the features with higher correlation, each feature is required to describe the grid independently and clearly, and if two variables with higher correlation are added into the model as the features at the same time, overfitting is caused, and the complexity of the model is increased. The accuracy of the result is also affected by the feature extraction method, and common methods include mapping method, selecting method, objective extraction method, and the like. The index objective extraction is carried out by adopting a correlation analysis and Laplace scoring method, and the analysis basic flow is as follows:
(1) And (5) data standardization processing. Each index extracted from the fact label library comprises an index value and a hierarchical semantic label value. The index value distributions are different, and in order to reduce the calculation error of the correlation coefficient, the deviation-kurtosis test or the approximate normal distribution processing is needed. After normalizing the normal distribution of the variable X, it is converted into X:
wherein: the mean value of the variable X, and sigma is the standard deviation of the variable X.
(2) And calculating the correlation coefficient between indexes. The Pearson correlation coefficient is a calculation method for measuring the linear correlation of two groups of discrete variables. The Pearson correlation coefficients for the two sets of variables x, y are defined as:
Wherein, sigmaxy is the covariance of the variables x and y, and sigma (x) and sigma (y) are the standard deviations of the variables x and y respectively.
(3) Laplace score calculation. Firstly, constructing a weight matrix Z, enabling LSr to be Laplace scoring of the r-th characteristic, fri represents an i-th sample of an r-th feature (i=1, 2., m). For each feature, an m×m adjacency matrix Z is respectively constructed, and the values of elements in Z are as follows:
Wherein xi,xj is the value of the ith and jth samples of the feature, and t is a proper constant.
Finally, calculating Laplace score:
Wherein fr＝[fr1,fr2,…,frm]T.
(4) And selecting characteristic indexes. Based on Laplace scoring and correlation analysis results, selecting a plurality of indexes with obvious differences of correlation strength with other indexes and higher scoring as main characteristic indexes.
(5) Model tag generation. The model labels are labeled texts which are further extracted through abstract clustering and other modes based on the extracted main characteristic indexes, can refine and accurately express grid characteristics, are easy to understand and apply, and are further generalized and summarized for operation characteristics. The model tag provides convenience for management and optimization services of equipment faults while identifying operating characteristics.
An index tag is generated based on the main feature index, the index tag including indexes corresponding to the region attribute, the natural attribute, and the operation attribute, respectively.
(6) And (5) constructing a portrait. An operation state portrait is constructed based on the index tag, and includes a region attribute, a natural attribute, and an operation attribute (refer to table 4 below).
TABLE 4 Table 4
And calculating the weight value of the index label under each attribute dimension by an entropy weight method, and carrying out weighting processing and sorting on the index label based on the weight value of the index label so as to obtain a corresponding label grade.
In summary, through modeling analysis, the running state portrait finally has the model tag attributes of three dimensions of region attribute, natural attribute and running attribute, and an entropy weight method is introduced on the basis of the model tag to calculate the model tag weight under each attribute dimension, and the entropy weight method weight calculation steps are as follows:
The calculation steps of the entropy weight method are roughly divided into three steps, namely 1) judging whether negative numbers exist in an input matrix, and if so, re-normalizing to a non-negative interval. 2) The specific gravity of the ith sample under the jth index is calculated and taken as the probability used in the calculation of the relative entropy. 3) And calculating the information entropy of each index, calculating the information utility value, and normalizing to obtain the entropy weight of each index. The corresponding label ranking is found in the percentile order after weighting the scores (see table 5 below).
TABLE 5
2. The main purpose of the correlation analysis is to study the degree of closeness of the relations between variables, such as the relations between height and weight, the wire quantity and tower quantity, the main transformer capacity and the distribution device. In statistical analysis, correlation is generally referred to as "linear correlation", the degree of closeness of which is represented by a correlation coefficient. The correlation coefficient is usually denoted as r, and takes a value between-1 and +1, and the closer the absolute value is to 1, the closer the relationship between the variables is. The absolute value is equal to 1, which indicates that the two variables are completely related, and the value of the B variable can be obtained by knowing the value of the A variable. The correlation coefficient is positive, which means that the B variable is increased simultaneously when the A variable is increased, and the B variable is decreased when the A variable is increased, and the B variable and the A variable are in negative correlation.
Pearson's simple correlation coefficient is used to measure the linear correlation of distance variables, and is most widely used in calculating the correlation coefficient. The calculation formula is as follows:
Where n is the number of samples, xi and yi are the values of two variables in different samples,AndThe mean of variables xi and yi, respectively. Since the calculation formula of the Pearson simple correlation coefficient is exactly in the form of a matrix product, it is also called a product distance correlation coefficient. The correlation coefficient can be expressed as the average of n products, which is calculated by normalizing xi and yi, respectively.
3. The corrected fault curve comprises the steps of carrying out correlation analysis on images of different running states and the fault rate of the transformer to obtain an adjustment coefficient g, and generating the following corrected fault rate curve considering the running images based on the adjustment coefficient, the fault rate curve of the transformer, the corrected fault rate curve considering familial defects and the corrected fault rate curve considering individual defects:
where g is an adjustment coefficient, HX1 is a failure rate of the corresponding image, and HX2 is an average failure rate.
The improvement of the step is that (1) more influencing factors are considered, namely, besides individual defects, overhaul behaviors and running states, other factors which influence the failure rate, such as environmental factors, load changes and the like, can be considered. (2) Dynamic correction, in which the fault rate curve is updated and corrected periodically as the running time of the equipment increases so as to reflect the change of the state of the equipment. (3) Uncertainty analysis is introduced, namely uncertainty analysis is carried out on the fault rate curve so as to evaluate the reliability and stability of the prediction result.
The method has the technical effects that (1) more accurate prediction is realized, and more influencing factors are considered, so that the accuracy of a fault rate curve can be improved. (2) And updating in real time, wherein the dynamic correction can ensure that the fault rate curve always reflects the latest state of the equipment. (3) Risk assessment-uncertainty analysis helps assess the risk and reliability of the predicted outcome.
In step S105, a maintenance strategy decision model is constructed to obtain an optimal maintenance time based on the corrected failure rate curve and the minimum cost and reliability constraints. Constructing an overhaul strategy decision model based on the corrected fault rate curve and minimum cost and reliability constraints to obtain a final overhaul time plan (constructing an optimized overhaul strategy decision model based on the corrected fault rate curve and double consideration of cost and reliability) further comprises constructing an overhaul strategy decision model based on overhaul risks and fault risks and constraint conditions thereof, wherein overhaul risks are determined based on the random loss of load of a power grid and the planned loss of load under an n-period overhaul mode m, and failure risks are determined based on the random loss of load of the power grid and the loss of the power grid under the n-period overhaul mode m, constraint conditions comprise simultaneous overhaul constraints, mutually exclusive overhaul constraints, overhaul resource constraints and power grid safety constraints, wherein the simultaneous overhaul constraints refer to constraints of simultaneously overhaul of the transformer to avoid repeated power failure caused by overhaul of the transformer, the mutually exclusive overhaul constraints refer to constraints of overhaul of the transformer not at the same time to avoid power failure, the overhaul resource constraints refer to the total number of overhaul transformers cannot exceed overhaul capacity of overhaul staff, and the power grid safety constraints refer to constraints of safety inspection through load flow calculation. And determining an overhaul time point which can ensure the safe and stable operation of the transformer and reach the optimal economic cost through solving the model.
(1) And the power grid overhaul risk is that the state overhaul is preventive overhaul and is carried out before equipment failure. Thus creating another loss, i.e. a risk of overhaul of the grid, while reducing equipment failure losses.
OM1＝pmPM1,mTmcm1;
Wherein OM is maintenance risk, OM (n) is maintenance risk of n time periods, T is time period number of one period, Mn is a T time period maintenance mode set, OM1,m and OM2,m are power grid random load loss and planned load loss under n time period maintenance mode M, Pm is probability of occurrence of the maintenance mode M, PM1,m is planned load loss caused by the T time period maintenance mode M, Tm is power outage time caused by maintenance, Cm1 is a unit pay of planned load loss on a demand side, Fn is a maintenance equipment set under the maintenance mode M, and Ck is maintenance cost of equipment k under the maintenance mode M.
(2) And the power grid fault risk is that other equipment except overhauling equipment can possibly break down at any time, so that power grid fault loss is caused.
Wherein OF is a fault risk, OF (N) is a t-period power grid fault risk, OF1,f is power grid random load loss in a t-period m overhaul mode, OF2,f is equipment self loss, GZX (t) is equipment fault rate, Pf,m is power grid random load loss amount, and NF is a fault equipment set. Mu is repair probability, Cmj is fault repair cost of the equipment j, COj is fault replacement cost of the equipment j, power failure time caused by Tf fault f, and Cm2 is the loss of load unit planning on the demand side.
(3) The power distribution network maintenance plan optimization model aims to minimize the maintenance risk of a power distribution transformer, wherein the maintenance risk OM and the fault risk OF exist during maintenance of the power distribution transformer, and the maintenance risk is increased by reducing the fault risk, namely the maintenance risk is contradictory. Both, however, are unified with safe and economical operation of the distribution network. Thus, the objective function expression is as follows:
F=min(OF+OM);
The constraint conditions are as follows:
1) While maintaining the constraints. The repeated power failure caused by equipment overhaul is avoided to improve the power supply reliability, the overhaul plan should avoid repeated overhaul operation as much as possible, and repeated power failure caused by the fact that one power failure can be solved is not allowed. Thus, some equipment must be serviced at the same time.
ttx＝tty;
2) Mutually exclusive service constraints. In order to avoid the power failure which can be avoided during maintenance, maintenance of some equipment is not suitable to be arranged at the same time.
tty＞ttx+Tx+1;
3) And (5) overhauling the resource constraint. The overhaul resource constraint refers to the constraint of the number of overhaul personnel, the overhaul technical capacity and the like. Meanwhile, the overhauling equipment cannot exceed the overhauling capacity of overhauling personnel.
In the formula, m is the total number of the devices, usn is a device maintenance state variable in a time period t, the value of the device in the maintenance state is 1, and otherwise, 0 is taken. M is the upper limit of the number of overhaul equipment per time period.
4) And (5) safety constraint of the power grid. The equipment overhauls and exits from running, and the power flow can be changed, so that some lines can be overloaded, and the node voltage is out of limit. Thus, a safety check must be performed by means of the load flow calculation.
Pl≤plmax;
Uqmin≤Uq≤Uqmax;
Wherein Pl is the power of the transformer l, Plmax is the maximum power allowed to pass through the transformer l, and Uq、Uqmin and Uqmax are the node q voltage and the upper and lower limit values of the node q voltage respectively.
(4) The marine predator algorithm is a novel meta-heuristic optimization algorithm, and various marine organisms continuously convert the identities of predators and prey objects of the marine organisms by simulating natural laws of survival of the marine organisms in the ocean, and the foraging strategies are converted according to different situations, so that the optimization process is carried out.
The marine predator algorithm considers that the foraging strategy of the marine predator changes between flying and Brownian walking, and selects between the two strategies according to different scenes, so that the optimal predation strategy is obtained.
First, the predators and the positions of the prey are initialized, the predators with optimal fitness construct elite matrices, and the prey matrices are constructed by evenly distributed prey, and the prey matrices are respectively shown in the following formulas.
The optimization process of MPA is divided into three phases according to different speed ratios.
Stage 1. Survey stage this stage, also known as the high speed ratio stage, is where the speed of the prey is much faster than the prey, who takes a stationary strategy and the prey performs brownian motion. This stage often occurs at the beginning of an algorithm optimization iteration, where a survey is made of global position information. The mathematical model of this stage is expressed as follows:
Wherein the method comprises the steps ofIs the moving step length; is a random vector based on Brownian walk normal distribution; p is a constant equal to 0.5; Is a uniform random vector in [0,1], iter is the current iteration number, max_Iter is the maximum iteration number, and N is the population number.
Stage 2 this stage is also known as the medium speed ratio stage, where the speed between predator and prey is comparable, both predator and prey are looking for their own prey. This stage generally occurs in the middle of the algorithm iteration, where the population is divided into two parts, where the prey makes levy flights, responsible for the development in space, the predator makes brownian motion, and responsible for the survey in space. The mathematical model expression at this stage is as follows:
Wherein, theRepresenting the lewy motion random vector, CF represents the adaptive parameters that control the predator movement step size.
The development stage is also called as a low custom ratio stage, wherein the speed of predators is higher than that of prey, mainly occurs in the later stage of algorithm iteration, and the predators adopt Levy flight strategies, so that development work of local areas is focused more. The mathematical model expression at this stage is as follows:
In addition, the algorithm also considers external environmental factors such as a shoal gathering device, a vortex effect and the like, and changes the foraging strategy of predators so as to jump out of a local extremum, thereby avoiding the problem of convergence and premature. The mathematical model expression is as follows:
wherein FADs are influence probabilities, and generally 0.2 is taken; Is a binary vector, r is a random number in [0,1], and r1、r2 is a random index of the hunting matrix respectively. And finally obtaining an optimal maintenance plan according to the solution.
Improvements to this step include (1) cost-benefit analysis, which may take into account long-term benefits of service, such as reduced downtime, improved equipment reliability, etc., in addition to the cost of service. (2) Multi-objective optimization-incorporating multiple objectives (e.g., cost, reliability, security, etc.) into an optimization model, seeking an overall optimal solution. (3) Algorithm optimization, namely adopting a more advanced optimization algorithm (such as a genetic algorithm, a simulated annealing algorithm and the like) to improve the solving efficiency and accuracy.
The method has the technical effects that (1) the comprehensive optimization and the cost-benefit analysis can ensure that the overhaul strategy achieves the optimal economic and long-term benefits. (2) Multi-objective balance multi-objective optimization can help find the best balance point among multiple objectives. (3) And the algorithm optimization can improve the solving efficiency and accuracy of the overhaul strategy decision model.
Referring to fig. 5, a transformer maintenance decision system based on state evaluation and fault rate correction is disclosed, which comprises a data acquisition and processing module 501 for acquiring and preprocessing transformer fault data to acquire feature values corresponding to oil chromatogram data, oil test data and electrical test data, a health evaluation interval acquisition module 502 for dividing the transformer fault data type into primary division intervals by using the preprocessed historical transformer fault data and then secondarily dividing the primary division intervals into health evaluation intervals based on the primary division intervals and setting parameters, an interval prediction module 503 for constructing and training a plurality of neural network models based on the feature values and the health evaluation intervals to acquire a plurality of prediction results of the transformer health conditions, a fault curve generation module 504 for generating a fault rate curve of the transformer based on the final prediction interval types and correcting the fault rate curve based on individual differences and running state images, and an optimal time determination module 505 for constructing an optimal maintenance time constraint model based on the corrected fault rate curve and minimum decision and the minimum maintenance cost.
Those skilled in the art will appreciate that all or part of the flow of the methods of the embodiments described above may be accomplished by way of a computer program to instruct associated hardware, where the program may be stored on a computer readable storage medium. Wherein the computer readable storage medium is a magnetic disk, an optical disk, a read-only memory or a random access memory, etc.
The present invention is not limited to the above-mentioned embodiments, and any changes or substitutions that can be easily understood by those skilled in the art within the technical scope of the present invention are intended to be included in the scope of the present invention.