CN111579731A

CN111579731A - Early transformer defect early warning method based on combined model

Info

Publication number: CN111579731A
Application number: CN202010598927.5A
Authority: CN
Inventors: 方瑞明; 张燕
Original assignee: Huaqiao University
Current assignee: Huaqiao University
Priority date: 2020-06-28
Filing date: 2020-06-28
Publication date: 2020-08-25
Anticipated expiration: 2040-06-28
Also published as: CN111579731B

Abstract

A transformer early defect early warning method of a combined model specifically comprises the following steps: screening abnormal characteristic gas based on a hidden Markov model by utilizing time sequence data of monitoring dissolved gas components in oil of a transformer to be monitored, wherein the screened abnormal gas is related to system state change and is used as a key node closely related to the system state change; on the basis of a key network consisting of the key nodes, judging whether the key network accords with critical characteristics according to a dynamic network marker theory so as to evaluate the health state of the transformer and early warn early defects; and the built model uses the self time sequence data of the transformer to be diagnosed, so that the generalization problem does not exist, and early defect early warning and identification are facilitated for the transformer.

Description

Early transformer defect early warning method based on combined model

Technical Field

The invention relates to the technical field of power transformer defect early warning and fault diagnosis, in particular to a transformer early defect early warning method based on a combined model.

Background

The power transformer is the most basic and important power grid equipment in a power system, and is also equipment which is very easy to cause accidents, defect early warning and fault diagnosis of the transformer are important contents for ensuring stable operation of the power system, wherein monitoring and analyzing Dissolved Gas (DGA) in oil plays an important role in ensuring safe and stable operation of the transformer.

The main problems of identifying the operating state of the transformer are as follows: 1) the factors influencing the insulation degradation of the transformer are more, such as environment, materials, working age and the like, and the individual differences caused by various influencing factors make it difficult to find universal standards or threshold values to identify the transition of the operation state of the transformer; 2) the mapping relation between the external observation amount (concentration change of dissolved gas in oil) and the internal defects of the transformer is complex, and accurate quantitative description and identification are difficult to carry out based on some existing experiences.

Hidden Markov Models (HMMs) are time-sequential probabilistic models that describe the process of generating a sequence of unobservable states from a hidden Markov chain and then generating an observation sequence from the sequence of states. Wherein, the conversion between the states and the observation sequence and the state sequence have certain probability relation. The state change of the transformer is a time-varying process, and the running state of the transformer can be indirectly judged according to the concentration change condition of the dissolved gas in the oil. If three operating states in the operation of the transformer are described as different markov processes, the transformer can be described as a discrete hidden markov model comprising two hidden states and eight observable states due to different dynamic characteristics between the normal operating state and the critical state, wherein the hidden markov process is a system state, and the observable states are the component and concentration changes of the gas dissolved in the oil corresponding to the normal operating state and the defect (critical) state.

In addition, scientists have discovered in recent years that Critical Slowing Down (CSD) has demonstrated significant potential in revealing whether complex power systems are prone to critical catastrophe. Critical slowness is a concept in statistical physics, which means that before a power system changes from one phase state to another phase state, the system approaches to the vicinity of a critical point, and particularly, a dispersion fluctuation phenomenon favorable for the formation of a new phase occurs at the critical point, and the phenomenon can cause three possible early warning signals in dynamics: the recovery of the disturbance becomes slow, the autocorrelation coefficient increases, and the variance increases. However, CSD is suitable for univariate systems with low noise interference, so experts have proposed to use Dynamic Network Markers (DNM) to describe the dynamics of multivariate or networks. The transformer fault occurrence process is a dynamic process, a critical state also exists in the process of converting from a healthy state to a fault state, and when the content of the characteristic gas dissolved in oil is increased to a certain degree and reaches a critical point, a system is suddenly converted from a stable state to an unstable state, and the critical conversion phenomenon occurs.

The two methods have the advantages that the dynamic process of the running state of the transformer can be judged by utilizing a DNM theory, the critical point of state transition is judged through the dynamic change of the correlation among the screened gases, the HMM is adopted to monitor the concentration of the characteristic gases related to the state change of the transformer, and the screening process can be carried out according to the self time sequence data of the transformer without collecting typical gas fault sample data or establishing a related gas prediction model. Therefore, the dynamic change of the concentration of the dissolved gas in the oil is monitored by the oil chromatogram online monitoring system of the transformer to be diagnosed, and the early warning signal appearing in the state transition of the transformer is detected based on the HMM and the DNM, so that early defect early warning and identification can be performed on the transformer.

Disclosure of Invention

The invention mainly aims to overcome the defects in the prior art and provides an early-stage defect early-warning method of a combined transformer based on a hidden Markov model and a dynamic network marker model, wherein the model is constructed based on the oil chromatogram online monitoring data of the transformer to be diagnosed, fault sample data in a typical state does not need to be collected, and the data is easy to obtain; and the built model uses the self time sequence data of the transformer to be diagnosed, so that the generalization problem does not exist, and early defect early warning and identification are facilitated for the transformer.

The technical scheme adopted by the invention for solving the technical problems is as follows:

a transformer early defect early warning method of a combined model comprises the steps of firstly screening characteristic gas with abnormal concentration based on an HMM, using the screened abnormal gas as a key node closely related to system state change and judging whether a sub-network formed by the screened abnormal gas meets critical characteristics or not according to a DNM theory, wherein the characteristic gas is related to the system state change, and the method comprises the following steps:

based on time sequence data of each characteristic gas monitored by an online monitoring device for dissolved gas in transformer oil, taking G sampling points including current time T as a starting point time period T, and reading data of G time periods (G is more than or equal to 1) forwards, wherein each time period comprises G sampling points (G is more than or equal to 4);

step II, taking the gas data group in the time period T-1 as a comparison group, screening the abnormal gas quantity of the gas data group in the time period T on each sampling point in the time period, and forming an abnormal group O in the time period T_T；

Step III, all the abnormal groups obtained from the time period T-G to the time period T-1 are combined into a new abnormal gas quantity group C_T-1(C_T-1＝{O_T-G,…,O_T-2,O_T-1) } taking the training set as a training set and establishing a hidden Markov model;

step IV, calculating the probability I of abnormal transition of the time period T of the current time T according to the built hidden Markov model_T：

I_T＝1-P_T(1)

Wherein P is_TRepresents from C_T-1Generating an exception group O for a time period T_TThe probability of (d);

step V, if I_TWhen the voltage rises suddenly, the system is probably close to a critical point, and at the moment, key nodes with large influence on the transformer dynamic network are further screened in the period, namely (I)_T-I_T-1)>Q (considering only abnormally rising gas, i.e. gas exceeding the upper limit), the transformer state is consideredAbnormal change occurs in the time interval T, the generality is not lost, and I exists when the transformer is in a normal state_TLess than or equal to 0.9, and when the transformer is in abnormal state, 0.9<I_TTaking Q as 0.1 because the ratio is less than or equal to 1;

step VI, calculating the average standard deviation of the critical network (DNM) composed of the critical nodes at each time interval

Average Pearson correlation coefficient among nodes in key network

And average Pearson correlation coefficient between nodes in the critical network and other non-critical nodes

Judging whether the critical characteristic is met:

1. mean standard deviation of critical networks

Is remarkably increased.

2. Average Pearson correlation coefficient between variables in critical network

And (4) increasing.

3. Average Pearson correlation coefficient between variable in critical network and non-critical network variable

And decreases.

And obtaining a quantized value I 'of the key network marker in each period (T) according to the calculation result'_T：

In the formula, the number is a small normal number used for avoiding the denominator being zero, and the denominator can be omitted if the denominator is not zero;

step VII, if I'_TMutation (without loss of generality of (I'_T-I′_T-1)>0.05 and (I'_T-I′_T+1)>0.05), meaning that the complex network corresponding to the transformer has critical change in the state of the period (T), from the steady state I'_T-1Through critical state I'_TConversion to Defect State I'_T+1And at the moment, sending out an early warning signal, and identifying the defect type of the transformer according to the formed dynamic network marker.

Further, taking the case plot method as an example, the step II is to set the gas data set N with the time interval T-1 and the time interval T being 7 × g respectively_T-1And N_TWherein 7 represents the amount of gas, detecting N_TWhether or not the (i, j) th data of (1) is abnormal data:

step II-1, forming a new time sequence by the (i, j) th data and the ith row sequence of the time period T-1, and arranging the data in a sequence from small to large;

and II-2, calculating an upper quartile Q3 and a lower quartile Q1 of the group of data, and calculating a quartile distance according to Q1 and Q3:

IQR＝Q3-Q1 (3)

wherein, Q1 (the first quartile) and Q3 (the third quartile) are respectively the 25 th and 75 th data after being arranged from small to large, and the difference between Q1 and Q3 is also called as quartile distance (IQR).

And II-3, if the detected data is smaller than the value of Q1-1.5 XIQR (lower limit) or larger than the value of Q3+1.5 XIQR (upper limit), the data is considered as abnormal data (the abnormal reduction of the gas can be properly ignored because the fault characteristic gas is abnormally increased due to the fault of the transformer, and only the value larger than the upper limit is considered), and the gas content is proved to be abnormally changed at the moment.

Further: step III is to specifically set two hidden states of the transformer as S₀(Normal) and S₁(Defect) (the state of time period T is denoted as s_T(s_T∈{S₀,S₁) }) state and state-to-state satisfying the markov property), the externally observable state is the abnormal change of the concentration of the characteristic gas of the transformer, and is respectively V ═ 0,1, …7 (for example, if there is an abnormal rise in 3 characteristic gases at the first sampling point of the period T, the number of abnormal changes in concentration at that time is 3, and V is 3), and the obtained observation sequence of the number of abnormal gases { O ═ 3 } is applied to all eight states_T-G,…,O_T-2,O_T-1Training HMM lambda of time T-1 by adopting Baum-Welch algorithm_T-1＝(A_T-1,B_T-1,π_T-1)：

Wherein: a denotes the transition probability matrix of the hidden state. It contains the probability from one hidden state i to another hidden state j, denoted as a ═ a_ij}_N×NWherein a is_ij＝P(s_T+1＝S_j|s_T＝S_i) (1. ltoreq. i, j. ltoreq. N) (N represents the number of hidden states) and has

B denotes an observed state probability matrix (confusion matrix). It represents the probability distribution of observed states observed in hidden state j of a given HMM over a period T, denoted B ═ B_jk}_N×MWherein b is_jk＝P(V_k＝O_T|s_T＝S_j) (1. ltoreq. i.ltoreq.N, 1. ltoreq. k.ltoreq.M) (M denotes the number of observable states corresponding to each state, and each observable state is denoted by V ═ V ≦ M₁,V₂,…,V_MThe observed state of the time period T is represented as O_T(O_T∈{V₁,V₂,…,V_M}) and have

And pi represents the probability matrix of the initial state. Probability matrix representing hidden state in initial period, noted

π＝{π_i＝P(s_T-G＝S_i)}_1×N,(1≤i≤N)。

Step III-1, initialization: when n is equal to 0 (note that the initial period is 0), give

And

giving an initial value, there is HMM lambda⁰＝(A⁰,B⁰,π⁰)；

Step III-2. recursion: when n is 1,2, …,

and

is calculated as follows:

wherein gamma is_T′(i) Indicating that for a given HMM and observation sequence O, the system is in state S for time period T_iProbability of (c):

ζ_T′(i, j) indicates that for a given HMM and observation sequence O, the system is in state S for time period T' -1_iAnd is in state S for a period T_jProbability of (c):

step III-3, terminating: when n is equal to H (the H-th calculation update), the recursion terminates, with

Then the trained HMM is

Further: said step IV is in particular an HMM lambda for a given model_T-1＝(A_T-1,B_T-1,π_T-1) And observation sequence O ═ O_T-G,…,O_T-1,O_TAnd if the transformer is still in a normal state in the time period T, calculating the probability of generating the sequence:

wherein

Let transformer HMM λ ═ (a, B, pi) (a ═ a_ij}_2×2，B＝{b_jk}_2×8，π＝{π_i}_{1×2,i∈{0,1}}) Observation sequence O ═ O_T-G,…,O_T-1,O_TThen forward probability α in equation:

α_T(i)＝α_T(s_T＝S_i)＝P(O_T-G,K,O_T-1,O_T,s_T＝S_i|λ),i∈{0,1} (13)

calculated by the forward algorithm:

step IV-1, initialization: when T is T-G, the compound is,

step IV-2. recursion: T-G +1, …, T-1,

step IV-3, terminating: when T is equal to T, the reaction solution is mixed,

then there is

Further: the step VI is specifically as follows:

for the first critical characteristic, if any

Adopting a multiple change method to judge whether the first critical characteristic is met:

wherein

Mean value of the key network standard deviation representing the time period (T +1), where

And

the average values of the standard deviations of the key networks in the time period (T) and the time period (T-1) respectively usually take 2-3 times of difference as a threshold value, and when F is reached_c<2, the increase is considered to be small, and 2<F_c<4 is considered to be increased when F_c>4, the increase is considered to be large;

if there is

And is

The average standard deviation of the critical network reaches a small peak and the increase is considered significant, meeting the first critical characteristic.

The invention has the following beneficial effects:

the hidden Markov model established according to the on-line monitoring data of the transformer detects the possible critical point of the gas change of the abnormal state transition of the system, further screens key gas molecules capable of reflecting the transition of the system to the fault state, serves as a dynamic network marker of the system, observes the change trend of the dynamic marker, judges whether the transformer approaches the critical transition point or not according to the key gas molecules, and identifies the possible defect causing the state transition of the transformer.

The invention is further described in detail with reference to the drawings and the embodiments, but the method for early warning the defect of the oil-immersed power transformer based on the hidden markov model and the dynamic network marker combination model is not limited to the embodiments.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a graph illustrating the HMM probability calculations for various time intervals according to the present invention; wherein FIG. 2(a) is the HMM probability calculation for 170 sets of chromatographic data; FIG. 2(b) is a calculation of HMM probability for 50 sets of chromatogram data truncated for the first occurrence of a large abnormal transition probability;

FIG. 3 shows the analysis result of the present invention with the anomalous gas of period 4 as DNM; wherein FIG. 3(a) is the P and I' values for each time period; FIG. 3(b) shows the respective periods

And

the value is obtained.

Detailed Description

The invention is further described below by means of specific embodiments.

Referring to fig. 1, the transformer used in the present embodiment is a main transformer of a 220kV substation No. 1 of taiyuan power supply branch company of power company in shanxi province. The main gas monitoring data are: h₂、CO、CH₄、C₂H₂、C₂H₄、C₂H₆、CO₂The monitoring step length is 24 hours in the early stage, the sampling interval is shortened in the later stage, and the monitoring step length is 7 to 10 hours. Table 1 shows a characteristic gas set of the oil-immersed transformer in different fault types, and lists the number of primary and secondary abnormal gases corresponding to each fault type. The invention is explained by taking MATLAB as a working platform, wherein G is 1 and G is 5. Selecting chromatographic monitoring data of the transformer main transformer from 8 months 16 days to 10 months 23 days in 2009 for analysis, wherein 170 groups of continuous sampling point chromatographic data are obtained; and setting every 5 continuous sampling points as a time period, and performing anomaly analysis on data of each time period by using an HMM (hidden Markov model) for 34 time periods. Table 2 shows the numbers of the gas monitoring items.

Table 1 variation of different fault gas compositions and quantities:

table 2 gas monitoring items:

based on each characteristic gas monitoring data collected by the online monitoring device for the dissolved gas in the transformer oil, the abnormal gas quantity on each sampling point in the time period T (T is more than or equal to 2) is screened out by adopting a box-line graph method to form an abnormal group O of the time period T_TThe specific process is as follows:

let the gas data set N have a period T-1 and a period T of 7 × g (g is 5 in this embodiment) respectively_T-1And N_TWherein 7 representsAmount of gas, detecting N_TWhether the data (i, j) of (1) is abnormal data (since the fault characteristic gas is abnormally increased due to the transformer fault, when the box plot method is adopted to judge the quantity of the abnormal gas, the abnormally reduced gas can be properly ignored):

step 1: forming a new time sequence array by the (i, j) th data and the ith row array of the time period T-1, and arranging the data and the ith row array in the order from small to large;

step 2: calculating an upper quartile Q3 and a lower quartile Q1 of the group of data, and calculating a quartile distance according to Q1 and Q3:

IQR＝Q3-Q1 (3)

And step 3: if the detected data is less than the value of Q1-1.5 XIQR (lower limit) or greater than the value of Q3+1.5 XIQR (upper limit), the data is considered abnormal data, indicating that the gas content has changed abnormally at that time.

All the abnormal groups obtained from the time interval T-G to the time interval T-1 are combined into a new abnormal gas quantity group C_T-1(in this example, G is 1, so C_T-1＝{O_T-G,…,O_T-2,O_T-1}＝{O_T-1And } taking the hidden Markov model as a training set to establish a hidden Markov model, specifically:

let two hidden states of the transformer be S₀(Normal) and S₁(Defect) (the state at time T is denoted as s_T(s_T∈{S₀,S₁) }), the states satisfy the markov property from one state to another), and the externally observable states are abnormal changes of the concentration of the characteristic gas of the transformer, namely, eight states in total, namely {0,1, …,7} respectively (for example: if there are 3 kinds of characteristic gas abnormal rises at the first sampling point of the time period T, the concentration abnormal change quantity at the moment is 3, and V is 3), and the observation sequence C of the obtained abnormal gas quantity_T-1Training HMM lambda of time T-1 by Baum-Welch algorithm_T-1＝(A_T-1,B_T-1,π_T-1)：

And pi represents the probability matrix of the initial state. A probability matrix representing the hidden state in the initial period, denoted as pi ═ pi { (pi)_i＝P(s_T-G＝S_i)}_1×N,(1≤i≤N)。

Step 1: initialization: when n is equal to 0, give

And

giving an initial value, there is HMM lambda⁰＝(A⁰,B⁰,π⁰)；

Step 2: recursion: when n is 1,2, …,

and

is calculated as follows:

and step 3: and (4) terminating: when n ═ H, the recursion terminates, this time with

Then the trained HMM is

Further according to the built hidden Markov model, calculating the probability I of abnormal transition of the time period T of the current time T_T：

I_T＝1-P_T(1)

Wherein P is_TRepresents from C_T-1Generating an exception group O for a time period T_TThe specific process of probability calculation is as follows:

HMM λ for a given model_T-1＝(A_T-1,B_T-1,π_T-1) And observation sequence O ═ O_T-G,…,O_T-1,O_TIn this embodiment, if the transformer is still in a normal state in the time period T, the probability of generating the sequence is calculated as follows:

wherein

Let transformer HMM λ ═ (a, B, pi) (a ═ a_ij}_2×2，B＝{b_jk}_2×7，π＝{π_i}_{1×2,i∈{0,1}}) Observation sequence O ═ O_T-G,…,O_T-1,O_TThen forward probability α in equation:

α_T(i)＝α_T(s_T＝S_i)＝P(O_T-G,K,O_T-1,O_T,s_T＝S_i|λ),i∈{0,1} (13)

calculated by the forward algorithm:

step 1: initialization: when T is T-G, the compound is,

step 2: recursion: T-G +1, …, T-1,

and step 3: and (4) terminating: when T is equal to T, the reaction solution is mixed,

then there is

Further, according to the calculated abnormal transition probability I_TWhen I is_TWhen the voltage rises suddenly, the content of various gases is abnormally changed in the time period T, which indicates that the system is possibly close to a critical point, and at the moment, a key node which has a large influence on the dynamic network of the transformer is further screened in the time period. Is provided when (I)_T-I_T-1)>Q (only considering abnormally rising gas, i.e. gas exceeding the upper limit), consider that the transformer state has changed abnormally in the period T, without loss of generality, and I is present when the transformer is in the normal state_TLess than or equal to 0.9, and when the transformer is in abnormal state, 0.9<I_TWhen Q is not more than 1, Q is 0.1 as a threshold value for judging abnormal transition of the transformer, and (I) is considered to be_T-I_T-1) When the gas content is more than or equal to 0.1, the abnormal change gas is obviously increased.

The probability calculation results (P values) and screening results for each period are shown in fig. 2 and tables 3 to 5. Wherein FIG. 2(a) is the HMM calculation result of 170 chromatographic data sets, and the analysis of the result shows that the probability I of the first-occurring larger abnormal transition is larger_TIs between the time interval 0 and the time interval 10, therefore the last sampling point of the time interval 10 is taken as an end point, and 50 sampling point data are intercepted forwards; assuming that every 5 continuous sampling points are in a time period and 10 time periods are total, the HMM is used to analyze the gas concentration data in each time period, and the calculation result is shown in fig. 2(b) (assuming that time period 1 is an initial time period and has no abnormal gas change, so the time period is taken as the abnormal gas quantity group C for training the hidden markov model_T-1Starting from time period 2 and calculating an abnormal transition profile starting from time period 3Rate). Table 3 shows the abnormal rising gas amount in each time period screened by the box plot method for the data of the set of 50 sampling points. Table 4 shows the screened abnormal gas in each time period.

Table 3 number of abnormal rising gas at sampling points of each period:

table 4 gas sampling points abnormally rising at each time interval:

as can be seen from the change in the value of P in fig. 2(b), when the group data is shifted from the group of the abnormal gas amounts in the period 4 to the group of the abnormal gas amounts in the period 5, there are: (I)₅－I₄)>0.1; period 5 probability P_TThe lowest probability of 0.088 is reached, and the normal transition probability is close to 0; the probability of an abnormal transition reaching a peak at time period 5: i is₅At 0.912, the probability of an abnormal transition is close to 1, indicating that a significant abnormal change in gas concentration occurs between period 4 and period 6, and the transformer is most likely to approach the system critical transition point. Comparing tables 3 and 4, the gas concentration variation in the time interval 4 is large, and the abnormal rising gas quantity of each sampling point in the time interval 4 is 0,1, 4, 4 and 4 respectively; referring to Table 2, the main gas in which the gas concentration abnormally increased in period 4 was CH₄、C₂H₄、C₂H₆、H₂The reason that the abnormal rise of the characteristic gas is caused by the change of the state of the transformer is considered to be that the abnormal change gas screened in the period 4 represents a key node which has a large influence on the change of the state of the transformer, a network formed by the abnormal change gas is a key network, and DNM analysis is carried out on the screened key network.

Computation of a Key node (CH)₄、C₂H₄、C₂H₆、H₂) Mean standard deviation at each time period of a constituent critical network (DNM)

Average Pearson correlation coefficient among nodes in key network

Judging whether the critical characteristic is met:

1. mean standard deviation of critical networks

Is remarkably increased.

And (4) increasing.

And decreases.

the specific determination of the first critical characteristic is:

if there is

wherein

And

if there is

And is

The calculation result is shown in fig. 3, from which it can be seen that:

as can be seen in FIG. 3(b), the DNM mean standard deviation at time period 6

Increase by

And is

Average Pearson correlation coefficient among nodes in key network

Rise up to

Average Pearson correlation coefficient between nodes in critical network and other non-critical nodes

Then decrease, there is

Critical characteristics are satisfied; as can be seen in figure 3(a), period 6 is l'_TObvious rising is generated, which means that the system state is abnormal, and the abnormal gas in the period 4 forms a sub-network with strong correlation in the period 6, which is a key network for the state transition of the transformer, and the comparison table 1 shows that CH₄、C₂H₄、C₂H₆、H₂These 4 gases are consistent with the major and minor gas components generated by the overheating fault, so it can be concluded that the transformer state is abnormal at time 6, and the reason for the abnormality is oil overheating fault.

According to the maintenance record, all chromatographic data of the main transformer before 8 month and 16 days in 2009 are normal, characteristic gases such as methane and ethylene are obviously increased beginning to be displayed at 8 month and 17 days in 2009, and the oil chromatographic data shows an obvious increasing trend; the on-line monitoring data and the laboratory chromatographic data have basically the same trend. Class A overhaul is carried out on the main transformer at 24 days 10 months 10 in 2009, and the fact that obvious overheating traces exist at the tail ends of a 220kV B-phase high-voltage lead and a sleeve of the main transformer, 2.5 strands of the lead are burnt out, 4 strands of the lead are damaged in total, and the insulating paper and the white cloth tape are locally burnt and carbonized is found. According to the on-line monitoring data analysis of the fault case, the method based on the hidden Markov and dynamic network marker combination model sends out the first early warning signal in the period 6, namely 9 months and 7 days in 2009.

The case analysis result shows that the combination method can send out early warning signals in time when abnormal state data appear, the method firstly establishes a hidden Markov model according to transformer on-line monitoring data, detects the possible critical point of gas change of system abnormal state conversion, further screens key gas molecules capable of reflecting the system conversion to a fault state, takes the key gas molecules as a dynamic network marker of the system, observes the change trend of the dynamic marker, judges whether the transformer approaches the critical conversion point or not according to the change trend, and identifies the possible defects causing the transformer state conversion.

Therefore, the early transformer defect early warning method based on the hidden Markov and dynamic network marker combined model provided by the invention utilizes the dynamic characteristics of the system and the interaction between gases to carry out early warning and identification on the system state, reduces the calculation difficulty of a single early warning model, improves the reliability of the model, and can provide technical scheme support for maintaining the safe operation of the transformer to a certain extent.

The above are merely preferred embodiments of the present invention, and should not be used to limit the scope of the present invention. The technical scope of the present invention is not limited to the contents described in the specification, and the related person can make modifications or additions within the scope not departing from the technical scope of the present invention and defined by the claims through the above description, and all modifications, additions, improvements, equivalents, and the like made to the above embodiments according to the technical spirit of the present invention are within the scope of the present invention.

The above description is only an embodiment of the present invention, but the design concept of the present invention is not limited thereto, and any insubstantial modifications made by using the design concept should fall within the scope of infringing the present invention.

Claims

1. A transformer early defect early warning method based on a combined model is characterized by comprising the following steps:

s1, based on time sequence data of each characteristic gas monitored by an online monitoring device for dissolved gas in transformer oil, taking G sampling points including current time T as a starting point time period T, and reading G time periods forwards, wherein G is more than or equal to 1, each time period comprises G sampling points, and G is more than or equal to 4;

s2, taking the gas data set in the time period T-1 as a comparison group, screening the abnormal gas quantity of the gas data set in the time period T on each sampling point in the time period, and forming an abnormal group O in the time period T_T；

S3, combining all the abnormal groups obtained from the time period T-G to the time period T-1 into a new abnormal gas quantity group C_T-1Taking the training set as a training set to establish a hidden Markov model;

s4, calculating the probability I of abnormal transition of the time period T of the current time T according to the built hidden Markov model_T：I_T＝1-P_T(ii) a Wherein P is_TRepresents from C_T-1Generating an exception group O for a time period T_TThe probability of (d);

step S5, if I_TSuddenly rising, and further screening abnormal change gas in the time period as a key node;

s6, calculating the average standard deviation of the key network consisting of the key nodes in each time period

Average Pearson correlation coefficient among nodes in key network

Judging whether the critical characteristic is met; and obtaining a quantized value I 'of the key network marker in each time period T according to calculation'_T：

Wherein, is a small normal number for avoiding denominator being zero;

step S7, if I'_TAbrupt transition occurs, i.e. transformer from steady state I'_T-1Through critical state I'_TConversion to Defect State I'_T+1And sending out an early warning signal, and identifying the defect type of the transformer according to the formed dynamic network marker.

2. The early-stage defect early-warning method of the transformer based on the combined model as claimed in claim 1, wherein: step S2 specifically includes:

step S21: setting time interval T-1 and time interval T as gas data set N_T-1And N_TDetecting N_TIf the (i, j) th data is abnormal data, forming a new time sequence by the (i, j) th data and the ith row number of the time period T-1, and arranging the data in a sequence from small to large;

s22, calculating an upper quartile Q3 and a lower quartile Q1 of the new time sequence number sequence, and calculating a quartile distance IQR according to Q1 and Q3:

IQR＝Q3-Q1

q1 is a first quartile, Q3 is a third quartile, Q1 and Q3 are numbers which are respectively in the 25 th percent and the 75 th percent after data are arranged from small to large, and the difference value between Q1 and Q3 is called an quartile distance IQR;

and S23, judging whether the detected data is smaller than Q1-1.5 multiplied by IQR or larger than Q3+1.5 multiplied by IQR, and if so, determining that the data is abnormal data.

3. The early-stage defect early-warning method of the transformer based on the combined model as claimed in claim 1, wherein: step S3 specifically includes:

step 31, initialization: when n is equal to 0, give

And

giving an initial value, there is HMM lambda⁰＝(A⁰,B⁰,π⁰)；

Step 32, recursion: when n is 1,2, …,

and

is calculated as follows:

and step 33, terminating: when n ═ H, the recursion terminates, this time with

Then the trained HMM is

H is the updating times; the state of the period T is denoted as s_T(s_T∈{S₀,S₁}，S₀Is a normal state of two hidden states of the transformer, S₁Is a defective state of the two hidden states of the transformer; a denotes a transition probability matrix of hidden states, including the probability from one hidden state i to another hidden state j, denoted as a ═ a_ij}_N×NWherein a is_ij＝P(s_T+1＝S_j|s_T＝S_i) I is more than or equal to 1, j is less than or equal to N, N represents the number of hidden states and has

B denotes an observed state probability matrix, representing the probability distribution of observed states observed under hidden state j of a given HMM over time period T, denoted as B ═ B_jk}_N×MWherein b is_jk＝P(V_k＝O_T|s_T＝S_j) I is more than or equal to 1 and less than or equal to N, k is more than or equal to 1 and less than or equal to M, M represents the number of observable states corresponding to each state, and each observable state is represented by V ═ V ≦ M₁,V₂,…,V_MThe observed state of the time period T is represented as O_T，O_T∈{V₁,V₂,…,V_MAnd is provided with

Pi represents the probability matrix of the initial state, represents the probability matrix of the hidden state in the initial period, and is marked as pi ═ pi_i＝P(s_T-G＝S_i)}_1×N,1≤i≤N。

4. The early-stage defect early-warning method of the transformer based on the combined model as claimed in claim 3, wherein: step S4 specifically includes:

for model HMM λ_T-1＝(A_T-1,B_T-1,π_T-1) And observation sequence O ═ O_T-G,…,O_T-1,O_TAnd if the transformer is still in a normal state during the time period T, calculating the probability of generating the sequence: