CN106596899A - Evaluation method of random error between oil-immersed transformer on-line monitoring data and on-line detection data - Google Patents

Abstract

The invention discloses a method for evaluating random errors between online monitoring data and electrification detection data of an oil-immersed transformer. First, the on-line monitoring data and electrification detection data are respectively obtained, and two sets of time series are obtained after preprocessing; secondly, the two The error sequence is obtained by subtracting the group time series, and multiple eigenmode functions are obtained after decomposing, and the eigenmode functions that meet the noise characteristics form a random error sequence; thirdly, the online monitoring time series is decomposed to obtain multiple eigenmode functions, and the eigenmode functions that meet the noise characteristics are eliminated. The real signal sequence is obtained after the eigenmode function of the noise feature, and the error reference sequence is constructed, and the sequence x _t and x _t ′ are respectively obtained according to the formula x _t =S _t +N _t and x _t ′=S _t +N _t ′, And calculate the corresponding signal-to-noise ratio d _t and d _t '; finally, according to the following formula to get the random error evaluation coefficient ρ; ρ=n _a /n.

Description

Evaluation of Random Error Between Online Monitoring Data and Live Detection Data of Oil-immersed Transformer price method

technical field

The invention relates to the field of signal analysis, in particular to an evaluation method for random errors between on-line monitoring data and electrification detection data of an oil-immersed transformer.

Background technique

There are two types of monitoring methods for oil-immersed transformers. One is to use online monitoring, that is, use infrared spectrometers to remotely monitor equipment online; the other is live detection, that is, professionals go to the site to sample and test transformer oil. Online data has the advantages of short monitoring interval and low cost, but its data noise and random errors are often the focus of power companies. However, at present, the noise and error estimation of transformer online monitoring data are still unresolved technical problems, and the research and development of random error evaluation methods for transformer online monitoring devices is of great value.

The online monitoring data of transformers are nonlinear and non-stationary time series. The analysis and processing of such data is very complicated, and traditional methods are difficult to apply. In the field of economics, HP filtering is often used to eliminate low-frequency long-term trends in non-stationary time series, and then measure high-frequency short-term random fluctuations (Hodrick R J, Prescott EC. Postwar U.S. Business Cycles: AnEmpirical Investigation[J]. Journal of MoneyCredit&Banking, 1997,29(1):1-16.); in the field of signal analysis, power spectral density (i.e. PSD method) is more used to analyze the noise power at different frequencies (John L, Hadley J.Correlated errors in geodetictime series: Implications for time-dependent deformation[M]//Journal of Geophysical Research: Solid Earth(1978–2012).1997:591–603.); these methods have a certain effect on the measurement of random fluctuations in online data, but their The theoretical basis comes from the spectral analysis of time series, which belongs to the frequency domain analysis method, and cannot accurately measure the time domain characteristics of the random fluctuation of the transformer online data.

In reality, we pay more attention to the fluctuation range of online data. If the random fluctuation range is too large, it will often affect the judgment of relevant personnel on the operation of the transformer. Therefore, the time-domain characteristics (local availability) of the random fluctuation of the online data of the transformer main measurement target. In the field of signal analysis, Fourier analysis is undoubtedly an important analysis method for signal characteristics, but its global characteristics make it relatively limited in dealing with non-stationary signals and describing the time-domain characteristics of signals. Therefore, on the basis of Fourier analysis, relevant scholars have proposed analysis methods for non-stationary signals, including short-time Fourier transform, Wigner-Ville distribution, wavelet transform, etc. The basic principle of the short-time Fourier transform is to add a window function to the signal sequence, convert the non-stationary signal into multiple stationary signals with short intervals, and then perform Fourier transform on the signal in the window, so as to obtain the time-varying spectrum of the original signal. The purpose of analyzing high-frequency signals (random noise) in different periods (Jurado F, Saenz J R. Comparison between discrete STFT and wavelets for the analysis of power quality events [J]. Electric Power Systems Research, 2002, 62 (3) :183-190.); Wigner-Ville distribution is to perform Fourier transform on the covariance function of the original signal, which solves the limitation of the window function on the short-term Fourier transform to a certain extent. At the same time, the Wigner-Ville distribution has many useful properties, including time-shift, symmetry, etc., which can better describe the time-varying characteristics of the signal (Martin W, FlandrinP. Wigner-Ville spectral analysis of nonstationary processes[J].IEEETransactions on Acoustics Speech&Signal Processing,1985,33(6):1461-1470.); wavelet transform proposed by Donoho is a multi-scale signal decomposition method, and the noise estimation process can be summarized as: (1) By setting the wavelet base The function performs wavelet decomposition on the original signal. (2) Use the median absolute variation of the first layer wavelet decomposition to estimate the noise (Johnstone I M, Silverman B W. Wavelet threshold estimators for data with correlated noise[J]. Journal of the royal statistical society: series B(statistical methodology) , 1997,59(2):319-351.) (Donoho D L. De-noising by soft-thresholding[J]. IEEE Transactions on Information Theory, 1995,41(3):613-627.). Compared with the above methods, wavelet transform has better time-frequency characteristics and noise metrics. The noise estimation method based on wavelet transform proposed by Donoho is currently the most widely used noise estimation method for nonlinear and non-stationary signals.

The above methods improve the traditional Fourier analysis and provide a scientific and systematic estimation method for nonlinear and non-stationary signal noise. However, these methods still have many problems in the case of random error estimation of transformer online data. The first is noise extraction. The above methods cannot achieve adaptive extraction of random noise, and the noise measurement effect depends on the selection of specific parameters. For example, the selection of wavelet basis functions has a great impact on its noise reduction ability, but in practical applications The determination of basis functions is a difficult problem. The second is noise estimation. The existing methods themselves have a certain time-varying nature, but the noise evaluation indicators start with more statistical properties, such as noise variance, median variation, Allan variance, TheoH variance, etc. (Allan variance and TheoH variance) Variance is an estimation method for specific noise, such as Gaussian white noise, flicker noise, random walk noise, etc. (Allan D W, Barnes J A.A modified "Allan variance" with increased doscillator characterization ability[C]//Thirty Fifth Annual Frequency ControlSymposium .1981.IEEE,1981:470-475.)(McGee J A,Howe D A.TheoH and Allandeviation as power-law noise estimators[J].ieee transactions on ultrasonics,ferroelectrics,and frequency control,2007,54(2) :448-452.)). These evaluation indicators are global and cannot reflect the locally available characteristics of online data. At the same time, these methods are mainly used for random error measurement of precision instruments, and are very sensitive to "gross errors (huge errors)", while the situation of random errors in transformer on-line monitoring devices is more complicated, and more universal evaluation indicators are needed.

To sum up, the existing methods are very limited in the case of random error estimation of transformer online data.

Contents of the invention

The invention provides an evaluation method for random errors between online monitoring data of oil-immersed transformers and live detection data, which can accurately measure the time-domain characteristics of random fluctuations in online monitoring data of transformers, and obtain evaluation coefficients of random errors after calculation ρ can accurately describe the measurement accuracy of the online monitoring device.

A method for evaluating random errors between on-line monitoring data and live detection data of an oil-immersed transformer, characterized in that it includes the following steps:

(1) The content of characteristic gas in oil-immersed transformer oil is obtained through remote online monitoring equipment, which is recorded as online monitoring data; the content of characteristic gas in oil-immersed transformer oil is obtained through manual sampling, which is recorded as live detection data; Preprocessing the online monitoring data and live detection data to obtain the online monitoring characteristic gas content-time series and the live detection characteristic gas content-time series with the same time interval and corresponding time points;

(2), subtract the online monitoring characteristic gas content-time series from the live detection characteristic gas content-time series to obtain the error sequence, and then use the ensemble empirical mode decomposition method to decompose the error sequence to obtain multiple eigenmode functions, which conform to The eigenmode function of the noise feature constitutes a random error sequence, denoted as N _t ;

(3) Use the ensemble empirical mode decomposition method to decompose the online monitoring characteristic gas content-time series to obtain multiple eigenmode functions, and get the real signal sequence S _t after removing the eigenmode functions that meet the noise characteristics. According to the real signal sequence S _t and the accuracy of the on-line monitoring device requires the construction of an error reference sequence N _t ′; and then according to the formulas x _t ＝S _t +N _t and x _t ′=S _t +N _t ′, the sequences x _t and x _t ′ are respectively obtained. After calculation Obtain the signal-to-noise ratios d and d' corresponding to the sequences x _t and x _t 'respectively;

(4) Considering the time-domain characteristics of the sequence x _t , use the rectangular window function to estimate the signal-to-noise ratio at each time point of x _t and x _t ′, respectively, to obtain d _t and d _t ′, and then according to the following formula (Ⅰ ) to obtain the evaluation coefficient ρ of the random error, so as to evaluate the measurement accuracy of the online monitoring device;

ρ=n _a /n (I);

In the formula, n is the total number of time points in the online monitoring characteristic gas content-time series, n _a is the number of time points where the random error is less than the actual demand value.

The present invention proposes a more practical online data random error evaluation scheme. For the noise extraction problem, we chose the empirical mode decomposition and its improved method to replace the existing wavelet transform. The reasons are: (1) The empirical mode decomposition is an adaptive noise extraction method, and no parameter setting is required before signal processing. Especially suitable for feature extraction of non-stationary time series (Wu Z, Huang N E.Ensemble empirical mode decomposition: anoise-assisted data analysis method[J].Advances in adaptive data analysis,2009,1(01):1-41 .) (Wu Z, Huang N E, Long S R, et al. On the trend, detrending, and variability of nonlinear and nonstationary time series [J]. Proceedings of the National Academy of Sciences, 2007, 104(38): 14889-14894. ). (2) A large number of empirical studies have shown that the empirical mode decomposition method is better than wavelet transform in terms of signal decomposition (Peng Z K, Peter W T, Chu F L. A comparison study of improved Hilbert–Huang transform and wavelet transform: application to fault diagnosis for rolling bearing[J].Mechanical systems and signal processing,2005,19(5):974-988.)(Labate D,La Foresta F,Occhiuto G,etal.Empirical mode decomposition vs.wavelet decomposition for the extraction of respiratory signal from single -channel ECG: A comparison[J].IEEE SensorsJournal,2013,13(7):2666-2674.)(Dai W J,Ding X L,Zhu J J,et al.EMD filtermethod and its application in GPS multipath[J]. Acta Geodaetica et Cartographica Sinica, 2006.). Therefore, compared with the existing methods, the empirical mode decomposition can help us extract the random error of the transformer online data more accurately; for the noise estimation problem, we choose the signal-to-noise ratio as the noise measurement index. On the one hand, the signal-to-noise ratio is more universal and applicable to any form of noise measurement. On the other hand, the signal-to-noise ratio has real physical meaning and can improve the interpretability of the model compared with statistical indicators. At the same time, in order to make the index have time-domain characteristics, we added a sliding time window to the extracted random error sequence, and took the signal-to-noise ratio of the signal in the window as the evaluation index of random noise at this moment, which reflected the partial availability of online data to the greatest extent. specialty.

In step (1), described characteristic gas comprises hydrogen, total hydrocarbon, carbon monoxide or carbon dioxide, and described total hydrocarbon can be hydrocarbons such as methane, ethane, ethylene, acetylene;

The content of the characteristic gas is the mass concentration or volume concentration of the characteristic gas.

The random error of the transformer online monitoring device is mainly used to describe the measurement accuracy of the online monitoring device. The stronger the fluctuation of random error, the lower the measurement accuracy and stability of the device. The State Grid has clear requirements for the stability of the online monitoring device, that is, the fluctuation of the online monitoring device should not exceed 10% of the measured value, and the monitoring device with poor measurement accuracy should be returned to the factory for correction.

In the process of random error level estimation, live detection data of oil-immersed transformers has important reference value. Therefore, we also use live detection data as a reference sequence, and perform data preprocessing on online monitoring data and live detection data, so that It becomes a time series with the same time interval and time points corresponding to each other.

In actual operation, the preprocessing is based on one day as the time interval, and the average number of multiple data in one day is taken, and linear interpolation is used to replace the data without data.

After a complete data preprocessing, we make a difference between the online monitoring data and the live detection data to obtain the error sequence of the online monitoring data, and decompose the error sequence into an ensemble empirical model. The ensemble empirical mode decomposition method is a signal processing method, and its main function is to decompose the original signal sequence into multiple eigenmode functions, and each eigenmode function has the characteristics of a stationary time series. The research shows that the eigenmode function of noise after empirical mode decomposition has certain statistical properties. According to the method of related literature, we calculated each eigenmode function decomposed from the online monitoring data, and took the first decomposed eigenmode function with a significantly 0 mean value as the noise sequence, thus obtaining the online monitoring The random error sequence N _t of the data.

The specific method of decomposing the error sequence by using the ensemble empirical mode decomposition method is as follows:

(a) Find all the maximum points and all the minimum points in the error sequence N _t (here denoted as x(t)), and connect all the maximum points with a curve to obtain the upper envelope e _max (t), and then get the lower envelope e _min (t) from all the minimum points;

(b) Calculate the average value m(t) of the upper envelope e _max (t) and the lower envelope e _min (t), and calculate the difference d(t) between the signal N _t and m(t);

(c) Check whether d(t) is an eigenmode function, the test standard is that the number of extreme points and zero-crossing points in the entire time range differs by at most one and the mean value of the upper and lower envelopes is zero: if d(t) is a The intrinsic modulus function, record imf(t)=d(t) as the intrinsic modulus function output of this process, let the residual r(t)=x(t)-d(t), the original sequence x(t) =r(t); if d(t) is not an eigenmode function, then let x(t)=d(t); repeat the above steps until the eigenmode function can no longer be decomposed; thus express the original signal as a cost eigen The sum of the modulo function and the residual term:

Among them, imf(t) represents the intrinsic modulus function item, i is the number of intrinsic modulus function items, N is the total number of intrinsic modulus function items, and r(t) represents the residual term.

At the same time, in order to conduct a better comparison and analysis, we use the same method to decompose the set empirical mode on the preprocessed online monitoring data, remove the eigenmode functions that conform to the statistical characteristics of noise, and obtain the real signal of the online monitoring device sequence S _t .

In order to know whether the online monitoring device meets the requirements, we also construct an error reference sequence N _t ′, record the accuracy of the online monitoring device as ε, construct the error reference sequence N _t ′, N _t ′=S _t ×m _t , where m _t is a randomly assigned white noise sequence with amplitudes between -ε and ε. Therefore, we only need to compare the fluctuations of N _t and N _t ′ to know whether the online monitoring device meets the requirements.

During the comparative analysis, we quantified the fluctuation levels of N _t and N _t ′ using the signal-to-noise ratio. If the known signal sequence x(t) is composed of noise sequence n(t) and signal sequence s(t), then the signal-to-noise ratio (signal-to-noise ratio, unit: dB) in x(t) is defined as follows:

In the formula, P _s is the signal power, P _n is the noise power, and their calculation formulas are:

In the formula, t represents the time point of the time series, and T represents the length of the signal sequence.

We substitute x _t =S _t +N _t and x _t ′=S _t +N _t ′ into the above formula to obtain the signal-to-noise ratio d and d'. If d>d', it means that the random error fluctuation of the online monitoring device is small, which can be use.

In addition, in order to investigate the local availability of online monitoring devices, we introduce the concept of rectangular window function to estimate the signal-to-noise ratio at each time point in the time domain. After adding the window function, we can get the distribution of the signal-to-noise ratio of the error sequence in the time domain. Let d _{t be the signal-to-noise ratio of the random error sequence of the online monitoring data, d t ′ the signal-to-noise ratio of the artificial reference noise sequence, and the number of time points when d t >d t ′ in the} time domain is n _a , and then substitute The evaluation coefficient ρ of random error can be obtained from formula (I). The calculation methods of d _t and d _t ′ are as follows:

In the formula, the calculation of P _st , P _nt , P _nt ′ is obtained from the rectangular window function, and the length of the rectangular window function is recorded as w, then the calculation formulas of P _st , P _nt , P _nt ′ are respectively:

In the formula, t represents the time point of the time series, and T represents the length of the signal sequence.

The value range of the evaluation coefficient ρ is [0,1]. The larger the value, the higher the availability of the online device in the time domain; when the evaluation coefficient ρ=1, it means that the online monitoring device is more efficient in the entire evaluation time range. The fluctuations within are in line with actual needs, and the availability is very high. When the result is 0, it means that the fluctuation of the online monitoring device in the entire evaluation time range does not meet the actual needs and should be corrected.

Compared with the prior art, the present invention has the following advantages:

(1) In terms of noise decomposition, the present invention is self-adaptive and can automatically decompose the online monitoring characteristic gas content-time series into noise series and signal series.

(2) The present invention considers the time-domain characteristics of the error sequence, has dynamic characteristics, and can estimate the local availability of the on-line monitoring characteristic gas content-time series.

Description of drawings

Figure 1 shows the characteristic gas content-time curve of online monitoring of an oil-immersed transformer infrared spectrometer after pretreatment and the characteristic gas content-time curve of live detection;

Fig. 2 is the random error sequence N _t decomposed on the basis of two groups of characteristic gas content-time series in the embodiment;

Fig. 3 is the real signal sequence S _t of the on-line monitoring data decomposed on the basis of the on-line monitoring characteristic gas content-time series in the embodiment;

Figure 4 shows the random error sequence (left picture) and real signal sequence (right picture) of online monitoring data decomposed by four different wavelet basis functions;

Fig. 5 is a comparison diagram of signal-to-noise ratio in the time domain calculated by using the random error sequence (upper left) and the artificial noise reference sequence (upper right).

detailed description

The present invention will be further described below in conjunction with specific examples, but the present invention is not limited to the following examples.

Example:

The random error of the transformer online monitoring device is mainly used to describe the measurement accuracy of the online monitoring device. The stronger the fluctuation of random error, the lower the measurement accuracy and stability of the device. The State Grid has clear requirements for the stability of the online monitoring device, that is, the fluctuation of the online monitoring device should not exceed 10% of the measured value, and the monitoring device with poor measurement accuracy should be returned to the factory for correction.

In the process of random error level estimation, live detection data of oil-immersed transformers has important reference value. Therefore, we also use live detection data as a reference sequence, and perform data preprocessing on online monitoring data and live detection data, so that It becomes a time series with the same time interval and time points corresponding to each other. In actual operation, the time interval is usually taken as one day. For the problem of missing online data or live data at each time point, we adopt a linear interpolation method to fill in the data so that the time span of the two is the same, and the time points are one by one. Correspondingly, the characteristic gas content-time curve of on-line monitoring after pretreatment and the characteristic gas content-time curve of charging detection are shown in Fig. 1 .

After a complete data preprocessing, we make a difference between the online monitoring data and the live detection data to obtain the error sequence of the online monitoring data, and decompose the error sequence into an ensemble empirical model. The ensemble empirical mode decomposition method is a signal processing method, and its main function is to decompose the original signal sequence into multiple eigenmode functions, and each eigenmode function has the characteristics of a stationary time series. The research shows that the eigenmode function of noise after empirical mode decomposition has certain statistical properties. We calculated each eigenmode function decomposed from the online monitoring data, and took the first decomposed eigenmode function with a significantly 0 mean value as the noise sequence, thus obtaining the random error sequence N of the online monitoring data _t , as shown in Figure 2.

At the same time, in order to conduct a better comparison and analysis, we use the same method to decompose the set empirical mode on the preprocessed online monitoring data, remove the eigenmode functions that conform to the statistical characteristics of noise, and obtain the real signal of the online monitoring device Sequence S _t , the corresponding time-domain distribution diagram is shown in Fig. 3 .

Because the State Grid has clear requirements on the stability of the online monitoring device, that is, the fluctuation of the online monitoring device should not exceed 10% of the measured value. In this embodiment, the accuracy requirement is 10% as an example. According to the real signal of the online monitoring data, An artificial noise reference sequence N _t ′ was constructed by using the white noise sequence m _t with an amplitude between -0.1 and 0.1, and a comparative analysis was made between the actual random error sequence N _t and the artificial noise reference sequence N _t ′, and the rectangular The window function calculates the distribution of the signal-to-noise ratio in the time domain. The window length of the rectangular window function is set to one month. The result is shown in Figure 5. The part circled in red in Figure 5 represents that the fluctuation of the random error of the online data is greater than that of the noise reference sequence, and these situations are well reflected in the SNR comparison graph.

In this example, in order to illustrate the advantages of the ensemble empirical decomposition method in noise extraction, we follow the same steps and use wavelet transform and four different wavelet basis functions to separate and extract the noise and real signals of the online monitoring data of the transformer , respectively get the random error sequence (left picture) and the real signal sequence (right picture) of four groups of online monitoring data, as shown in Figure 4, it can be found that different wavelet basis functions have an important impact on noise extraction and signal decomposition, wavelet The difficulty of determining the basis function is the key to the practical application of this method.

Claims (9)

1. between a kind of oil-filled transformer online monitoring data and live detection data random error evaluation methodology, its feature exists In comprising the steps：

(1) content of characteristic gas in oil-filled transformer oil, is obtained by remote online monitoring equipment, is designated as monitoring number on-line According to；The content of characteristic gas in oil-filled transformer oil is obtained by manual sampling, is designated as in live detection data；Respectively to Line Monitoring Data and live detection data carry out pretreatment, obtain that time interval is identical and time point mutually corresponding on-line monitoring Characteristic gas content-time serieses and live detection characteristic gas content-time serieses；

(2), on-line monitoring characteristic gas content-time serieses and live detection characteristic gas content-time serieses are subtracted each other, is obtained To error sequence, recycle set ensemble empirical mode decomposition method that decomposition is carried out to error sequence and obtain multiple intrinsic mode functions, accord with The intrinsic mode functions composition random error series of noise characteristic are closed, N is designated as_t；

(3) decomposition is carried out to on-line monitoring characteristic gas content-time serieses using set ensemble empirical mode decomposition method and obtains many Individual intrinsic mode functions, to reject and obtain authentic signal sequence S after the intrinsic mode functions for meeting noise characteristic_t, according to actual signal sequence Row S_tAnd the required precision instrument error reference sequences N of on-Line Monitor Device_t′；Again according to formula x_t=S_t+N_tWith x_t'=S_t+N_t′ Respectively obtain sequence x_tAnd x_t', it is computed respectively obtaining sequence x_tAnd x_t' corresponding signal to noise ratio d and d '；

(4) sequence x is considered_tTemporal signatures, using rectangular window function respectively to x_tWith x_tSignal to noise ratio on ' each time point is entered Row estimation, obtains d_tWith d_t', the evaluation coefficient ρ of random error is obtained further according to lower formula I, with this survey to on-Line Monitor Device Accuracy of measurement is evaluated；

ρ=n_a/n (Ⅰ)；

In formula, n is time point sum in on-line monitoring characteristic gas content-time serieses, n_aActual demand is less than for random error The number of the time point of value.

2. random error is commented between oil-filled transformer online monitoring data according to claim 1 and live detection data Valency method, it is characterised in that in step (1), described characteristic gas include hydrogen, methane, ethane, ethylene, acetylene, an oxidation Carbon or carbon dioxide；

The content of the characteristic gas is characterized the mass concentration or volumetric concentration of gas.

3. random error is commented between oil-filled transformer online monitoring data according to claim 1 and live detection data Valency method, it is characterised in that in step (1), described pretreatment is：

With one day as time interval, there is taking the mean for multiple data in one day, being replaced with linear interpolation without data.

4. random error is commented between oil-filled transformer online monitoring data according to claim 1 and live detection data Valency method, it is characterised in that in step (2), using set ensemble empirical mode decomposition method error sequence is decomposed it is concrete Method is as follows：

(a) error identifying sequence N_tAll maximum points and all minimum points in (being designated as x (t) herein), and will be all very big Value o'clock is coupled together with a curve and obtains coenvelope line e_max(t), then obtain lower envelope line e by all minimum points_min(t)；

B () calculates coenvelope line e_max(t) and lower envelope line e_minMeansigma methodss m (t) of (t), and signal calculated N_tWith the difference of m (t) d(t)；

C whether () inspection d (t) be intrinsic mode functions, and touchstone is the number of extreme point and zero crossing in whole time range The average of at most difference one and upper and lower envelope is zero：If d (t) is an intrinsic mode functions, note imf (t)=d (t) is The intrinsic mode functions output of this process, makes residual error r (t)=x (t)-d (t), original series x (t)=r (t)；If d (t) is not It is intrinsic mode functions, then makes x (t)=d (t)；Repetition above step till it can not decomposite intrinsic mode functions again；So as to incite somebody to action Primary signal is expressed as intrinsic mode functions and residual error item sum：

$x\left(t\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{imf}}_i\left(t\right)+r\left(t\right);$

Wherein imf (t) represents intrinsic mode functions item, and i is the item number of intrinsic mode functions, N for intrinsic mode functions total item, r (t) Represent residual error item.

5. random error is commented between oil-filled transformer online monitoring data according to claim 1 and live detection data Valency method, it is characterised in that in step (2), it is significantly 0 that the specific features for meeting noise are the averages of intrinsic mode functions.

6. random error is commented between oil-filled transformer online monitoring data according to claim 1 and live detection data Valency method, it is characterised in that in step (3), error reference sequences N_t' building method it is as follows：

The precision of on-Line Monitor Device is designated as into ε, instrument error reference sequences N_t', N_t'=S_t×m_t, wherein, m_tIt is amplitude in-ε To the white noise sequence being randomly assigned between ε.

7. random error is commented between oil-filled transformer online monitoring data according to claim 1 and live detection data Valency method, it is characterised in that in step (3), the computational methods such as following formula (II -1) of signal to noise ratio d：

$d=10\log \left(\frac{P_s}{P_n}\right)---(II-1);$

In formula, P_sFor signal power, P_nFor noise power, its computing formula is respectively：

$P_s=\frac{1}{T}{\mathrm{\Σ}}_{t=1}^T{S_t}^2,P_n=\frac{1}{T}{\mathrm{\Σ}}_{t=1}^T{N_t}^2;$

In formula, t represents seasonal effect in time series time point, the length of T representation signal sequences；

The computational methods of signal to noise ratio d ' such as following formula (II -2)：

$d^{\′}=10\log \left(\frac{P_s}{{P_n}^{\′}}\right)---(II-2);$

In formula, P_n' also it is noise power, its computing formula is：

${P_n}^{\′}=\frac{1}{T}{\mathrm{\Σ}}_{t=1}^T{N_t}^{\′2}.$

8. random error is commented between oil-filled transformer online monitoring data according to claim 1 and live detection data Valency method, it is characterised in that in step (4), d_tWith d_t' computational methods it is as follows：

$d_t=10log\left(\frac{P_{st}}{P_{nt}}\right)---(II-3);$

${d_t}^{\′}=10log\left(\frac{P_{st}}{{P_{nt}}^{\′}}\right)---(II-4);$

In formula, P_st, P_nt, P_nt' calculate by rectangular window function gained, the length of rectangular window function is designated as into w, then P_st, P_nt, P_nt′ Computing formula be respectively：

$P_{st}=\frac{1}{w}{\mathrm{\Σ}}_t^{t+w}{S_t}^2,t=1,2...T-w$

$P_{nt}=\frac{1}{w}{\mathrm{\Σ}}_t^{t+w}{N_t}^2,t=1,2...T-w$

${P_{nt}}^{\′}=\frac{1}{w}{\mathrm{\Σ}}_t^{t+w}{N_t}^{\′2},t=1,2...T-w$

In formula, t represents seasonal effect in time series time point, the length of T representation signal sequences.

9. random error is commented between oil-filled transformer online monitoring data according to claim 8 and live detection data Valency method, it is characterised in that meet d_t>d_t' time point think random error less than actual demand value time point, number It is designated as n_a。