CN105550943A - Method for identifying abnormity of state parameters of wind turbine generator based on fuzzy comprehensive evaluation - Google Patents

Abstract

The invention relates to a method for identifying abnormity of state parameters of a wind turbine generator based on fuzzy comprehensive evaluation. The method comprises the following steps: S1, obtaining the mean absolute error of a test sample according to a selection result of a state parameter generalized fuzzy abnormity identification model so as to obtain weights of various prediction models; S2, realizing state parameter prediction in a time interval to be analyzed through the various prediction models; S3, realizing condition analysis of state parameters through the residual error of the various predication models so as to obtain residual error abnormity indexes of various models; S4, calculating fuzzy membership degrees of various indexes to form a fuzzy evaluation matrix, and calculating an output layer evaluation value; and S5, inputting an evaluation result according to the membership degree maximum principle, and taking a corresponding comment as the evaluation result. The method disclosed by the invention is based on SCADA data of a wind power plant; the method is easy in programming realization; abnormity of parameters can be reflected accurately and effectively; and the identification accuracy of abnormity of the state parameters can be increased by comprehensively considering the abnormity identification results of the plurality of prediction models.

Description

The abnormal discrimination method of a kind of Wind turbines state parameter based on fuzzy comprehensive evoluation

Technical field

The invention belongs to new forms of energy power equipment safety assessment technology field, relate to the abnormal discrimination method of a kind of Wind turbines state parameter based on fuzzy comprehensive evoluation.

Background technology

In wind energy turbine set SCADA (SupervisoryControlAndDataAcquisition) system, the status monitoring parameter of unit is not only the embodiment of equipment working condition, also contains the relevant information of unit health status simultaneously.Based on the set state abnormal parameters operating mode's switch of SCADA system, it is the important channel obtaining Wind turbines reliability information.But the state parameter in SCADA system is subject to the impact of wind speed, environment temperature, the abnormal information of unit is easily covered.

The main method of at present carrying out anomalous identification to set state parameter is the forecast model of the parameter setting up Wind turbines state, and according to predicting that the residual error produced carries out identification to set state.The validity of these class methods and accuracy mainly based on: 1) forecast model can realize the prediction of state parameter accurately.2) data sample of forecast model all derives from unit when normally running.Existing research mainly reaches the object of precision of prediction raising by the input parameter of optimal prediction model, but considers less on the impact of data sample.In practical application, the selection of model sample data all has larger impact to the prediction of state parameter and abnormal identification.Based on the forecast model of historical data as training sample, because ageing equipment and maintenance can have an impact to model accuracy, based on the forecast model of Recent data as training sample, due to larger impact can be had on the sensitive of abnormal identification containing data during misoperation.Therefore, in the uncertain situation of sample validity, how to improve the accuracy of forecast model to the abnormal identification of state parameter, be the problem needing solution at present badly.

Summary of the invention

In view of this, the object of the present invention is to provide the abnormal discrimination method of a kind of Wind turbines state parameter based on fuzzy comprehensive evoluation, the method, according to the incidence relation of the incidence relation of set state parameter and unit operating characteristic, set state parameter and physical environment, sets up the forecast model of state parameter.State parameter forecast model is set up, comparative analysis precision of prediction respectively with historical data sample, Recent data sample, similar unit data sample.And based on this, the system of selection of forecast model and the quantitative analysis method of prediction residual exception of Wind turbines state parameter are proposed, construct the abnormal identifying index system of state parameter, adopt the method for fuzzy evaluation to combine the abnormal identification result of multiple forecast model.The method can reflect the exception of parameter accurately and effectively; Consider the accuracy that the abnormal identification result of multiple forecast model can improve the abnormal identification of state parameter.

For achieving the above object, the invention provides following technical scheme:

The abnormal discrimination method of Wind turbines state parameter based on fuzzy comprehensive evoluation, comprises the following steps:

S1: the mean absolute error obtaining test sample book according to the selection result of the abnormal identification model of state parameter generalized fuzzy, and obtain each forecast model weights omega by formula (6) _i:

S2: realize predicting the state parameter of period to be analyzed by each forecast model; The prediction of model adopts the mode of window rolling, and the state parameter abnormal identification time interval is set as 6 hours;

S3: realize the status analysis to state parameter by the residual error of each forecast model; The data selection of residual error characteristic statistics analysis only considers the state parameter point be in sample span, and all the other future positions are as invalid prediction point; Choose the cumulative data of 24 hours effective residual errors, quantitative analysis each model state abnormal parameters degree, obtain each model residual error abnormal index EAI _i;

S4: the fuzzy membership calculating each index according to formula (3)-formula (5), forms fuzzy matrix for assessment, adopts formula (7) to calculate output layer assessed value;

S5: the maximum principle of input foundation degree of membership of assessment result, namely with b _max=max (b _i) corresponding comment l _ias evaluation result.

Further, in step sl, the framework of the abnormal identification model of state parameter generalized fuzzy is as follows:

Corresponding dissimilar data sample, set up state parameter forecast model corresponding with it respectively, specific implementation process is divided into following 4 steps:

1) Monitoring Data of certain period in current SCADA system is gathered, and obtain the time series of the state parameter time series needing to carry out analyzing and the forecast model input parameter corresponded;

2) corresponding each forecast model obtains the predicted value of state parameter under this data sample, and obtains the time series of state parameter prediction residual;

3) calculate the abnormal index of each residual sequence according to the abnormal quantization method of residual sequence, characterize the state parameter intensity of anomaly under each forecast model;

4) with the abnormal index of all residual sequences structure index system, fuzzy comprehensive evaluation method is adopted to carry out abnormal identification to state parameter.

Further, in step s 2, the Forecasting Methodology of described state parameter specifically comprises:

The forecast model of set state ginseng is for reflecting the incidence relation in SCADA system between each state parameter, choose the time period of Water demand, with forecast model, state parameter is predicted, if set state parameter is in normal condition, then this residual error feature predicted the outcome conforms to the residual error feature of data sample; If set state parameter is in abnormal conditions, then the residual error feature of this residual error predicted the outcome and data sample has significant difference; The forecast model construction method of state parameter comprises the following steps: the input parameter selecting forecast model; Extract the training sample of forecast model and the test sample book of residual error statistical characteristic analysis; The type of analyses and prediction model, structure and training method, adopt training sample to train;

Specific as follows:

1) selection of input parameter

In forecast model, the selection of input parameter is divided into two types, and the 1st class state parameter and wind speed and environment temperature correlativity are comparatively large, the prediction of this type of state parameter using wind speed, environment temperature and upper a period of time temperature as the input parameter of model; 2nd class state parameter and wind speed correlativity are comparatively large, and the prediction of this type of state parameter is using wind speed, propeller pitch angle and crab angle error as input parameter;

2) extraction of sample data

The training sample of forecast model takes three kinds of extracting modes: the 1st kind for the recent sample of the machine, data sample derives from this TV station unit self, and data sample time section to be taken in recent one month and not containing time period data to be analyzed; 2nd kind is the machine historical sample, and data sample derives from the machine, and data sample time section is taken at same month in last year; Be the recent sample of similar unit in 3rd, data sample derives from wind field other units of the same type, and data sample time section is recent one month data and comprises time period data to be analyzed;

In addition, the extraction of all sample datas does not all comprise forced outage event, if this operation period contains forced outage event, should remove the data of stopping transport in first 2 days; For ensureing that residual distribution modeling contains sufficient statistics, the ratio arranging training sample and test sample book is 6:4;

3) data prediction

First data sample carries out pre-service to data before training, and to ensure validity and the accuracy of data, mainly comprises: reject invalid data, average value processing and normalized; In addition, change parameter faster for state parameter, adopt 1 minute average treatment, as rotating speed and power etc.; For the parameter that state parameter change is slower, adopt 10 minutes average treatment, as temperature mode parameter;

4) training of forecast model

Reverse transmittance nerve network (BPNN) is utilized to set up the forecast model of state parameter; Its structure of BPNN model consists of: 1 input layer, 1 hidden layer and 1 output layer; Wherein, the number of input state parameter equals the nodes of mode input layer, and hidden layer neuron number is determined by 20 chi independent experiment training networks, and output layer nodes is 1, and hidden layer transport function is tansig type, and output layer transport function is logsig type.

Further, following methods is adopted to carry out the foundation of the abnormal identification model of state parameter generalized fuzzy:

1) intensity of anomaly of prediction residual quantizes

If set state parameter is normal, then in predicted time section, residual distribution feature conforms to the residual distribution of training sample; With this, set up the abnormal index based on residual distribution model: if prediction residual is in the larger span of probability density value in residual distribution characteristic, then this parameter is normal; If prediction residual is in low probability density interval in residual distribution characteristic, then this parameter is abnormal; According to above-mentioned principle to divide residual error span: with the residual values of fractile 0.025,0.25,0.75,0.975 correspondence, residual error span is divided into 3 intervals;

Definition residual error abnormal index (ErrorAbnormalIndex, EAI) characterizes the intensity of anomaly of residual error, and computing formula is as follows:

$EAI=1-\frac{N_1*C_1}{\underset{i=1}{\overset{3}{\Σ}}{N}_i*C_i}---\left(1\right)$

In formula, N _ifor value in residual sequence is in the number in interval i, C _ifor value is in the penalty factor in interval i, be set to [1,3,5]; EAI is larger, shows that residual error intensity of anomaly is higher;

2) the fuzzy abnormal identification of state parameter

Adopt fuzzy comprehensive evaluation method to carry out the abnormal identification of parameter, the judgment index system of the method is made up of output layer and indicator layer: indicator layer is the residual error abnormal index EAI that forecast model obtains _i, output layer is abnormal evaluation result; In addition, be divided into " normally ", " attention " and "abnormal" three kinds of situations to the exception level of state parameter, namely comment is: L=[normal, note, abnormal]=[l ₁, l ₂, l ₃]; Residual error abnormal index EAI _ito the state l in Comment gathers _idegree of membership be v _ij, then available membership grade sets V _i=[v _i1, v _i2, v _i3] represent with residual error abnormal index EAI _icarry out the result assessed; The fuzzy matrix for assessment of output layer is:

$V=\left[\begin{array}{c}{V}_1\\ V_2\\ M\\ V_n\end{array}\right]=\left[\begin{array}{ccc}{v}_{11}& v_{12}& v_{13}\\ v_{21}& v_{22}& v_{23}\\ M& M& M\\ v_{n1}& v_{n2}& v_{n3}\end{array}\right]---\left(2\right)$

Adopt the distribution membership function of triangle and half trapezoidal combination, the concrete defining method of degree of membership is:

$v_1\left({\mathrm{EAI}}_i\right)=\left\{\begin{array}{cc}1& {\mathrm{EAI}}_i<0.4\\ 3.67-6.67{\mathrm{EAI}}_i& 0.4\&le;{\mathrm{EAI}}_i<0.55\\ 0& {\mathrm{EAI}}_i\&GreaterEqual;0.55\end{array}\right.---\left(3\right)$

$v_2\left({\mathrm{EAI}}_i\right)=\{\begin{array}{cc}0& {\mathrm{EAI}}_i<0.4or{\mathrm{EAI}}_i\&GreaterEqual;0.7\\ 6.67{\mathrm{EAI}}_i-2.67& 0.4\&le;{\mathrm{EAI}}_i<0.55\\ 4.67-6.67{\mathrm{EAI}}_i& 0.55\&le;{\mathrm{EAI}}_i<0.7\end{array}---\left(4\right)$

$v_3\left({\mathrm{EAI}}_i\right)=\left\{\begin{array}{cc}0& {\mathrm{EAI}}_i<0.55\\ 6.67{\mathrm{EAI}}_i-3.67& 0.55\&le;{\mathrm{EAI}}_i<0.7\\ 1& {\mathrm{EAI}}_i\&GreaterEqual;0.7\end{array}\right.---\left(5\right)$

Set the weight of each index with the prediction accuracy of each forecast model to test sample book, computing method are as follows:

${\mathrm{\&omega;}}_i=(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})/\underset{i=1}{\overset{N}{\&Sigma;}}(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})=(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})/(N-1)---\left(6\right)$

In formula, MAE _ifor the mean absolute error value that forecast model i predicts test sample book;

Calculate the fuzzy membership of each index according to formula (3)-formula (5), form fuzzy matrix for assessment, adopt formula (7) to calculate output layer assessed value.

$B=\left[\begin{array}{ccc}{b}_1& b_2& b_3\end{array}\right]\left[\begin{array}{cccc}{\mathrm{\&omega;}}_1& {\mathrm{\&omega;}}_2& L& {\mathrm{\&omega;}}_n\end{array}\right]\left[\begin{array}{ccc}{v}_{11}& v_{12}& v_{13}\\ v_{21}& v_{22}& v_{23}\\ M& M& M\\ v_{n1}& v_{n2}& v_{n3}\end{array}\right]---\left(7\right).$

Beneficial effect of the present invention is: the method that the present invention adopts, based on wind energy turbine set SCADA data, is easy to programming realization, can reflects the exception of parameter accurately and effectively; Consider the accuracy that the abnormal identification result of multiple forecast model can improve the abnormal identification of state parameter.

Accompanying drawing explanation

In order to make object of the present invention, technical scheme and beneficial effect clearly, the invention provides following accompanying drawing and being described:

Fig. 1 is the construction method schematic diagram of state parameter forecast model;

Fig. 2 is the interval division schematic diagram of prediction residual;

Fig. 3 is the half trapezoidal Membership Function Distribution figure combined with triangle;

Fig. 4 is the abnormal identification flow process of state parameter.

Embodiment

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described in detail.

1, Wind turbines state parameter classification

The Wind turbines state parameter of the present embodiment research comes from wind energy turbine set SCADA system.By such environmental effects whether according to, state parameter is divided into 2 classes by the present invention, and the 1st class state parameter comprises gear case oil temperature, generator unit stator winding temperature, wheel speed etc., this type of parameter and physical environment closely related.2nd class parameter comprises oil pressure, hydraulic pressure, yawing velocity etc., and this type of state parameter is only relevant with the performance of equipment self.In above-mentioned two class state parameters, the abnormal identification of the 2nd class state parameter only can need be realized by the setting of threshold value, and the 1st class state parameter cannot be judged by value due to the identification affecting its state parameter by physical environment.Therefore, the present invention only carries out abnormal discrimination method research for the 1st class state parameter.With certain wind energy turbine set SCADA system domestic for example, Wind turbines state parameter and classify as shown in table 1.

Table 1 Wind turbines state parameter

2, the framework of the abnormal identification model of state parameter generalized fuzzy

The data sample of the 1st class state parameter and this moment unit operation mode and physical environment correlativity are comparatively large, cannot directly be differentiated its abnormal conditions by the value of state parameter.Therefore, the data sample that the present embodiment is corresponding dissimilar, sets up state parameter forecast model corresponding with it respectively, and propose the abnormal identification model of generalized fuzzy of state parameter, specific implementation process is divided into following 4 steps:

1) Monitoring Data of certain period in current SCADA system is gathered, and obtain the time series of the state parameter time series needing to carry out analyzing and the forecast model input parameter corresponded.

2) corresponding each forecast model obtains the predicted value of state parameter under this data sample, and obtains the time series of state parameter prediction residual.

3) calculate the abnormal index of each residual sequence according to the abnormal quantization method of residual sequence, characterize the state parameter intensity of anomaly under each forecast model.

4) with the abnormal index of all residual sequences structure index system, fuzzy comprehensive evaluation method is adopted to carry out abnormal identification to state parameter.

3, the Forecasting Methodology of state parameter

In the present invention, the forecast model of set state ginseng is for reflecting the incidence relation in SCADA system between each state parameter.Choose the time period of Water demand, predict with forecast model to state parameter, if set state parameter is in normal condition, then this residual error feature predicted the outcome conforms to the residual error feature of data sample; If set state parameter is in abnormal conditions, then the residual error feature of this residual error predicted the outcome and data sample has significant difference.The forecast model construction method of state parameter is as shown in Figure 1: the input parameter 1) selecting forecast model; 2) training sample of forecast model and the test sample book of residual error statistical characteristic analysis is extracted; 3) type of analyses and prediction model, structure and training method, adopts training sample to train.

1) selection of input parameter

In forecast model, the selection of input parameter is divided into two types, and the 1st class state parameter and wind speed and environment temperature correlativity are comparatively large, the prediction of this type of state parameter using wind speed, environment temperature and upper a period of time temperature as the input parameter of model; 2nd class state parameter and wind speed correlativity are comparatively large, and the prediction of this type of state parameter is using wind speed, propeller pitch angle and crab angle error as input parameter.

2) extraction of sample data

The training sample of forecast model takes three kinds of extracting modes: the 1st kind for the recent sample of the machine, data sample derives from this TV station unit self, and data sample time section to be taken in recent one month and not containing time period data to be analyzed; 2nd kind is the machine historical sample, and data sample derives from the machine, and data sample time section is taken at same month in last year; Be the recent sample of similar unit in 3rd, data sample derives from wind field other units of the same type, and data sample time section is recent one month data and comprises time period data to be analyzed.

In addition, the extraction of all sample datas does not all comprise forced outage event, if this operation period contains forced outage event, should remove the data of stopping transport in first 2 days.For ensureing that residual distribution modeling contains sufficient statistics, the ratio that the present invention arranges training sample and test sample book is 6:4.

3) data prediction

First data sample carries out pre-service to data before training, and to ensure validity and the accuracy of data, mainly comprises: reject invalid data, average value processing and normalized.In addition, change parameter faster for state parameter, adopt 1 minute average treatment, as rotating speed and power etc.For the parameter that state parameter change is slower, adopt 10 minutes average treatment, as temperature mode parameter.

4) training of forecast model

The present invention utilizes reverse transmittance nerve network (BPNN) to set up the forecast model of state parameter.Its structure of BPNN model consists of: 1 input layer, 1 hidden layer and 1 output layer.Wherein, the number of input state parameter equals the nodes of mode input layer, and hidden layer neuron number is determined by 20 chi independent experiment training networks, and output layer nodes is 1, and hidden layer transport function is tansig type, and output layer transport function is logsig type.

4, the method for building up of the abnormal identification model of state parameter generalized fuzzy

1) intensity of anomaly of prediction residual quantizes

If set state parameter is normal, then in predicted time section, residual distribution feature conforms to the residual distribution of training sample.With this, set up the abnormal index based on residual distribution model: if prediction residual is in the larger span of probability density value in residual distribution characteristic, then this parameter is normal; If prediction residual is in low probability density interval in residual distribution characteristic, then this parameter is abnormal.According to above-mentioned principle to divide residual error span: with the residual values of fractile 0.025,0.25,0.75,0.975 correspondence, residual error span is divided into 3 intervals.As shown in Figure 2.

Definition residual error abnormal index (ErrorAbnormalIndex, EAI) characterizes the intensity of anomaly of residual error, and computing formula is as follows:

$EAI=1-\frac{N_1*C_1}{\underset{i=1}{\overset{3}{\&Sigma;}}{N}_i*C_i}---\left(1\right)$

In formula, N _ifor value in residual sequence is in the number in interval i, C _ifor value is in the penalty factor in interval i, be set to [1,3,5].EAI is larger, shows that residual error intensity of anomaly is higher.

2) the fuzzy abnormal identification of state parameter

The present invention adopts fuzzy comprehensive evaluation method to carry out the abnormal identification of parameter.The judgment index system of the method is made up of output layer and indicator layer: indicator layer is the residual error abnormal index EAI that forecast model obtains _i, output layer is abnormal evaluation result.In addition, the exception level of the present invention to state parameter is divided into " normally ", " attention " and "abnormal" three kinds of situations, and namely comment is: L=[normal, note, abnormal]=[l ₁, l ₂, l ₃].Residual error abnormal index EAI _ito the state l in Comment gathers _idegree of membership be v _ij, then available membership grade sets V _i=[v _i1, v _i2, v _i3] represent with residual error abnormal index EAI _icarry out the result assessed.The fuzzy matrix for assessment of output layer is:

$V=\left[\begin{array}{c}{V}_1\\ V_2\\ M\\ V_n\end{array}\right]=\left[\begin{array}{ccc}{v}_{11}& v_{12}& v_{13}\\ v_{21}& v_{22}& v_{23}\\ M& M& M\\ v_{n1}& v_{n2}& v_{n3}\end{array}\right]---\left(2\right)$

Adopt the distribution membership function of triangle and half trapezoidal combination, as shown in Figure 3.The concrete defining method of degree of membership is:

$v_1\left({\mathrm{EAI}}_i\right)=\left\{\begin{array}{cc}1& {\mathrm{EAI}}_i<0.4\\ 3.67-6.67{\mathrm{EAI}}_i& 0.4\&le;{\mathrm{EAI}}_i<0.55\\ 0& {\mathrm{EAI}}_i\&GreaterEqual;0.55\end{array}\right.---\left(3\right)$

$v_2\left({\mathrm{EAI}}_i\right)=\{\begin{array}{cc}0& {\mathrm{EAI}}_i<0.4{\mathrm{orEAI}}_i\&GreaterEqual;0.7\\ 6.67{\mathrm{EAI}}_i-2.67& 0.4\&le;{\mathrm{EAI}}_i<0.55\\ 4.67-6.67{\mathrm{EAI}}_i& 0.55\&le;{\mathrm{EAI}}_i<0.7\end{array}---\left(4\right)$

$v_3\left({\mathrm{EAI}}_i\right)=\left\{\begin{array}{cc}0& {\mathrm{EAI}}_i<0.55\\ 6.67{\mathrm{EAI}}_i-3.67& 0.55\&le;{\mathrm{EAI}}_i<0.7\\ 1& {\mathrm{EAI}}_i\&GreaterEqual;0.7\end{array}\right.---\left(5\right)$

Set the weight of each index with the prediction accuracy of each forecast model to test sample book, computing method are as follows:

${\mathrm{\&omega;}}_i=(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})/\underset{i=1}{\overset{N}{\&Sigma;}}(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})=(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})/(N-1)---\left(6\right)$

In formula, MAE _ifor the mean absolute error value that forecast model i predicts test sample book.

Comprehensive foregoing, as shown in Figure 4, concrete steps are as follows for the process of the fuzzy abnormal identification of state parameter:

S1: the mean absolute error obtaining test sample book according to the selection result of the abnormal identification model of state parameter generalized fuzzy, and obtain each forecast model weights omega by formula (6) _i:

S2: realize predicting the state parameter of period to be analyzed by each forecast model; The prediction of model adopts the mode of window rolling, and the state parameter abnormal identification time interval is set as 6 hours;

S3: realize the status analysis to state parameter by the residual error of each forecast model; The data selection of residual error characteristic statistics analysis only considers the state parameter point be in sample span, and all the other future positions are as invalid prediction point; Choose the cumulative data of 24 hours effective residual errors, quantitative analysis each model state abnormal parameters degree, obtain each model residual error abnormal index EAI _i;

S4: the fuzzy membership calculating each index according to formula (3)-formula (5), forms fuzzy matrix for assessment, adopts formula (7) to calculate output layer assessed value;

$B=\left[\begin{array}{ccc}{b}_1& b_2& b_3\end{array}\right]\left[\begin{array}{cccc}{\mathrm{\&omega;}}_1& {\mathrm{\&omega;}}_2& L& {\mathrm{\&omega;}}_n\end{array}\right]\left[\begin{array}{ccc}{v}_{11}& v_{12}& v_{13}\\ v_{21}& v_{22}& v_{23}\\ M& M& M\\ v_{n1}& v_{n2}& v_{n3}\end{array}\right]---\left(7\right)$

S5: the maximum principle of input foundation degree of membership of assessment result, namely with b _max=max (b _i) corresponding comment l _ias evaluation result.

Embodiment:

This example is with certain wind energy turbine set SCADA system data instance, elaborate the process of the abnormal discrimination method of the state parameter based on fuzzy comprehensive evoluation that the present invention proposes, and contrast with the abnormal identification result of Individual forecast model, the validity of checking the inventive method and accuracy.

Choose the data instance of wind energy turbine set unit forced outage, this unit is reported to the police on July 15th, 2013 and is shut down.By transferring the record of examination table of this unit, find this unit due to the collector ring oxidation of set generator system serious, and frequency converter excitation invalid can not realize grid-connected.This is shut down and amounts to 72 hours, and loss electricity reaches 10000kWh.

This example sets forth the implementation procedure of abnormal identification with wind power generating set generator system front bearing temperature in SCADA system.First carry out choosing of sample data, the time period of data sample is set as that on April 1st, 2013 was to May 30, by each forecast model to generator system front bearing temperature prediction.After obtaining the result of forecast model, state grade division is carried out to the prediction residual statistical distribution characteristic of each model.Calculate the abnormal index of each model to parameter prediction residual error according to formula (1), then calculate the degree of membership of each residual error abnormal index according to formula (3) ~ formula (5), obtain Judgement Matrix:

$V=\left[\begin{array}{ccc}0& 0.87& 0.13\\ 0& 0& 1\\ 0& 0& 1\\ 0& 0& 1\end{array}\right]$

Calculate each Model Weight by formula (6), result is [0.28,027,0.23,0.22].The fuzzy membership calculating output layer according to formula (7) is [0,0.24,0.76].It can thus be appreciated that parameter is in abnormality.And if only adopt the forecast model based on Recent data, abnormal parameters identification result is " attention ".Therefore, adopt the abnormal discrimination method of the state parameter of multi-model prediction residual distribution more accurate.

What finally illustrate is, above preferred embodiment is only in order to illustrate technical scheme of the present invention and unrestricted, although by above preferred embodiment to invention has been detailed description, but those skilled in the art are to be understood that, various change can be made to it in the form and details, and not depart from claims of the present invention limited range.

Claims (4)

1., based on the abnormal discrimination method of Wind turbines state parameter of fuzzy comprehensive evoluation, it is characterized in that: comprise the following steps:

S1: the mean absolute error obtaining test sample book according to the selection result of the abnormal identification model of state parameter generalized fuzzy, and obtain each forecast model weights omega by following formula _i:

${\mathrm{\&omega;}}_i=(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})/\underset{i=1}{\overset{N}{\&Sigma;}}(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})=(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})/(N-1)$

In formula, MAE _ifor the mean absolute error value that forecast model i predicts test sample book;

S2: realize predicting the state parameter of period to be analyzed by each forecast model; The prediction of model adopts the mode of window rolling, and the state parameter abnormal identification time interval is set as 6 hours;

S3: realize the status analysis to state parameter by the residual error of each forecast model; The data selection of residual error characteristic statistics analysis only considers the state parameter point be in sample span, and all the other future positions are as invalid prediction point; Choose the cumulative data of 24 hours effective residual errors, quantitative analysis each model state abnormal parameters degree, obtain each model residual error abnormal index EAI _i;

S4: the fuzzy membership calculating each index, forms fuzzy matrix for assessment, calculates output layer assessed value;

S5: the maximum principle of input foundation degree of membership of assessment result, namely with b _max=max (b _i) corresponding comment l _ias evaluation result.

2. the abnormal discrimination method of a kind of Wind turbines state parameter based on fuzzy comprehensive evoluation according to claim 1, is characterized in that: in step sl, and the framework of the abnormal identification model of state parameter generalized fuzzy is as follows:

Corresponding dissimilar data sample, set up state parameter forecast model corresponding with it respectively, specific implementation process is divided into following 4 steps:

1) Monitoring Data of certain period in current SCADA system is gathered, and obtain the time series of the state parameter time series needing to carry out analyzing and the forecast model input parameter corresponded;

2) corresponding each forecast model obtains the predicted value of state parameter under this data sample, and obtains the time series of state parameter prediction residual;

3) calculate the abnormal index of each residual sequence according to the abnormal quantization method of residual sequence, characterize the state parameter intensity of anomaly under each forecast model;

4) with the abnormal index of all residual sequences structure index system, fuzzy comprehensive evaluation method is adopted to carry out abnormal identification to state parameter.

3. the abnormal discrimination method of a kind of Wind turbines state parameter based on fuzzy comprehensive evoluation according to claim 2, it is characterized in that: in step s 2, the Forecasting Methodology of described state parameter specifically comprises:

The forecast model of set state ginseng is for reflecting the incidence relation in SCADA system between each state parameter, choose the time period of Water demand, with forecast model, state parameter is predicted, if set state parameter is in normal condition, then this residual error feature predicted the outcome conforms to the residual error feature of data sample; If set state parameter is in abnormal conditions, then the residual error feature of this residual error predicted the outcome and data sample has significant difference; The forecast model construction method of state parameter comprises the following steps: the input parameter selecting forecast model; Extract the training sample of forecast model and the test sample book of residual error statistical characteristic analysis; The type of analyses and prediction model, structure and training method, adopt training sample to train;

Specific as follows:

1) selection of input parameter

In forecast model, the selection of input parameter is divided into two types, and the 1st class state parameter and wind speed and environment temperature correlativity are comparatively large, the prediction of this type of state parameter using wind speed, environment temperature and upper a period of time temperature as the input parameter of model; 2nd class state parameter and wind speed correlativity are comparatively large, and the prediction of this type of state parameter is using wind speed, propeller pitch angle and crab angle error as input parameter;

2) extraction of sample data

The training sample of forecast model takes three kinds of extracting modes: the 1st kind for the recent sample of the machine, data sample derives from this TV station unit self, and data sample time section to be taken in recent one month and not containing time period data to be analyzed; 2nd kind is the machine historical sample, and data sample derives from the machine, and data sample time section is taken at same month in last year; Be the recent sample of similar unit in 3rd, data sample derives from wind field other units of the same type, and data sample time section is recent one month data and comprises time period data to be analyzed;

In addition, the extraction of all sample datas does not all comprise forced outage event, if this operation period contains forced outage event, should remove the data of stopping transport in first 2 days; For ensureing that residual distribution modeling contains sufficient statistics, the ratio arranging training sample and test sample book is 6:4;

3) data prediction

First data sample carries out pre-service to data before training, and to ensure validity and the accuracy of data, mainly comprises: reject invalid data, average value processing and normalized; In addition, change parameter faster for state parameter, adopt 1 minute average treatment, as rotating speed and power etc.; For the parameter that state parameter change is slower, adopt 10 minutes average treatment, as temperature mode parameter;

4) training of forecast model

Reverse transmittance nerve network (BPNN) is utilized to set up the forecast model of state parameter; Its structure of BPNN model consists of: 1 input layer, 1 hidden layer and 1 output layer; Wherein, the number of input state parameter equals the nodes of mode input layer, and hidden layer neuron number is determined by 20 chi independent experiment training networks, and output layer nodes is 1, and hidden layer transport function is tansig type, and output layer transport function is logsig type.

4. the abnormal discrimination method of a kind of Wind turbines state parameter based on fuzzy comprehensive evoluation according to claim 3, is characterized in that: adopt following methods to carry out the foundation of the abnormal identification model of state parameter generalized fuzzy:

1) intensity of anomaly of prediction residual quantizes

If set state parameter is normal, then in predicted time section, residual distribution feature conforms to the residual distribution of training sample; With this, set up the abnormal index based on residual distribution model: if prediction residual is in the larger span of probability density value in residual distribution characteristic, then this parameter is normal; If prediction residual is in low probability density interval in residual distribution characteristic, then this parameter is abnormal; According to above-mentioned principle to divide residual error span: with the residual values of fractile 0.025,0.25,0.75,0.975 correspondence, residual error span is divided into 3 intervals;

Definition residual error abnormal index (ErrorAbnormalIndex, EAI) characterizes the intensity of anomaly of residual error, and computing formula is as follows:

$EAI=1-\frac{N_1*C_1}{\underset{i=1}{\overset{3}{\&Sigma;}}{N}_r*C_i}---\left(1\right)$

In formula, N _ifor value in residual sequence is in the number in interval i, C _ifor value is in the penalty factor in interval i, be set to [1,3,5]; EAI is larger, shows that residual error intensity of anomaly is higher;

2) the fuzzy abnormal identification of state parameter

Adopt fuzzy comprehensive evaluation method to carry out the abnormal identification of parameter, the judgment index system of the method is made up of output layer and indicator layer: indicator layer is the residual error abnormal index EAI that forecast model obtains _i, output layer is abnormal evaluation result; In addition, be divided into " normally ", " attention " and "abnormal" three kinds of situations to the exception level of state parameter, namely comment is: L=[normal, note, abnormal]=[l ₁, l ₂, l ₃]; Residual error abnormal index EAI _ito the state l in Comment gathers _idegree of membership be v _ij, then available membership grade sets V _i=[v _i1, v _i2, v _i3] represent with residual error abnormal index EAI _icarry out the result assessed; The fuzzy matrix for assessment of output layer is:

$V=\left[\begin{array}{c}{V}_1\\ V_2\\ M\\ V_n\end{array}\right]=\left[\begin{array}{ccc}{v}_{11}& v_{12}& v_{13}\\ v_{21}& v_{22}& v_{23}\\ M& M& M\\ v_{n1}& v_{n2}& v_{n3}\end{array}\right]---\left(2\right)$

Adopt the distribution membership function of triangle and half trapezoidal combination, the concrete defining method of degree of membership is:

$v_1\left({\mathrm{EAI}}_i\right)=\left\{\begin{array}{cc}1& {\mathrm{EAI}}_i<0.4\\ 3.67-6.67{\mathrm{EAI}}_i& 0.4\&le;{\mathrm{EAI}}_i<0.55\\ 0& {\mathrm{EAI}}_i\&GreaterEqual;0.55\end{array}\right.---\left(3\right)$

$v_2\left({\mathrm{EAI}}_i\right)=\left\{\begin{array}{cc}0& {\mathrm{EAI}}_i<0.4{\mathrm{orEAI}}_i\&GreaterEqual;0.7\\ 6.67{\mathrm{EAI}}_i-2.67& 0.4\&le;{\mathrm{EAI}}_i<0.55\\ 4.67-6.67{\mathrm{EAI}}_i& 0.55\&le;{\mathrm{EAI}}_i<0.7\end{array}\right.---\left(4\right)$

$v_3\left({\mathrm{EAI}}_i\right)=\left\{\begin{array}{cc}0& {\mathrm{EAI}}_i<0.55\\ 6.67{\mathrm{EAI}}_i-3.67& 0.55\&le;{\mathrm{EAI}}_i<0.7\\ 1& {\mathrm{EAI}}_i\&GreaterEqual;0.7\end{array}\right.---\left(5\right)$

Set the weight of each index with the prediction accuracy of each forecast model to test sample book, computing method are as follows:

${\mathrm{\&omega;}}_i=(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})/\underset{i=1}{\overset{N}{\&Sigma;}}(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})=(1-\frac{{\mathrm{MAE}}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{MAE}}_i})/(N-1)---\left(6\right)$

In formula, MAE _ifor the mean absolute error value that forecast model i predicts test sample book;

Calculate the fuzzy membership of each index according to formula (3)-formula (5), form fuzzy matrix for assessment, adopt formula (7) to calculate output layer assessed value.

$B=\left[\begin{array}{ccc}{b}_1& b_2& b_3\end{array}\right]=\left[\begin{array}{cccc}{\mathrm{\&omega;}}_1& {\mathrm{\&omega;}}_2& L& {\mathrm{\&omega;}}_n\end{array}\right]\left[\begin{array}{ccc}{v}_{11}& v_{12}& v_{13}\\ v_{21}& v_{22}& v_{23}\\ M& M& M\\ v_{n1}& v_{n2}& v_{n3}\end{array}\right]---\left(7\right).$