3.2 Feature extraction The following arethe features that are extracted from the signal after processing with DWT. 1.
Standard deviation: It is a value expressing that, by whatextent the signal differs from the mean.2. Energy: The total content energyof the current signal. 3. Kurtosis:Bigger kurtosis point representsmore outlier in the signal 15. 4.
Skewness: It is a measure of the irregularity of theprobabilitydistribution about its mean 15. 3.3 Data mining model forclassification The objective of a data miningmodel is to create an understandable structure by taking data set into account.Based on the model, the system behaviour can be identified. Usually, the datamining model may be classified as descriptive model or predictive model. In the design of HIF detection method,predictive data mining model is preferred because of the requirement of thework i.
e classification. There are several data miningmodels reported 16-18, However, DT model is chosen because of its transparency,efficiency and popularity. The proposed method utilizesopen source data mining package ‘R’ for creating the DT model 19. For HIFdetection, DT model is generated by considering feature set into account. There are four features set training the DT against the targetoutput of 1 for HIF and 0 for non-HIF. From the simulated data set, 70% usedfor training and 30% for testing. The data mining generated DT model with threefeature set for HIF detection shown in Fig.
5. As a result, these features hasmore discriminating capability in the classification of HIF and non-HIF. 4.
Results and discussions: The behaviour of phase currentwith HIF inception time of 0.065 sec isshown in Fig. 6. The high frequency transient due to HIF are clearly seen inwavelet level D1 to D3. The response ofcurrent due to non-linear load switching at time of 0.065 sec is shown inFig.
7. The behaviour of non-linear load show that, it has high peak valuecompared to HIF which can be seen from the wavelet level D1 to D3. Anothernon-HIF event considered in the work is capacitance switching, the decomposedsignal after this event inception time of 0.065 sec is shown in Fig.8. It isvery clear from the wavelet level D2 to D4, there is no significant change in oscillationfew cycles after event inception. After calculating features from the signal,subsequently apply the same to DT modelto take final decision on relay signal. Theproposed method is evaluated through the following three parameters:1.
Dependability:Predicted HIF against total HIFconditions. 2. Security: PredictedNon-HIF against total non-HIF conditions.
3. Accuracy: Actual predicted against total number of conditionsconsidered. Testing was carried out against 278 cases comprises 243HIF and 35 non-HIF cases.
The confusion matrix of the DT model given in Table 3infers that all the HIF and non-HIF cases are correctly classified.Moreover, Power system always susceptible to noise, it is necessary to investigate the proposed methodwith noise. Table 4 compares the performance under noisy atmosphere.
It isobserved that the performance is almost similar till 20 db noise on the signal.Additional boost in SNR, reduces the performance. Hence, the proposed method issuitable for SNR of 20 or less.Conclusions:In this paper, a new methodology based on the WT and data miningbased DT model is presented.
The proposed method uses data mining that hascapability distinguishing HIF pattern. The proposed model extensively tested onthe actual electric power system model with real data using MATLAB with widerange of variation in operating conditions and providing exceptional detectionrate under noisy environment. This indicates that the detection scheme proposedfor HIF is highly reliable and secure.