Abstract – With the development of new functionalitiessolar and wind energy based hybrid system areupcoming energy source with higher efficiency. Solar andwind energy being naturally available in abundance andnon-polluting is one of the most promising sources. Dueto the development of modern power electronics devicesthe power quality of wind solar hybrid system getsaffected. Hence due to the increasing usage of sensitiveelectronic equipments in wind solar hybrid system PowerQuality has become a major concern. Sag, swell andharmonics are the critical aspect of Power Quality issues.This project presents an approach that is able to providethe detection and identification of power qualityproblems. This method is developed by using discretewavelet transform (DWT) analysis. The given signal isdecomposed through wavelet transform. Later, using thewavelet coefficients, Feature Extraction is done and anArtificial Neural Network is developed to classify thepower quality disturbances. The training and testing datarequired to develop the ANN model is generated throughsimulation. In this project, it is demonstrated that eachpower quality disturbance has unique deviations from thepure sinusoidal waveform and this is adopted to provide areliable classification of the type of disturbance. Thecombined mathematical transformation and artificialneural network-based approach is able to classify thepower quality disturbances accurately.Key words: Power Quality, Wavelet Transform, ArtificialNeural Network(ANN), Back propagation Algorithm.I. INTRODUCTIONThe quality of electric power has become an important issuein wind solar hybrid system because, with the introductionand wide spread use of sensitive electronic equipment,customers have become much more aware and sensitive tosag, swell, harmonics and other power anomalies. In order toimprove the power quality of the wind solar hybrid system,the power disturbances should be monitored continuously3.Power quality monitoring and analysis must be able to detect,localize, estimate and classify the disturbances on the hybridsystem.2On the other hand, the wavelet transform has been adoptedin different fields, such as telecommunications and acoustics.In the last decade, the wavelet transform has been studied toanalyze voltages and currents during short durationdisturbances. The main purpose of this paper is to propose anapproach in which the MATLAB wavelet transform isadopted not only to detect power quality problems in thewind solar hybrid system but also to classify them.One of the most important issues in power qualityproblems is how to detect and classify disturbance waveformsautomatically in an efficient manner. Automatic disturbancerecognition can be carried out with the help of wavelettransform1.Using the properties of WT and the features of thedecomposed waveform, along with the ANN algorithm, it ispossible to extract important information from a disturbancesignal and determine what type of disturbance has caused apower quality problem to occur2.II. PROPOSED METHODOLOGYFig. 1. Block Diagram of Proposed MethodologyPower disturbance signals in wind solar hybrid system areusually generated by the change in the wind speed, variationin the solar irradiation and change in the frequency. Thedisturbances can be classified into categories that can vary ineffect, duration and intensity.There are currently many power disturbance-recordingdevices for power system, which are being used to monitorpower disturbance signal, resulting a massive increase in theamount of stored data. Therefore, we have applied WaveletTransform to the compression of power disturbance signalsbased on the properties of decomposition and reconstruction.ANN can be used for classifying the types of disturbancesbased on Wavelet Coefficients7-8. This Paper focuses onAutomatic disturbance recognition using the combination ofwavelet transform and ANN.III. WAVELET TRANSFORMWavelet analysis is the new tool for monitoring powerSagHarmonicsswellFeaturessExtractionClassifier DisturbancecategoryDisturbanceDatabasequality problems. Wavelet transformation has the ability toanalyze different power quality problems simultaneously intime and frequency domain. The wavelet transform is usefulin detecting and extracting features of various types ofelectric power quality disturbances in wind solar hybridsystem because it is sensitive to signal irregularities butinsensitive to the regular signal behaviour.Wavelet analysis deals with expansion of functions interms of a set of basis functions, like Fourier analysis.However, wavelet analysis expands functions not in terms oftrigonometric polynomials but in terms of wavelets, whichare generated in the form of translations and dilations of afixed function called mother wavelet. Compared with Fouriertransform, wavelet can obtain both time and frequencyinformation of signal, while only frequency domaininformation can be obtained from Fourier transform.The signal can be represented in terms of both the scalingand wavelet function as follows.f(t)= ?cj(n)? (t-n)+? ? dj(n)2j/2 ? (2jt-n) _________(1)n n j=0where cj is the J level scaling coefficient,dj is the j level wavelet coefficient,? is scaling function,?(t) is wavelet function,J is the higher level of Wavelet transform,T is timeIV. DISCRETE WAVELET TRANSFORMEqn.(2) has theoretical interest for the development andcomprehension of its mathematical properties. However, itsdiscretization is necessary for practical applications. Fordiscrete time systems, the discretization process leads to thetime discrete wavelet series asDWT ?x(m,n) = ?? x(t) ?*m,n (t)dt __________(2)where, ?m,n(t) = a0-m/2 Ø(t-nb0a0m/a0m),a= a0m and b=nb0 a0m.In power quality analysis, discrete wavelet transformationis adequate to classify the power quality problems. It moves atime domain discritized signal into its corresponding waveletdomain. This is done through a process called “sub-bandcodification”, which is done through digital filter techniques.In the signal processing theory, to filter a given signal f(n)means to make a convolution of this signal. Basically, theDWT evaluation has two stages.Fig. 2. Sub-band Codification of a signalThe first level is the determination of wavelet coefficients.These coefficients represent the given signal in the waveletdomain. From these coefficients, the second stage isachieved with the calculation of both the approximated andthe detailed version of the original signal, in different levelsof resolutions, in the time domain.At the end of the first level of signal decomposition (asillustrated in Fig. 2), the resulting vectors yh(k) and yg(k)will be, respectively, the level 1 wavelet coefficients ofapproximation and of detail. In fact, for the first level , thesewavelet coefficients are called cA1(n) and cD1(n),respectively, as stated below:cA1(n) = (n).hd(- k + 2n) _____(3)cD1(n) = (n).gd(- k + 2n) _____(4)Next, in the same way, the calculation of the approximated(cA2(n)) and the detailed (cD2(n)) version associated to thelevel 2 is based on the level 1 wavelet coefficient ofapproximation (cA1(n)). The process goes on, alwaysadopting the “n-1” wavelet coefficient of approximation tocalculate the “n” approximated and detailed waveletcoefficients.Once all the wavelet coefficients are known, the discretewavelet transform in the time domain can be determined.This is achieved by “rebuilding” the corresponding waveletcoefficients, along the different resolution levels. Thisprocedure will provide the approximated (aj(n)) and thedetailed (dj(n)) version of the original signal as well as thecorresponding wavelet spectrum.V. FEATURE EXTRACTIONThe purpose of the feature selection and extraction is toreduce the dimension of the data. The feature extraction isbased on the transformations that make it possible to reducethe dimension of the data. Principal component analysis ishd(n) gd(n)yh(k) yg(k)2 2f(n)the most commonly used feature extraction method.The principal components are the linear combinations ofthe original variables. They represent a new co-ordinatesystem where the new axes represent the directions withmaximum variability. Thus the total system variability can bepresented with a smaller number of principal components.The following steps are used to extract the features.The following seven features of different types of powerquality events have been extracted: Mean (?), standarddeviation (?2), skewness (g1), kurtosis (g2), RMS, form factor(FF) and Fast Fourier Transform(FFT).??? ?Ntt t Ex N x1? 1/ _________________(5)2112 ) ( / 1 ) ( ? ? ? ? ? ? ? ???Ntt t t E x Ex N x _____(6)??? ?Ntt g N x131 1/ 6 (( ? ) /? ) _______________(7)/ 24{1/ (( ) / ) 3} 412 ? ? ? ??? ?Ntt g N N x _____(8)???Ntt RMS N x12 1/ _________________(9)FF ? ? / rms __________________(10)Thus, the important features of each type of fault eventsare extracted from the wavelet coefficients and used to trainthe neural network.VI. DEVELOPMENT OF ANN MODELThe developed ANN model has an input layer with 7 nodesand the hidden layer has 10 nodes, while the output layer has3 nodes representing the type of disturbances: voltage sag,voltage swell and harmonics11.The Feed Forward Artificial Neural Network is trained byback propagation method. The simulated events are dividedinto two sets, one of which is to be used for neural-networktraining and the other for testing. The training set consistedof 150 training samples with three disturbance classes. Thetest set consisted of 50 samples (10 samples per class).The training is set for maximum epochs=500 and learningrate=1*103. If the training process cannot reach the targeterror, the other set of input signal for training is selected.Verification of training results is performed in a way that theANN is first trained with data patterns, and then tested withsamples, which are not used for training.VII. SIMULATION RESULTSThis section discusses the details of simulation carried outusing wavelet transform as features and ANN as classifier.The signals with various power quality disturbances aregenerated using standard equations in MATLABenvironment. In this project, three power quality disturbancesare generated. A total of 150 samples with 50 from each classwere generated to develop the neural network.Fig. 3 output of ANNThe above mentioned figure represents the outputfrom the ANN which contains the five classes of powerquality disturbances and its mismatch in few signals. Theresult shows that the network is trained properly and it canclassify different types of disturbances correctly.Table 1. Classification result of power qualityThe seven features were extracted from thedisturbance signals using Wavelet Transforms. Thedeveloped ANN model has an input layer with 7 nodes. Thehidden layer has 10 nodes and the output layer has 3 nodesrepresenting the various disturbances voltage sag, swell andharmonics. The inputs for training ANN are the five featuresextracted as mentioned in the previous section. The trainingis set with learning rate 1e-006, maximum epochs 500 andtarget error 0.0001. After training of the ANN, the weightand bias can be used to detect the power quality problems.The randomly selected 50 signals from each power qualityevent is used to test the ANN.Table 2. Distribution of misclassified dataTable I and II gives the simulation result for five classes ofPQ disturbance problem using the proposed method . Thediagonal elements in table II represent the correctly classifiedPQ types. The off diagonal elements represent themisclassification. From the table II, it is clear that out of 150sag, swell and harmonics nothing is misclassified. Hence theproposed methodology is able to detect and classify 100% ofthe power quality problems correctly.VIII. CONCLUSIONThe method proposed in this paper can effectivelyclassify different kinds of PQ disturbances in the wind solarhybrid system. Totally 150 samples were generated using thisprocess. Thus the Wavelet transform is used for detectingand extracting disturbance features. ANN is used forclassifying various types of fault and the result shows that itcan classify all types of disturbances successfully. Thepercentage of accuracy is also calculated.