Power signal classification using dynamic wavelet network
Introduction
Recently power quality [1], [2], [3], [4], [5], [6], [7] (PQ) and related power supply issues have become quite a serious problem for both the end users and the utilities. According to established reports, the industry is losing huge amount of resources due to power outages and PQ problems. PQ may be defined as the continuous availability of electric power that confirms to accepted standards of phase, frequency, and voltage. Power system disturbance is defined as a variation in these standards. The PQ issues and related phenomena can be attributed to the use of solid-state switching devices, unbalanced and non-linear loads, industrial grade rectifiers and inverters, computer and data processing equipments, etc. which are being increasingly used in both the industry and home appliances. These devices introduce distortions in the phase, frequency and amplitude of the power system signal thereby deteriorating PQ. Subsequent effects could range from overheating, motor failures, inoperative protective equipment to power inrush, quasistatic harmonic distortions and pulse type current disturbances. Therefore it is essential to monitor these disturbances.
There are different types of disturbances which occur in power systems due to various reasons. These disturbances can be broadly classified into two types namely, steady-state and short duration disturbances. Power System disturbances like voltage sag, voltage swell, harmonics, sag with harmonics, swell with harmonics, voltage flicker, momentary interruption, momentary interruption with harmonics, Flicker, Flicker with harmonics, etc. are the constituents of steady-state class. These types of disturbances are short-term under or over-voltage conditions that can last from one cycle to several cycles of the 50 Hz AC mains signal. In fact harmonics introduced by power electronic switching devices can persist in the power signal permanently. Disturbance signals like spike, notch, transient, chirp increasing and decreasing, etc. are categorized under short duration disturbances. These disturbances appear in the power line for very short durations typically in the microseconds range. These disturbances occur and decay rapidly with time and hence are difficult to monitor as well as remove. These are always unpredictable in nature. Thus steady-state disturbances account for the maximum PQ distortion and hence these must be effectively detected and classified in an online system.
One of the important issues in power signal disturbance analysis is to detect and classify the disturbance waveforms automatically in an efficient manner and in this regard wavelet transform (WT) [1], [2], [3], [4] plays a significant role in the detection and localization of the power signal disturbances. However, WT alone cannot accurately classify the various types of disturbances occurring in the power signal. Therefore to effectively identify a disturbance in the power signal waveform several methods based on WT and artificial neural network (ANN) has been presented in the last few years [5], [6], [7]. In these methods the power signal is processed using WT to provide several features that are used by the ANN algorithm to produce effective recognition of the power signal disturbances. Traditionally the neural networks have been used to carry out the classification task for the PQ disturbances. However, the neural networks have some drawbacks that include the determination of network architecture, network parameter assignment, and the failure to adapt in a dynamic environment. Thus adaptive networks like the probabilistic neural network (PNN) [8], [9], [10], [11], [12], and general regression neural networks (GRNN) [13] having expandable or reducible network architectures, and fast learning speed are suitable for online applications in PQ monitoring.
A combination of wavelet transform and PNN has been presented recently [5] using the energy distribution and time duration at each of the 13 decomposition levels as inputs to the PNN for classification of seven types of power signal disturbance events. Seven types of power signal disturbances have been classified by using the features obtained from WT as inputs to a self-organizing learning array [6]. Wavelet-based online disturbance detection for PQ applications are discussed in [14], but the method is not suitable for the classification of various types of disturbances that occur in the power signal under dynamic environment. To achieve robust steady-state power signal detection and classification process in a dynamic environment, an integrated wavelet network known as dynamic wavelet network (DWN) comprising an input wavelet transform layer and adaptive PNN layers is used. The wavelet layer transforms the time domain signal into a time-scale domain, thereby giving scale and time localized features that are used by the neural network for the pattern classification task. Non-orthonormal basis functions have been used as the activation functions of the wavelet layer. The PNN is implemented using a probabilistic model such as Bayesian classifiers and gives the probability of belongingness of a particular input pattern towards all the classes.
The term ‘Dynamic’ is justified as the DWN has the capability of automatic adjustment of learning cycles for different classes of non-stationary signals, for minimizing error. In this model the number of learning cycles is variable and it need not be a predetermined fixed value for all kinds of signal combinations. Depending on the type of disturbance signals provided and the learning rate, the network decides the number of learning cycles required in order to obtain minimal error. This helps in avoiding unnecessary learning iterations hence providing fast learning.
Section snippets
Wavelet layer
The Fourier transform (FT) decomposes a time series signal in terms of a sum of sinusoidal basis functions. The FT provides a continuous spectrum of the frequency components present in the signal which can be recombined to obtain the original time domain signal. Thus the FT provides an efficient tool for the harmonic analysis of stationary signals. But one demerit that can be attributed to the FT is that, once the signal has been transformed into the frequency domain none of the time
Creation of training patterns
For the purpose of analysis, we have used synthetic signals generated using MATLAB. The fundamental sinusoidal signal is taken to be of 50 Hz and it is sampled at 64 points per cycle. Thus the sampling frequency is 3.2 kHz (50 × 60). The signals are generated for different levels as well as starting positions. Hence the proposed model can be effectively put into use in an online signal detection system (OSDS) for accurate detection and classification of various steady-state non-stationary power
Discussion
In the simulation process we have applied different types and combinations of steady-state power network disturbance signals to the DWN. It was observed that DWN achieved remarkable efficiency and accuracy. Irrespective of the training patterns given, the network tuned the smoothing parameter sigma, thereby adjusting the network weights and activation functions of the hidden layer. To achieve this, the network automatically decides the number of learning cycles required, e.g. when harmonics,
Test results
The test results for PQ disturbance signals with parameters like dilation, translation learning rate are shown in Table 2, Table 3, Table 4, Table 5, Table 6, Table 7.
Conclusion
The DWN is a robust pattern classifier and has higher pattern recognition accuracy in classifying various power signal disturbances in comparison to wavelet-based PNN. Therefore the OSDS based on DWN has been developed in this paper for power signal disturbance classification. The DWN combines the use of Morlet wavelet and adaptive probabilistic network for the detection of different PQ disturbances. The proposed model has a dynamic and fast adaptation algorithm, which tunes the parameters
References (16)
- et al.
Mining for similarities in time series data using wavelet based feature vectors and neural networks
Eng. Appl. Artif. Intell.
(2007) - et al.
ower quality assessment via wavelet transform analysis
IEEE Trans. Power Del.
(1996) - et al.
Application of wavelet to model short-term power system disturbances
IEEE Trans. Power Syst.
(1996) - et al.
Power quality detection and classification using wavelet-multiresolution signal decomposition
IEEE Trans. Power Deliv.
(1999) - et al.
A measurement method based on the wavelet transform for power quality analysis
IEEE Trans. Power Deliv.
(1998) Wavelet-based neural network for power disturbance recognition and classification
IEEE Trans. Power Deliv.
(2004)- et al.
A self-organizing learning array system for power quality classification based on wavelet transform
IEEE Trans. Power Deliv.
(2006) Adaptive probabilistic neural networks for pattern classification in time-varying environment
IEEE Trans. Neural Netw.
(2004)
Cited by (29)
A critical and comprehensive review on power quality disturbance detection and classification
2020, Sustainable Computing: Informatics and SystemsCitation Excerpt :Over 140 papers are reviewed critically and classified into four categories. The first category includes topics on general power quality, their standards, economic considerations and automatic power classifier model [1–28]. The second category [29–83] includes feature extraction techniques such as Fourier transform (FT) based methods [29–42], S transform (ST) based methods [43–54], Hilbert Huang transform (HHT) based methods [55–61], Wavelet transform (WT) based methods [62–72], miscellaneous feature extraction techniques [73,30–83].
A comprehensive overview on signal processing and artificial intelligence techniques applications in classification of power quality disturbances
2015, Renewable and Sustainable Energy ReviewsCitation Excerpt :A combination of fully informed particle swarm (FIPS) and adaptive PNN approach known as PNN based feature selection (PFS) was proposed in [136] to identify the PQ disturbances where the statistical features were extracted by ST and time–time (TT) transform (extension of ST) for training the PNN classifier. An integrated approach of wavelet layer and adaptive PNN known as Dynamic Wavelet Network (DWN) was proposed in [137] particularly suitable for the classification of PQ signals in a dynamic environment with time varying non-stationary PQ signals. The DWN is the combination of the two sub-networks consisting of input DWT layer and adaptive PNN layers.
An integrated PSO for parameter determination and feature selection of ELM and its application in classification of power system disturbances
2015, Applied Soft Computing JournalCitation Excerpt :As always, the PQ engineer is inundated with an enormous amount of data for inspect. Thus it is desirable to develop automatic methods for detecting, identifying, and analyzing various disturbances [1–3]. Such classification methods include two major steps namely the feature extraction using signal processing techniques and classification using application of a statistical or neural classifier to the constructed feature vectors.
Classification of power quality events - A review
2012, International Journal of Electrical Power and Energy SystemsHybrid soft computing techniques for feature selection and parameter optimization in power quality data mining
2011, Applied Soft Computing JournalCitation Excerpt :Traditionally, PQ was judged by visual inspection of the disturbance waveforms, so the engineer's knowledge plays a critical role. As always, the PQ engineer is inundated with an enormous amount of data for inspection it is desirable to develop automatic data mining methods for detecting, classifying, and analyzing various disturbances [1,2]. Such classification methods include two major steps namely the feature extraction using signal processing techniques and classification using application of a statistical or neural classifier to the constructed feature vectors.
Modified wavelet neural network in function approximation and its application in prediction of time-series pollution data
2011, Applied Soft Computing JournalCitation Excerpt :Wavelet neural networks (WNNs) have been introduced as an alternative to MLPs that overcome their shortcomings [10,11]. Due to the advantages of WNNs as universal approximators, the fact that they have more compact topology than other neural networks and their fast learning speed owing to the constitution of the localized wavelet activation function in the hidden layer [12], WNNs had received much attention from other researchers and have been used extensively to solve numerous real world problems such as face recognition, time-series prediction, pattern classification and system identification [13–30]. Various issues have been addressed in WNN studies, which include adjusting the connection weights by employing different learning algorithms for accelerating the convergence of WNNs [9,17,21,52], making alterations in the network architecture [22], introducing variation in the types of activation functions used in the hidden layer [23,24], and modifying the wavelet function parameters, a process that involves proper initialization of the translation and dilation vectors [25–27].