Signal Processing and Machine Learning for Power Quality Disturbance Detection and Classification Roberto Togneri (Signal Processing and Recognition Lab)
Power Quality (PQ) disturbances are broadly classified into three categories: magnitude variations: o e.g. sag, swell, interruption o caused by inception of line faults and penetration of heavy or light loads
Power Quality (PQ) disturbances are broadly classified into three categories: sudden transients o e.g. spikes, notch, oscillatory o caused by network switching effects and natural lightning events
Power Quality (PQ) disturbances are broadly classified into three categories: steady-state harmonics o e.g. harmonics, flicker o caused by nonlinear loads and power converters.
Why is Power Quality disturbance (PQD) classification and detection important? Increasing incidence of PQD due to: o switching of renewable energy sources (wind, solar) sag, swell, interruption, flicker o power grids utilising (nonlinear) power electronic technology harmonics, flicker Importance underscored by IEEE Std. 1459-2010 and UNE standard
Block Diagram of PQD classifier This follows the usual pattern recognition paradigm commonly deployed for object recognition and biometric recognition consisting of two key components which are common to all such systems: Feature extraction using some form of short-time or frame based spectral transform (i.e. signal processing) Classification using standard static generative (probabilistic) or discriminative classification systems (i.e. pattern recognition and machine learning)
Feature Extraction There is large variety of possible transform methods which have been investigated o The most common used techniques are derived from the Wavelet transform, S-transform and Hibert-Huang transforms (aka Empirical Mode Decomposition or EMD). These features are typical used for multiresolution time-frequency analysis common in physical systems generating signal analysis such PQD, vibration analysis, etc. o Recent research has also started to exploit advanced approaches like compressed sensing (CS) and non-negative matrix factorisation (NMF) and neural network or deep feature learning methods which operate directly on the time-domain power signal.
Classification The usual suspects: o The classic ANNs and Fuzzy systems (so 90s) o The very popular Support Vector Machine (SVM) o The ubiquitous Gaussian Mixture Model (GMM) The new kids on the block (yet to be realised) o Deep Neural Networks (DNNs) and variants: CNN, DBM, RNN o Random Forest Classifiers (RFC) o Sparse Representation Classifiers (SRC) and Linear Regression Classifiers (LRC)
Feature Selection? Feature selection can be used after feature extraction and before classification to remove redundant feature components and reduce the feature dimension more compact, discriminative features o Mapping techniques: PCA (principal components analysis) and LDA (linear discriminant analysis), common in speech/image feature selection o Optimization techniques: GA (genetic algorithms) and PSO (particle swarm optimisation) o Information Theoretic techniques: mrmr (minimum redundancy, maximum relevance), common in bioinformatics and biomedical feature selection.
The state of play? With such good performance this is seems to be a solved problem? But not really since these results: o Are for single fault classification rather than multiple faults but how common are multiple faults and do faults overlap and do they need to be individually detected and classified? o Are for relatively clean power signals but what type of noise or channel distortion do power signals have (unlike communications and audio signals which are buried in much noise and distortions). o Are for fault event classification but what about fault event detection where the onset and offset times of the fault are important, is this a concern for PQD? There is more that can be done when more complex PQD scenarios arise.
Jan 2014 IEEE TSE
Jan 2014 IEEE TSE
Jan 2014 IEEE TSE
Jan 2014 IEEE TSE CONCLUSION? Results pretty good, we need to complicated things some more!
May 2014 IEEE TIE Now it gets more fancy o ADALINE NN with adaptive-step LMS provides harmonic estimates which are used to calculate the Vrms, THD from the enhanced spectrum; histogram counts of the voltage levels over a half-cycle are taken directly from the voltage signal. o Broad class faults are detected and classified using different systems (Vrms thresholding, THD thresholding and FFNN classification)
May 2014 IEEE TIE
May 2014 IEEE TIE
May 2014 IEEE TIE We note the following o For single faults accuracy is quite good (> 85%) and as expected performance degrades with multiple faults o Noise degrades performance as one would expect.
JAN 2015 IEEE TIE Now it gets conceptually more interesting o Use Sparse Signal Decomposition (SSD) using overcomplete hybrid dictionaries o The rms envelope of the detail coefficients localise (detect) and characterise (classify) transient faults o The rms envelope of the approximation coefficients capture and better characterise long-term harmonic behaviours (e.g.. flicker) for classification Comparison between SSD, WT and ST: (a) Uncategorised disturbance (mainly spike with harmonic disturbance): SSD: detail is the spike, approx captures the harmonic change WT: poor characterisation of the two effects ST: requires further interpretation (b) 65% sag and transient events: SSD: detail is the transient/switch, approx is the sag WT: 1-4 is the transient, 5-7 is the sag (with some transient) ST: no idea!
JAN 2015 IEEE TIE (c) flicker and transient events: SSD: detail is the oscillatory transient, approx captures the harmonic flicker WT: dual characterisation of transient (c1 and c2) and distorted flicker (c3) ST: requires further interpretation
JAN 2015 IEEE TIE What can we conclude? The proposed method is competitive with the ADADLINE+FFNN approach for single faults The proposed method fails when confronted with dual faults which involve a harmonic event (reason not clear) Other Observations: The feature extraction is conceptually attractive but the classification based on decision rules using different feature attributes and numerical ranges and thresholds is not (too much hand-crafted or expert knowledge required) A more sophisticated machine learning classification on SVMs or RFCs should provide superior performance.
DISCUSSION AND CONCLUSION PQD is likely to become a more significant research area with more complex grids and microgrids which will increase the number and type of faults and undesirable signal conditions that can occur. Current research in the area has only recently considered multiple fault conditions Evaluations presented only measure accuracy and do not evaluate any aspect of detection of fault onsets and offsets via precision and recall as measured by the F1 score. Only additive noise disturbances have been considered but are these realistic for power signals? Are there other types of disturbances that are more realistic to do with measurement distortion or uncertainty? CAN WE LEARN FROM AUDIO SIGNAL PROCESSING? Audio signal processing considers other problems which may be of interest to PQ: signal enhancement, signal source separation, true signal detection and more sophisticated signal classification paradigms The emerging area of Acoustic Event Detection (AED) can offer baseline systems and protocols for PQD detection and classification The multi-resolution CQT has proved popular for AED as the log-spaced spectral features better capture harmonics; in ASR the so-called cepstral features are uncorrelated and better separate long-term channel effects; and the use of feature normalisation and dynamic features improve robustness. Are these worth considering for PQD especially for more complex fault conditions? THE PQD CHALLENGE AED has DCASE how about creating our own challenge for PQD using synthetic and real data? See how the audio people do this: http://www.cs.tut.fi/sgn/arg/dcase2016/
REFERENCES [Survey Paper 1] Khokhar, Suhail, et al. "A comprehensive overview on signal processing and artificial intelligence techniques applications in classification of power quality disturbances." Renewable and Sustainable Energy Reviews 51 (2015): 1650-1663. [Survey Paper 2] Mahela, Om Prakash, Abdul Gafoor Shaik, and Neeraj Gupta. "A critical review of detection and classification of power quality events." Renewable and Sustainable Energy Reviews 41 (2015): 495-505. [Jan 2014 IEEE TSE] Ray, Prakash K., et al. "Optimal feature and decision tree-based classification of power quality disturbances in distributed generation systems." IEEE Transactions on Sustainable Energy 5.1 (2014): 200-208. [May 2014 IEEE TIE] Valtierra-Rodriguez, Martin, et al. "Detection and classification of single and combined power quality disturbances using neural networks." IEEE Transactions on Industrial Electronics 61.5 (2014): 2473-2482. [Jan 2015 IEEE TIE] Manikandan, M. Sabarimalai, S. R. Samantaray, and Innocent Kamwa. "Detection and classification of power quality disturbances using sparse signal decomposition on hybrid dictionaries." IEEE Transactions on Instrumentation and Measurement 64.1 (2015): 27-38.