Study Of Multiple Power Quality Disturbance Classification Methods Based On Extreme Learning Machine

Power quality disturbance classification is the basis of power quality monitoring and analysis.The main objective of power quality disturbance classification(PQDC)is accurately recognizing the type of disturbance in order to provide a reference for power system parameter estimation and control strategies.Therefore,PQDC is an important research problem.Power quality events are often associated with multiple disturbances and their classification is a typical multi-label learning problem.Designing efficient and accurate methods for recognizing multiple power quality disturbances(MPQDs)is of great importance.However,the traditional multi-label learning algorithms employed for PQDC have slow training speeds and they require the adjustment of complex structural parameters,especially those based on back-propagation neural networks and support vector machines.Extreme learning machine(ELM)is a single hidden layer,feedforward neural network learning algorithm and it has become very popular in recent years.ELM only requires the initialization of the number of hidden layer nodes and there is no need to repeatedly adjust the input weight of the network and the bias of the hidden layer.ELM can generate a unique optimal solution.ELM has attracted much attention from researchers because of their learning speed and good generalization performance.This study was supported by two grants from the National Natural Science Foundation of China.The multi-label learning characteristics were analyzed with ELM,especially by exploiting the label dependency.Various investigations were conducted in terms of the algorithm design and MPQD recognition application due to the difficulties caused by overlapping MPQD features,the redundant or incomplete features obtained by artificial selection,and the inadequate number of marked MPQD samples.In this dissertation,we present three improved ELM algorithms with a good capacity for generalization and classification performance.We compared the effectiveness of our proposed algorithms with several state-of-the-art algorithms,where they were verified based on a simulation signal dataset and hardware dataset.The main contributions of this dissertation are summarized as follows.(1)We analyzed the MPQD classification as a multi-label learning problem and proposed a new method based on variational mode decomposition(VMD)and a random discriminative projection ELM for multi-label learning(RDPEML).First,VMD was employed to extract the features of MPQDs to form the input vectors for the classifier.Second,a novel multi-label classifier called RDPEML was constructed for multi-label learning.By combining a random discriminative projection multi-class ELM with a thresholding learning method-based radial basis function kernel ELM,we extended the traditional multi-class ELM to multi-label learning applications.Finally,a test study was conducted using Matlab synthetic signals and real signals sampled from a three-phase standard source under different noise conditions.RDPEML achieved better classification performance with superior computational speed compared with several recent state-of-the-art multi-label learning algorithms.(2)For the first time,the multi-label active learning method was applied in PQDC to solve the problem of inadequately labelled MPQDs and we obtained a large number of unlabeled MPQD data from a power grid detection device.We proposed a new active learning strategy for working on MPQDs.By exploiting the label exclusiveness of MPQDs,a new uncertainty sampling strategy based on label exclusiveness and ranking scores was defined to select the most informative example.The labelling cost of unknown MPQDs was reduced by combining the new sampling strategy and the base classifier RDPEML.(3)A novel deep ELM algorithm called deep Hessian ELM for multi-label learning(DHELMML)was developed to reduce the redundancy and incompleteness of the artificial selection features.DHELM-ML stacks an ELM-based denoising auto-encoder(ELM-DAE)series to create a multi-layer neural network.The adoption of ELM-DAE as the basic module accelerates the training process for the multi-layer neural network,which is the limitation of the traditional deep learning algorithms due to the back-propagation tuning method.Furthermore,Hessian regularization is introduced into the ELMDAE model to explore the intrinsic geometric structure of the data and to enhance the classification performance.RDPEML replaces the multi-class ELM as the supervised model in the highest layer of the deep neural network.Thus,the original hierarchical ELM algorithm is extended to multi-label learning.In conclusion,a new multi-label learning algorithm based on ELM was developed in this study to help address the low efficiency of the state-of-the-art multi-label learning algorithms employed for MPQD recognition.Due to the lack of adequate training data to match disturbances with their corresponding types and the difficulty of artificial feature extraction of some multiple disturbances,we proposed an innovative ELM-based active learning algorithm and an improved multi-label deep ELM algorithm.These algorithms have both academic and practical value.