A novel model based on wavelet LS-SVM integrated improved PSO algorithm for forecasting of dissolved gas contents in power transformers

• Published in: Electric power systems research Vol. 155; pp. 196 - 205
• Main Authors: Zheng, Hanbo, Zhang, Yiyi, Liu, Jiefeng, Wei, Hua, Zhao, Junhui, Liao, Ruijin
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.02.2018; Elsevier Science Ltd
• Subjects: Algorithms; Artificial neural networks; Back propagation; Back propagation networks; Basis functions; Correlation coefficients; Dissolved gases; Electric power; Electric utilities; Error analysis; Forecasting; Least squares support vector machine; Natural gas; Neural networks; Oil-immersed power transformers; Optimization algorithms; Particle swarm optimization; Radial basis function; Regression analysis; Studies; Support vector machines; Transformers; Wavelet analysis; Wavelet technique; Dissolved gases; Wavelet technique; Oil-immersed power transformers; Least squares support vector machine; Forecasting; Particle swarm optimization
• ISSN: 0378-7796; 1873-2046
• DOI: 10.1016/j.epsr.2017.10.010

•An approach combing wavelet technique with least squares support vector machine is proposed.•A mutation operation with certain probability is applied to the traditional particle swarm optimization algorithm.•The integrated approach is capable of forecasting the dissolved gas contents more accurately. Finding out the transformer incipient faults and their development trend has always been a central issue for electric power companies. In this paper, a novel approach combing wavelet technique with least squares support vector machine (LS-SVM) for forecasting of dissolved gases in oil-immersed power transformers has been proposed. The algorithm of particle swarm optimization (PSO) with mutation is developed to optimize the parameters of constructed wavelet LS-SVM regression (W-LSSVR). The existence of admissible wavelet kernels is proven by theoretic analysis. Evaluation of forecasting performance is based upon the measures of mean absolute percentage error (MAPE) and squared correlation coefficient (r2). On the basis of the proposed approach, a procedure is put forward to serve as an effective tool and experimental results show that this approach is capable of forecasting the dissolved gas contents accurately. Comparing with the back propagation neural network (BPNN), the radial basis function neural network (RBFNN), the generalized regression neural network (GRNN), and the SVM regression (SVR) in two practical cases (taken hydrogen as an example here), the MAPEs of the proposed approach are significantly better than that of the four methods (5.4238% vs 19.1458%, 11.7361%, 7.7395%, 8.3248%; 2.1567% vs 18.9453%, 10.2451%, 7.8636%, 2.4628%) respectively.