Journal of Shanghai Jiao Tong University ›› 2020, Vol. 54 ›› Issue (4): 376-386.doi: 10.16183/j.cnki.jsjtu.2020.04.006

Previous Articles     Next Articles

Hybridizing Multivariate Empirical Mode Decomposition and Extreme Learning Machine to Predict Non-Stationary Processes

LI Chunxiang,ZHANG Haoyi   

  1. Department of Civil Engineering, Shanghai University, Shanghai 200444, China
  • Online:2020-04-28 Published:2020-04-30

Abstract: The lack of sensor layout and the deficiency of sensor data are the key problems in the study of wind pressure measurement. The spatial prediction of wind pressure can restore the missing data and expand the air pressure information, and help to establish the wind pressure distribution on the structure surface. In this paper, a spatial prediction algorithm based on multivariate empirical mode decomposition (MEMD) and extreme learning machine (ELM) is proposed. The MEMD method is used to decompose the multi-channel non-stationary signals. The intrinsic mode functions and the residue are obtained with the same number and similar frequency. The decomposed data are restructured by frequency and ELM is used to train the restructured data and make predictions. The effectiveness and accuracy of the algorithm are verified by simulated data from autoregressive moving average (ARMA) model and observed wind pressure data. At the same time, the least squares support vector machine using radial basis kernel function (RBF-LSSVM) and the basic ELM method are introduced as the comparison. It is proved that the prediction error of the MEMD-ELM method is smaller and the correlation with the real value is higher. The modes decomposed through MEMD preserve the correlation from origin data, so a more accurate result is obtained, which proves that MEMD-ELM is an effective multi-variety prediction method.

Key words: multivariate empirical mode decomposition (MEMD), extreme learning machines (ELM), non-stationary, forecasting

CLC Number: