An optimized twin support vector regression algorithm enhanced by ensemble empirical mode decomposition and gated recurrent unit

Despite the rapid development of support vector regression (SVR), it costs unacceptable training time in large-scale datasets and is hard to fit complex, high frequency oscillating, and non-stationary time series data. SVRs are still perplexed by the selection of critical parameters and hidden noise...

Full description

Saved in:

Bibliographic Details
Published in:	Information sciences Vol. 598; pp. 101 - 125
Main Authors:	Ding, Shifei, Zhang, Zichen, Guo, Lili, Sun, Yuting
Format:	Journal Article
Language:	English
Published:	Elsevier Inc 01.06.2022
Subjects:	Cloud theory Ensemble empirical mode decomposition (EEMD) Gated recurrent unit (GRU) Salp swarm algorithm (SSA) Twin support vector regression (TWSVR) Cloud theory Gated recurrent unit (GRU) Ensemble empirical mode decomposition (EEMD) Twin support vector regression (TWSVR) Salp swarm algorithm (SSA)
ISSN:	0020-0255, 1872-6291
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	Despite the rapid development of support vector regression (SVR), it costs unacceptable training time in large-scale datasets and is hard to fit complex, high frequency oscillating, and non-stationary time series data. SVRs are still perplexed by the selection of critical parameters and hidden noise in input data. This work proposes a hybrid model to overcome these issues that need to be resolved, namely EEMD-GRU-TWSVRCSSA. The proposed model utilizes twin support vector regression (TWSVR) to overcome the shortcomings of the SVR in terms of training time and fitting accuracy. A novel meta-heuristic algorithm, cloud salp swarm algorithm (CSSA), is employed to automatically select the optimal hyper parameters for the TWSVR. The ensemble empirical mode decomposition (EEMD) reduces the influences of hidden noise in the input data, meanwhile splitting the high-frequency and low-frequency sub-datasets and feeding them to the gated recurrent unit (GRU) and TWSVR-based model, respectively. The forecasting of the proposed algorithm and other alternative algorithms are conducted on three real-world electric load datasets from the National Electricity Market (NEM), Queensland and New South Wales regions, Australia, and the well-known National Grid UK. Experimental results demonstrate the superiority and competitiveness of the proposed algorithm.
ISSN:	0020-0255 1872-6291
DOI:	10.1016/j.ins.2022.03.060