Time Series Prediction With Incomplete Dataset Based on Deep Bidirectional Echo State Network

Qiang Wang,Wei Wang,Ying Liu,Jun Zhao,Linqing Wang

doi:10.1109/access.2019.2948367

Qiang Wang, Wei Wang + Show 3 more

Open Access

https://doi.org/10.1109/access.2019.2948367

Copy DOI

Abstract

In the complex industrial environment, data missing situation is often occurred in the process of data acquisition and transition. The major contribution of the paper is the proposal of a deep bidirectional echo state network (DBESN) framework for time series prediction with such incomplete dataset. Instead of data imputation methodology, a bidirectional fusion reservoir is here designed to extract the deep bidirectional feature along with forward and backward time scales, based on which a deep autoencoder echo state network (DAESN) and a deep bidirectional state echo state network (DBSESN) are constructed for the incomplete output and input samples, respectively. As for such two networks, a bidirectional echo state network (BESN) is proposed for connecting them to constitute the DBESN framework for prediction. To verify the effectiveness of the proposed method, one synthetic time series as well as two real-world industrial datasets are employed to conduct the comparative experiments. The experimental results demonstrate that the proposed method outperforms other comparative ones at various missing rates.

Highlights

With the development of modern industrial information technology, the amount of the industrial data accumulated in process of manufacturing is increasing at an unprecedented rate [1]
Aiming at time series prediction with incomplete dataset, a deep bidirectional echo state network (DBESN) framework is proposed in this study
DEEP AUTOENCODER ECHO STATE NETWORK In order to extract the deep feature of the output sample with missing points of the DBESN framework, a deep autoencoder echo state network (DAESN) model is proposed in this study, which stacks multilayer autoencoder echo state network (AESN) to form deep structure for feature extraction

Summary

INTRODUCTION

With the development of modern industrial information technology, the amount of the industrial data accumulated in process of manufacturing is increasing at an unprecedented rate [1]. Aiming at time series prediction with incomplete dataset, a deep bidirectional echo state network (DBESN) framework is proposed in this study. In order to dispose the missing points in the output and input samples of the proposed framework instead of data imputation, a deep autoencoder echo state network (DAESN) and a deep bidirectional state echo state network (DBSESN) are constructed based on a designed bidirectional fusion reservoir for extracting the deep bidirectional feature of the samples in both past and future time. B. DEEP AUTOENCODER ECHO STATE NETWORK In order to extract the deep feature of the output sample with missing points of the DBESN framework, a DAESN model is proposed in this study, which stacks multilayer autoencoder echo state network (AESN) to form deep structure for feature extraction. By multi-layer feature extraction of the DAESN optimized by the improved unsupervised wake-sleep algorithm, the feature of the topmost layer is the deep feature Ldf of the output samples, which is expressed as

DEEP BIDIRECTIONAL STATE ECHO STATE NETWORK

BIDIRECTIONAL ECHO STATE NETWORK

PROCEDURE OF THE PROPOSED PREDICTION MODEL

EXPERIMENTS AND ANALYSIS

MACKEY-GLASS TIME-SERIES DATA

CONCLUSION