Process performance prediction based on spatial and temporal feature extraction through bidirectional LSTM

Abstract

Abstract With the development of deep learning, it has been a trend to build data driven soft sensors in process industries with neural networks. There are a number of networks proposed to deal with time series prediction, such as Recurrent Neural Network (RNN) and Long Short-Term Memory Network (LSTM). However, it is a critical part to extract nonlinear and dynamic characteristics hiding in process data collected from industrial production. This paper proposes a novel approach for performance prediction based on the spatial and temporal feature extraction through bidirectional LSTM networks (BiLSTM) for a reactor network. Due to the superiority of processing sequences from both directions, BiLSTM are utilized to simulate the physical structure of the reactor network. With both spatial and temporal feature extraction, the deep learning model through BiLSTM achieves nice prediction performance.