A novel bidirectional DiPLS based LSTM algorithm and its application in industrial process time series prediction

Abstract

The development of industry has led to the big data revolution in manufacturing industry, which has greatly improved production quality and efficiency. A accuracy prediction model is established by using deep learning method, especially short and long term memory (LSTM). However, the application field of LSTM is usually very limited, and the interpretability of LSTM is weak. In this paper, a novel bidirectional DiPLS based LSTM algorithm (BiDiPLS-LSTM) is proposed to solve this problem. First, DiPLS is used to process the forward time series data to obtain the most predictable dynamic latent variable (DLV) of the target variable. At the same time, DiPLS method is also used to process the reverse time series data and extract the reverse DLV. Then, the forward and reverse dynamic features of the same location will form new input data, and this new input data will be input into the LSTM network to realize the prediction of time series data. In order to verify the effectiveness of the proposed BiDiPLS-LSTM method, the real 660 MW coal-fired boiler process data is used, and the emission of oxides of nitrogen (NOx) is regarded as the output. The results show that the extracted dynamic features can be used to build a very accuracy model, and the results outperform the traditional LSTM method, the PLS-LSTM method, the unidirectional DiPLS-LSTM method, and traditional RNN method in the comparison. For the first time, we used the bidirectional dynamic features of the input data as the real input to predict the output, which will obtain a good prediction effect. What is more, the use of dynamic latent variables made the result more explanatory.