Industrial Soft Sensor Prediction based on Multi-model Integrated Method

Abstract

The industrial processes are commonly characterized by nonlinearities and dynamics. Therefore, long short-term memory (LSTM) networks are often adopted to extract the nonlinear dynamic features for the prediction of industrial quality indicators. However, traditional LSTM only captures the temporal characteristics of input variables but ignores the output variables. Therefore, a multi-model integrated method (MMIM) is proposed for simultaneously extracting the input and output temporal characteristics in this study. In the MMIM, a LSTM and other static models are used to collect the temporal and static characteristics for the inputs, while a RNN is applied to predict the output variable. The effectiveness and performance are verified on an industrial hydrocracking plant for the prediction of light naphtha isopentane and heavy naphtha quality.