Grade transition optimization by using gated recurrent unit neural network for styrene-acrylonitrile copolymer process

Abstract

The melt index (MI) of polymer products is an important quality reference for the product properties. However, MI cannot be measured in real-time, and the current value of MI can only be obtained by laboratory analysis after several hours, which leads to unsatisfactory quality control results. To solve the problem, this paper adopts the styrene-acrylonitrile (SAN) copolymers process as a target process and uses the Gated Recurrent Unit (GRU) to establish a MI dynamic prediction model for different grades of SAN copolymer to estimate the current and future MI values, which ultimately improve the MI quality control performance. In addition, to solve the quality fluctuation caused by the difficulty of fine tune the chain modifier feed flow during the grade transition. Therefore, this paper also combines the GRU dynamic model and a virtual controller to provide recommended operating values for the chain modifier to reduce the transient time during grade transition. The simulation results in this paper show that the predicted value of MI is in agreement with the actual measured value. In addition, the recommended value of the chain modifier feed flow rate in comparison to actual manual control can significantly reduce about 28.6 hours of the grade transition time.