Control of melt index in an industrial ethylene-vinyl acetate process using a recurrent neural network soft sensor and multiple virtual control strategies

This study concerns the development of a gated recurrent units (GRU) artificial neural network (ANN) model and three virtual controllers for controlling melt index (MI) in an industrial ethylene-vinyl acetate (EVA) resin production process. Candidate input variables (features) were selected using the eXtreme Gradient Boosting (XGBoost) method and also operator experience and engineering knowledge. Bayesian optimization was applied to determine the optimal values of hyperparameters. Model performance was quantified using the mean absolute percentage error (MAPE). Step tests were performed to ensure process gain consistency. The predictive model was used to create virtual controllers using three control architectures: virtual PID, fuzzy, and model predictive control. Results show that the model can accurately predict melt index for most EVA grades. Furthermore, all virtual control systems can control the melt index to the setpoint for most grades, completing grade changeover faster and with less off-spec production than manual control.