Model Predictive Control of Grade Transition with Attention Base Sequence-to-Sequence Model

In industry, process input-output data exhibit complex nonlinear dynamics. Such behavior must be modeled by nonlinear time series for use in model-based control, optimization, and monitoring. In this work, a sequence-to-sequence (StS) model was developed for the ASPEN Polymer Plus simulator of an industrial high-density polyethylene (HDPE) slurry reactor. Inclusion of attention mechanism and elastic net (EN) training was found to substantially improve the gain consistency and time dynamics of the model. The resulting model was utilized as a non-linear model predictive control (NLMPC) to control the hydrogen to ethylene ratio (HER) and pressure. The NLMPC can navigate the grade transition of the reactor as well as maintaining the steady state.