A flexible multi-step prediction architecture for process variable monitoring in chemical intelligent manufacturing

Multi-step prediction models can forecast variables ahead of time, which is valuable for process variable monitoring. Although deep learning (DL) is promising in multi-step prediction, its compatibility, interpretability and practicality, which are crucial for chemical applications, have received little attention. Thus, a DL architecture—Light Attention-Mixed Base Target Autoregression Unit (LAMB-TAU) is proposed. It utilizes special-designed networks to simulate process driving forces, and wraps these networks with a decoder, delivering an interpretable and high-accuracy multi-step prediction on process variables. Moreover, an adaptable sampling procedure is proposed to free the multi-step predictions on difficult-to-measure variables from high-cost experiments. The effectiveness of LAMB-TAU is verified by modeling studies on chemical processes including esterification and formaldehyde production. Besides, model practicalities including extended multi-step prediction, uncertainty and computational cost are explored. The proposed LAMB-TAU is instructive for DL multi-step prediction studies toward chemical process monitoring, which promotes the development of intelligent chemical industry.