Abstract
This paper proposes an AIPC-based controller integrating a PINN surrogate with monotonicity constraints, a TD3 agent, and a steady-state detection module, along with an operator experience integration strategy (AIPC-OE) that activates rule-based corrective actions near constraint boundaries. Three reward functions are evaluated to guide reward design. Validated on a rigorous dynamic model of an industrial pre-distillation column, the controller reduces side-stream FBP overshoot by 89% vs. PID and achieves 44% faster settling than PID in the baseline scenario. Under crude property variation, AIPC-OE further reduces side-stream overshoot by 44% compared to standalone AIPC. Under simultaneous variation of crude properties and throughput, both AIPC and AIPC-OE substantially suppress side-stream FBP overshoot relative to MPC (2.0°C and 2.3°C vs. 7.3°C). To our knowledge, this is the first validation of a PINN-augmented DRL controller on a rigorous first-principles dynamic model of an industrial distillation column, demonstrating its potential for practical deployment in complex chemical processes.

This publication has 12 references indexed in Scilit: