Real-time optimization of continuous catalytic reforming using transfer learning and reinforcement learning for feedstock variability

Abstract

Background

Real-time optimization (RTO) of continuous catalytic reforming (CCR) is hindered by model mismatch caused by feedstock property variations, which degrade the performance of existing optimization strategies.

Methods

To address this challenge, this study proposes an RTO method integrating transfer learning (TL) and reinforcement learning (RL). A high-precision surrogate model of the CCR process is constructed as the agent's interactive environment. A TD3 agent is designed based on optimization objectives, incorporating dropout layers during training to enhance robustness. The trained critic network serves as a monitor, detecting feedstock changes via absolute temporal difference error.

Findings

TL enables the agent to adapt efficiently to varying feedstock properties. Experimental results demonstrate that the proposed method rapidly identifies optimal operating conditions, with the monitor effectively detecting feedstock changes. Across multiple test cases, the TL strategy reduces average training time to one-tenth of training from scratch, without compromising performance, demonstrating its effectiveness and potential for RTO applications.