Model-free adjustment of reducing agent for SCR device under label deficiency: Regulation-oriented stage-wise reward deep Q-learning with transfer-learned state

Abstract

Data-driven methods of nitrogen oxides (NOX) soft-sensing and selective catalytic reduction (SCR) operation for fluid catalytic cracking (FCC) process have two terms of issues. Firstly, labeled data might be deficient to train a prediction model due to the lack of monitoring devices. Secondly, the operational data can not be directly used to learn a reinforcement learning model. To address these issues, a latent temporal feature adaptation transfer learning and long-short reward deep q-learning network (LTFATL-LSRDQN) is proposed. It transfers the knowledge from another similar FCC process to realize the soft-sensing of NOX. Maximum mean discrepancy loss is introduced to the objective function of autoencoder (AE) to unify the probability distribution of transformed latent features. The operation of the treatment device is abstracted to a Markov decision process. A long- and short-term reward mechanism is introduced to DQN to constrain the selection of action. The effectiveness of LTFATL-LSRDQN is verified with the data from industrial FCC processes. The introduction of domain adaptation successfully aligns the latent features, and achieves higher soft-sensing accuracy than some state-of-the-art methods. Using the results from LTFATL as inputs, LSRDQN accomplishes more enduringly continual operation of SCR device on the premise of the regulations and constrained actions.