文献类型：期刊文献

英文题名：A fine-tuned RNN model for accurately predicting the spatial distribution of parameters in light hydrocarbon cracking tubular reactor

作者：Tang, Shiyi[1];Duan, Zhaoyang[1];Tian, Zhou[1];Du, Wenli[1];Qian, Feng[1]

机构：[1]East China Univ Sci & Technol, Key Lab Smart Mfg Energy Chem Proc, Minist Educ, Meilong Rd 130, Shanghai 200237, Peoples R China

年份：2025

卷号：504

外文期刊名：CHEMICAL ENGINEERING JOURNAL

收录：;EI(收录号：20250117623354);WOS:【SCI-EXPANDED(收录号:WOS:001403020200001)】；

语种：英文

外文关键词：Light hydrocarbon cracking; Recurrent neural network; PSO optimization; Custom loss function

摘要：The plug flow reactor (PFR) is atypical example of a distributed parameter system. Modeling of a PFR usually requires to develop a one-dimensional reactor model with rigorous kinetics. However, the high computational cost of such models restricts their practical applications. In this work, we develop a gated recurrent unit (GRU) based-recurrent neural network (RNN) model to predict the cracking process of all light hydrocarbons along the coil length under different operating conditions. The training data are generated using Coilsim1D, a software that combines a complex single-event microkinetic model, covering a comprehensive range of light hydrocarbon feedstock compositions and reaction conditions to ensure the broad applicability of the model. Subsequently, we define a custom loss function and further optimize the weight coefficients within it using the PSO algorithm to enhance the model's prediction accuracy. The optimal model is able to predict the yield distribution of all key products accurately, as well as the temperature and pressure profiles of the process gas along the reactor length under different cracking feedstocks and operating conditions. The average MAE calculated on the original scale between the predictions and the true data is approximately 0.04wt%, 238.80Pa, and 0.44K for the yields of the key products, process gas pressure, and process gas temperature, respectively. The RNN model reduces prediction time by approximately two orders of magnitude compared to running a single cracking reaction simulation in Coilsim1D.