Abstract:

Rectisol process is a relatively mature purification process in China's coal chemical production, and its modeling research is conducive to the intelligent construction and development of the plant. In this paper, based on the need of intelligent construction of rectisol plant in a petrochemical enterprise, real-time data preprocessing rules were established, and the CO₂ content prediction model of purified gas was built based on the Bayesian optimization (BO) long short-term memory (LSTM) neural network. The results showed that after preprocessing the real-time data of the device by manual screening and correction and the maximum information coefficient (MIC) method, the number of variables of the collected real-time data could be reduced from 84 to 22, which reduced the redundancy of the data. The six hyper-parameters of the LSTM model were tuned by using the BO, and the BO-LSTM prediction model was established based on the optimized hyper-parameter combinations. The evaluation indexes of RMSE, MAE and R² of the prediction model were 0.0395, 0.0275 and 0.9843, respectively, which were higher in precision and better in regression effect than the traditional BP and LSTM model, proving the feasibility and accuracy of the model in the prediction of the rectisol purified gas composition, and it could guide the optimization of the production of purified gas products. The product optimization modeling method based on the CO₂ content prediction model of rectisol purified gas could provide ideas for the digital and intelligent construction of related processes.

Key words: rectisol, Bayesian optimization, neural network, computer simulation, prediction

摘要：

低温甲醇洗工艺作为我国煤化工生产中较为成熟的净化工艺，对其进行建模研究有利于工厂智能化建设与发展。本文基于某石化企业低温甲醇洗装置智能化建设的需要，建立了实时数据预处理规则，并基于贝叶斯优化（BO）的长短时记忆（LSTM）神经网络建立净化气CO₂含量预测模型。结果表明，通过人工筛选及校正和最大信息系数（MIC）方法对装置的实时数据预处理后，可将采集的实时数据变量数由84个降低到22个，降低了数据冗余度；运用BO算法对LSTM模型的6个超参数进行调优，根据优化后的超参数组合建立BO-LSTM预测模型，得到BO-LSTM预测模型的评价指标均方根误差（RMSE）、平均绝对误差（MAE）、回归系数（R²）分别为0.0395、0.0275、0.9843，相比于传统的反向传播（BP）与LSTM模型精度更高、回归效果更好，证明了该模型在低温甲醇洗净化气组成预测运用中的可行性与准确性，能够为净化气产品的生产优化做指导。基于低温甲醇洗净化气CO₂含量预测模型的产品优化建模方法可为相关工艺数字化和智能化建设提供思路。

关键词: 低温甲醇洗, 贝叶斯优化, 神经网络, 计算机模拟, 预测