GA辅助BP神经网络预测催化裂化装置汽油产率  

GA-ANN METHOD FOR PREDICTION OF GASOLINE YIELD OF RFCCU

作　　者：张忠洋 李泽钦[2] 李宇龙[3] 李国庆[2] 

机构地区：[1]中国石油抚顺石化分公司,辽宁抚顺113008 [2]华南理工大学强化传热与过程节能教育部重点实验室 [3]中国石油规划总院

出　　处：《石油炼制与化工》2014年第7期91-96,共6页Petroleum Processing and Petrochemicals

摘　　要：催化裂化反应-再生系统是一个高度非线性和强耦合的操作系统,用传统建模方法很难描述。鉴于人工神经网络(ANN)非线性预测和自学习自适应能力强,而遗传算法(GA)全局寻优能力强的特点,将两者结合,先通过GA寻得BP神经网络最优的权值和阈值初值,再赋予BP,从而改善BP模型随机不确定选择初值的方法,提高其映射精度。以某炼油厂2.8Mt/a MIP装置反应-再生系统为研究对象,选取第一反应区温度、第二反应区温度、第一再生器温度、第二再生器温度、反应器压力、再生器压力等6个变量为神经网络的输入变量,汽油产率为输出变量,建立6-11-1的BP神经网络,并采用GA来对BP神经网络的权值和阈值进行优化。结果表明,未经GA优化时BP神经网络对催化裂化汽油产率的预测数据的均方误差为5.16,而经GA优化后预测数据的均方误差为4.92。The system of reaction and generation unit of RFCCU is a highly non-linear and strong coupled operation system and is too hard to be described by traditional model. The combination of the artificial neural network （ANN） with strong nonlinear prediction and self-learning ability and the genetic algorithm （GA） with global optimization ability provides a promising way to solve the problem. The optimal initial weights and threshold value are calculated by GA for the BP neural network firstly and feeded back to BP model to improve the method for random uncertain choice of initial value and the map- ping accuracy. In a practical application of this method for a 2.8 Mt/a MIP unit, a 6-11-1 type of BP neural network where the GA is used to optimize the weights and values of the BP network was established using the temperatures of two reaction zones and two regenerators along with the pressures of the reactor and regenerator as six input variables to predict the output variable gasoline yield. The results show that the predictive gasoline yield by BP neural network without GA has the mean squared error （MSE） of 5.16 while the one with GA optimization has the MSE of 4. 92.

关 键 词：催化裂化 反应-再生系统 神经网络 遗传算法 产率 

分 类 号：TE624.41[石油与天然气工程—油气加工工程]