遗传算法的聚丙烯熔融指数最优预报  

The optimal prediction of melt index of polypropylene based on genetic algorithm

作　　者：张志猛[1,2] 楼巍[1] 刘兴高[1] 

机构地区：[1]浙江大学工业控制技术国家重点实验室,浙江杭州310027 [2]浙江大学软件学院,浙江杭州310027

出　　处：《计算机与应用化学》2011年第7期843-846,共4页Computers and Applied Chemistry

基　　金：国家自然科学基金(50876093);浙江省杰出青年科学基金(R4100133);浙江省科技厅国际合作(2009C3408);国家863计划(2006AA05Z226)

摘　　要：熔融指数只能通过人工取样、离线化验分析获得,时间滞后大,难以满足实时控制的要求,其软测量预报意义很大。本文提出了一种基于遗传算法的聚丙烯熔融指数最优预报方法。径向基函数(RBF)神经网络用来拟合输入与输出之间的非线性关系,用遗传算法对RBF神经网络权值进行优化。基于某石化企业聚丙烯生产过程采集的历史数据进行研究。根据反应机理以及流程工艺分析,将温度、压力、液位、氢气气相百分数、3股丙烯进料流率、2股催化剂进料流率等9个过程变量作为模型的输入。经过优化的预报模型的均方根误差从优化前的0.0107减小为0.0057;平均绝对误差从原来的0.0392减小为0.0220;平均相对误差从优化前的1.49%减小为0.94%,表明优化后模型精度大大提高。优化后的标准差从优化前的0.0496减小为0.0271,表明优化后模型具有更好的预报稳定性。希尔不等系数从优化前的0.0096减小为优化后的0.0051,表明优化后模型与实际过程具有更好的一致性。优化后的RBF神经网络模型预测结果的各项指标都明显优于优化前的RBF模型,表明遗传算法提高了原模型的预报准确性、稳定性和可靠性。Melt index （MI） is considered an important indicator of polypropylene product quality, which is usually evaluated off-line with an analytical procedure that takes almost two hours to complete in the laboratory, leaving the process without any real-time quality indicator during this period. An alternative is to develop on-line soft sensor of MI. An optimal melt index prediction modeling based on genetic algorithm is proposed. Radical basis function （RBF） neural network is used to characterize the nonlinearity between input data and output data. The genetic algorithm （GA） is employed to optimize the weights of the RBF neural network. A detailed comparison research among the standard RBF and optimized RBF is carried out based on the data from one real plant of petrochemical company. According to the analysis of reaction mechanism, nine process variables, such as the temperature, pressure, liquid level, percentage of hydrogen gas, flow rate of three propylene feeds and two catalyst feeds, are chosen as the input variables. The root mean square error of optimized RBF model is reduced from 0.0107 without optimization to 0.0057, the mean absolute error is reduced from 0.0392 to 0.0220, the mean relative error is reduced from 1.49% to 0.94%. The standard error is reduced from 0.0496 to 0.0271. The Theil＇s inequality coefficient is reduced from 0.0096 to 0.0051. These results show that the proposed prediction model with GA algorithm established a good prediction of the polypropylene process.

关 键 词：遗传算法 径向基函数神经网络:熔融指数 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置] TQ018[自动化与计算机技术—控制科学与工程]