摘要
针对芳烃异构化过程(AH IP)中影响对二甲苯(PX)产率的因素众多且复杂等特点,提出一种自适应神经网络(Adaptive ANN)以建立AH IP的各因素与PX产率的关联模型。Adaptive ANN将样本分成训练样本和校验样本,并设计过拟合判据参数。通过训练样本对网络进行训练,训练过程中以模型对校验样本的预测性能为指标,通过过拟合判据参数的计算自适应地在获得具有最佳预测性能模型时终止网络训练,克服了传统的神经网络以模型的拟合精度为指标,造成训练时间过长和过拟合等缺点。
According to the character of aromatic hydrocarbon isomerization process （AHIP）that there are many factors influencing the productive ratio of p-xylene （PX）in AHIP, a novel adaptive artificial neural network （ANN）is proposed to model the AHIP. The adaptive ANN divides sample into training sample and testing sample, and an over fitting criterion of model is proposed for the adaptive ANN. When the training sample is employed to train the ANN, the testing sample is applied to surveil the predict ability of ANN during the whole training process. According to the cal- culated result of the over fitting criterion, the adaptive ANN is able to end ANN training immediately after the optimal predict ability model is obtained, and hence to overcome the flaws of the over fitting of traditional ANN model and time-consuming training process. The predict precision of AHIP model obtained by the adaptive ANN is higher than the traditional ANN.
出处
《化工自动化及仪表》
EI
CAS
2006年第5期6-8,共3页
Control and Instruments in Chemical Industry
基金
国家自然科学基金项目(20506003)
上海启明星项目(04QMX1433)
国家"973"计划项目(2002CB312200)
关键词
反传算法
自适应神经网络
芳烃异构化
过拟合判据
建模
back propagation
adaptive artificial neural network
aromatic hydrocarbon isomerization
over fitting criterion
modeling
