摘要
针对原油蒸馏过程常规软测量模型难以适应原油进料性质变化的问题,提出Bootstrap多神经网络的非线性软测量处理策略。通过Bootstrap算法复制出训练集样本空间上的多个样本子空间,训练出多神经网络模型,避免了单个神经网络易于陷入局部最优及过度训练的弱点,具有较高的准确率和泛化能力。本处理策略用于建立常压塔一线干点的软测量模型,仿真结果表明模型预测准确率和鲁棒性较好,对原油性质变化具有较好的适应性。该方法将会改进实际蒸馏过程在进料性质变化情况下的产品质量指标的软测量精度。
A nonlinear soft-sensing strategy with bootstrap aggregated neural network is proposed to solve the poor adaptability of conventional soft-sensor methods when teedstock varies in crude oil distillation. A bootstrap aggregated neural network shows better accura- cy and generalization capability than a single neural network which can be trapped in a local minimum or over-fitted the training data. The proposed strateg) is used for developing a soft-sensor for the end point of kerosene product of a simulated atmospheric tower. The simulation results show that the bootstrap aggregated neural network soft-sensor possesses high predictive accuracy and robustness and the proposed soft-sensor gives good performance even under severe feedstock variations. It is helpful to improve the soft-sensor precision of product quality index when feedstock varies in crude oil distillation.
出处
《控制工程》
CSCD
北大核心
2009年第4期475-477,506,共4页
Control Engineering of China
基金
国家863高技术研究计划基金资助项目(2007AA04Z193
2006AA04Z168)