论文部分内容阅读
针对具有强非线性、复杂的化工过程软测量建模,提出一种基于平方根容积卡尔曼滤波(SCKF)的递归神经网络方法.基于Elman递归神经网络,首先构建状态空间模型,然后应用SCKF算法进行训练,所有网络的权值将作为系统的状态进行更新.容积卡尔曼滤波(CKF)通过三阶Spherical-Radial容积准则生成容积点,利用容积点逼近状态的后验分布,使得高维非线性滤波中的多变量积分数值求解成为可能.在CKF的基础上,SCKF采用预测及后验误差协方差矩阵的平方根因子进行递推运算,进一步改进了算法的数值稳定性.将该方法应用于脱丁烷塔底部丁烷组分含量以及硫回收装置尾气中SO2和H2S含量的软测量动态建模实例中,在同等条件下,还与基于EKF、SCKF的前馈神经网络,基于EKF的递归神经网络等其它方法对比.结果表明,本文的方法能够获得很好的建模精度,显示出其有效性.
Aiming at the soft nonlinear modeling process, a new recurrent neural network based on square root volume Kalman filter (SCKF) is proposed.Based on the Elman recurrent neural network, the state space model is constructed firstly and then the SCKF algorithm is used Training, the weights of all networks will be updated as the state of the system Volumetric Kalman Filter (CKF) Volumetric points are generated by the third-order Spherical-Radial volumetric criterion, and the posterior distribution of states is approximated by the volumetric point, Based on CKF, SCKF uses the square root factor of prediction and a posteriori error covariance matrix to recursively calculate, which further improves the numerical stability of the algorithm.The method is applied to the debutanizer In the case of the soft-sensing dynamic modeling of the content of butane and the content of SO2 and H2S in the tail gas of the sulfur recovery unit, under the same conditions, it is also compared with other feedforward neural networks based on EKF and SCKF, recurrent neural networks based on EKF, The results show that the proposed method can obtain good modeling accuracy and show its effectiveness.