基于FNN-UKF神经网络的氧化铝浓度动态预测模型

Dynamic Prediction Model Based on FNN-UKF Neural Networks for Alumina Concentration

针对影响氧化铝浓度的因素较多,彼此相关性强,难以建立动态预测模型的问题,提出一种基于FNN-UKF神经网络的动态预测方法。首先考察各原始变量置零前后在特征空间主元投影上的相似度,判断其对氧化铝浓度的解释能力,由此进行原始变量选择;再将约简后的原始变量输入基于UKF算法的神经网络预测模型,通过更新状态估计值和方差矩阵提高模型的泛化能力。对取自某厂160KA大型预焙槽的247组样本数据进行检验:228组样本的预测误差在±1%之内,计算量减少52.07%,表明该方法在保证预测精度的同时,有效降低了模型学习的计算量。

Based on false nearest neighbors and unscented kalman filter （FNN-UKF）, a dynamic prediction method for alumina concentration was proposed. In the new KPCA feature subspace, it was inspired by FNN that interpretation of alumina coneentration would be estimated by calculating the variables mapping distance in the KPCA space to select secondary variables. Selected variables were introduced into BP neural networks as input vector. UKF algorithm, in which estimated value and variance matrix of state were updated to improve the generalization capability of the networks, was used to train weight values and threshold values. By using 247 samples of 160KA operating aluminum cell from a factory, experimental results demonstrated that the forecast error of 228 samples was ± 1%, the computation was decreased to 52.07%. The method in which the computation time was reduced effectively can surely accuracy of parameter estimation.