基于平方根无迹卡尔曼神经网络的铝电解工耗模型

Model of aluminum electrolysis based on square root unscented Kalman neural network

铝电解过程中传统机理建模及静态建模等方法难以建立精确过程模型。采用平方根无迹卡尔曼神经网络算法构建铝电解工耗模型。无迹卡尔曼神经网络滤波与平方根滤波理论相结合,改进无迹卡尔曼神经网络滤波算法,利用协方差矩阵的平方根代替无迹卡尔曼算法中的协方差矩阵参与递推运算,解决铝电解建模过程中出现误差协方差矩阵非正定型而导致滤波发散的问题,并且提高了模型的自适应能力和精确度。通过对某铝厂出铝情况的日报样本进行验证,对比神经网络模型和无迹卡尔曼神经网络模型,平方根无迹卡尔曼神经网络提高了铝电解工耗模型精度和可靠性,表明了该方法的有效性、先进性和可靠性。

In the process of aluminum electrolysis,it is difficult to establish the precise process model,such as the traditional mechanism modeling and the static modeling.In this paper,the use of the square root of the unscented Kalman neural network algorithm to build a model of aluminum electrolysis.Combination of the unscented Kalman neural network filter and square root filtering theory,the improved Kalman neural network filtering algorithm,instead of covariance matrix of Kalman algorithm in the recursion with the square root of the covariance matrix,solve the non-positive type and cause filtering divergence problem of error covariance matrix of aluminum electrolysis in the process of modeling,and to improve the model accuracy and adaptive ability.Verified by an aluminum on the daily samples,compared with neural network model and unscented Kalman neural network model,the square root unscented Kalman neural network improves the aluminum electrolytic power consumption model precision and reliability,verify the validity of the proposed method,advanced and reliable.