强跟踪平方根UKFNN的铝电解槽工耗动态演化模型

摘要

The aluminum electrolysis process has multiple characteristics including multivariate, strong coupling, strong interference and time-varying parameters. Therefore, its model development is technically difficult. According to the characteristics of the process, an improved unscented Kalman filter neural network based on strong tracking filter and square root filter (STR-UKFNN) is proposed in this paper. Then, the STR-UKFNN is used to create the dynamic evolutionary model for energy consumption of aluminum reduction cell. Firstly, the state covariance matrix of the UKFNN algorithm is replaced by its square root to participate in recursive operations; Secondly, the filter gain matrice in the algorithm of UKFNN is adjusted by introducing the time-varying fading factor and the diminishing factor. A series of experiments have been conducted by using the daily samples from the 170 kA new pre-baked cell. The experimental results show that the method improves the precision of the energy model and the real-time tracking ability for the abrupt state change of the aluminum reduction cell. So the method is helpful to guide the optimization of operating parameters in the aluminum electrolysis process.%铝电解过程具有多变量、强耦合、强干扰、参数时变等特征,故其模型开发是一个技术难点。根据该过程的特点,本文提出强跟踪平方根无迹Kalman 神经网络(Strong tracking square root unscented Kalman filter neural network, STR-UKFNN),并用其建立铝电解槽工艺能耗的动态演化模型。该方法利用误差协方差矩阵的平方根代替UKFNN 算法中的协方差阵,避免误差协方差矩阵可能出现负定而导致滤波发散,并在UKFNN 算法中引入渐消因子和弱化因子,实时调整滤波增益,提高模型收敛速度和其对突变状态的跟踪能力。通过某铝厂170 kA 预焙槽的日报样本验证表明,该方法提高了能耗模型的精度和对电解槽突变状态的实时跟踪能力,有助于指导铝电解过程操作参数的优化。