An improved feed-forward neural network based on UKF and strong tracking filtering to establish energy consumption model for aluminum electrolysis process

作者：Yao, Lizhong; Li, Taifu; Li, Yanyan^*; Long, Wei; Yi, Jun

来源：Neural Computing & Applications, 2019, 31(8): 4271-4285.

DOI：10.1007/s00521-018-3357-9

• Feed-forward neural network
• Strong tracking filtering
• Unscented Kalman filtering
• Energy consumption model
• Aluminum electrolysis

The paper presents a modeling method about the energy consumption of aluminum electrolysis process based on a new neural network. The proposed neural network (NN) is built by combining two theories of unscented Kalman filtering (UKF) and strong tracking filtering (STF), which is shortened as STUKFNN in this study. Moreover, the new training algorithm and robustness analysis of the STUKFNN are presented. The final section of the paper shows an illustrative example regarding the application of the new training algorithm to estimate the technical energy consumption of the aluminum electrolysis process, compared with the modeling methods of back-propagation neural network (BPNN), extended Kalman filtering neural network (EKFNN) and unscented Kalman filtering neural network (UKFNN). The analysis and results show that the method improves the real-time tracking ability of dynamic interference in aluminum electrolysis process, and the accuracy of STUKFNN is better than the other three modeling methods. The average indicators MAE, MSE, R of the STUKFNN based on 30 runs are 15.4793, 1862.65 and 0.9966, respectively, which are all superior to other methods. The proposed method also shows better performance compared with UKFNN, EKFNN and BPNN by the proportion of relative error (RE) in the interval vertical bar RE vertical bar <0.1% based on all samples, 76, 38, 22 and 74%, respectively.

• 出版日期2019-8
• 单位重庆科技学院; 四川大学