Prediction of component concentrations in sodium aluminate liquor using stochastic configuration networks

• Published in: Neural computing & applications Vol. 32; no. 17; pp. 13625 - 13638
• Main Authors: Wang, Wei, Wang, Dianhui
• Format: Journal Article
• Language: English
• Published: London Springer London 01.09.2020; Springer Nature B.V
• Subjects: Aluminum oxide; Artificial Intelligence; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Configurations; Data Mining and Knowledge Discovery; Image Processing and Computer Vision; Liquor; Model accuracy; Neural networks; Optimization; Original Article; Probability and Statistics in Computer Science; Regression analysis; Sodium; Sodium aluminate; Sodium carbonate; Titration; Stochastic configuration networks; Industrial data modelling; Sodium aluminate liquor; Component concentrations
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-020-04771-4

Online measuring of component concentrations in sodium aluminate liquor is essential and important to Bayer alumina production process. They are the basis of closed-loop control and optimization and affect the final product quality. There are three main components in sodium aluminate liquor, termed caustic hydroxide, alumina and sodium carbonate (their concentrations are represented by c K , c A and c C , respectively). They are obtained off-line by titration analysis and suffered from larger time delays. To solve this problem, a hybrid model for c K and c A is proposed by combining a mechanism model and a stochastic configuration network (SCN) compensation model. An SCN-based model for c C is also proposed using the estimated values of c K and c A from the hybrid model. A real-world application conducted in Henan Branch of China Aluminum Co. Ltd demonstrates the effectiveness of the proposed modelling techniques. Experimental results show that our proposed method performs favourably in terms of the prediction accuracy, compared against the regress model, BP neural networks, RBF neural networks and random vector functional link model.