An integrated machine learning model for calcium carbonate scaling monitoring in a closed-cycle seawater cooling system

• Published in: Journal of the Taiwan Institute of Chemical Engineers 2024-04, Vol.157, p.105434, Article 105434
• Main Authors: Li, Zhijie, Hu, Mingming, Zhang, Lianqiang, Chen, Chong, Xu, Kai, Kong, Qingling, Li, Zhuoxiao, Yin, Jianhua
• Format: Article
• Language: English
• Subjects: Gaussian anomaly detection; Machine learning; Random forest; Scaling monitoring; Seawater cooling
• ISSN: 1876-1070; 1876-1089
• DOI: 10.1016/j.jtice.2024.105434

•The establishment of an integrated calcium carbonate scaling monitoring system in a closed-cycle seawater cooling system, employing a combination of mechanistic and data-driven approaches, circumventing limitations associated with simple data or mechanistic models.•The successful development and validation of an integrated machine learning model using seven years of operational data from a closed-cycle seawater cooling system.•This model replaces complex scaling monitoring methods, such as △A, and enhances the efficiency of scale inhibitor utilization while reducing chemical residues. Continuous and real-time monitoring of calcium carbonate scaling of closed-cycle seawater cooling system can improve the efficiency of scale inhibitor utilization and reduce chemical residues. In this study, an integrated machine learning model was established for the monitoring and anomaly detection of calcium carbonate scaling. A novel approach is proposed, wherein changes in pH value are analyzed as an indicator of the scaling problem during the seawater circulation cooling process. The proposed framework leverages a Random Forest (RF) model to establish a predictive model for normal pH levels in the absence of scaling. By computing the difference between online pH measurements and the predicted normal pH values, the system enables online monitoring of calcium carbonate scaling. Evaluation of the RF model using key performance metrics, including the coefficient of determination (R2) and root mean square error (RMSE), demonstrates its high predictive accuracy. Furthermore, an anomaly detection algorithm based on the Gaussian distribution is developed to monitor abnormal conditions, including scaling. This model could replace complex scaling monitoring methods, such as △A, and enhances the efficiency of scale inhibitor utilization as a promising method to reduce chemical residues. [Display omitted]