Abstract: Based on the operating data from the cycle-cooling water system of a petrochemical enterprise, 899 sets of valid data samples were obtained by data preprocessing; the input variables for corrosion rate prediction (FSSL) model and adhesion rate prediction (NFSL) model were selected by using maximum mutual information coefficient and Pearson correlation coefficient methods. With 3 machine learning algorithms including BP neural network, KNN regression, and XGBoost, the prediction models for FSSL and NFSL of the cycle-cooling water system were established respectively. The evaluation results of the prediction accuracy and warning effectiveness of three models showed that the average relative error of the all three models was below 9%, and the three models had good fitting effect and generalization ability. The model based on XGBoost method had the best performance, the average relative error for both FSSL and NFSL was less than 5%, the decision coefficient R² was over 0.9, and the early warning accuracy was over 91.5% and 97.3% respectively.

Key words: cycle-cooling water system, corrosion rate, adhesion rate, machine learning algorithm