Abstract: In order to improve prediction accuracy of corrosion rate of submarine pipelines, a submarine pipeline corrosion prediction model based on improved Random Forest was established. First, Spearman correlation coefficient was used to analyze correlation of corrosion data of real Marine hanging plates and factor analysis was performed to reduce dimensionality. Then, the K K-fold cross-Nalidation was set to five, K-fold cross-validation was set up, and RFR model was established to train dimensionality reduction data. Finally, output model's cross-validation score, and compare and evaluate model's highest score prediction result with other models. The research results show that average score of improved RFR model is 0.912, which is higher than BP neural network(BPNN) model and support vector regression(SVR) model. Meanwhile, root mean square error(RMSE) and mean absolute error(MAE) of the highest score prediction results of five-fold cross-validation are 1.441 and 1.3 respectively, which are better than corresponding values of contrast model.

Key words: submarine pipeline, corrosion prediction, random forest regression(RFR), Spearman correlation coefficient, factor analysis, K-fold cross-validation