Prediction of Internal Corrosion in Oil and Gas Piping Systems—An Independent Analysis Using Machine Learning

This work presents machine learning (ML) algorithm development and testing for internal corrosion rate prediction in oil and gas (O&G) piping systems. The ML models were trained using TWI’s dataset of environmental and fluid parameters, operational and flow parameters, as well as pipeline and material parameters at 35000+ condition monitoring locations (CMLs). We have trained two ML models based on different artificial intelligence (AI) approaches: model 1, a direct regression model, and model 2, a composite classification-regression model. The ML models were created using data that posed remarkable challenges, such as biasedness, skewness, and duplicate values. We observed better statistical metrics for predicting using model 2 compared to prediction using model 1. Systematic parameter selection, power transformation, and hyperparameter optimization using the grid search technique were employed in the model development. Although the model 2 predicted well for lower corrosion rate (0.05mm/year), it fell short in predicting accurately for higher corrosion rates (>0.2mm/year) as the uncertainty increases at higher corrosion rates. We have also provided a discussion on the uncertainty in the model, targeted performance metrics, use of correction factors, improving model accuracy through model refinement, and reducing bias nature of the dataset by adding more diverse values at higher corrosion rates.