A multi-fluid machine learning framework for leakage prediction in pressurized pipeline systems

Pipeline leakage to this day is considered one of the most serious operational issues on the municipal, industrial, chemical, and fuel-transport networks. Conventional leak modeling methods use a model that is water-based in nature and thus such models cannot be used to model fluids with different density-viscosity behavior. This research suggests a hybrid approach of predicting multi-fluids leaks through the combination of laminar outputs simulations, density-augmented analytical equations, and machine learning using multiple models. To isolate hydraulic effects of the fluids, four fluids, industry-influenced water, ethanol, diesel and glycerin, have been simulated at the same geometries of a pipeline consisting of a stainless-steel pipeline. The classical equations of leak size and leak position were re-parameterized to include fluid density and characterize the leak in fluid formations besides water-based models. Ridge regression, random forest, gradient boosting regression (GBR), support vector regression (SVR) and multi-layer perceptron (MLP) which are the five supervised regression models were trained to predict pressure and flow rate produced by leaks. Findings indicate that ensemble models, especially the random forest and GBR are the most accurate models and have better cross-fluids stability ensured by low error variance as well as high agreement between feature-importance with the hydraulic theory.