A Bi-objective Optimization Approach to Reducing Uncertainty in Pipeline Erosion Predictions

Abstract Uncertainty presented in pipeline erosion predictions may limit the field application of these models. Especially when the to-be-tested regions highly extrapolate the prediction model, large prediction uncertainties can be witnessed. We previously developed a Gaussian process (GP) model based framework to estimate the erosion prediction uncertainty. There are two major goals in the quantification of erosion prediction uncertainty. The first is to minimize the prediction discrepancy of the model uncertainty. The second is to enhance the reliability of the predicted model uncertainty. GP modeling, as a kernel-based approach, relies on the proper selection of hyperparameters. The hyperparameters are generally optimized by maximum marginal likelihood (MLE) using conjugated gradient approach. However, using MLE as the objective function may not satisfy both goals. Furthermore, for non-convex marginal likelihood functions, optimization approaches like conjugated gradient is sensitive to the selection of initial values, and may lead to local optima. Here, we present a bi-objective optimization approach for training the GP model of erosion prediction uncertainty. GP models trained using bi-objective optimization outperform the GP model trained using the MLE based approach in terms precision and reliability for the prediction of erosion model uncertainty.