Predicting Wax Deposition in Deepwater Pipelines Using a Genetic Algorithm-Optimized Multilayer Perceptron Model

Abstract

Wax accumulation in deepwater crude-oil conduits continues to be a significant flow-assurance issue, as relatively little accumulation can raise pressure loss, limit flow and prompt unplanned pigging or shut-downs. Accurate prediction of deposition rates is challenging due to the highly nonlinear interaction between fluid properties, thermal gradients, and near-wall hydrodynamics. In this paper, a GA-optimized Multilayer Perceptron (GA-MLP) is presented in order to reduce reliance on manual trial and error process on MLP design in neural network and to establish a more systematic modeling workflow which is able to repeat anytime. A dataset comprising 149 experimentally measured cases was utilized, with an 80/20 split to form training and test subsets while maintaining balanced distributions of the target variable across both sets. Model architecture selection was guided by a genetic algorithm, which identified an optimal multilayer perceptron (MLP) with two hidden layers and appropriate regularization parameters. To enhance model robustness, an ensemble of networks was trained, each initialized with distinct random seeds. Model performance was evaluated in both the transformed feature space and after back-transformation to the original engineering units, ensuring practical relevance of the results. In the transformed domain, the GA-MLP ensemble achieved a training coefficient of determination (R2) of 0.9083 with corresponding RMSE and MAE values of 0.3016 and 0.2133, respectively, while the test set yielded an R2 of 0.5092, RMSE of 0.6798, and MAE of 0.5242. When predictions were converted back to engineering units, the model achieved training R2, RMSE, and MAE values of 0.8339, 4.77 g/m2·h, and 2.77 g/m2·h, respectively, with test values of 0.4512, 9.72 g/m2·h, and 6.74 g/m2·h. Further calibration, performed exclusively using training data and with predictions constrained to physically realistic ranges, improved real-unit performance to R2 values of 0.8735 for training and 0.3809 for testing. Overall, the genetic algorithm-optimized MLP framework offers a systematic and reproducible methodology for modeling wax deposition in deepwater environments. Future work will involve benchmarking these results against Radial Basis Function Neural Networks (RBFNN) and other conventional regression approaches, providing a comprehensive assessment of the relative strengths and trade-offs associated with each method.