Categorical Boosting-PSO: A SHAP-Interpretable Hybrid Framework for High-Fidelity Wax Appearance Temperature Prediction in Subsea Pipelines

Abstract

Wax deposition in subsea pipelines poses significant flow assurance challenges, yet accurate wax appearance temperature (WAT) determination remains constrained by resource intensive experiments and the absence of a robust, interpretable machine learning framework. This study proposes Categorical Boosting (CatBoost) optimized via particle swarm optimization (CatBoost-PSO) as a first of its kind, high-fidelity WAT predictor. It was rigorously evaluated against six competing models, including XGBoost, GBDT, XGB-GBDT-SE, and three additional CatBoost metaheuristic variants, on the most comprehensive reported WAT dataset (98 samples: 82 for model development, 16 unseen data for validation) using oil density, wax content, and pour point temperature as inputs. CatBoost-PSO demonstrated superior test-set performance (R² = 0.9760, RMSE = 1.8106 K, AARE = 0.5063%), surpassing established benchmarks, including MLP-LMA, MLP-BR, GEP, ANFIS, ET, RF, and DT, with a 33.3% RMSE reduction (from 2.7164 K to 1.8106 K), over closest competitor, the ANFIS model, although comparisons are indicative due to differences in datasets and validation protocols. Predictive reliability was established through external validation (AARE ≤ 1%) on unseen crude samples, trend analysis, k-NN distance assessment and Williams' plot. SHAP analysis identified pour point temperature as the dominant predictor, with wax content × pour point revealed as the strongest synergistic interaction pair, collectively consistent with paraffin crystallization thermodynamics. CatBoost-PSO constitutes a deployable, physically grounded screening tool, trained in 0.340 s with a 0.3 MB memory footprint, for high-throughput WAT screening in subsea pipeline flow assurance applications.