Improving Offshore Condensate Production Optimization Processes with Prediction of Reid Vapor Pressure Using Machine Learning

Operators need to know condensate Reid vapor pressure (RVP) to keep exports safe without over-heating the stabilizer and giving away liquid yield, but the platform's online analyzer was unreliable. This study built and benchmarked data-driven RVP soft sensors against a first-principles correlation on cleaned plant and laboratory data, then chose a model that is accurate and also simple to deploy and maintain. The result lets operations push condensate production closer to the specification limit with more confidence, and shows which process variables control RVP.

Key results

The online analyzer's error against lab samples was normally 1.5–2.5 psia RMSE, which justified building a soft sensor
Phase 1 showed ML beat the first-principles model, and the DCS predictor was replaced with a depth-4 decision tree
In phase 2, XGBoost and LightGBM scored best overall and Lasso/Ridge came second; the cleaned dataset held 51 features and 10,529 records
A linear model was deployed in a real-time visualization tool with rolling retraining and e-mail alerts on error drift, because it extrapolates better and needed no new infrastructure; stabilizer temperature was the strongest driver of RVP, followed by pressure and overhead gas flow

Approach

Joined more than 10 years of daily lab RVP results (interpolated to hourly) with plant historian data on flows, pressures, temperatures and sales-gas composition, then removed shutdown/upset periods and records that subject matter experts flagged as invalid
Phase 1: compared seven ML models with the in-use first-principles model over 33 time-based folds (300-day training, 90-day validation) using RMSE
Phase 2: tested tree-based, linear, simple deep-learning and distance-based models (XGBoost, LightGBM, Random Forest, Extra Trees, GBM, Decision Tree, ElasticNet, Ridge, Lasso, with and without PCA) using 10 back-tests plus a hold-out set, scored on RMSE, MAE and MAPE
Weighed accuracy against deployment limits in the DCS, historian and visualization tools, and analysed variable importance and coefficients to find what drives RVP

The problem

The stabilizer's online RVP analyzer was unreliable (its sample filter clogged and readings differed from lab results), so operators could not control RVP tightly and gave up condensate yield to stay within specification.

Tools & methods

XGBoostLightGBMRandom ForestExtra TreesGradient Boosting MachineDecision TreeLassoRidgeElasticNetPCAKorsten first-principles modelDCStime-series cross-validation

Summary written for this site from the published paper. For figures, data and the full method, read the paper on OnePetro.