The Reid Vapor Pressure (RVP) of condensate is an important factor in condensate production due to its connection with safety of transportation. Despite their significance, research attempts to precisely estimate RVP in condensate production are not widely studied. This research provides a methodology for assessing estimation algorithms and a scenario for deploying machine learning techniques to reliably forecast the condensate RVP for a production facility.
Several RVP prediction approaches are used to evaluate the accuracy and field optimization prospects. Machine learning methods and first-principles models, such as Korsten's, are evaluated to assess accuracy based on operating parameters, laboratory tests, and online analyzer results. The operating data from the field has also been joined with laboratory testing data from the export floating and storage offloading vessel and served as a basis for this study.
This project examines RVP prediction, feature engineering/selection, cross validation, modelling, error metric evaluation, and deployment system determination results, emphasizing on modelling accuracy, simplicity, explain ability, model life cycle management, and deployment impacts to current system infrastructure. The study reveals that statistical linear models perform well, even if machine learning methods like adaptive boosting regressor are more accurate. The prediction from the traditional empirical model is inferior to the machine learning results. Investigation indicates that the instrument selection does not account for sample contaminants, resulting in wrong instrument reading. The first-principle model (i.e., Korsten's model) provides data with a high bias but a low variance.
To validate the most influential characteristics, the selected features recommended by machine learning models are also discussed with subject matter experts. Eliminating data during process disturbance events is one effective key feature engineering technique identified from this study. Adoption of selective time-based conditioning and periodic process upset periods (e.g., compressor shutdown, abnormal maintenance activities). This significantly improves model correctness. Based on the results of the study, several deployment options have been evaluated, and real-time data visualization tool has been selected for deployment due to its ease of use and outstanding user interface function, as well as its periodic model evaluation and real-time data streaming in the backend.
This study researches several solutions and proposes a successful technique for predicting the condensate RVP. The methodology and results used in this study are helpful for both petroleum industry and academic research, especially when it comes to predicting condensate RVP in full-scale condensate production and understanding the key stabilization processing qualities that affect RVP. This will help to guarantee that the field maximizes economic benefits and complies with safety standards for condensate production.
