Abstract
Accurate knowledge of reservoir fluid properties, especially reservoir saturation pressures and reservoir type, is paramount for the estimation of reservoir volumetrics, well design and placement, well and reservoir performance management, field development planning and ultimately economic evaluation of the reservoir. However, most of the existing methods such as PVT fluid sampling, compositional grading, and empirical models have proven to be either ineffective or expensive and sometimes, leads to ungeneralizable results.
This paper discusses the application of machine learning (ML) techniques to develop a robust model for prediction of reservoir fluid properties such as saturation pressures in a Niger Delta Field and the subsequent classification of the reservoirs as under-saturated or saturated. Reservoir data including the PVT data, compositions (C1-C7+), temperature & pressure data, fluids contacts of known reservoir type were considered initially in order to train a model using a multi-features regression ML algorithms for the determination of the saturation pressure and a Two-class boosted decision ML algorithms to determine the determine the type of reservoir (saturated or undersaturated). Of the 34 parameters considered, it was found that the reservoir fluid composition has significant impact in determining the accuracy of the Pb-hypothesis. The key performance indicators such as MAE, RMSE, and RSE are within 0.02-0.05 and Co-efficient of determination of about 95% for Pb determination. When compared with the Standing, Glaso, Petrosky-Farshad etc. correlation, the AAE was significantly less than both cases. AAE for the Standing and Glaso correlations were respectively18.5% and 25.1% while that of the ML model was 2.3% using both the data from training set and test set. For the classification algorithms in determining type of reservoir, the model performed within the 73-100% accuracy, precision & recall. The Area under the curve (AUC) of the Receiver Operator Characteristic (ROC) chart of approximately 97% indicated the robustness of the model. The results showed that the use of a properly trained and accurately validated ML model can deliver better predictions of reservoir fluid properties and subsequent reservoir type when compared to conventional methods.
