Prediction efficiency of immiscible Water Alternating Gas Performance by LSSVM-PSO algorithms

Abstract

In this work the aim is developing LSSVM-PSO model capable of capturing the interplay between the most influential parameters (mechanisms) and recovery factor (RF) of WAG process in layered reservoirs. In a previous work 1840 Black Oil Model simulations were run for a 2D model with multiple layers, an injector and a producer, and used to derive a dimensionless number correlating reservoir heterogeneity, WAG hysteresis, gravity, mobility ratio and WAG ratio to predict recovery factor (as measured after 1.5 injected pore volumes). Given that only one parameter, the dimensionless number, was used to correlate RF, a significant data scatter was observed. In this work the database is expanded by running 824 new simulations using new hysteresis parameters values. The Machine Learning algorithm Least Squares Support Vector Machine (LSSVM) is used to correlate RF with representative input parameters, such as characteristic mobility ratios, gravity numbers, heterogeneity factor and more. The appropriate number of effective input parameters was obtained by reducing the set of independent input parameters to dimensionless groups. Particle Swarm Optimization was used to optimize the LSSVM algorithm parameters. The trained LSSVM-PSO model could serve as an effective screening tool in uncovering important trends of parameter variation and improve the efficacy of uncertainty analyses.