Strategic Well Test Planning Using Random Forest

Abstract

Abstract Optimizing a huge oilfield with hundreds of wells would involve thorough analysis. Here, the performance of a subset of those wells might be a critical factor. Therefore, it is important to obtain accurate information on that subset. Obtaining accurate information on an individual well is usually done via a proper well test analysis. Although performing a well test is relatively expensive, it can provide important information such as the skin, reservoir pressure (i.e. BHP), permeability, and nearby faults. In this study, we would like emphasize on the measurement of BHP. This emphasis is chosen due to the fact that the BHP from a subset of all wells can be a contributing factor towards the field oil recovery factor. Hence, the objective of this study is to find that subset of wells based on uncertainty studies conducted via reservoir simulations. Here, the uncertainty study involves performing 100 Latin Hypercube Monte Carlo (LHMC) sampling of the reservoir simulation input parameters. For each of the 100 runs, the uncertain input parameters and simulated BHP values/trends for each well are recorded. Thus, a dataset consists of the input parameters, simulated BHP for each well, and the field oil recovery factor as columns will be generated. The random forest (RF) algorithm was implemented in order to find the subset of wells mentioned. RF algorithm is simply a collection of decision trees (DT), where the results generated by RF were obtained by using a voting mechanism among the trees involved. Several trees were sampled from the generated forest, and reveal interactions between the monitored BHP value/trend with other parameters. Combination of LHMC and RF can also be seen a global approach to sensitivity analysis, since RF can reveal the rank of importance of the RF’s independent variables based on a collection of uncertainty runs. In the exploration and production (E&P) applications, DT has been used extensively; appraising, real option valuation selecting an artificial lift method, predicting permeabilities and detecting fracture corridors. However, to the best of the authors’ knowledge, RF had never been implemented in any E&P activities, particularly in strategic well test planning.