A comparison of machine learning algorithms as surrogate model for net present value prediction from wells arrangement data

Net Present Value (NPV) measures whether an investment will be profitable within a given period of time. In oil production planning, it consists of an important indicator to evaluate different production strategies. The NPV estimate is calculated on the basis of the production data which are generally obtained by means of numerical simulations, which consider the strategy details and the physical reservoir model. However, the simulator demands high computational resource, which may take hours or days of processing time to evaluate a single strategy, depending on the size of the reservoir model. To speed up this process a simpler model, referred to as a surrogate model, can be used to approximate the simulator output. In this work, we hypothesize that it is possible to predict the NPV using only wells arrangement data as predictors. Moreover, we present a comparison among six machine learning algorithms used as a surrogate model: Linear Regression, K-Nearest Neighbor, Multi-Layer Perceptron, Kernel Ridge Regression, Support Vector Regression, and Gradient Tree Boosting. Results confirm it is viable to predict NPV from wells arrangement data, in special with kernel-based methods.