A machine learning-driven stochastic simulation of underground sulfide distribution with multiple constraints

Abstract

The increase of sulfide (S2−) during the water flooding process has been regarded as an essential and potential risk for oilfield development and safety. Kriging and stochastic simulations are common methods for assessing the element distribution. However, these traditional simulation methods are not able to predict the continuous changes of underground S2− distribution in the time domain by limited known information directly. This study is a kind of attempt to combine stochastic simulation and the modified probabilistic neural network (modified PNN) for simulating short-term changes of S2− concentration. The proposed modified PNN constructs the connection between multiple indirect datasets and S2− concentration at sampling points. These connections, which are treated as indirect data in the stochastic simulation processes, is able to provide extra supports for changing the probability density function (PDF) and enhancing the stability of the simulation. In addition, the simulation process can be controlled by multiple constraints due to which the simulating target has been changed into the increment distribution of S2−. The actual data test provides S2− distributions in an oil field with good continuity and accuracy, which demonstrate the outstanding capability of this novel method.

Keywords

• machine learning
• stochastic simulation
• sulfide
• chemical estimation
• multiple constraints