Permeability prediction using hybrid techniques of continuous restricted Boltzmann machine, particle swarm optimization and support vector regression

Abstract

How to obtain reliable permeability data is universally considered as one of the critical work that guides geologists to explore oil-gas accumulation zones underground. Many significant researches related to permeability prediction have revealed that permeability can be directly calculated from logging data under usage of some complex non-linear equations. In this way, the key of permeability prediction is how to establish relational expression between permeability and logging data. Support vector regression is one of the best mathematical models using to explain complex mapping relationship between independent and dependent variables, and thus it can be viewed as an ideal approach to predict permeability. However, such model cannot be effective when different kinds of input data have high correlation or network parameters are not evaluated well. Then other two mathematical models, continuous restricted Boltzmann machine and particle swarm optimization, are referred to use to support the application of SVR. CRBM is functional to make a new data separation from the raw data, and network parameters can be optimized after PSO process. Therefore a new data-driven permeability prediction model CRBM-PSO-SVR is provided in this article. Data source used for method validation derives from five coring wells of the IARA oilfield, Santos Basin, Brazil. In two self-designed experiments, the accuracy rates of new method are respectively 67.34% and 76.67%, both of which are higher than those of other comparison methods. Experiment results well demonstrate the effectiveness of new method in permeability prediction when only logging data is available.