Mixture of relevance vector regression experts for reservoir properties prediction

One of the most indispensable works in the oil and gas exploration and exploitation is reservoir properties forecast. We develop a reservoir properties prediction method based on the mixture of relevance vector regression (RVR) experts. For reservoir properties prediction, an individual machine learning model is probably insufficient with limited training data. The mixture of experts can decompose a complicated problem of reservoir properties prediction into several relatively simple sub-problems by incorporating multiple learning models, where each model processes specific parts of the data. In the proposed method, RVR has been chosen as the expert because it results in sparser regressors and determines hyperparameters automatically. It can also project data into a high-dimensional space through kernel functions to resolve the non-linear problem and avoid the curse of dimensionality. At first, a mixture of RVR experts model is trained on the well data samples. The input features are elastic properties and the output is a reservoir property. Then, the learning model is applied to test dataset and the corresponding reservoir properties are generated. The proposed method is applied to two field data, in which the learning model is obtained by training on the well data and is tested on the well and seismic data, respectively. Compared with an individual RVR expert, some metrics, such as mean absolute derivation (MAD), root mean square error (RMSE), coefficient of determination (R-2) and Akaike information criterion (AIC) etc., are effectively improved. The successful implementation of the method demonstrates its feasibility, certifying the superiority of the new method in the aspect of likelihood of fit and accuracy again.

Automated summary

This paper proposes a reservoir properties prediction method based on a mixture of relevance vector regression experts. By incorporating multiple learning models, the method decomposes the complicated problem into simple sub-problems and improves the accuracy of prediction.