Lithology spatial distribution prediction based on recurrent neural network with Kriging technology

Abstract

Deep learning technology can fit the non-linear and non-stationary characteristics in geological statistics. It has become an important tool for predictive modeling. However, the classical deep neural network cannot integrate the spatial dependence directly from UTM_X, UTM_Y and depth when applied to spatial prediction and lacks the reliability analysis of the results. This study proposed a flexible spatial prediction method based on gated recurrent unit (GRU) neural network with Kriging technology. Spatial dependent deep learning, spatial constraint weights and related reservoir information are used to complete the prediction of lithology spatial distribution. Furthermore, Bayesian theory is integrated to quantify the reliability of the model. Seismic information is used as the spatial constraint of GRU neural network to realize the fusion of cross-domain knowledge, which can improve the accuracy and reliability of the prediction. Compared with the traditional Kriging method and deep learning method, the prediction accuracy (R2 = 95.071%) of the proposed method is improved by 8.642% and 3.034% in the field data. The method can be potentially applied to the spatial distribution prediction of related geological conditions.