Attention-based bidirectional gated recurrent unit neural networks for well logs prediction and lithology identification

Abstract

Many old oilfields have missed or distorted well logs data, which is due to long history of shutdown, poor borehole conditions, damaged instrument, and other reasons. These bring great difficulties to redevelopment of oil and gas field. In this study, a novel method was proposed to complete well logs prediction and quantitative lithology identification. A new model based on gate recurrent unit (GRU) neural network and attention mechanism was constructed in our study. A bidirectional GRU network was designed to extract key features from forward and backward well logs data along depth direction, and attention mechanism was introduced to assign different weights to each hidden layer to improve prediction accuracy. We fully consider the trend of well logs data with depth, the correlation of different log series and the actual depth accumulation effect. Compared with other four traditional intelligent models, experiments on actual reservoir dataset show that the constructed model can obtain lower error and higher fitting degree. Implementation of the proposed method can serve as an economical and reliable alternative for oil and gas industry and provides fast and effective data for further geological research combined with artificial intelligence.