Quantitative characterization of shale gas reservoir properties based on BiLSTM with attention mechanism

Evaluating the potential of shale gas reservoirs is inseparable from reservoir properties prediction. Accurate characterization of total organic carbon, porosity and permeability is necessary to understand shale gas reservoirs. Seismic data can help to estimate these parameters in the area crossing-wells. We develop an improved deep learning method to achieve shale gas reservoir properties estimation. The relationship between elastic attributes and reservoir properties is built up by training a deep bidirectional long short-term memory network, which is suitable for time/depth sequence prediction, on the logging and core data. Except some commonly used technologies, such as layer normalization and dropout, we also introduce attention mechanism to further enhance the prediction accuracy. Besides, we propose to carry on the normal scores transform on the input features, which aims to make the relationship between inputs and targets clear and easy to learn. During the training process, we construct quantile loss function, then use Adam algorithm to optimize the network. Not only the characterization results, but also the confidence interval can be output that is meaningful for uncertainty analysis. The well experiment indicates that the method is promising for reducing prediction errors when training samples are insufficient. After analyzing in wells, the established model is acted upon seismic inverted elastic attributes to characterize shale gas reservoirs in the whole studied area. The estimation results coincide well with the actual development results, showing the feasibility of the novel method on the characterization for shale gas reservoirs.