Oil production forecasting using deep learning for shale oil wells under variable gas-oil and water-oil ratios

Abstract

Deep learning approaches can be utilized for production forecasting in cases with complex production profiles. In this work, a deep learning algorithm was developed to use the pertinent production profiles such as oil rate, GOR, WOR to train the model. Using reservoir simulation, the deep learning model was first benchmarked using synthetic for two cases: (i) decline oil rate under constant GOR and (ii) decline oil rate under variable GOR. The architecture of the recurrent neural network was structured by one LSTM layer and four dense layers. Also, the public dataset for monthly production of shale wells published by the North Dakota Commissioner was also used. The model could capture the incremental oil rate rise when the GOR or WOR changed. Comparison between decline-curve and deep learning indicated that deep learning could be more accurate. The results manifested that, deep learning performance can be categorized into healthy, moderate, and poor based on the magnitude of training loss. Poor performance was obtained for the data with multiple sudden rises and falls of the oil production history. But for the majority of the wells, the healthy type of model was achieved with consistently low losses for training and validation datasets.