Predicting Oil Production in Single Well using Recurrent Neural Network

Single well production prediction is an essential task for oilfield development planning and analysis. Existing methods used for such prediction suffer from a few problems. In particular, current methods do not consider large-scale data labeling or production prediction in different water cut phases. To this end, we propose to holistically use the static, historical data of a single well, such as its geological and production data to enable data labeling in different phases via our labelling tool. In addition, we use Long Short-Term Memory (LSTM), a well-known Recurrent Neural Network, to build a predictive model for single-well production. The proposed model uses dominating features on well production and can train multiple wells together, which can generalize the application of the model. The model has also been fine-tuned to speed up training via the use of batch normalization. Compared with Random Forest (RF) and Support Vector Machine (SVM), our proposed LSTM model demonstrates better prediction accuracy and strong generalization capability and thus lends itself nicely to single well production prediction in various water saturation phases.