Forecasting Oil Recovery Using Long Short Term Memory Neural Machine Learning Technique

Abstract

Abstract This paper focuses on using a timeseries neural network to forecast the oil recovery of a mature oil reservoir undergoing tertiary CO2 water alternating gas (WAG) enhanced oil recovery (EOR). Estimating future oil recovery is a necessity for planning an effective EOR strategy. Because of the high uncertainty associated with numerical modeling input parameters, modeling is not necessarily an accurate predictor of future performance for a specific well or even an entire field. The evolution of machine learning algorithms has shown that data-driven models can make decisions based on trends and pattern recognition to achieve tractable, robust, and cost-effective solutions. The methodology is validated by analyzing a five-spot pattern from the study field. The one injector well and four producers within the pattern are considered to be mutually connected. The multivariate timeseries (MTS) field data utilized in the model construction include production bottom-hole pressure, injection pressure, WAG cycles, and injection volumes. These MTS input data were preprocessed into a format that is more understandable and useful for the model. A Long-Short-Term Memory (LSTM) neural network model was established to determine patterns and trends, discover relationships from MTS data, and subsequently predict oil recovery through model-fitting. During the model construction, the preprocessed dataset was split into training and testing sets based on production time periods. The largest portion of the data set is apportioned to train the model, and it also corresponds to the earliest part of the production. The model is tested on the remaining data set chronologically. Analysis of field history calibration through loss iteration of the training dataset shows a low mean squared error of 7.16 and a relatively high R-squared value of 0.92. The developed model was validated using a test set, and results showed high-level model predictability of an R-squared value of 0.88. Additional model validation was performed using other wells’ information within the pattern as a blind test dataset. An average R-squared of 0.88 was observed for the other producing wells. The validated model was used to forecast oil recovery into the future with a reasonable outcome. From the forecast, uncertainty increased with the increasing length of time in the future, and the alteration of the WAG cycle significantly impacted the oil recovery. The LSTM model can predict oil recovery with a high level of accuracy. The successful predictions and reasonable forecasting of the oil recovery prove the effectiveness and usefulness of data-driven models. The workflow presented in this paper predicts the oil recovery without a detailed geological model and/or numerical simulation; it only considers time-changing parameters. Analyzing the LSTM model's results provides robust guidance to adjust real-time field development plans.