Prediction of well performance in SACROC field using stacked Long Short-Term Memory (LSTM) network

Abstract

Forecasting oil production can be very challenging, especially for reservoirs with sparse data or other complexities. If traditional decline curve analysis or time series models fail to capture production rate variabilities, a machine learning model for time series data may be effective. A temporal machine learning model, a long short-term memory network model (LSTM) in specific, may be trained to predict oil, gas, and water production rates. We develop an LSTM for such an application and evaluate its efficacy with a case study of the Scurry Area Canyon Reef Operators Committee (SACROC) unit, an active CO2 enhanced oil recovery (EOR) field in western Texas, USA. The monthly averaged production rates of oil, gas and water from 23 producers are obtained from simulations of the SACROC reservoir model that also includes 22 injectors for 5-spot injection of water and CO2 (alternating annually). The bottom hole pressure (BHP) of the producers, BHPs of surrounding injectors, and historical production rates are used as input data for the LSTM. Blind predictions of LSTM test sets show promising outcomes for data that otherwise traditional time series models are not effective. Stacked LSTM models are efficient in multi-step-ahead forecasting. Such an LSTM approach may also be effective for quantitative analysis of unconventional oil and gas reservoirs like shales or other tight formations. Critical aspects of the LSTM workflow include optimization of machine learning parameters and quantification of the relative impacts of different variables on forecasted outcomes.