Forecasting cumulative oil recovery from waterflood using a deep proxy model

The fluid injection rate, injection period, and well placement during waterflood govern the fluid displacement from the reservoir. In general, the available traditional reservoir simulators are utilized for attaining near-optimal environments to maximize oil recovery. However, this approach desires numerous iterations of long simulation runs which in turn, consume significant time and increase cost. In this context, a deep learning proxy model consisting of long short-term memory and densely connected neural networks has been proposed to predict cumulative oil production in the Waraga field, Uganda. The proposed model uses minimal information including the injection rate, injection period, and well pattern. One hundred and twelve cases validate the model's performance with only 1% average relative prediction errors. The proposed method allows for selecting the best field development scenario within minutes compared to several hours to days using the traditional approach.