A Novel Decision-Making Framework for Waterflooding Optimization using Long and Short-Term Memory Models and Metaheuristics

Abstract Meeting future energy demands in the low-carbon emissions paradigm requires improved decision-making processes. Waterflooding plays a vital role in obtaining increased oil recovery factors and minimizing undesirable water production. However, waterflooding optimization involves costly well control management optimization methods especially when numerical simulation is used. Alternatives such as data-driven proxy modeling can overcome the computation complexity drawbacks. In this paper, we develop a decision-making waterflooding framework, where an optimization component has embedded financial and machine learning models, to establish the wells operational plan obtaining the maximum profit and the best oilfield management. In this work, we use a reduced-order predictive numerical model to generate synthetic data to train machine learning algorithms to be used in the optimization framework. We develop our methodology to find the optimal strategy to drive a waterflooding project using a black-oil reservoir model. We implement the proper orthogonal decomposition (POD)-based model reduction to evaluate the reservoir dynamics and calculate the historical fluid production based on an operational plan, reducing the time consumption and demand for computational resources. Based on these results, we train and test machine learning models to predict oil and water production rate for each well (output data) in which the operational wells’ constraints change over time (input data), and select which of them has higher accuracy in the forecast. We evaluate the LSTM (long short-term memory) which are focusing in time series forecasting, using a multivariate model, analyzing Vanilla, Stacked and Bidirectional. Lastly, iteratively, the LSTM selected are embedded into a non-linear optimization component to define the best operational strategy in an oilfield with waterflooding, considering the reservoir's physics and a financial evaluation in a short- to mid-term planning horizon. The proposed making decision framework is applied to a two-phase heterogeneous waterflooding reservoir with a 5-point inverted injection pattern. Then we conduct the comparison between the multivariate LSTM model tested, selecting for each producer well two LSTM models, one for oil and one for water production rate. These models were integrated in an optimization component, which use metaheuristics and an iterative methodology to maximize the Net Present Value (NPV), considering the oil and energy prices fluctuation. The novel framework presented combines multiple algorithms in a seamless fashion. It allows us to define the operational plan strategy in an efficient manner, seeking a cost-efficient strategy over the mid and short-term. The methodology uses the LSTM models to learn about the historical reservoir behavior. Whit these we can accurately predict fluid production for each producer well. Both financial and LSTM models are embedded into an iterative optimization component which uses metaheuristics to maximize the NPV, establishing the bottom-hole pressure in each producer well and the water injection rate in the injector well.