Multi-objective optimization of reservoir development strategy with hybrid artificial intelligence method

Optimization of subsurface hydrocarbon production holds paramount importance for decision-makers as it de-termines crucial development strategies such as optimal well placement and well control parameters (e.g. in-jection/production rates of injectors and producers). Despite the availability of numerous established optimization methods in this field, traditional reservoir production optimization methods face challenges in simultaneously addressing multiple development objectives and coordinating the interaction of well control parameters in different control steps. In this work, we construct a hybrid artificial intelligence method to jointly optimize well placement and well control parameters, taking into account two development objectives and dynamic optimization. It consists of two stages. First, a reservoir potential map is generated with the production potential formula that considers reservoir pressure, remaining oil saturation and other reservoir properties (permeability, hydrocarbon column height) etc. The reservoir potential map provides guidance for placing well in medium to high potential areas and engineering constraints for the optimization process. Then, a hybrid artificial intelligence method that couples deep learning method (Long and Short Term Memory (LSTM)) and multi-objective optimization algorithm (Non-dominated Sorting Genetic Algorithm II (NSGA-II)) is established to seek a compromise between the two objectives in water-flooding processes. The LSTM neural network is trained as the surrogate model to replace the high-fidelity simulator to achieve high efficiency of overall optimization workflow. The NSGA-II algorithm is employed for handling the joint optimization problem of well placement and well control parameters by maximizing the cumulative oil production and minimizing the water cut. The per-formance of the proposed method is tested on one benchmark function and two reservoir models. On the 2D synthetic reservoir model the optimized scheme leads to a notable increase of 3 x 104 m3 in cumulative oil production, accompanied by 17 % reduction in water cut when contrasted with the base scheme. Similarly, within the 3D reservoir model, the optimized scheme results in a substantial enhancement, boosting cumulative oil production by 14 x 104 m3 and reducing water cut by 20% compared to the base scheme. Moreover, the proposed method surpasses alternative multi-objective optimization (MOO) algorithms by demonstrating 82% and 95% reduction in optimization time, respectively. The results demonstrate that this method can provide the optimal water-flooding strategies under the premise of different development objectives. The Pareto front (or optimal solutions) generated by the hybrid method offers a variety of diverse water-flooding strategies to assist subsurface engineers in making informed decisions.