Investigating the performance of proxy and artificial intelligence models to optimize the injection and production strategy in modified Punq-S3 benchmark reservoir under water flooding operation

Today, even after decades of oil production from hydrocarbon reservoirs and practicing various EOR methods in the industry, water flooding is still one of the most widely used enhanced recovery techniques. Because real reservoir models are complex and their implementation leads to high computational cost and is time-consuming, hence researchers use artificial intelligence and neural network models as proxy models to solve this problem. The use of proxy models solves the problem of high computational cost and time consumption, but due to the uncertainty in predicting the behavior of the reservoir, it becomes a challenge. In this study, considering the recent challenge, the performance of proxy models has been investigated. To investigate this issue, a Punq-S3 benchmark reservoir model was selected to optimize well control parameters and maximize the net present value, and optimization was carried out using two approaches: optimization of control parameters of production and injection wells using proxy models and optimization using a real reservoir model. Three deep learning algorithms (ANN, LSTM and GRU) and two design of experiment method (Taguchi and Latin hypercube sampling) were used to create six proxy models. The particle swarm optimization algorithm was selected to optimize the injection and production strategy for the two approaches, and after optimization and comparison of the results, it was observed that despite the high accuracy of the proxy models, the optimal states in the proxy model-based methods and the real reservoir model-based method were completely different from each other, and the actual net current values ​​for these three cases indicate that the proxy models cannot accurately predict the optimization behavior. The use of proxy models cannot necessarily be reliable and trustworthy. Despite the existing challenge, by examining and analyzing proxy models and optimization results, proxy models with higher reliability can be created, and it was concluded that using the Taguchi design of experiment method and the use of artificial neural networks can achieve more reliable proxy models.