Stratified allocation method for water injection based on machine learning: A case study of the Bohai A oil and gas field

The Bohai A oil and gas field is a natural gas and oil coproduction reservoir in the southern Bohai Sea, with an average gas–oil ratio of approximately 65 m 3 /m 3 . The oil and gas field has now entered the high water-cut stage, and in it, ineffective water circulation has intensified. Meanwhile, the process of adjusting the injection volume of water injection wells is overly complicated and relies on the experience of reservoir engineers. This paper established an automatic allocation method aimed at optimizing injection strategies based on the reservoir injection allocation scheme and utilizing real-time online data from intelligent layered injection wells by combining numerical simulation with artificial intelligence and machine learning algorithms. First, according to the basic parameters of block B in the Bohai A oil and gas field, a reservoir numerical simulation model was established, and historical fitting was carried out. The calculation found that the natural gas production of the A oil field would increase over time, although its oil production showed a decreasing trend. Using this model, finite group calculations were performed to establish an effective dataset. Second, the training and prediction effects of three machine learning prediction models—support vector machine, BP neural network, and random forest—were compared, and the BP neural network was selected as the machine learning mathematical model for injection allocation optimization. Third, 300 neurons and three hidden layers were used in the optimized neural network. The number of training set samples used was 185, and the number of test set samples was 19. Fourth, the optimized BP neural network model exhibits enhanced prediction accuracy, improved generalization capabilities, and superior dynamic relationship–capturing abilities. It effectively establishes a relatively accurate complex nonlinear relationship between the injected water volume and the production of natural gas and oil, providing valuable guidance for layered allocation in injection wells. The relative error of the calculation results of the optimized neural network prediction model is approximately ±2.3 %. This model can be utilized to simulate the injection allocation of injection wells, potentially increasing natural gas and oil production by over 4 %.