Machine Learning-Assisted Injection-Production Optimization for Offshore Low-Permeability Reservoirs

Offshore oilfields with complex well patterns and heterogeneous formations often experience unbalanced injection-production systems, early water breakthrough, and low recovery efficiency, making the optimization of well operating parameters particularly challenging. To address this, a hybrid optimization workflow is developed by coupling surrogate modeling with a Genetic Algorithm (GA) to maximize cumulative oil production in offshore low-permeability reservoirs. The workflow integrates Least-Squares Support Vector Machine (LS-SVM) and eXtreme Gradient Boosting (XGBoost) models as fast, accurate surrogates for numerical reservoir simulations. LS-SVM is selected for its robustness and numerical stability under data-scarce conditions, while XGBoost is employed for its strong nonlinear representation capability when greater data variability is available. The ZJ10 reef limestone reservoir in the South China Sea is used as a representative case to evaluate the workflow. Two optimization scenarios of increasing complexity are investigated: (1) optimization of injection and production rates, and (2) joint optimization involving both rate adjustment and production-to-injection well conversion. The LS-SVM model exhibits superior predictive performance compared to XGBoost when trained on small data sets. Integrating the surrogate models with the GA yields a 4–5% improvement in 10-year cumulative oil production compared with the current schedule. Further optimization with well conversion increases the injection-to-production ratio to nearly 1.0, enhancing pressure maintenance and achieving an additional 16% recovery improvement. This study demonstrates that the proposed LS-SVM-GA framework provides a computationally efficient and physically interpretable tool for optimizing complex offshore reservoir operations, offering a promising approach for data-constrained, field-scale decision-making in offshore oilfield development.