Graph Neural Networks and Multi-Objective Hybrid Optimization: A Review of Intelligent Decision-Making Frameworks for Waterflood Reservoir Injection–Production Regulation

Abstract

Real-time regulation of water-flooding injection–production systems is a crucial factor in enhancing reservoir development efficiency. However, in practical water-flooding operations, reservoir heterogeneity causes the inter-well fluid connectivity to vary dynamically over time. As a result, conventional regulation methods often require a considerable amount of time to adjust control schemes, leading to significant lag effects. When oilfields enter the medium- to high-water-cut stage, the geological structures and flow patterns become increasingly complex, making it difficult for traditional approaches to maintain real-time management. In recent years, Graph Neural Networks (GNNs) and hybrid multi-objective optimization algorithms such as Genetic Algorithm–Multi-Objective Particle Swarm Optimization (GA-MOPSO) have emerged as promising technologies in the field of reservoir development, demonstrating frontier applications in dynamic optimization and intelligent regulation of water-flooding systems [1-3]. GNNs can effectively model the graph structure of dynamic inter-well connectivity, accurately capturing the spatial–temporal evolution characteristics of the reservoir, while GA-MOPSO achieves a robust trade-off between maximizing the Net Present Value (NPV) and minimizing the Injection–Production Difference (IPD) [4]. In addition, recent studies have shown that surrogate models exhibit strong performance in predicting unseen geological configurations during heterogeneous multiphase flow simulations [5]. This paper provides a comprehensive review of GNN- and GA-MOPSO-based hybrid approaches for water-flooding regulation, with particular emphasis on the advantages of the Heterogeneous Spatiotemporal Fusion Model (HSTMF) in water-cut prediction. Furthermore, it discusses the key challenges and potential directions for future research in terms of model generalization, algorithmic robustness, and engineering deployment [6].