Graph neural networks and hybrid optimization for water-flooding regulation

Injection-production regulation in water-flooding reservoirs is a critical aspect for enhancing reservoir development efficiency and economic benefits. However, it faces multi-dimensional technical challenges in practical applications. In the dimension of dynamic response characterization, traditional well test analysis requires well shut-in testing, which disrupts the normal production process. Tracer monitoring is costly and has poor timeliness, making it difficult to meet the requirements of real-time regulation. Numerical simulation is constrained by grid precision and computational resources, making real-time model updates challenging. In the dimension of intelligent decision-making, the adjustment of injection-production parameters lags behind reservoir dynamic changes, resulting in low displacement efficiency. Moreover, conventional data-driven methods struggle to handle multi-objective optimization problems. To address the above-mentioned issues, this paper proposes a dynamic regulation method that integrates graph neural networks with hybrid optimization algorithms and constructs a closed-loop optimization framework for reservoir production dynamics. The main work is as follows: (1) Based on the relational graph convolutional network, an attention mechanism is introduced. Geological features and production data are utilized to calculate the attention weights between injection and production wells, achieving adaptive weight allocation for injection-production units and constructing inter-well topological relationships. Combined with the Transformer encoder, long-period injection-production correlation features are extracted, and the long short-term memory captures the dynamic evolution patterns of production time series. With geological features and production schemes as inputs and the prediction of inter-well water cut as the objective, the heterogeneous spatiotemporal multimodal fusion network model is established. Compared with the existing methods, this model does not interfere with the production schemes of on-site injection and production wells and significantly reduces the computational load. Through validation with two numerical examples, the usability and accuracy of the model are fully demonstrated, with the mean squared error below 0.001 and a coefficient of determination (R2) above 0.98. (2) A hybrid optimization strategy is proposed, combining the genetic algorithm (GA) with the multi-objective particle swarm optimization (MOPSO) algorithm to construct the GA-MOPSO hybrid optimization algorithm. Through the prediction model and the optimization algorithm, a closed-loop “monitoring-prediction-optimization” management system is constructed. Practical applications show that this framework increases the net present value by 4.83% and the cumulative oil production by 2.7% and reduces the block water cut by 4.02%, significantly improving the economic benefits and recovery factor of oilfield development and providing strong support for the sustainable development of oilfields.