A Machine Learning-Based Production Prediction Method for Multi-Type Fracturing in Waterflood Reservoirs

Abstract Resolving injection–production conflicts is essential to enhance the productivity of waterflood reservoirs. Fracturing is widely used as a key technique for regulating injection and production. Numerical simulation, commonly used for conflict management, often suffers from high computational cost, high dimensionality, and strong nonlinearity. To address these issues, machine learning is used to construct surrogate models for simulating fracturing performance. These models offer fast computation and strong flexibility. As a result, they provide an effective alternative to numerical simulation for production forecasting. This study proposes a reservoir production forecasting model that integrates a Graph Attention Network (GAT) and a multivariate Long Short-Term Memory (LSTM) network. The model is designed to enable fast and accurate prediction of post-fracturing productivity. A typical waterflood reservoir (Reservoir A) in Northwest China is selected as the case study. Key production-controlling factors are identified using the Random Forest algorithm. To account for the uncertainty in geological and development parameters, a prediction dataset that incorporates temporal dependencies is generated through numerical simulation. The GAT module is utilized to learn the spatial characteristics of well placement and distribution, while the LSTM is used to capture temporal characteristics in production dynamics. A production prediction model incorporating physical constraints is constructed by combining spatial and temporal learning. Model performance was assessed based on RMSE, MAPE, and the coefficient of determination (R2). Hyperparameters are optimized, and the model's accuracy and robustness under various fracturing conditions are verified. Several critical geological and engineering parameters in Waterflood Reservoir A predominantly influence production performance. These parameters are the effective reservoir thickness, water cut at the time of treatment, fracture half-length, SRV permeability, fracture conductivity, and initial reservoir pressure. A total of 50,000 synthetic samples of oil and water well treatments were generated using numerical simulation. These samples exhibit diverse geological characteristics and fracturing intensities. According to the error analysis, the traditional LSTM model is consistently outperformed by the GAT-LSTM model under various development stages and fracturing intensities. It demonstrates superior generalization ability and predictive accuracy in capturing complex nonlinear injection–production relationships and temporal patterns. This confirms the model's applicability and stability in complex waterflood systems.