A Mechanism-Data Hybrid Approach for Enhanced Gas-Oil Ratio Prediction in CO2 Flooding Using Water Cut-Constrained Long Short-Term Memory

Summary Accurate forecasting of the gas-oil ratio (GOR), the most direct and sensitive indicator of CO2 breakthrough and gas channeling, is critical for mitigating channeling risks and optimizing CO2 flooding performance in enhanced oil recovery (EOR) operations. However, GOR time series often exhibit complex behavior, characterized by extended periods of stability followed by abrupt surges due to CO2 breakthrough. This irregular pattern challenges conventional data-driven methods [e.g., univariate long short-term memory (LSTM) models], which frequently struggle to capture such rapid transitions effectively. To overcome these limitations, this study proposes a novel multi-input LSTM framework that incorporates water-cut dynamics as a physical constraint. Applied to data from 26 production wells (2000–2023) in low-permeability CO2 flooding reservoirs, the model identifies a consistent precursor pattern: a decline in water cut precedes GOR surges by approximately 4 months. Leveraging this lagged correlation, we design a hybrid model that integrates auxiliary sequences, sliding time windows (STWs), and multivariate LSTM. The model dynamically balances historical GOR trends and real-time water-cut signals to improve responsiveness to abrupt changes. Overall, the proposed method reduces the mean absolute error (MAE) by 34% (from 260.60 to 172.89) and the RMS error (RMSE) by 27% (from 522.87 to 382.15). The improvement is even more pronounced during GOR surge periods, where the model achieves a 48.5% reduction in MAE, demonstrating significantly enhanced sensitivity to rapid fluctuations. In stable phases, MAE is reduced by 83.8%, highlighting the model’s adaptability across varying dynamic regimes. Feature attribution analysis further shows that water cut contributes significantly to long-term GOR trend modeling. This work demonstrates that integrating reservoir physics, specifically water-cut dynamics, into deep learning (DL) architectures significantly improves GOR prediction accuracy. It represents a shift from purely data-driven models to physics-informed artificial intelligence solutions in reservoir management, with promising implications for intelligent gas channeling surveillance in CO2-EOR operations.