Precise forecasting of Gas-Oil Ratio (GOR) curves is crucial for the effective and safe exploitation of reservoirs influenced by CO2 flooding. On a macroscopic level, the GOR curve typically exhibits a rapid rise, while on a microscopic level, it shows significant fluctuations. These characteristics make it challenging for conventional prediction methods to capture these dynamics, resulting in notable deficiencies in existing univariate models in terms of rapid response to changes and anomaly detection.
To address the challenges in predicting GOR curves, this study employs impulse response functions and cross-correlation functions to identify the lagged correlation between water cut and GOR curves. Based on these findings, a prediction strategy incorporating water cut constraints is proposed. This approach utilizes a multi-input Long Short-Term Memory (LSTM) network to balance the long-term trends and short-term fluctuations in GOR. The model is trained on well production data from 26 production wells in a CO2 pilot area in Northeast China, capturing monthly production indicators to improve prediction accuracy and enhance early warning capabilities for CO2 breakthrough events.
The research results indicate that incorporating water cut as a constraint variable significantly improved the accuracy of GOR curve predictions, particularly enabling predictions 60 days in advance in multi-step forecasting. The model's mean absolute error (MAE) decreased from 260.60 to 172.89, and the root mean square error (RMSE) reduced from 522.87 to 382.15, demonstrating a significant enhancement in the model's prediction accuracy and performance.
Incorporating water cut as a constraint variable into the deep learning prediction strategy significantly improves GOR curve trend accuracy and sensitivity to fluctuations. This provides engineers with an early warning tool for CO2 breakthrough events, reducing uncertainty and risk in CO2 injection. Additionally, introducing lagged correlation variables enhances the model's ability to capture complex interactions in time series data, offering valuable insights and methodological references for future productivity prediction research.