Sensitivity Analysis and Potential Prediction of Heavy Oil Reservoirs Under Different Steam Flooding Methods

Heavy oil reservoirs often enter a high-water-cut and low-production stage after multiple cycles of steam stimulation. Converting to steam flooding can enhance recovery, yet the reliable prediction of incremental production potential and optimal design of injection–production parameters remain limited. In this study, a real heavy oil reservoir block was selected to develop a hybrid modeling framework integrating numerical simulation and machine learning for predicting steam flooding performance. A conceptual model was established on a numerical simulation platform to reproduce the transition from cyclic stimulation to continuous steam flooding, analyzing temperature, oil saturation, and recovery evolution under different geological, operational, and process conditions. Sensitive parameters were identified through single- and multi-factor analyses, and mathematical models for multiple injection–production schemes—continuous, cyclic, and asynchronous—were constructed for optimization. A comprehensive multi-scenario dataset combining simulation and field data was used to train and validate several machine learning models, including artificial neural networks, gradient boosting decision trees, XGBoost, and LightGBM. Among them, the LightGBM model achieved the highest predictive accuracy (R2 = 0.99) and computational efficiency. The proposed framework enables the rapid and reliable prediction of incremental oil potential and provides a robust tool for optimizing steam flooding parameters, offering significant value for field-scale heavy oil development.