Inter-well connectivity is a key parameter for evaluating the effectiveness of the CO 2 flooding development and optimizing injection-production strategies. However, existing research methods face challenges in dynamic modeling, real-time prediction, and adaptability to complex reservoirs. To address these issues, this study proposes a method based on the Long Short-Term Memory-Graph Attention Network (LSTM-GAT) to quantitatively analyze inter-well connectivity and predict future well production. First, a geological model is constructed, and numerical simulations are conducted to obtain dynamic production data and flow field distribution data of injection and production wells, serving as the foundation for connectivity analysis. Then, an LSTM-GAT model is developed based on the simulation data, where Long Short-Term Memory (LSTM) captures the temporal evolution of connectivity by learning long-term dependencies from time-series production data. Meanwhile, Graph Attention Network (GAT) extracts spatial dependencies by dynamically adjusting the importance of connections between wells through an attention mechanism. This enables the model to effectively integrate both temporal and spatial characteristics of inter-well connectivity, leading to a more accurate quantitative evaluation of connectivity strength and precise prediction of future production changes in production wells. The results demonstrate that this method effectively captures the dynamic connectivity between injection and production wells during CO 2 flooding development and provides scientific support for optimizing injection-production strategies. Compared with traditional methods, the LSTM-GAT model improves prediction accuracy, enhances adaptability to complex reservoir conditions, and enables real-time analysis. This study offers a novel approach for CO 2 flooding development in complex reservoirs, thereby improving precision and economic efficiency.
Li et al. (Fri,) studied this question.