Interpretable causal-based temporal graph convolutional network framework in complex spatio-temporal systems for CCUS-EOR

Global climate change has escalated in recent years. Carbon dioxide capture, enhanced oil recovery (EOR)-utilization and storage (CCUS-EOR) has the potential to significantly mitigate greenhouse gas emissions by storing CO2 underground, all while providing economic advantages. Hence, it is essential to comprehensively understand the CCUS-EOR mechanism and properly forecast its efficiency to mitigate the greenhouse effect. Nonetheless, current intelligent approaches are typically data-driven only and are unable to reveal the CCUS-EOR's physical mechanism. Therefore, we propose an interpretable causal-based temporal graph convolutional neural network (TGCN) framework, called Causal TGCN (CTGCN). The proposed framework integrates the causal graph into TGCN, based on Peter and Clark Momentary Conditional Independence plus (PCMCI+) for causal discovery. We comprehensively evaluated the effectiveness for causal discovery and prediction performance of the CTGCN model for CO2-EOR and storage. The findings indicate that the proposed model can appropriately forecast subsurface hydrocarbon production and CO2 storage capacity. In addition, it can effectively identify the CO2-WAG dynamic mechanism and the hydrocarbon causal flow pathways. Consequently, it effectively investigates the potential causal physical mechanisms of CCUS-EOR. It is noteworthy that we discovered that the predictive performance of deep learning methods can be significantly enhanced even if the model is embedded with a static causal discovery graph. To the best of our knowledge, our study is the first to use causal deep learning methods in the field of CCUS to reveal the causal mechanism of CO2-EOR and storage. This has immense importance in enhancing our comprehension of the operations of the complex spatio-temporal systems in CCUS and energy engineering in the future.