Spatiotemporal multimodal data fusion transformer-based gas channeling prediction model

Since gas injection for enhanced oil recovery (EOR), gas channeling is a key factor restricting development of gas injection projects. Accurate prediction of gas channeling directions provides critical basis for adjusting development plans, but the challenge of multimodal data fusion in complex reservoirs has constrained prediction accuracy. Aiming at the challenges of multimodal data fusion in dynamic gas channeling prediction for complex reservoirs, this article proposes a gas channeling direction prediction model based on spatiotemporal multimodal data fusion Transformer. By integrating permeability field images, well location distribution images, fixed geological parameters, and dynamic injection-production data, a multi-branch spatiotemporal collaborative prediction framework is constructed to realize the prediction of gas channeling direction. Experimental results show that the structural similarity between the prediction results (based on the structural similarity index measure [SSIM]) and those from the commercial numerical simulation software (computer modeling group [CMG]) is close to 90%. Compared with traditional methods, the model exhibits significant advantages in the continuity of gas channeling paths and details of spatial distribution. This research provides an efficient technical tool for optimizing gas injection development plans and preventing gas channeling risks, and its multimodal fusion framework has promotional value for similar spatiotemporal prediction problems.