EUVIMEDEUVIMED entdecken
Crossref

Physics-Informed Latent Potential Graph Flow Transformer for Efficient Spatiotemporal Forecasting and Optimization in Layered Waterflooding

Xianing Li, He Liu, Jia Liu, Jianqiao Liu, Jing Zhang, Bing Hou, Kai Du, Huanquan Pan

SPE Journal · 11.11.2026 · 31 (11) · 7263-7279

DOI: 10.2118/236596-pa ↗

Kurz gefasst

Fachartikel in der Zeitschrift SPE Journal, online erschienen am 31.08.2026. Verfasst von Xianing Li und 7 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Layered water injection optimization in mature high water-cut reservoirs remains a challenging problem because repeated full-physics numerical simulations are computationally prohibitive, while conventional data-driven surrogate models often fail to capture essential reservoir physics, such as interwell pressure communication and physically consistent flow behavior. To address these limitations, we propose a physics-guided spatiotemporal surrogate model termed the latent potential graph flow transformer (LP-GFT), which integrates transmissibility-driven graph diffusion with a causal autoregressive transformer for rapid production forecasting and well control optimization. Unlike conventional graph neural networks (GNNs) that rely on normalized adjacency matrices for feature smoothing, the proposed framework constructs a physics-informed graph topology based on interwell connectivity and transmissibility. Starting from the weakly compressible pressure diffusion equation, reservoir dynamics are discretized into a graph Laplacian operator, which serves as the physical foundation of the spatial encoding module. A latent hydraulic potential field is introduced as a reduced-order proxy of pressure-driven flow potential, enabling multistep graph diffusion while preserving geological connectivity and physically consistent flow behavior. The extracted spatial representations are subsequently processed by a causal transformer with autoregressive attention masking to capture long-term temporal dependencies and delayed injection/production responses. To further improve model robustness, a physics-informed loss function is designed to enforce data fidelity and stabilize the latent potential representation through gauge fixing. The proposed framework is validated using a layered waterflooding case from Block X in eastern China. On the test set, LP-GFT achieves high-fidelity predictions at field operational scale, with root mean square errors (RMSEs) of 0.0862 m3/d for oil production and water production. The average coefficient of determination (R2) exceeds 0.97 for oil production and reaches 0.99 for water production, demonstrating strong predictive robustness. Furthermore, the trained surrogate model is coupled with the simultaneous perturbation stochastic approximation (SPSA) algorithm for well control optimization. The optimized strategy improves the net present value (NPV) from 4.07 × 107 to 4.63 × 107 USD. These results demonstrate that the LP-GFT-based framework provides an efficient, physically consistent, and practically applicable solution for reservoir forecasting and rapid decision-making in mature layered waterflooding systems.

Zitieren

APA

Li, X., Liu, H., Liu, J., Liu, J., Zhang, J., Hou, B., Du, K., & Pan, H. (2026). Physics-Informed Latent Potential Graph Flow Transformer for Efficient Spatiotemporal Forecasting and Optimization in Layered Waterflooding. SPE Journal, 31(11), 7263-7279. https://doi.org/10.2118/236596-pa

Vancouver

Li X, Liu H, Liu J, Liu J, Zhang J, Hou B, et al. Physics-Informed Latent Potential Graph Flow Transformer for Efficient Spatiotemporal Forecasting and Optimization in Layered Waterflooding. SPE Journal. 2026;31(11):7263-7279. doi:10.2118/236596-pa

BibTeX
@article{Li2026,
  author = {Li, Xianing and Liu, He and Liu, Jia and Liu, Jianqiao and Zhang, Jing and Hou, Bing and Du, Kai and Pan, Huanquan},
  title = {Physics-Informed Latent Potential Graph Flow Transformer for Efficient Spatiotemporal Forecasting and Optimization in Layered Waterflooding},
  journal = {SPE Journal},
  year = {2026},
  volume = {31},
  number = {11},
  pages = {7263--7279},
  doi = {10.2118/236596-pa}
}
Kostenlos starten →