Hybrid forecasting of oil and water production using long short‑term memory networks guided by reservoir simulation outputs

Prospective forecasting of oil and water production remains challenging because numerical simulators, although grounded in multiphase‑flow physics, may accumulate deviations during long‑range prediction, while purely data‑driven models lack forward‑looking information and often become unstable when extrapolating beyond historical data. This creates a methodological gap: despite advances in hybrid and physics‑guided learning, existing approaches are typically retrospective, rely on online assimilation, or require architectural modifications that break temporal causality. This study proposes a fully prospective, simulation‑guided forecasting workflow in which an LSTM model is trained solely on historical production and guided during extrapolation by pre‑executed numerical‑simulation forecasts routinely generated for field‑management. The workflow is causal, non‑assimilative, and free of surrogate construction, offering a practical and deployable integration between simulation outputs and deep learning. Using ten years of field data from a carbonate reservoir and evaluating 59 forecast start dates across two production units, the proposed hybrid approach consistently improves curve‑wise accuracy, volumetric alignment, and long‑range stability relative to the numerical simulator and purely autoregressive LSTM extrapolation. Improvements are most pronounced where simulator deviations are larger, demonstrating that even imperfect physical guidance can enhance prospective data‑driven forecasting. The workflow provides a transparent and operationally realistic pathway for hybrid prediction in complex reservoirs.