Enhancing Short-Term Production Forecast in Oil Fields: Integrating Data-Driven and Model-Based Approaches to Reduce Uncertainty

Abstract

Abstract Reservoir simulation models are usually applied to optimize oil field production across its life cycle but face challenges in short-term forecasting. Data-driven techniques (DD) show promise for short-term predictions but lack reliability over extended periods. This study introduces a Hybrid Methodology (HyM) combining the optimal features of model-based (MB) with DD approaches to select the best simulation models to make short-term decisions, effectively reducing uncertainty in short-term production forecasts for a real field. Using DD techniques, such as Transformers, we forecast 150 days of cumulative water, oil, and gas production for each well of a real field based only on historical production. We also run production forecasts using 200 history-matched simulation models for this field. In our HyM methodology, we propose to select the best simulation models according to their fit to the DD forecasts, assuming the DD forecasts to be more representative of the short-term behavior of the real field. We compare our HyM to a conventional MB model selection based on history-matching quality, referred to as MB. The proposed approach was applied to a heavy oil field from the Campos basin (Brazil), a deep-water turbiditic reservoir. We separated five months of real history data as the ground truth to rank the selected models through the forecast error and rank similarity. With this evaluation, we measured how accurate the forecast of the chosen simulation models was and the efficacy of each approach (HyM and MB) in ranking the simulation models. Our results showed that the subset of models selected using HyM successfully outperformed those chosen by the MB approach since they are better fitted for short-term forecasts for oil, water, and gas rates. The proposed HyM method yields a 75% error reduction, considering the rank similarity, with respect to the MB approach. Consequently, we conclude that better results may be achieved by using a DD forecast as a reference to reduce the uncertainty of oil, water, and gas rates (HyM approach) instead of using the actual history (MB approach). The new methodology showed promising results by selecting models with attenuated errors for oil/water rates during the history-forecast transition, complementing the data assimilation procedure. This study presents an innovative methodology merging the strengths of machine learning (ML) and simulation models to enhance the reliability of short-term production forecasts. The proposed approach is versatile, not tied to any specific ML algorithm, and effectively minimizes uncertainties, particularly in complex fields with numerous wells.