A New Hybrid Data-Driven and Model-Based Methodology for Improved Short-Term Production Forecasting

Abstract

Abstract Model-based (MB) solutions are widely used in reservoir management and production forecasting throughout the life-cycle of oil fields. However, such approaches are not often used for short-term (up to six months) forecasting due to the immediate-term productivity missmatch and the large number of models required to honor uncertainties. Recently developed data-driven (DD) techniques have shown promising performance in immediate term forecasting (from days to months) while losing performance as the timeframe increases. This work, proposes and investigates a hybrid methodology (HM) that combines MB and DD techniques focusing on improving the short-term production forecast. A common practice in reservoir management to understand the impact of uncertainties, is to build an ensemble of simulation model scenarios to assess the impact of these uncertainties on production forecasts. The proposed HM relies on the DD-assisted selection of a subset of models from the set of assimilated (posterior) models. Specifically, the pool of MB models is ranked based on their similarities to the DD production forecasts in the immediate term (e.g., one month), followed by the selection of the top models. The selected MB models are then used in the short-term forecasting task. In a case study for an offshore pre-salt reservoir benchmark, the proposed HM is compared to two baselines: one purely DD and another fully MB. The case study considered two forecasting conditions: human intervention-free with restrictions (HIF-R), with no intervention in the controls except to follow physical restrictions, and with human interventions (WHI), following optimization rules. Our results showed that the HM significantly outperformed the MB baseline, regardless of forecasting condition (HIF-R and WHI) or variables (pressure and oil/water/gas rates) for all evaluation metrics (time series similarity and rank-based) and top-selected models tested. The hybrid approach also helped improve the well productivity uncertainty that emerged from the data assimilation. Such results indicate that the performance of MB short-term forecasts can be enhanced when assisted by DD techniques, such as in our proposed HM. Comparing these two approaches, the best forecasts were split between the HM and the DD baseline. In the partially idealized HIF-R conditions, the DD baseline was best when the forecast trend was steady. However, the HM was superior for the more complex production behaviors. In the more realistic WHI conditions, the HM outperformed the DD baseline in almost every aspect tested given the inability of the chosen DD technique to leverage known interventions. This work is the first effort to improve MB short-term production forecasts, using production data, with a machine learning technique through a proposed HM. The proposed DD-assisted selection of models proved successful in a benchmark case study, which means it is promising for application in other fields and for further development.