Robust Adaptive Model Aggregation for Well Production Forecasting

Summary Recent advances in closed‑loop reservoir management enable the optimization of the production of the entire reservoir life cycle using short- and medium-term production forecasts. Within this context, strategies based on sequential model aggregation have gained prominence for generating production forecasts in reservoirs at a fraction of the computational cost of those used by industry nowadays. However, mitigating the influence of field data measurement noise on forecasts can be very challenging from a practical standpoint. With this work, we propose a robust adaptive model aggregation (AMA) method that incorporates an adaptive exponential moving average-based filter to mitigate the effects of measurement noise on forecast accuracy. The effectiveness of the proposed approach is demonstrated through simulations on synthetic reservoirs (Egg and OLYMPUS) and on a carbonate reservoir in a Brazilian pre-salt field and is compared with the results of three established model aggregation methods from the related literature—ridge, lasso, and exponentially weighted average (EWA). Findings indicate that the proposed approach is consistently the most accurate predictor of production rates, particularly oil production rates (OPRs), in the presence of measurement noise compared with other forecasting methods considered.