Learning-based pattern-data-driven forecast approach for predicting future well responses

Abstract Forecasting well responses, such as flow rates and pressures, is crucial for effective reservoir management and investment decision-making in the development of subsurface reservoir resources. Recently, data-driven forecast methods, such as data-space inversion (DSI) and a learning-based data-driven forecast approach (LDFA), have been introduced to mitigate the computational cost and geological constraint issues of history-matching methods. However, DSI and LDFA have extrapolation, conditioning, and prediction variance issues. In this study, we propose two simpler alternatives, a learning-based pattern-data-driven forecast approach (LPFA) and an ensemble conditioning step (ECS), to resolve the issues associated with DSI and LDFA. LPFA mitigates the extrapolation issue by scaling well responses for each instance using the mean and variance of an observation period. ECS addresses the conditioning and prediction variance issues of LDFA and LPFA by combining predictions of multiple learning models and screening out predictions that do not sufficiently honor observed data. The prediction performances of DSI, LDFA, LPFA, and ECS were compared using two benchmark models: Brugge and Olympus models. Among these methods, LPFA provided the most accurate predictions for future well responses and reasonable uncertainty intervals, achieving an error of 2.3%. Even when predicting well responses outside the range of prior data, LPFA maintained a prediction error of 4.8%, unlike the other methods whose performance significantly declined. ECS improved the prediction accuracy by 1–2% and reduced the uncertainty in the predictions of future well responses by approximately 50%. Our approaches are generic and can be integrated with other data-driven forecast methods to enhance prediction performance.