Deep-Learning-Based Autoregressive Framework for Rapid Production Forecasting in Mature Waterflooding Fields

Abstract Mature oilfields under waterflooding commonly suffer from high water cut and inefficient water cycling, posing significant challenges for production optimization. Conventional reservoir simulation–based optimization requires detailed geological modeling and extensive history matching, which are both labor-intensive and time-consuming for fields with decades of operation. To overcome these challenges, this study proposes a deep-learning-based autoregressive framework for production forecasting and injection optimization in mature waterflooding fields. In this framework, a neural model captures the nonlinear relationships between injection parameters and production responses using only historical field data. Both static features (e.g., inter-well distances) and dynamic time-series data (e.g., oil and water production rates, water injection rates) are jointly embedded to characterize the interactions among injector–producer pairs. An autoregressive deep neural network is developed and trained to predict future production profiles and guide injection optimization. Multiple architectures, including GRU, LSTM, and Transformer, are implemented and systematically compared within the same forecasting framework to evaluate their accuracy, temporal stability, and robustness across different prediction horizons. The framework was validated using synthetic data from a 3D numerical model with 17 producers and 18 injectors arranged in a five-spot pattern. Ten years of production data were used for training, and three subsequent years for blind testing. During six-month continuous forecasts, the trained model achieved mean absolute percentage error (MAPE) values below 10.39% for oil and 8.12% for water production rates. This study demonstrates that the proposed deep-learning-based framework can achieve reliable production forecasting without the need for detailed reservoir simulations or large synthetic training datasets, providing an efficient and scalable approach for data-driven management of mature waterflooding fields.