Data-Driven Proxy Models Based on Recurrent Neural Network for Reservoir Simulators

• Juan A. R. Tueros UFPE
• Johnattan D. F. Viana UFPE
• Victor B. Silva UFPE
• Aluízio F. R. Araújo UFPE

Resumo

Management of oil reservoirs involves conducting numerous simulations to understand the oil field, which results in high computational costs. These costs have an adverse effect on activities related to controlling production and injection wells for optimizing production. This study aims to develop a data-driven proxy model that uses only historical production data to replace the reservoir simulator. The proxy models must accurately predict responses and capture reservoir dynamics, i.e., a controlled time series that includes the time series of fluid phase flow rates and bottom hole pressure, based on the specified controls of injection and production wells. For this purpose, in this work we apply four neural network strategies to capture reservoir dynamics: Long Short-Term Memory (LSTM), LSTM with attention mechanism (LSTM-A), Gated Recurrent Unit (GRU), and GRU with attention mechanism (GRU-A). To assess the performance of these strategies, we used a synthetic reservoir model from the literature with 4 production wells and 8 injection wells. GRU-A demonstrated superior effectiveness and exhibited lower variability compared to other approaches, achieving up to 13% improvement in error reduction on the test set. Thus, even simpler neural network models prove useful for replacing simulators, saving time and computational resources. The source code is available here.