Russian Engineers Train AI to Estimate Reservoir Pressure at Oil Fields
The new method allows engineers to evaluate reservoir pressure without shutting down production.
Researchers at the TatNIPIneft Institute, part of the oil company Tatneft, have proposed estimating reservoir pressure using machine-learning algorithms. The approach could provide fast and reliable estimates without the labor-intensive field measurements traditionally required, according to the industry publication Neftyanye Vesti.
Estimating reservoir pressure is one of the key tasks in oil-field development, especially at late stages of production. This parameter determines well operating modes, the efficiency of water-injection systems and the validity of engineering decisions.
Reservoir pressure is usually determined through hydrodynamic well testing. But those measurements require temporarily shutting down production, which can result in major financial losses. At mature fields where wells have been operating for decades, obtaining stable and accurate measurements becomes even more difficult.
To address this challenge, specialists from the institute, together with the Almetyevsk-based Vysshaya shkola nefti (Higher School of Oil), developed a method for predicting reservoir pressure using machine-learning algorithms. Their study focused on the Bobrikov deposits of the Romashkinskoye oil field in Tatarstan – one of the most extensively studied yet technically complex reservoirs in the region. The model was trained using operational data collected over nearly 24 years of field exploitation. The analysis included well operating parameters that describe production regimes and operating conditions.
Training the Model
Researchers first analyzed how reservoir pressure correlates with various geological and operational parameters. The results showed that bottomhole pressure, pump installation depth and water cut had the strongest correlation. These and other indicators formed the basis for training the model.
The team plans to expand the set of input parameters and develop more comprehensive models that combine several machine-learning methods.
Earlier we reported that a digital registry of geological maps was created in Russia’s Samara region.
