Proposing predictions of the evolution of the pressure in a field is usually done using dynamic models and thus it requires a proper model to be built and matched. This process takes several months to years depending on the complexity of the field fluids, geology, the number of wells and available data.
Our objectives are to evaluate if the pressure answer of a field regarding several production/injection schemes can be estimated with neural networks in situations of fast decision-making or when the model is not available or under construction.
For this demonstration project, python libraries are used to train models (keras) and build an interface (streamlit). The training process is done using a Jupyter notebook: 4 main families of models are available, each one with its own set of hyperparameters. The interface allows the user to navigate through the trained models and to select one of these for making predictions. In order to make a prediction, several boxes must be filled, field liquid production, field liquid injection and forecasted period.
Results are obtained using the Volve field (North Sea) open-source dataset and show that neural networks are able to capture the evolution of pressure depending on the production/injection patterns.
A comparison with the open-source Volve dynamic model must still be carried on, and a comparison with material balance could also be interesting. Testing the methodology on other fields would also be very useful to further validate the approach.