Dynamic Reservoir Prediction and Well Control Optimization Using Physics-Informed Framework
Résumé
Deep learning-based surrogates such as Physics-informed neural networks (PINNs) are gaining more and more attention in reservoir simulation workflows. However, conventional PINN frameworks often face challenges in handling discontinuities in reservoir data, emerging from geological heterogeneities or temporal variations in well controls, which hinders their effectiveness in dynamic reservoir management. To address this limitation, we propose an enhanced neural network framework, that can dynamically adapt to changes of reservoir behavior in response to well controls adjustments. To this end, our model incorporates well control features, into the neural network architecture, enabling the model to better predict 3D pressure maps, saturation maps and their corresponding production rates over time. Our framework is based on a physically aware encoder-decoder with a CNN backbone that can extract locally relevant geological features while keeping track of the global reservoir’s behavior. The governing Darcy’s law and mass conservation equations are enforced in their discretized form by the network. This allows us to handle the strong local heterogeneities within the reservoir in addition to the sink term, corresponding to the production rates, which adds a layer of complexity to the simulation due to its spatio-temporal dependence on the well coordinates and on the well controls. We evaluate our model on a 3D synthetic channelized reservoir presented in previous work with two-phases (oil & water) characterized by a heterogeneous permeability field with mud (lower porosity) and sand (higher porosity) facies. Our model produces very accurate results in terms of prediction accuracy respect to state-of-the-art numerical simulator, being much more computationally efficient when predicting new well controls schedules. Due to these large gains in computational speed the model can be coupled with an optimizer to optimize long term production. In our synthetic example we show that significant gains in terms of production can be obtained respect to more standard approaches.