Reservoir simulation is an important tool for decision making and field development management. It enables reservoir engineers to predict reservoir production performance, update an existing model to reproduce monitoring data, assess alternative field development scenarios and design robust production optimization strategies by taking into account the existing uncertainties. A big obstacle in automating model calibration and production optimization approaches is the massive computation required to predict the response of real reservoirs under proposed changes in the model inputs. To speed up reservoir response predictions without compromising accuracy, fast surrogate models have been proposed. These models are either derived by preserving the physics of the involved processes (e.g. mass balance equations) to provide reliable long-range predictions or are developed based solely on statistical relations, in which case they can only provide short-range predictions due to the absence of the physical processes that govern the long-term behavior of the reservoir. We present an alternative solution that combines the advantages of both statistics-based and physics-based methods by deriving the flow predictions in complex two-dimensional models from one-dimensional flow network models. The existing injection/production wells in the original model form the nodes or vertices of the flow network. Each pair of wells (nodes) in the flow network is connected using a one-dimensional numerical simulation model; hence, the entire reservoir is reduced to a connected network of one-dimensional simulation models where the coupling between the individual one-dimensional models is enforced at the nodes where network edges intersect. The proposed flow network model provides a useful and fast tool for characterizing inter-well connectivity, estimating drainage volume between each pair of wells, and predicting reservoir production over an extended period of time for optimization purposes. We estimate the parameters of the flow network model using a robust training approach to ensure that the flow network model reproduces the response of the original full model under a wide range of development strategies. This step helps preserve the flow network model's predictive power during the production optimization when development strategies can change at different iterations. The robust networks training and the subsequent production optimization iterations are computationally efficient as they are performed with the faster flow network model. We demonstrate the effectiveness and applicability of our proposed flow network modeling approach to rapid production optimization using two-phase waterflooding simulations in synthetic and benchmark models.
