Summary: A new deep-learning-based surrogate model is developed and applied for predicting dynamic oil rate and water rate with different well controls. The surrogate model is based on the graph neural networks (GNNs) and long-short-term memory (LSTM) techniques. The GNN models are used to characterize the connections of injector-producer pairs and producer-producer pairs, while an LSTM structure is developed to simulate the evolution of the constructed GNN models over time. In this way, we use geological attributes at wells and well controls with different times as input data. The oil rates and water rates at different times are generated. In this study, the GNN-LSTM surrogate model is applied to a high dimensional oil-gas-water field case with flow driven by 189 wells (i.e., 96 producers and 93 injectors) operating under time-varying control specifications. A total of 500 high-fidelity training simulations are performed in the offline stage, out of which 450 simulations are used for training the GNN-LSTM surrogate model, which takes about 150 minutes on an RTX2060 GPU. The trained model is then used to provide production forecasts under various well control scenarios, which are shown to be consistent with those obtained from the high-fidelity simulations (e.g., around 4.8% and 4.3% average relative errors for water production rates and oil production rates, respectively). The online computations from our GNN-LSTM model take about 0.3 seconds per run, achieving a speedup of over a factor of 1,000 relative to the high-fidelity simulations, which takes about 363 seconds per run. Overall, this model is shown to provide reliable and fast predictions of oil rates and water rates with a large level of perturbations in the well controls. Finally, the proposed GNN-LSTM model, in conjunction with the particle swarm optimization (PSO) technique, is applied to optimize the field oil production by varying the well control schedule of all injectors. Due to the significant speedup and high accuracy of the proposed surrogate model, the improved well-control strategies can be efficiently obtained.
Closed loop reservoir management is challenged with building reliable and fast predictive reservoir models to make field decisions. Traditional numerical simulation models can be difficult to characterize, tedious to build and calibrate, and at times computationally prohibitive for short term decision cycles in field applications. On the other hand, pure data-driven methods often lack physical insights and have limited range of applicability. For operational scenarios such as short-term production forecasting, waterflood optimization, production control and understanding major reservoir mechanisms, it is desirable to use a reservoir modeling methodology that is easy to build, history match, compute and interpret. In this work, we propose to use a hybrid and efficient reservoir graph network (RGNet) modeling approach based on time of flight concept that can be built using routinely measured field measurements (such as pressure and rates) and can be used for real-time forecasts, scenario modeling, production optimization and control. We propose a gridding method based on discretized time of flight for multi-well scenario with interference. It simplifies the 3D reservoir flow problem into a graph network representation that can be solved with any commercial reservoir simulator, which enables the RGNet model to be readily applied for various types of fluid physics. The parameters in RGNet model are obtained through assimilating observed data. The RGNet model has a very compact model representation that requires significantly less complexity compared with full-physics 3D models, which leads to very fast simulation. The efficiency of RGNet makes it appealing for applications where many simulation runs are needed. We applied the proposed approach on SPE benchmark reservoir simulation models for single well, multi-well with interference and injector-producer pairs. The calibrated models are used to quantify uncertainty for production forecasting. In all cases, the range of uncertainty is reduced effectively and efficiently with data assimilation. The posterior RGNet models are also shown to provide reasonable estimates of reservoir and well drainage volumes. By virtue of the reduced complexity, the modeling methodology is highly scalable while still retains physical interpretability (in terms of pore volume and transmissibility). We also discuss the potential applications of the method such as reservoir connectivity analysis and well control optimization. The proposed reservoir graph network (RGNet) modeling approach provides a unique and sustainable way to combine advanced analytics and physics to develop an explainable dynamic reservoir model that can be effectively used to understand reservoir behavior and optimize performance. The lightweight model lends itself naturally to fast computation that are required for scenario analysis and optimization.