Enhanced well-based surrogate reservoir modeling with integrated streamlines simulation data

Implementing deep learning-based surrogate reservoir models (SRMs) on an operational field scale and replacing them with numerical industrial simulators has yet hindered challenges in acquiring high-quality and fast-supply data for both the initial training and ongoing usage phases of SRM. These data challenges make conventional SRMs inefficient for field-scale reservoirs due to their dependence on time-consuming finite difference (FD) data. This research introduces a new generation of well-based intelligent models that tackle the challenges posed by conventional methods by integrating dynamic data from StreamLines (SL) as a reliable and fast-supply data source. SL data is utilized for the first time in developing a well-based SRM, which accelerates the data supply process and enhances the model's efficiency during simulation. The swift accessibility of SL data enables the use of a direct training strategy in the LSTM network, effectively mitigating typical training error accumulation. By introducing a new Streamline-based data sampling technique, data continuity and quality were vastly enhanced, resulting in higher accuracy of models with lower data requirements during the training phase and faster convergence. Additionally, leveraging SL data, the introduced database allows the intelligent models to operate independently from FD data during the ongoing usage phase. This independence eliminates the time load of finite difference data provision in conventional SRMs, greatly enhancing its functionality. This breakthrough in developing well-based SRMs offers improved data acquisition, simulation efficiency, and facilitates real-time decision-making in dynamic reservoir simulation applications. Promising results are achieved in predicting well oil/water production rates, with an average absolute relative deviation of less than 4% for two heterogeneous benchmark reservoir models. Furthermore, the simulation run time and overall process duration are significantly reduced compared to high-fidelity FD simulators and conventional proxy models.