A physics-based data-driven model is proposed for forecasting of subsurface energy production. The model fully relies on production data and does not require any in-depth knowledge of reservoir geology or governing physics. In the proposed approach, we use the Delft Advanced Reservoir Terra Simulator (DARTS) as a workhorse for data-driven simulation. DARTS uses an operator-based linearization technique that exploits an abstract interpretation of physics benefiting computational performance. The physics-based data-driven model is trained to fit data increasing the fidelity of the model forecast and reflecting significant changes in reservoir dynamics or physics over its history. The model is examined and validated for both synthetic and real field production data. We demonstrate that the developed approach is capable of providing accurate and reliable production forecast on a daily basis, even if the exact geological information is not available.

A physics-based data-driven model is proposed in this study for the forecasting of secondary oil recovery. The model fully relies on production data and does not directly requires any in-depth knowledge of the reservoir geology or governing physics. In the proposed approach, we utilise Delft Advanced Reservoir Terra Simulator (DARTS) as a base for data-driven simulation. DARTS uses an Operator-Based Linearization technique which exploits an abstract interpretation of physics benefiting computational performance for a forward simulation. The proposed strategy was evaluated first on the two synthetic data ensembles and showed good prediction accuracy for a significantly reduced model size. Besides, the data-driven proxy methodology was compared with an advanced flow-based upscaling technique and demonstrated an improved accuracy for both ensembles. Besides, the proposed data-driven approach was examined on two realistic data sets. For the first case, the methodology demonstrates advanced predictive performance for training based on synthetic data generated from a high-fidelity simulation model with imposed random noise. To check the robustness of the proposed methodology, the control parameters for a forecast period were significantly changed in comparison to the training period. The data-driven model still manages to predict the forecast production quite close to the reference high-fidelity results. However, the training performed on another data set based on historical production from a real brownfield was not fully successful. We relate a bigger error in both training and forecast period for this model to poor data quality. The training procedure for this model led to a moderate accuracy in history matching for a long production period, where general production trends have resembled true data and water breakthrough time was restored in nearly all wells. However, there are still periods of poor accuracy, especially where shark peaks and falls are experienced.