Large data volumes and complex physical processes coupling multiphase flow, numerics and production assessment, as in the case of reservoir simulation, are difficult to analyze rapidly but are needed to guide operators for designing production strategies and subsequent reservoir management. The traditional reservoir simulation process is time consuming and alternatives such as data-driven proxy modeling can overcome the computation complexity drawbacks. A machine learning technique called recurrent neural network (RNN) has been proved useful for reservoir modeling with sequence data. In this paper, we develop a novel end-to-end production prediction workflow that can be used to fast guide reservoir development when production begins.
In this work, we apply RNN on analyzing control parameter data and synthetic historical production data for better reservoir characterization and production prediction. More specifically, we would like to build a model to directly link the control parameters (flow rate and bottom hole pressure) with desired production output, e.g. production rate and water cut. One immediate benefit of the model is to avoid the state variable calculation (pressure/saturation). In addition, as this is a data-driven end-to-end production prediction model, it will not require the numerical iteration and gradient calculation once the training is completed. We explore two types of RNN based structure: cascaded LSTM and Ensemble Kalman filter (EnKF) enhanced LSTM. The LSTM (long short-term memory) is used to compensate the weakness of standard RNN for preserving long time information dependencies. The structure of LSTM takes into account the memory of previous calculation when modeling the current response. Our cascaded LSTM is an improvement to regular LSTM as it incorporates physical quantities of interest such as water breakthrough. The model is conceived with two consecutive networks: one network for breakthrough time estimation with output being fed into a second network that reconciles other features important for oil production prediction. The EnKF enhanced LSTM has the capability of performing data assimilation based on real time production data, thus providing a way to update our model constantly.
In this work, we first show the methodology applied to the two-phase water flooding reservoir with five spot production scenarios. Then we conduct the comparison of Bayesian optimization tuned cascaded LSTM vs. standard LSTM. Finally, we showcase the usefulness of Ensemble Kalman filter in improving and updating current model.
The method presented in this paper uses RNN (specifically cascaded LSTM) to learn the pattern from sequence data and identify the reservoir simulation proxy model, which can accurately predict surface production rate and water cut without the state variable calculation. The study also shows improved accuracy (over standard methods) for EnKF trained RNN and its capability of updating flow rate prediction based on new observation data.