The addition of Machine Learning (ML) into traditional petroleum engineering workflows has gained influence over the years. Incorporating ML models can speed up computations significantly, and hence they are vital in situations when fast decisions are needed. It takes a long time to history match and update the reservoir model. Therefore, the usage of physics-based simulators is limited. The implementation of ML models addressed this issue.
The goal of this thesis is the development of the ML model that can forecast oil production. The main focus is put on utilizing the Long Short-Term Memory (LSTM) cells and 1-D convolutions for forecasting. LSTM’s are better equipped to handle this problem compare to simple Neural Networks (NN) because each data point is appropriately treated as a time series observation instead of an independent entity.
The work has been conducted on the synthetic dataset generated in the Petrel simulator on the 2D heterogeneous reservoir model. The production rates together with the complex injection schedule have been extracted and analyzed using purely data-driven ML model.
Different combinations of input parameters have been investigated to find the optimum setup of features for production forecasting. The univariate ML models that use only past oil production data as an input have demonstrated reasonable performance in short-term predictions (several days). For the longer-range forecasts (up to 1 year), the ML model will need the injection rates as an input in addition to production rates. That complex models require more computational power, but they can give longer forecasts comparing to univariate ML models.