Predicting EUR in unconventional tight-shale reservoirs with prolonged transient behavior is a challenging task. Most methods used in predicting such long-term behavior have shown certain limitations..
Available unconventional tight-shale reservoir data is analyzed by an artificial recurrent neural network (RNN) architecture such as LSTM and a decision tree method called XGBoost is used and predictions are obtained.
The forecasts from the LSTM and XGBoost machine learning models and the physics-based reservoir simulation models are compared. Four different reservoir simulation models have been created for different hydrocarbon types; these are dry gas, condensate, light oil, and volatile oil, respectively. An analysis of the comparison shows that the LSTM and XGBoost machine learning models have some forecasting capabilities, but this capability is highly dependent on the input data. In addition, predictions have also been made based on the decline curve analysis (DCA). A detailed analysis is done using the forecast results from LSTM, XGBoost and the DCA.
Machine learning applications are growing rapidly in the oil and gas industry. However, this does not mean every situation needs a machine learning solution. As per this study, classical methods might perform better and gives faster results. Notably, in case of limited data, the machine learning methods can underperform, and the importance of traditional techniques arise again. This study uses only synthetic/publicly-available data to generate data through reservoir simulation runs built with publicly-available Eagle Ford-like data to for analysis with different operational scenarios.
Number of Pages
Looking for more?
Some of the OnePetro partner societies have developed subject- specific wikis that may help.