Application of Deep Learning for Understanding Dynamic Well Connectivity
Abstract
Artificial intelligence and machine learning have transformed many industries. However, the oil and gas industry is lagging in AI adaption. Currently, with the low oil prices and a considerable performance gap in the oil and gas industry, companies are looking for new ways to improve their operational efficiency. We have a promising proposition to apply state-of-the-art deep learning algorithms to reservoir management to understand the dynamic well-connectivity of reservoirs.
The deep learning algorithms, Long Short-Term Memory (LSTM) and Gated Recurrent Network (GRU) have a successful history in applying to many complex sequential and time series problems. In this thesis, we formulate the problem as a supervised deep-learning problem and use the LSTM and GRU algorithms to train a model that could identify well-connectivity. We model a single layer LSTM and GRU model with cell states (memory cells) to match the historical production rate by providing the input as the injection rate. For training purposes, we split the available data into training, validation, and testing datasets. We have also applied the Early Stopping criteria to prevent the underfitting and overfitting of the model. In the Early Stopping criteria, we monitor the error of the model in the validation dataset and select the model with minimum error in the validation set. The hyperparameters, cell size and window size, are optimized by the Grid Search method.
Although deep learning models work well, they are black-box models and do not provide any interpretability between the input features and the outputs. So, we have applied the Permutation Feature Importance method for the model interpretability. This method calculates the reservoir connectivity by permuting or shuffling the inputs (water injection rates) one by one to the trained model and calculating the increase in the root mean square error (RMSE).
The deep learning workflow is applied to two cases: First, to a synthetic high permeability streak reservoir for proof of concept; second, to a field-scale model of the Brugge reservoir. The normalized streamline flux allocation factor validates the reservoir connectivity from the deep learning model.
Citation
Kareepadath Sajeev, Shyam (2020). Application of Deep Learning for Understanding Dynamic Well Connectivity. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /192552.