Machine learning approaches for modeling spatio-temporal patterns in subsurface energy production

Abstract:

Subsurface energy production is governed by complex, multiscale processes which are difficult to simulate using current tools due to computational cost and inflexibility. Moreover, current methods either simplify or ignore the added complexity associated with modeling these processes in the context of fractured, porous media. We implement and compare the performance of three machine learning models that explicitly leverage the geological properties associated with fractured, porous media. The data were simulated using the MRST-shale framework, an open-source approach to numerical modeling of shale transport and storage that allows for natural fractures at multiple scales. The simulated data represent the cumulative production and spatial pressure distribution timeseries associated with a horizontal well with hydraulic fractures. Additional geological properties were simulated to serve as predictive input features to the models, including permeability, porosity, hydraulic conductivity, bottom hole pressure, and fracture aperture. We compared the performance of three candidate models implemented in our research, including an artificial neural network (ANN), convolutional long-short term memory network (convLSTM), and a hybrid convolutional neural network-LSTM model (CNN-LSTM). All three models are able to leverage the spatio-temporal relationships present in the data. The ANN is well-suited for use with spatio-temporal data since it learns non-linear relationships well and makes no assumption regarding stationarity in time and space. The convolutional components of the other two models extract the spatial patterns in the data and treat them as features, and the LSTM components of those models learn long-term dependencies in these features over time. The three models predicted cumulative production and the spatial pressure distribution reasonably, with varying degrees of accuracy and efficiency. We will discuss the relative strengths and weaknesses of the three models by performing the sensitivity of key hyperparameters and model architectures and present an optimal model architecture through the hyperparameter optimization.