Crude Predictions

Steam-assisted gravity drainage (SAGD) has been widely shown to enhance the recovery of heavy oil in an effective manner. However, analyzing the performance of SAGD in heavy oil reservoirs usually involves time-consuming and tedious static and dynamic numerical simulations, making it impractical for decision-making and forming predictions. However, our design solution overcomes this through deep learning techniques, providing a viable backbone for both data analysis and decision-making, expediting the SAGD workflow through our easy-to-use user interface.

Our design is innovative in its crossover between front-end development and advanced machine learning techniques, bridging the gap between data sciences and artificial intelligence. Furthermore, through deep learning we successfully harness neural networks in providing time-series forecasting for a unique utility in optimization, in addition to making predictions.

What makes our model effective, not only as a proof-of-concept for the usage of deep learning in SAGD, but enables it to excel as a design tool, is its huge performance leap in simulations and analysis. The workflow of our design solution foregoes the tedious manual calibration involved in traditional solutions that commonly employ mass and energy balance systems by instead utilizing deep learning. For the initial trade-off of training a model, our workflow enables the rapid generation of predictions and visualizations, outperforming computationally intensive simulations. Lastly, our design solution highlights the potential of machine learning and artificial intelligence in SAGD, demonstrating the ability to make hands-free inferences without the dependency for predetermined information.

In validating our design solution, we have undergone multiple benchmarks of our deep learning model, employing iterative development to narrow down the most advantageous solution through the usage of prototyping technologies such as Kaggle, Jupyter, and Weights and Biases. Moreover, we have pitted our solutions against traditional techniques such as Auto-regressive Integrated Moving Average (ARIMA), and several other baselines in time-series techniques through Darts in Python.

Ultimately, while the main goal of our design challenge was in creating a proof-of-concept solution to optimizing SORs in SAGD through deep learning techniques, our final Long Short-Term Memory model has proven to be a strong contender in generating robust and accurate predictions over a relatively large time-frame. While deep learning frameworks are unlikely to replace mass and energy balance systems and other simulation tools, in analyzing the complexity of SAGD in heavy oil reservoirs, our project demonstrates its strong ability to create inferences, proving its ability for making predictions in a new environment, and demonstrating its use case as a supplemental resource in SAGD.