Manuscript Title

Field development planning and economic analysis require reliable forecasting of bitumen production. Forecasting at the field level may be done using reservoir simulations, type-curve analysis, and other (semi-)analytical techniques. Performing reservoir simulation is usually computationally expensive and the non-uniqueness of a history-matched solution leads to uncertainty in the model predictions and production forecasts. Analytical proxies, such as Butler's model and its various improvements, allow for sensitivity studies on input parameters and forecasting under multiple operational scenarios and geostatistical realizations to be conducted rather quickly, despite being less accurate than reservoir simulation. Similar to their reservoir simulation counterparts, proxy models can also be tuned or updated as more data are obtained. Type curves also facilitate efficient reservoir performance prediction; however, in practice, the performance of many SAGD well-pairs tends to deviate from a set of pre-defined type curves. Historical well data is a digital asset that can be utilized to develop machine learning or data-driven models for the purpose of production forecasting. These models involve lower computational effort compared to numerical simulators and offer better accuracy compared to proxy models based on Butler's equation. Furthermore, these data-driven models can be used for automated optimization, quantification of geological uncertainties, and "What If" scenario analysis. This paper presents a novel machine learning workflow that includes a predictive model development using the random forest algorithm, clustering, Bayesian updating, Monte Carlo sampling, and genetic algorithm for accurate forecasting of real-world SAGD injection and production data, and optimization. The training dataset involves field data that is typically available for a SAGD well-pair (e.g. operational data, geological, and well design parameters). Just as importantly, this machine learning workflow can update predictions in real-time, be applied for the quantification of the uncertainties associated with the forecasts, and optimize steam allocation, making it a practical tool for development planning and field-wide optimization. To the best of the author's knowledge, this is the first time that machine learning algorithms have been applied to a SAGD data set of this size.