The prediction of the production performance of the Steam Assisted Gravity Drainage (SAGD) process under different conditions is a significant task. The conventional analysis tools for the SAGD process, numerical simulations, for instance, are time-consuming with high computational costs. In addition, studies on the dynamics of the SAGD process are still rare, which hinders the exercises to systematically characterize the entire SAGD process, objectively reveal the resultant production behavior, and reliably provide the basis for practical decisions. As a result, it is important to find a feasible method that can dynamically predict SAGD production performances. This thesis develops a number of data-driven model-based workflows for the dynamic prediction of SAGD production performance under different conditions by various machine learning algorithms. First, optimal data-driven models are set up based on different data groups to predict SAGD production performance with the relevant parameters. Second, through a comprehensive analysis of the SAGD production period, three SAGD production period indicators are collected based on the simulated production performances. Data-driven models are constructed to establish the relationship between those SAGD production period indicators and the relevant parameters. Third, a study is conducted to analyze the SAGD production performance with infill wells based on data-driven models. Next, the effects of heterogeneous features on the SAGD process are discussed and the relevant parameters of those features are considered in the data-driven model prediction workflow. Finally, a data-driven model-based implementation is built to predict future production performance by correlating previous time-series data. Based on the real field data which is modified by several data preprocessing methods, a series of different data-driven models are employed to approximate the inverse relationship between the future production performances and the previous relevant information. Those data-driven models result in complementary workflows that can successfully serve as predictive tools for SAGD processes with high predictive accuracy. The developed data-driven workflow can also be extended in different SAGD problems derived from other data sets. But the predictive ability of a data-driven model largely depends on the data size and feature diversity of the data sample used for training, and different machine learning algorithms selections.