Real-time realization of Dynamic Programming using machine learning methods for IC engine waste heat recovery system power optimization

This study aims to present a method for real-time realization of Dynamic Programming algorithm for power optimization in an organic Rankine Cycle waste heat recovery system. Different from existing studies, for the first time machine learning algorithms are utilized to extract the rules from offline Dynamic Programming results for optimal power generation. In addition, for the first time a single state Proper Orthogonal Decomposition and Galerkin Projection based reduced order model is combined with Dynamic Programming for its high accuracy and low computation cost. For a transient driving cycle, Dynamic Programming algorithm is utilized to generate the optimal working fluid pump speed. A total of eleven state-of-art machine learning algorithms are screened to predict this pump speed. Random Forest algorithm is then selected for its best pump speed prediction accuracy. A rule-based method is added to the Random Forest model to improve energy recovery. As one of the main discoveries in this study, in the rule extraction process, the Random Forest model reveals that the time delayed exhaust gas mass flow rate and exhaust temperature improve the rule extraction accuracy. This observation points out the difference between steady state and transient optimization and that the steady state optimization results do not necessarily hold true in transient conditions. Another key observation is that Random Forest - rule-based method retrieves 97.2% of the energy recovered by offline Dynamic Programming in a validation driving cycle. In addition, the inclusion of rule-based method significantly increases the Random Forest model's energy recovery from 66.5% to 97.2%. This high accuracy means that the machine learning models can be used to extract Dynamic Programming rules for real-time application.