Closed-loop forced heat convection control using deep reinforcement learning.

• A deep reinforcement learning based forced convection control method is proposed. • A novel deep Q-network (DQN) integrating with three advanced techniques is used. • The method shows efficiency and robustness on learning nonlinear control strategy. • The trained agent has generalization ability on geometric configuration. In this paper, deep reinforcement learning (DRL) is applied on forced convection control of conjugate heat transfer systems governed by the coupled Navier-Stokes and heat transport equations. A novel value-based deep Q-network (DQN) integrating with three advanced techniques is utilized for identifying underlying heat and mass transfer mechanism in the interaction of close-loop active control. The effectiveness and feasibility of DRL based forced convection control is firstly demonstrated by studying a two-dimensional cooling problem, where a single heat source immersed in an open cavity. A more complex testbed with multiple immersed heat source and multiple degree-of-freedom control is investigated further, which shows outstanding capability of DRL algorithm to learn strong nonlinear control strategy, and compared to conventional control method, the novel DRL approach has ability to obtain better cooling effect of about 8 degrees lower. Then the robustness of the method is verified by applying trained agent on several situations with unknown physical conditions and adding additional customized control requirements/aims. Moreover, sensitivity analysis confirms that the trained agent has a certain generalization ability on geometric configuration, which strengthens the confidence of applying the DRL-based active heat transfer control method on practical applications. Current research demonstrates the efficiency and applicability of the DRL based active thermal control on strong-nonlinearity system, and also encourages further investigations on more complex and practical problems. [ABSTRACT FROM AUTHOR]