This article focuses on closed-loop active control of flow past a circular cylinder, where the reinforcement learning (RL) is utilized to obtain the intelligent self-adaptive control strategy. In the closed loop, six velocity sensors are used to provide feedback signals and establish the state space, and the real-time adjustable rotations are used to mediate the flow surrounding the cylinder such that the lift fluctuations can be mediated. In order to acquire high-fidelity data of flow dynamics in both temporal and spatial space, the lattice Boltzmann method is adopted, which couples the multi-block mesh partition technique so as to obtain a sufficiently large computational domain while providing sufficiently fine mesh in the near field, as well as multi-force immersed boundary method for accurately dealing with the rotating wall boundary. Based on this control loop, as well as the numerical environment, the lift fluctuation of an isolated cylinder is mitigated by 92.5%, the drag fluctuation is mitigated by 44.3%. In the meantime, the recirculation bubble is enlarged by 36.4%, implying smaller pressure drop between the front and rear side of the circular cylinder. Through the local linear stability analysis, the length of absolutely unstable region is elongated by 36%, while the frequency of the least stable perturbation deviates from the vortex shedding frequency. Through comparisons of uncontrolled and RL-controlled instantaneous pressure fields that corresponds to the maximum lift, it is directly revealed that the RL-based control successfully balances the pressure distributed near the upper and lower surface. Furthermore, based on the transfer learning, the lift fluctuation of a circular cylinder interfered by an upstream cylinder is mitigated by 95.9%, on the cost of a doubled mean drag. Through this study, an intelligent flow self-adaptive control for lift mitigations is proposed, and a detailed case study is provided for future flow control of more complicated flows.