Deep reinforcement learning (DRL) for process control is one of challenging applications of state-of-art artificial intelligence (AI). It has been proven that DRL has a strong ability to learn superior strategies for complex tasks such as igo, video game playing, automated drive, and so on. For many years, model predictive control (MPC) has been a main successful control method in industrial control. And the authors have proposed an approach for MPC combined with DRL to achieve more precise and adaptive control. In this paper, we extend the approach to avoid oscillation of manipulated input. We propose a novel policy which blends actions according to the trained Q-function in DRL.