Abstract: To improve the moisture control accuracy in output tobacco from loosening and conditioning machines, an intelligent moisture predictive control method based on moisture absorption mechanism during conditioning and deep reinforcement learning was proposed. Firstly, the Long Short-Term Memory (LSTM) model was employed as the base model to predict moisture content in output tobacco. Secondly, an improved prediction model was formed by integrating the moisture absorption mechanism with the LSTM model. Then LSTM-Actor-Critic (LAC) model was built by embedding the predictive model into the Actor-Critic network architecture through deep reinforcement learning to achieve dynamic optimization control. Finally, based on the model prediction and control results, both the prediction and the control models were iteratively adjusted and optimized. The results showed that, compared with the conventional proportional integral derivative (PID) control method: 1) the proposed intelligent control strategy was able to obtain a relatively small range of the standard deviations (σ) of the moisture content or temperature of the output tobacco from the production data of the vast majority of 50 consecutive batches selected; 2) the process capability (CP) index of the moisture content in the output tobacco increased by 48% with a 17% reduction in standard deviation, while the CP for temperature of the output tobacco increased by 17% with a 16% reduction in standard deviation. Both the moisture content and temperature fluctuated slightly, indicating more precise control. This method achieved collaborative optimization control for both the temperature and humidity indexes without compromising the control effectiveness in the output tobacco, providing technical support for intelligent control of leaf moisture content during loosening and conditioning process for cigarette manufacturing.