Abstract: In the tobacco loosening and conditioning system with non-linearity and time lag, a model predictive control method is proposed to address the problems of the traditional control methods with low prediction accuracy and low control stability. First, to improve the model prediction accuracy, the convolutional neural network is integrated with the gated recurrent unit network according to the NARMAX model to build a prediction model with multi-input and multi-output systems of the loosening and conditioning process. Then, a teaching and learning salp swarm optimization algorithm is proposed to perform rolling optimization, ensuring the recirculated air temperature and the outlet moisture consistently and accurately meet the set values. The results show the model achieves synchronized control of recirculated air temperature and outlet moisture, and performs better prediction and control than other models. The average root-mean-square error of the prediction model is 0.027, the average overshoot of the controller is 0.118%, and the average CPK value is as high as 2.45.