In practical industrial-scale paddy drying production, manual empirical operation is still widely used for process control. This often leads to poor uniformity in the moisture content distribution of discharged grains, affecting product quality. Model Predictive Control (MPC) is considered the most effective control method for paddy drying, but its implementation in industrial-scale drying is hindered by its high computational cost. This study aims to address this challenge by proposing a deep-learning-based model predictive control (DL-MPC) strategy for paddy drying. By establishing a mapping relation between the inlet and outlet paddy moisture content and paddy flow velocity, a DL-MPC strategy suitable for multistage counter-flow paddy drying systems is proposed. DL-MPC systems are developed using long short-term memory (LSTM) neural networks and trained using datasets from single-drying-stage and multistage drying systems. Simulation and analysis are conducted, followed by verification experiments on a 5HNH-15 multistage counter-flow paddy dryer. The results show that the DL-MPC system significantly improves computational speed while achieving satisfactory control performance. The predicted paddy flow velocity exhibits a smooth variation and matches field data obtained from multiple transition points, confirming the effectiveness of the designed DL-MPC system. The mean absolute error between the predicted and actual paddy moisture content under the DL-MPC system is 0.190% d.b., further supporting the effectiveness of the control system.
Keywords: computational cost; deep-learning; industrial-scale drying; model predictive control; paddy drying.