Hybrid CNN–LSTM–Linformer Driven Adaptive NMPC for Environmental Control in Pig Housing

Accurate prediction and energy-efficient control of pig-house environments remain challenging because temperature, relative humidity, NH3, and CO2 are nonlinear, strongly coupled, and affected by ventilation and seasonal conditions. This study proposes an intelligent prediction-control framework integrating hybrid deep learning prediction with nonlinear model predictive control for multivariable pig-house environmental regulation. Several hybrid deep learning models were compared, including CNN-LSTM-standard transformer, CNN-LSTM-lightweight transformer, and the proposed CNN-LSTM-linformer-style model. The CNN-LSTM-lightweight transformer achieved the highest overall prediction accuracy, whereas the proposed CNN-LSTM-linformer-style model provided the most compact structure by using separable convolution, global average pooling, and linformer-style attention, reducing training time and memory usage by approximately 53% compared with the CNN-LSTM-standard transformer. The prediction model was integrated with FLC, NMPC, and ANMPC for closed-loop ventilation control. ANMPC adjusts control weights online according to environmental deviations to balance environmental regulation and energy use under disturbances. In the nominal closed-loop simulation, NMPC and ANMPC reduced ventilation energy consumption by approximately 44% compared with FLC, while ANMPC achieved a 4.35% lower NH3 steady-state error and a 3.5% faster NH3 recovery response than NMPC under disturbance conditions. In 24 -h pre-field verification, NMPC and ANMPC reduced energy consumption by 35.8% and 27.4%, respectively, while maintaining pollutant safety.