We demonstrate real-time surrogate-model-based temperature control of a co-rotating twin-screw extruder. A one-dimensional mechanistic extruder model describing mass transport, pressure dynamics, and heat transfer is extended and experimentally validated. To enable real-time optimization, neural network surrogate models in a nonlinear autoregressive with exogenous inputs formulation are trained on simulation data from the mechanistic model. Given the sparsity of available measurements, a surrogate is combined with a particle filter for robust state observation, reconstructing the complete melt and barrel temperature profiles from only three melt temperature sensors. A Bayesian optimization algorithm is proposed for optimal sensor placement that maximizes state observation accuracy under hardware constraints. A surrogate is integrated into a model predictive control formulation that simultaneously maximizes melt temperature, tracks throughput references, and enforces constraints imposed by material characteristics and process limitations for industrially relevant product cases. The complete closed-loop scheme is validated experimentally on a pilot-scale extrusion process.
This paper studies the modeling and optimization of microbial fermentation processes based on metabolic networks. Existing modeling methods rely heavily on large amounts of data and expert knowledge. They often struggle to capture the strong nonlinear dynamics of fermentation processes, and their predictive performance decreases under changing operating conditions. To improve model accuracy and generalization, we propose the metabolic reaction neural network (MRNN) method that embeds microbial metabolic reaction information into the network structure. The design enables automatic learning of microbial metabolic evolution. Compared with conventional mechanistic models, MRNN reduces the dependence on prior knowledge and can learn complex reaction networks and metabolic kinetics. Compared with general neural network models, MRNN adopts a neural ODE framework and incorporates stoichiometric matrices and reaction pathway information to preserve metabolic constraints while learning nonlinear fermentation dynamics. An optimization framework is also developed to improve product synthesis by optimizing substrate feed rate and pH based on pathway-level metabolic flux objectives. A penicillin fermentation case study verifies the effectiveness of the proposed method for microbial fermentation modeling and metabolic flux optimization.
Natural gas pipeline networks are increasingly required to respond to fast fluctuations in integrated energy systems. A major obstacle to real-time pressure control is the propagation delay of pressure waves, which is often neglected or absorbed into inertial dynamics in conventional control-oriented models. This study develops a delay-aware transfer-function model and reformulates it into an extended state-space representation by explicitly separating propagation delay from inertial response. Based on this model, and to constrain the rate of change of control actions and improve actuator smoothness under transient disturbances, an incremental-input delay-compensated model predictive controller (DC-MPC) is developed and compared with delay-ignored MPC (DI-MPC) and conventional operating strategies. Step-response validation against a high-fidelity transient gas-network simulator shows that the proposed model captures delayed pressure-wave propagation more accurately than the classical second-order transfer-function model, reducing the cumulative pressure-response deviation by 49.48%. In the simple pipeline case, DC-MPC achieves a mean absolute percentage deviation of 1.631% while reducing the total variation of compressor outlet pressure by 32.38% relative to DI-MPC, thereby suppressing overshoot and oscillatory control actions. In the 47-node natural gas network case, DC-MPC provides smoother control and lower tuning sensitivity, whereas DI-MPC achieves better tracking only at the expense of aggressive actuator movements and higher tuning sensitivity. These results demonstrate that explicitly accounting for propagation delay, together with incremental-input control design, improves prediction consistency, control smoothness, and the engineering feasibility of real-time pressure control in natural gas pipeline networks.
The wastewater treatment process (WWTP) is a critical component driving sustainable industrial development. The monitoring and control of the dissolved oxygen (DO) concentration play a vital role in the WWTP. An adaptive utility function, a control barrier function, and an adaptive dynamic programming approach based on a lifelong learning mechanism are constructed in this paper to realize the online control of the DO concentration. First, the magnitude of the tracking error is adopted for control input modulation to enhance the adaptability of the utility function. Second, an adjustable control barrier function is employed to constrain the tracking error within a specific range. Then, a constraint term is introduced into the process of updating the weights by using weights of the critic and action networks from the previous time step, which enhances the approximation performance of the proposed controller. A stability criterion related to the learning law is also established and the uniform ultimate boundedness of the weight estimation error is proven. Finally, the established method is applied to the WWTP and its tracking performance is validated under various weather conditions.
Achieving rapid heating, low overshoot, and spatial temperature uniformity simultaneously remains challenging for large-volume industrial ovens because of strong thermal inertia, transport delay, nonlinear thermophysical behavior, and nonuniform convective heat distribution. This paper presents a control-oriented oven design and a practical coordinated control strategy for heater power and airflow regulation. A 270-L forced-convection oven with rotationally symmetric baffle openings and isomorphic airflow pathways is first developed to improve the open-loop thermal characteristics, thereby facilitating closed-loop regulation of the transient heating response and spatial temperature uniformity. A lumped-parameter nonlinear thermal model is then derived from energy conservation and expressed in a compact state-space form for predictive control. Based on this model, a finite-control-set model predictive controller regulates the quantized heater power through direct enumeration under explicit temperature and input constraints. Meanwhile, a model-free adaptive controller regulates the fan mass-flow command without requiring an explicit flow-field model. The complete scheme is implemented on an industrial PLC and experimentally validated on the developed oven platform. Rolling-horizon prediction validation based on successive 30-s-ahead predictions yields an RMSE of 2.01 °C and an R2 of 0.96. In closed-loop experiments, the oven reaches 95% of the 85 °C setpoint in 733.66 s, with an average steady-state deviation of 0.19 °C and a maximum overshoot of 0.46 °C. Under a severe door-opening disturbance, the temperature and spatial-uniformity indicators recover to their prescribed ranges within 113.58 s and 21.74 s, respectively. Comparative experiments against a model-optimized two-degree-of-freedom PID controller and constant-airflow operation further demonstrate the effectiveness and practical implementability of the proposed coordinated framework.
