Deep Learning Model Comparison Study on Temperature Control in Electric Facilities

Control systems have evolved along with industrial electric facility development and are very important in the manufacturing processes. This paper compared various temperature control models: proportional, integral and derivative control (PID) model, model-predictive control (MPC) model, and long short-term memory (LSTM) models. The micro temperature control lab (TClab) was using for evaluating the performance of those control models. First, temperature predictive model for model predictive control (MPC) was proposed. We compared physics-based energy balance model, first order plus dead-time (FOPDT) model and neural network-based LSTM model for the predictive model of the future temperature of TClab. The model was used to identify future temperature by using the heater's response and current temperature. The FOPDT and LSTM models have 97-98% of accuracy as the predictive model. Second, the performance of the various control models was compared with a specific temperature profile. The conventional PID control has 1.988 & DEG;C/sec errors between the target temperature and the actual temperature in the designated temperature profile. The average error value of the model is lower (1.766 ?/sec errors) due to the predictive model and optimization. The accuracy of MPC is 12.4% higher than that of the PID model. Neural network based deep learning control model is clearly more accurate the PID model. When LSTM model was trained with optimal parameters, The average error between the target temperature and the actual temperature of LSTM model is 1.752 ?/sec. This is 13.5% higher accuracy than that of the conventional PID model. Our experiment results well explained the difference of various temperature control models and they showed that the LSTM model can greatly improve the accuracy with optimized parameters and showed the possibility for the next intelligent control system with massive amounts of data accumulated.