Integration of the NNARX Model in Python/Keras for Predictive Microclimate Control of Industrial Premises

Abstract

The paper presents an approach to integrating an NNARX neural network model into the Python/Keras software environment for implementing predictive control of the microclimate in industrial premises as part of a cyber-physical system. The choice of the NNARX architecture as the predictive core of the control system is justified, and the specifics of its software implementation using the TensorFlow/Keras library are described. The structure of the software is presented, including a module for data acquisition and preprocessing of sensor data, a prediction module based on the trained NNARX model, and a control action generation module using a fuzzy logic controller. The results of experimental studies are provided, confirming the effectiveness of the proposed solution: the root mean square error of temperature prediction is 0.302 °C, the mean absolute error is 0.248 °C, and the temperature deviation from the setpoint under real operating conditions does not exceed 0.5 °C. It is shown that integrating the predictive model into a closed-loop control system reduces the stabilization time by 10–15% compared to classical PI controllers.