FREQUENCY DRIVEN SCREW COMPRESSORS UNLOAD STATE PREDICTION IN MODERN AIR CONTROL SYSTEM

Paper refers to authors’ previous studies in this subject and proposes simulation method for neural networks adaptation and testing in field of compressed air systems. Simulated supervisory controller system using real time gathered data allowed to test solutions prior to commissioning and minimize industrial implementation of the control system faulty decision risk. Since with limited amount of gathered data creating digital twin is very difficult (because of multiple physical unknown factors) a different approach is proposed. The subject of a study is container-based air station consisting of three screw compressors equipped with permanent magnet motors with different electrical power. Data was acquisitioned over one-month period. For aiding supervisory control system a straightforward solution was used consisting of neural network based on Long Short-Term Memory (LSTM) architecture, featuring two LSTM layers with a dedicated loss function. The study showed that it was possible to correctly detect around 90% of negative operating states however a longer time for consideration appeared to be less reliable.