Improved control of compressed air networks using machine learning

Abstract

Mining companies in South Africa are under financial strain and so need to minimise expenses to stay profitable. Improving the efficiency of compressed air networks could lead to a significant reduction in electricity costs, which will benefit mines. Some mines rely on a central control room that monitors and controls compressed air networks remotely. Due to the complexity of compressed air networks, some software-based solutions have been used to assist with improved supply control. The Dynamic Compressor Selector (DCS) controller is an existing compressor control system that has been used to achieve electricity cost savings. This controller relies on accurate matching of compressed air supply and demand to improve control of a compressed air network. Simulation software has also been applied to evaluate and implement energy-saving initiatives on mining compressed air networks. The current software-based solutions are based on fundamental principles and physical relationships. These solutions rely on real-time instrumentation availability and accuracy to function properly, which is problematic on mines. Furthermore, due to the complexity of some mining compressed air networks, expert knowledge is often required for the setup of existing solutions. These challenges could delay or hinder the implementation of existing solutions. From the literature, machine learning models such as artificial neural networks (ANNs) have been applied for the prediction of parameters in complex systems. A need was identified to investigate the feasibility of applying a machine learning model for the prediction of parameters in a mining compressed air network. A need was also identified to apply a machine learning model for real-time prediction to improve compressed air network control.

A platinum mining complex with a large compressed air network was identified as a suitable case study. A method, based on common steps in research, was applied to develop an ANN for predicting pressures on the mining compressed air network. The developed ANN achieved a mean absolute percentage error (MAPE) of 4.20% for all predicted parameters, thus verifying it to be accurate. This application of machine learning was combined with a real-time energy management system (EMS) to develop a compressed air network control solution that could improve the compressed air network supply. This was achieved by implementing the EMS to trigger the start and stop commands of compressors and machine learning to predict pressures and verify whether suggested compressor combinations would work. The developed compressed air network control solution was implemented at the case study mining complex. An accurate simulation model was developed and applied to assist in validating that the solution could improve compressed air network control. The simulation software helped to identify some shortcomings of the solution, which were addressed by updating the solution. A real-time test was performed which indicated a potential electricity cost saving of R 1.6 million per annum. The application of an ANN to predict pressures in a mining compressed air network successfully mitigated issues faced regarding instrumentation and complexity. Implementing this system to aid control room operators managed to achieve electricity cost savings whilst delivering the required pressure to critical demand points. This validates the possibility of applying a machine learning model for real-time prediction to improve compressed air network control.