Machine learning application in deep-level gold mine chiller control improvement

Abstract

The cost of electricity in South Africa has increased substantially, causing deep-level gold mines to reduce energy demand and operational costs. Refrigeration systems account for approximately 23% of deep-level gold mine energy budgets, and reducing refrigeration energy demand presents an opportunity to significantly reduce operating costs.

Dynamic water temperature set-point control has been used to reduce refrigeration system energy consumption. This method involves varying the set point for the bulk air cooler supply water temperature based on important parameters such as ambient conditions, time of day and demand statistics.

Machine learning has been widely applied and studied in building refrigeration control and analysis to create intelligent control systems that implement dynamic chiller temperature setpoint control. However, this solution has not yet been thoroughly explored in mining refrigeration systems, leading to a lack of knowledge on model selection, model evaluation and data requirements. The objective of this study was to expand the understanding of machine learning methods in the control and characterisation of deep-level gold mine refrigeration systems. This objective involved evaluating machine learning algorithms, including those that have been previously implemented and those that have not yet been applied to deep-level gold.

Deep-level gold mine refrigeration system data was gathered, processed, and used to create features and train machine learning models. The study compared support vector machines, multiple linear regression, artificial neural networks, random forest regression and ensemble learning methods. The model performance was analysed through k-fold cross-validation, and the effect of data dispersion on the model performance was analysed to reveal best practices in data handling.

The best performing models were investigated further to determine their suitability for use in chiller water set-point control. The investigation involved using the trained machine learning models to suggest new water temperature set points and compare them to the current water temperature supply to estimate the energy demand reduction. It was estimated that using this method would save 1.42 MWh and 2.65 MWh of electricity daily on the two case study systems. This would reduce the annual operational cost by R385 000 and R700 000 on the two case study systems, respectively. mine refrigeration characterisation. The study further investigated the effect of data dispersion when creating characteristic models of refrigeration systems.