Probabilistic machine learning for enhanced chiller sequencing: A risk-based control strategy

Abstract

Multiple-chiller systems are widely adopted in large buildings due to their high flexibility and efficiency in providing cooling capacity. A reliable and robust chiller sequencing control strategy is crucial to ensure the energy efficiency and stability of the multiple-chiller systems. However, conventional chiller sequencing control strategies are usually based on real-time measured cooling load without considering the cooling load changes in the following hours. Conventional rule-based strategy may result in unnecessary switching on and off, leading to energy waste and impairing system stability. Therefore, this study proposes a robust chiller sequencing control strategy that utilizes probabilistic cooling load predictions. 1h-ahead probabilistic cooling load prediction in the form of the normal distribution is made using natural gradient boosting (NGBoost). Compared to conventional machine learning algorithms, NGBoost can predict not only the future cooling load but also the uncertainty of the predicted cooling load, which enables the load prediction to handle the uncertainties associated with the data/measurements adequately. A novel risk-based sequencing strategy is developed based on the probabilistic cooling load predictions. The data experiment shows that the proposed strategy can significantly improve the stability and reliability of the chiller plant by reducing the total switching number by up to 43.6 %.