As data centres become integral to modern infrastructure, their energy consumption, particularly in cooling systems, presents a critical challenge for sustainability. This paper addresses this issue by applying time-series machine learning models to forecast the performance of a highly efficient 100 kW evaporative cooling system applied in a real-world data centre. Using a dataset spanning 4 months, we developed and optimised two predictive models based on XGBoost and Random Forest, to estimate cooling capacity and Coefficient of Performance (COP). Initial results showed suboptimal performance, with the XGBoost model achieving a Mean Absolute Error (MAE) of 1.34 for cooling capacity and 6.50 for COP, alongside a negative R-squared, indicating poor fit. However, after hyperparameter tuning, the Random Forest model significantly improved the predictions, achieving an MAE of 0.39 and an R-squared of 0.85 for cooling capacity, and an MAE of 2.21 and an R-squared of 0.54 for COP. These findings underscore the potential of these models to optimise cooling efficiency, offering valuable insights for reducing energy consumption and operational costs in data centre operations. This research paves the way for more sustainable data centre designs and operations across diverse climatic conditions. Practical application: The predictive models developed in this study enable building environment professionals to optimise data centre cooling systems. By accurately forecasting cooling capacity and Coefficient of Performance (COP) under varying environmental conditions, these models allow for proactive adjustments to cooling strategies, ensuring efficient operation and minimising energy waste. This research provides a practical tool for enhancing the sustainability of data centres, directly supporting industry efforts to meet stringent energy efficiency targets and reduce the carbon footprint of critical infrastructure.