The reliability of building energy prediction results is often threatened by lack of comprehensive and continuous data, especially when dealing with older buildings that are not furnished with building energy management systems. In order to investigate the performance of building energy prediction models under limited data, this paper utilises four distinct machine learning methods - decision tree (DT), support vector machine (SVM), random forest (RF) and voting regression (VR) to predict energy consumption of the Chemistry building of a prominent higher institution, based on just meteorological data. The results indicate that SVM is unable to accurately predict building energy consumption based on the prescribed input variables alone. However, in general, DT, RF and VR offered far more reliable and accurate energy consumption prediction outcomes with the same training and testing data sets. More specifically, RF outperformed all other included methods. It was also observed that the extension of the time span for the training data sets offered insignificant improvement to the prediction accuracy as postulated by some earlier studies. With regards to overall generalisation capability, VR outperformed all approaches, with outcomes from RF also marginally better than those from DT.