Data availability has triggered the development of implementing artificial intelligence methods on building energy consumption analysis of prediction. Recent studies have also continuously proved the excellent performance of artificial intelligence methods in this regard. However, there is a lack of investigation of the impact of building types on model prediction performance, especially for buildings without obvious energy usage patterns. In this study, the use of long short-term memory networks (LSTMs) model is proposed to predicted energy consumption of classroom, library and student hall buildings. The results indicated that LSTM showed the best performance when predicting building with obvious energy usage pattern. The accuracy was impeded when it came to buildings that did not show obvious usage pattern. In this study, better prediction results can be achieved when feeding LSTMs model with longer training data sets.