Buildings are major energy consumer worldwide, accounting for 20%-40% of the total energy production. Efficient energy management in buildings is important for effective energy saving. In this study, we propose and develop a five-step machine learning and artificial intelligence approach for high-precision energy forecasts in buildings. First, a feature database of potential energy predictors is constructed. Then, for a given building, its historical data is compared against the feature database to extract the features that best fit the observed consumption patterns. Afterwards, historical data is grouped by daily consumption pattern similarities and a machine is trained on each cluster to make local or cluster specific predictions. Finally, these local predictions are combined to generate the global precise energy forecast of the building. Tested on a set of buildings geographically distributed in Canada and in the USA, our method shows improved performance compared to traditional approaches.