Researchers in many countries are developing surrogate models to estimate the energy performance of the building stock. In Brazil, the building energy labelling system can be performed using a simplified method which is based on a basic surrogate model using multiple linear regressions. Based on the limitations associated with this model the aim of this study was to develop a more accurate surrogate model to predict the annual cooling energy demand of commercial buildings. The combination of all possible variations of the properties and their values resulted in more than 2.5 quadrillion cases. One million cases sampled by Latin Hypercube method were considered. Several statistical modelling techniques were tested to generate the surrogate model: multiple linear regression, multivariate adaptive regression splines, support vector machines, the Gaussian process, random forests and artificial neural networks. The surrogate model was applied into a medium office to observe the difference between building energy simulation results. The results showed that the artificial neural network method presented the best performance, with a NRMSE below 1%. The validation procedure indicates that the novel surrogate model is able to describe the relation between inputs data and cooling energy demand. (C) 2016 Elsevier B.V. All rights reserved.