Energy consumed in buildings represents a challenge in the context of reduction of greenhouse gases emission. For this reason and due to the growing interest in operative costs reduction the energy used by buildings (tertiary and privates) for heating, ventilating, and air conditioning (HVAC) is even more investigated. Due to the nature of the energy consumption profile a predictive optimization method is one of the solution the scientific literature spreads even more. However optimization techniques need a good and reliable prediction of the variables of interest over a time horizon. This work focuses on methods to obtain a robust and reliable predictor based on artificial neural networks. For the optimization purposes the neural model predicts total heating energy consumption (gas), internal air temperature and aggregated thermal discomfort 12 h ahead. Training and testing data are simulated using a simulator based on heat, air and moisture model for building and systems evaluation (HAMBASE), by which a real office building was modeled. Influence of training data sample size and selection of predictor inputs is examined. Several combinations of early stopping condition and network complexity are tested for different training sample sizes. It is observed that the early stopping mechanism is crucial especially but not only for small training data, because it reliably overcomes overfitting problems. Surprisingly, relatively small networks were sufficient or performed best, although examined range of training sample covered up to five heating seasons. The use of a model tuning is thus supported by the results. Further, two strategies of selection of suitable input variables are demonstrated. While the input selection does not degrade the prediction performance, it is able to reduce the dimensionality and thus to save computational, communication, time, and data acquisition demands. The importance of inputs selection in HVAC modeling is thus pointed out and demonstrated. (C) 2015 Elsevier B.V. All rights reserved.