The ability to accurately forecast the building energy consumed can be valuable in a number of contexts. In Energy Performance Contracting and retrofitting actions, a key requirement for determining the Return on Investment is a precise and accurate quantification of the energy and concomitant cost savings resulting from the implementation of a set of Energy Conservation Measures. The precise quantification of energy savings is critical for Energy Savings Companies (ESCOs), under the guaranteed or shared savings contracting models; typically, this is done following a Measurement and Verification (M&V) protocol where the pre-intervention energy, computed using a regression model is compared against actual post-intervention measured energy values. Another example, during the building operational phase, is the use of operational intelligence platforms to compare between expected and actual performance, and inform the facility manager when deviations or degradations are observed so that actions can be taken. Common to both these examples, is the need for a model capable of predicting with sufficient accuracy the energy consumption, while taking into account key factors that affect the overall performance, like prevailing weather conditions, occupancy patterns, etc. In the literature two classes of such models are often encountered: i) bottom-up, physics-based whole building simulation models; and ii) models based on statistical or machine learning techniques, requiring data typically obtained from the building monitoring system. The development of physics-based models necessitates a laborious and costly design and calibration process requiring expert knowledge; in the work presented here, data-driven models of the latter category are investigated since they have less stringent requirements, typically the availability of good-quality data. Different types of machine-learning models have been reported in literature, such as Neural Networks, Support Vector Machines, Gaussian Mixtures, etc. These are regression models aiming at identifying the underlying relationship between the depended variables (in our case the total energy consumption of the building) and the independent or explanatory variables, such as outside temperature levels, solar radiation, occupancy levels, etc. From the available model types, we adopt Gaussian Processes (GPs) since they do not require the excessive fine-tuning other models necessitate (e.g. selecting the number of hidden layers and nodes of a Neural Network or the hyper-parameters for the Support Vector Machines). GP models are purely data-oriented in the sense that an a-priori definition of a structural relationship between the dependent and explanatory variables is not required; what is required instead is the selection/specification of the covariance structure of the independent variables to explain the interaction with the dependent variable. An added benefit of GP regression models is that an estimate on the prediction uncertainty is also available. It has been shown that GPs require less data and lead to more accurate uncertainty estimates compared to standard regression methods, like the ones typically used in the application of the International Performance Measurement and Verification Protocol (IPMVP). The proposed methodology has been applied using measurements obtained from a study building. The task at hand is the modelling and prediction of both the daily and hourly total building energy consumption. At each time resolution, we have identified (different) suitable explanatory variables and performed a sensitivity analysis to evaluate the contribution of each variable on the final accuracy of the model. Two sets of data are used: training data for constructing the regression model; and test data for evaluating the accuracy of the prediction, utilizing common statistical indices such as Mean Squared Error (MSE) and coefficient of determination (R2) to evaluate the quality of the predictions. The experiments performed indicate the importance of selecting proper independent variables for each task, since the quality of the prediction largely depends on the ability of the explanatory variables to describe the behaviour of the dependent variable. In addition, the prediction quality and robustness of the GP regression was evident in all our tests, which combined with the laborious-free tuning of the framework make this approach highly-attractive for the task.