IMPROVING WHOLE BUILDING ENERGY SIMULATION WITH ARTIFICIAL NEURAL NETWORKS AND REAL PERFORMANCE DATA Raymond Sterling 1 , Daniel Coakley 1 , Thomas Messervey 2 , Marcus M. Keane 1 1 Informatics Research Unit for Sustainable Engineering, NUI Galway, Ireland 2 R2M Solution, Catania, Italy raymond.sterling@nuigalway.ie ABSTRACT This paper proposes an approach that combines detailed building energy simulation (BES) programs (such as Energy Plus) with real consumption data used to train an artificial neural network (ANN). The goal is to adapt the EnergyPlus simulation to provide a better representation of the real performance data of a building. The approach has the advantage of being able to extract relevant information from the BES model while at the same time producing an accurate model with a clear applicative view. The new model has the possibility to evolve and improve with the building’s operation adapting the simulation’s performance over time to account for modelling error, potential aging of the building, changes in use, or changes in its infrastructure. This work also shows comparative results where grey box models reach similar accuracies to black box models with the added advantage of being able to extract relevant information from them. Finally, the approach is applied to a naturally ventilated building in order to demonstrate the applicability and performance of the proposed methodology. INTRODUCTION Estimations indicate that buildings are responsible for over 40% of energy consumption in Europe and the United States and CO2 emissions of approximately 30% (EUROSTAT 2010; European Environment Agency 2011). To tackle these problems, several initiatives propose to deliver significant improvements in energy efficiency and the reduction of carbon emissions in buildings (e.g. Energy Performance in Buildings Directive (European Parliament and Council of Europe 2010)). These initiatives have increased the need for applying energy conservation measures in new and old buildings. However, given the practical limitations of rigorously testing different energy conservation measures (ECM’s) in existing buildings, and even more in planned buildings, one important area of focus is on developing and improving whole-building energy simulation. BES models serve as test beds on which testing different approaches to energy efficiency can be conducted until the desired reductions are reached and before applying these measures in the real building. This new focus on building simulation is driving research towards improving the simplicity and accuracy of the simulations using different calibration methodologies. Different techniques can be found in literature that are applied to the problem of using modelling and simulation combined with real operation data to accurately represent the energy behaviour of the building, with applications that vary from small residential buildings to big office complexes. These techniques may be grouped into the following branches: • Engineering Methods; • Black Box Models; • Grey box Models. Simplified engineering methods and black box models based on artificial intelligence offer the better trade- off between accuracy and simplicity of the model while grey box models can combine these two and leverage on the advantages of each to further improve accuracy (Zhao and Magoulès 2012). This paper presents an approach for integrating a detailed engineering model (engineering method) with an artificial neural network (ANN) model (black box method) to represent the energy behaviour of a natural ventilated building. The reason behind this approach is to reduce the development time of an accurate model while still allowing the extraction of relevant information and data from the engineering model. Another advantage of the approach is the possibility of automating the learning process such that the most recent data may be incorporated. By presenting a case study using real data, the efficacy of the approach can be assessed based on two standard metrics used for evaluating the goodness of building energy simulation for measuring energy demand and savings (ASHRAE 2002). These metrics are the coefficient of variation of the root mean square error (CV-RMSE) and the normalised mean square error (NMBSE) both comparing simulated with real measured data.