Using neural networks to predict hourly energy consumptions in office and industrial buildings as a function of weather data

Office and industrial premises are among the most energy consuming type of buildings. Compared to residential buildings, they are characterized by more regular occupation patterns and stricter control of building systems. Under these conditions, it is expected that energy consumptions may be more easily predictable and may be significantly influenced by outdoor conditions more than by individual preferences. This may result in availability of straightforward predictions of energy use (at daily or hourly basis) which may contribute to trade energy at lower costs, make a better use of renewable energies, while balancing energy saving and occupants’ comfort. An essential contribution to the ability to easily and accurately predict energy consumptions, is given by the ever-increasing number of smart and IoT-based devices that collect data inside and outside buildings and consequently make them available for processing. Taking advantage of such data, it is worth investigating if advanced artificial intelligence methods (like neural networks and machine learning) are capable of yielding predictions of energy consumptions and, ideally, indoor conditions. For the purpose of the present paper, the dataset (including both input and output parameters) was obtained through simulation (using the popular EnergyPlus tool), including one office and one industrial reference building, and using three different climatic datasets. Finally, artificial neural networks were trained assuming daily and hourly energy consumptions (subdivided by category) as the target variable, showing that in most of the cases very accurate predictions could be obtained.