Artificial Neural Networks to Support Energy Efficiency Planning in Wide City-Built Compartments

The purpose of this research is to develop a highly reliable forecasting tool based on machine learning techniques to assess the energy consumption of buildings before and after the implementation of energy retrofit projects. In relation to the research field of energy efficiency in the built environment, several approaches have been tested and successfully applied. However, the literature still lacks a shared approach that can be fast and flexible enough to be used on the large-city scale, but at the same time very precise and accurate in energy assessments. Therefore, this paper develops a set of neural networks to forecast the energy demand of buildings due to heating, cooling, hot water, and electricity as a function of a bundle of buildings descriptive features, such as, among the others, envelope characteristics, installations, climatic area, or occupation schedule. The training of the neural networks is performed on 100,000 observations produced by means of the software EnergyPlus. The archetype parametric buildings, created to produce the training database, have been calibrated on real data. The neural networks produced allow to experiment and test different energy retrofit options on a multiplicity of buildings, confronting several design scenarios, while also comparing the environmental benefits produced. The neural networks developed in this research are validated on a set of 11 buildings in North Italy, producing reliable results, with average discrepancies below 10% when compared to real data. Therefore, among the main results of this research, there is the flexibility and versatility of the forecasting tool developed, speeding up the management of energy retrofit projects.