Automatic energy demand assessment in low-carbon investments: a neural network approach for building portfolios

Abstract

Purpose This paper aims to develop a forecasting tool for the automatic assessment of both environmental and economic benefits resulting from low-carbon investments in the real estate sector, especially when applied in large building stocks. A set of four artificial neural networks (NNs) is created to provide a fast and reliable estimate of the energy consumption in buildings due to heating, hot water, cooling and electricity, depending on some specific buildings’ characteristics, such as geometry, orientation, climate or technologies. Design/methodology/approach The assessment of the building’s energy demand is performed comparing the as-is status (pre-retrofit) against the design option (post-retrofit). The authors associate with the retrofit investment the energy saved per year, and the net monetary saving obtained over the whole cost after a predetermined timeframe. The authors used a NN approach, which is able to forecast the buildings’ energy demand due to heating, hot water, cooling and electricity, both in the as-is and in the design stages. The design stage is the result of a multiple attribute optimization process. Findings The approach here developed offers the opportunity to manage energy retrofit interventions on wide property portfolios, where it is necessary to handle simultaneously a large number of buildings without it being technically feasible to achieve a very detailed level of analysis for every property of a large portfolio. Originality/value Among the major accomplishments of this research, there is the creation of a methodology that is not excessively data demanding: the collection of data for building energy simulations is, in fact, extremely time-consuming and expensive, and this NN model may help in overcoming this problem. Another important result achieved in this study is the flexibility of the model developed. The case study the authors analysed was referred to one specific stock, but the results obtained have a more widespread importance because it ends up being only a matter of input-data entering, while the model is perfectly exportable in other contexts.