Leveraging spatial data infrastructure for machine learning based building energy performance prediction

Abstract

The calculation, management and maintenance of energy performance of buildings (EPBs) are significant in increasing energy efficiency in buildings and reducing greenhouse gas emissions since it is estimated that approximately one third of energy consumption is associated with buildings, and furthermore, three-quarters of the existing building stock is characterized by energy inefficiency. However, in many cases, EPBs are either not calculated or not integrated in a register within national spatial data infrastructure (NSDI). This complicates policy development and planning for both local and national governments, which may result in numerous complications. The objective of this paper is twofold: firstly, to design a building energy data model as an extension of NSDI in Türkiye and then implementing and populating it with real data taken from energy performance certificates from the Tuzla District in Istanbul; and secondly, to develop energy performance prediction models with Machine Learning (ML) algorithms (i.e., Random Forest (RF), Gradient Boosting Machine (GBM), Light Gradient Boosting Machine (LightGBM) and Extreme Gradient Boosting (XGBoost) in order to estimate the overall performance scores of the buildings. The model’s findings demonstrated robust predictive accuracy, achieving an R2 of 0.818 (XGBoost) and performance metrics of RMSE=5.153, MAE=2.886, and MAPE=3.369. These results substantiate the model’s reliability in estimating targeted building energy performance scores. These predictions can be used to provide a comprehensive overview of districts in terms of EPB and inform the development of road maps at the district, city, or national level. Furthermore, the predictions can support the development of EPB-related legislation, facilitate the design of incentive and sanction mechanisms, and promote broader sustainability and climate mitigation goals in a practical manner. Nevertheless, as a limitation of this study, the model has only been tested in a single district, which restricts its generalizability; it should therefore be evaluated in other areas to confirm its applicability.