A Novel Hybrid Technique For Building Demand Forecasting Based On Data-driven And Urban Scale Simulation Approaches

This paper presents a novel feature engineering procedure to generate case study specific input variables for the training of data-driven models used to predict the heating demand of blocks of buildings. Traditionally, predictive model training is performed using sets of data from sensors (e.g. weather stations, metering systems). Feature engineering procedures such as the inclusion of innovative predictive variables in the forecasting framework are generally not considered. The method presented in this paper exploits results of calibrated physics-based building energy models that are included as an additional independent variable in combination with the traditional sets of predictors in an innovative forecasting framework. The method is tested on a district case study of the city of Geneva (CH) served by a district heating network. Results show that the presented approach improves the quality of the forecasting outcomes of state-of-the-art predictive algorithms. In this context, the accuracy of the simulation outputs affects the predictive capability of the presented forecasting procedure. In addition, normalised information derived from substation of the heating network of the district are informative for the predictive model.