Role of input features in developing data-driven models for building thermal demand forecast

Abstract

The energy consumption of buildings accounts for a major share in the modern society. Accurate forecast of building thermal demand is of great significance to both building management systems and heat distribution networks. Machine learning models driven by abundant load data have demonstrated their great capability in predicting real-world consumption patterns and trends. A large number of input features have been considered in the literature for developing data-driven models. However, a thorough analysis regarding their importance is currently lacking. This work first presents a review on the commonly considered features in building thermal demand prediction models, and focuses particularly on their influences. To further facilitate investigating the impacts of various input features, based on a four-year dataset collected from a district heating system with 13 input features, a deep learning model, the long short-term memory (LSTM) network, is employed for a real-world case study. Our results suggest that the past load, outdoor temperature, and hour index have the greatest influence, and should be primarily considered in building thermal demand forecast models. For the studied case, they lead to an RMSE of 12.231 MW and a CV-RMSE of 5.814 %. Additionally involving wind speed and day index is also useful, which improves the RMSE to 11.971 MW and CV-RMSE to 5.691 %. On the contrary, including all available features does not achieve a bettery accuracy, in which RMSE and CV-RMSE are 12.349 MW and 5.871 %.