Forecasting District Heating Demand using Machine Learning Algorithms

Abstract

Short-term forecasting of thermal energy demand is critical to optimally manage on-site renewable energy generation and the charge and discharge of energy storage devices in district heating and cooling (DHC) systems. As part of a larger study on advanced predictive control for a solar district heating system with 52 homes – the Drake Landing Solar Community (DLSC) – this paper investigates the use of Machine Learning algorithms to predict the aggregated heating load of the community.In this study, the initial approach to estimate the heating load of the DLSC employed a piecewise linear regression based on the outdoor air temperature. Such an approach yields significant errors, in particular when weather forecasts are used instead of actual outdoor air temperature measurements. It has been found that Machine Learning algorithms, such as decision trees, can significantly improve the accuracy of predicted heating loads by incorporating the effect of additional influencing factors (e.g., time of the day, day of the week, solar radiation, etc.). In this study, the predicted heating demand obtained from different algorithms are compared under two different scenarios; (a) by using actual weather conditions from measured data; (b) by using weather forecasts. The potential implementation of such models for control purposes is discussed.