DAY-AHEAD FORECASTS OF EXCHANGED HEAT IN A DISTRICT HEATING SUBSTATION WITH ENSEMBLE METHODS

Abstract

Short-term forecasts of heating loads are very important for planning the operation of energy supply systems successfully. The heat demand of a building depends on various parameters that include the materials, geometry, occupancy, type of activity, etc. There are a number of machine learning methods and approaches used in the literature to accomplish these tasks. This paper presents a segment of broader research. The objective is to forecast the amount of heat exchanged in a district heating substation between the primary network and the secondary heating system of a multi-story residential building. Forecasts are performed 24 hours ahead, with the resolution of one hour. The paper applies and compares multiple ensemble regression methods based on decision trees. The input parameters are heat demand lags, time-related variables, e.g. hour of day, day of week and month, and dry bulb temperature as the most important weather variable. The time-series problem is transformed into a classical supervised machine learning problem. The models are trained and tested with the actual measured data collected over five heating seasons. The paper examines the performance of four methods: gradient boosting, histogram gradient boosting, extremely randomized trees and random forest. The applied evaluation metrics for the models are the coefficient of determination, root mean squared error and mean absolute error. All methods used have very similar prediction performance. Random forest has the smallest root mean square error (43.56 kWh), while extremely randomized trees have the lowest mean absolute error (27.34 kWh).