Data-Driven Analysis of Heating Demand in Buildings Connected to District-HeatingPattern Recognition and Demand Prediction

  1. Mikel Lumbreras 1
  2. Koldobika Martin-Escudero 1
  3. Gonzalo Diarce 1
  4. Roberto Garay- Martinez 2
  1. 1 ENEDI Research Group, Department of Energy Engineering, Faculty of Engineering of Bilbao, University of the Basque Country UPV/EHU
  2. 2 TECNALIA, Basque Research and Technology Alliance (BRTA)
Libro:
Renovation wave: 12º Congreso Europeo sobre Eficiencia Energética y Sostenibilidad en Arquitectura y Urbanismo – 5º Congreso Internacional de Construcción Avanzada: Bilbao, 29-30 Septiembre 2021
  1. Rufino J. Hernández Minguillón (ed. lit.)

Editorial: Servicio Editorial = Argitalpen Zerbitzua ; Universidad del País Vasco = Euskal Herriko Unibertsitatea

ISBN: 978-84-1319-374-8

Año de publicación: 2021

Páginas: 47-56

Congreso: Congreso Europeo sobre Eficiencia Energética y Sostenibilidad en Arquitectura y Urbanismo (12. 2021. Bilbao)

Tipo: Aportación congreso

Resumen

This paper presents a novel framework for the analysis of heat consumption data of buildings connected to a district-heating network using machine learning techniques. The high variability and uncertainty of energy production in new district-heating networks make highly important to have deep insightful knowledge of the instant demand of all the buildings connected to the grid. Thus, the present paper presents a methodology for discovering heat consumption patterns in the buildings as well as a black-box model for heat load prediction. The approach to analyzing the consumption data is carried out by a combination of unsupervised and supervised learning models. The unsupervised learning of the heat consumption patterns is carried out using the widely used k-means algorithm, whereas supervised random-forest algorithm is applied for heat-load forecasting. The proposed framework is applied to a real residential building located in Tartu (Estonia) and connected to a subnetwork of the district-heating network of this location. The unsupervised clustering results in three main day-types with different consumption patterns throughout these days. Silhouette index is used for the validation of the clusters. The outcome from the heat load prediction model results in prediction accuracy over 0.95 for the R2 value