A Data-Driven Approach for Discovery of Heat Load Patterns in District Heating

Ece Calikus, Slawomir Nowaczyk, Anita Pinheiro Sant'Anna, Henrik Gadd, Sven Werner

Abstract

Understanding the heat usage of customers is crucial for effective district heating operations and management. Unfortunately, existing knowledge about customers and their heat load behaviors is quite scarce. Most previous studies are limited to small-scale analyses that are not representative enough to understand the behavior of the overall network. In this work, we propose a data-driven approach that enables large-scale automatic analysis of heat load patterns in district heating networks without requiring prior knowledge. Our method clusters the customer profiles into different groups, extracts their representative patterns, and detects unusual customers whose profiles deviate significantly from the rest of their group. Using our approach, we present the first large-scale, comprehensive analysis of the heat load patterns by conducting a case study on many buildings in six different customer categories connected to two district heating networks in the south of Sweden. The 1,222 buildings had a total floor space of 3.4 million square meters and used 1,540 TJ heat during 2016. The results show that the proposed method has a high potential to be deployed and used in practice to analyze and understand customers' heat-use habits.

Introduction

District heating (DH) systems are essential components of sustainable urban energy infrastructure. However, their efficiency is often hampered by a lack of visibility into the "demand side"—the buildings themselves. Traditionally, DH operators have treated customer buildings as a "black box," focusing primarily on production and distribution. To transition toward 4th Generation District Heating, which requires lower distribution temperatures and higher efficiency, operators must understand how heat is actually consumed.

Manual analysis of thousands of buildings is impossible, and simple metrics like total annual demand fail to reveal temporal inefficiencies or faulty equipment. This paper introduces an automated, data-driven framework to analyze heat usage across entire networks, identifying behavioral patterns and anomalies at scale.

Methodology

The researchers developed a three-step pipeline based on the Knowledge Discovery in Databases (KDD) process to transform raw meter data into actionable insights.

The three-step Knowledge Discovery in Databases (KDD) pipeline for heat load analysis.

Step 1: Data Preprocessing

Raw hourly meter data often contains noise and sensor errors. The authors utilized Median Absolute Deviation (MAD) to filter out sudden sensor "jumps." Following cleaning, the data underwent z-normalization. This step is critical because it ensures the clustering algorithm focuses on the shape of the heat load (the temporal pattern) rather than the absolute magnitude of consumption, allowing buildings of different sizes to be compared directly.

Step 2: Pattern Discovery and Data Representation

To handle the massive volume of data, the researchers aggregated hourly readings into representative weekly profiles for different seasons (Winter, Early Spring/Late Autumn, Late Spring/Early Autumn, and Summer). This creates a seasonal "fingerprint" for each building.

Aggregation of hourly heat load data into representative weekly profiles across different seasons.

For clustering, the study employed the k-shape algorithm. Unlike standard k-means, which uses Euclidean distance and often fails with time-series data, k-shape utilizes a "Shape-Based Distance" (SBD) metric based on cross-correlation. This allows it to group sequences with similar peaks and valleys even if they are slightly shifted in time.

Step 3: Anomaly Detection and Visual Exploration

After clustering, the authors identified "customers of interest" using Cantelli’s inequality to calculate an abnormality threshold. Profiles deviating more than three standard deviations from their cluster mean were flagged. Finally, domain experts inspected the clusters to map them to known control strategies, such as Night Setback (NSB) or Continuous Operation (COC).

Case Study: Swedish District Heating Networks

The framework was tested on a real-world dataset from two networks in south-west Sweden (Helsingborg and Ängelholm). The study analyzed 10.7 million hourly measurements from 1,222 buildings across six categories: multi-dwelling, commercial, public administration, health/social, school, and industrial.

Key Results

  • 15 Distinct Clusters: These were successfully mapped to four primary control strategies: Continuous Operation (COC), Night Setback (NSB), and two types of Time Clock operations (TCO5 for workdays, TCO7 for full weeks).
  • Prevalence of Inefficiency: A striking 68% of all analyzed buildings used Continuous Operation, accounting for 73% of total heat demand. In the service and industrial sectors, 45% of buildings ran at full capacity 24/7, representing a significant opportunity for energy savings through simple controller reprogramming.
  • Actionable Anomalies: The framework flagged 163 buildings (approximately 12%) as abnormal. Expert validation revealed these weren't just errors but specific issues: faulty substation valves causing high summer loads, or unique usage patterns like nightclubs with massive weekend spikes.

Discussion and Future Work

The study demonstrates that a building’s control strategy leaves a distinct visual signature. By focusing on shape rather than magnitude, the researchers proved that diverse building types often share identical behavioral patterns, enabling personalized demand management at scale.

However, the high data rejection rate (38%) due to meter failures highlights ongoing physical reliability challenges in smart grids. Future work will focus on integrating weather data more explicitly and moving toward automated diagnostics that can "learn" to identify specific types of faults, such as leaking valves, directly from the heat load signature.

Summary updated

Abstract

Understanding the heat usage of customers is crucial for effective district heating operations and management. Unfortunately, existing knowledge about customers and their heat load behaviors is quite scarce. Most previous studies are limited to small-scale analyses that are not representative enough to understand the behavior of the overall network. In this work, we propose a data-driven approach that enables large-scale automatic analysis of heat load patterns in district heating networks without requiring prior knowledge. Our method clusters the customer profiles into different groups, extracts their representative patterns, and detects unusual customers whose profiles deviate significantly from the rest of their group. Using our approach, we present the first large-scale, comprehensive analysis of the heat load patterns by conducting a case study on many buildings in six different customer categories connected to two district heating networks in the south of Sweden. The 1222 buildings had a total floor space of 3.4 million square meters and used 1540 TJ heat during 2016. The results show that the proposed method has a high potential to be deployed and used in practice to analyze and understand customers' heat-use habits.
BibTeX
@inproceedings{calikus2019data,
  author = {Ece Calikus and Slawomir Nowaczyk and Anita Pinheiro Sant'Anna and Henrik Gadd and Sven Werner},
  title = {{A Data-Driven Approach for Discovery of Heat Load Patterns in District Heating}},
  booktitle = {Applied Energy},
  year = {2019},
  doi = {10.1016/j.apenergy.2019.113409},
}
Figure