Concept drift detection in district heating: A characterization framework and detector-centric benchmark

Energy and AI, vol. 25, pp. 100788

Abstract

Machine learning models for district heating load forecasting operate in non-stationary data stream environments where concept drift causes progressive performance deterioration. Systematic approaches to detect drift remain limited due to ad hoc algorithm selection, predominant evaluation within active model adaptation contexts, and lacking application-specific benchmarks with documented ground truth drift occurrences. This paper develops a benchmarking framework for concept drift detection using district heating as the application domain. We present a multidimensional characterization framework employing dissimilarity measurement as the primary dimension to enable cross-category detection method comparison. A standardized evaluation protocol combining detector-centric drift detection metrics, model performance metrics, and computational effort metrics isolates detection accuracy from downstream model adaptation effectiveness. Eight representative drift detection methods spanning performance-based and data distribution-based categories are benchmarked on the DHDrift dataset comprising nine scenarios with sudden and incremental drift at three severity levels, complemented by hyperparameter robustness, noise robustness, and real-world case studies. Maximum MCC of 0.32 highlights the inherent difficulty of detecting drift in seasonal thermal load data. Kolmogorov–Smirnov Windowing achieves highest drift detection accuracy, while Paired Learners demonstrates statistically superior load forecasting improvement. Performance-based methods achieve detection latencies below four days for sudden drift, whereas data distribution-based methods require longer evidence accumulation. All methods operate within real-time computational constraints. However, incremental drift affecting load pattern shapes emerges as a critical capability gap among the evaluated methods. The complementary studies show that leading algorithm rankings persist under suboptimal hyperparameter configurations, measurement noise, and a real-world case, providing initial evidence for transferability.

Authors 3

1. K Derzsi corresponding Aachen E.ON Energy Research Center Institute for Energy Efficient Buildings and Indoor Climate

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, E.ON Energy Research Center, Institute for Energy Efficient Buildings and Indoor Climate, Mathieustraße 10, Aachen, 52074, Germany

2. Jonas Klingebiel Aachen E.ON Energy Research Center Institute for Energy Efficient Buildings and Indoor Climate

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, E.ON Energy Research Center, Institute for Energy Efficient Buildings and Indoor Climate, Mathieustraße 10, Aachen, 52074, Germany

3. Dirk Müller Aachen E.ON Energy Research Center Institute for Energy Efficient Buildings and Indoor Climate

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, E.ON Energy Research Center, Institute for Energy Efficient Buildings and Indoor Climate, Mathieustraße 10, Aachen, 52074, Germany