Identifying the validity domain of machine learning models in building energy systems

Abstract

The building sector significantly contributes to climate change. To improve its carbon footprint, applications like model predictive control and predictive maintenance rely on system models. However, the high modeling effort hinders practical application. Machine learning models can significantly reduce this modeling effort. To ensure a machine learning model’s reliability in all operating states, it is essential to know its validity domain. Operating states outside the validity domain might lead to extrapolation, resulting in unpredictable behavior. This paper addresses the challenge of identifying extrapolation in data-driven building energy system models and aims to raise knowledge about it. For that, a novel approach is proposed that calibrates novelty detection algorithms towards the machine learning model. Suitable novelty detection algorithms are identified through a literature review and a benchmark test with 15 candidates. A subset of five algorithms is then evaluated on building energy systems. First, on two-dimensional data, displaying the results with a novel visualization scheme. Then on more complex multi-dimensional use cases. The methodology performs well, and the validity domain could be approximated. The visualization allows for a profound analysis and an improved understanding of the fundamental effects behind a machine learning model’s validity domain and the extrapolation regimes.

Keywords

• Extrapolation detection
• Validity domain
• Novelty detection
• Machine learning
• Artificial neural network
• Data-driven model predictive control