Fault Detection in Solar Thermal Systems using Probabilistic Reconstructions

As heating is the largest factor of Greenhouse gases in the household sector, it should
be the focus of our decarbonisation efforts. Solar Thermal Systems (STS), which provide
heat based on solar energy, are a promising technology in this regard. However, STS
are prone to faults due to improper installation, maintenance, or operation, often leading
to a substantial reduction in efficiency, damage to the system, or even an increase in
energy cost. As individual monitoring is economically prohibitive for small-scale systems,
automated monitoring and fault detection should be used to address this issue.
We propose a data-driven neural network approach for fault detection in small-scale
STS, utilising probabilistic reconstructions from a long short-term memory (LSTM) based
Variational Autoencoder (VAE). Key factors in our approach are generalising from faultless
data to previously unseen systems and an anomaly score derived from an ensemble of
reconstructions. We apply this to an operational dataset provided by our industry partner,
which includes systems with different types of faults.
Our results show that our model can detect faults in STS with comparable performance
to the state-of-the-art expert-based system used by our industry partner. Furthermore, our
model can detect previously undetected faults, specifically those resulting from unexpected
behaviour in the control software or behaviours that were entirely unexpected and not
considered in the expert-based system. Thus, a combination of our model and the expert-
based system covers a broader range of faults than either system and is proposed for
further use in the industry partner’s application. Additionally, other providers without a
functioning expert-based system could build upon our work to get a minimal viable product
for fault detection in STS, purely based on data from existing systems and without the
need to install additional sensors or domain-specific knowledge.