Fault Detection in Solar Thermal Systems using Probabilistic Reconstructions

TL;DR

The paper tackles fault detection in domestic solar thermal systems by leveraging a probabilistic reconstruction-based framework. It introduces an LSTM-based VAE that provides probabilistic reconstructions and both homoscedastic and heteroscedastic uncertainty estimates, trained only on nominal data and evaluated on the PaSTS dataset. Results show the reconstruction-based approach, especially with heteroscedastic uncertainty, outperforms simple baselines like PCA and more complex DL models, highlighting robust generalization to unseen systems. The work emphasizes practical deployment considerations, suggesting that while DL can offer improvements, simple and well-calibrated models may provide a favorable trade-off between performance and engineering overhead, with real-world integration into service workflows demonstrated.

Abstract

Solar thermal systems (STS) present a promising avenue for low-carbon heat generation, with a well-running system providing heat at minimal cost and carbon emissions. However, STS can exhibit faults due to improper installation, maintenance, or operation, often resulting in a substantial reduction in efficiency or even damage to the system. As monitoring at the individual level is economically prohibitive for small-scale systems, automated monitoring and fault detection should be used to address such issues. Recent advances in data-driven anomaly detection, particularly in time series analysis, offer a cost-effective solution by leveraging existing sensors to identify abnormal system states. Here, we propose a probabilistic reconstruction-based framework for anomaly detection. We evaluate our method on the publicly available PaSTS dataset of operational domestic STS, which features real-world complexities and diverse fault types. Our experiments show that reconstruction-based methods can detect faults in domestic STS both qualitatively and quantitatively, while generalizing to previously unseen systems. We also demonstrate that our model outperforms both simple and more complex deep learning baselines. Additionally, we show that heteroscedastic uncertainty estimation is essential to fault detection performance. Finally, we discuss the engineering overhead required to unlock these improvements and make a case for simple deep learning models.

Figures (14)

• Figure 1: Schematic overview of our LSTM-based VAE for anomaly detection. The input time series is tokenized into fixed-length segments and passed through an encoder LSTM, which outputs parameters of a Gaussian latent distribution. A latent sample drawn from this distribution is then fed into a decoder LSTM, which reconstructs the tokens and maps them to a Gaussian distribution, approximating the original input data.
• Figure 2: Collector temperature for a high variance and low variance day of the same system. Illustrating the large difference in variance both within a day and between different days.
• Figure 3: Performance of Our Model and the Rescaled PCA-R model on different systems. The systems are sorted by performance of that model. The systems are sorted by their F1 score, and on the x-axis is the rank of the model. On the y-axis is the F1 score the model achieves on the corresponding system. The average of these is the system-wise F1 score. Note that the system ranks can differ between the two models.
• Figure 4: Qualitative results for two sample systems. The x-axis represents the day index, and the y-axis is the anomaly score (values above 3 are capped for visibility). Individual dots denote a day with a fault annotation (Fault), a potential anomaly marked by the expert system (Merk) or nominal behaviour (Normal) according to its colour.
• Figure 5: Impact of the number of principal components on anomaly detection performance. The x-axis depicts the number of components, and the y-axis shows both the resulting F1 score and the cumulative variance explained by those components.