Neural network synthetic dataset generation for fault detection in district heating substations

District heating systems (DHS) play a vital role in sustainable heating solutions and the decarbonization of the energy sector. However, inefficiencies due to undetected faults in substations result in high return temperatures, increasing heat losses, and limiting the integration of renewable energy sources. The lack of publicly available labeled datasets poses a significant challenge for fault detection using supervised learning models. To address this issue, this study explores three machine learning-based synthetic data generation techniques – time series forecasting, generative adversarial networks (GANs), and fault signature transfer. These methods aim to increase publicly available data either by sharing the generating model or a synthetic dataset. The novelty lies in the combination of advanced supervised machine learning methods being applied to a large, fully labeled data set to create new, equally labeled data for publication, as, to our knowledge, no such dataset has been compiled before. We evaluate our methods on the first-of-its-kind ILSE dataset, which includes real-world smart meter data from 547 substations and 1,162 reviewed faults from a German DHS network, including detailed root cause information. Overall, time series forecasting achieves an MAPE of 3% to 10% for inlet and outlet temperature and 25% to 40% for heat load and flow rate, both of which are within year-to-year variance. For GANs, specifically TimeGAN, we found a discriminative score of about 0.10 compared to 0.24 in the original publication when tested on Energy benchmark data. Fault signature transfer has yet to yield usable results, most likely due to the high variance in the fault signatures, fault duration, and overlapping or multiple root causes. Finally, fault data in the synthetic data is not yet good enough for practical use, e. g. training a fault detector. • The study focuses on improving fault detection in district heating substations. • A key challenge is the lack of publicly available and labeled datasets, which limits the development of effective machine learning models. • The paper explores three synthetic data generation methods: time series forecasting, generative neural networks (TimeGAN), and fault signature transfer. • The three data generation methods are tested on a large labelled dataset containing up to 547 substations, 1,162 fault reports, and 1,208 anomalies.