AI Spots Heating Faults Before Customers Complain

District heating systems, which supply warmth to buildings from a central source, are crucial for integrating renewable energy and reducing carbon emissions. However, faults in these systems often go unnoticed until customers report issues like no heat or hot water, leading to energy inefficiency and costly emergency repairs. A new study addresses this problem by introducing a publicly available dataset and an AI framework that can detect faults early, potentially transforming how utilities maintain these vital networks. This research, published by Fraunhofer IEE, provides a practical solution to a long-standing in the energy sector, where the lack of shared data has hindered progress in predictive maintenance.

The researchers developed a framework that combines a labelled dataset with an evaluation to benchmark early fault detection in district heating substations. They found that their AI models, using a technique called autoencoders, could identify 60% of faults before customers reported problems, with an average lead time of 3.9 days for those detected faults. This means utilities could address issues like malfunctioning pumps or incorrect settings proactively, reducing return temperatures and improving overall system efficiency. The dataset includes time series from 93 substations, annotated with fault descriptions and maintenance actions, making it the first public, real-world resource of its kind for this application.

To achieve these , the team used an open-source Python framework called EnergyFaultDetector, which employs autoencoders—neural networks that learn normal behavior from operational data. These models were trained on data representing fault-free periods, such as the last two years up to two weeks before an incident, and then monitored for deviations that signal anomalies. The evaluation focused on three metrics: Accuracy for recognizing normal behavior, Reliability for minimizing false alarms, and Earliness for detecting faults as soon as possible. For example, they set a desired detection window of 24 hours before a report to ensure actionable lead time, and used a criticality threshold to filter out sporadic false positives.

The data shows that the conditional autoencoder model, which accounts for time-of-day and day-of-week patterns, performed best overall, achieving a normal-behavior accuracy of 0.98 and a reliability score of 0.83 across both manufacturers' datasets. In specific cases, such as a fault where a domestic hot water controller was mistakenly set to night mode, the model detected anomalies 24 hours before the report, with feature analysis correctly identifying the dropped temperature setpoint as the root cause. For faults with high monitoring potential—those detectable with existing sensors—the models showed improved earliness and reliability, though some issues like incorrect parametrisation from commissioning were excluded as they require different checks.

This framework has significant for energy utilities and cities aiming to reduce waste and operational costs. By enabling predictive maintenance, it helps prevent comfort issues for customers and supports lower-temperature district heating networks, which are key to decarbonizing heat supply. The public dataset and open-source code allow other researchers and companies to build on this work, fostering innovation in fault detection. However, the study notes limitations, such as the 10-minute data resolution potentially missing brief faults like valve oscillations, and the need for more diverse fault types and transfer learning to handle limited training data in some cases. Future research could explore physics-informed models or higher-frequency data to further enhance detection capabilities.