AI Detects Faulty Radiators and Fixes Them in Real Time

District heating is one of the quiet workhorses of urban decarbonization, and it is also one of the most fragile. Across Europe, the share of final heat consumption delivered by district heating networks is projected to climb from 15 percent in 2020 to 31 percent by 2050 as the continent pursues climate neutrality. Yet the performance of these networks ultimately depends on what happens at the end of the pipe, inside buildings where radiators and substations extract heat from the supply water. A new study published in Results in Engineering by Nima Monghasemi, Stavros Vouros, Konstantinos Kyprianidis, and Amir Vadiee of Mälardalen University and collaborators argues that this building-level bottleneck has been badly underserved by fault-detection research, and it offers something most prior systems lack: a closed loop that not only spots a broken radiator but immediately compensates for it.

The scale of the problem is striking. Surveys of Swedish district heating utilities have found that leakages and faults in building internal heating systems are the most frequently reported fault categories. Studies indicate that up to 74 percent of buildings connected to district heating networks contain faults that elevate return temperatures, and an analysis of hourly meter readings from 135 substations in two Swedish networks found that roughly three quarters exhibited at least one fault or sub-optimal behavior. These customer-side faults compress the supply–return temperature differential, force higher distribution flow rates, and undermine the energy and exergy efficiency that low-temperature district heating is designed to achieve. In fourth-generation systems, which operate with narrower temperature margins than the 80–100 °C networks most buildings still connect to, such disturbances are even more consequential.

What makes the new framework unusual is that it embeds unsupervised machine learning models directly inside the building energy simulation loop. The researchers trained three anomaly detectors—a One-Class Support Vector Machine, an Isolation Forest, and a Local Outlier Factor model—exclusively on fault-free baseline data from a full heating season of a modified U.S. Department of Energy Retail Standalone commercial building, simulated in EnergyPlus with a 10-minute timestep and located in Västerås, Sweden. Six engineered features, including the instantaneous temperature error, a normalized hot-water mass flow ratio, a two-hour rolling mean of the comfort deviation, the zone temperature rate of change, outdoor temperature, and the hot-water supply–return differential, define the operating point. Because the models learn only the boundaries of normal operation, they can flag emerging deviations without requiring labeled examples of specific faults, which are scarce and expensive to collect in real buildings.

Once a sustained anomaly is confirmed by a temporal persistence filter, a signature-based diagnosis ladder classifies the fault as stuck-closed, stuck-open, or a supply-setpoint bias, and the supervisory compensation module ramps the supply-temperature setpoint toward a fault-specific target. The ramp follows a discrete first-order decay with a gain of 0.2, meaning roughly 90 percent of a requested setpoint change completes within 110 minutes. For under-delivery faults the target is 70 °C, boosting heat extraction through a restricted radiator; for over-delivery the target is 50 °C, throttling the thermal driving force. The compensated setpoint feeds back into the simulation at the next timestep, completing the closed loop with a one-step causal delay that a physically realizable controller could replicate.

The fault taxonomy deliberately spans physical and control failure mechanisms. A stuck-closed valve is a partial mechanical obstruction parameterized by a flow-restriction fraction, clamping hot-water flow to as little as 10 percent of the reference capacity. A stuck-open valve is implemented as a positive bias on the zone heating setpoint, reproducing the sustained overheating and compressed differential a mechanically seized valve would cause. The supply-setpoint bias extends the taxonomy to a system-level failure, depressing the hot-water loop temperature across all zones simultaneously, as might result from a drifted outdoor-temperature sensor or an incorrect reset-curve slope. Across 45 distinct simulations, the framework was tested at five severity levels per fault type, with parameter sensitivity sweeps and compound-fault stress tests probing the boundaries of supervisory controllability.

The headline results reveal a fault-type dichotomy in supervisory recovery potential. The stuck-open proxy fault achieved the highest comfort recovery at 68.8 percent of fault-induced discomfort degree-hours, while cutting hydronic energy use by 15 percent relative to the uncompensated fault and lowering the return-water temperature 4.2 K below baseline, a secondary benefit for network performance. The supply-setpoint bias achieved 57.2 percent recovery at a near-zero energy penalty, though compensation raised the return temperature 6.0 K above baseline because the unrestricted hydraulic path drives excess heat through the radiators. The physically constrained stuck-closed valve achieved only 38.6 percent recovery, because the mechanical flow restriction imposes a hydraulic ceiling that no supply-temperature boost can breach.

Detection performance proved strongly fault-dependent. The stuck-closed fault, with its pronounced combination of anomalously low flow and sustained underheating, was detected by a majority-vote ensemble with perfect recall and an F1-score of 0.90 within 20 minutes of onset, at a seasonal false-positive rate of just 0.18 percent. The stuck-open fault, whose overheating signal builds gradually across multiple features, required 100 minutes and achieved an F1 of 0.86. The supply-setpoint bias was the hardest to detect, with diffuse thermal signatures overlapping normal low-load conditions; none of 168 evaluated configurations exceeded an F1 of 0.52, though high recall ensured the compensation loop could still engage reliably.

Perhaps the most counterintuitive finding concerns feedback stability. For the stuck-open fault, successful compensation removes the overheating signature that triggered detection, causing the controller to release, the biased setpoint to reassert, and the cycle to repeat four to five times per occupied day—a sawtooth oscillation that caps net recovery. For the supply-setpoint bias, an asymmetric release logic with an extended hold sustains a single uninterrupted compensation block, eliminating cycling. The authors conclude that release logic design is as consequential for supervisory recovery as the detection algorithm itself, and that faults where compensation suppresses its own anomaly signature should employ fixed-duration holds rather than continuous re-detection.

Severity sweeps exposed practical detection floors and recovery ceilings. The mildest stuck-open bias of +0.5 °C is entirely undetectable, lying within normal zone-temperature variability, while detection becomes effective at +1.0 °C. For the stuck-closed fault, recovery peaks at a moderate restriction of φ = 0.20 and collapses at the most severe φ = 0.10, where the supply-temperature boost provides insufficient additional thermal driving force. Counterintuitively, detection F1 for the supply-setpoint bias peaks at moderate bias and declines at extreme severity, because very large depressions push the heating system into a regime resembling cold-weather normal operation rather than a distinct fault signature. The fixed anomaly thresholds also provide no severity estimation capability, a limitation for maintenance prioritization.

Compound-fault stress tests in diagnosis-only mode revealed the framework’s single-fault operational scope. When two under-delivery faults reinforce each other, the diagnosis remains action-consistent, correctly calling for a supply-temperature boost. But when a stuck-open bias partially cancels a supply-setpoint depression, the fixed-priority ladder commits to the wrong verdict on most alarmed steps, and parallel signature evaluation reveals conflicting control actions—reduce versus boost—on 80.7 percent of alarmed timesteps. The authors argue that an explicit abstain state with temporal confirmation is necessary to prevent erroneous compensation under concurrent faults. For physically constrained faults, the framework functions primarily as a diagnostic triage tool, providing rapid classification for priority maintenance dispatch, whereas control-type faults are viable candidates for immediate autonomous relief until maintenance is scheduled. Field validation in instrumented district heating substations remains the critical next step to assess sensor noise and unmodeled hydraulic dynamics before such systems can move from simulation into operational buildings.

Subject of Research: Building energy systems and district heating

Article Title: Integrated fault detection and supervisory compensation for hydronic radiator faults in district-heated buildings

Article References: Monghasemi, N., Vouros, S., Kyprianidis, K., & Vadiee, A. (2026). Integrated fault detection and supervisory compensation for hydronic radiator faults in district-heated buildings. Results in Engineering, 32, Article 113471. https://doi.org/10.1016/j.rineng.2026.113471

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113471

Keywords: Integrated, fault, detection, supervisory, compensation, hydronic, radiator, faults, district-heated, buildings, scientific research