Dear Reader,

In this 60th issue of our heatbeat Research Newsletter, we spotlight two current studies at the intersection of asset management and operational optimization for district heating networks (DHNs). The first paper demonstrates how machine learning (ML) can prioritize pipe replacements with limited data, and the second distills what truly needs to be modeled when bringing Model Predictive Control (MPC) to decentralized DHNs.

1) Predicting Pipe Failures: A Machine Learning Approach to Asset Management in District Heating from Jensen et al.

What they did

Researchers from DTU and SDU developed a per-pipe fault-vulnerability model that combines geospatial layers (soil, land use, road class, precipitation, topography) with operational pressure information, trained on HOFOR data from a Danish DHN. Because detailed hydraulics were not available, the team introduced a lightweight linear pressure approximation (pressure drop per meter plus elevation term), enabling pressure features without a digital twin. Class imbalance (≈1:215) was handled via over-/down-sampling and feature selection (Boruta). A Random Forest achieved the best performance.

Key results

• Ranking performance: The model achieves Area Under the Curve (AUC) ≈0.72, and captures >40% of historical failures by inspecting only ~10% of network length—on par with or better than prior literature.
• Pressure matters: Even the simple pressure proxy improved AUC10 by ~4.5% versus no-pressure features, showing operational data adds real signal beyond static GIS.
• Absolute replacement decisions remain hard: Under severe class imbalance, a best-tuned threshold still produced many false positives—underscoring the need for more labeled failures and richer features (e.g., number of joints, internal water temperature) to support dig/no-dig decisions.

Why it matters

Utilities can start prioritizing inspections and targeting hotspots using data that many already have: GIS layers, seasonal pressure at source, and simple elevation data—without waiting for a full hydraulic twin. At the same time, operators should treat ML outputs as ranked risk lists (where it excels) rather than definitive replacement commands.

Requirements analysis for Model Predictive Control in a decentralized DHN from Zoschke et al.

What they did

Fraunhofer ISE, University of Freiburg, and partners assessed which thermohydraulic effects must be included when deploying MPC in a multi-producer DHN (Weil am Rhein, Germany). They compared a heat-based MILP (no hydraulics) against evidence from hydraulic simulation (pandapipes) and monitoring—then derived the minimum set of non-linearities needed for practical, reliable MPC.

Key results & implications

• MPC reduces costs—already with a heat-based MILP: Using one year of data, expected operating cost reductions were ~17.8% (perfect forecast) and ~14.3% with ±20% noise, mainly by reducing fossil backup and increasing CHP usage.
• Hydraulic limits are the deal-breaker: Pump head constraints can prevent certain producer mixes at high load (e.g., one site limited to ~3bar head), so pressure-loss constraints and pump limits must be included to avoid infeasible schedules. Pumping costs themselves were minor (~0.6-1.3€/MWh) and don’t change dispatch priorities—they’re optional in the objective.
• Supply temperature optimization: small but real: Data suggested ~1.8% extra savings from flow-temperature optimization; dwell times were long (winter mean ≈38min; summer ≈3h), emphasizing time-delay modeling if temperature becomes a decision variable. Network storage effect was limited (≈0.5MWh winter, 2MWh summer) compared to the existing TES (≈6MWh).

What to implement first

• Must-have for reliability: Add pressure-loss constraints and pump head limits into MPC.
• Next steps for efficiency: If problem size remains tractable, add supply-temperature optimization with appropriate delay handling—expect incremental gains vs. the large benefits already achieved by anticipatory, heat-based MPC.

Further Information

We recommend reading both articles in full for methodology and context (MPC: DOI in pre-proof; Pipe Failures: IEEE Access—accepted). Also, as mentioned in our last issue, don’t miss our monthly heatbeat Digital Twin feature updates and new webinar format on our blog: https://heatbeat.de/en/blog/83/ .

The next issue of our newsletter will be published on November 5th, 2025. Until then, you can meet us at the following events https://heatbeat.de/en/events/ :

• 01.10.2025: heatbeat digital Twin - Feature Update Live
• 22.10.2025 - 23.10.2025: KOMMUNALE - Nationwide Trade Fair and Congress for Municipal Supplies in Nürnberg
• 23.10.2025: Nahwärme kompakt 2025 in Stuttgart