Machine learning for district heating and cooling: A review of open problems and deployment barriers

District heating and cooling (DHC) systems are central to urban decarbonization, yet they present machine learning (ML) challenges distinct from other energy domains. Thermal inertia creates multi-hour prediction delays, pipe networks impose graph-structured hydraulic coupling, seasonal demand varies by orders of magnitude, and safety-critical operation demands guarantees most data-driven methods cannot provide. This review analyzes seven open problems the field must resolve for real-world impact. For load forecasting, feature engineering consistently outweighs architectural innovation, ensemble methods match deep learning for sub-day horizons at far lower cost, and transformers demonstrate superiority only beyond 48 hours. For control, MPC improvements range from 5% to 75% depending on untapped flexibility, while deep reinforcement learning excels in hybrid discrete–continuous action spaces under forecast uncertainty. For fault detection, labeled data scarcity is the fundamental bottleneck. Supervised methods achieve 95–99% accuracy, but synthetic data generation remains insufficient for practical use. The most critical structural finding is the underrepresentation of district cooling despite fundamentally different physics (nonlinear chiller thermodynamics, phase-change storage, solar-driven demand) requiring purpose-built ML approaches. Few studies validate their methods through field deployment, revealing a persistent gap between algorithmic innovation and operational impact. A research roadmap poses 13 specific questions the field must answer.