District Heating Networks (DHNs) are a key component of thermal energy system decarbonization, and their planning, design, and operation place conflicting demands on simulation tools with respect to accuracy, flexibility, and computational efficiency. This thesis proposes a hybrid modeling strategy that combines physics-based simulation with machine learning surrogates, selectively applied across different phases of the DHN lifecycle to balance these requirements.
For the detailed design phase, the thesis argues that high-fidelity, component-based physical modeling remains necessary to ensure accuracy, interpretability, and robustness. This is achieved through the development of PyDHN, an open-source Python framework emphasizing modularity and component flexibility. The framework is systematically verified against benchmarks, experimental data, and operational measurements from the DHN of Verbier (Switzerland), showing strong agreement despite measurement uncertainty. Although computationally intensive, PyDHN provides reliable reference solutions for engineering analysis and for generating high-quality training data.
The thesis then focuses on the operational phase, where computational speed becomes a primary constraint. A hybrid differentiable framework is introduced in which the hydraulic solver is replaced by a topology-specific Graph Neural Network (GNN) surrogate, while thermal dynamics are retained in a fully differentiable physical model optimized for GPU execution. A novel cycle-based graph representation allows the surrogate to capture global pressure-flow dependencies in meshed networks. Compared to fully physics-based simulations, the proposed approach reduces computational time by a factor of 100 while maintaining high predictive accuracy. Its differentiable structure further supports gradient-based optimization of operational control strategies, which outperform heuristic and gradient-free methods.
Finally, the thesis considers the pre-design phase, characterized by the need to rapidly explore alternative network layouts. A topology-agnostic GNN surrogate is developed and trained on a large dataset of synthetic networks. By incorporating physics-informed inductive biases, the model generalizes to unseen topologies and exhibits zero-shot transfer to the Verbier network. While slightly less accurate than topology-specific surrogates, it provides sufficient fidelity for early-stage planning at minimal computational cost.
Overall, this thesis shows that hybrid physics-machine learning modeling offers a practical and scalable approach to DHN simulation across the entire lifecycle. Rather than replacing physical models, data-driven surrogates complement them, supporting real-time optimization and large-scale design exploration for sustainable thermal energy systems.
