Machine learning (ML) has become a powerful and versatile tool capable of modeling a wide range of engineering systems, from materials to large‑scale infrastructure networks. In hydraulics, where water supply systems (WSS) exhibit strong nonlinearities, heterogeneous components, and tightly coupled operational dynamics, ML offers a promising alternative to traditional physics‑based simulators that are computationally demanding and difficult to calibrate. The potential of differential machine learning [1] —models that learn instantaneous temporal derivatives rather than absolute values — for accurate, time‑flexible simulation of WSS dynamics is demonstrated as well as its extension to composite, physically structured architectures capable of modeling complex, real‑world water supply networks. The proposed framework decomposes the WSS into a hierarchy of sub‑models representing pumps, valves, and local hydraulic processes, which are then integrated into a system‑level model through end‑to‑end training. By assigning each sub‑model physically meaningful inputs and outputs, the architecture enhances interpretability and reduces the risk of learning spurious correlations. At the same time, the global optimization ensures consistent interaction between subsystems, addressing limitations of independently trained modular approaches. This bypasses explicit time‑series modeling and offers flexibility regarding temporal resolution—crucial for real‑world datasets with heterogeneous sampling frequencies. The methodology is validated on a large‑scale, operational WSS featuring variable‑speed pumps, actively controlled valves, and mixed gravitational‑pumping regimes. Models are trained using both synthetic and real operational data, supported by a transfer‑learning procedure. Results show high predictive accuracy for electrical power and flowrates, with improved consistency and stability relative to monolithic differential ML models. Challenges remain in long‑horizon tank‑level prediction due to error accumulation during numerical integration, but the findings highlight the potential of differential and composite ML strategies to enable real‑time simulation, digital‑twin development, and operation‑aware decision support in modern hydraulic systems.