District heating networks have proven their higher conversion efficiency, economic viability and environmental benefits when compared to decentralized and individual heating systems. These benefits are achieved through the ability to incorporate a wide variety of production means, including renewable intermittent sources but also via the use of short-term and/or inter-seasonal storage. Due to the numerous interactions between these components, their different dynamic aspects and operating constraints, physical simulations are computationally heavy so that running optimization tasks become prohibitively expensive and time consuming. Therefore, new control optimization schemes need to be drawn up to accelerate the predictive control and to facilitate the decision-making process. In the present work, we assess the application of geometric deep learning as a surrogate modeling framework for district heating simulations. Beyond processing non-Euclidian data, this deep learning approach aims to encode geometric and topological understandings of data as inductive biases in deep learning models. More precisely we trained Graph Neural Networks to emulate a thermo-hydraulic simulator of district heating network. This statistical inference method allows us to drastically reduce simulation time, hence unlocking further optimization loops and parametric space exploration. In addition, their permutation equivariance and stability to perturbations are assessed to discuss their scalability to more complex network topologies and control schemes.