Efficient training of learning-based thermal power flow for 4th generation district heating grids

Computing the thermal power flow, i.e., determining the grid state consisting of temperatures, pressures, and mass flow rates for given heat flows of consumers and suppliers, is an important element of various applications in district heating grids. It is classically done by solving the nonlinear heat grid equations iteratively but can be sped up by orders of magnitude using learned surrogate models such as neural networks. In this paper, we propose to speed up the model-building process for such learned models via a novel scheme for generating a suitable training data set. By sampling exemplary consumer or supplier mass flow rates instead of their heat flows, we avoid the iterative solution process during training data generation. We show with simulations for test settings with typical features of 4th generation district heating grids, such as multiple decentral heat sources and meshed grid structures, that the new approach can reduce training set generation times by one to two orders of magnitude compared to sampling heat flows, without loss of relevance of the training data set. Moreover, we show that training a surrogate model with a training data set significantly outperforms sample-free, physics-aware training approaches.