Optimization of district heating system is computationally cost intensive as it requires an important number of simulations for such complex systems, at each time steps, over long periods of time. This issue is a bottleneck clearly identified in the literature. A solution consists in simulating the system with several short time series, representative/typical of the whole period analyzed. The challenge is then to define the typology of data supposed to be representative. In this paper, a method is proposed to defined a typology of some inputs for DH modeling and optimization. The methodology is based on the daily demands of the DH substations characterized by aggregated key parameters which quantify the load and its variability. A k-means based clustering is done where k is determined by implementing a multi-criteria decision aiding algorithm. At this stage, the criteria used are the intra and inter- clusters distances and the similarity of the demands in a same cluster. To generate typical days a compromise between all these criteria needs to be done. The impact of this choice on the accuracy of simulations using time series archetypes is assessed.This methodology is tested on a part of Nantes’ district heating in France. It is shown the accuracy of simulations drops when the number of clusters is too low, whereas the computational time growths with the number of clusters. The comparison with real data confirms the usefulness of combining the clustering method with MCDA to detect the best compromise between computational time and accuracy.