9 Future perspectives

The first future perspectives of this work are to address the limitations mentioned earlier. In the first instance, obtaining actual data on the transformers should be easy. Then, the computation method for hot water demand in QBuildings should be considered. A characterisation by building category and construction period, such as the one conducted by Mr Girardin for the SH demand, could be done with field measurements. Lastly, the testing of the starting hypothesis - REHO results are assimilable according to the clusters should be implemented. The last should conclude on what is the best methodology: clustering algorithm - Kmedoids or GaussianMixture -, input data - Base scenario or Per ERA scenario-, and whether the outliers should be removed or not.
It would also be interesting to do clustering on another region and compare the results with those obtained for Geneva. One way of comparing could be to use another type of ML algorithm family: classification algorithms. With the archetypes obtained in Geneva, a classification method can be developed to attribute labels to the transformer districts from another city. The labels from the classification are then put in perspective with the ones from the clustering.

If despite the fact that these limitations have been overcome, the results are not convincing, in order to facilitate district energy optimisation, other approaches may be necessary. Thus, when looking at the results of one optimisation, some links between the solutions appear, especially between the units chosen. For example, when solar panels are installed as the main electricity supply (with a nearly maximised capacity), it is very likely to also have in the solution the installation of a heat pump for the heating needs. The duo is often coupled with an electric radiator or a water heater for the too extreme periods. Instead of clustering upstream of REHO to reduce the calculation time, the idea would be to cluster downstream on the solutions. The aim is to define a set of solutions, among which the model must then decide on the best one according to the scenario. The clustering of solution allows to move from a MILP model, with a high number of decision variables, to an MILP problem, where only the sizing is a non-integer variable. It would allow for less optimised but much faster solutions at the building scale. The main shortcoming of this approach would probably be its difficulty to scale up by doing a centralised optimisation. It would also require many optimisations with different inputs to consider that the proposed solution set provides a sufficiently satisfactory solution for any of the building situations it might be asked to solve.

Master thesis, Spring 2022

Joseph Loustau