Train a single model per cluster

Currently we build one model per city. which is trained with the combined data from the cluster to which the city belongs to.
Although it sees the entire data (once) it is trained to minimize the prediction loss of a single city.

An alternative would be to using the same model train it to predict all the cities in the cluster, by running it through the data multiple times (once per each city)

Another possibility, would be to have it predict all cities at once by defining a multidimensional target of dimension `n_cities by pred_window`.