Learned upper bounds for the Time-Dependent Travelling Salesman Problem