A machine-learning-based column generation heuristic for electric bus scheduling

Bus scheduling in public transit consists in determining a set of bus schedules to cover a set of timetabled trips at minimum cost. This planning process has evolved recently with the advent of electric buses that introduce constraints related to vehicle autonomy and battery charging process. In particular, column-generation algorithms have regained popularity for solving problems similar to the one considered in this paper, namely, the multi-depot electric vehicle scheduling problem (MDEVSP) with a piecewise linear charging function and capacitated charging stations. To tackle large-scale MDEVSP instances, we design a column generation (CG) heuristic that relies on reduced-sized networks to generate the bus schedules. The reduction is achieved by selecting a priori a subset of the arcs. Multiple selection techniques are studied: some are based on a greedy heuristic and others exploit a supervised learning algorithm relying on a graph neural network. It turns out that combining both selection types yields the best computational results. On 405 artificial instances involving between 568 and 1474 trips and generated from real bus lines in Montreal, the network reduction technique produced an average computational time reduction of 71.6% (compared to the same CG heuristic but without network reduction), while deteriorating solution cost by an average of 2.2%. On 8 larger instances containing more than 2500 trips on average, the proposed solution method also provided an average time saving of 52.5% with an average gap of 4.2% thanks to a transfer learning approach.