Container cars on a freight train might be following one another in the same direction, but no two are alike. Not only are they different sizes and weights, but their origin and final destination points may vary as well. For Canadian National (CN), which operates thousands of trains every week, train configuration and load planning are akin to solving a giant jigsaw puzzle. For help with that task, CN partnered with the Université de Montréal Department of Computer Science and Operations Research (DIRO), one of our members, to create the Chair on Optimization of Railway Operations.
Cars are loaded onto a train according to the destinations of the containers they carry. On a train leaving Montreal for Vancouver, for example, a block of cars bound for Toronto will be unloaded first, so they must be loaded at the end of the train. Once in Toronto, however, more freight will be added to the train, and those cars need to be loaded based on their destinations. Added to that are size and weight constraints: for example, a heavy container must not be stacked atop a lighter one. “It is all these constraints in combination that make intermodal terminal operations so complex,” notes Johanne Dandurand, Director of Innovation and Multimodal Optimization at CN. There are two steps to solving the jigsaw puzzle.
The first is to analyze demand, to determine the number and types of cars that will be required to convey the various containers to their respective destinations. This amounts to a prediction problem, which can be solved using machine-learning algorithms that are fed with historical data on freight conveyances. CN therefore supplied the team led by Emma Frejinger, associate professor at the DIRO and holder of the CN Chair, with four years’ worth of data on container destinations, dates and characteristics. Accounting for external factors likely to influence freight transportation—such as economic cycles, days of the week, and even the variable date of the Chinese New Year—algorithms identified trends including changes over time in the number of containers in circulation and their characteristics. Using those trends, they generated predictions of demand for train cars.
With the demand estimated, the second step consists in determining the optimum train configuration as well as the best container-loading sequence. These two factors will enable optimized intermodal terminal operations. Using the predictions provided by machine-learning algorithms, operations research algorithms can be used in turn to help define that optimal solution.
“This helps crane operators make decisions, by telling them where to put a given container,” explains Ms. Frejinger, who is also a member of the Interuniversity Research Centre on Enterprise Networks, Logistics and Transportation (known by its French acronym, CIRRELT), which is itself a member centre of IVADO.
The two steps are therefore closely linked, as the degree of optimization of the solutions identified depends on the degree of accuracy of the predictions. The primary challenge is therefore to obtain the most accurate predictions possible. Unforeseen circumstances can arise, however, and distort the predictions. This leads to a further challenge: defining robust solutions that ensure optimization even when the unexpected happens.