Passenger flow pattern learning based on trip counting in lift systems combined with real-time information

Conventional Control in vertical transportation systems may use information about passenger flows in order to estimate the number of passengers behind each landing call and to assess the destination of these possible passengers. This information supports the lift dispatching algorithm by giving it the opportunity to implement specific strategies for different circumstances. This paper proposes a new method to identify passenger flows in advance, using historical trip counting information summarized into origin destination matrices for short periods of time. Using these matrices, a clustering procedure can identify periods of homogeneous flow present in the data, learning the main traffic flow and providing a long-term view about the traffic profile in which the system is working. Real-time information about the traffic measurements extracted from the information transmitted to the dispatching algorithm can provide the short-term view. By mixing long-term and short-term information it is possible to estimate the expected values of the unknown quantities. The benefits of this process are tested against the Multiple Travelling Salesman Problem (MTSP) where the salesman corresponds to cars and the cities correspond to landing and car calls. The MTSP is the core of a stochastic bi-level optimization problem when the genetic algorithms are applied to the lift dispatching problem.