Dynamic Clustering of PI-Hubs Based on Forecasting Demand in Physical Internet Context

Abstract

This paper proposes an approach to reduce the complexity of hub’s connection. The approach is based on the concept of dynamic clustering for the retailers’ demand, solving retailers to PI-hubs’ clusters assignment problem, and then tackling a routing problem for each cluster. The dynamic clustering is based on a forecasted demand calculated from learning algorithm, Long Short-Term Memory (LSTM) Recurrent Neural Network. After implementing the clustering method, a Mixed Integer Linear Programming model is used to solve the problem of assigning retailers to clusters. Besides, this paper evaluates the clustering performance using Hopkins statistic and Silhouette width scores. The experiments and the results show that the dynamic clustering by using K-Medoid models provide a better performance and reduce the complexity of the transportation problem between PI-Hubs and retailers in the context of the Physical Internet.