AN MCDM-GOVERNED END-TO-END MACHINE LEARNING FRAMEWORK FOR TRUCK DELAY PREDICTION

Abstract

Delays in the arrival of trucks lead to queues at the entrance, idle resources and knock-on disruptions; however, in most studies, optimisation is carried out in a standalone manner based on a single model. This paper proposes an end-to-end architecture combining a leak-proof data pipeline, a suite of delay prediction models, a minimal MLOps cycle and a two-level multi-criteria decisionmaking (MCDM) layer. At the first level, model configurations are ranked according to accuracy, stability, latency and operational complexity. Level 2 selects a response policy for each predicted latency range based on operational key performance indicators (KPIs). Weights are determined using the AHP method, whilst alternatives are ranked using the TOPSIS method. The article describes the problem formulation and each level of the conceptual model, and sets out an evaluation framework based on 85,410 trips.