On Sustainable Supply Chains: Optimal Design of a Multimodal Logistics Network with Shipment Consolidation, Stochastic Demand, and Machine Learning

Abstract

In this thesis, we consider the logistics network of a multi-echelon multimodal supply chain with multiple products and components taking economic and environmental sustainability, and shipment consolidation into consideration. Procedures for calculating and using both the water and carbon footprints of the network as metrics for its environmental sustainability are also explored. The supply chain logistics network is modelled as a Mixed Integer Linear Program (MILP) and then tested on randomly generated but realistic test instances. The effects of shipment consolidation on the economic and environmental cost of operations are analysed with results showing that consolidation decreases the supply chain (SC) cost especially when the distance between the shipper and receiver is significant. Considering that in reality, some of the parameters of supply chain network models might be stochastic, experiments are carried out with the designed MILP model having its demand parameter as stochastic. With the continual digitalization of supply chain processes leading to the automatic generation of data, machine learning (ML) has evolved as a methodology with the potential to help optimize stochastic models with its increasingly accurate predictions of future occurrences due to the continuous innovation of new algorithms. ML approaches to predicting stochastic parameters using historical data are evaluated in comparison to the more traditional stochastic programming approaches over multiple prediction periods. The three ML models utilized, Attention CNN-LSTM (AC-LSTM), Attention ConvLSTM (ACV-LSTM) and an ensemble of both models using Support Vector Regression (Ensemble-SVR), performed significantly better than the stochastic programming approaches considered (Simple recourse programming and Chance-constrained programming) in all scenarios. The MILP models using the predictions from the ML algorithms obtained the highest value of stochastic solution (VSS) and had the lowest expected value of perfect information (EVPI). This makes a case for the continued integration of ML prediction methodologies into stochastic optimization modelling.