Integrating Predictive and Optimization Model for Intelligent Schedule Management based on Real Time ETA Information

Integrating Predictive and Optimization Model for Intelligent Schedule Management based on Real Time ETA Information: A Concept of Machine Learning and Exact Solution in Petrochemical Loading Facility

Emanuel Febrianto Prakoso, Emanuel (TU Delft Civil Engineering and Geosciences)

Tavasszy, Lorant (graduation committee)
Maknoon, M.Y. (graduation committee)
Pel, A.J. (graduation committee)
Vanga, R. (mentor)

Delft University of Technology

Transport, Infrastructure and Logistics

2021-10-06

This study addresses the truck rescheduling problem as the consequence of uncertain arrival time. It proposes an integrated system of predictive model powered by machine learning algorithm and exact optimization model such that it is distinct from most existing literatures in this domain. The uncertainty of truck arrival time is captured as presence probability by the developed predictive model. Subsequently, a Mixed-Integer Quadratic Programming (MIQP) is built to solve the Probabilistic Slot Rescheduling Problem (P-SRP) in which the rescheduling is subject to expected value constraint that incorporates the presence probability of incoming trucks. The objective is to minimize the expected cost of rescheduling that would lead to more efficient and robust operation. In regard to the predictive model, evaluation according to the standardized KPI shows the ANN is the best algorithm to fit the input historical data with overall F1 score of 73%. Moreover, adding real-time elements enhances the prediction result by 20%. In regard to optimization model, the P-SRP model results on expected cost that is 42% lower than the P-SRP model. Numerical experiment based on multiple scenarios indicates that the proposed solution yields optimal added values in situation characterized by high number of reschedule being required, large scale of operation in terms of quantity of loading bay and utilization rate, the priority is to maintain initial schedule, and there is no stand-in loading bay available. Lastly, it is found that increasing operational efficiency comes at the expense of the schedule robustness, although the value is negligible. This study is limited to the conceptual setting and the use of synthetic data; therefore it should be extended to include empirical result.

Schedule Management
Machine Learning
Exact Method
Robust Optimization
Mixed Integer Programming
Decision Support System
Slot Allocation
Prediction Model
Probabilistic classification

http://resolver.tudelft.nl/uuid:95c0caf0-bced-442e-8442-3c9db710ef8a

Student theses

master thesis

© 2021 Emanuel Emanuel Febrianto Prakoso