Proceedings of ICLT 2025
PREDICTING LOGISTICS DELAYS FOR RESILIENT SUPPLY CHAINS USING DEEP LEARNING
Panupong Wanjantuk; Ruth Banomyong
Department of Computer Engineering, Khon Kaen University, Thailand; Center of Excellence in Connectivity, Thammasat Business School, Thammasat University, Thailand
International Conference on Logistics & Transport 2025, Tokyo, Japan, pp. 167-176
Abstract
Purpose: Building resilience in logistics systems is essential to navigating disruptions, reducing risk, and ensuring operational continuity. Despite growing disruptions, existing delay prediction models often lack interpretability and adaptability. This study addresses that gap by developing and explaining a deep learning-based model to predict delivery delays, thereby supporting proactive logistics decision-making and enhancing supply chain resilience. Design/methodology/approach: We use a real-world logistics dataset containing over 32,000 hourly records from a Southern California network collected between 2021 and 2024. Data features span traffic congestion, ETA variation, port activity, vehicle behavior, and IoT sensor data. After preprocessing and feature selection, we trained 10 Multi-Layer Perceptron (MLP) models with varying depths and dropout rates. Binary classification was performed using the delay_probability variable, with class imbalance handled through class weighting in the loss function. Model performance was evaluated via accuracy, precision, recall, and AUC. The best-performing model was further explained using Captum’s Integrated Gradients method to identify key contributing features. Findings: The optimal MLP model achieved 72.8% accuracy and high recall, effectively identifying delayed deliveries. Using Captum’s Integrated Gradients, input features were ranked by their contribution to model predictions. The top features included ETA variation, port congestion level, weather severity, and driver fatigue score. These insights enhanced model transparency and supported operational trust and interpretability. Research limitations/implications (if applicable): This study is based on a historical static dataset; future research may incorporate live data feeds and spatiotemporal event modeling for real-time prediction capabilities. Extension to multi-class delay severity or integration with route optimization engines is also suggested. Practical implications (if applicable): The proposed model serves as an early-warning tool for logistics planners, enabling timely identification of potential delivery delays. By providing interpretable predictions, the tool supports proactive decision-making such as rerouting shipments, reallocating warehouse resources, and adjusting driver assignments. These capabilities directly contribute to maintaining business continuity, minimizing service disruptions, and implementing adaptive logistics strategies in the face of uncertainty—whether caused by traffic congestion, port delays, or extreme weather events. The approach helps organizations transition from reactive to predictive logistics management, strengthening resilience across the supply chain. Originality/value: This research presents a practical, explainable application of deep learning to supply chain risk prediction, contributing to the literature on resilient logistics systems and the role of AI- driven decision-support tools.
Keywords
Resilient Logistics; Delay Prediction; Deep Learning; Explainable AI; Supply Chain Analytics
Citation
Panupong Wanjantuk; Ruth Banomyong (2025). PREDICTING LOGISTICS DELAYS FOR RESILIENT SUPPLY CHAINS USING DEEP LEARNING. Proceedings of the International Conference on Logistics & Transport (ICLT 2025), Tokyo, Japan, pp. 167-176.