Citation

Abdulelah, Aymen Jalil and Sonuç, Emrullah and Yassen, Esam Taha and Al-Andoli, Mohammed Nasser (2026) Uncertainty-aware deep reinforcement learning for sustainable electric vehicle routing: A hybrid optimization framework. Knowledge-Based Systems, 350. p. 116546. ISSN 09507051

Text main.pdf - Published Version Restricted to Repository staff only Download (1MB)

Abstract

Urban freight accounts for approximately 20% of transport-related greenhouse gas emissions, yet electric vehicle (EV) routing faces critical barriers: limited range, charging constraints, and energy uncertainty. Existing methods cannot quantify stranding risk or support chance-constrained battery feasibility, as energy consumption exhibits 30%–40% variability. We address the EVRPTW under operational uncertainty via probabilistic energy forecasting and hybrid optimization. Bidirectional LSTM networks deliver calibrated uncertainty quantification (94.7% empirical coverage for 95% confidence intervals), enabling model-based chance-constrained battery feasibility. Graph neural networks encode battery-aware spatial–temporal dependencies, while Proximal Policy Optimization with optional mixed-integer programming refinement balances real-time adaptability (0.3–0.8 s inference) with offline solution quality. Evaluation on 30 Solomon-derived EVRPTW benchmark instances (60–100 customers) shows an 8.8% cost reduction over the strongest learning baseline (MVMoE), rising to 12.2% under dynamic conditions. Real-world deployment with a 25-vehicle electric fleet over 13 weeks validates €143,000 in projected annual savings, 98.1% on-time delivery, 15.5% energy reduction, and 96% planning time reduction, with fleet emissions at approximate carbon parity with diesel on Italy’s transitional grid alongside elimination of local air pollutants. Transfer learning protocols reduce region-adaptation retraining effort by ∼95% (from ∼72 h to 3–5 h of fine-tuning), requiring only 400–600 local routes. This work demonstrates a pathway toward more sustainable urban logistics through uncertainty-aware optimization with measurable environmental, economic, and operational benefits.

Item Type:	Article
Uncontrolled Keywords:	Graph attention networks
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines > QA75.5-76.95 Electronic computers. Computer science
Divisions:	Faculty of Computing and Informatics (FCI)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	31 Jul 2026 05:16
Last Modified:	31 Jul 2026 05:16
URII:	http://shdl.mmu.edu.my/id/eprint/16394

Downloads

Edit (login required)