Unmanned Aerial Vehicle Delivery Route Planning in Cold Weather Using Neural Combinatorial Optimization and Adaptive Large Neighborhood Search
- 1School of Mechanical and Electrical Engineering, Shihezi University, Shihezi 832003, China
- 2College of Information Science and Engineering, Northeastern University, Shenyang 110169, China
Abstract
As the low-altitude economy grows rapidly, unmanned aerial vehicle (UAV) logistics is expanding into cold regions. Such operations must account for the coupled effects of low temperature, wind speed and direction, and decreasing payload on energy consumption and endurance; using conservative or average weather parameters may waste resources or underestimate operational risks. This study proposes a method for multi-UAV, multi-sortie delivery planning in cold regions that minimizes total operating cost while serving all customers and ensuring safe depot returns. The method formulates a multi-sortie energy consumption and endurance model accounting for temperature, wind fields, and payload changes, and couples neural sequence generation conditioned on known weather time series with continuous-physics route splitting and adaptive large neighborhood search to perform customer sequencing, sortie partitioning, and schedule improvement. Across test instances of three customer sizes, schedules generated under different planning conditions are evaluated under the same simulated dynamic weather. Results show that Dynamic planning completes 72/72 benchmark instances and reduces mean penalized cost by 32.1–65.9% compared with Typical planning (which completes 57/72), increases mean completion rate by 16.7–37.5 percentage points across three training seeds, and maintains higher completion rates on both in-range and extreme historical cold days. This study provides a planning method that balances operational feasibility and cost efficiency and quantifies the operational value of weather time-series information.
Related
- A Novel Framework for Energy-Aware Optimal Path Planning of Unmanned Aerial Vehicles
- EnerGNN: Learning Optimization-Compatible Energy Functions for Exact Constrained Combinatorial Inference
- On Sparse Graph Limits for Neural Combinatorial Optimization
- Super-Quadratic Quantum Speedups for Combinatorial Optimization via Tilted Walks
- Generic Machine Learning-Based Sparsification for Arc Flow Formulation of Combinatorial Optimization Problems
BibTeX
@article{fang2026unmanned,
title = {{Unmanned Aerial Vehicle Delivery Route Planning in Cold Weather Using Neural Combinatorial Optimization and Adaptive Large Neighborhood Search}},
author = {Jiankai Fang and Jianyou Xu and Hongkun Wang and Yan Qian and Shuo Zhang and Yuwei Song},
journal = {Mathematics},
year = {2026},
month = oct,
doi = {10.3390/math14213975},
url = {https://www.mdpi.com/2227-7390/14/21/3975},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Unmanned Aerial Vehicle Delivery Route Planning in Cold Weather Using Neural Combinatorial Optimization and Adaptive Large Neighborhood Search
AU - Jiankai Fang
AU - Jianyou Xu
AU - Hongkun Wang
AU - Yan Qian
AU - Shuo Zhang
AU - Yuwei Song
PY - 2026
DA - 2026/10/31
JO - Mathematics
DO - 10.3390/math14213975
UR - https://www.mdpi.com/2227-7390/14/21/3975
AB - As the low-altitude economy grows rapidly, unmanned aerial vehicle (UAV) logistics is expanding into cold regions. Such operations must account for the coupled effects of low temperature, wind speed and direction, and decreasing payload on energy consumption and endurance; using conservative or average weather parameters may waste resources or underestimate operational risks. This study proposes a method for multi-UAV, multi-sortie delivery planning in cold regions that minimizes total operating cost while serving all customers and ensuring safe depot returns. The method formulates a multi-sortie energy consumption and endurance model accounting for temperature, wind fields, and payload changes, and couples neural sequence generation conditioned on known weather time series with continuous-physics route splitting and adaptive large neighborhood search to perform customer sequencing, sortie partitioning, and schedule improvement. Across test instances of three customer sizes, schedules generated under different planning conditions are evaluated under the same simulated dynamic weather. Results show that Dynamic planning completes 72/72 benchmark instances and reduces mean penalized cost by 32.1–65.9% compared with Typical planning (which completes 57/72), increases mean completion rate by 16.7–37.5 percentage points across three training seeds, and maintains higher completion rates on both in-range and extreme historical cold days. This study provides a planning method that balances operational feasibility and cost efficiency and quantifies the operational value of weather time-series information.
SN - 2227-7390
ER -