Capacity Evaluation and Configuration of Low-Altitude Medical-Logistics UAVs Using a Semi-Open Queueing Network and a Deep Neural Network
Abstract
This paper proposes a capacity-planning and performance-evaluation framework for low-altitude UAV medical logistics by integrating a semi-open queueing network (SOQN), approximate mean value analysis (AMVA), and a deep neural network (DNN). First, a multi-branch SOQN model is constructed according to the practical operating process of UAV medical delivery, including UAV flight, blood-bank loading, hospital unloading, return flight, and base inspection. Stability conditions, maximum throughput boundaries, and approximate order response-time expressions are then derived. Latin hypercube sampling (LHS) and AMVA generate supervised labels for an approximate fixed-topology DNN surrogate. An independently implemented discrete-event simulation (DES) comparison over 34 analytically stable scenarios gives overall AMVA response-time and utilization MAPEs of 3.73% and 0.247%. Response-time MAPE rises to 21.49% at or above 90% utilization, so those high-load magnitudes are treated as boundary stress tests rather than precise numerical validation. At $\lambda =0.40$ orders/min, six UAVs violate the stated limits, whereas seven satisfy them. On 5,014 held-out feasible cases, the tuned DNN achieves 93.62% Top-1 accuracy, 99.70% Top-2 accuracy, and Macro-F1 of 0.9257. The generator retained 50,000 feasible scenarios after rejecting 30,229 infeasible candidates before reaching its target. Because the DNN lacks a no-feasible-fleet (NF) output class, DNN-only results are provisional; certified decisions require complete AMVA enumeration over $R\in \{5,\ldots,20\}$ , which returns NF when the feasible set is empty. DNN performance claims apply only to the declared topology and sampled domain. Operational logs were unavailable; near-saturation, changed-topology, and deployment decisions require simulation and operational verification.