A two-stage last-mile delivery platform: supervisor-guided transfer learning with LSTM and attention for demand forecasting, fleet calibration, and VRPTW optimisation

Benhra, Jamal, Labtiti, Aymane
OPSEARCH · Springer Nature · 2026-11-03 · DOI: 10.1007/s12597-026-01317-5

Abstract

Last-mile delivery planning depends on short-term demand forecasts and vehicle routing, but the step that turns a forecast into a fleet and a routing instance is rarely specified or evaluated. We present FALCON , a two-stage platform in which a rule-based supervisor classifies delivery zones as forecastable, baseline or rejected; a transfer-learned long short-term memory (LSTM) network with multi-head attention produces seven-day hourly forecasts with an upper bound calibrated on a held-out period; a sequential heterogeneous fleet-sizing rule converts payload and stop-time limits into orders per delivery wave; and a fully specified vehicle routing problem with time windows is solved by a genetic algorithm, particle swarm optimisation and continuous-domain ant colony optimisation sharing one random-key encoding, decoder, fitness and a repair with a proved coverage guarantee. On the public LaDe corpus (1.86 million orders, four other cities for pre-training, a leakage-free chronological protocol), the supervisor avoids large errors on sparse zones (99% vs. 85% weighted MAPE), whereas the network does not beat a four-week hour-of-week mean (47% vs. 34%) and transfer does not improve on training from scratch. Planning on the calibrated upper bound reduces orders beyond planned capacity from 21.5 to 6.8% at 63% higher fleet cost, with 80% test coverage against an 85% target. On 30 instances with 10–200 customers, the three metaheuristics perform similarly over repeated runs, while Google OR-Tools is better on every instance and matches exact optima on small ones.

Read at publisher →

Similar articles