Percorrer por autor "VAZ, ANA RITA SIMÕES DO PAÇO E"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- A hierarchical data-driven approach for last-mile delivery route planning using LSTM-based district sequencingPublication . VAZ, ANA RITA SIMÕES DO PAÇO E; Ramos, António José Galrão; Salimi, FarzamLast-mile delivery route planning in large-scale urban contexts is a complex problem, where direct optimisation at the stop level can become computationally demanding and difficult to interpret operationally. This dissertation develops and evaluates an integrated hierarchical data-driven approach to route planning, combining capacity-based grouping, LSTM-based district sequencing, and stop-level route construction. The study uses historical operational data from a parcel delivery company operating in the Porto region, structured according to a three-level district hierarchy. The use of historical routes is motivated by the assumption that they contain not only spatial visitation patterns but also tacit knowledge accumulated by drivers through daily operational experience. After data preparation, cleaning and validation, the final dataset comprised 507 624 records, 8 236 routes and 156 operational days. The proposed grouping strategy generated capacity-compatible groups while preserving the hierarchical district structure, producing 701 groups across the 17 test days, all respecting the maximum vehicle capacity. District sequencing was formulated as a hierarchical sequential prediction task, with independent LSTM models trained at the L1, L2, and L3 levels. The results showed that the model was especially effective at the most aggregated level and remained competitive at lower levels. The final integrated solution transformed the generated groups into complete stop-level routes, combining learned historical patterns with spatial continuity rules. Compared with the historical routes, the proposed solution reduced the estimated total distance, although the CVRP baseline produced lower distances by directly optimising at the stop level. Overall, the results show that the proposed approach delivers a fast, modular, and interpretable solution to large-scale last-mile delivery planning, balancing capacity feasibility, territorial coherence, and historical operational knowledge.