Napovedovanje naročil za problem optimizacije v zadnjem kilometru : magistrsko delo

Gaberšček, Maj (avtor), Vidali, Janoš (mentor)

V magistrski nalogi obravnavamo napovedovanje števila naročil oziroma povpraševanja znotraj logističnega procesa dostave v zadnjem kilometru. Cilj naloge je generiranje natančnih napovedi števila naročil po regijah in časovnih oknih ter nato uporaba napovedi na praktičnem primeru problema optimizacije poti (reševanje problema VRP). Vhodni podatki so realni podatki anonimne spletne trgovine, ki so agregirani po regijah in časovnih oknih. Zaradi zakonodaje GDPR in interesa podjetja pravih podatkov ne smemo objaviti, zato jih je bilo treba najprej anonimizirati. V postopku anonimizacije smo standardizirali časovna okna, zapolnili luknje v podatkih in dodali šum. Nato smo definirali novo dostavno območje in na njem samodejno definirali smiselne regije (večkotniki) s pomočjo gručenja. Del anonimizacije je zajemal tudi smiselno dodelitev časovnih vrst na nove regije s pomočjo skaliranja in drugih transformacij. Anonimizirane podatke smo nato analizirali z namenom vpogleda v strukturo in odvisnosti, saj to omogoča bolj optimalne definicije napovednih modelov. Za napovedovanje smo najprej definirali nekaj statističnih modelov, s pomočjo katerih smo kasneje lahko primerjali dodano vrednost nekaterih kompleksnejših modelov. Od statističnih modelov smo primerjali model zadnjega tedna, model povprečja po dnevih v tednu, model uteženega tedenskega povprečja in agregacijski model Divide et impera. Od kompleksnejših modelov sta bila implementirana model naključnega gozda in model LSTM. Izkazalo se je, da je model Divide et impera dosegel najboljše rezultate med vsemi modeli. Iz tega lahko sklepamo, da kompleksnejši modeli ne presegajo rezultatov statističnih, zato njihova uporaba v praksi ni smiselna. Prav tako vidimo, da se nam bolj splača napovedovati število naročil za celotno območje in nato napoved razdeliti po regijah glede na njihovo velikost. Pomembnost predikcije v praksi smo preverili z uporabo simulacije problema optimizacije poti (VRP). Primerjali smo uspešnost in kvaliteto razvoza v primeru, ko kupcem dovolimo izbirati dostavna časovna okna poljubno in v primeru, ko nekaterim kupcem omejimo izbiro na podlagi napovedi. Pokazalo se je, da sta dobra napoved in omejitev dostavnih časovnih oken vplivali na več dostavljenih naročil, boljšo izkoriščenost vozil in manjše stroške na naročilo.

• Jezik: Slovenski jezik
• Ključne besede: napovedovanje povpraševanja, strojno učenje, časovne vrste, optimizacija dostave
• Vrsta gradiva: Magistrsko delo/naloga
• Tipologija: 2.09 - Magistrsko delo
• Organizacija: FMF - Fakulteta za matematiko in fiziko
• Leto izida: 2026
• PID: 20.500.12556/RUL-189767
• UDK: 519.8
• COBISS.SI-ID: 300613379
• Datum objave v RUL: 15.10.2026

Sekundarni jezik

• Jezik: Angleški jezik
• Naslov: Predicting orders for the last-mile delivery optimization problem

Izvleček

In this master’s thesis, we address predicting the store order demand within the last-mile delivery logistics process. The goal is to generate accurate predictions of the number of orders by regions and time slots and later apply the predictions to a practical example of the vehicle routing problem (VRP). The input data consists of real-world data from an anonymous online store, which is aggregated by regions and time slots. Due to GDPR and the company’s privacy concerns, we cannot publish the real data, so it was necessary to anonymize it first. The anonymization process involved standardizing time slots, filling empty gaps in the data and adding noise. We defined a new delivery area and used clustering on addresses to automatically define regions (polygons). Anonymized data was then scaled and applied to the newly defined regions. The final dataset was then analyzed to gain insight into the structure and dependencies, which later helped define more accurate prediction models. The first step of predicting was to define some baseline statistical models, which we later used for comparison with more complex ones. Statistical models included the last week model, the average by days of the week model, the weighted weekly average model and the Divide et impera aggregation model. We also implemented a random forest model and an LSTM model. Best results on the test dataset were achieved by the Divide et impera model, which shows that in this case, more complex models do not outperform simple statistical ones. That means that complex models are not worth the effort in practice. We can also see that due to the data structure (small regions with high noise), it is better to predict the number of orders for the whole area and then split the prediction by regions based on their size. The importance of prediction in the whole logistics process was verified by simulating the vehicle routing problem. We compared delivery success and VRP solution quality between a scenario in which customers could choose delivery time slots arbitrarily and a scenario in which some customers had limited choices based on the previously generated prediction. Tests confirmed that good prediction combined with limiting delivery time slots led to more delivered orders, better vehicle utilization and lower total cost per order for the logistics company.

• Ključne besede: demand prediction, machine learning, time series, delivery optimization