Gaussian process modelling of petrol sales in gas stations in Finland

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis

Department

Mcode

SCI3044

Language

en

Pages

37+5

Series

Abstract

Petrol delivery is an important challenge for the gas station networks operating hundreds of stations. At the moment, a common approach is to react to the low product level in the tanks using some IoT device and request a delivery based on an alert. The same devices are usually capable of recording the daily sales of petrol. This collected data can be used for analysis and forecasting. The array of daily petrol sales can be analyzed as a time series with modern machine learning methods. The machine learning methods can be used to model sales and accurately predict future demand. With that knowledge, the delivery can be scheduled in advance, minimizing the logistics cost for a compound product delivery problem. Bayesian machine learning, a special class of machine learning algorithms, is also capable of estimating the confidence levels of the prediction, allowing to calculate and account for the risks in case of the series deviation from the predicted values. We propose a Gaussian process model from the Bayesian machine learning methods for forecasting petrol demand for a week forward that can be used to create an optimal delivery schedule to minimize business expenses. Several forms of the model are described, and the model performance is compared to a popular model used in time series analysis SARIMA, which is used as a baseline in this thesis. The models are evaluated on a sample petrol station in Helsinki, Finland.

Description

Supervisor

Marttinen, Pekka

Thesis advisor

Poryaz, Onir
Talvitie, Ossi

Other note

Citation