Using Spatial and Temporal Context for Predicting Energy Consumption of Electric Vehicles
Authors
Madsen, Bjarke Bak Toxværd ; Nørskov, Christoffer Popp ; Bundgaard, Jacob Hjort
Term
4. term
Education
Publication year
2019
Submitted on
2019-06-11
Pages
68
Abstract
Antallet af elbiler vokser år for år. Trods bedre batterier og længere rækkevidde oplever mange stadig bekymring for at løbe tør for strøm, kaldet rækkeviddeangst. Mere præcise forudsigelser af energiforbrug kan afhjælpe dette ved at give et mere sikkert skøn over, hvor langt man kan køre. I afhandlingen udvikler vi flere maskinlæringsmodeller, der forudsiger elbilers energiforbrug ved at udnytte kontekstuel information om de enkelte ture. Grundlaget er et stort datasæt med elbilkørsel på det danske vejnet, som vi beriger med eksterne kilder, bl.a. højdeoplysninger fra et højderaster over Danmark og hastighedsdata fra Vejdirektoratets speedmap. Vores resultater viser, at et rekurent neuralt netværk (RNN) – en model, der kan lære af sekvenser af data – udnytter den sekventielle struktur i konteksten og forbedrer præcisionen markant i forhold til en simpel baseline-model.
The number of electric vehicles is growing every year. Despite longer-range batteries, many drivers still worry about running out of charge—a concern known as range anxiety. More precise predictions of energy use can reduce this worry by giving drivers a clearer estimate of how far they can drive. This thesis builds several machine learning models that predict the energy consumption of electric vehicles by using contextual information about each trip. The work is based on a large dataset of EV driving on the Danish road network and is enriched with external sources, including elevation from a national height raster and speed information from Vejdirektoratet (the Danish Road Directorate). We find that a Recurrent Neural Network (RNN)—a model designed to learn from sequences—takes advantage of the sequential nature of the context data and performs significantly better than a simple baseline model.
[This abstract was generated with the help of AI]
Keywords
Documents