Forecasting Session Characteristics for Electric Vehicle Charging Stations

Pavšič, Amadej (author), Demšar, Jure (mentor)

The rapid adoption of electric vehicles (EVs) strains existing power grids, creating a need for intelligent Energy Management Systems (EMS) driven by accurate forecasting. This thesis addresses the stochastic nature of EV charging by systematically evaluating three forecasting paradigms – aggregated grid occupancy, individual session parameters at plug-in, and generative session sequencing. Utilizing real-world datasets from workplace charging environments and applying transfer learning strategies, we evaluate a suite of models ranging from classical machine learning to state-of-the-art transformer architectures. The results demonstrate that aggregated forecasting using fine-tuned transformers achieves high accuracy and is viable for immediate deployment in EMS. Conversely, session-level prediction remains difficult due to feature scarcity and behavioral variance. However, the proposed autoregressive sequence model successfully captures macro-level charging patterns, proving its value for stochastic grid simulation despite limited deterministic accuracy.

• Language: English
• Keywords: electric vehicle charging, deep learning, transformers, time series, energy management system, transfer learning
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2025
• PID: 20.500.12556/RUL-177209
• COBISS.SI-ID: 262700547
• Publication date in RUL: 17.12.2025

Secondary language

• Language: Slovenian
• Title: Napovedovanje značilnosti sej za polnilnice električnih vozil

Abstract

Hitra širitev polnilnih postaj in uporabe električnih vozil (EV) predstavlja obremenitev za obstoječa elektroenergetska omrežja, kar narekuje potrebo po pametnih sistemih za upravljanje energije (ang. EMS), ki temeljijo na natančnem napovedovanju. Magistrsko delo raziskuje stohastično naravo polnjenja EV s sistematičnim vrednotenjem treh paradigm napovedovanja – agregirane zasedenosti omrežja, parametrov posamezne seje ob priklopu in generativnega zaporedja sej. Na podlagi realnih podatkovnih zbirk s polnilnih postaj ob poslovnih stavbah in z uporabo strategij prenesenega učenja smo ovrednotili nabor modelov – od klasičnega strojnega učenja do najsodobnejših transformerskih arhitektur. Rezultati kažejo, da agregirano napovedovanje s prilagojenimi transformerji dosega visoko natančnost in je zrelo za implementacijo v EMS. Nasprotno pa ostaja napovedovanje na ravni posamezne seje zahtevno zaradi vedenjske variabilnosti in pomanjkanja značilk. Vendar pa predlagani avtoregresijski model uspešno zajame makronivojske vzorce polnilnih sej, kar je kljub omejeni deterministični natančnosti ključno za stohastične simulacije omrežja.

• Keywords: polnjenje električnih vozil, globoko učenje, transformerji, časovne vrste, sistem za upravljanje energije, preneseno učenje