Capturing Transformer Overloads through Ensemble-Based Forecasting of EV Charging Station Consumption: An Evaluation of Model Diversity and Its Impact on Detection Performance

Søgård, Casper ; Truong, Sebastian

4. term

Software, Master

2025

2025-06-06

100 pages

Givet den store udvikling af elbiler og de globale FN's verdensmål om Bæredygdige byer og samfund, stiger antallet af ladestandere markant, hvilket belaster elnetværket mere for hvert år som går. Derfor vil studiet imødekomme udfordringerne ved at forudsige ladestandernes elforbrug for at mindske de risici som overbelastningerne kan forårsage på transformere. Ved at bruge tidsserie data aggregeret på time basis, vil dette studie analysere data fra virkelighedens verden fra byen Boulder i Colorado samt for et syntetisk dataset for danske hjemmeladestandere. Rapporten vil gennemgå to forsøg, hvilket er: (1) at forudsige elforbruget vha. MAE og Huber loss samt (2) opfange overbelastninger vha. recall og MAE/Huber loss. Til dette har rapporten udarbejdet en ensemble model, som med forskellige base learners (xPatch, PatchMixer, GRU, LSTM, AdaBoost, Random Forest, Gradient Boosting) trænes med bootstrap sampling. Studiets resultater peger på at deep learning modeller, især PatchMixer og xPatch, udkonkurrerer ensemble modellen og de individuelle modeller, både i forudsigelser og i at opfange overbelastninger, dog med udfordringer i at fange høje værdier. Idet ensemble modellens resultater ikke var fyldestgørende, var problemer såsom ustabile konfigurationer samt dårligt ydende base learners en af hovedårsagerne til dens dårlige præstation.

The rapid adoption of electric vehicles (EVs), driven by global sustainability goals, has led to widespread deployment of electric vehicle charging stations (EVCS), raising concerns about the stability of existing power infrastructure. This study addresses the challenge of forecasting EVCS electricity consumption to help mitigate transformer overload risks. Using time series data aggregated at an hourly resolution, we analyze a real-world dataset from Boulder, Colorado and a synthetic Danish residential dataset. Two main tasks are explored: (1) consumption forecasting using MAE and Huber loss, and (2) overload detection using recall and MAE/Huber Loss. An ensemble model is proposed, combining diverse base learners (e.g. xPatch, PatchMixer, GRU, LSTM, AdaBoost, Random Forest, Gradient Boosting) trained on bootstrap samples. Results show that deep learning models, particularly PatchMixer and xPatch, outperform ensemble and standalone models in both forecasting and overload detection, though all models struggled with peak consumption values. While ensemble performance was mixed, issues were largely attributed to unstable configurations and underperforming base learners.

ladestandere ; ensemble ; forecasting ; AI ; deep learning ; ensemble learning ; transformer overload ; power grid

Download
View record in AAU Student Projects