Capturing Transformer Overloads through Ensemble-Based Forecasting of EV Charging Station Consumption: An Evaluation of Model Diversity and Its Impact on Detection Performance
Authors
Søgård, Casper ; Truong, Sebastian
Term
4. term
Education
Publication year
2025
Submitted on
2025-06-06
Pages
100
Abstract
The rapid growth of electric vehicles has led to many new charging stations, putting extra stress on the power grid, especially local transformers. This thesis examines how to forecast electricity use at EV charging stations to reduce the risk of transformer overloads. We use hourly time series data from two sources: a real-world dataset from Boulder, Colorado, and a synthetic dataset representing Danish residential charging. The study focuses on two tasks: (1) forecasting consumption, evaluated with MAE (mean absolute error) and Huber loss, which reduces the influence of extreme values; and (2) detecting overloads, evaluated with recall (the share of overload events correctly identified) and the same error metrics. We propose an ensemble model that combines diverse base learners (e.g., xPatch, PatchMixer, GRU, LSTM, AdaBoost, Random Forest, and Gradient Boosting) trained on bootstrap samples, i.e., random resamples of the training data. Results show that deep learning models—particularly PatchMixer and xPatch—perform best for both forecasting and overload detection. However, all models struggled with peak consumption. The ensemble delivered mixed performance, largely due to unstable configurations and weak base learners.
Den hurtige udbredelse af elbiler har ført til mange nye ladestandere, hvilket lægger pres på elnettet og især på lokale transformere. Denne afhandling undersøger, hvordan man kan forudsige elforbrug ved ladestandere for at mindske risikoen for, at transformere bliver overbelastet. Vi arbejder med tidsseriedata på timebasis fra to kilder: et virkeligt datasæt fra Boulder, Colorado, og et syntetisk datasæt, der repræsenterer danske boliger. Der fokuseres på to opgaver: (1) at forudsige forbrug, vurderet med MAE (gennemsnitlig absolut fejl) og Huber-tab, som dæmper indflydelsen af ekstreme værdier; og (2) at opdage overbelastninger, vurderet med recall (andelen af overbelastninger, der bliver korrekt fundet) samt de samme fejlmål. Vi foreslår en ensemblemodel, der kombinerer forskellige grundmodeller (fx xPatch, PatchMixer, GRU, LSTM, AdaBoost, Random Forest og Gradient Boosting) trænet på bootstrap-prøver, dvs. tilfældige genudtræk af data. Resultaterne viser, at dybe læringsmodeller—især PatchMixer og xPatch—klarer sig bedst både i forbrugsforudsigelse og i at opdage overbelastninger. Alle modeller havde dog svært ved forbrugstoppe. Ensemblet gav blandede resultater, primært på grund af ustabile konfigurationer og underpræsterende grundmodeller.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords