A Comparative Analysis of Machine Learning Algorithms for Aggregated Electric Chargepoint Load Forecasting

Karlsruhe Institute of Technology, Institute for Automation and Applied Informatics, Hermann-von-Helmholtz-Platz 1, 76344 Eggenstein-Leopoldshafen, Germany

^* Corresponding author: chang.li@kit.edu

Abstract

With the development of electric vehicles in the last years, the number of electric chargepoints are expanding rapidly. Accordingly, the aggregated load demand from different electric chargepoints is increasing significantly. Due to the unpredictability of charging behaviour, it is difficult to build white-box models to analyse the patterns and to predict the load profiles, which is essential for other tasks such as demand side management. Thus, in this work, four different models based on machine learning and deep learning algorithms namely Random Forest (RF), Support Vector Regression (SVR), Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) are applied to a massive real-world open dataset from the UK, published in 2018, to compare the forecast performance of each algorithm with the modified persistence model as the baseline. The raw data are first pre-processed to generate the aggregated load demand by hour and then used for training and forecasting with a predictive horizon of 72 hours. The results are compared by using two common descriptive statistics, i.e., normalized Root-Mean-Square Error (nRMSE) and Mean Absolute Percentage Error (MAPE). In comparison we find that the GRU generates the lowest prediction error with 5.12% MAPE and 8.24% nRMSE in January 2017 and the modified persistence model generates the overall lowest prediction error with 2.88% MAPE and 3.76% nRMSE in July 2017.