Publication

Abstract :

Electric Vehicles (EVs) play a very important role in the modern-day electrical ecosystem. Throughout the world, there is a surge in the number of EVs in the market owing to an increase in public demand and as a general policy by governments and manufacturers in order to foster sustainable, clean energy sources for transportation, which has minimal impact on the environment. While this takes place, on the other hand increase in demand of EVs over time has resulted in considerable stress in existing power systems and grids, which need to be analysed. This paper aims to perform electrical vehicle load forecasting using data from multiple charging stations which are split into a data-rich source domain and a data-scarce target domain. Transformer-based models are trained on the entire source domain and its respective clusters, obtained through Kmeans clustering based on usage patterns. A base model (BM1) is first trained on the entire source domain and another fine tuned model (FTM) is trained on clustered data. These models are then applied to the data-scarce target domain to analyse the impact of transfer learning in such scenarios. Based on the result obtained, the transfer learning model which was fine tuned using cluster based FTM in the source domain, i.e Transfer learning model 2 (TLM2) was able to outperform models trained from scratch on the target domain,i.e Base model 2 (BM2) and model trained on the entire source domain Base model (BM1),i.e Transfer learning model 1 (TLM1) by almost 3 percent in R2R2 metrics and the results show an improvement of 1.011.01 RMSE compared to 3.4 using traditional Bi-LSTM based models. This shows how the integration of attention-based architectures with cluster-based knowledge provides an effective mechanism for load forecasting.