Short-term aggregate electric vehicle charging load forecasting in diverse conditions with minimal data using transfer and meta-learning

Abstract

AbstractThe proliferation of electric vehicles (EVs) necessitates accurate EV charging load forecasting for demand-side management and electric-grid planning. Conventional machine learning-based load forecasting methods like long short-term memory (LSTM) neural networks rely on large amounts of historical data, which can be resource-intensive and time-consuming to collect. In this study, we employ Transfer Learning (TL) and Model-Agnostic Meta-Learning (MAML) for short term EV charging load forecasting. These methods involve pre-training a base model on a larger comprehensive EV charging dataset followed by fine-tuning using a few days’ worth of EV charging data in our target location. We find that the performance of both the TL and MAML models outperform traditional LSTM models and other classic machine learning models in the context of forecast accuracy when working in three different settings with limited data , load variance, and diverse geographical locations. The error metrics from TL and MAML are up to 24% and 61% lower than deep learning and classic machine learning models respectively.