A Meta-Learning Approach for Electric Vehicles' Charging Load Demand Forecasting

Abstract

The increasing adoption of Electric Vehicles (EVs) offers a cleaner alternative to traditional means of transportation but creates significant challenges for power grid management. As a consequence, accurate EV charging load forecasting is essential. We propose using Meta-Learning, specifically First-Order Model-Agnostic Meta-Learning (FOMAML), to improve generalizability in forecasting. While some regions have ample data, others lack sufficient data, requiring models that can converge in a few iterations. Forecasting models that can generalize well can help city planners and businesses with infrastructure rollout and power purchase planning. We compared the accuracy of FOMAML models with traditional RNN, LSTM, and Encoder models, finding that FOMAML provides better generalization across different datasets.