Short-Term EV Load Forecasting Using Kolmogorov Arnold Networks

As global electric vehicle (EV) adoption accelerates, accurate EV load forecasting has become essential for optimizing grid management and integrating renewable energy sources effectively. Despite its importance, many studies fail to address the significant differences between private and public EV charging scenarios, which vary in user behavior, charging patterns, and infrastructure requirements. To bridge this gap, this paper introduces a novel short-term EV load forecasting technique leveraging Kolmogorov-Arnold Networks (KAN). By replacing traditional linear weights with spline-parametrized univariate functions, the proposed model dynamically learns and adapts to diverse activation patterns, offering enhanced forecasting accuracy. Our model was evaluated using datasets from both public and private EV charging stations, with Sample Convolution and Interaction Network (SCINet) and Mogrifier Long ShortTerm Memory serving as benchmark comparisons. The results highlight the proposed model's ability to surpass all baseline methods, showcasing exceptional accuracy and scalability across a variety of EV load forecasting scenarios.