Kolmogorov-Arnold Networks for Supervised Learning Tasks in Smart Grids

Multi-layer Perceptron (MLP) is a foundational and powerful model for supervised learning tasks, particularly in forecasting and classification tasks in power systems. However, MLP models face challenges such as the requirement of large training datasets, extensive parameters, and high computational costs. Today, with the popularity of renewable energy on both the supply and demand sides, the power systems become more complex and the challenges will become bigger for MLP models. Recently, a new machine learning model based on the Kolmogorov-Arnold (KAN) representation theorem has emerged as a promising alternative to classical MLP models to address these issues. By utilizing learnable activation functions B-spline on neurons, the KAN model can quickly detect and learn hidden features, reducing the need for large datasets and computational resources while achieving higher accuracy. In this paper, we aim to demonstrate the performance of the KAN networks compared to the MLP networks using real-world data on Lithium-ion battery capacity and electric vehicle (EV) energy consumption. The numerical results demonstrate that the KAN model is a potentially powerful tool for data analysis in the power sector.