A Comparison Between Kolmogorov-Arnold and Multilayer Perceptron Networks for State of Health Estimation of Lithium-Ion Batteries

Among the several methods proposed for state of health estimation are those based on machine learning. Within this field, there are the so-called artificial neural networks which are successfully used for modeling complex problems with nonlinearities. The recently proposed Kolmogorov-Arnold network (KAN) is a promising alternative to the long-established multilayer perceptron network (MLP); therefore, a fair comparison between the two approaches is presented here. The training time with KAN was 7.8 times faster than with MLP for the battery B005 of the NASA dataset and 2.5 times faster for battery B006, while the mean regression time for the battery B005 of the with a symbolic formula are about three orders of magnitude faster in comparison with a MLP network, and 45 times faster for the battery B006. The overall training times and regression times results for the battery B007 are similar to those of B005. The mean MAEs for the batteries B005, B006, and B007 with symbolic regression are 0.972%, 1.111%, and 0.948%, respectively. The most interesting aspect of this paper is the use of the underexplored symbolification of the KAN, which enables faster regression compared with MLPs.