Kolmogorov-Arnold networks for algorithm design in battery energy storage system applications
Rolando Antonio Gilbert Zequera, Anton Rassõlkin, Toomas Vaimann, Ants Kallaste
Abstract
Energy technologies and Artificial Intelligence (AI) are essential for the energy transition to a carbon-free future through decarbonization, digitalization, and decentralization. The Kolmogorov-Arnold Network (KAN) is a promising new type of Neural Network (NN) that can improve Deep Learning models and serve as an alternative to the Multilayer Perceptron (MLP) for complex tasks. This paper proposes using KANs to design algorithms for Battery Energy Storage System (BESS) applications, focusing on state estimation, Remaining Useful Lifetime (RUL), and charging management. A wide range of datasets is collected by performing extensive testing on battery cells to demonstrate the robustness of the algorithms, in addition to advanced techniques like cross-validation, Regularization, Bayesian optimization, and Fine-tuning to improve Model Performance Analysis. The resulting network architectures were designed using Keras and PyTorch APIs, stored in PyTorch state dictionaries and Hierarchical Data Format (HDF5) files, and tested on new battery datasets. The final KANs achieved over 96 % accuracy, outperforming Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and MLPs in algorithm design and BESS applications. • Kolmogorov-Arnold Networks for different Battery Energy Storage System applications. • Stochastic network architectures are generated by Fine-tuning and Bayesian optimization. • Algorithm design based on AI methods and energy frameworks. • Deep Learning methodology to improve the robustness, adaptability, and effectiveness of Neural Networks.