Koopman Theory Inspired Neural Network for State of Charge Estimation

Abstract

Koopman theory offers a potent framework for mode-by-mode analysis of system dynamics. Inspired by this theory, this work introduces a deep-learning framework utilizing autoencoders and a customized loss function for time series prediction. The work demonstrates the effectiveness of the model through practical application on state of charge (SoC) estimation of a battery. Model based SoC estimation techniques like Extended Kalman Filter(EKF) require complex models and technical data to estimate the SoC of a battery. The proposed model is able to predict the state of charge upto 10 time stamps in the future. This model is able to generalize the system dynamics such that a model trained for T time stamps is seen to give RMSE lesser than 0.01 for all tested temperatures at a future time stamp lesser than T. These findings demonstrate the superior performance of the proposed Koopman-inspired neural network(KoNN) compared to the traditional time series estimation technique EKF, Multivariate Linear Regression(MVLR), Extra Tree Regression(ETR), and Neural Network(NN) showcasing its versatility for various predictive tasks. © 2024 IEEE.