A Training/Inference Strategy Based on Restart: Improvement of Performance of Koopman Autoencoder for Future Estimation

Digital twin has gained attentions in the literature because it enables us to either estimate future states of a system or prevent its fault failures. To successfully apply digital twin into a system, a model of the system has to be accurately identified. Koopman Operator Theorey proved in 1931 sheds light on the system identification because it makes us interpret a nonlinear dynamical system as a linear system. Nowdays, deep learning technoloiges with the theory have been used for the system identification. We herein aim to review a model named as Koopman Autoencoder (KAE), which is considered to be baseline in the literature, and propose a new training/inference strategy to improve the performance of KAE for future estimation. To demonstrate advantages of our strategy, we develop two KAE models with/without our strategy, using synthetic datasets generated with the IEEE 3-Bus system. We show that KAE with the strategy can achieve a better performance with respect to mean squared error and relative root mean squared error, compared to KAE without it.