Model adaptive phase space reconstruction

Abstract

Phase space reconstruction (PSR) methods allow for the analysis of low-dimensional data with methods from dynamical system theory, but their application to prediction models, like those from machine learning (ML), is limited. Therefore, we present here a model adaptive phase space reconstruction (MAPSR) method that unifies the process of PSR with the modeling of the dynamical system. MAPSR is a differentiable PSR method that enables the use of ML methods and is based on the idea of time delay embedding. For achieving differentiable, continuous, real-valued delays, which can be optimized using gradient descent, the discrete time signal is converted to a continuous time signal. The delay vector, which stores all potential embedding delays and the trainable parameters of the model, are simultaneously updated to achieve an optimal time delay embedding for the observed system. MAPSR does not rely on any threshold or statistical criterion for determining the dimension and the set of delay values for the embedding process. The quality of the reconstruction is evaluated by the prediction loss. We apply the proposed approach to uni- and multivariate time series stemming from regular and chaotic dynamical systems and a turbulent combustor to test the generalizability of the method and compare our results with established PSR methods. We find that for the Lorenz system, the model trained with the MAPSR method is able to predict chaotic time series for nearly 7 to 8 Lyapunov time scales, which is found to be much better compared to other PSR methods (AMI-FNN and PECUZAL methods). For the univariate time series from the turbulent combustor, the long-term prediction error of the model trained using the MAPSR method stays in between that of AMI-FNN and PECUZAL methods for the regime of chaos, and for the regime of intermittency, the MAPSR method outperforms the AMI-FNN and PECUZAL methods.