Hidden Data Recovery and Forecasting via Next-Generation Reservoir Computing With Multiscale Delay Selection

Reconstructing hidden dynamics and forecasting nonlinear time series remain central challenges in machine learning and nonlinear system modeling. Next-generation reservoir computing (NG-RC) provides an efficient framework for these tasks, yet its standard formulation with uniform delays can be suboptimal for signals with multiple characteristic time scales. We propose NG-RC with multiscale delay selection (NGRC-MDS), in which irregular delay vectors are constructed directly from the temporal structure of the signal. The method combines empirical mode decomposition (EMD) with kernel density estimation (KDE) to extract dominant temporal scales and form compact, interpretable delay sets. Experiments on the adaptive Kuramoto network, the Lorenz system, and the Mackey-Glass system show that the proposed approach consistently matches or exceeds optimized NG-RC accuracy while requiring minimal tuning. In particular, it improves hidden-state recovery in multiscale dynamical systems and remains effective for both short-term and long-term forecasting of delayed chaotic dynamics. These results show that multiscale delay selection provides a practical extension of NG-RC for nonlinear systems with complex temporal organization.