Presentation Information
[17p-A33-9]Physics-Guided Clustered Echo State Network for Prediction of Large spatiotemporally chaotic Dynamics
〇Kuei-Jan Chu1, Nozomi Akashi1, Akihiro Yamamoto1 (1.Kyoto Univ.)
Keywords:
Reservoir Computing,Physics-Informed Machine Learning,Large spatiotemporally chaotic Dynamics
Chaos is aperiodic long-term behavior in a deterministic system. Predicting chaotic dynamics, such as weather forecasting, is essential yet challenging. Data-driven machine learning techniques have recently shown promise in this area. Echo State Networks (ESNs), a type of recurrent neural network, have achieved model-free attractor reconstruction of chaotic systems. However, predicting complex dynamics requires vast training data, leading to inefficiency. To address this, we propose a model called physics-guided clustered echo state network (PGClustered ESN), which leverages spatial structures in the target dynamical systems.
This model uses multiple clusters in a large reservoir, each corresponding to a variable in the target system, connected only to clusters with coupled variables. We demonstrate the efficacy of our model on two chaotic systems, the coupled map lattice (CML) and Lorenz 96. PGClustered ESNs outperform standard ESNs in both short-term prediction and attractor reconstruction, highlighting the value of incorporating coupling structures in spatiotemporal dynamics prediction.
This model uses multiple clusters in a large reservoir, each corresponding to a variable in the target system, connected only to clusters with coupled variables. We demonstrate the efficacy of our model on two chaotic systems, the coupled map lattice (CML) and Lorenz 96. PGClustered ESNs outperform standard ESNs in both short-term prediction and attractor reconstruction, highlighting the value of incorporating coupling structures in spatiotemporal dynamics prediction.
Comment
To browse or post comments, you must log in.Log in