NAR Broad Learning System for dynamical systems prediction.

Area of Science:

• Dynamical Systems and Chaos Theory
• Machine Learning and Artificial Intelligence
• Complex Systems Analysis

Background:

• Dynamical systems exhibit complex spatiotemporal relationships, making accurate prediction challenging.
• Existing data-driven models struggle with simultaneous temporal and spatial feature extraction and fast training.
• Efficiently analyzing and predicting dynamical systems remains a significant bottleneck in scientific research.

Purpose of the Study:

• To propose a novel shallow network, the Nonlinear Autoregression Broad Learning System (NAR-BLS), for enhanced dynamical system prediction.
• To address the limitations of current data-driven methods in terms of speed and feature extraction capabilities.
• To develop a model capable of simultaneously capturing temporal dynamics and spatial features for improved prediction accuracy.

Main Methods:

• Developed NAR-BLS, a shallow randomized flatten network integrating a temporal feature capture branch.
• Employed a separated-aggregated mapping of feature and enhancement nodes for spatial feature extraction.
• Utilized ridge regression solely for output layer weight computation, ensuring rapid training and updating.

Main Results:

• NAR-BLS demonstrated superior performance in predicting two chaotic systems and four real-world datasets.
• The model successfully extracted both temporal dynamic and spatial features concurrently.
• Achieved rapid training and updating speeds due to the simplified weight computation.

Conclusions:

• NAR-BLS offers a highly effective and efficient solution for dynamical system prediction.
• The proposed architecture overcomes key limitations of existing data-driven approaches.
• NAR-BLS shows significant potential for applications in various scientific domains requiring dynamical system analysis.