Recurrent Broad Learning Systems for Time Series Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Complex Systems Modeling

Background:

• Broad Learning System (BLS) offers an efficient approach to modeling complex systems.
• BLS utilizes feature and enhancement nodes for nonlinear transformations.
• Existing BLS models can be extended for various applications.

Purpose of the Study:

• Introduce a novel Recurrent Broad Learning System (RBLS) for enhanced time series analysis.
• Incorporate recurrent connections in enhancement nodes to capture temporal dynamics.
• Utilize sparse autoencoders for improved feature extraction.

Main Methods:

• Developed RBLS by adding recurrent connections to BLS enhancement nodes.
• Employed a sparse autoencoder for initial feature extraction from input data.
• Implemented conjugate gradient methods for weight updates based on prediction errors (fine-tuning).

Main Results:

• RBLS demonstrated effective modeling of chaotic time series.
• The model achieved significantly small prediction errors on benchmark chaotic datasets.
• Satisfactory performance was observed on a real-world dataset.

Conclusions:

• RBLS is a powerful and efficient model for processing sequential and time series data.
• The proposed RBLS effectively captures dynamic characteristics crucial for time series forecasting.
• The model's architecture and learning mechanisms contribute to its high accuracy and efficiency.