Online dynamic ensemble deep random vector functional link neural network for forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Traditional ensemble deep random vector functional link (edRVFL) models are not optimized for online learning.
• The inherent randomness in edRVFL features can hinder the extraction of meaningful temporal patterns.
• Existing methods lack robust mechanisms for adapting to evolving time series data distributions.

Purpose of the Study:

• To propose a novel three-stage online deep learning model for time series analysis.
• To extend the capabilities of the edRVFL model for dynamic and adaptive learning.
• To improve the extraction of temporal features and handle data distribution shifts in time series.

Main Methods:

• Development of a dynamic edRVFL model incorporating online decomposition, online training, and online dynamic ensemble components.
• Utilizing an online decomposition technique for feature engineering tailored to the edRVFL architecture.
• Designing an online learning algorithm for efficient edRVFL model training and an online dynamic ensemble for output aggregation.

Main Results:

• The proposed dynamic edRVFL model demonstrates effective performance in online time series analysis.
• The three-stage online approach successfully addresses limitations of the original edRVFL model.
• Comparative evaluations on sixteen time series datasets show competitive results against state-of-the-art methods.

Conclusions:

• The dynamic edRVFL model offers a robust and adaptive solution for online time series learning.
• The integration of online decomposition, training, and ensemble methods enhances temporal feature representation.
• This research contributes a significant advancement in applying deep learning models to dynamic time series data.