Related Experiment Videos
Dual-orthogonal radial basis function networks for nonlinear time series prediction
1Department of Automatic Control and Systems Engineering, University of Sheffield, Sheffield, UK
Summary
A novel Dual-orthogonal Radial Basis Function (RBF) Network (DRBF) improves nonlinear time series prediction by modifying distance metrics. This new approach enhances accuracy for complex, correlated data, outperforming conventional methods.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Computational Neuroscience
Background:
- Conventional Radial Basis Function (RBF) networks use Euclidean distance, which is suboptimal for highly correlated time series data.
- Lagged system outputs in time series prediction often exhibit strong correlations, challenging standard distance-based RBF network approaches.
Purpose of the Study:
- Introduce a new Dual-orthogonal RBF Network (DRBF) structure for improved nonlinear time series prediction.
- Address the limitations of Euclidean distance metrics in RBF networks when dealing with correlated time series inputs.
Main Methods:
- The DRBF network modifies the distance metric using a classification function based on estimation data.
- Training involves a two-stage process: learning classification basis functions and input nodes, followed by regressor selection and weight learning.
- A forward Orthogonal Least Squares (OLS) selection procedure is employed for selecting important input nodes and centers.
Main Results:
- Simulation results demonstrate the effectiveness of the DRBF network for both single-step and multi-step ahead predictions.
- The DRBF approach shows improved performance on a test dataset compared to conventional methods.
Conclusions:
- The DRBF network offers a more appropriate and effective approach for nonlinear time series prediction, especially with correlated data.
- The modified distance metric and two-stage training process contribute to enhanced predictive accuracy.