ScaleNet--multiscale neural-network architecture for time series prediction

Area of Science:

• Computational Science
• Applied Mathematics
• Artificial Intelligence

Background:

• Nonlinear dynamic systems present significant challenges for accurate time series prediction.
• Traditional methods often struggle to capture complex temporal patterns and trends effectively.

Purpose of the Study:

• To investigate the effectiveness of a multiscale neural-network (NN) architecture for time series prediction.
• To improve prediction accuracy for nonlinear dynamic systems by leveraging multiscale analysis.

Main Methods:

• Decomposition of time series data into different scales using wavelets.
• Prediction of wavelet coefficients at each scale using separate multilayer perceptron NNs.
• Integration of predictions from all scales using an expert NN.
• Training networks with backpropagation and Levenberg-Marquardt, initialized using a novel clustering algorithm.

Main Results:

• The multiscale NN architecture demonstrated superior prediction accuracy compared to single-scale architectures.
• The method effectively analyzes both short-term details (high frequencies) and long-term trends (low frequencies).
• Clustering-based initialization improved prediction results over random initialization.

Conclusions:

• The proposed multiscale NN architecture is highly effective for time series prediction of nonlinear dynamic systems.
• Coordinated analysis across multiple time and frequency scales provides richer information for prediction.
• Further improvements in prediction accuracy are possible with enhanced learning methods.