MultiScaleWave: a wavelet-based multiscale framework for univariate time series forecasting.

Area of Science:

• Data Science
• Machine Learning
• Signal Processing

Background:

• Accurate time series forecasting is crucial across various domains.
• Real-world data often presents challenges like noise, non-stationarity, and multiscale dependencies, hindering prediction accuracy.
• Existing methods struggle to effectively model these complex temporal characteristics.

Purpose of the Study:

• To propose a novel deep learning framework, MultiScaleWave, for univariate time series forecasting.
• To address the limitations of existing methods in handling noisy and non-stationary time series data.
• To improve forecasting performance by leveraging time series decomposition.

Main Methods:

• MultiScaleWave employs a deep learning approach based on time series decomposition.
• It utilizes multi-level discrete wavelet transforms to decompose the time series into multiscale temporal components.
• Each component is processed by a granularity-adaptive module, with outputs fused for final forecasting.

Main Results:

• The MultiScaleWave model demonstrated superior performance on benchmark datasets.
• It outperformed competitive baseline models in forecasting accuracy.
• Validation confirmed the model's effectiveness and generalizability across different datasets.

Conclusions:

• MultiScaleWave offers an effective solution for univariate time series forecasting.
• The framework successfully addresses challenges posed by noise and non-stationarity.
• The results highlight the potential of wavelet-based decomposition in deep learning for time series analysis.