WHiAR-Net: an interpretable multi-scale forecasting framework via Wavelet-Hilbert feature engineering

Accurate forecasting of non-stationary time series remains challenging due to complex volatility. We introduce WHiAR-Net, an interpretable framework integrating wavelet theory with Hilbert spectral analysis to separate long-term trends from transient fluctuations. By structurally embedding operator error bounds, this transparent architecture offers a principled alternative to conventional black-box models. Systematic experiments on electricity and environmental datasets demonstrate that our method achieves highly competitive accuracy, outperforming modern deep learning baselines, with promising future applications in smart energy grids and climate monitoring.