Nonstationary Sequence Learning with Probabilistic Deep Models for Phase Segmentation and Multi-Scale Load Forecasting

Nonstationary time series often exhibit latent phase structures and abrupt transitions, posing significant challenges for modeling and analysis. Traditional approaches typically focus either on coarse-grained phase segmentation or on localized change-point detection, but seldom address both in a unified manner. This separation limits the generality and interpretability of existing methods. To overcome these limitations, we investigate a unified framework that integrates wavelet-based multiscale time-frequency representation, latent phase discovery, and entropy-driven stability-based change-point detection, aiming to capture both long-range behavioral regimes and fine-grained temporal discontinuities. The central problem addressed is how to jointly model dynamic patterns and abrupt variations in nonstationary sequences in a computationally efficient and interpretable way. Extensive experiments show that the proposed method improves F1-score by 6.5% on average and reduces false positives by 12% compared with state-of-the-art baselines, while maintaining comparable inference speed.