HTV-Trans addresses the non-stationarity issue in multivariate time series forecasting by combining a hierarchical probabilistic generative module with a transformer. The model considers the inherent non-stationarity and stochasticity characteristics within MTS, providing expressive representations for forecasting tasks. By recovering intrinsic non-stationary information into temporal dependencies, HTV-Trans shows efficiency in diverse datasets. Previous methods primarily adopt stationarization techniques to handle non-stationarity, but HTV-Trans introduces a powerful probabilistic generative module to address this challenge. The hierarchical structure of HTV-Trans allows for multi-scale representation of original time series data, enhancing predictive capabilities. The model's architecture includes a transformer encoder to capture dynamic information and an MLP decoder for forecasting tasks.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Muyao Wang,W... at arxiv.org 03-11-2024
https://arxiv.org/pdf/2403.05406.pdfDeeper Inquiries