What is the role of normalization in time series deep learning architectures?

Normalization stabilizes training by scaling inputs and activations, improving convergence and performance.

Time series data often has varying scales across features and time. Normalization methods like z-score scaling, batch normalization, and layer normalization reduce internal covariate shift and stabilize gradient flow. In multivariate forecasting, normalization per series or global normalization strategies significantly affect model performance.