Research map / Time series / Long-term forecasting

Channel Normalization for Time Series Channel Identification

ICML 2025 · Time series › Long-term forecasting · approach: Spatiotemporal forecasting models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Channel Normalization (CN) and its adaptive (ACN) and prototypical (PCN) variants, which assign channel-specific affine normalization parameters to enhance channel identifiability in time series models.

Problem

Time series models often cannot distinguish between channels (non-channel-identifiable), producing identical outputs for identical inputs regardless of channel identity, and existing channel-identifier methods give limited gains or require architecture changes; per-channel parameters also do not sca

Approach

Replace the shared affine parameters of layer normalization with channel-specific affine parameters (CN); extend with input-dependent local parameters attended via channel similarity (ACN) and with learnable prototype parameters attended via channel-prototype similarity so the method works with any number of channels (PCN).

Key results

• ACN on iTransformer reduces average MSE across 12 datasets and 4 horizons from 0.275 to 0.241 (12.4% improvement), versus 11.3% for CN.
• On RMLP, ACN improves average MSE by 11.8% (ACN additional gain of 6.1 points over CN), with 46 of 48 best counts.
• PCN with K=5 improves iTransformer by 8.4% on average across datasets, smaller than CN and ACN gains.

Measured effects reported in the paper

Surprising findings

• Simply adding distinct constant vectors to each channel token of iTransformer improves MSE from 0.275 to 0.252, showing channel distinction alone helps.
• Channel-specific parameters in C-token and C-project baselines degrade performance of CID models.

Limitations

• Gains on CID models such as S-Mamba and TSMixer are smaller than on non-CID models.
• PCN gives a weaker form of channel identifiability and its gains are smaller than CN/ACN on single-task models.
• Number of prototypes K must be set manually; automatic selection is left for future work.

Experimental setup

Datasets: ETTh1 ETTh2 ETTm1 ETTm2 PEMS03 PEMS04 PEMS07 PEMS08 Exchange Weather ECL Solar-Energy

Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 PEMS03 PEMS04 PEMS07 PEMS08 Exchange Weather ECL Solar-Energy

Metrics: MSE MAE

Baselines: iTransformer RMLP S-Mamba TSMixer channel identifier C-LoRA Instance Normalization C-token C-project constant vector

Builds on: Layer Normalization iTransformer UniTS C-LoRA RevIN

Closest papers by meaning

• Channel Matters: Estimating Channel Influence for Multivariate Time Series (NeurIPS 2025 · similarity 0.84)
• TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · similarity 0.84)
• From Similarity to Superiority: Channel Clustering for Time Series Forecasting (NeurIPS 2024 · similarity 0.83)
• SIN: Selective and Interpretable Normalization for Long-Term Time Series Forecasting (ICML 2024 · similarity 0.82)
• CATCH: Channel-Aware Multivariate Time Series Anomaly Detection via Frequency Patching (ICLR 2025 · similarity 0.82)
• CARD: Channel Aligned Robust Blend Transformer for Time Series Forecasting (ICLR 2024 · similarity 0.82)