> ML_ALGORITHM // PATCHTST-CHANNEL-INDEPENDENT-PATCHING_v1.0
PatchTST (Channel-Independent Patch Time Series Transformer)
Highly effective time series Transformer that segments continuous series into subseries patches and applies channel-independent self-attention.
Transformer-Based Forecastingtime-series-forecastingmoderate-posthoclarge (>100k)
Back to All AlgorithmsComputational Complexity
Training Complexity:O(epochs * batch_size * channels * (L/P)^2 * d)
Inference Complexity:O(channels * (L/P)^2 * d)
Hardware Profile
CPU Friendly:No
Requires GPU:Yes
Memory Footprint:moderate
Interpretability & Data
Interpretability Tier:moderate-posthoc
Training Data Needs:large (>100k)
Interpretability Assessment
Channel independence prevents spurious cross-variate correlations while patch attention preserves temporal memory.
Suitable Tasks & Supported Modalities
Suitable Tasks:
long horizon forecastingmultivariate time series forecasting
Supported Modalities:
time-series
Implementing Libraries
Foundational Literature
Common Pitfalls & Warnings
- Selecting a patch size P incompatible with underlying periodicity distorts Fourier spectral characteristics