Skip to contentScholarGate
LibraryBookshelfDeskReview StudioAssistant
Sign in
On this page
IntuitionHow it worksWhen to use itStrengths & limitationsCommon pitfallsApplicationsFrequently asked🔒 Read the full methodSourcesRelated methods
Cite this pageSpotted an issue on this page? Report or suggest a fix →
Home›Deep learning›MICN: Multi-scale Isometric Convolution Network for Long-term Time-series Forecasting
Machine learningTime-series forecasting

MICN: Multi-scale Isometric Convolution Network for Long-term Time-series Forecasting

MICN (Multi-scale Isometric Convolution Network) · Also known as: Multi-scale Isometric Convolution Network, Multi-scale Local and Global Context Model, MICN Forecaster, Çok Ölçekli İzometrik Evrişim Ağı

MICN (Multi-scale Isometric Convolution Network) is a convolutional neural network architecture for long-term time-series forecasting introduced by Huiqiang Wang and colleagues at ICLR 2023. Its central idea is to capture both local temporal patterns and global seasonal dependencies simultaneously through multi-scale isometric convolutions combined with a merge attention mechanism, enabling efficient and expressive modeling of complex temporal dynamics without the quadratic cost of full self-attention.

ScholarGate
  1. Machine learning
  2. v1
  3. 1 Sources
  4. PUBLISHED
Cite this page →
Tools & resources
Download slides
Learn & explore

Read the full method

Members only

Sign in with a free account to read this section.

Sign in

Method map

The neighbourhood of related methods — select a node to explore.

MICN
SCINetTimesNet

When to use it

MICN is well suited for long-horizon multivariate or univariate time-series forecasting tasks where capturing multi-scale temporal dependencies matters — for example, energy consumption, weather, or financial data with overlapping seasonalities. It assumes the series contains meaningful structure at multiple resolutions. It is less appropriate when the series is very short, when stationarity cannot be approximated, or when interpretability of individual component contributions is paramount. Alternatives include PatchTST, TimesNet, and SCINet for similar settings.

Strengths & limitations

Strengths
  • Efficiently captures both local and global temporal dependencies through multi-scale isometric convolution without full self-attention's quadratic cost
  • Isometric kernel design preserves scale-appropriate receptive fields, reducing information distortion across resolutions
  • Direct multi-step forecasting head avoids autoregressive error accumulation in long-horizon predictions
  • Competitive performance on standard long-term forecasting benchmarks with relatively lightweight architecture
Limitations
  • Performance can degrade on series with highly irregular or non-stationary patterns that do not exhibit consistent multi-scale structure
  • The number of scales K and downsampling rates are hyperparameters that require tuning and domain knowledge
  • As a purely convolutional model it may underperform Transformer-based methods on datasets where long-range sparse dependencies dominate
  • Limited interpretability: multi-scale representations are merged into a black-box latent space with no explicit decomposition output

Frequently asked

What makes MICN 'isometric' compared to standard multi-scale CNNs?

In MICN, the convolutional kernel size at each scale is set equal to the downsampling stride applied to produce that scale's sub-sequence. This isometric constraint ensures that the receptive field of the convolution precisely matches the spacing between retained samples, so temporal structure is neither over-smoothed nor aliased — each scale sees exactly the granularity appropriate to its resolution.

Does MICN require decomposition preprocessing (e.g., trend-seasonal splitting) before training?

No explicit decomposition preprocessing is required. MICN implicitly captures trend and seasonal components across different scales through its multi-scale architecture. However, standard input normalization (such as instance normalization or reversible instance normalization) is typically recommended to improve training stability and cross-domain generalization.

How should the number of scales K and downsampling rates be selected in practice?

Common practice is to align downsampling rates with the dominant periods in the dataset — for example, rates of 1, 2, and 4 for hourly data with daily and weekly seasonality. K is typically kept small (2–4 scales) to balance expressiveness and computational cost. Validation set performance and frequency-domain analysis of the target series are the primary guides for this choice.

Sources

  1. Wang, H., Peng, J., Huang, F., Wang, J., Chen, J., & Xiao, Y. (2023). MICN: Multi-scale local and global context modeling for long-term series forecasting. ICLR. link ↗

How to cite this page

ScholarGate. (2026, June 2). MICN (Multi-scale Isometric Convolution Network). ScholarGate. https://scholargate.app/en/deep-learning/micn

Related methods

SCINetTimesNet

Which method?

Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.

  • SCINetDeep learning↔ compare
  • TimesNetDeep learning↔ compare
Compare side by side →

Similar methods

SCINetTimeMixerTimesNetCrossformerPyraformerPatchTSTTSMixerLightTS

Related reference concepts

Convolutional and Sequence ModelsDeep Generative ModelsDeep LearningNeural Network ArchitecturesSelf-Supervised and Representation LearningSequence-to-Sequence Models and Transformers

Spotted an issue on this page? Report or suggest a fix →

ScholarGate — MICN (MICN (Multi-scale Isometric Convolution Network)). Retrieved 2026-07-21 from https://scholargate.app/en/deep-learning/micn · Dataset: https://doi.org/10.5281/zenodo.20539026
Quick facts
Originator
Huiqiang Wang et al.
Year
2023
Type
CNN-based time-series forecasting architecture
Subfamily
Time-series forecasting
Venue
ICLR 2023
Input
Multivariate or univariate time series
Related methods
SCINetTimesNet
ScholarGate

A content-first reference library for research methods — what each one is, how it works, and where it comes from.

Open data (CC-BY)

Explore

  • Library
  • Search the library…
  • Browse by field
  • Fields
  • Journey
  • Compare
  • Which method?

Reference

  • Subjects
  • Atlas
  • Glossary
  • Methodology
  • Philosophy

Your tools

  • Bookshelf
  • Desk
  • Chat

Company

  • About
  • Pricing
  • Contact
  • Suggest a method

Entries are compiled from published sources for reference. Verifying the accuracy and suitability of any information for your own use remains your responsibility.

© 2026 ScholarGate · A research-method reference library
  • Privacy
  • Cookies
  • Terms
  • Delete account