AdaptiveFrequencyNet: a long-term time series forecasting approach based on learnable frequency decomposition and cross-scale attention fusion

Description

Cui, Li, and Wang present AdaptiveFrequencyNet, a forecasting model for long runs of time-stamped data such as electricity use or weather. Long-range forecasts are hard because patterns repeat at several scales at once. Older methods split the data into fixed frequency bands that cannot adapt. Their 2026 design learns its own split using adjustable filters, then blends information across scales with attention, the technique that lets a model weigh which parts of the data matter most.

Tested on six standard benchmarks against nine rival models, the new network posted competitive errors while keeping a moderate size. Results varied by dataset: it matched the leading transformer model on one electricity series, while rivals edged it on a weather benchmark. The authors are candid that rankings depend on the dataset and forecast horizon. Their claim is narrower than a clean sweep: learnable decomposition holds its own without rigid frequency splits.

Catalog record

Source: https://europepmc.org/articles/PMC13493831