(AAAI 2026) Towards Non-Stationary Time Series Forecasting with Temporal Stabilization and Frequency Differencing

This code is a PyTorch implementation of our ICLR'24 paper "Towards Non-Stationary Time Series Forecasting with Temporal Stabilization and Frequency Differencing". [arXiv]

Introduction

DTAF, a dual-branch framework that addresses non-stationarity in both the temporal and frequency domains. For the temporal domain, the Temporal Stabilizing Fusion (TFS) module employs a non-stationary mix of experts (MOE) filter to disentangle and suppress temporal non-stationary patterns while preserving long-term dependencies. For the frequency domain, the Frequency Wave Modeling (FWM) module applies frequency differencing to dynamically highlight components with significant spectral shifts. By fusing the complementary outputs of TFS and FWM, DTAF generates robust forecasts that adapt to both temporal and frequency domain non-stationarity.

Quickstart

1. Requirements

Install the dependencies with the following command:

pip install -r requirements.txt

2. Data preparation

You can obtain the well-preprocessed datasets from Google Drive or Baidu Drive. Then, place the downloaded data under the folder ./dataset.

3. Train and evaluate the model

• To see the model structure of DTAF, click here.
• We provide all the experiment scripts for DTAF and other baselines under the folder ./scripts/DTAF.

Results

We utilize the Time Series Forecasting Benchmark (TFB) code repository as a unified evaluation framework, providing access to all baseline codes, scripts, and results. Following the settings in TFB, we do not apply the "Drop Last" trick to ensure a fair comparison.

Results of comprehensive parameter searches

Results from comprehensive parameter searches for the long-term forecasting task. The look-back window underwent testing with lengths 36 and 104 for NN5, ILI, Covid-19, and Wike2000, and 96, 336, and 512 for all other datasets. We search for the best results from these look-back windows and report the best results.

Extensive experiments on 11 real-world datasets from 6 different application domains demonstrate that DTAF achieves state-of-the-art(SOTA) performance. We show the full main results of the 11 below:

Citing DTAF

If you find this resource helpful, please consider to cite our research:

@inproceedings{
    lu2025dtaf,
    title={Towards Non-Stationary Time Series Forecasting with Temporal Stabilization and Frequency Differencing},
    author={Junkai Lu and Peng Chen and Chenjuan Guo and Yang Shu and Meng Wang and Bin Yang},
    booktitle={AAAI},
    year={2026}
}