DisenTS (TPAMI)

---

This repo is the official PyTorch implementation of our TPAMI paper: DisenTS: Disentangled Channel Evolving Pattern Modeling for Multivariate Time Series Forecasting.

Introduction

Multivariate time series forecasting plays a crucial role in many real-world applications. Existing forecasting models usually adopt either a unified channel-mixing paradigm or a channel-independent paradigm. However, real-world multivariate series often contain heterogeneous channels with distinct evolving patterns, including different seasonalities, trend shifts, and local dynamics. A single shared model may therefore struggle to capture all channel-wise patterns properly.

To alleviate this issue, we propose DisenTS, a model-agnostic framework for disentangled channel evolving pattern modeling. DisenTS utilizes multiple forecasting models to uncover diverse latent evolving patterns and adaptively aggregates their predictions for each input channel. In this way, DisenTS preserves the flexibility of multi-pattern modeling while avoiding the prohibitive cost of training an independent model for every channel.

Framework Overview

• Disentangled Forecasting Model Pool: DisenTS employs multiple forecasting models as experts, where each forecaster is encouraged to specialize in a distinct evolving pattern.
• Forecaster Aware Gate (FAG): The gate adaptively generates routing signals according to both the input series characteristics and the current states of the forecasting models.
• Linear Weight Approximation (LWA): LWA quantizes complex forecasting models into compact matrices, enabling the gate to be aware of each forecaster's functional behavior.
• Similarity Constraint (SC): SC regularizes the approximated forecaster states to promote specialization and disentanglement among different forecasting experts.
• Weighted Prediction Aggregation: The final forecast is produced by a weighted sum of different expert predictions, allowing each channel to select suitable evolving patterns dynamically.

Experiments

We conduct extensive experiments on 14 real-world datasets, covering both long-term and short-term multivariate forecasting settings. DisenTS is evaluated with mainstream channel-independent backbones such as DLinear, SparseTSF, PatchTST, and ConvTimeNet, and is further studied with channel-dependent methods.

The results show that DisenTS consistently improves forecasting accuracy across diverse data scenarios. In particular, it brings clear gains on large-scale heterogeneous multivariate datasets such as PEMS, validating the effectiveness of disentangled channel evolving pattern modeling.

Usage

Running

We provide ready-to-use scripts for DisenTS-enhanced backbone models.

sh scripts/run_linear.sh     # DisenTS + DLinear
sh scripts/run_sparsetsf.sh  # DisenTS + SparseTSF
sh scripts/run_tst.sh        # DisenTS + PatchTST
sh scripts/run_conv.sh       # DisenTS + ConvTimeNet

You can also run a single experiment with run_longExp.py. For example:

python -u run_longExp.py \
  --is_training 1 \
  --root_path ./dataset/exchange_rate \
  --data_path exchange_rate.csv \
  --model_id exchange_96_96 \
  --model Disen \
  --disen_backbone DLinear \
  --data custom \
  --features M \
  --seq_len 336 \
  --label_len 168 \
  --pred_len 96 \
  --enc_in 8 \
  --dec_in 8 \
  --c_out 8 \
  --learning_rate 0.01 \
  --similar_weight 0.01 \
  --disen_times 4 \
  --sample_num 128 \
  --d_gate 64 \
  --itr 3

Tuning

For other forecasting tasks and backbones, the main DisenTS-related hyper-parameters are:

• disen_backbone: the forecasting backbone used in the expert pool, e.g., DLinear, SparseTSF, PatchTST, or ConvTimeNet.
• disen_times: the number of forecasting experts in DisenTS.
• similar_weight: the weight of the Similarity Constraint.
• sample_num: the number of samples used for estimating the approximated forecaster state.
• d_gate: the hidden dimension of the Forecaster Aware Gate.

Acknowledgement

This repo is built on the pioneer works (Time-Series-Library). The environment requirements and datasets can be found in their original repo. Many thanks to their efforts and devotion!

Citation

If you find this repository useful for your work, please consider citing it as follows:

@article{liu2026disents,
  title={DisenTS: Disentangled Channel Evolving Pattern Modeling for Multivariate Time Series Forecasting},
  author={Liu, Zhiding and Yang, Jiqian and Mao, Qingyang and Zhao, Yuze and Cheng, Mingyue and Li, Zhi and Liu, Qi and Chen, Enhong},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2026}
}