Welcome to the official repository of the TQNet paper: "Temporal Query Network for Efficient Multivariate Time Series Forecasting".
🚩 News (2025.05): TQNet has been accepted to ICML 2025 and the paper and code is currently available.
Please note that SparseTSF, CycleNet, and TQNet represent our continued exploration of leveraging periodicity for long-term time series forecasting (LTSF). The differences and connections among them are as follows:
| Model | Use of Periodicity | Technique | Effect | Efficiency | Strengths | Limitation |
|---|---|---|---|---|---|---|
| SparseTSF (ICML 2024 Oral) |
Indirectly via downsampling | Cross-Period Sparse Forecasting | Ultra-light design | < 1k parameters | Extremely lightweight, near SOTA | Fails to cover multi-periods (solved by CycleNet) |
| CycleNet (NeurIPS 2024 Spotlight) |
Explicit via learnable parameters | Residual Cycle Forecasting (RCF) | Better use of periodicity | 100k ~ 1M parameters | Strong performance on periodic data | Fails in multivariate modeling (solved by TQNet) |
| TQNet (ICML 2025) |
Serve as global correlations | Temporal Query in attention mechanism | Robust inter-variable correlation modeling | ~1M parameters | Enhanced multivariate forecasting performance | Hard to scale to ultra-long look-back inputs |
TQNet is the powerful successor to CycleNet (our previous work, NeurIPS 2024 Spotlight). While CycleNet introduces the use of learnable vectors to model periodic patterns within individual variables, it fails to effectively capture inter-variable dependencies.
To bridge this gap, TQNet leverages the same periodically shifted learnable vectors from CycleNet as queries in an attention mechanism, allowing the model to capture global inter-variable correlations. Meanwhile, keys and values are derived from the raw input data to encode local sample-level information, thereby fusing global priors with local observations.
TQNet adopts a single-layer attention mechanism and a lightweight MLP, resulting in a compact and efficient model design. Extensive experiments demonstrate that TQNet learns more robust multivariate dependencies, achieving state-of-the-art forecasting accuracy across 12 challenging real-world datasets.
An ablation study on different Q-K configurations reveals that the TQNet strategy, which considers both global and per-sample correlations, consistently yields the best performance. In contrast, variants that rely solely on either global or per-sample information lead to suboptimal results.
Further ablations highlight the core contribution of the proposed TQ technique. In addition, integration studies confirm the portability and effectiveness of TQ in enhancing the forecasting capability of other existing models.
The figure below visualizes the t-SNE projections of the learned TQ vectors after training. Strikingly, channels that are closer in the t-SNE space tend to share similar temporal patterns, while those farther apart exhibit distinct dynamics. This indicates that the TQ technique effectively captures intrinsic multivariate correlations.
To further validate its effectiveness, we evaluate TQNet on a multivariate-to-univariate forecasting task—predicting one target variable using a varying number of input channels. As shown in the figure below, incorporating even a moderate amount of covariate information significantly improves TQNet’s performance, highlighting its strength in modeling robust cross-variable dependencies.
Finally, thanks to its powerful TQ technique and lightweight architecture, TQNet delivers exceptional forecasting accuracy with minimal computational overhead. Even when scaling to datasets with more variables, TQNet remains highly efficient, closely matching the speed of linear-based models like DLinear.
To get started, ensure you have Conda installed on your system and follow these steps to set up the environment:
conda create -n TQNet python=3.8
conda activate TQNet
pip install -r requirements.txt
All the datasets needed for TQNet can be obtained from the [Google Drive] that introduced in previous works such as Autoformer and SCINet.
Create a separate folder named ./dataset and place all the CSV files in this directory.
Note: Place the CSV files directly into this directory, such as "./dataset/ETTh1.csv"
The training scripts for replicating the TQNet results are available at:
./scripts/TQNet
You can reproduce all the main results of TQNet with the following code snippet.
conda create -n TQNet python=3.8
conda activate TQNet
pip install -r requirements.txt
sh run_main.sh
For your convenience, we have provided the execution results of "sh run_main.sh":
./result.txt
You can also reproduce all the ablation results of TQNet and the other variants with the following script.
sh run_ablation.sh
If you find this repo useful, please cite our paper.
We extend our heartfelt appreciation to the following GitHub repositories for providing valuable code bases and datasets:
https://github.com/ACAT-SCUT/CycleNet
https://github.com/lss-1138/SparseTSF
https://github.com/thuml/iTransformer
https://github.com/lss-1138/SegRNN
https://github.com/yuqinie98/patchtst
https://github.com/cure-lab/LTSF-Linear
https://github.com/zhouhaoyi/Informer2020
https://github.com/thuml/Autoformer
https://github.com/MAZiqing/FEDformer
https://github.com/alipay/Pyraformer