TCDM: A Temporal Correlation-Empowered Diffusion Model for Time Series Forecasting

Huibo Xu, Likang Wu, Xianquan Wang, Zhiding Liu, Qi Liu

Introduction

Diffusion models have emerged as a powerful class of generative models, particularly in the domain of image synthesis. However, their application to time series forecasting is complicated by the inherent nature of the diffusion process. Standard diffusion models typically add independent, identically distributed (i.i.d.) Gaussian noise to data until it reaches a state of pure chaos, subsequently learning to reverse this process. In time series, the most critical information is often the temporal correlation—the relationship between successive observations. The addition of i.i.d. noise acts as a "shredder" for these relationships, particularly affecting low-frequency trend components. This loss of structural priors makes it difficult for models to reconstruct the original temporal "rhythm" during denoising, leading to directional biases and poor long-term predictive performance.

Methodology

The proposed Temporal Correlation-Empowered Diffusion Model (TCDM) addresses these limitations through a "divide-and-conquer" approach, utilizing three core components to preserve temporal structure.

The overall architecture of TCDM, illustrating the trend-residual decomposition and the integration of the MTC and RIM modules.

Trend-Residual Decomposition

TCDM decomposes the input time series into a smooth, long-term trend and high-frequency residuals. A frequency-domain MLP, acting as the Base Model, captures the trend component. This allows the diffusion model to focus exclusively on the more complex residual patterns, which are often harder to model using traditional methods.

Maintaining Temporal Correlation (MTC) Module

The MTC module is the central innovation of TCDM. Instead of adding completely random noise at each diffusion step, TCDM introduces Temporally Correlated Noise. This noise is generated via an autoregressive process, ensuring that the noise added to a specific point $i$ is dependent on the noise added to point $i-1$. This mechanism forces the model to maintain local temporal correlations even in highly noised states, providing a structural constraint that guides the reverse denoising process.

The MTC module's forward and reverse processes, highlighting the autoregressive noise generation.

Redesigned Initial Module (RIM)

Standard diffusion models typically begin the reverse process from pure Gaussian noise. TCDM introduces the Redesigned Initial Module (RIM), which truncates this process to preserve more of the natural data structure. RIM starts the reverse process from a state that is not fully "shredded." To bridge the gap between pure noise and this structured starting point, a GAN-based generator is employed to perform the transition in a single step, improving both efficiency and accuracy.

Experimental Results

The TCDM framework was evaluated on nine standard multivariate time series benchmark datasets across various domains, including energy (ETTh1, ETTm1, Electricity), environment (Weather, Wind), and finance/transport (Exchange, Traffic).

Forecasting Accuracy

TCDM achieved state-of-the-art results, recording the lowest Mean Squared Error (MSE) on six out of the nine datasets. The model attained an average rank of 1.333, significantly outperforming the next best baseline, TimeDiff, which had an average rank of 3.111.

Temporal Pattern Fidelity

To measure the model's ability to capture temporal dependencies, the authors utilized the Normalized Partial Autocorrelation Difference Index (N-PADI). TCDM demonstrated a significant advantage; for instance, on the ETTh1 dataset, TCDM achieved an N-PADI of 0.559, whereas TimeDiff scored 60.057. This indicates that TCDM is far more effective at preserving the underlying temporal "rhythm" of the data.

Discussion and Conclusion

TCDM demonstrates that by engineering the noise process to be autoregressive, diffusion models can be made to respect the inherent sequential nature of time series data. The combination of frequency-domain trend modeling and structured noise generation allows TCDM to overcome the "identity crisis" faced by standard diffusion models in forecasting tasks.

While the model shows significant promise, future work will address potential limitations, such as the sensitivity of the moving-average decomposition to extremely strong periodicities and the training complexity introduced by the GAN-based RIM module. Nonetheless, TCDM represents a major step forward in applying generative diffusion models to complex time series forecasting.

Summary updated

Abstract

Although previous studies have applied diffusion models to time series forecasting, these efforts have struggled to preserve the intrinsic temporal correlations within the series, leading to suboptimal predictive outcomes. This failure primarily results from the introduction of independent, identically distributed (i.i.d.) noise. In the forward process, the addition of i.i.d. noise to the time series gradually diminishes these temporal correlations. The reverse process starts with i.i.d. noise and lacks priors related to temporal correlations, which can result in directional biases during sampling. From a frequency-domain perspective, noise disrupts the low-frequency-dominated structure of trend components, making it difficult for the model to learn long-term temporal dependencies. To address these limitations, we introduce a decomposition prediction framework to complement the novel Temporal Correlation-Empowered Diffusion Model. Overall, We decompose the time series into trend and residual components, predict them using a base model and a diffusion model, and then combine the results. Specifically, a frequency-domain MLP model was adopted as the base model due to its not distorting the original sequence, and better the capture of long-range temporal dependencies. The diffusion model incorporates two key modules to capture short- and mid-range temporal correlations: the Maintaining Temporal Correlation Module and the Redesigned Initial Module. Extensive experiments across multiple datasets demonstrate that the proposed method significantly outperforms related strong baselines.
BibTeX
@inproceedings{xu2025tcdm,
  author = {Huibo Xu and Likang Wu and Xianquan Wang and Zhiding Liu and Qi Liu},
  title = {{TCDM: A Temporal Correlation-Empowered Diffusion Model for Time Series Forecasting}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/749},
}
Figure