Abstract:Time series forecasting is vital in numerous web applications, influencing critical decision-making across industries. While diffusion models have recently gained increasing popularity for this task, we argue they suffer from a significant drawback: indiscriminate noise addition to the original time series followed by denoising, which can obscure underlying dynamic evolving trend and complicate forecasting. To address this limitation, we propose a novel flexible decoupled framework (FDF) that learns high-quality time series representations for enhanced forecasting performance. A key characteristic of our approach leverages the inherent inductive bias of time series data of its decomposed trend and seasonal components, each modeled separately to enable decoupled analysis and modeling. Specifically, we propose an innovative Conditional Denoising Seasonal Module (CDSM) within the diffusion model, which leverages statistical information from the historical window to conditionally model the complex seasonal component. Notably, we incorporate a Polynomial Trend Module (PTM) to effectively capture the smooth trend component, thereby enhancing the model's ability to represent temporal dependencies. Extensive experiments validate the effectiveness of our framework, demonstrating superior performance over existing methods and highlighting its flexibility in time series forecasting. The source code is available at <a class="link-external link-https" href="https://github.com/zjt-gpu/FDF" rel="external noopener nofollow">this https URL</a>.

What problem does this paper attempt to address?

This paper attempts to solve an important problem in time - series prediction: Existing diffusion models have a significant drawback in time - series prediction, that is, adding noise to the original time - series indiscriminately and then performing denoising processing. This method may mask the underlying dynamic evolution trend and complicate the prediction. Specifically, these methods do not take into account the unique inductive biases (such as trend smoothness and seasonal regularity) in time - series data during the process of adding and removing noise, resulting in the possible loss or distortion of key time structures during the noise - adding and - removing stages, thus affecting the prediction accuracy. To overcome this challenge, the authors propose a novel Flexible Decoupled Framework (FDF), which exploits the inherent inductive biases of time - series data by decomposing it. FDF decomposes the time - series into trend and seasonal components and models each component separately, thereby achieving more accurate analysis and modeling. Specifically, the paper proposes two core modules: 1. **Conditional Denoising Seasonal Module (CDSM)**: This module introduces an innovative conditional denoising strategy in the diffusion model, using statistical information in the historical window to conditionally model complex seasonal components. This enables the model to capture and process seasonal patterns in the time - series more effectively. 2. **Polynomial Trend Module (PTM)**: This module uses a polynomial linear layer to effectively capture the smooth trend component, thereby enhancing the model's ability to represent time - dependence. Through this decoupled method, FDF can not only capture complex patterns in the time - series more accurately, but also retain key trend information and improve the prediction accuracy. The paper verifies the effectiveness of this framework through extensive experiments and demonstrates its superior performance in time - series prediction tasks.

FDF: Flexible Decoupled Framework for Time Series Forecasting with Conditional Denoising and Polynomial Modeling

FDF: Flexible Decoupled Framework for Time Series Forecasting with Conditional Denoising and Polynomial Modeling

Denoising Diffusion Straightforward Models for Energy Conversion Monitoring Data Imputation

Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting

Disentangled Interpretable Representation for Efficient Long-term Time Series Forecasting

FDNet: Focal Decomposed Network for Efficient, Robust and Practical Time Series Forecasting

A Financial Time Series Denoiser Based on Diffusion Model

Not All Frequencies Are Created Equal:Towards a Dynamic Fusion of Frequencies in Time-Series Forecasting

Channel-aware Contrastive Conditional Diffusion for Multivariate Probabilistic Time Series Forecasting

Non-autoregressive Conditional Diffusion Models for Time Series Prediction

Generative Time Series Forecasting with Diffusion, Denoise, and Disentanglement

FreDF: Learning to Forecast in Frequency Domain

DisenTS: Disentangled Channel Evolving Pattern Modeling for Multivariate Time Series Forecasting

Differential Convolutional Fuzzy Time Series Forecasting

DEFM: Delay E mbedding based Forecast Machine for Time Series Forecasting by Spatiotemporal Information Transformation

Latent Diffusion Transformer for Probabilistic Time Series Forecasting

TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting

Retrieval-Augmented Diffusion Models for Time Series Forecasting

REDI: Recurrent Diffusion Model for Probabilistic Time Series Forecasting

TFDNet: Time-Frequency Enhanced Decomposed Network for Long-term Time Series Forecasting

Periodicity Decoupling Framework for Long-term Series Forecasting