Domain Fusion Controllable Generalization for Cross-Domain Time Series Forecasting from Multi-Domain Integrated Distribution

TL;DR

TimeControl introduces a diffusion-based Domain-Fusion paradigm for cross-domain time series forecasting. By learning a shared, domain-fused prior through a multi-scale condition-denoising network and enabling efficient fine-tuning with a plug-and-play adapter, TimeControl achieves strong zero-shot and domain-transfer generalization across 49 benchmarks. The method simultaneously handles flexible-length forecasting via a patch-based tokenizer and a generation-style guidance mechanism that balances conditional fidelity with sampling diversity. Empirical results demonstrate superior zero-shot and fine-tuned forecasting, strong generation quality (Context-FID, QICE, CRPS, PICP), and notable efficiency gains over diffusion baselines and LLM-based approaches, underscoring the practical value of cross-domain diffusion for time series analysis.

Abstract

Conventional deep models have achieved unprecedented success in time series forecasting. However, facing the challenge of cross-domain generalization, existing studies utilize statistical prior as prompt engineering fails under the huge distribution shift among various domains. In this paper, a novel time series generalization diffusion model (TimeControl) that pioneers the Domain-Fusion paradigm, systematically integrating information from multiple time series domains into a unified generative process via diffusion models. Unlike the autoregressive models that capture the conditional probabilities of the prediction horizon to the historical sequence, we use the diffusion denoising process to model the mixed distribution of the cross-domain data and generate the prediction sequence for the target domain directly utilizing conditional sampling. The proposed TimeControl contains three pivotal designs: (1) The condition network captures the multi-scale fluctuation patterns from the observation sequence, which are utilized as context representations to guide the denoising network to generate the prediction sequence; (2) Adapter-based fine-tuning strategy, the multi-domain universal representation learned in the pretraining stage is utilized for downstream tasks in target domains; (3) A novel hybrid architecture is designed to align the observation and prediction spaces, enabling TimeControl to generate prediction sequences of arbitrary lengths with flexibility. We conduct extensive experiments on mainstream 49 benchmarks and 30 baselines, and the TimeControl outperforms existing baselines on all data domains, exhibiting superior zero-shot generalization ability.

Figures (19)

• Figure 1: Illustration of the paradigm in the domain-fused pre-training, fine-tuning and inference stages of the proposed TimeControl architecture. ① In the domain-fused training stage, all modules of TimeControl are trained end-to-end on the fusion dataset with the forecasting task as the metric. ② In the fine-tuning stage, only the adapter module is allowed to continue training on a specific dataset. ③ Finally, all the weights were frozen, and the prediction sequence was generated after $T$rounds of iterative denoising. Furthermore, ④ presents the TimeControl's architecture. (a) In the diffusion process, the original input sequence $X_0$ is progressively noised until degenerating into the Gaussian noise $X_T$. (b) In the context learning phase, mixed different domain sequences serve as inputs. The condition net captures cross-domain temporal fluctuation patterns as conditional variables to guide the generation process. (c) In the denoising phase, model accepts representations from multiple domains to reconstruct the fusion distribution. Forecasting by iterative denoising process.
• Figure 2: The overall framework of TimeControl . Specifically, the observation sequence $X_{-L+1:0}^{0}$and the diffusion timestep $t$ are processed by the input instance module to obtain the lookback embedding, trend-prompt embedding, and time embedding, which serve as inputs to the condition-denoising net and the adapter. The condition net captures the multi-scale representations $h_{a,b,c,m}$and the adapter transforms those into the context variables $\overline{h}_{a,b,c,m}$which utilized to guide conditional generation process for the prediction task.
• Figure 3: (a): Illustration of three sub-modules, where Transformer1D contains two versions that are utilized to establish condition-denoising net, respectively. Downsample1D and Upsample1D contain different linear interpolation layers. (b): In the non-conditional generation paradigm, the TimeControl architecture can be divided into there components, where the locked grey block shows the backbone of the pre-trained obtained TimeControl . based on this, several trainable consecutive adapter blocks (blue blocks) are added with a set of zero convolution layer (white) to construct the fine-tuned network. Which is used to align the multi-domain uniform space with the specified domain representation space.
• Figure 4: Visualisation on the validity of proposed condition net. Where the upper part shows the visualisations on ETTm1 and ETTm2 datasets.
• Figure 5: Demonstration of TimeControl prediction results on real long-term multi-periodic sequences sampled from Traffic dataset. Demonstration of TimeControl prediction results on real short-term non-periodic sequences sampled from ETT dataset. The blue and red lines respectively represent the actual values and predicted values of the future time series.