D-CTNet: A Dual-Branch Channel-Temporal Forecasting Network with Frequency-Domain Correction

Abstract: Accurate Multivariate Time Series (MTS) forecasting is crucial for collaborative design of complex systems, Digital Twin building, and maintenance ahead of time. However, the collaborative industrial environment presents new challenges for MTS forecasting models: models should decouple complex inter-variable dependencies while addressing non-stationary distribution shift brought by environmental changes. To address these challenges and improve collaborative sensing reliability, we propose a Patch-Based Dual-Branch Channel-Temporal Forecasting Network (D-CTNet). Particularly, with a parallel dual-branch design incorporating linear temporal modeling layer and channel attention mechanism, our method explicitly decouples and jointly learns intra-channel temporal evolution patterns and dynamic multivariate correlations. Furthermore, a global patch attention fusion module goes beyond the local window scope to model long range dependencies. Most importantly, aiming at non-stationarity, a Frequency-Domain Stationarity Correction mechanism adaptively suppresses distribution shift impacts from environment change by spectrum alignment. Evaluations on seven benchmark datasets show that our model achieves better forecasting accuracy and robustness compared with state-of-the-art methods. Our work shows great promise as a new forecasting engine for industrial collaborative systems.

• The paper introduces a dual-branch forecasting network that decouples temporal dynamics and inter-variable correlations for enhanced accuracy.
• It utilizes patch embedding, global self-attention fusion, and frequency-domain correction to mitigate non-stationarity in MTS data.
• Evaluations on seven benchmark datasets demonstrate significant improvements over state-of-the-art models, underlining its industrial applicability.

D-CTNet: A Dual-Branch Channel-Temporal Forecasting Network with Frequency-Domain Correction

Introduction

The increasing complexity of collaborative industrial systems necessitates accurate forecasting of Multivariate Time Series (MTS) data. The dependence on vast sensor networks for real-time data collection imposes a dual challenge on forecasting models: accurately modeling complex inter-variable dependencies while addressing the non-stationary distribution shifts due to dynamic environmental changes. "D-CTNet: A Dual-Branch Channel-Temporal Forecasting Network with Frequency-Domain Correction" proposes an innovative architecture to tackle these challenges and improve reliability in collaborative industrial environments (2512.00925).

Methodology

The proposed D-CTNet integrates a novel patch-based dual-branch design that decouples and learns intra-channel temporal patterns and multivariate correlations through a streamlined architecture.

Figure 1: Overall workflow of the proposed D-CTNet. It integrates four stages to model channel-temporal dependencies and mitigate non-stationarity. (1) Patch Embedding and Representation Learning. (2) Dual-Branch Channel-Temporal Modeling: Channel Attention and Temporal Featuring. (3) Global Inter-patch attention. (4) Stationarity Correction.

Patch Embedding and Representation Learning: Input time series are preprocessed using Reversible Instance Normalization and segmented into patches to facilitate local semantic representation and reduce redundancy.

Dual-Branch Channel-Temporal Modeling: This parallel architecture consists of a linear temporal modeling branch for capturing intra-channel temporal dynamics and a channel attention branch to model inter-variable correlations. This design ensures a decoupling in the feature learning process, enhancing both temporal and variable correlation modeling capabilities.

Global Inter-Patch Attention Fusion: Extending beyond local scopes, this module implements a global self-attention mechanism to fuse patch-level information, capturing long-range dependencies critical for industrial MTS data.

Frequency-Domain Stationarity Correction: Adaptive correction mechanisms are employed to align the spectral characteristics of prediction outputs with input data, mitigating non-stationarity-induced generalization losses effectively.

Results and Comparative Analysis

The D-CTNet was evaluated across seven benchmark datasets, including ETTm2, ETTh1, Electricity, and others, demonstrating superior performance in terms of forecasting accuracy. Notably, D-CTNet achieved the lowest Mean Squared Error (MSE) and Mean Absolute Error (MAE) across most datasets and forecast horizons, outperforming state-of-the-art models such as Transformers, GNNs, and MLPs.

Figure 2: Visualization comparison of forecasting results.

Especially in scenarios with long prediction horizons, such as in the Exchange dataset, the model's ability to accurately forecast demonstrates the effectiveness of the dual-branch and global fusion strategies. The ablation studies further confirm the importance of each component, showing significant drops in performance when modules like global attention fusion or frequency-domain correction are removed.

Implications and Future Directions

The proposed D-CTNet advances time series forecasting in collaborative industrial settings by significantly enhancing the accuracy and robustness against non-stationary data. These advancements provide a robust foundation for applications such as Digital Twin construction and predictive industrial maintenance.

Future work should investigate the deployment of D-CTNet on edge devices to accommodate real-time processing needs in Industrial IoT scenarios, possibly integrating cross-domain functionalities with LLMs to further expand its applicability in smart manufacturing ecosystems and beyond.

Conclusion

D-CTNet represents a significant step forward in addressing the dual challenge of modeling complex temporal and multivariate patterns while mitigating the effects of non-stationarity in MTS forecasting. Its dual-branch architecture, enhanced by global and spectral correction modules, sets a new benchmark for accuracy and reliability in the field. The model promises reliable data-driven insights essential for the advancement of Industry 4.0 technologies and collaborative industrial systems (2512.00925).