.png){kind=icon}
.png){kind=icon}
Return
Time series decomposition network for long-term series forecasting of non-stationary industrial processes
.png){kind=icon}
PRE
AI
Y
X
Z
H
Y
DOI：10.1016/j.jprocont.2026.103699.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Long-term series forecasting (LTSF) plays a crucial role in energy efficiency analysis and operational optimization in industrial production processes. However, achieving accurate LTSF in the industrial domain remains a formidable challenge, primarily due to two inherent data characteristics: strong non-stationarity, which undermines the stability of temporal dependencies, and pronounced lag effects, which obscure the causal relationships in trend evolution. The existing LTSF models often fail to concurrently solve these interwoven dynamic problems, leading to suboptimal performance. Therefore, a novel non-stationary time series decomposition network (NSDnet) is proposed for the LTSF of industrial processes, which consists of the non-stationary attention mechanism and the time delay aggregation (TDA). The non-stationary attention mechanism reintegrates non-stationary features into the temporal dependency modeling after data stabilization to address the non-stationary issues faced by seasonal forecasting. Then, the TDA performs rolling aggregation on the time series based on the selected delay factor to address the potential lag characteristics in trend prediction of industrial process data. Meanwhile, by utilizing unsupervised pre-training of the NSDnet to extract seasonal and trend features, and then using the next moment data as the supervised training label for the NSDnet, ultra-realtime long-term prediction is achieved. Finally, the NSDnet is validated on five benchmarks and an actual industrial production dataset. The experimental results show that the NSDnet achieves state-of-the-art (SOTA) results, where the average mean square error and average mean absolute error are reduced by 4.31% and 3.94%, compared with the previous SOTA model TimesNet, demonstrating the outstanding potential in guiding complex industrial production.
Keywords:
Neural network
Time series forecasting
Series Decomposition
Production prediction
Petrochemical industrial production
Journal
IF:
3.9
Papers:
3.4K
Citations:
7.3K
Organization
Cited Papers