Temporal structure-preserving transformer for industrial load forecasting.

Area of Science:

• Energy Systems Engineering
• Artificial Intelligence
• Data Science

Background:

• Industrial park energy management requires accurate power load forecasting.
• Existing models face challenges with multi-target load series and integrating diverse exogenous variables.
• Optimizing energy efficiency and operational decisions hinges on precise forecasting.

Purpose of the Study:

• To introduce a novel architecture, the Temporal Structure-Preserving Transformer (TSPT), for industrial power load forecasting.
• To address the limitations of existing models in handling complex, multi-target series and integrating exogenous data.
• To enhance forecasting accuracy by preserving temporal structures and fusing multiscale patterns.

Main Methods:

• Decomposition of multi-target series into univariate series for parallel processing.
• Integration of exogenous variables (weather, production, efficiency data) using Gated Feature Fusion (GFF).
• Application of structure-preserving transformations to capture multiscale temporal patterns.

Main Results:

• TSPT demonstrated superior performance compared to state-of-the-art methods on a real-world industrial park dataset.
• The model effectively handled complex, multi-target forecasting tasks with integrated exogenous variables.
• Preserving temporal structure and parallel processing significantly enhanced forecasting accuracy.

Conclusions:

• TSPT offers a scalable and accurate solution for industrial power load forecasting.
• The proposed method improves energy management and operational decision-making in industrial settings.
• Effective integration of domain-specific knowledge enhances forecasting capabilities.