A Feature Enhancement Framework Based on Pre-trained Foundation Models for Time-series Forecasting in EIoT

With the advancement of the energy internet of things (EIoT), accurate predictive data in industrial power systems (IPS) has become crucial for fault detection, energy management and system optimization. In recent years, TimeSeries Pre-trained Foundation Models (TSPFMs) have demonstrated considerable potential in this application. However, their predictive performance remains constrained when confronted with domain-specific data with high noise, multi-periodic, and complex features. To improve the forecasting ability of TSPFMs, we propose a novel framework that merges a Frequency Enhanced Module (FEM) and a high-dimensional Temporal Position Attention (TPA) mechanism for feature enhancement. To be specific, the FEM processes multi-channel data in the frequency domain, effectively capturing feature independencies. The TPA generates channel-specific attention maps, thereby integrating a broader range of contextual information into local features. Additionally, we introduce a dynamic loss function that adjusts penalties according to the forecasting step size, thereby optimizing the long-term prediction performance. By performing experiments on multiple IPS datasets, the results indicate that the optimized model achieves a $12.50 \%$ to $16.82 \%$ improvement in prediction accuracy, demonstrating enhanced robustness and accuracy across various prediction lengths.