Accurate forecasting of pipeline displacement in thermal power plants is essential for deep peak regulation, operational safety, and structural health monitoring. However, hybrid-driven methods based on mathematical decomposition often involve high computational costs, while conventional models that rely on single loss functions struggle to capture long-term trends in multi-step forecasts. To address these challenges, this study proposes a Variational Mode Decomposition-inspired forecasting network that integrates a time–frequency-aware embedding module with a trend-aware loss function. The network employs a patch-based linear embedding to efficiently extract multi-scale temporal features and introduces a customized loss to improve trend alignment across forecasting horizons. The model is evaluated using real-world displacement data from a thermal power plant. The results demonstrate that it achieves higher forecasting accuracy, stronger trend preservation, and substantially better computational efficiency than both decomposition-based approaches and purely data-driven neural networks.
Dai et al. (Mon,) studied this question.