A transformer-based spatiotemporal forecasting for overheating risks in flexible operation of thermal power system

Accurate short-term forecasting of tube overheating risks in flexible thermal power plants is essential for the safe integration of high shares of renewable energy. Data-driven approaches that enable real-time proactive warnings offer a promising path toward intelligent power systems. However, the field is challenged by the multivariate nature of superheater tubes and their spatial correlations, while available external operating parameters provide only spatially lumped information. Here, through a systematic correlation analysis of 235 superheater tubes, we reveal pronounced spatial heterogeneity in the tube-wise dependence on externally measured steam and flue gas temperatures. In contrast, strong positive correlations are consistently observed among spatially adjacent tubes. Inspired by this finding, we develop a tailored Transformer-based model that synergistically captures both temporal dynamics and spatial correlation structures. On the test set covering all 235 tubes, the proposed model achieves a mean absolute error (MAE) of 0.42 °C for a 2-min prediction horizon, with spatial information improving prediction accuracy by 27.6%. Due to its high accuracy and multi-tube forecasting capability, the model enables reliable early warning of overheat events, attaining a hit rate of 97% with a false alarm ratio of 3.3% over 1302 overheat events.