Data-driven modeling of ultra-supercritical unit based on improved PatchTST network with multi-task balancing

Ultra-supercritical (USC) units play a critical role in modern power system due to their high efficiencies and low emissions, yet their complex nonlinear characteristics and long-term temporal dependence pose significant challenges for accurate modeling and control. Traditional models fail to capture both local patterns and long-term temporal dependencies in USC unit operational data. This leads to the model's poor generalization ability under different operating conditions. Existing methods typically adopt fixed weighting strategies in their loss functions that cannot adapt to the importance of different output variables during dynamic load changes, resulting in poor model generalization abilities under different operating conditions. In this study, a novel data-driven modeling approach is proposed that integrates an improved patch time series transformer network with adaptive multi-task balancing for accurate USC boiler–turbine coupled process modeling. First, a specialized patch-based strategy is established that effectively captures local patterns and long-term temporal dependencies through an enhanced self-attention mechanism, overcoming the inability of traditional methods to model multi-scale temporal relationships in USC units. Second, an improved reversible instance normalization module is designed to mitigate the problem of data distribution shift and improve the generalization across the varying operating conditions of USC units. Third, a multi-task adaptive weight learning loss function is designed to adjust task-specific weights during training, improving the response speed during rapid load changes. Comprehensive experimental validation with data from a 1000 MW USC unit revealed that the proposed method is superior to other modeling methods in terms of modeling accuracy and computational efficiency.