Data-driven modeling for ultra-supercritical unit based on bidirectional test-time training and improved temporal convolutional network

Ultra-supercritical (USC) unit is extensively used to address the challenges posed by integrating significant amounts of renewable energy into the electricity grid, placing greater demands on the unit's flexible control capabilities. An accurate system model is essential for ensuring effective controller design. However, accurate dynamic modelling using traditional system identification methods under fast load varying of the unit is quite difficult. The unit generates a huge amount of data throughout its operational lifecycle. By establishing advanced data analytical techniques, this data can be utilized to assess the system's performance and decode the complex principles governing its internal behavior. In this research a novel data driven modeling strategy is developed, utilizing an integrated neural network framework. In this integrated neural network, a novel bidirectional RNN i.e. bidirectional test-time training (BiTTT) is applied as a means to extract the long-term dependence from the input data, while the improved temporal convolutional network (ITCN) handles extraction of complex local feature information within the data. Additionally, the discrete wavelet transform (DWT) method is utilized to mitigate noise within operational data. This methodology improves modeling precision by differentiating between valid information and noise. Finally, simulations and comparative experiments are carried out utilizing the actual operating data of a 1000 MW USC power unit. The proposed model accurately represents the operational features of the unit, and the mean square errors of the output power, separator outlet temperature, and main steam pressure are 1.86E+00, 8.75E-02, and 6.74E-04, respectively. Therefore, the proposed model can be applied to simulation analysis and controller design.