火电机组的大范围变负荷运行对过热蒸汽温度(SST)的稳定性提出了严峻挑战。准确有效地预测SST对过热系统的控制起着重要作用。本文提出一种基于多元时序集成方法的SST预测模型,在时间卷积神经网络(TCN)基础上引入注意力机制(Attention TCN,TCAN),提升对SST多元时序数据的特征提取能力,再将能量守恒约束作为等式约束引入TCAN架构的数据驱动模型损失函数中,增强模型泛化能力。并将SST与喷水流量之间的物理关系作为不等式约束引入上述损失函数中,使混合模型可表征正确的物理特征。为验证所提方法的优越性,基于火电厂采集的大范围负荷变化的运行数据进行了仿真测试,对比结果表明,基于物理损失函数的混合模型的预测效果不仅具有通用性,而且与阶跃响应的动态特性保持科学一致。

The large-scale variable load operation of thermal power units poses a severe challenge to the stability of superheated steam temperature ( SST ). Accurate and effective prediction of SST plays an important role in the control of superheated system. In this paper, a SST prediction model based on multivariate time series integration method is proposed. Attention TCN ( TCAN ) is introduced on the basis of time convolution neural network ( TCN ) to improve the feature extraction ability of SST multivariate time series data. Then, energy conservation constraints are introduced into the data-driven model loss function of TCAN architecture as equality constraints to enhance the generalization ability of the model. The physical relationship between SST and water jet flow is introduced into the above loss function as an inequality constraint, so that the hybrid model can characterize the correct physical characteristics. In order to verify the superiority of the proposed method, the simulation verification is carried out based on the operation data of large-scale load changes collected by thermal power plants. The comparison results show that the prediction effect of the hybrid model based on the physical loss function is not only universal, but also scientific and consistent with the dynamic characteristics of the step response.

TP183? ???