【摘要】 提出一种基于机理与数据混合的深度神经网络(PFNN)辨识方法，用于过热器温度被控对象的多变量回归建模，进而发展深度神经网络(DNN)动态前馈控制策略，抑制过热器温度的频繁波动.PFNN辨识方法融合过热器动态机理与长短期记忆神经网络，对强耦合、大时延的过热器温度对象具有很强的泛化能力和稳定性，可实现动态特性的精确预测.结果表明，机组灵活调峰时，过热器导前区温度和滞后区温度的PFNN辨识结果与现场实际温度的平均绝对偏差在1℃以内，模型可靠性较高.采用DNN动态前馈控制后，过热器出口蒸汽温度与设定值偏差维持在±2.5℃以内，有效抑制了因负荷变化引起的过热器蒸汽出口温度波动.更多还原

【Abstract】 An identification method based on physics-fusion neural network(PFNN) is proposed to establish a multivariable regression model for the superheater temperature, and then a dynamic feed-forward control strategy based on deep neural network(DNN) is studied to suppress the frequent fluctuation of the superheater temperature. The PFNN identification method is used to combine the dynamic mechanism of the superheater with the neural network of long-short term memory, it had strong generalization ability and stability for the superheater temperature with strong coupling and large time delay. It can precisely predict the dynamic characteristics of the superheater temperature. The results show that when the unit is in flexible peak regulation, the difference of the PFNN prediction results and the actual operation data is within 1 ℃ for the leading and lagging regions of the superheater temperature, validating the reliability of the PFNN model. When the DNN dynamic feed-forward control is used, the deviation of the steam temperature of the superheater outlet from the set value is within ±2.5 ℃, indicating that the proposed strategy can efficiently suppress the temperature fluctuations of the superheater steam outlet caused by load variations.更多还原