The ultra-supercritical (USC) coal fired boiler-turbine unit is an advanced power generation system with low emissions and high efficiency. It is also a typical multivariable nonlinear system with great inertia. Generally, building an accurate analytic model using the conventional system identification methods are quite difficult. However, the big data generated by the monitoring system can reflect the USC unit's operation status and reveal the internal mechanism, if appropriate data analysis methods are developed. A deep neural network (DNN) is proposed in this paper to model a 1000 MW USC unit. In this DNN, stacked denoising auto-encoder is adopted to obtain the intrinsic features from the input data, while the long short-term memory network is in charge of outputting the expected normal behaviors of USC system along the time axis. Furthermore, to guarantee the convergence of this network, a reasonable intensity of added noise is identified via Lyapunov stability method. The DNN model is compared with the traditional multi-layer perception network, the stacked denoising auto-encoder, and two other random neural networks, to show the advantages in forecasting the dynamic behavior of USC unit.

超超临界（USC）燃煤锅炉-涡轮机组是一种先进的发电系统，具有低排放和高效率的特点。它也是具有大惯性的典型多变量非线性系统。通常，使用常规系统识别方法构建准确的分析模型非常困难。但是，如果开发了适当的数据分析方法，则监视系统生成的大数据可以反映USC单元的运行状态并揭示内部机制。本文提出了一种深度神经网络（DNN），以对1000 MW超超临界机组进行建模。在该DNN中，采用堆叠式降噪自动编码器从输入数据中获取固有特征，而长短期存储网络负责沿时间轴输出USC系统的预期正常行为。此外，为了保证该网络的收敛性，通过Lyapunov稳定性方法确定了合理的附加噪声强度。将DNN模型与传统的多层感知网络，堆叠式去噪自动编码器和其他两个随机神经网络进行比较，以显示预测USC单元动态行为的优势。