针对T60重型燃气轮机存在的动态特性强非线性、多变量耦合及工况大范围变化等特点，提出一种基于长短期记忆网络（LSTM）的非线性广义预测控制器设计方法。构建了燃气轮机多输入多输出LSTM动态模型，通过互信息法确定时间窗口，采用四分位距剔除异常值，并对数据进行Z-score标准化处理。进一步设计重型燃气轮机的非线性广义预测控制器，利用LSTM模型在线预测自由响应并采用数值逼近强制响应以构建动态矩阵，提出基于局部线性化的实时梯度优化策略，将非凸优化问题转化为可实时求解的二次规划子问题，并利用投影梯度下降法实时求解控制率。在MATLAB平台开展仿真实验，选取大范围阶跃负荷为扰动输入、控制周期100ms。通过与现场运行数据对比，实验结果表明：相比T60燃机在用的Min-Max控制器，NGPC控制器在负荷突变时超调量降低35%，响应曲线平滑度提高42%，控制误差显著降低，转速与温度约束违规率为零，单步计算耗时7-9ms（占控制周期9%以内），满足重型燃气轮机实时控制要求。

To address the strong nonlinear dynamics, multivariable coupling, and wide-ranging operating conditions of the T60 heavy-duty gas turbine, this study proposes a nonlinear generalized predictive controller design method based on Long Short-Term Memory (LSTM) networks. A multi-input multi-output LSTM dynamic model of the gas turbine is constructed. The time window is determined using the mutual information method, outliers are removed via the interquartile range, and data undergo Z-score normalization. A nonlinear generalized predictive controller for the heavy-duty gas turbine is further designed. The LSTM model predicts the free response online, while the forced response is approximated numerically to construct the dynamic matrix. A real-time gradient optimization strategy based on local linearization is proposed, transforming the nonconvex optimization problem into a solvable quadratic programming subproblem. The projection gradient descent method is employed to solve the control rate in real time. Simulation experiments were conducted on the MATLAB platform, employing wide-range step load disturbances as input with a 100-millisecond control cycle. Comparing results withfield operational data, the experiments demonstrate that compared to the Min-Max controller currently used on the T60 gas turbine, the NGPC controller reduces overshoot by 35% during load transients, improves response curve smoothness by 42%, significantly lowers control error, achieves zero violation rate for speed and temperature constraints, and completes single-step calculations in 7-9 ms (within 9% of the control cycle), meeting the real-time control requirements for heavy-duty gas turbines.

国家管网集团西部管道公司科研项目