重型燃气轮机具有参数耦合关系高度非线性的特点，对传统机理建模与控制设计方法造成了较大的困难。基于神经网络方法的建模方式，以黑箱原理实现建模过程，提供了新的解决途径。生成的燃气轮机神经网络模型可嵌入预测控制算法中作为预测模型，为预测控制的实现提供新的思路。本文首先选取NARX神经网络和Elman网络两种前向型动态网络作为建模工具，利用重型燃气轮机现场运行数据，对起动过程进行动态工况建模，对运行参数进行预测。测试结果表明，合理范围内的隐含层神经元数量设置均可以获得可接受的预测模型。对于NARX神经网络，闭环运行模式应用价值更高，但预测误差大于开环运行模式。开环模式测试中表现更优秀的网络个体，不一定在闭环预测中保持更佳的预测精度。由于训练误差函数与测试误差评估函数的差异，对特定部分过程预测精度有更高要求的场景下，可以通过修改训练误差函数形式，对网络训练梯度下降方向进行引导。本文通过对训练误差函数中的输出参数高数值点赋予更高权重值，牺牲低数值区域的预测精度，获得了对高数值区域预测相对误差更小的网络个体。对于多组样本序列数据的训练问题，相比于交替训练、多组网络输出加权的处理方式，样本序列拼接、拼接点权重值置零的训练方式生成的网络模型预测精度更高，训练工作量更少。Elman网络不存在开环运行模式，在全过程闭环模式预测中表现优于NARX网络。但是由于承接层输入参数的初始化收敛问题，对于非训练样本起始状态的预测测试，Elman网络在前10步的预测精度远劣于NARX网络。验证NARX网络模型在燃气轮机动态过程的预测能力后，将其嵌入至非线性预测控制算法中，承担预测控制中非线性预测模型的角色。神经网络预测模型与传热搜索滚动优化算法和目标函数的结合，在仿真测试中实现了对重型燃气轮机Simulink模型对象的多输入多输出预测控制。在完成调峰任务的同时，兼顾对透平排气温度的约束。通过设计不同的目标函数排气温度项，可以实现升负荷过程中循环效率优先和升负荷率优先两种控制模式，证明了这一预测控制方法设计的灵活性。此外，经测试验证，该预测控制方法具备一定的抗扰动能力。

Highly nonlinear couple relationship among parameters of heavy-duty gas turbine confronts traditional mechanism modeling and control design with significant difficulty. The method of modeling based on neural network, which achieves the modeling work by black box principle, provides a new way to solve the problem. The generated gas turbine neural network model can be inserted into predictive control algorithm, inspiring a new way to achieving predictive control.In this paper, two types of forward dynamic network, NARX neural network and Elman network, were utilized to model heavy-duty gas turbine start-up transient condition on the basis of operation data, for the purpose of predicting operating parameters. Testing result revealed that, reasonable configuration of number of hidden layer neurons can all lead to acceptable prediction models. For NARX network, parallel operation mode has higher application value. But its prediction error is larger than series-parallel operation mode’s. The individuals which have better performance in series-parallel operation mode cannot promise better performance in parallel operation mode.Because of the difference between training error function and testing error function, under the circumstance of requiring higher prediction performance in particular part of process, gradient descent direction of network training can be guided by revising training error function. In this paper, training samples which have higher numerical value output parameters were assigned greater weight in training error function, to acquire network individuals whose performance in high numerical value zone was better, at the price of performance in low numerical value zone. As for the issue of multiple training sample sequences, split joint and zeroing the weight of junction operation in network training can acquire better network models and simplify training step, compared the method of training in turns and weighting the outputs of several networks.Elman network cannot work in series-parallel mode, but has better performance in whole sequence prediction compared to NARX network. Because of the convergence problem of context layer input parameters, Elman network is overwhelmed by NARX network in first 10 steps if the start point of testing does not match the initial state of training sequence.After the test of gas turbine dynamic prediction, NARX network model was inserted into nonlinear predictive control algorithm, playing the role of nonlinear prediction model in the algorithm. By combining neural network prediction model with heat transfer search receding horizon optimization algorithm and objective function, multi-input-multi-output predictive control of a heavy-duty gas turbine Simulink model object was achieved in simulation test. When the peak regulation task was completed, the constrain of turbine exhaust temperature was given consideration at the same time. In the method of regulating the term of turbine exhaust temperature in objective function, load increasing process can be control in cycle efficiency priority mode and load increasing rate priority mode, proving the flexibility of this predictive control design. Moreover, this predictive control method has the ability of rejecting disturbances to some extent, which was verified by test.