针对重型燃气轮机负荷系统具有复杂耦合关系、高度非线性以及内部特性难以获取等特点,设计了一种基于LSTM-XGBoost集成模型的非线性模型预测控制策略,用于提高燃气轮机负荷系统的设定值跟踪能力。结合LSTM和XGBoost两种不同的网络构建集成模型,对燃气轮机负荷系统的输出功率及排气温度进行预测,两种输出参数的测试集均方根误差分别为0.060 3和0.064 1。仿真结果表明,LSTM-XGBoost集成模型可以实现两种输出参数在时间序列上的多步预测,且极大程度上提高了单一模型的预测精度。之后,利用该模型设计了基于数据驱动的模型预测控制器,并采用蛇优化(SO)算法与目标函数结合,进行滚动优化。与常规预测控制策略相比,引入SO算法的LSTM-XGBoost控制器在不同功率指令下的超调量分别缩小至3.2%和0.2%,排气温度控制的超调量为0,实现了重型燃机负荷系统的多输入多输出非线性预测控制,提高了设定值跟踪的准确性及快速性。

A nonlinear model predictive control strategy based on LSTM-XGBoost integrated model is designed to improve the setpoint tracking capability of the gas load turbine system for the heavy-duty gas turbine load system, which is characterized by complex coupling relationships, high nonlinearity, and difficulty in obtaining internal characteristics. The integrated model is constructed by combining two different networks, LSTM and XGBoost, to predict the output power and exhaust temperature of the gas turbine load system, and the root-mean-square errors of the test sets of the two output parameters are 0.060 3 and 0.064 1, respectively. The simulation results show that the LSTM-XGBoost integrated model can realize multi-step prediction of the two output parameters in the time series, and it can greatly improve the predicting capability of the gas turbine system. multi-step prediction, and it greatly improves the prediction accuracy of a single model. Afterwards, a data-driven model-based predictive controller is designed using this model, and the snake optimization (SO) algorithm is combined with the objective function for rolling optimization. Compared with the conventional predictive control strategy, the overshoot of the LSTM-XGBoost controller with the introduction of the SO algorithm is reduced to 3.2% and 0.2% under different power commands, and the overshoot of the exhaust temperature control is zero, which realises the multi-input and multi-output nonlinear predictive control of the load system of the heavy-duty combustion turbine, and improves the accuracy and rapidity of the setpoint tracking.

国家自然科学基金项目