提出一种自适应学习率记忆递归神经网络预测控制器及自适应学习率方法，它由用于预测和控制的子神经网络组成，预测子网络向控制子网络提供控制灵敏度信号；并证明了记忆递归神经网络学习的收敛性和稳定性条件．仿真结果表明控制器在线实时控制具有非线性、时变、多变量特性的水轮发电机组，对各种工况具有良好的性能． A self-adaptive learning rate memory recursive neural network predictive controller and adaptive learning rate method are proposed, which are composed of sub-neural networks used for prediction and control. The predictive sub-networks provide control sensitivity signals to the control sub-networks. The memory Convergence and Stability Conditions of Recurrent Neural Network Learning. The simulation results show that the controller real-time control of hydroelectric generator sets with nonlinear, time-varying and multivariable characteristics has good performance for various working conditions.