Observer-based adaptive neural network control for PEMFC air-feed subsystem

Abstract

Oxygen excess ratio (OER) tracking of polymer electrolyte membrane fuel cell (PEMFC) air-feed subsystem is an interesting topic to obtain a high energy conversion ratio. The control purpose is challenging due to the unknown nonlinear function, time-varying external disturbance and unmeasured variable. To deal with the above problems, a neural network control scheme based on tracking error observer and disturbance observer is proposed. Firstly, a canonical form of the PEMFC air-feed subsystem is developed based on the input–output feedback linearization method. Meanwhile, a tracking error observer which only uses the system output is developed for the reconstruction of the unmeasured variable. Secondly, the neural network based on a novel adaptive learning law is proposed to approximate the unknown nonlinear function. An identification model with a second-order low-pass filter is designed such that the adaptive learning law is adjusted by the observer error and modeling error, simultaneously. Finally, to reduce the effect caused by compound disturbance including the external disturbance and neural network approximation error, a nonlinear disturbance observer is proposed. The main contribution of this study is not only avoiding the prior knowledge of system uncertainty, external disturbance and the unmeasured variable but also improving the OER tracking performance under the measurement noise. Lyapunov theory analysis shows that the system tracking error is uniformly ultimately bounded. Numerical simulations and hardware-in-loop (HIL) experiments show that the proposed neural network control algorithm can maintain OER at the target value even under external disturbance. The experimental results show that the performance indexes including the mean absolute error (MAE), the root mean square error (RMSE) and the standard deviation (SD) of the proposed controller during the whole process are 0.0213, 0.0547 and 0.0538, respectively. Compared with the proportional–integral–derivative (PID) controller (MAE 0.0466, RMSE 0.0762 and SD 0.0745), robust adaptive radial basis function neural network (RBFNN-H∞) controller (MAE 0.0841, RMSE 0.0839 and SD 0.0813) and hybrid robust adaptive radial basis function neural network (HARBFNN-1st) controller (MAE 0.0255, RMSE 0.0613 and SD 0.0603), the proposed controller performs a better performance.