Adaptive neural network observer for proton-exchange membrane fuel cell system

This paper develops an adaptive neural network (NN) observer for proton-exchange membrane fuel cells (PEMFCs). Indeed, information on the oxygen excess ratio (OER) value is crucial to ensure optimal management of the durability and reliability of the PEMFC. The OER indicator is computed from the mass of oxygen and nitrogen inside the PEMFC cathode. Unfortunately, the measurement process of both these masses is difficult and costly. To solve this problem, the design of a PEMFC state observer is attractive. However, the behaviour of the fuel cell system is highly non-linear and its modelling is complex. Due to this constraint, a multilayer perceptron neural network (MLPNN)-based observer is proposed in this paper to estimate the oxygen and nitrogen masses. One notable advantage of the suggested MLPNN observer is that it does not require a database to train the NN. Indeed, the weights of the NN are updated in real time using the output error. In addition, the observer parameters, namely the learning rate and the damping factor, are online adapted using the optimization tools of extremum seeking. Moreover, the proposed observer stability analysis is performed using the Lyapunov theory. The observer performances are validated by simulation under MATLAB (R)/Simulink (R). The supremacy of the proposed adaptive MLPNN observer is highlighted by comparison with a fixed-parameter MLPNN observer and a classical high-gain observer (HGO). The mean relative error value of the excess oxygen rate is considered the performance index, which is equal to 1.01% for an adaptive MLPNN and 3.95% and 9.95% for a fixed MLPNN and HGO, respectively. Finally, a robustness test of the proposed observer with respect to measurement noise is performed.