The RUL Prediction Of Hydrogen Fuel Cells Based On Recurrent Neural Network

The accurate prediction of the remaining useful life(RUL)of hydrogen fuel cells has always been an important direction of the research field of fuel cells,which is of great significance for commercial promotion of fuel cells.At present,the RUL prediction methods of hydrogen fuel cells can be divided into 3 categories: the model-driven methods,the data-driven methods and the hybrid methods.However,due to some limitations of the model-driven methods and the hybrid methods,the data-driven methods are most widely appiled at present.Data-driven methods mainly include support vector machine(SVM),the grey model,neural networks and so on.Among them,neural networks related methods have good generalization ability,and can learn the nonlinear characteristics of the system autonomously,so they are widely used in this field.Especially the recurrent neural network(RNN),which has memory for data,is particularly suitable for processing time series data such as fuel cells aging data.Therefore,in this paper,the RUL prediction of hydrogen fuel cells is studied by some neural networks.The main contents are as follows:Firstly,the research status of the RUL prediction of hydrogen fuel cells at home and abroad is introduced in this paper.Besides,the advantages and disadvantages of some proposed methods are analyzed.For example,the limitation of SVM and the grey model to large-scale data,the memoryless of traditional neural networks and so on.Secondly,the basic structure and working principle of hydrogen fuel cells are introduced in this paper.And the factors that may affect the current operation state of the fuel cells are pointed out.Then,according to the data adopted in this paper,some physical parameters of the fuel cell stack are explained.At the same time,the percentage of stack voltage drop is determined as the criterion to judge the end of life.At last,the principle of back-propagation neural network(BPNN)is introduced in this paper,and some shortcomings of BPNN for RUL prediction are pointed out.Then,the index to judge the superiority and inferiority of the prediction model are given.After removing the abnormal points and smoothing the ageing data of the fuel cell,the BPNN is applied for the RUL prediction of the hydrogen fuel cell.And the RNNs are proposed in the next.Because of the risk of gradient disappearance in the normal RNNs,its variants LSTM(Long Short-term Memory)and GRU networks are proposed to predict the short-term RUL of the fuel cell at different starting points in this paper.The short-term RUL prediction could be mainly divided into single-step prediction and multi-step prediction.The prediction results show that LSTM and GRU have higher prediction accuracy than BPNN.And the convergence speed of GRU is faster due to the neuron of GRU is much simpler than LSTM neuron,so it is more suitable for the demand of online RUL prediction.Then,GRU is applied to predict the long-term RUL of the fuel cell in this paper,which verifies the effectiveness of GRU network in the field of hydrogen fuel cell life prediction.