Study On Deep Learning-Based Dynamic Performance Modeling Of Auto-Motive Proton Exchange Membrane Fuel Cells

In the past two years,countries successively proposed their own carbon neutrality goals,and most countries planned to accelerate the development of the hydrogen energy industry chain in their technical routes to achieve carbon neutrality.Proton exchange membrane fuel cell(PEMFC),which is the downstream in the hydrogen energy industry chain,has the advantages of zero emission,high energy conversion efficiency,no mechanical vibration and low noise.Therefore,it is regarded as one of the green energy conversion devices with great application prospects.In order to accelerate the development and application of PEMFC and reduce time and economic costs,mathematical models are widely used to simulate the PEMFC.Physical modeling has always been the mainstream modeling method,which is based on the basic physical and chemical processes inside the PEMFC and solving the complex coupled governing equations.The model complexity would increase exponentially for increasing the model accuracy.In contrast,the data-driven modeling focuses on finding the input-output relationships from experimental datasets collected from the operation of a fuel cell system,without the in-depth knowledge of these physical and chemical processes,and thus generally has the high computational efficiency and accurate results.Therefore,the main purpose of this study is to use deep learning algorithms instead of existing physical models to obtain a reliable and fast computing surrogate model.The research can be summarized as follows:(1)A quasi-2D transient model of PEMFC is constructed by electrochemical mechanism,which can simulate the dead-ended anode and anode recirculation mode of the PEMFC.The relationship model of volume flowrate,rotational speed and outlet pressure of hydrogen circulating pump,the inertial link model of the drive motor and the proportional-integral-derivative(PID)control model of the control voltage of the drive motor are built,and the parameters of PID control are optimized by the genetic algorithm,which greatly improves the response speed of the hydrogen circulating pump.The three models are combined to build a transient model of the hydrogen circulating pump that can automatically adjust the rotational speed and control the outlet pressure.Then,a pipeline model with time delay is constructed by using the pressure drop formula and the velocity formula to simulate the flow process of gas and liquid in the pipeline and the response lag problem in the operation of the fuel cell system.Finally,the above models are coupled with the air compressor model,membrane humidifier model and radiator model of the research group to obtain a transient system simulation model,which provides a solid simulation and data basis for the fuel cell dynamic performance prediction research.(2)The M-ANN method is proposed to predict the short-term performance degradation behaviors of PEMFC under different operating conditions.The main idea is using the multivariate polynomial regression(MPR)method to predict the initial value under different conditions,and using artificial neural network(ANN)to predict the variation value over time.Comparing the M-ANN method with the MPR method and ANN method in predicting short-term performance degradation behaviors,it is found that the M-ANN method is better than the other two methods.The reason is that M-ANN method can achieve complementary advantages between the algorithms by decomposing the problem,and finally achieves higher prediction accuracy.On this basis,the effects of different hidden layers and activation functions on the prediction accuracy and performance prediction curves are compared,and the current optimal number of hidden layers is found.In addition,it is found that the predicted curve obtained by Sigmoid activation function is smoother than that obtained by Re LU(Rectified Linear Unit)activation function.Although there is no obvious difference in error,it is obvious that the smoothed prediction result curve is more consistent with the real situation.(3)MPR,support vector regression(SVR)and LSTM-MLP are used to monitor the health status of fuel cell stacks during long-term operation.The LSTM-MLP method with the highest prediction accuracy is selected.This is because the area specific resistance(ASR)of the predicted target is not only related to the current working condition,but also greatly affected by the historical working condition and has a certain time lag effect,so the LSTM-MLP method with memory function is more appropriate.On this basis,the structure of LSTM-MLP model is optimized,and then the optimal model structure is used to perform sensitivity analysis on the input features.It is found that the unimportant features are the same whether predicting ASR or predicting voltage.After removing these three feature inputs,the prediction accuracy is even slightly improved.Then the optimal input sequence time for the predicted target value is obtained,which improves the prediction accuracy effectively.Finally,the sampling frequency is reduced to 0.025 to 0.05 times the original one with minimal reduction in prediction accuracy,which greatly reduces the model complexity and the amount of input data.