Performance Prediction And Multi-objective Optimization Of Fuel Cell Based On Neural Network

Fuel cell is an alternative energy conversion device.Because of its advantages of high efficiency,zero emission,and high reliability,it has been regarded as a promising new generation of power generation device in recent years.However,there are still two problems exist during the commercialization process of fuel cells.The first one is when operating conditions change in practical engineering,the performance indexes should be predicted quickly and precisely.The second one is that operating and geometric parameters should be properly adjusted to optimize the multiple performance indexes of the fuel cell simultaneously.Thanks to the advancement in machine learning,adopting intelligent algorithms such as artificial neural network(ANN)and genetic algorithm to model fuel cells has become a new research hotspot.Based on the two problems explored above,the application of ANN in fuel cell area is studied in this work,so that the prediction accuracy of fuel cell performance and the multi-objective optimization research of fuel cell can be further improved.Firstly,a three-dimensional proton exchange membrane fuel cell(PEMFC)model is developed through the computational fluid dynamics(CFD)technology.The model is validated based on the experiment parameters in literature.Results show that the simulation results are all in good agreement with the experimental data under three different working conditions.Therefore,the model is verified and can be utilized as the source of training data and the basic model for multi-objective optimization.Secondly,the wildly used back-propagation neural network(BPNN)is easy to fall into local optimum and its prediction accuracy needs to be further improved when it is applied to predict fuel cell performance.To solve this problem,the deep belief network(DBN)is employed in this study.The DBN model is trained through 1500 data which are obtained by the developed CFD model.Meanwhile,the data preprocessing method,fine-tuning epochs,and neural network topology are discussed to obtain the most suitable DBN model.Results indicate that the DBN model can predict the PEMFC performance precisely,and the DBN prediction accuracy is superior to that of other two intelligent methods.Moreover,training DBN with only 1500 data can complete the prediction of 41,600 working conditions in seconds,thus obtaining the optimal operating condition corresponding to the maximum power density quickly.Finally,the ANN is combined with the multi-objective genetic algorithm to conduct the multi-objective optimization of PEMFC.The decision variables are selected from eleven common PEMFC parameters using variance analysis.Then three data-driven ensemble learning models are trained as surrogate models to calculate fitness values of the optimization algorithm.The optimization algorithm is implemented to optimize the power density,system efficiency,and cathode oxygen distribution uniformity simultaneously.Results show that among numerious PEMFC parameters,operating pressure/temperature,anode stoichiometry,gas diffusion layer thickness,membrane thickness and channel width are the most significant parameters affecting the optimization objectives.The introduced optimization procedure combing optimization algorithm and surrogate models takes only 9 min 37 s to obtain the Pareto front,it can significantly reduce the optimization time.Compared with the basic model,the power density of the optimal model is improved by 0.117 W/cm~2,the system efficiency is increased by 11.04%,and the oxygen distributes more uniform on the cathode catalyst layer.