Research Of Fuel Cell Based On Support Vector Regression

Fuel cell (FC), be capable of directly converting the chemical energy of fuel toelectrical energy and thermal energy by electrochemical reaction, is regarded as thefourth-generation of devices for generating electricity following hydroelectric power，thermal power and nuclear energy. Because it is not limited by the Carnot cycle, itsenergy conversion efficiency can be more than60%and is2~3times as that ofinternal-combustion engine. In addition, because there is no burning process duringworking, fuel cell does not emit harful gases such as sulfur oxide (SOx) and the nitrogenoxide (NOx) so that it has little pollution to environment. Therefore, as a kind of highefficient and clean energy, fuel cell has been competing for development by differentcountries in the21stcentury.The Support Vector Machine (SVM) brought forward by Vapnik and others, is anew machine learning method based on the VC dimension theory and the StructuralRisk Minimization Principle of the Statistical Learning Theory. It integratedtechnologies in maximum interval hyper-plane, Mercer kernel, convex quadraticprogramming, sparse solution and slack variable, and solved various practical problemssuch as small samples, nonlinearity, over learning/fitting, high dimension and localminimization problem, etc. In recent years, SVM has been successfully employed tosolve classification and regression problems in many fields. When SVM is applied toregression, it is called support vector regression (SVR).The FC system is a multi-input and multi-output system, a sound model can helpus to simulate, optimize and evaluate for FC. For example, one can foresee the output ofFC under different operating conditions by using a correlation model. However, most ofthe existing models for FCs are either too complicated, or the accuracy is not highenough for researchers and users. In this thesis, the nonlinear and offline models for FCsphysical properties are constructed by using the SVR approach combined with particleswarm optimization (PSO) algorithm for its parameter optimization. The research offuel cell based on PSO-SVR, is helpful to improve the experimental efficiency, can savea lot of manpower, valuable time and financial resources, provides a new clue for FCresearch, and would significantly promote the development of FC technology progressand FC development.The main contents of this thesis are as follows: (1) The basic information of the fuel cells are briefly summarized and analyzed,including their working principles, types, characteristics, development and theirapplications, etc.(2) The theory of SVM has been introduced briefly, which contains machinelearning theory, statistical learning theory and kernel function theory, etc.(3) According to experimental membrane water content dataset of proton exchangemembrane fuel cell (PEMFC), which was measured under different operatingtemperature and membrane impedance, a PSO-SVR model was established tomodel/predict the PEM membrane water content for PEMFC. The relationship betweenPEM membrane water content and two factors (cell temperature, membrane impedance)is very complicated and exists high-nonlinear, but the predicted value of PSO-SVRmodel can match the experimental value very well. The maen absolute error(MAE)=0.01, mean absolute percentage error (MAPE)=0.16%, correlation coefficient(R2)≈1.00, respectively. In addition, the available maximum PEM membrane watercontent and minimum PEM membrane water content are predicted by using theestablished PSO-SVR model, i.e., when the operating temperature is51.5℃andmembrane impedance is1.96m, the maximum PEM membrane water content wouldbe λmax=9.73; when the operating temperature is24.0℃, membrane impedance is27.20m, the minimum PEM membrane water content would be λmin=1.84.(4) As the electrical power of PEMFC can be affected by the operatingtemperatures, operating pressures, anode/cathode humidification temperatures,anode/cathode stoichiometric flow ratios, these factors were acted as input variables andthe electrical power of PEMFC as output, a PSO-SVR model was constructed to predictthe electrical power of PEMFC. The results illuatrate that predicted the MAE is0.156W,the MAPE is0.68%and R2reaches0.998. These results demonstrate the calculatedvalues via the PSO-SVR model are quite agreement with the measured indices.(5) By taking the direct methanol fuel cells (DMFC) cell temperature and cellcurrent density as input parameters, a PSO-SVR model was built to forecast the DMFCoutput voltage, the predicted performance was compared with that of Artificial NeuralNetwork (ANN) model. The result reveals that the predicted MAE and MAPE ofPSO-SVR reaches0.004990V and0.93%, which are superior to those(MAE=0.009943V and MAPE=2.23%) of ANN, respectively; the correlation coefficientR2=0.995of POS-SVR is also greater than that (0.991) of ANN model. This confirmedthat the PSO-SVR model regression/prediction ability surpasses that of ANN model, thus the PSO-SVR model is more suitable for modeling/predicting the DMFC cellvoltage.(6) Based on the measured electrical conductivity dataset of BSCF-SSC compositecathode, a PSO-SVR model is constructed and employed to predict the BSCF-SSCcomposite cathode electrical conductivity. The results reveal that calculated MAE,MAPE and R2by the PSO-SVR model are0.0467S/cm,0.09%and0.999, respectively.These means that the predicted value of PSO-SVR model is tallied well with theexperimental value, and the PSO-SVR model can be used for BSCF-SSC compositecathode conductivity prediction. Use PSO-SVR model to predicted BSCF-SSCcomposite cathode conductivity, an available maximum BSCF-SSC composite cathodeconductivity was predicted via the PSO-SVR model, i.e., when the operatingtemperature is344℃and SSC content is39wt%, it would reach242.9S/cm.