【作者】 刘云龙；

【导师】 李辉；

【作者基本信息】 东北师范大学 ， 计算机应用技术， 2018， 硕士

【摘要】 太阳能光电转化效率(Power Conversion Efficiency,PCE)是评价染料敏化太阳能电池(Dye Sensitized Solar Cell,DSSC)性能的一项至关重要的参数,因此,光电转化效率预测的精度直接影响到太阳能电池的性能。但是由于电池器件结构复杂,很难利用量子化学计算方法直接从分子结构性质准确计算出光电转化效率的计算值。而机器学习方法能有效地绕过复杂的实验过程,直接构造出染料分子结构性质与太阳能电池光电转化效率定量关系。本文结合量子化学方法(Quantum Mechanics,QM)与机器学习方法(Machine Learning,ML)建立量子化学与机器学习校正模型QM/ML模型,结合两者的优势来预测染料敏化太阳能电池光电转化效率。首先,利用量子化学B3LYP方法在STO-3G和6-31G*基组上分别计算出染料分子的物理化学性质;其次,采用3种机器学习方法:支持向量机(Support Vector Machine,SVM)、广义回归神经网络(General Regression Neural Network,GRNN)、分类与回归树(Classification and Regression Tree,CART)构建了集成学习器SVM-GRNN-CART(SGC);最后SGC结合三种特征选择方法增L去R选择算法(Plus-L Minus-R Selection algorithm,+L-R),随机Lasso算法(Randomized Lasso algorithm),遗传算法(Genetic algorithm)建立了集成学习级联回归模型。级联回归模型第一级输出为短路电流(Jsc)、开路电压(Voc)和填充因子(FF),它们作为第二级回归模型的输入来预测光电转化效率PCE。实验结果表明,与单一学习器SVM、GRNN、CART和同质集成学习器随机森林(Random Forest,RF)对比,集成学习器(SGC)级联回归模型在预测能力、拟合优度和模型的稳定性方面有明显优势。尤其在STO-3G基组上,SGC结合+L-R方法来预测PCE得到了最好的预测结果,其中平均绝对误差为0.37(%),均方根误差为0.50(%),决定系数为0.89。研究表明SGC可以更有效地预测有机染料敏化太阳能电池的光电转换效率,尤其在耗时较低的小基组STO-3G下获得较好的预测结果,为预测与设计合成新的染料分子提供了一个有效的工具,从而为实验合成节约了成本。更多还原

【Abstract】 Power Conversion Efficiency(PCE)is the most important parameter for the Dye Sensitized Solar Cell(DSSC)performance evaluation,So the prediction accuracy of PCE directly affects the performance of DSSC.However,due to the complex structure of the solar cell device,it is difficult to use the quantum chemistry calculation method to directly calculate the value of PCE from the molecular structure properties.The machine learning methods can effectively bypass the complicated experimental process and directly construct the quantitative relationship between the structure of the fuel molecule and PCE.In this research,combined with quantum mechanics(QM)and machine learning(ML)to establish a quantum chemistry and machine learning correction model,referred to as QM/ML model,which can combine with their advantages to predict PCE of DSSC.Firstly,the properties of fuel molecules are calculated based on the STO-3G and 6-31G*basis sets using the density functional method B3LYP.Then,three machine learning methods are used:Support Vector Machine(SVM),General Regression Neural Network(GRNN),Classification and Regression Tree(CART)to construct a ensemble learner SVM-GRNN-CART(SGC).Finally,SGC combines three kinds of feature selection methods Plus-L Minus-R Selection algorithm,Randomized Lasso algorithm,Genetic algorithm to construct ensemble learner cascaded regression model.The output of the first stage of the cascaded regression model is Short-Circuit Photocurrent Density(Jsc),Open Circuit Voltage(Voc),and Fill Factor(FF),which are used as input to the second level regression model to predict the power conversion efficiency.Experimental results show that compared with the single learner SVM,GRNN,CART and homogeneity ensemble learner Random Forest(RF),ensemble learner(SGC)cascaded regression model have obvious advantages over single learner regression model and homogeneity ensemble learner RF in predicting ability,goodness of fit,and stability.Especially on the STO-3G basis sets,SGC combined with+L-R to predict the PCE get the best prediction results,in which the mean absolute error(MAE)is 0.37(%),the root of mean square error(RMSE)is 0.50(%)and the coefficient of determination(R~2)is 0.89.This study shows that the ensemble learner cascade model(SGC)can more effectively predict the PCE of DSSC,especially on the time-consuming low STO-3G basis set,we get better prediction results which provide an effective tool for predicting and designing new dye molecules and save the cost for experimental synthesis.更多还原