Machine Learning Based CdTe Solar Cell Simulation And Maximum Efficiency Point Searching

In the 21 st century,energy crisis and environmental problem are increasingly serious.Solar energy have been used widely as a huge amount of environmentally sustainable energy,and many high efficiency solar cells have been developed,and CdTe solar cell is one of them.However,due to the high cost of photovoltaic research,we need to improve the photovoltaic conversion efficiency of solar cells to lower the costs.There are many numerical simulation methods to explore the cells' performance,but the simulation speed based on the implicit nonlinear J-V curve is slow.Therefore,we need to build a faster and more efficient CdTe solar cell model,and then search for the maximum efficiency point,that is to find the parameter configuration of the highest efficiency CdTe solar cell.The data set modeled in this paper comes from the solar cell simulation software SCAPS,and there are 20002 data,8 input features,1 output value(conversion efficiency).Among them,the eight inputs are characterized by: CdTe layer thickness,CdTe acceptor concentration,CdTe layer defect level,CdTe layer defect concentration,concentration of CdS layer thickness,the CdS layer benefactor,the CdS layer defect level,the CdS layer defect concentrations.Using the grid search of machine learning,aiming at the material properties of emission and absorption layer of CdTe solar cell,we finally choose the LASSO regression model(1286 polynomials)with fifth order polynomial expansion to simulate the solar cell.The calculation time is five times faster than that of the general simulation software(SCAPS),and the prediction error was only about 1%.The LASSO model also has good performance in predicting the variation trend of efficiency,and it can be very precise to describe the efficiency trend of CdTe cell based on the defect distribution of CdS layer,the defect distribution of CdTe layer and the thickness of CdTe layer.However,the trend prediction is poor when it comes to the acceptor concentration of CdTe layer,the thickness of CdS layer and the donor concentration of the CdS layer.Through the analysis of the feature importance of the decision tree,we have found that the deviation of the efficiency trend prediction is due to the incomplete training data.For there are some deviation of the efficiency trend prediction based on certain characteristics,we can only find out the local maximum efficiency point based on the defect distribution of CdS layer,the defect distribution of CdTe layer and the thickness of CdTe layer through Q-learning model,and the local maximum efficiency point could be the global maximum efficiency point.