Solar cell performance analysis method using a neural network-based optoelectronic model

Abstract

The present invention discloses a method for analyzing the performance of solar cells using a photoelectric model based on a neural network, which solves the problem that in the process of perovskite cell performance analysis, the existing model parameter calibration time is long, the simulation efficiency is low, and the performance analysis requirements cannot be met quickly. The method includes establishing a thin film wave optical model with a matrix analytical model; constructing and training an artificial neural network model; merging the thin film wave optical model and the artificial neural network model to form a photoelectric model; inputting measured data into the photoelectric model to obtain multiple battery performance parameters, and performing corresponding battery performance analysis according to the battery performance parameters. The performance analysis parameters of the present invention are calculated based on the photoelectric model. The photoelectric model adopts an artificial neural network model, abandons the finite element solver with high calculation difficulty and long time consumption, and replaces it with a machine learning method. The parameter calibration time is short, the simulation efficiency of the photoelectric model is improved, and the demand for rapid analysis of solar energy performance is met.

Description

Technical Field

[0001] The invention relates to the technical field of perovskite cell photoelectric simulation, and in particular to a solar cell performance analysis method using a photoelectric model based on a neural network. Background Art

[0002] In the scenario of photovoltaic development and application, we usually use current-voltage (iV) curve characterization technology to measure the performance of the battery, including short-circuit current (i sc )、Open circuit voltage(V oc ), fill factor (FF) and photoelectric conversion efficiency (PCE), etc. These battery performances are closely related to the carrier transport properties of each layer of the battery thin film material, and can be analyzed through photoelectric model simulation.

[0003] In optics, the Beer-Lambert Law, various matrix algorithms, one-dimensional and higher numerical methods, etc., which can be used to calculate the short-circuit current of perovskite cells, are all extensions of the Maxwell equations. When applied to thin film structures with flat surfaces, the one-dimensional model used is fast and requires less calculation time.

[0004] However, in terms of electrical simulation, the methods that can be used to calculate the current-voltage curve of solar cells are usually the dual diode model and the carrier drift diffusion model. Both require parameter calibration, and the time usually takes hours or days, or even weeks or months in the case of large computational load or high precision, and it is impossible to quickly complete the calculation and predict and optimize the performance of the device. The complete carrier drift diffusion model contains multiple types of input parameters and is highly nonlinear. It requires practical finite element solution, which leads to an increase in the amount of calculation. The time required to calculate a single voltage is measured in seconds, and the simulation efficiency is low. Summary of the invention

[0005] The present invention mainly solves the problems in the performance analysis process of perovskite batteries that the existing model parameter calibration time is long, the simulation efficiency is low, and the performance analysis requirements cannot be met quickly. A solar cell performance analysis method using a photoelectric model based on a neural network is provided.

[0006] The above technical problem of the present invention is mainly solved by the following technical solution: A method for analyzing solar cell performance using a photoelectric model based on a neural network, comprising the following steps:

[0007] S1. Establish thin film wave optics model using matrix analytical model;

[0008] S2. construct and train an artificial neural network model, where the artificial neural network model is a final artificial neural network composed of a device performance artificial neural network, a maximum power point composite loss artificial neural network, and an opening voltage composite loss artificial neural network;

[0009] S3. Combining the thin film wave optics model and the artificial neural network model to form an optoelectronic model;

[0010] S4. The data to be tested is input into the photoelectric model to obtain multiple battery performance parameters, and corresponding battery performance analysis is performed according to the battery performance parameters.

[0011] The parameters used for solar cell performance analysis in the present invention are calculated according to a photoelectric model based on a neural network. The photoelectric model is a combined model of a thin film wave optics model and an artificial neural network model for carrier drift diffusion. The photoelectric model abandons the finite element solver with high computational difficulty and long time consumption, and replaces it with a machine learning method. The parameter calibration time is short, the simulation efficiency of the photoelectric model is improved, and the demand for solar energy performance analysis is quickly met.

[0012] The optoelectronic model obtained by combining the thin film wave optics model and the artificial neural network model is essentially an analytical model. The operation speed is faster than that of the one-dimensional / two-dimensional model, while maintaining reliable accuracy. When applied to large-scale or high-accuracy calculations and optimization, the time consumption is greatly reduced, and the requirements for computer hardware are also greatly reduced. The optoelectronic model avoids the use of software such as finite element solvers and can be stored on each server in the form of analytical solutions, which is practical, convenient and fast.

[0013] As a preferred solution, the construction and training of the artificial neural network model specifically includes the following process:

[0014] S21. Obtain input parameter data, sample and preprocess the input parameter data, and construct model sample data;

[0015] S22. Sample data is input into the carrier drift diffusion model to calculate the output parameters, and the sample data is classified according to the output parameter type;

[0016] S23. Establish an artificial neural network model, train the artificial neural network model with sample data, and obtain a final artificial neural network model that is a combination of the device performance artificial neural network, the maximum power point composite loss artificial neural network, and the open-voltage composite loss artificial neural network.

[0017] The artificial neural network model contains three branches, which are distinguished by different output parameters: battery performance parameters, battery compound loss at maximum power, and battery compound loss at open circuit voltage. These three branches can be used separately or simultaneously. The three branches of the artificial neural network model have all been optimized for the parameters of the perovskite solar cell device, reducing the error of the artificial network and improving the accuracy of the final artificial neural network model.

[0018] As a preferred solution, step S21 specifically includes:

[0019] Obtain parameters from the experimental characterization database and the material parameter database and confirm the parameter range;

[0020] The Latin hypercube sampling method is used to sample the parameters to obtain the input parameters;

[0021] The input parameter preprocessing includes:

[0022] Compare the maximum and minimum values ​​of the input parameter. If the difference between the maximum and minimum values ​​exceeds the set difference threshold, the input parameter is sampled using the decimal logarithm method.

[0023] Normalize the input parameters;

[0024] Sample data are constructed according to the preprocessed input parameters, including training data, validation data and test data.

[0025] This scheme obtains each parameter and determines the range of parameters from the experimental characterization database and material parameter database of perovskite solar cells, and ensures that these ranges cover existing materials and possible new materials in the future as much as possible. The input parameters are sampled from the parameters. The sampling method adopts the Latin hypercube sampling method. All input parameters change at the same time, so that the input parameters can be evenly distributed, and the number of samples required in theory is reduced, saving data acquisition and artificial neural network training time. The preprocessing of the input parameters includes judging the difference between the maximum and minimum values ​​of each input parameter. When there is a huge difference between the maximum and minimum values, the input parameters are sampled using the decimal logarithm method to ensure the uniform distribution of the data. The preprocessing also includes normalizing the input parameters to evenly distribute the data in the range of -1 and 1. The preprocessed input parameters are used to construct sample data, and a total of four sets of data are established, one for artificial neural network training training data, the sampling number is, for example, 100,000, one for verification verification data, the sampling number is 10,000, and two for testing test data, the sampling number is 10,000 each.

[0026] As a preferred solution, after step S21, the step further includes matching the input parameters with the energy level requirements of the materials / transmission layers of the solar cell. If the matching is successful, the input parameters are retained; if the matching is unsuccessful, the input parameters are discarded.

[0027] Since the energy levels of each material / transport layer of a perovskite solar cell need to match in order to generate electricity normally, energy level requirements need to be matched. When the energy level in an input parameter data does not match, the input parameter will be discarded and the remaining matching data will be retained.

[0028] As a preferred solution, step S22 specifically includes:

[0029] Input the sample data into the carrier drift diffusion model, solve it through the finite element solver, and output the current-voltage curve;

[0030] Analyze the current-voltage curve to obtain the battery open circuit voltage, fill factor, conversion efficiency, and radiation recombination loss, non-equilibrium carrier recombination loss, Auger recombination loss, and interface recombination loss at the maximum power point and open circuit voltage.

[0031] The input parameters are classified into three parts: performance parameters of output parameters, maximum power point compound loss, and open circuit voltage compound loss.

[0032] The training data source of the artificial neural network model in this scheme is the carrier drift diffusion model using the finite element solver, so that the final photoelectric model takes into account a wider range of perovskite solar devices and thin film parameters, has a higher data source accuracy, and has a larger adjustable range and degree of freedom of data compared to the double diode model.

[0033] As a preferred solution, an artificial neural network model is established, the structure of which includes an input layer, three intermediate hidden layers, a fixed hidden layer before the output layer, and an output layer. Bayesian regularization is used for training, the maximum number of iterations is one thousand, the maximum number of iterations allowed for the increase in verification error is ten, and the hyperparameters are determined by a Bayesian optimization program.

[0034] The structure of the established artificial neural network model is an input layer, three middle hidden layers, a fixed hidden layer before the output layer, and an output layer. The training logic is Bayesian regularization, the maximum number of iterations is one thousand, and the maximum allowed number of iterations for increasing the verification error is ten. The model sets a maximum number of iterations higher than the general requirement value and a lower allowed number of iterations for increasing the maximum verification error. In addition, due to the use of a large sample size and fixed training, verification and test data generated by different seeds, rather than splitting the data according to the usual K-fold cross-validation, overfitting is prevented while ensuring the training consistency and accuracy of the resulting neural network.

[0035] Bayesian optimization procedure for hyperparameters: The optimal hyperparameters of the artificial neural network are determined by Bayesian optimization. The hyperparameters that can be optimized include but are not limited to the initial learning rate, minimum gradient, momentum, activation function of each intermediate hidden layer, and number of neurons. The cost function used is the final mean square error of the artificial neural network training. The optimization is terminated by limiting the maximum optimization time, the target minimum value and minimum gradient of the cost function, and the maximum number of iterations.

[0036] The artificial neural network model finally obtained is used to achieve rapid battery device characterization, and its calculation results are synchronized to the database. At the same time, whenever a new batch of experimental data is input, the model can be directly retrained. Due to the retrainable nature of the artificial neural network model, after a certain number of experimental characterizations and calibrations, the model can be retrained regularly to ensure that the optoelectronic model keeps up with the development of materials and technology. Usually, retraining takes less time than initial training, and the optimized hyperparameters can be used directly.

[0037] As a preferred solution, after the photoelectric model is constructed, the photoelectric model is also calibrated, including the following steps:

[0038] Obtain short-circuit current, open-circuit voltage, fill factor and conversion efficiency from the experimental voltage-current curve, and confirm the material parameters or parameter ranges based on the materials used and the preparation method;

[0039] According to the short-circuit current, the optical parameters are calibrated in combination with the thin film wave optics model, including the real and imaginary parts of the refractive index of the material at each absorption wavelength, as well as the thickness of each film layer;

[0040] The undetermined parameters are calibrated based on the open circuit voltage, fill factor, conversion efficiency and artificial neural network model.

[0041] This solution can realize rapid calibration of optical parameters and film thickness, and rapid calibration of material and composite parameters. Calibration is performed as needed. When calibration is required, the input source of the artificial neural network model is the approximate parameters of the battery material and the current-voltage curve. Therefore, only current-voltage characterization is required in the laboratory, and no other instrument is required to assist in calibration. The calibration process is convenient and fast. The present invention solves the problem that the existing perovskite solar cell model parameter calibration takes a long time, and the simulation efficiency is low and cannot quickly meet actual needs.

[0042] As a preferred solution, the thin film wave optics model is:

[0043] X=A\B

[0044] Where X is the column matrix of the optical electric field intensity to be solved, A is the diagonal matrix composed of wave number and film thickness, and B is the column matrix composed of the light source used as excitation between each film.

[0045] The film wave relationship model of this scheme adopts a matrix analytical model, where matrix A is composed of wave number and film thickness, taking into account the refraction properties of light in each film and between each film layer, and matrix B represents the light source used as excitation between each film layer. The optical electric field intensity is calculated by dividing the data of each layer of matrix A by the data of the corresponding layer of matrix B by the left division.

[0046] As a preferred solution, the data to be measured is input into the photoelectric model to obtain multiple battery performance parameters, which specifically includes the following process:

[0047] The data to be measured are input into the thin film wave optics model. The method of obtaining the sample data is the same as that of obtaining the sample data. The optically generated carrier generation rate G is calculated as:

[0048] G=f(E z )=f optics (l,m,κ,I0,R a )

[0049] Among them, E z is the intensity of the optical electric field, l is the film thickness, m is the real part of the refractive index, κ is the imaginary part of the refractive index, I0 is the incident spectrum, R a is the surface reflectivity;

[0050] The data to be measured is combined with the carrier generation rate and input into the artificial neural network model to calculate the output parameter Output.

[0051] Output=f NN (G,W,χ,μ,N, ε,A aug ,B rag ,τ,v,V a ,T)

[0052] Where W is the metal work function, χ is the material energy level, μ is the carrier mobility, N is the material energy level density, ε is the material dielectric function, A aug is the Auger recombination constant, B rag is the radiative recombination constant, τ is the carrier lifetime, v is the interface recombination constant, V a is the applied voltage and T is the temperature.

[0053] As a preferred solution, the data to be tested is input into the photoelectric model to obtain battery performance parameters including device performance, maximum power point composite loss, and opening voltage composite loss data, and corresponding performance analysis is performed based on the performance parameters.

[0054] The output parameters of the optoelectronic model include three parts, namely device performance, maximum power point compound loss, and open voltage compound loss data. The output parameters can be used for parameter sensitivity analysis, compound loss analysis, and battery performance optimization. Among them, parameter sensitivity analysis can use the three parts of device performance, maximum power point compound loss, and open voltage compound loss data, and use Monte Carlo and other simulation methods to calculate the sensitivity of each output parameter to the final output parameter locally or globally, which is used to find the thin film layer or surface in the perovskite battery that is most likely to affect its performance. Compound loss analysis can use the maximum power point compound loss and open voltage compound loss data, and use the output parameters to calculate the proportion of each compound loss of the perovskite battery device at the maximum power point and open voltage, so as to determine how to effectively improve the fill factor and open circuit voltage. Battery performance optimization can use device performance data to directly apply the output parameters to various large-scale optimization algorithms.

[0055] Therefore, the advantages of the present invention are:

[0056] The photoelectric model used in the present invention for simulating and obtaining battery performance parameters adopts a combined model of a thin film wave optics model and an artificial neural network model for carrier drift diffusion, abandons the finite element solver with high computational difficulty and long time consumption, and replaces it with a machine learning method. The parameter calibration time is short, the simulation efficiency of the photoelectric model is improved, and the needs for solar energy performance analysis are quickly met.

[0057] The artificial neural network model uses Latin hypercube sampling valve, and the input parameters are evenly distributed, which reduces the samples required for theory and saves data acquisition and neural network training time. Instead of splitting data according to the usual K-fold cross-validation, a large sample size and fixed training, validation and test data generated by different seeds are used to ensure the consistency and accuracy of the training of the obtained neural network and prevent overfitting.

[0058] It can realize the rapid calibration of optical parameters and film thickness, and the rapid calibration of material and composite parameters. When calibration is needed, the input source of the artificial neural network model is the approximate parameters of the battery material and the current-voltage curve. It only needs to perform current-voltage characterization in the laboratory, and no other instruments are required for auxiliary calibration. The calibration process is convenient and fast, which solves the problem of time-consuming calibration of perovskite solar cell model parameters and low simulation efficiency in the process of battery performance analysis, which cannot meet the needs of rapid performance analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic diagram of a process of the present invention.

[0060] Figure 2 It is a flow chart of the calibration and use of the photoelectric model in the present invention.

[0061] Figure 3 It is a flow chart of the training of the artificial neural network module in the present invention. DETAILED DESCRIPTION

[0062] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0063] Embodiment 1:

[0064] This embodiment is a method for analyzing solar cell performance using a photoelectric model based on a neural network, such as Figure 1 As shown, the following steps are included:

[0065] S1. Establish thin film wave optics model using matrix analytical model.

[0066] As a preferred solution of this embodiment, the thin film wave optics model is constructed using a matrix analytical model. Through a series of numerical simplifications, a matrix composed mainly of 0 and 1 is obtained, focusing on versatility, program simplification and computational efficiency. The solution is stored in the form of mathematical symbols. The structure of the thin film wave optics model is:

[0067] X=A\B

[0068] Where X is the column matrix of the optical electric field intensity to be solved, A is the diagonal matrix of wave number and film thickness, taking into account the refraction properties of light in each film and between each film layer, and B is the column matrix of the light source used as excitation between each film. The optical electric field intensity is calculated by left-dividing the data of each layer of matrix A by the data of the corresponding layer of matrix B.

[0069] The thin film wave optics model is specifically expressed by the following formula:

[0070] X=A\B

[0071] A= ,

[0072] , ,

[0073] , ,

[0074] k0=2π / λ, k n = k0(m n -iκ n ),

[0075] .

[0076] Where N is the number of film layers, n is 1~N, L is the total length of the film, e is the base of the natural logarithm, i is the imaginary unit, E0 is the optical electric field intensity on the incident surface, E 0,b is the optical electric field intensity of the reflecting surface, l is the film thickness, m is the real part of the refractive index, κ is the imaginary part of the refractive index, x is the distance between the coordinate and the incident surface, S n A is a matrix that reflects the propagation properties of light in each film. n is the intensity of light propagating in the x direction at each film interface, B n is the intensity of light propagating in the negative x direction at each film interface.

[0077] S2. Construct and train an artificial neural network model, which is a final artificial neural network composed of the device performance artificial neural network, the maximum power point composite loss artificial neural network, and the open voltage composite loss artificial neural network. Figure 3 As shown, the specific steps include:

[0078] S21. Obtain input parameter data, sample and preprocess the input parameter data, and construct model sample data.

[0079] As a preferred solution of this embodiment, the following process is specifically included:

[0080] S211. Obtain parameters from the experimental characterization database and the material parameter database and confirm the parameter range; specifically, confirm the range of each input parameter from the experimental characterization data and the material parameter data, and try to ensure that these ranges cover all existing materials and possible new materials in the future.

[0081] S212. Use Latin hypercube sampling method to sample from parameters to obtain input parameters; specifically, the sampling method uses Latin hypercube sampling method, and all input parameters change at the same time, so that the input parameters can be evenly distributed, and the number of samples required in theory is reduced, saving data acquisition and artificial neural network training time.

[0082] S213. Input parameter preprocessing, including difference processing and normalization

[0083] The difference is handled as:

[0084] The maximum and minimum values ​​of the input parameters are compared. If the difference between the maximum and minimum values ​​exceeds the set difference threshold, the decimal logarithm method is used to sample the input parameter to ensure uniform distribution of the data.

[0085] The normalization process is:

[0086] Each input parameter is normalized to evenly distribute the data between -1 and 1.

[0087] S214. Construct sample data according to the preprocessed input parameters, including training data, verification data and test data.

[0088] As a preferred solution of this embodiment, a total of four sets of data are established here, including one set of training data for artificial neural network training, with a sampling number of, for example, 100,000, one set of verification data for verification, with a sampling number of 10,000, and two sets of test data for testing, with a sampling number of 10,000 each.

[0089] S215. Match the energy level requirements of each material / transmission layer of the solar cell to each input parameter. If the match is successful, the input parameter is retained. If the match is unsuccessful, the input parameter is discarded.

[0090] Since the energy levels of each material / transport layer of the perovskite solar cell need to match in order to generate electricity normally, energy level matching is required. When the energy level in an input parameter data does not match, the input parameter will be discarded and the remaining matching data will be retained.

[0091] S22. The retained sample data is input into the carrier drift diffusion model to calculate the output parameters, and the sample data is classified according to the output parameter type.

[0092] As a preferred solution of this embodiment, the following steps are specifically included:

[0093] S221. Input the sample data into the carrier drift diffusion model, solve it through the finite element solver, and output the current-voltage curve. In this process, the sample data is firstly imported into the carrier drift diffusion model using the finite element solver, and solved it through the finite element solver, and output the current-voltage curve.

[0094] S222. Analyze the performance parameters of the battery, such as open circuit voltage, fill factor, conversion efficiency, and radiation recombination loss, non-equilibrium carrier recombination loss, Auger recombination loss and interface recombination loss at the maximum power point and open circuit voltage, from the current-voltage curve;

[0095] S223. Classify the input parameters according to the performance parameters of the output parameters, the maximum power point composite loss, and the open circuit voltage composite loss.

[0096] The training data source of the artificial neural network model in the above scheme is the carrier drift diffusion model using a finite element solver, so that the final photoelectric model takes into account a wider range of perovskite solar devices and thin film parameters, has a higher data source accuracy, and has a larger adjustable range and degree of freedom of data compared to the double diode model.

[0097] S23. Establish an artificial neural network model, train the artificial neural network model with sample data, and obtain a final artificial neural network model that is a combination of the device performance artificial neural network, the maximum power point composite loss artificial neural network, and the open-voltage composite loss artificial neural network.

[0098] The structure of the artificial neural network model established includes an input layer, three middle hidden layers, a fixed hidden layer before the output layer, and an output layer. The training logic is Bayesian regularization, the maximum number of iterations is one thousand, and the maximum number of iterations allowed for the verification error increase is ten. The model sets a maximum number of iterations higher than the general required value and a lower maximum number of iterations allowed for the verification error increase. In addition, a large sample size and fixed training, verification, and test data generated by different seeds are used, rather than splitting the data according to the usual K-fold cross-validation, thereby ensuring the training consistency and accuracy of the obtained neural network and preventing overfitting.

[0099] Bayesian optimization procedure for hyperparameters: The optimal hyperparameters of the artificial neural network are determined by Bayesian optimization. The hyperparameters that can be optimized include but are not limited to the initial learning rate, minimum gradient, momentum, activation function of each intermediate hidden layer, and number of neurons. The cost function used is the final mean square error of the artificial neural network training. The optimization is terminated by limiting the maximum optimization time, the target minimum value and minimum gradient of the cost function, and the maximum number of iterations.

[0100] The artificial neural network model includes three branch neural networks, which are trained according to the input parameters after the above classification. The final artificial neural network model is composed of three optimized artificial neural networks with the same input and different outputs, which can be used to calculate single, multiple or all output parameters. The three neural network branches are device performance artificial neural network, maximum power point compound loss artificial neural network, and open voltage compound loss artificial neural network.

[0101] The final artificial neural network model is used to achieve fast battery device characterization, and its calculation results are synchronized to the experimental characterization database. At the same time, every time a new batch of experimental data comes in, the model can be directly retrained. Because the artificial neural network model is retrainable, after a certain number of experimental characterizations and calibrations, the model can be retrained regularly to ensure that this optoelectronic model follows the development of materials and technologies. Usually, retraining takes less time than initial training, and the optimized hyperparameters can be used directly.

[0102] S3. Combine the thin film wave optics model and the artificial neural network model to form an optoelectronic model.

[0103] The photovoltaic model abandons the finite element solver, which is difficult to calculate and takes a long time, and replaces it with an artificial neural network model based on machine learning, which improves the simulation efficiency of the photovoltaic model and quickly meets the needs of solar energy performance analysis. This new photovoltaic model is essentially an analytical model, and its computing speed is much better than that of one-dimensional / two-dimensional models, while maintaining reliable accuracy. Therefore, when calculating and optimizing large quantities or high-accuracy, the time consumption is greatly reduced, and the requirements for computer hardware are also greatly reduced. The photovoltaic model avoids the use of software such as finite element solvers, and can be stored on each server in the form of analytical solutions, which is convenient and quick to use.

[0104] S4. The data to be tested is input into the photoelectric model to obtain multiple battery performance parameters, and corresponding battery performance analysis is performed according to the battery performance parameters.

[0105] Specifically, the process includes the following:

[0106] S41. Input the measured data into the thin film wave optics model to calculate the optically generated carrier generation rate G:

[0107] G=f(E z )=f optics (l,m,κ,I0,R a )

[0108] Among them, E z is the intensity of the optical electric field, l is the film thickness, m is the real part of the refractive index, κ is the imaginary part of the refractive index, I0 is the incident spectrum, R a is the surface reflectivity;

[0109] S42. The measured data is combined with the carrier generation rate and input into the artificial neural network model to calculate the output parameter Output.

[0110] Output=f NN (G,W,χ,μ,N, ε,A aug ,B rag ,v, τ,V a ,T)

[0111] Where W is the metal work function, χ is the material energy level, μ is the carrier mobility, N is the material energy level density, ε is the material dielectric function, A aug is the Auger recombination constant, B rag is the radiative recombination constant, τ is the carrier lifetime, v is the interface recombination constant, V a is the applied voltage and T is the temperature.

[0112] S43. Obtain battery performance parameters including device performance, maximum power point composite loss, and open voltage composite loss data, and perform corresponding performance analysis based on the performance parameters. The output parameters can be used for parameter sensitivity analysis, composite loss analysis, and battery performance optimization applications. Among them, parameter sensitivity analysis can use three parts of data, namely device performance, maximum power point composite loss, and open voltage composite loss, and use simulation methods such as Monte Carlo to calculate the sensitivity of each output parameter to the final output parameter locally or globally, which is used to find the thin film layer or surface in the perovskite battery that is most likely to affect its performance. Composite loss analysis can use maximum power point composite loss and open voltage composite loss data, and use output parameters to calculate the proportion of each composite loss of the perovskite battery device at the maximum power point and open voltage, so as to determine how to effectively improve the fill factor and open circuit voltage. Battery performance optimization can use device performance data to directly apply output parameters to various large-scale optimization algorithms.

[0113] Embodiment 2:

[0114] This embodiment provides another method for analyzing solar cell performance using a photoelectric model based on a neural network, comprising the following steps:

[0115] S1. Establish thin film wave optics model using matrix analytical model;

[0116] S2. construct and train an artificial neural network model, where the artificial neural network model is a final artificial neural network composed of a device performance artificial neural network, a maximum power point composite loss artificial neural network, and an opening voltage composite loss artificial neural network;

[0117] S3. Combining the thin film wave optics model and the artificial neural network model to form an optoelectronic model;

[0118] S4. Rapidly calibrate the optoelectronic model;

[0119] S5. The data to be tested is input into the photoelectric model to obtain multiple battery performance parameters, and corresponding battery performance analysis is performed according to the battery performance parameters.

[0120] As a preferred solution of this embodiment, after the photoelectric model is constructed, the photoelectric model is calibrated, such as Figure 2 As shown, the specific steps include:

[0121] S401. Obtain short-circuit current, open-circuit voltage, fill factor and conversion efficiency from the experimental voltage-current curve, and confirm material parameters or parameter ranges based on the materials used and the preparation method;

[0122] S402. Calibrate optical parameters based on the short-circuit current and the thin film wave optics model, including the real and imaginary parts of the refractive index of the material at each absorption wavelength, and the thickness of each film layer;

[0123] S403. Calibrate the undetermined parameters based on the open circuit voltage, fill factor, conversion efficiency and the artificial neural network model, and the calibrated data will also be stored in the experimental characterization database. Finally, the calibrated optoelectronic model is obtained.

[0124] This solution can realize the rapid calibration of optical parameters and film thickness, as well as the rapid calibration of material and composite parameters. When calibration is required, the input source of the artificial neural network model is the approximate parameters of the battery material and the current-voltage curve. Therefore, only the current-voltage characterization is required in the laboratory, and no other instruments are required to assist in calibration. The calibration process is convenient and fast. This solution solves the problem that the existing perovskite solar cell model parameter calibration takes a long time, and the simulation efficiency is low and cannot quickly meet actual needs.

[0125] The remaining steps of this embodiment are the same as those of Embodiment 1, and Step S1 specifically includes:

[0126] The thin film wave optics model is constructed using a matrix analytical model. Through a series of numerical simplifications, a matrix composed mainly of 0 and 1 is obtained. It focuses on versatility, program simplicity and computational efficiency, and its solutions are stored in the form of mathematical symbols.

[0127] Step S2 specifically includes:

[0128] S21. Obtain input parameter data, sample and preprocess the input parameter data, and construct model sample data.

[0129] As a preferred solution of this embodiment, S21 specifically includes the following process:

[0130] S211. Obtain parameters from the experimental characterization database and the material parameter database and confirm the parameter range; specifically, confirm the range of each input parameter from the experimental characterization data and the material parameter data, and try to ensure that these ranges cover all existing materials and possible new materials in the future.

[0131] S212. Using Latin hypercube sampling method to sample from the parameters to obtain input parameters;

[0132] S213. Input parameter preprocessing, including difference processing and normalization

[0133] The difference is handled as:

[0134] The maximum and minimum values ​​of the input parameters are compared. If the difference between the maximum and minimum values ​​exceeds the set difference threshold, the decimal logarithm method is used to sample the input parameter to ensure uniform distribution of the data.

[0135] The normalization process is:

[0136] Each input parameter is normalized to evenly distribute the data between -1 and 1.

[0137] S214. Construct sample data according to the preprocessed input parameters, including training data, verification data and test data.

[0138] As a preferred solution of this embodiment, a total of four sets of data are established here, including one set of training data for artificial neural network training, with a sampling number of, for example, 100,000, one set of verification data for verification, with a sampling number of 10,000, and two sets of test data for testing, with a sampling number of 10,000 each.

[0139] S215. Match the energy level requirements of each material / transmission layer of the solar cell to each input parameter. If the match is successful, the input parameter is retained. If the match is unsuccessful, the input parameter is discarded.

[0140] S22. The retained sample data is input into the carrier drift diffusion model to calculate the output parameters, and the sample data is classified according to the output parameter type.

[0141] As a preferred solution of this embodiment, the following steps are specifically included:

[0142] S221. Input the sample data into the carrier drift diffusion model, solve it through the finite element solver, and output the current-voltage curve. In this process, the sample data is firstly imported into the carrier drift diffusion model using the finite element solver, and solved it through the finite element solver, and output the current-voltage curve.

[0143] S222. Analyze the performance parameters of the battery, such as open circuit voltage, fill factor, conversion efficiency, and radiation recombination loss, non-equilibrium carrier recombination loss, Auger recombination loss and interface recombination loss at the maximum power point and open circuit voltage, from the current-voltage curve;

[0144] S223. Classify the input parameters according to the performance parameters of the output parameters, the maximum power point composite loss, and the open circuit voltage composite loss.

[0145] S23. Establish an artificial neural network model, train the artificial neural network model with sample data, and obtain a final artificial neural network model that is a combination of the device performance artificial neural network, the maximum power point composite loss artificial neural network, and the open-voltage composite loss artificial neural network.

[0146] The structure of the artificial neural network model established includes an input layer, three intermediate hidden layers, a fixed hidden layer before the output layer, and an output layer. The training logic is Bayesian regularization, the maximum number of iterations is one thousand, and the maximum number of iterations allowed for the verification error increase is ten.

[0147] Bayesian optimization procedure for hyperparameters: The optimal hyperparameters of the artificial neural network are determined by Bayesian optimization. The hyperparameters that can be optimized include but are not limited to the initial learning rate, minimum gradient, momentum, activation function of each intermediate hidden layer, and number of neurons. The cost function used is the final mean square error of the artificial neural network training. The optimization is terminated by limiting the maximum optimization time, the target minimum value and minimum gradient of the cost function, and the maximum number of iterations.

[0148] The artificial neural network model includes three branch neural networks, which are trained according to the input parameters after the above classification. The final artificial neural network model is composed of three optimized artificial neural networks with the same input and different outputs, which can be used to calculate single, multiple or all output parameters. The three neural network branches are device performance artificial neural network, maximum power point compound loss artificial neural network, and open voltage compound loss artificial neural network.

[0149] The final artificial neural network model is used to achieve fast battery device characterization, and its calculation results are synchronized to the experimental characterization database. At the same time, every time a new batch of experimental data comes in, the model can be directly retrained. Because the artificial neural network model is retrainable, after a certain number of experimental characterizations and calibrations, the model can be retrained regularly to ensure that this optoelectronic model follows the development of materials and technologies. Usually, retraining takes less time than initial training, and the optimized hyperparameters can be used directly.

[0150] S3. Combine the thin film wave optics model and the artificial neural network model to form an optoelectronic model.

[0151] S4. The data to be tested is input into the photoelectric model to obtain multiple battery performance parameters, and corresponding battery performance analysis is performed according to the battery performance parameters.

[0152] Specifically, the process includes the following:

[0153] S41. Input the measured data into the thin film wave optics model to calculate the optically generated carrier generation rate G:

[0154] G=f(E z )=f optics (l,m,κ,I0,R a )

[0155] Among them, E z is the intensity of the optical electric field, l is the film thickness, m is the real part of the refractive index, κ is the imaginary part of the refractive index, I0 is the incident spectrum, R a is the surface reflectivity;

[0156] S42. The measured data is combined with the carrier generation rate and input into the artificial neural network model to calculate the output parameter Output.

[0157] Output=f NN (G,W,χ,μ,N, ε,A aug ,B rag ,v, τ,V a ,T)

[0158] Where W is the metal work function, χ is the material energy level, μ is the carrier mobility, N is the material energy level density, ε is the material dielectric function, A aug is the Auger recombination constant, B rag is the radiative recombination constant, τ is the carrier lifetime, v is the interface recombination constant, V a is the applied voltage and T is the temperature.

[0159] S43. Obtain battery performance parameters including device performance, maximum power point composite loss, and open voltage composite loss data, and perform corresponding performance analysis based on the performance parameters. The output parameters can be used for parameter sensitivity analysis, composite loss analysis, and battery performance optimization applications. Among them, parameter sensitivity analysis can use three parts of data, namely device performance, maximum power point composite loss, and open voltage composite loss, and use simulation methods such as Monte Carlo to calculate the sensitivity of each output parameter to the final output parameter locally or globally, which is used to find the thin film layer or surface in the perovskite battery that is most likely to affect its performance. Composite loss analysis can use maximum power point composite loss and open voltage composite loss data, and use output parameters to calculate the proportion of each composite loss of the perovskite battery device at the maximum power point and open voltage, so as to determine how to effectively improve the fill factor and open circuit voltage. Battery performance optimization can use device performance data to directly apply output parameters to various large-scale optimization algorithms.

[0160] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0161] Although this article uses more terms such as thin film wave optics model, artificial neural network model, optoelectronic model, performance parameter, maximum power point composite loss, open circuit voltage composite loss, etc., it does not exclude the possibility of using other terms. These terms are used only to more conveniently describe and explain the essence of the present invention; interpreting them as any additional limitations is contrary to the spirit of the present invention.

Claims

1. A method for analyzing solar cell performance using a photoelectric model based on a neural network, characterized in that: The following steps are involved: S1. Establish thin film wave optics model using matrix analytical model. X=A\B Where X is the column matrix of the optical electric field intensity to be solved, A is the diagonal matrix composed of wave number and film thickness, and B is the column matrix composed of the light source used as excitation between each film; S2. Establish an artificial neural network model, input sample data classified into three parts: performance parameters, maximum power point composite loss, and open circuit voltage composite loss to train the artificial neural network model, and obtain a final artificial neural network that combines the device performance artificial neural network, the maximum power point composite loss artificial neural network, and the open circuit voltage composite loss artificial neural network; S3. Combining the thin film wave optics model and the artificial neural network model to form an optoelectronic model; S4. The data to be tested is input into the photoelectric model to obtain multiple battery performance parameters, and corresponding battery performance analysis is performed according to the battery performance parameters.

2. The method for analyzing solar cell performance using a photoelectric model based on a neural network according to claim 1, characterized in that The establishment of the artificial neural network model specifically includes the following processes: S21. Obtain input parameter data, sample and preprocess the input parameter data, and construct model sample data; S22. Sample data is input into the carrier drift diffusion model to calculate the output parameters, and the sample data is classified according to the output parameter type; S23. Establish an artificial neural network model, train the artificial neural network model with sample data, and obtain a final artificial neural network model that is a combination of the device performance artificial neural network, the maximum power point composite loss artificial neural network, and the open-voltage composite loss artificial neural network.

3. The method for analyzing solar cell performance using a photoelectric model based on a neural network according to claim 2, characterized in that Step S21 specifically includes: Obtain parameters from the experimental characterization database and the material parameter database and confirm the parameter range; The Latin hypercube sampling method is used to sample the parameters to obtain the input parameters; The input parameter preprocessing includes: Compare the maximum and minimum values ​​of the input parameter. If the difference between the maximum and minimum values ​​exceeds the set difference threshold, the input parameter is sampled using the decimal logarithm method. Normalize the input parameters; Sample data are constructed according to the preprocessed input parameters, including training data, validation data and test data.

4. The method for analyzing solar cell performance using a photoelectric model based on a neural network according to claim 2 or 3, characterized in that Step S21 also includes matching the energy level requirements of each material / transmission layer of the solar cell to each input parameter. If the matching is successful, the input parameter is retained; if the matching is unsuccessful, the input parameter is discarded.

5. The method for analyzing solar cell performance using a photoelectric model based on a neural network according to claim 2, characterized in that The step S22 specifically includes: Input the sample data into the carrier drift diffusion model, solve it through the finite element solver, and output the current-voltage curve; Analyze the current-voltage curve to obtain the battery open circuit voltage, fill factor, conversion efficiency, and radiation recombination loss, non-equilibrium carrier recombination loss, Auger recombination loss, and interface recombination loss at the maximum power point and open circuit voltage. The input parameters are classified into three parts: performance parameters of output parameters, maximum power point compound loss, and open circuit voltage compound loss.

6. The method for analyzing solar cell performance using a photoelectric model based on a neural network according to claim 2, characterized in that The artificial neural network model structure includes an input layer, three intermediate hidden layers, a fixed hidden layer before the output layer, and an output layer. Bayesian regularization is used for training, the maximum number of iterations is one thousand, the maximum number of iterations allowed for the increase in validation error is ten, and the hyperparameters are determined by the Bayesian optimization procedure.

7. The method for analyzing solar cell performance using a photoelectric model based on a neural network according to claim 1, 2 or 3, characterized in that After the photoelectric model is constructed, the photoelectric model is also calibrated, including the following steps: Obtain short-circuit current, open-circuit voltage, fill factor and conversion efficiency from the experimental voltage-current curve, and confirm the material parameters or parameter ranges based on the materials used and the preparation method; According to the short-circuit current, the optical parameters are calibrated in combination with the thin film wave optics model, including the real and imaginary parts of the refractive index of the material at each absorption wavelength, as well as the thickness of each film layer; The undetermined parameters are calibrated based on the open circuit voltage, fill factor, conversion efficiency and artificial neural network model.

8. The method for analyzing solar cell performance using a photoelectric model based on a neural network according to claim 1, 2 or 3, characterized in that The data to be measured is input into the photoelectric model to obtain multiple battery performance parameters, which specifically includes the following process: The measured data is input into the thin film wave optics model to calculate the optically generated carrier generation rate G: G=f(E z )=f optics (l,m,κ,I0,R a ) Among them, E z is the intensity of the optical electric field, l is the film thickness, m is the real part of the refractive index, κ is the imaginary part of the refractive index, I0 is the incident spectrum, R a is the surface reflectivity; The measured data is combined with the carrier generation rate and input into the artificial neural network model to calculate the output parameter Output. Output=f NN (G,W,χ,μ,N, ε,A aug ,B rag ,v, τ,V a ,T) Where W is the metal work function, χ is the material energy level, μ is the carrier mobility, N is the material energy level density, ε is the material dielectric function, and A aug is the Auger recombination constant, B rag is the radiative recombination constant, τ is the carrier lifetime, v is the interface recombination constant, V a is the applied voltage and T is the temperature.

9. The method for analyzing solar cell performance using a photoelectric model based on a neural network according to claim 1, 2 or 3, characterized in that The data to be tested is input into the optoelectronic model to obtain the battery performance parameters including device performance, maximum power point compound loss, and opening voltage compound loss data, and corresponding performance analysis is performed based on the performance parameters.

Images