Solar spectrum simulation method based on BP neural network

Abstract

The invention discloses a solar spectrum simulation method based on a BP neural network, belongs to the technical field of solar spectrum simulation, and aims to solve the problem that no LED solar simulator can simulate AM0G and AM1.5 G solar spectrums at the same time. The method comprises the steps of an LED solar spectrum simulation framework giving consideration to AM0G and AM1.5 G, an NSGA-II-assisted solar spectrum LSTM simulation algorithm, a solar spectrum simulation instance, conclusion analysis and the like. According to the method, the spectrum matching degrees of the AM0G and the AM1.5 G are superior to + / -10.5% and + / -9.3% respectively, the simulation effect of A +-level solar spectrum simulation giving consideration to the AM0G and the AM1.5 G is achieved, and the method has the advantages of being high in practicability and wide in application range. And theoretical basis and technical support are provided for high-precision solar spectrum reconstruction.

Description

Technical Field

[0001] The present invention relates to the technical field of solar spectrum simulation, in particular to a solar spectrum simulation method based on BP neural network. Background Art

[0002] Direct solar radiation is crucial in fields such as aerospace, meteorology and environmental science, as well as solar energy conversion, storage and application. However, direct solar radiation in nature is subject to intermittent weather conditions and is subject to unpredictable weather conditions. The stability of the environment tested using direct solar radiation is difficult to guarantee. Therefore, a solar simulator that can provide a controllable, reliable and repeatable direct solar radiation illuminance and spectrum experimental environment indoors has become a crucial research equipment.

[0003] With the continuous advancement of LED technology, the type of LED is no longer the key factor limiting the spectral performance of LED solar simulators. LED solar simulators have made extremely significant progress and breakthroughs in spectral matching. However, there is currently no LED solar simulator that can achieve simultaneous simulation of AM0G and AM1.5G solar spectra.

[0004] Aiming at the above problems, a solar spectrum simulation method based on BP neural network is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide a solar spectrum simulation method based on BP neural network, and to use this device to work, thereby solving the problem in the above background that no LED solar simulator can achieve simultaneous simulation of AMOG and AM1.5G solar spectra.

[0006] To achieve the above object, the present invention provides the following technical solution: a solar spectrum simulation method based on BP neural network, the specific steps are as follows:

[0007] S1: LED solar spectrum simulation framework taking into account both AM0G and AM1.5G;

[0008] S2: NSGA-II assisted solar spectrum LSTM simulation algorithm;

[0009] S3: Solar spectrum simulation example;

[0010] S4: Conclusion analysis.

[0011] Furthermore, the specific steps of the LED solar spectrum simulation framework taking into account both AM0G and AM1.5G described in S1 are as follows:

[0012] S101: Discrete and Reconstruct;

[0013] S102: Reconstruction quality analysis;

[0014] S103: Simulation algorithm strategy.

[0015] Furthermore, the specific steps of the NSGA-II-assisted solar spectrum LSTM simulation algorithm described in S2 are as follows:

[0016] S201: Solar spectrum simulation training dataset generated based on NSGA-II;

[0017] S202: LSTM-based solar spectrum simulation neural network.

[0018] Furthermore, the specific steps of the solar spectrum simulation example described in S3 are as follows:

[0019] S301: LED type selection and training data set generation;

[0020] S302: Neural network training and iteration;

[0021] S303: Solar spectrum simulation and matching.

[0022] Furthermore, the principles of discretization and reconstruction described in S101 are as follows:

[0023] The uniform discrete sampling frequency of the solar spectrum is set to m, and the sampling wavelength λ solari Composition matrix The corresponding solar spectrum illuminance γ i The matrix ω=[γ1,γ2,…,γ m ], presenting the solar spectrum distribution in matrix form, namely X solar =[Δ T ,ω T ], assuming that there are m single-wavelength spectrum reconstruction units Recorded as Among them, O i Indicates wavelength The corresponding spectral illumination, assuming the spectral reconstruction unit The weight coefficient is The formula followed by the discretized solar spectrum reconstruction can be expressed as:

[0024]

[0025] in, is the adjustment coefficient matrix,

[0026] The spectral distribution of narrowband LED is According to the wavelength matrix Sampling is performed with a peak wavelength of λ i The spectral illuminance of the narrowband LED at discrete sampling positions is β i , βi Can be expressed as Therefore, the spectral illuminance distribution of narrowband LED can be expressed as Spectral reconstruction unit Replaced with a peak wavelength of λ i The narrow-band LED is set to The solar spectrum distribution is reconstructed using narrowband LEDs. The formula is as follows:

[0027]

[0028] Furthermore, the reconstruction quality analysis in S102 is based on the following:

[0029] Spectral curve of narrowband LED It shows the characteristics of normal distribution. In order to improve the matching speed and generalization ability of the algorithm, the narrow-band LED spectrum curve is normalized. The formula is as follows:

[0030]

[0031] where μ i is the peak wavelength, η i is the standard deviation of the spectral curve, the full width at half maximum ω is used to characterize the width of the narrowband LED, and η≈0.425ω;

[0032] The reconstructed spectrum of two adjacent narrowband LEDs is The formula is as follows:

[0033]

[0034] Where μ1 and μ2 represent the peak wavelengths of the two narrowband LEDs, and ω1 and ω2 represent the full width at half maximum of the two narrowband LEDs.

[0035] The spectral reconstruction quality is evaluated by spectral intensity uniformity (SD) and spectral purity (SMSR), which are expressed as follows:

[0036]

[0037] Where N is the number of sampling wavelengths for spectral intensity uniformity, is the mean irradiance corresponding to all sampling wavelengths, is the maximum irradiance of the reconstructed spectral curve, is the sub-peak radiance of the reconstructed spectral curve.

[0038] Furthermore, the specific steps of the simulation algorithm strategy described in S103 are as follows:

[0039] In the stage of generating training data sets using a multi-objective genetic algorithm, a multi-objective genetic algorithm based on the non-dominated sorting genetic algorithm II is used. With a specific spectral curve as the target, the optimal solution for spectrum simulation based on multiple LEDs is generated to form a solar spectrum simulation training data set.

[0040] In the solar spectrum simulation neural network stage, a network architecture in the form of a long short-term memory network is adopted, and the LSTM neural network is trained using the solar spectrum simulation training data set.

[0041] Furthermore, the solar spectrum simulation training based on NSGA-II described in S201 generates a data set, and the specific steps are as follows:

[0042] Step 1: Initialize the population and randomly generate the initial population;

[0043] Step 2: Fitness evaluation, calculate the objective function value of each individual;

[0044] Step 3: Non-dominated sorting and crowding calculation, perform non-dominated sorting on the population and calculate the crowding of each individual;

[0045] Step 4: Selection operation, selecting the next generation population from the parent population and the child population to retain excellent individuals and guide the population to evolve towards the optimal solution;

[0046] Step 5: Crossover and mutation. Perform crossover and mutation operations on the selected population to generate a new offspring population. The crossover operation combines the excellent genes of the parent individuals to produce better offspring. The mutation operation is used to introduce new genetic mutations in the offspring individuals, further increasing the diversity of the population and preventing the algorithm from falling into a local optimum.

[0047] Step 6: Merge the parent and child generations to form a temporary population. Then, select the next generation population through non-dominated sorting and crowding calculation to retain excellent individuals and guide the population to evolve towards the optimal solution.

[0048] Step 7: Iteration, repeat steps 2 to 6 until the preset number of iterations is reached;

[0049] The 6500K standard blackbody spectrum curve S is close to the solar spectrum 6500K (λ), spectral curve shift coefficient b and spectral curve S of narrow-band LED LEDi (λ) as the input variable, and the weight coefficient L i As the output result, the root mean square error (RMSE) of the spectrum simulation was used as the evaluation function to construct a multi-objective linear programming model. The formula is as follows:

[0050]

[0051] Where n is the type of LED light source, and m is the number of wavelength samples in the evaluation process.

[0052] Furthermore, the LSTM-based solar spectrum simulation neural network described in S202 works as follows:

[0053] In the structure of LSTM unit, x t and h t Represent the input vector and output vector of the LSTM unit, respectively, t 、i t and o t Represent the activation values of the forget gate, input gate and output gate respectively, C t and Represent the unit state and its candidate value respectively, the subscript t indicates the time step; σ and tanh represent the sigmoid and tanh activation functions respectively; W and b represent the weight matrix and bias vector respectively. ⊙ represents the multiplication of the corresponding elements of the two matrices;

[0054] The update mechanism of the LSTM unit at each time step t is described by the relevant vectorized equation. The specific calculation process is as follows:

[0055] Step 1: Calculate the forget gate activation value f at time step t t , the formula is as follows:

[0056] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0057] Step 2: Confirm to add to cell state C t The new information, including the activation value of the input gate and the candidate state, is expressed as follows:

[0058] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0059]

[0060] Step 3: Combine the cell state C of the previous time step t-1 , forget gate activation value f t , input gate activation value i t and candidate status Update cell status C t , the formula is as follows:

[0061]

[0062] Step 4: Calculate the output value h of the LSTM unit t , through the activation of the output gate, the cell state C of the current time step is determined t The output flow of information in is expressed as follows:

[0063] o t =σ(X o ·[h t-1 ,x t ]+b o );

[0064] h t =o t ⊙tanh(C t ).

[0065] Furthermore, the principles of LED type selection and training data set generation described in S301 are as follows:

[0066] The selected narrowband LED light source is introduced into an integrating sphere with a radius of 100mm. The coating material of the integrating sphere is polytetrafluoroethylene. Polytetrafluoroethylene has extremely high reflectivity in the ultraviolet-visible-near-infrared band of 0.25μm-2.50μm, allowing the light to undergo multiple diffuse reflections in the integrating sphere to achieve a uniform spectral distribution. The spectrum at the exit of the integrating sphere is then collected using a fiber optic spectrometer.

[0067] The 6500K standard blackbody spectrum curve and the spectra of 29 selected narrowband LEDs were discretized at 1nm intervals within the range of 300nm to 1100nm. The spectrum curve offset coefficient b was set to a random number between 1 and 1.5. 5000 operations were performed to generate a training dataset for the LSTM-based solar spectrum simulation neural network.

[0068] The principles of the neural network training and iteration described in S302 are as follows:

[0069] Using the training data set obtained in S301, AM0G and AM1.5G are used as input to train the LSTM-based solar spectrum simulation neural network;

[0070] The RMSE value of the AM0G model stabilized after 200 iterations, while the RMSE value of the AM1.5G model reached a stable state after 250 iterations. The simulated solar spectrum distributions with significant RMSE fluctuations during the iteration process were selected as representatives for in-depth analysis.

[0071] The solar spectrum simulation matching described in S303 is based on the following principles:

[0072] The spectral matching degree of the solar simulator adopts the evaluation standard of "interval energy percentage", and the specific calculation process follows the following formula:

[0073]

[0074] Where, SPD AM0 and SPD AM1.5 is the spectral matching degree between AM0G and AM1.5G, E SIM (λ) is the simulated solar spectrum illuminance with wavelength λ, E AM0 (λ) and E AM1.5 (λ) is the solar spectrum illuminance at the theoretical wavelengths of AM0G and AM1.5G, and Δλ is the wavelength collection interval.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] 1. This paper proposes a solar spectrum simulation framework for LEDs that takes both AM0G and AM1.5G into account, consisting of solar spectrum discretization and reconstruction, a basis for analyzing reconstruction quality, and a solar spectrum simulation algorithm strategy. The principles of solar spectrum discretization and reconstruction are analyzed, and a discretized solar spectrum reconstruction model is established. Using the normal distribution characteristics of narrowband LED spectral distribution as an example, a spectrum reconstruction quality assessment basis based on SD and SMSR is established. Finally, a solar spectrum simulation algorithm strategy, assisted by NSGA-II, a multi-objective genetic algorithm training set generation and a solar spectrum simulation neural network, is proposed.

[0077] 2. The present invention designs a solar spectrum simulation training data set generation method based on NSGA-II, which uses a 6500K standard blackbody spectrum curve close to the solar spectrum, a spectrum curve offset coefficient, and the spectrum curves of various narrow-band LEDs as input variables, and uses the root mean square error of the spectrum simulation as the evaluation function; by adding a fully connected layer on the basis of the classic LSTM neural network, the solar spectrum simulation neural network of LSTM is completed.

[0078] 3. The spectral matching degrees of AM0G and AM1.5G were better than ±10.5% and ±9.3% respectively, completing the A+ level solar spectrum simulation effect that takes into account both AM0G and AM1.5G, providing a theoretical basis and technical support for high-precision solar spectrum reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 The spectrum reconstruction effect of the present invention Figure 1 ;

[0080] Figure 2 The spectrum reconstruction effect of the present invention Figure 2 ;

[0081] Figure 3 The standard blackbody 6500K, AM0G and AM1.5G solar spectra of the present invention are shown below:

[0082] Figure 4 Schematic diagram of the LSTM unit structure of the present invention;

[0083] Figure 5 The normalized spectral power distribution diagram of the selected 29 narrow-band LEDs of the present invention;

[0084] Figure 6 : This is a graph showing the relationship between RMSE and training times for the simulated AMOG and AM1.5G solar spectra of the present invention;

[0085] Figure 7 A solar spectrum diagram of the number of iterations of the AMOG feature of the present invention;

[0086] Figure 8 The solar spectrum diagram of the AM1.5G characteristic iteration number of the present invention;

[0087] Figure 9 Schematic diagram of solar spectrum matching according to the characteristic iteration number of the present invention;

[0088] Figure 10 It is a schematic diagram of the overall process of the present invention;

[0089] Figure 11 Schematic diagram of the specific steps of the LED solar spectrum simulation framework taking into account both AM0G and AM1.5G of the present invention;

[0090] Figure 12 Schematic diagram of the specific steps of the NSGA-II-assisted solar spectrum LSTM simulation algorithm of the present invention;

[0091] Figure 13 Schematic diagram of the specific steps of the solar spectrum simulation example of the present invention. DETAILED DESCRIPTION

[0092] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0093] In order to solve the technical problem that the current LED solar simulator lacks a spectrum simulation algorithm that can take into account both AM0G and AM1.5G solar spectra, such as Figures 1-13 As shown, the following preferred technical solutions are provided:

[0094] 1. LED solar spectrum simulation framework taking into account both AM0G and AM1.5G:

[0095] 1. Discrete and reconstruction principle:

[0096] First, the uniform discrete sampling frequency of the solar spectrum is set to m, and the sampling wavelength λ solari Composition matrix The corresponding solar spectrum illuminance γ i The matrix ω=[γ1,γ2,…,γ m ], so that the solar spectrum distribution state can be presented in matrix form, that is, X solar =[Δ T ,ω T ]. Then assume that there are m single-wavelength spectrum reconstruction units. Recorded as Among them, O i Indicates wavelength Corresponding spectral illumination, set the spectral reconstruction unit The weight coefficient is The formula followed by the discretized solar spectrum reconstruction can be expressed as:

[0097]

[0098] in, is the adjustment coefficient matrix, Furthermore, the narrowband LED spectral distribution is According to the wavelength matrix For sampling, the peak wavelength is λ i The spectral illuminance of the narrowband LED at the discrete sampling position is set to β i , β i It can be expressed as Therefore, the spectral illuminance distribution of narrowband LED can be expressed as Finally, the spectrum reconstruction unit Replaced with a peak wavelength of λ i The narrow-band LED is set to The reconstruction of the solar spectrum distribution using these narrowband LEDs can be expressed as:

[0099]

[0100] According to formula (2), the weight coefficient of the narrowband LED is precisely adjusted through the algorithm to achieve high-precision matching between the reconstructed spectrum and the solar spectrum.

[0101] 2.Reconstruction quality analysis basis:

[0102] While narrowband LED technology can achieve high-precision reproduction of the solar spectrum from 300nm to 1100nm, the peak wavelength distribution of narrowband LEDs is uneven. Therefore, by analyzing the quality of the reconstructed spectra of two adjacent narrowband LEDs, the selection of narrowband LEDs was optimized.

[0103] Typically, the spectral curve of a narrowband LED It shows the characteristics of normal distribution. In order to improve the matching speed and generalization ability of the algorithm, the narrow-band LED spectrum curve needs to be normalized, which can be expressed as:

[0104]

[0105] where μ i is the peak wavelength, η i is the standard deviation of the spectral curve. The full width at half maximum (FWHM) ω is used to characterize the width of the narrowband LED, and η ≈ 0.425ω.

[0106] Reconstructed spectrum of two adjacent narrowband LEDs for:

[0107]

[0108] Where μ1 and μ2 represent the peak wavelengths of the two narrowband LEDs, and ω1 and ω2 represent the full width at half maximum of the two narrowband LEDs.

[0109] To achieve high-quality solar spectrum reconstruction, the number of narrowband LEDs used should be minimized to cover the required spectral band without generating spectral gaps. Therefore, the present invention comprehensively evaluates the quality of spectral reconstruction using two indicators: spectral intensity uniformity (SD) and spectral purity (SMSR). The formula is:

[0110]

[0111] Where N is the number of sampling wavelengths for spectral intensity uniformity, is the mean irradiance corresponding to all sampling wavelengths, is the maximum irradiance of the reconstructed spectral curve, is the sub-peak radiance of the reconstructed spectral curve.

[0112] The full width at half maximum of (two adjacent) narrowband LEDs is usually in the range of 10nm to 40nm, that is, the range of ω1 and ω2 is 10nm to 40nm, and in order to ensure that the reconstructed spectrum avoids discontinuity, the peak interval Δλ ranges from 0nm to (ω1+ω2)nm. The spectrum reconstruction effect is as follows Figure 1-Figure 2 shown.

[0113] The SD after spectral reconstruction, such as Figure 1 As shown in (a), (b), (c) and (d), ω1 is the data generated when 10nm, 20nm, 30nm and 40nm respectively, and ω2 is in the range of 10nm to 40nm. SD value of ; SMSR after spectral reconstruction, such as Figure 2 As shown in (e), (f), (g) and (h), ω1 is the data generated when 10nm, 20nm, 30nm and 40nm respectively, and ω2 is in the range of 10nm to 40nm.

[0114] SMSR value.

[0115] The SD index measures the uniformity of the reconstructed spectrum. Lower values indicate greater uniformity, meaning smaller intensity variations between wavelengths. The SMSR index measures the purity of the reconstructed spectrum. Higher values indicate fewer spurious peaks beyond the main peak, meaning higher spectral purity. Therefore, by comprehensively considering both SD and SMSR, it is possible to optimize the selection of narrowband LED light sources, achieving higher-quality solar spectrum reconstruction.

[0116] 3. Simulation algorithm strategy:

[0117] Following the standard's provisions for spectral irradiance distribution, the wavelength range of the AM0G solar spectrum is defined as 300nm to 1100nm, while the wavelength range of the AM1.5G solar spectrum is 400nm to 1100nm. The wavelength ranges of these two solar spectra are different, and compared with the smooth spectral curve of standard blackbody radiation (such as 6500K), both exhibit complex spectral detail characteristics and have obvious differences in details. The distribution of the standard blackbody 6500K, AM0G and AM1.5G solar spectrum curves is shown in the figure below. Figure 3 shown.

[0118] Since there are few publicly available training datasets for spectral curve simulation, in order to give full play to the advantages of various LED spectrum reconstruction methods, achieve efficient and balanced simulation of AMOG and AM1.5G solar spectra, and enhance the robustness and universality of the system, this paper proposes a NSGA-II-assisted solar spectrum LSTM simulation algorithm strategy consisting of two stages: generating a training dataset using a multi-objective genetic algorithm and a solar spectrum simulation neural network.

[0119] In the stage of generating training data sets by multi-objective genetic algorithm, a multi-objective genetic algorithm based on non-dominated sorting genetic algorithm II (NSGA-II) is adopted. With a specific spectral curve as the target, a series of diversified optimal solutions of spectral simulation based on multiple LEDs are generated to form a solar spectrum simulation training data set.

[0120] The solar spectrum simulation neural network stage adopts a long short-term memory (LSTM) network architecture, and uses a solar spectrum simulation training dataset to train the LSTM neural network, aiming to improve the generalization and efficiency of the LSTM-based solar spectrum simulation neural network.

[0121] 2. NSGA-II assisted solar spectrum LSTM simulation algorithm:

[0122] 1. Method for generating datasets based on NSGA-II solar spectrum simulation training:

[0123] NSGA-II is an efficient multi-objective genetic algorithm designed based on the concept of genetic algorithms. It maintains the diversity and uniform distribution of dominant solutions through non-dominated sorting and crowding calculation. It has excellent fitness distribution and is highly advantageous in maintaining population diversity and avoiding the loss of excellent solutions. The basic process of NSGA-II is as follows:

[0124] Step 1: Initialize the population and randomly generate the initial population;

[0125] Step 2: Fitness evaluation, calculate the objective function value of each individual;

[0126] Step 3: Non-dominated sorting and crowding calculation, perform non-dominated sorting on the population and calculate the crowding of each individual;

[0127] Step 4: Selection operation, selecting the next generation population from the parent population and the child population to retain excellent individuals and guide the population to evolve towards the optimal solution;

[0128] Step 5: Crossover and mutation. Perform crossover and mutation operations on the selected population to generate a new offspring population. The crossover operation combines the excellent genes of the parent individuals to produce better offspring. The mutation operation is used to introduce new genetic mutations in the offspring individuals, further increasing the diversity of the population and preventing the algorithm from falling into a local optimum.

[0129] Step 6: Merge the parent and child generations to form a temporary population. Then, select the next generation population through non-dominated sorting and crowding calculation to retain excellent individuals and guide the population to evolve towards the optimal solution.

[0130] Step 7: Iteration, repeat steps 2 to 6 until the preset number of iterations is reached.

[0131] In the present invention, the 6500K standard black body spectrum curve S close to the solar spectrum is used. 6500K (λ), spectral curve shift coefficient b and spectral curve of narrowband LED As input variables, and the weight coefficient L i As the output result, the root mean square error (RMSE) of the spectrum simulation is used as the evaluation function to construct a multi-objective linear programming model, which is expressed as:

[0132]

[0133] Where n is the type of LED light source, and m is the number of wavelength samples in the evaluation process.

[0134] According to the principle of multi-objective linear programming, by increasing the number of iterative operations of the multi-objective genetic algorithm and adjusting the spectral curve shift coefficient b, the scale of the training data set can be effectively controlled (the number of training data sets can be effectively controlled) while ensuring the diversity of the data set.

[0135] 2. LSTM-based solar spectrum simulation neural network:

[0136] Long Short-Term Memory (LSTM) is an advanced recurrent neural network. LSTM adds memory storage units and gate structures to traditional recurrent neural networks. The unit state is responsible for storing the state information of neurons, and the gate structure selectively transmits information through neural layers and element-by-element multiplication mechanisms. LSTM uses three gate structures to protect and control the flow of information: the forget gate is responsible for determining the information discarded in the unit, the input gate is responsible for determining the new information in the unit, and the output gate is responsible for filtering the final output information. LSTM can express long-term dependency information in input features through the gate structure and effectively alleviate the problems of gradient vanishing and gradient exploding. The specific structure of the LSTM unit is as follows: Figure 4 shown.

[0137] In the structure of LSTM unit, x t and h t Represent the input vector and output vector of the LSTM unit, respectively, t 、i t and o t Represent the activation values of the forget gate, input gate and output gate respectively, C t and where σ and tanh represent the cell state and its candidate value, respectively. The subscript t indicates the time step. σ and tanh represent the sigmoid and tanh activation functions, respectively. W and b represent the weight matrix and bias vector, respectively. ⊙ represents the multiplication of corresponding elements of two matrices.

[0138] The update mechanism of the LSTM unit at each time step t is described by the relevant vectorized equation. The specific calculation process is as follows:

[0139] Step 1: Calculate the forget gate activation value f at time step t t , the formula is as follows: (8);

[0141] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0142] Step 2: Confirm to add to cell state C t The new information, including the activation value of the input gate and the candidate state, is expressed as follows: (9);

[0144] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0145]

[0146] Step 3: Combine the cell state C of the previous time step t-1 , forget gate activation value f t , input gate activation value i t and candidate status Update cell status C t , the formula is as follows:

[0147]

[0148] Step 4: Calculate the output value h of the LSTM unit t , through the activation of the output gate, the cell state C of the current time step is determined t The output flow of information in is expressed as follows:

[0149] o t =σ(X o ·[h t-1 ,x t ]+b o )(12);

[0150] h t =o t ⊙tanh(C t )(13);

[0151] Based on the classic LSTM neural network, a LSTM-based solar spectrum simulation neural network is constructed by adding a fully connected layer to ensure that the number of nodes in the fully connected layer is consistent with the number of LED types in the training dataset.

[0152] On this basis, the LSTM-based solar spectrum simulation neural network is trained using a training data set based on the NSGA-II multi-objective genetic algorithm, and a solar spectrum simulation algorithm that can take into account both AM0G and AM1.5G using a set of multiple LEDs can be obtained.

[0153] 3. Solar spectrum simulation example:

[0154] 1. LED type selection and training data set generation:

[0155] In order to ensure the accuracy of solar spectrum reconstruction, according to the theoretical model described by equations (5) and (6), and combined with the existing narrow-band LED resources, this paper selects 29 narrow-band LEDs as the basic light source. The normalized spectral power distribution (NSPD) of these narrow-band LEDs is as follows: Figure 5 The selected narrowband LED light source is introduced into an integrating sphere with a radius of 100 mm. The sphere is coated with polytetrafluoroethylene (PTFE), which has extremely high reflectivity in the UV-visible-near-infrared band (0.25 μm-2.50 μm). This ensures that the light undergoes multiple diffuse reflections within the integrating sphere, achieving a uniform spectral distribution. The spectrum at the exit of the integrating sphere is collected using an AvaSpec-ULS2048XL-EVO fiber optic spectrometer.

[0156] Since the simulated spectral range is required to cover 300nm to 1100nm, the 6500K standard blackbody spectral curve and the spectra of the selected 29 narrow-band LEDs are discretized at intervals of 1nm from 300nm to 1100nm. The spectral curve offset coefficient b is set to a random number between 1 and 1.5. Through 5000 operations, a training data set for the LSTM-based solar spectrum simulation neural network is generated.

[0157] 2. Neural network training and iterative simulation effects:

[0158] Since the present invention selects 29 kinds of narrow-band LEDs, the number of network nodes in the fully connected layer of the constructed LSTM-based solar spectrum simulation neural network is 29. Using the above training data set, AMOG and AM1.5G are used as input to train the LSTM-based solar spectrum simulation neural network. The relationship between the RMSE of the AMOG and AM1.5G solar spectra and the number of training times is shown as follows: Figure 6 shown.

[0159] During the iteration process, the RMSE value of the AMOG model stabilized after 200 iterations, while the RMSE value of the AM1.5G model reached a stable state after 250 iterations. In order to further analyze the simulation performance of the solar spectrum during the iteration process, the present invention selected the simulated solar spectrum distribution with significant fluctuations in the RMSE value during the iteration process as a representative, and the corresponding number of iterations is shown in Table 1:

[0160] Table 1 Feature iteration times for AMOG and AM1.5G

[0161] Spectral type Feature iterations AMOG 5、18、200 AM1.5G 5、20、250

[0162] According to the weight coefficients of 29 narrow-band LEDs in the iterative process, the light source is adjusted to achieve an irradiance of 1 sun constant at the exit of the integrating sphere. The spectrum is collected by a spectrometer to obtain the simulated distribution of the AMOG and AM1.5G solar spectra, as shown in the following figure: Figure 7 and Figure 8 shown.

[0163] 3. According to the IEC 60904-9:2020 international standard, the spectral matching degree of the solar simulator adopts the evaluation standard of "interval energy percentage". The specific calculation process follows formulas (12) and (13), and the formula is as follows:

[0164]

[0165] Where, SPD AM0 and SPD AM1.5 is the spectral matching degree between AMOG and AM1.5G, E SIM (λ) is the simulated solar spectrum illuminance with wavelength λ, E AM0 (λ) and E AM1.5 (λ) is the solar spectrum illuminance at the theoretical wavelength of AMOG and AM1.5G, and Δλ is the wavelength collection interval.

[0166] Figure 7 and Figure 8 According to formulas 14 and 15, the spectrum matching error and the solar spectrum matching of the characteristic iteration times of AM0G and AM1.5G are as follows: Figure 9The solar spectrum matching results for each iteration are shown in Tables 2 and 3, with (a) representing AM0G and (b) representing AM1.5G. The results of the spectrum matching to the A-level standard (error bar ±25%) or the A+-level standard (error bar ±12.5%) are listed in Tables 2 and 3, respectively.

[0167] Table 2 AM0G feature iteration times and spectrum matching

[0168]

[0169] Table 3 AM1.5G feature spectral matching results of iteration times

[0170]

[0171] It can be seen from Tables 2 and 3 that in the early stage of iteration (5 iterations for both AM0G and AM1.5G), the spectrum simulations of both AM0G and AM1.5G did not meet the A-level standard. In the middle stage of iteration (18 iterations for AM0G and 20 iterations for AM1.5G), the AM0G solar spectrum met the A-level standard except for the range of 900nm-1100nm. The rest met the A+ level standard. The spectrum simulations of AM1.5G all met the A+ level standard, but did not meet the A+ level standard in the range of 400nm-500nm and the range of 900nm-1100nm. The tolerance limit for the A+ standard is approaching. In the later stages of iteration (200 iterations for AM0G and 250 iterations for AM1.5G), all bands of both AM0G and AM1.5G met the A+ standard. For AM0G, the spectral simulation deviation in the 300nm-900nm band was within ±4.5%, while the simulation deviation in the 900nm-1100nm band was -10.5%. For AM1.5G, the simulation deviation in the 900nm-1100nm band was better than ±3.6%, with the exception of the -9.3% deviation in the 900nm-1100nm band. This larger deviation is due to the fact that among the 29 narrow-band LED types, the number of narrow-band LEDs in the 900nm-1100nm wavelength range is relatively small. Although this has met the A+ standard, it still lags significantly behind other bands. Therefore, with the continuous advancement of LED technology, spectral matching will be further improved by expanding the number of narrow-band LEDs in the 900nm-1100nm band.

[0172] 4. Conclusion and Analysis:

[0173] During the AM0G spectrum simulation process, when the number of iterations reached 200, the RMSE tended to stabilize. At this time, the spectral simulation deviations of all bands from 300nm to 900nm were less than ±4.5%, and the spectral simulation deviation from 900nm to 1100nm was -10.5%. The spectral matching of all bands reached A+ level; during the AM1.5G spectrum simulation process, when the number of iterations reached 250, the RMSE difference tended to stabilize, and the spectral matching of all bands reached A+ level. At this time, except for the light level simulation deviation of 900-1100nm at -9.3%, the spectral simulation deviations of the remaining bands were all better than ±3.6%.

[0174] Since the 900-1100nm wavelength is relatively rare among the 29 narrow-band LEDs used, the spectral simulation error of AM0G and AM1.5G in the 900nm-1100nm range can still reach A+ level. However, it is still larger than that in other bands. In the future, the number of narrow-band LEDs in this band can be increased to achieve better spectral simulation effects.

[0175] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0176] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A solar spectrum simulation method based on BP neural network, characterized in that: The specific steps are as follows: S1: LED solar spectrum simulation framework taking into account both AM0G and AM1.5G; S2: NSGA-II assisted solar spectrum LSTM simulation algorithm; S3: Solar spectrum simulation example; S4: Conclusion analysis.

2. The solar spectrum simulation method based on BP neural network according to claim 1, characterized in that: The specific steps of the LED solar spectrum simulation framework that takes into account both AM0G and AM1.5G as described in S1 are as follows: S101: Discrete and Reconstruct; S102: Reconstruction quality analysis; S103: Simulate algorithm strategy.

3. The solar spectrum simulation method based on BP neural network according to claim 2, characterized in that: The specific steps of the NSGA-II-assisted solar spectrum LSTM simulation algorithm described in S2 are as follows: S201: Solar spectrum simulation training dataset generated based on NSGA-II; S202: LSTM-based solar spectrum simulation neural network.

4. The solar spectrum simulation method based on BP neural network according to claim 3, characterized in that: The specific steps of the solar spectrum simulation example described in S3 are as follows: S301: LED type selection and training data set generation; S302: Neural network training and iteration; S303: Solar spectrum simulation and matching.

5. The solar spectrum simulation method based on BP neural network according to claim 4, characterized in that: The principles of discretization and reconstruction described in S101 are as follows: The uniform discrete sampling frequency of the solar spectrum is set to m, and the sampling wavelength λ solari Composition matrix The corresponding solar spectrum illuminance γ i The matrix ω=[γ1,γ2,…,γ m ], presenting the solar spectrum distribution in matrix form, namely X solar =[Δ T ,ω T ], assuming that there are m single-wavelength spectrum reconstruction units Recorded as Among them, O i Indicates wavelength The corresponding spectral illumination, assuming the spectral reconstruction unit The weight coefficient is The formula followed by the discretized solar spectrum reconstruction can be expressed as: in, is the adjustment coefficient matrix, The spectral distribution of narrowband LED is According to the wavelength matrix Sampling is performed with a peak wavelength of λ i The spectral illuminance of the narrowband LED at discrete sampling positions is β i , β i Can be expressed as Therefore, the spectral illuminance distribution of narrowband LED can be expressed as Spectral reconstruction unit Replaced with a peak wavelength of λ i The narrow-band LED is set to The solar spectrum distribution is reconstructed using narrowband LEDs. The formula is as follows:

6. The solar spectrum simulation method based on BP neural network according to claim 5, characterized in that: The basis for the reconstruction quality analysis described in S102 is as follows: Spectral curve of narrowband LED It shows the characteristics of normal distribution. In order to improve the matching speed and generalization ability of the algorithm, the narrow-band LED spectrum curve is normalized. The formula is as follows: where μ i is the peak wavelength, η i is the standard deviation of the spectral curve, the full width at half maximum ω is used to characterize the width of the narrowband LED, and η≈0.425ω; The reconstructed spectrum of two adjacent narrowband LEDs is The formula is as follows: Where μ1 and μ2 represent the peak wavelengths of the two narrowband LEDs, and ω1 and ω2 represent the full width at half maximum of the two narrowband LEDs. The spectral reconstruction quality is evaluated by spectral intensity uniformity (SD) and spectral purity (SMSR), which are expressed as follows: Where N is the number of sampling wavelengths for spectral intensity uniformity, is the mean irradiance corresponding to all sampling wavelengths, is the maximum irradiance of the reconstructed spectral curve, is the sub-peak radiance of the reconstructed spectral curve.

7. The solar spectrum simulation method based on BP neural network according to claim 6, characterized in that: The specific steps of the simulation algorithm strategy described in S103 are as follows: In the stage of generating training data sets using a multi-objective genetic algorithm, a multi-objective genetic algorithm based on the non-dominated sorting genetic algorithm II is used. With a specific spectral curve as the target, the optimal solution for spectrum simulation based on multiple LEDs is generated to form a solar spectrum simulation training data set. In the solar spectrum simulation neural network stage, a network architecture in the form of a long short-term memory network is adopted, and the LSTM neural network is trained using the solar spectrum simulation training data set.

8. The solar spectrum simulation method based on BP neural network according to claim 7, characterized in that: The solar spectrum simulation training based on NSGA-II described in S201 generates a data set, and the specific steps are as follows: Step 1: Initialize the population and randomly generate the initial population; Step 2: Fitness evaluation, calculate the objective function value of each individual; Step 3: Non-dominated sorting and crowding calculation, perform non-dominated sorting on the population and calculate the crowding of each individual; Step 4: Selection operation, selecting the next generation population from the parent population and the child population to retain excellent individuals and guide the population to evolve towards the optimal solution; Step 5: Crossover and mutation. Perform crossover and mutation operations on the selected population to generate a new offspring population. The crossover operation combines the excellent genes of the parent individuals to produce better offspring. The mutation operation is used to introduce new genetic mutations in the offspring individuals, further increasing the diversity of the population and preventing the algorithm from falling into a local optimum. Step 6: Merge the parent and child generations to form a temporary population. Then, select the next generation population through non-dominated sorting and crowding calculation to retain excellent individuals and guide the population to evolve towards the optimal solution. Step 7: Iteration, repeat steps 2 to 6 until the preset number of iterations is reached; The 6500K standard blackbody spectrum curve S is close to the solar spectrum 6500K (λ), spectral curve shift coefficient b and spectral curve of narrowband LED As input variables, and the weight coefficient L i As the output result, the root mean square error (RMSE) of the spectrum simulation was used as the evaluation function to construct a multi-objective linear programming model. The formula is as follows: Where n is the type of LED light source, and m is the number of wavelength samples in the evaluation process.

9. The solar spectrum simulation method based on BP neural network according to claim 8, characterized in that: The LSTM-based solar spectrum simulation neural network described in S202 works as follows: In the structure of LSTM unit, x t and h t Represent the input vector and output vector of the LSTM unit, respectively, t 、i t and o t Represent the activation values of the forget gate, input gate and output gate respectively, C t and Represent the unit state and its candidate value respectively, the subscript t indicates the time step; σ and tanh represent the sigmoid and tanh activation functions respectively; W and b represent the weight matrix and bias vector respectively. ⊙ represents the multiplication of the corresponding elements of the two matrices; The update mechanism of the LSTM unit at each time step t is described by the relevant vectorized equation. The specific calculation process is as follows: Step 1: Calculate the forget gate activation value f at time step t t , the formula is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ); Step 2: Confirm to add to cell state C t The new information, including the activation value of the input gate and the candidate state, is expressed as follows: i t =σ(W i ·[h t-1 ,x t ]+b i ); Step 3: Combine the cell state C of the previous time step t-1 , forget gate activation value f t , input gate activation value i t and candidate status Update cell status C t , the formula is as follows: Step 4: Calculate the output value h of the LSTM unit t , through the activation of the output gate, the cell state C of the current time step is determined t The output flow of information in is expressed as follows: the t =σ(X o ·[h t-1 ,x t ]+b o ); h t =o t ⊙tanh(C t )。 10. The solar spectrum simulation method based on BP neural network according to claim 9, characterized in that: The principles of LED type selection and training data set generation described in S301 are as follows: The selected narrowband LED light source is introduced into an integrating sphere with a radius of 100mm. The coating material of the integrating sphere is polytetrafluoroethylene. Polytetrafluoroethylene has extremely high reflectivity in the ultraviolet-visible-near-infrared band of 0.25μm-2.50μm, allowing the light to undergo multiple diffuse reflections in the integrating sphere to achieve a uniform spectral distribution. The spectrum at the exit of the integrating sphere is then collected using a fiber optic spectrometer. The 6500K standard blackbody spectrum curve and the spectra of 29 selected narrowband LEDs were discretized at 1nm intervals within the range of 300nm to 1100nm. The spectrum curve offset coefficient b was set to a random number between 1 and 1.

5. 5000 operations were performed to generate a training dataset for the LSTM-based solar spectrum simulation neural network. The principles of the neural network training and iteration described in S302 are as follows: Using the training data set obtained in S301, AM0G and AM1.5G are used as input to train the LSTM-based solar spectrum simulation neural network; The RMSE value of the AM0G model stabilized after 200 iterations, while the RMSE value of the AM1.5G model reached a stable state after 250 iterations. The simulated solar spectrum distributions with significant RMSE fluctuations during the iteration process were selected as representatives for in-depth analysis. The solar spectrum simulation matching described in S303 is based on the following principles: The spectral matching degree of the solar simulator adopts the evaluation standard of "interval energy percentage". The specific calculation process follows the following formula: Where, SPD AM0 and SPD AM1.5 is the spectral matching degree between AM0G and AM1.5G, E SIM (λ) is the simulated solar spectrum illuminance with wavelength λ, E AM0 (λ) and E AM1.5 (λ) is the solar spectrum illuminance at the theoretical wavelengths of AM0G and AM1.5G, and Δλ is the wavelength collection interval.

Images