A method for evaluating the dust accumulation degree of photovoltaic arrays based on GS-SVM

Abstract

The present invention discloses a method for assessing the dust accumulation degree of a photovoltaic array based on GS-SVM. The method comprises the following steps: first, the short-circuit current of the photovoltaic array under different operating conditions is collected. Then, the short-circuit current under different operating conditions, the corresponding light intensity and ambient temperature, and the corresponding dust accumulation density are used as labels to construct a training data set. Then, hyperparameters to be optimized for a support vector machine (SVM) are set, and the training data set is updated. The GS-SVM model is then trained with the updated data set to obtain a GS-SVM model with the highest classification accuracy. Finally, the GS-SVM model is used to perform real-time assessment of the dust accumulation degree of the photovoltaic array.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation, and more specifically, relates to a method for evaluating the dust accumulation degree of a photovoltaic array based on GS-SVM. Background Art

[0002] Photovoltaic power generation has seen rapid growth in the clean energy sector in recent years. By the end of 2022, China's cumulative installed photovoltaic capacity reached 87.41 GW. Analysis of photovoltaic power generation operations and maintenance in recent years reveals that photovoltaic arrays are susceptible to natural factors such as windblown sand, dust, dirt, rain, and fog. Dust accumulation on array surfaces can be a serious problem, potentially reducing power generation efficiency by 20%-40%. Therefore, detecting and assessing dust accumulation has become an essential and challenging task in the operation and maintenance of photovoltaic power plants.

[0003] Current research on photovoltaic array dust accumulation assessment focuses on dust accumulation model detection methods that combine theoretical and field testing. These methods typically use engineering models of the photovoltaic array as a reference to analyze and assess actual conditions. This requires a precise engineering model, but due to array characteristics and regional variations, modeling is challenging and difficult to achieve widespread applicability. In recent years, a number of computer vision-based methods have emerged, leveraging the differences in color and texture features of array surface images under varying dust accumulation conditions to identify dust accumulation levels. However, this incurs additional equipment costs and the image data is highly noisy. Furthermore, methods using machine learning to train models and identify dust accumulation levels based on collected dust accumulation assessment features, such as light intensity, temperature, voltage, current, and power, are gradually gaining momentum. However, when dust-accumulated arrays exhibit other common faults, such as local shorts, local opens, global aging, and local shadowing, the accuracy of these machine learning methods is significantly reduced, making them difficult to apply in engineering research. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a photovoltaic array dust accumulation degree assessment method based on GS-SVM, which can solve the problems of the current photovoltaic array dust accumulation degree assessment model based on machine learning algorithm collecting too many feature quantities and the assessment accuracy being easily disturbed by other faults.

[0005] To achieve the above-mentioned object of the invention, the present invention provides a method for evaluating the dust accumulation degree of a photovoltaic array based on GS-SVM, which is characterized by comprising the following steps:

[0006] (1) Collect the short-circuit current of the photovoltaic array under different conditions;

[0007] (1.1) Set the dust density interval Δd, divide the dust density into levels, and set the level number k, k = 1, 2, ..., K, K is the number of dust density levels;

[0008] (1.2) Collect the photovoltaic array at different light intensities G i , different ambient temperatures T j And different dust accumulation density ρ k The short-circuit current under the i-th light intensity, j-th ambient temperature and k-th dust density is recorded as I i,j,k , i=1,2,…,n,j=1,2,…,m,n is the level of light intensity, m is the level of ambient temperature;

[0009] (2) Constructing a training dataset;

[0010] In each set of collected data, the short-circuit current I i,j,k The corresponding light intensity G i and ambient temperature T j Composition of training data x d =(G i ,T i ,I i,j,k ), the corresponding dust accumulation density ρ k As label y d , thus constructing the training data set {(x1,y1),(x2,y2),…,(x d ,y d ),…(x D ,y D )};

[0011] (3) Set the hyperparameters of the support vector machine (SVM) to be optimized;

[0012] (3.1) Select N kernel functions, set the value of the kernel parameter g of each kernel function, and the value of the penalty factor C of each kernel function;

[0013] (3.2), introduce the lth kernel function There is a mapping from input space to feature space Find the maximum-margin separating hyperplane in a high-dimensional feature space:

[0014]

[0015] Where w, b are hyperparameters, d1≠d2, d1, d2∈[1,D];

[0016] (3.2) Using the Lagrangian function and strong dual transformation, by eliminating w and b, the above equation is transformed into the following target optimization problem:

[0017]

[0018] in, is the Lagrangian operator under the corresponding dust accumulation degree, C is the penalty factor of the kernel function;

[0019] (4) Update the training data set;

[0020] (4.1) Substitute the training data (x1, x2) and the corresponding labels into the formula in step (3.2), calculate the target value minL, and then replace the kernel function Finally, we get N groups of target values ​​minL, and then we sum up the N groups of target values ​​minL and replace them with x1 in the training dataset. The updated x1 is recorded as y1 remains unchanged;

[0021] (4.2) Process the labels corresponding to the training data (x2, x3) according to step (4.1) to obtain the updated y2 remains unchanged; and so on until the last set of training data The corresponding tag continues to be processed according to step (4.1) to obtain the updated y D Save unchanged;

[0022] (5) Construct GS-SVM model;

[0023] The GS-SVM model includes a three-dimensional coordinate system grid GS and a support vector machine SVM, wherein the three-dimensional coordinate system grid GS model is (K, C, g), K represents the introduced kernel function, C is the penalty factor of the kernel function, and g is the kernel parameter of the kernel function;

[0024] (6) Training the GS-SVM model;

[0025] (6.1), the updated training data set is divided into M training subsets, where the κth training subset is denoted as σ κ ,κ=1,2,…,M;

[0026] (6.2) Initialize κ = 1, input the κth training subset into the GS-SVM model, and then select a set of hyperparameters (K, C, g) according to the three-dimensional coordinate system grid GS and substitute them into the SVM model. The SVM model processes each set of training data in the κth training subset and obtains the corresponding prediction label;

[0027] (6.3) According to the predicted label and the true label of each set of training data, the classification accuracy of each set of training data is calculated. If the predicted label is the same as the true label, the classification accuracy is recorded as 1, otherwise it is recorded as 0;

[0028] (6.4) Calculate the average classification accuracy of all training data in the κth training subset, denoted as p κ ;

[0029] (6.5) Determine whether the current number of iterations κ has reached the maximum value M. If κ < M, proceed to step (6.6); otherwise, proceed to step (6.7);

[0030] (6.6), let κ = κ + 1, and then return to step (6.2);

[0031] (6.7) Find the maximum value max(p κ ), and then find max(p κ ) The set of hyperparameters (K, C, g) is the optimal hyperparameter, and the trained GS-SVM model is obtained;

[0032] (7) Real-time assessment of the degree of dust accumulation in photovoltaic arrays;

[0033] The short-circuit current I of the PV array is collected in real time when the light intensity G and ambient temperature T are unknown but the dust accumulation density is unknown. The light intensity G, ambient temperature T, and short-circuit current I are then combined into test data x = (G, T, I). The test data x = (G, T, I) is then input into the trained GS-SVM model to output the predicted label value and obtain the dust accumulation level of the PV array.

[0034] The object of the invention of the present invention is achieved like this:

[0035] The present invention is a photovoltaic array dust accumulation degree assessment method based on GS-SVM. First, the short-circuit current of the photovoltaic array under different working conditions is collected. Then, the short-circuit current under different working conditions and the corresponding light intensity and ambient temperature, as well as the corresponding dust accumulation density are used as labels to construct a training data set. Then, the hyperparameters to be optimized of the support vector machine (SVM) are set, and the training data set is updated. The GS-SVM model is then trained with the updated data set to obtain the GS-SVM model with the highest classification accuracy. Finally, the GS-SVM model is used to perform real-time assessment of the dust accumulation degree of the photovoltaic array.

[0036] At the same time, the photovoltaic array dust accumulation degree assessment method based on GS-SVM of the present invention also has the following beneficial effects:

[0037] (1) The present invention uses the GS algorithm to search and optimize the hyperparameters in SVM, which can avoid problems such as overfitting, long training time and poor training results caused by inappropriate hyperparameter settings;

[0038] (2) The present invention simplifies the number of characteristic quantities for evaluating the dust accumulation degree of photovoltaic arrays, and the amount of input data is smaller than that of traditional recognition models, thereby reducing the cost and threshold of data collection;

[0039] (3) The photovoltaic array dust accumulation assessment method proposed in the present invention can resist the interference of faults such as short circuit, open circuit, aging and partial shading on the dust accumulation assessment, ensuring that the algorithm can obtain high-accuracy dust accumulation prediction results under the relatively complex operation and maintenance conditions of photovoltaic modules;

[0040] (4) The present invention analyzes the output characteristics of photovoltaic array dust accumulation and other faults, and reveals that the electrical parameter of short-circuit current can reflect the dust accumulation of photovoltaic array and is not easily disturbed by other faults; then the three parameters of short-circuit current, light intensity and temperature are used as input feature quantities and input into the SVM dust accumulation degree evaluation model containing GS hyperparameter optimization technology. After experimental test and analysis, it is proved that the accuracy of this method is higher than that of DecisionTree, GS-DecisionTree and XGBoost evaluation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of the photovoltaic array dust accumulation degree assessment method based on GS-SVM of the present invention;

[0042] Figure 2 It is a test model for the operating conditions of a dust-accumulated photovoltaic array;

[0043] Figure 3 It is the volt-ampere characteristic curve of the photovoltaic array with different dust accumulation degrees;

[0044] Figure 4 It is the evaluation accuracy of different algorithms on the test set when the training set data accounts for different proportions of the total data set;

[0045] Figure 5 This is the classification effect of GS-SVM with 6 features and 3 features on the test set. DETAILED DESCRIPTION

[0046] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0047] Example

[0048] Figure 1 This is a flow chart of the photovoltaic array dust accumulation degree evaluation method based on GS-SVM of the present invention.

[0049] In this embodiment, if Figure 1 As shown, the present invention provides a method for evaluating the dust accumulation degree of a photovoltaic array based on GS-SVM, comprising the following steps:

[0050] S1. According to Figure 2 The dust accumulation photovoltaic array operating condition test model shown is used to collect the short-circuit current of the photovoltaic array under different operating conditions;

[0051] S1.1. Set the dust density interval Δd, divide the dust density into levels, and set the level number k, k = 1, 2, ..., K, where K is the number of dust density levels;

[0052] S1.2, collecting data of photovoltaic array at different light intensities G i , different ambient temperatures T j And different dust accumulation density ρ k The short-circuit current under the i-th light intensity, j-th ambient temperature and k-th dust density is recorded as I i,j,k , i=1,2,…,n,j=1,2,…,m,n is the level of light intensity, m is the level of ambient temperature;

[0053] In this embodiment, the dust density range is 0-15g / m 2 , the interval is 5g / m 2 , so K = 4; the range of light intensity is 200~1200W / m 2 , the interval is 50W / m 2 , so n = 21; the ambient temperature is 20 ~ 50 ° C, the interval is 5 ° C, so m = 7; a total of 588 groups of short-circuit currents of photovoltaic modules are collected.

[0054] In this embodiment, the volt-ampere characteristic curves of the photovoltaic array under different dust accumulation densities are measured as follows: Figure 3 As shown, in Figure 3 In the short-circuit current I i,j,k (The dust accumulation density from top to bottom is 0, 5, 10, 15) and the maximum power point current I m The change is obvious, and it decreases with the increase of dust density; the open circuit voltage U oc There is basically no change, the maximum power point voltage U m It increases with the increase of dust accumulation density, but the change amplitude is very weak. Both of them cannot obviously reflect the change of the dust accumulation degree of the photovoltaic array. Figure 3 The volt-ampere characteristic curve of the photovoltaic array when mixed faults occur is shown in Figure 2. Taking the volt-ampere characteristic curve of the photovoltaic array with a single dust accumulation fault as a reference, when other faults occur in the dust accumulation array, the occurrence of other faults will cause I m Significantly changing and irregular, and I i,j,k 、U oc 、U m There is basically no change. Therefore, I i,j,k As a characteristic quantity for evaluating the dust accumulation degree of the photovoltaic array, the interference of other faults on the evaluation of the dust accumulation degree of the photovoltaic array can be eliminated.

[0055] S2, build training data set;

[0056] In each set of collected data, the short-circuit current I i,j,k The corresponding light intensity G i and ambient temperature T j Composition of training data x d =(G i ,T i ,I i,j,k ), the corresponding dust accumulation density ρ k As label y d , thus constructing the training data set {(x1,y1),(x2,y2),…,(x d ,y d ),…(x D ,y D )};

[0057] S3. Set the hyperparameters of the support vector machine (SVM) to be optimized.

[0058] S3.1. Select three kernel functions and set the value of the kernel parameter g and the penalty factor C of each kernel function.

[0059] In this embodiment, the selected kernel functions include: linear kernel function, polynomial kernel function and Gaussian kernel function, and the value ranges of the corresponding parameters and penalty factors of each kernel function are: the penalty factor C of the linear kernel function ranges from [1,100], with a value interval of 5; the kernel parameter g of the polynomial kernel function ranges from [2,5], with a value interval of 1, and the penalty factor C ranges from [1,100], with a value interval of 5; the kernel parameter g of the Gaussian kernel function is [-10,10], with a value interval of 0.5, and the penalty factor C ranges from [1,100], with a value interval of 5;

[0060] S3.2. Introducing the lth kernel function There is a mapping from input space to feature space Find the maximum-margin separating hyperplane in a high-dimensional feature space:

[0061]

[0062] Where w, b are hyperparameters, d1≠d2, d1, d2∈[1,D];

[0063] S3.2. Using the Lagrangian function and strong dual transformation, we can eliminate w and b and transform the above equation into the following target optimization problem:

[0064]

[0065] in, is the Lagrangian operator under the corresponding dust accumulation degree, C is the penalty factor of the kernel function;

[0066] S4, update the training data set;

[0067] S4.1. Substitute the training data (x1, x2) and the corresponding labels into the formula in step S3.2, calculate the target value minL, and then replace the kernel function Finally, we get three sets of target values ​​minL, and finally we sum the three sets of target values ​​minL and replace them with x1 in the training dataset. The updated x1 is recorded as y1 remains unchanged;

[0068] S4.2, continue processing the labels corresponding to the training data (x2, x3) according to step S4.1 to obtain the updated y2 remains unchanged; and so on until the last set of training data The corresponding tag continues to be processed according to step S4.1 to obtain the updated y D Save unchanged;

[0069] The outputs of multiple kernel functions are stacked to create a higher-dimensional feature representation. Each kernel function output can be used as a new feature dimension and combined with the original features to construct a richer feature space. This allows for more connections between data when training data is limited, which can improve model performance.

[0070] S5. Construct GS-SVM model;

[0071] The GS-SVM model includes a three-dimensional coordinate system grid GS and a support vector machine SVM, wherein the three-dimensional coordinate system grid GS model is (K, C, g), K represents the introduced kernel function, C is the penalty factor of the kernel function, and g is the kernel parameter of the kernel function;

[0072] S6. Train the GS-SVM model;

[0073] S6.1. Divide the updated training data set into M training subsets, where the κth training subset is denoted as σ κ ,κ=1,2,…,M;

[0074] S6.2. Initialize κ = 1 and input the κth training subset into the GS-SVM model. Then, select a set of hyperparameters (K, C, g) based on the 3D coordinate grid GS and substitute them into the SVM model. The SVM model processes each set of training data in the κth training subset to obtain the corresponding predicted label.

[0075] S6.3. Calculate the classification accuracy of each training data set based on the predicted label and the true label. If the predicted label is the same as the true label, the classification accuracy is recorded as 1, otherwise it is recorded as 0.

[0076] S6.4. Calculate the average classification accuracy of all training data in the κth training subset, denoted as p κ ;

[0077] S6.5. Determine whether the current number of iterations κ has reached the maximum value M. If κ < M, proceed to step S6.6; otherwise, proceed to step S6.7.

[0078] S6.6. Set κ = κ + 1, and then return to step S6.2;

[0079] S6.7. Find the maximum value max(p κ ), and then find max(p κ ) The set of hyperparameters (K, C, g) is the optimal hyperparameter, and the trained GS-SVM model is obtained;

[0080] S7, real-time assessment of the degree of dust accumulation in photovoltaic arrays;

[0081] The short-circuit current I of the PV array is collected in real time when the light intensity G and ambient temperature T are unknown but the dust accumulation density is unknown. The light intensity G, ambient temperature T, and short-circuit current I are then combined into test data x = (G, T, I). The test data x = (G, T, I) is then input into the trained GS-SVM model to output the predicted label value and obtain the dust accumulation level of the PV array.

[0082] Effect evaluation

[0083] In order to verify the proposed i,j,k )'s GS-SVM model for evaluating the dust accumulation degree of photovoltaic arrays. This paper evaluates and compares the performance of the proposed method from two aspects: evaluation accuracy and anti-interference ability.

[0084] 1) Comparison of the GS-SVM model and other models in terms of performance evaluation accuracy

[0085] This paper uses the test set data to test the accuracy of the dust accumulation degree assessment model based on GS-SVM, and compares its performance with other common classification algorithms, including decision tree and XGBoost algorithms. When the total number of data sets is fixed, this paper trains the model by feeding different proportions of data into the model as the training set. After the training is completed, the remaining data is used as the test set to test the model performance. The test results are as follows: Figure 4 shown.

[0086] from Figure 4 As can be seen, the evaluation accuracy of the SVM model after GS hyperparameter optimization is significantly higher than that of the SVM model without hyperparameter optimization. This demonstrates that the GS algorithm plays a crucial role in searching for optimal hyperparameters for SVM models and improving model performance. Furthermore, compared with the DecisionTree, GS-DecisionTree, and XGBoost models, the GS-SVM model also demonstrates superior performance on small sample sizes. During model training, when the training set data proportion reached 40%, the DecisionTree, GS-DecisionTree, XGBoost, and GS-SVM models achieved a fit of approximately 90% to the training set. However, their evaluation accuracy on the test set was below 80%, indicating that these algorithms suffer from severe overfitting when training on small sample sizes. However, the GS-SVM and SVM achieved similar classification accuracy on both the training and test sets, regardless of the training set sample proportion, demonstrating the SVM's superior robustness compared to the other models. Furthermore, only when the training set samples account for 90% can the DecisionTree model achieve the same high classification accuracy as the GS-SVM model. The classification accuracy of the SVM, GS-DecisionTree, and XGBoost algorithms is lower than that of the GS-SVM model. This shows that GS-SVM has better accuracy performance than other classification algorithms on small sample sets.

[0087] 2) Performance comparison of the GS-SVM algorithm using 6 features and the GS-SVM algorithm using 3 features

[0088] In order to evaluate the health status of photovoltaic arrays, many researchers use I i,j,k 、U oc 、U m , I m Four electrical parameters and two environmental parameters G and T are used as the model input for research. i,j,k Three feature quantities, one electrical parameter and two environmental parameters, G and T, serve as model inputs. To compare the performance of two parameter feature selection methods for assessing the degree of dust accumulation in photovoltaic arrays, this paper trained the GS-SVM algorithm model using the training data set, which accounted for 90% of the total dataset. The performance of the two models was then tested using 14 test data sets. The first 10 data points represent a single fault associated with dust accumulation in the photovoltaic array; F1 represents a combined fault associated with dust accumulation and a short circuit; F2 represents a combined fault associated with dust accumulation and an open circuit; F3 represents a combined fault associated with dust accumulation and aging; and F4 represents a combined fault associated with dust accumulation and partial shading.

[0089] Figure 5The following are the evaluation results of the two models on the test set. As can be seen from the figure, the accuracy of the GS-SVM algorithm using 6 features on the 14 test sets is only 80%, which is much lower than the GS-SVM using 3 features. Moreover, during the simulation, the former takes 182 seconds to run, while the latter only takes 18 seconds. This shows that reducing the selection of features can not only improve the generalization performance of the model to a certain extent, but also speed up the model training speed. In addition, it can be seen that the GS-SVM algorithm model with 3 features proposed in this paper can still accurately identify the degree of dust accumulation in the photovoltaic array when other faults are introduced, and has very good anti-interference ability; while the GS-SVM model with 6 features has a further decrease in accuracy when other faults are introduced, which is only 25%. The reason is that the introduction of short circuit, open circuit, aging and shadow faults will affect the U oc 、U m , I m This will have an impact, thereby interfering with the evaluation and judgment of the model.

[0090] In summary, the GS-SVM using three features proposed in this paper not only has a high evaluation accuracy on the small sample set, but also can resist the interference of faults such as short circuit, open circuit, and aging on the dust accumulation degree evaluation results, and the training speed is also greatly improved.

[0091] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A method for evaluating the dust accumulation degree of photovoltaic arrays based on GS-SVM, characterized in that: The following steps are involved: (1) Collect the short-circuit current of the photovoltaic array under different conditions; (1.1) Set the dust density interval Δd, divide the dust density into levels, and set the level number k, k = 1, 2, ..., K, K is the number of dust density levels; (1.2) Collect the photovoltaic array at different light intensities G i , different ambient temperatures T j And different dust accumulation density ρ k The short-circuit current under the i-th light intensity, j-th ambient temperature and k-th dust density is recorded as I i,j,k , i=1,2,…,n,j=1,2,…,m,n is the level of light intensity, m is the level of ambient temperature; (2) Constructing a training dataset; In each set of collected data, the short-circuit current I i,j,k The corresponding light intensity G i and ambient temperature T j Composition of training data x d =(G i ,T i ,I i,j,k ), the corresponding dust accumulation density ρ k As label y d , thus constructing the training data set {(x1,y1),(x2,y2),…,(x d ,y d ),…(x D ,y D )}; (3) Set the hyperparameters of the support vector machine (SVM) to be optimized; (3.1) Select N kernel functions, set the value of the kernel parameter g of each kernel function, and the value of the penalty factor C of each kernel function; (3.2), introduce the lth kernel function There is a mapping from input space to feature space Find the maximum-margin separating hyperplane in a high-dimensional feature space: Where w, b are hyperparameters, d1≠d2, d1, d2∈[1,D]; (3.2) Using the Lagrangian function and strong dual transformation, by eliminating w and b, the above equation is transformed into the following target optimization problem: in, is the Lagrangian operator under the corresponding dust accumulation degree, C is the penalty factor of the kernel function; (4) Update the training data set; (4.1) Substitute the training data (x1, x2) and the corresponding labels into the formula in step (3.2), calculate the target value minL, and then replace the kernel function Finally, we get N groups of target values ​​minL, and then we sum up the N groups of target values ​​minL and replace them with x1 in the training dataset. The updated x1 is recorded as y1 remains unchanged; (4.2) Process the labels corresponding to the training data (x2, x3) according to step (4.1) to obtain the updated y2 remains unchanged; and so on until the last set of training data The corresponding tag continues to be processed according to step (4.1) to obtain the updated y D Save unchanged; (5) Construct GS-SVM model; The GS-SVM model includes a three-dimensional coordinate system grid GS and a support vector machine SVM, wherein the three-dimensional coordinate system grid GS model is (K, C, g), K represents the introduced kernel function, C is the penalty factor of the kernel function, and g is the kernel parameter of the kernel function; (6) Training the GS-SVM model; (6.1), the updated training data set is divided into M training subsets, where the κth training subset is denoted as σ κ ,κ=1,2,…,M; (6.2) Initialize κ = 1, input the κth training subset into the GS-SVM model, and then select a set of hyperparameters (K, C, g) according to the three-dimensional coordinate system grid GS and substitute them into the SVM model. The SVM model processes each set of training data in the κth training subset and obtains the corresponding prediction label; (6.3) According to the predicted label and the true label of each set of training data, the classification accuracy of each set of training data is calculated. If the predicted label is the same as the true label, the classification accuracy is recorded as 1, otherwise it is recorded as 0; (6.4) Calculate the average classification accuracy of all training data in the κth training subset, denoted as p κ ; (6.5) Determine whether the current number of iterations κ has reached the maximum value M. If κ < M, proceed to step (6.6); otherwise, proceed to step (6.7); (6.6), let κ = κ + 1, and then return to step (6.2); (6.7) Find the maximum value max(p κ ), and then find max(p κ ) The set of hyperparameters (K, C, g) is the optimal hyperparameter, and the trained GS-SVM model is obtained; (7) Real-time assessment of the degree of dust accumulation in photovoltaic arrays; The short-circuit current I of the PV array is collected in real time when the light intensity G and ambient temperature T are unknown but the dust accumulation density is unknown. The light intensity G, ambient temperature T, and short-circuit current I are then combined into test data x = (G, T, I). The test data x = (G, T, I) is then input into the trained GS-SVM model to output the predicted label value and obtain the dust accumulation level of the PV array.

Images