A hybrid prediction method for direct irradiance based on multi-feature extraction of ground-based cloud images

Abstract

The present invention relates to a hybrid prediction model for direct irradiance based on multi-feature extraction of ground-based cloud maps, which combines knowledge-based cloud detection with an improved deep convolutional neural network. The extracted global features and local features are used as inputs to the feature fusion and irradiance prediction module, together with the solar altitude angle as input data. The features are comprehensively processed through a multi-layer fully connected network, thereby learning the complex relationship between the features; finally, the predicted value of direct irradiance is given through the regression output layer. The present invention fully considers and addresses the problem of direct solar radiation prediction under complex meteorological conditions. Through the joint extraction of global and local features, the model can more comprehensively describe the cloud characteristics, thereby achieving high-precision irradiance prediction; experimental results verify the effectiveness and superiority of the present invention, especially under complex and changeable weather conditions, showing high prediction accuracy and stability.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy prediction, and in particular relates to a hybrid prediction method for direct irradiance based on multi-feature extraction of ground-based cloud images. Background Art

[0002] In recent years, with the increasing global dependence on fossil fuels and the growing problem of environmental pollution, solar energy has gradually attracted widespread attention as a sustainable and clean energy source. Photovoltaic power generation, due to its environmentally friendly and resource-rich characteristics, is considered a key means of achieving carbon neutrality. However, complex meteorological conditions, especially changes in cloud cover, cause significant fluctuations in direct normal irradiance (DNI), affecting the power generation stability and efficiency of photovoltaic systems. Therefore, how to effectively and accurately predict direct normal irradiance has become a key component of photovoltaic power generation system prediction and scheduling.

[0003] Currently, mainstream direct solar irradiance (DNI) prediction methods can be divided into knowledge-based cloud detection methods and deep learning-based feature extraction methods. Knowledge-based cloud detection methods mainly rely on manually extracted cloud features, such as cloud coverage and transparency. These methods perform well under relatively simple weather conditions, but they show obvious shortcomings when dealing with changing and complex meteorological conditions and lack the ability to adapt to the diversity and randomness of clouds. Deep learning methods, such as convolutional neural networks (CNNs), have excellent performance in image recognition and prediction due to their powerful feature extraction and nonlinear fitting capabilities. However, due to the complexity of cloud changes, local features extracted by deep learning methods alone are often insufficient for CNN models to accurately describe the overall weather conditions. In particular, there are certain limitations when capturing long-span, large-scale cloud information.

[0004] Cloud variations can cause fluctuations in direct solar irradiance (DNI), leading to significant changes in photovoltaic (PV) power generation. Therefore, DNI prediction based on cloud characteristics is crucial for accurate PV power forecasting. However, modeling the relationship between DNI and cloud conditions remains challenging due to complex cloud variations and the ambiguity of key characteristics.

[0005] Therefore, the purpose of this paper is to propose a novel hybrid model that integrates knowledge-based cloud detection with an improved convolutional neural network, which can simultaneously extract global and local features of clouds, thereby improving the prediction accuracy of direct solar radiation under complex weather conditions. Summary of the Invention

[0006] In response to the problems existing in the existing technology, the present invention proposes a hybrid prediction model for direct irradiance based on multi-feature extraction of ground-based cloud images. By integrating knowledge-based cloud detection technology and an improved convolutional neural network (DC-CNN), comprehensive extraction of global and local features is achieved, and an end-to-end mapping of ground-based cloud images and irradiance is constructed to achieve high-precision prediction of direct solar irradiance.

[0007] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0008] A hybrid prediction method for direct irradiance based on multi-feature extraction of ground-based cloud images includes the following steps:

[0009] Step 1: Obtain historical ground-based cloud images and historical irradiance data and pre-process them, and calculate the corresponding historical solar zenith angle data for model prediction;

[0010] Step 2: Using a knowledge-based cloud detection feature extraction module, based on the red-to-blue ratio (RBR) threshold segmentation method, and taking into account the influence of aerosol optical depth, the clear sky library method is used to further extract thin and thick clouds, resulting in a segmented cloud map that distinguishes thin clouds, thick clouds, and clear sky.

[0011] Step 3: Based on the segmented cloud image, a thick and thin cloud matrix U is obtained to distinguish thick clouds, thin clouds, and clear sky pixels. The cloud amount characteristics are extracted according to the proportion of cloud pixels.

[0012] Step 4: Further refine the description of cloud amount. Based on the segmented cloud image, different cloud pixels are assigned different weights. Combined with the brightness matrix, the cloud thickness coefficient feature is obtained. The knowledge-based cloud detection module is responsible for extracting global features, including cloud amount and cloud thickness coefficient features.

[0013] Step 5: Using an improved convolutional neural network (DC-CNN) feature extraction module to better extract local features of the cloud layer and increase the comprehensiveness of cloud image features; the improved convolutional neural network combines dilated convolution with a coordinate attention mechanism;

[0014] Step 6: After the two feature extraction modules complete the extraction of global features and local features, the feature fusion and irradiance prediction module uses a multi-layer fully connected network and a nonlinear activation function (ReLU function) to comprehensively process the features, thereby learning the complex relationship between the features. Finally, the predicted value of direct irradiance is obtained through the regression output layer;

[0015] Preferably, the final values ​​of the thin cloud threshold and the thick cloud threshold are determined through experience and repeated experiments. Based on the final values, the thick and thin cloud matrix U that distinguishes thick clouds, thin clouds and clear sky pixels is used to calculate the cloud amount f according to the following formula:

[0016]

[0017]

[0018] Where m and n represent the pixel size of the image, i and j represent the pixel coordinates in the image, i represents the row index in the image, j represents the column index in the image, Cloud(i,j) and Clear(i,j) are divided into 0 or 1 according to the pixel classification, and pixel(i,j) represents the type of pixel at position (i,j) in the image.

[0019] Preferably, in step 4, considering that the clouds in the ground-based cloud image are generally white, and thicker clouds are relatively dark, the cloud image is analyzed using the HSV color model, and the V value representing the image brightness (brightness value V) is extracted as a direct indicator of the sky brightness:

[0020] V=max(R,G,B) (4)

[0021] Based on the thick and thin cloud matrix U, different weight coefficients are assigned to the corresponding positions of thin cloud, thick cloud, and clear sky pixels. A matrix with the same dimensions as the cloud image pixels is established, with a size of m×n, where m and n are the pixel sizes of the image, respectively. This gives a refined description of the cloud amount matrix W. Combined with the clear sky library method, the V value that best matches the clear sky cloud image at the current moment is extracted to form the brightness matrix V, with a size of m×n.

[0022] Each cloud image (if its size is 227*227) corresponds to a pixel matrix (also 227*227 in size). Matrix U refers to the matrix that is converted from the pixel matrix corresponding to the original cloud image through the aforementioned steps to distinguish between thick clouds, thin clouds, and clear sky pixels (the original pixel range of each pixel is 0-255. After conversion, the original pixel values ​​are divided into three fixed values ​​corresponding to the three different sky types: clear sky, thin clouds, and thick clouds).

[0023] The cloud cover matrix W is a matrix obtained by directly assigning different weight coefficients to three different types of pixels on the matrix U, and its size is the same as U.

[0024] By calculating the V value of each pixel point according to the formula of the brightness value V at each pixel position, the V value of this pixel position is taken as the V value of this pixel position. All pixel positions adopt the above processing method to form a brightness matrix V together.

[0025] Furthermore, to reflect the impact of clouds on irradiance under different atmospheric aerosol concentrations, the following cloud thickness coefficient calculation formula is used: The cloud thickness of any pixel in the ground-based cloud map can be regarded as the multiplication of the brightness (considering the influence of aerosols) and cloud cover under ideal cloudless conditions. CTAs shown in the following formula:

[0026]

[0027] Where Tr represents the trace of the matrix, T represents the transpose symbol, V represents the brightness matrix under ideal cloud-free conditions, which is obtained using the clear sky cloud map that best matches the current cloud map, W is the cloud amount matrix for fine-grained description, and υ ij is the V value of the pixel with coordinate (i, j) in the matrix V.

[0028] Preferably, in step 5, the improved convolutional neural network (deep neural network) combines dilated convolution (DC) and coordinate attention mechanism (CA) to enhance the ability to capture multi-scale features. Dilated convolution expands the receptive field by introducing intervals in the convolution kernel, allowing the network to capture cloud feature information in a larger range without significantly increasing computational complexity. At the same time, the coordinate attention mechanism effectively enhances the model's attention to features at different positions in the feature map by introducing spatial coordinate information. Because it not only focuses on the feature relationship between channels, but also combines the spatial position of the features, it performs well in processing tasks involving complex spatial structures and enhances the network's perception of long-distance dependent features of the cloud layer.

[0029] As a preferred approach, the proposed DC-CNN feature extraction module consists of 11 layers, including 5 dilated convolution layers, 3 average pooling layers, 1 coordinate attention mechanism layer, a tiling layer, and a fully connected layer. DC-CNN uses dilated convolution to extract more receptive fields from the input tensor and adopts a hierarchical feature extraction design. In the early stages, larger convolution kernels are used to extract more spatial information and primary features, such as cloud edges and large-area cloud structures. Subsequent layers of convolution continue to deepen the extraction of more complex and advanced features, and add a coordinate attention mechanism module deep in the network. Combining spatial coordinate information and the feature relationship between channels, it improves the ability to capture information about large-span clouds in the image. After the tiling layer, the extracted features are obtained through the fully connected layer.

[0030] Preferably, in step 6, three fully connected layers are used to learn the nonlinear combination of features. The first two fully connected layers use the ReLU activation function and have 32 and 64 neurons respectively. The final output layer has 1 neuron to obtain the solar irradiance prediction result.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] (1) By combining knowledge-based cloud detection with an improved convolutional neural network feature extraction module, this paper can simultaneously extract global cloud features (such as cloud amount and cloud thickness coefficient) and local features (such as cloud edges and morphology), greatly improving the ability to describe complex cloud features. This combined global and local feature extraction method can effectively address the problem of direct solar irradiance prediction under variable weather conditions.

[0033] (2) Traditional DNI prediction methods are unstable when weather conditions change dramatically. However, this paper enhances the perception of large-scale cloud features through dilated convolution and coordinate attention mechanisms, enabling the model to maintain high prediction accuracy even when dealing with rapid cloud changes. Experiments show that the proposed model exhibits good robustness and adaptability under complex meteorological conditions, significantly outperforming existing baseline models.

[0034] (3) Through comparative experiments, it can be seen that the present invention has significant improvements in nRMSE and nMAE, which are reduced by at least 15.40% and 27.98% respectively compared with the traditional benchmark model. 2 The value increased by at least 3.5%, indicating that the accuracy of the model in predicting direct solar irradiance has been significantly improved, which can better support the optimal scheduling of photovoltaic power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a cloud segmentation image generated using the knowledge-based cloud detection module;

[0036] Figure 2 It is a schematic diagram of the overall structure of the improved convolutional neural network DC-CNN;

[0037] Figure 3 It is a bar chart of the overall performance evaluation indicators of four different prediction models;

[0038] Figure 4 It is the direct irradiance prediction map for high-variability days;

[0039] Figure 5 It is a direct irradiance prediction map for low-variability days. DETAILED DESCRIPTION

[0040] The technical solutions of the present invention are further specifically described below through examples. These examples are provided for the purpose of illustrating the present invention and are not intended to limit the present invention. All other examples obtained by persons of ordinary skill in the art based on the examples in this application without creative work are intended to fall within the scope of protection of this application.

[0041] A hybrid prediction method for direct irradiance based on multi-feature extraction of ground-based cloud images includes the following steps:

[0042] Step 1: Obtain historical ground-based cloud images and historical irradiance data and pre-process them, and calculate the corresponding historical solar zenith angle data for model prediction;

[0043] Step 2: Using a knowledge-based cloud detection feature extraction module, based on the red-to-blue ratio (RBR) threshold segmentation method, and taking into account the influence of aerosol optical depth, the clear sky library method is used to further extract thin and thick clouds, resulting in a segmented cloud map that distinguishes thin clouds, thick clouds, and clear sky.

[0044] Step 3: Based on the segmented cloud image, a thick and thin cloud matrix U is obtained to distinguish thick clouds, thin clouds, and clear sky pixels. The cloud amount characteristics are extracted according to the proportion of cloud pixels.

[0045] Step 4: Further refine the description of cloud amount. Based on the segmented cloud image, different cloud pixels are assigned different weights. Combined with the brightness matrix, the cloud thickness coefficient feature is obtained. The knowledge-based cloud detection module is responsible for extracting global features, including cloud amount and cloud thickness coefficient features.

[0046] Step 5: Use the improved convolutional neural network (DC-CNN) feature extraction module to better extract the local features of the cloud layer and increase the comprehensiveness of the cloud image features;

[0047] Step 6: After the two feature extraction modules complete the extraction of global features and local features, the feature fusion and irradiance prediction module uses a multi-layer fully connected network and a nonlinear activation function (ReLU function) to comprehensively process the features, thereby learning the complex relationship between the features. Finally, the predicted value of direct irradiance is obtained through the regression output layer;

[0048] The final values ​​of the thin cloud threshold and thick cloud threshold are determined through experience and repeated experiments. Based on the final values, the thick and thin cloud matrix U that distinguishes thick clouds, thin clouds and clear sky pixels is used to calculate the cloud amount f according to the following formula:

[0049]

[0050] Where m and n represent the pixel size of the image, i and j represent the pixel coordinates in the image, i represents the row index in the image, j represents the column index in the image, Cloud(i,j) and Clear(i,j) are divided into 0 or 1 according to the pixel classification, and pixel(i,j) represents the type of pixel at position (i,j) in the image.

[0051] In step 4, considering that the clouds in the ground-based cloud image are generally white, and thicker clouds are relatively dark, the cloud image is analyzed using the HSV color model, and the V value representing the image brightness is extracted as a direct indicator of the sky brightness:

[0052] V=max(R,G,B) (4)

[0053] Based on the thick and thin cloud matrix U, different weight coefficients are assigned to the corresponding positions of thin cloud, thick cloud, and clear sky pixels. A matrix with the same dimensions as the cloud image pixels is established, with a size of m×n, where m and n are the pixel sizes of the image, respectively. This gives a refined description of the cloud amount matrix W. Combined with the clear sky library method, the pixel value V that best matches the clear sky cloud image at the current moment is extracted to form the brightness matrix V, with a size of m×n.

[0054] To reflect the impact of clouds on irradiance under different atmospheric aerosol concentrations, the following cloud thickness coefficient calculation formula is used: The cloud thickness of any pixel in the ground-based cloud map can be regarded as the multiplication of the brightness (taking into account the influence of aerosols) and cloud cover under ideal cloudless conditions. CT As shown in the following formula:

[0055]

[0056] Where Tr represents the trace of the matrix, T represents the transpose symbol, V represents the brightness matrix under ideal cloud-free conditions, which is obtained using the clear sky cloud map that best matches the current cloud map, W is the cloud amount matrix for fine-grained description, and υ ij is the V value of the pixel with coordinate (i, j) in the matrix V.

[0057] In step 5, the improved convolutional neural network (DC-CNN) combines dilated convolution (DC) with the coordinate attention mechanism (CA) to enhance its ability to capture multi-scale features. Dilated convolution expands the receptive field by introducing spacing in the convolution kernel, allowing the network to capture cloud feature information over a wider range without significantly increasing computational complexity. Simultaneously, the coordinate attention mechanism effectively enhances the model's attention to features at different locations in the feature map by introducing spatial coordinate information. Because it not only focuses on the feature relationships between channels but also incorporates the spatial location of features, it excels in tasks involving complex spatial structures and enhances the network's perception of long-range dependent features in the cloud layer.

[0058] The proposed DC-CNN feature extraction module consists of 11 layers, including five dilated convolutional layers, three average pooling layers, one coordinate attention mechanism layer, a tiling layer, and a fully connected layer. DC-CNN uses dilated convolution to extract more receptive fields from the input tensor and adopts a hierarchical feature extraction design. In the early stages, larger convolution kernels are used to extract more spatial information and primary features, such as cloud edges and large-area cloud structures. Subsequent convolutional layers continuously deepen the extraction of more complex and advanced features. A coordinate attention mechanism module is added deep into the network, combining spatial coordinate information with feature relationships between channels to improve the ability to capture information about large-span clouds in the image. After the tiling layer, the extracted features are passed through the fully connected layer.

[0059] In step 6, three fully connected layers are used to learn the nonlinear combination of features. The first two fully connected layers use the ReLU activation function and have 32 and 64 neurons respectively. The final output layer has 1 neuron to obtain the solar irradiance prediction result.

[0060] The effectiveness of the direct irradiance hybrid prediction method based on multi-feature extraction of ground-based cloud images in this paper was verified using ground-based cloud images and solar measurement data from the National Renewable Energy Laboratory (NREL). For the processing of ground-based cloud images, the main task is to extract the region of interest (ROI) of the RGB image so as to remove unnecessary pixels and retain only the pixels of the sky image. The mask is a binary matrix corresponding to the positions of the original ground-based cloud image. The pixel values ​​in the mask are only 0 or 1. 1 corresponds to the position of the sky image in the original ground-based cloud image where edge obstacles are removed, i.e., the ROI, and 0 corresponds to the remaining positions of non-interest in the original ground-based cloud image. After multiplying the original ground-based cloud image by the mask, the processed image is obtained, which is scaled and used as the input of the model with a resolution of 227×227 pixels. Data from 61 full days from April 1 to May 31, 2021 were selected (the data set contains various weather types such as sunny, cloudy, and overcast days). After aligning the image with the numerical data, two typical days were selected as the data source for the subsequent case study. The remaining 4,760 data points were available. To ensure effective training and reliable validation, the data was randomly shuffled and divided into training, validation, and test sets. The training dataset consists of 2,856 data points, the validation set consists of 476 data points, and the test set consists of 1,428 data points.

[0061] Figure 1The cloud segmentation images generated by the knowledge-based cloud detection method under different weather types are shown. The original ground-based cloud images (a, c, e, g) and the segmentation images obtained by the knowledge-based cloud detection method (b, d, f, h) are shown: (1) sunny (a, b); (2) few clouds (c, d); (3) cloudy (e, f); (4) overcast (g, h). In the cloud segmentation images, dark gray indicates thick clouds, light gray indicates thin clouds, and dark blue indicates clear sky. Based on the segmented cloud images, the two global features of cloud amount and cloud thickness coefficient are calculated according to formulas (1), (2), (3) and (4), (5), respectively, providing important support for the establishment of subsequent models.

[0062] For local feature extraction of sky images, an improved convolutional neural network DC-CNN feature extraction module is proposed, which combines dilated convolution to expand the receptive field and coordinate attention mechanism to enhance long-range dependencies. Figure 2 This is a schematic diagram of the overall DC-CNN architecture. Colors represent different layer types. The top section shows the DC-CNN layers, and the bottom section shows the detailed hyperparameters for each layer. The ground cloud image is used as the input for the DC-CNN. The DC-CNN hierarchical feature extraction design extracts cloud image features layer by layer. After the dilated convolutional layer, average pooling layer, coordinate attention mechanism layer, and tiling layer, the fully connected layer extracts local features.

[0063] The global and local features extracted by combining the knowledge-based cloud detection feature extraction module and the improved convolutional neural network (DC-CNN) feature extraction module serve as input to the feature fusion and irradiance prediction module. Together with the solar altitude angle, these features are processed through a multi-layer fully connected network to learn the complex relationships between the features. Finally, a regression output layer provides a predicted value for direct irradiance.

[0064] In order to verify the contribution of different modules to the overall model performance, the present invention conducted multiple sets of ablation experiments to evaluate the impact of the knowledge-based cloud detection feature extraction module, the DC-CNN feature extraction module, and the combination of the two on the final accuracy of the model. The experimental results are shown in Table 1, where (1) Baseline: The baseline model proposed in the present invention contains the main part of the proposed method, but removes the DC and AC components in the DC-CNN feature extraction module and the knowledge-based feature extraction module. (2) Baseline+A: It means that the complete DC-CNN feature extraction module is restored based on the baseline model. (3) Baseline+B: It means that the knowledge-based feature extraction module is embedded in the baseline model. (4) Baseline+A+B: The knowledge-based feature extraction module and the complete DC-CNN feature extraction module, including the AC and DC modules, are simultaneously applied in the baseline model.

[0065] Table 1 Ablation experiment results on the test set

[0066] Model <![CDATA[R 2 ]]> nMBE nMAE nRMSE Baseline 0.8657 -0.1452 0.1987 0.3935 Baseline+A 0.8979 0.0321 0.1731 0.3430 Baseline+B 0.9005 0.0775 0.1645 0.3386 Baseline+A+B 0.9218 0.0098 0.1377 0.3000

[0067] After introducing the knowledge-based cloud detection feature extraction module, the R 2 The value of R is increased by 4.02%, and nRMSE and nMAE are reduced by 13.95% and 17.21% respectively; the DC-CNN feature extraction module improves the ability to capture multi-scale features and the perception of large-span cloud features, making R 2 The improvement is 3.72%, and the nRMSE and nMAE are reduced by 12.83% and 12.88% respectively.

[0068] In order to compare with the proposed hybrid model, three classic model methods for solar radiation prediction based on ground-based cloud imagery were selected as benchmark models. They include (1) the SolarNet model, which is based on the VGG-16 architecture and captures the potential relationship between sky images and irradiance. It only requires a single image to achieve end-to-end irradiance prediction. (2) the 3D-CNN model, which models the dynamic changes and spatial distribution of clouds, captures the dynamic characteristics of clouds changing over time, and achieves irradiance prediction. (3) the CNN-MLP model, which combines convolutional neural networks with multi-layer perceptrons (MLPs). The cloud features extracted by CNN are fused with other meteorological variables to improve the accuracy of irradiance prediction. The experimental results are shown in Table 2.

[0069] Table 2 Statistical results of overall performance evaluation indicators of four different prediction models

[0070]

[0071]

[0072] The corresponding overall performance index bar chart is drawn as follows Figure 3 As shown in the figure, it can be found that the overall performance of the hybrid model proposed in this invention is the most superior. Among all the methods, the nRMSE and nMAE values ​​of the method proposed in this invention are the smallest, which are at least reduced by 15.40%, 27.98% and R 2 The value is the largest, which is at least improved by 3.5%. The corresponding nRMSE is 0.300, nMAE is 0.1377, and R 2 It is 0.9218.

[0073] The model is case-validated on a typical day with high variability.

[0074] A high-variability day is one in which a high-frequency irradiance change suddenly occurs after a period of clear skies. Figure 4A high variability daily direct irradiance forecast is shown. This case is challenging due to the rapid and sometimes difficult to estimate changes in DNI. The 3D CNN model performs the worst in this case, with nRMSE 3D CNN =0.3397, nRMSE given by CNN-MLP model and SolarNet model CNN-MLP =0.2498, nRMSE SolarNet =0.2697; the model proposed in this invention performs best when used, nRMSE HM =0.1647, nMAE HM =0.1362, R 2 HM =0.8916. In this complex cloud weather situation, the proposed model accurately estimates the sudden drop in DNI caused by clouds.

[0075] The model is case-validated on a typical day with low variability.

[0076] A low-variability day is defined as a mostly clear sky with some cloud at certain times of the day. Figure 5 A low-variability daily direct irradiance forecast is shown. The CNN-MLP model performs worst in this case, with nRMSE CNN-MLP =0.1754, nRMSE given by 3D CNN model and SolarNet model 3DCNN =0.1425, nRMSE SolarNet =0.1313, the model proposed in this invention obtains the best result among the four methods, nRMSE HM =0.0962,nMAE HM =0.0709,R 2 HM = 0.9390. As can be seen from the sky images taken at 2:10 PM and 2:20 PM, the sun is obscured by approaching thin clouds, causing a sharp drop in DNI. Among the four methods, the proposed model accurately captures the drop in DNI, demonstrating the effectiveness of the model.

[0077] This paper proposes a hybrid direct irradiance prediction model based on ground-based cloud images using multi-feature extraction. This model combines knowledge-based cloud detection with an improved deep convolutional neural network to fully address the problem of direct solar radiation prediction under complex meteorological conditions. By jointly extracting global and local features, the model can more comprehensively describe cloud characteristics, thereby achieving high-precision irradiance prediction. Experimental results demonstrate the effectiveness and superiority of this method, demonstrating high prediction accuracy and stability, particularly under complex and changing weather conditions.

[0078] The present invention has been described in detail above with reference to the embodiments. However, the contents described are only preferred embodiments of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A hybrid prediction method for direct irradiance based on multi-feature extraction of ground-based cloud images, characterized by The following steps are involved: Step 1: Obtain historical ground-based cloud images and historical irradiance data and pre-process them, and calculate the corresponding historical solar zenith angle data for model prediction; Step 2: Using a knowledge-based cloud detection feature extraction module, based on the red-to-blue ratio (RBR) threshold segmentation method, and taking into account the influence of aerosol optical depth, the clear sky library method is used to further extract thin and thick clouds, resulting in a segmented cloud map that distinguishes thin clouds, thick clouds, and clear sky. Step 3: Based on the segmented cloud image, a thick and thin cloud matrix is ​​obtained to distinguish thick clouds, thin clouds, and clear sky pixels. U , extract cloud cover features based on the proportion of cloud pixels; Step 4: Further refine the description of cloud amount, and segment the cloud map according to the thick and thin cloud matrix. U , different weights are assigned to the corresponding positions of thin cloud, thick cloud and clear sky pixels, and combined with the brightness matrix to obtain the cloud thickness coefficient feature. The knowledge-based cloud detection module is responsible for extracting global features, including cloud amount and cloud thickness coefficient features; Step 5: Using an improved convolutional neural network feature extraction module to extract local features of the cloud layer; the improved convolutional neural network combines dilated convolution with a coordinate attention mechanism; Step 6: After completing the extraction of global features and local features, the feature fusion and irradiance prediction module is used, a multi-layer fully connected network is adopted, and the features are comprehensively processed using a nonlinear activation function to learn the complex relationship between the features. Finally, the predicted value of direct irradiance is obtained through the regression output layer.

2. The hybrid direct irradiance prediction method based on multi-feature extraction of ground-based cloud images according to claim 1 is characterized by: In step 3, the final values ​​of thin cloud threshold and thick cloud threshold are determined through experience and repeated experiments, and the thick and thin cloud matrix of thick cloud, thin cloud and clear sky pixels is distinguished according to the final values. U , the cloud cover is calculated according to the following formula f : ； Among them, m and n represent the pixel size of the image, i 、 j represents the pixel coordinates in the image, i represents the row index in the image, j Represents the column index in the image, Cloud( i,j ) and Clear( i,j ) is divided into 0 or 1 according to the pixel classification. pixel (i,j) Represents the image at position ( i , j ) pixel type.

3. The hybrid direct irradiance prediction method based on multi-feature extraction of ground-based cloud images according to claim 1 is characterized by: In step 4, the cloud image is analyzed using the HSV color model, and the V value representing the image brightness is extracted as a direct indicator of sky brightness: ； According to the thick and thin cloud matrix U , assign different weight coefficients to the corresponding positions of thin cloud, thick cloud and clear sky pixel points, establish a matrix with the same pixel dimension as the segmented cloud image, with a size of m×n, where m and n are the pixel sizes of the segmented cloud image, and obtain a cloud amount matrix with refined description W , combined with the clear sky library method, extract the V value that best matches the clear sky cloud image at the current moment and form a brightness matrix V , the matrix size is m×n.

4. The hybrid direct irradiance prediction method based on multi-feature extraction of ground-based cloud images according to claim 3 is characterized in that Cloud thickness coefficient characteristics As shown in the following formula: ； in, Tr represents the trace of the matrix, T represents the transpose symbol, V represents the brightness matrix, W is the cloud cover matrix described in detail, is a matrix V The median coordinates ( i , j )The V value of the pixel.

5. The hybrid direct irradiance prediction method based on multi-feature extraction of ground-based cloud images according to claim 1 is characterized by: In step 5, the improved convolutional neural network feature extraction module consists of 11 layers, including 5 dilated convolutional layers, 3 average pooling layers, 1 coordinate attention mechanism layer, tiling layer and fully connected layer.

6. The hybrid direct irradiance prediction method based on multi-feature extraction of ground-based cloud images according to claim 5 is characterized by: The improved convolutional neural network uses dilated convolution to extract more receptive fields from the input tensor and adopts a hierarchical feature extraction design. In the early stage, larger convolution kernels are used to extract more spatial information and primary features. Subsequent layers of convolution continue to deepen the extraction of more complex and advanced features. A coordinate attention mechanism module is added deep in the network to combine spatial coordinate information and the feature relationship between channels to improve the ability to capture information of large-span clouds in the image. After the tiling layer and the fully connected layer, the extracted features are obtained.

7. The hybrid direct irradiance prediction method based on multi-feature extraction of ground-based cloud images according to claim 1 is characterized by: In step 6, three fully connected layers are used to learn nonlinear combinations of features.

8. The hybrid direct irradiance prediction method based on multi-feature extraction of ground-based cloud images according to claim 7 is characterized by: The first two fully connected layers use nonlinear activation functions and have 32 and 64 neurons respectively. The final output layer has 1 neuron to obtain the solar irradiance prediction results.

Images