Photovoltaic power generation control method and system

Abstract

The invention relates to the field of photovoltaic power generation, in particular to a photovoltaic power generation control method and system. A photovoltaic power generation control system comprises an actual output power monitoring module, a to-be-detected photovoltaic panel image acquisition module, an environment feature vector establishing vector, an environment feature graph establishing vector, an analog switch tube duty ratio output module, a photovoltaic power generation delay control strategy establishing module and a maximum output power updating module. According to the method, the environment feature vector corresponding to each photovoltaic panel is generated, the environment feature graph is constructed based on the environment feature vectors corresponding to all the photovoltaic panels, the maximum output power of the photovoltaic system under the current environment condition is predicted based on the environment feature graph, and the corresponding photovoltaic power generation delay control strategy is generated. Therefore, the illumination condition of each photovoltaic panel is considered, and the influence of a local extreme point caused by local shielding on the maximum output power is avoided.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic power generation, and more specifically, to a photovoltaic power generation control method and system. Background Art

[0002] Photovoltaic (PV) power generation is a technology that uses solar cells to convert solar radiation into electrical energy. The key to this technology lies in the efficiency of photovoltaic cells, that is, their ability to convert sunlight into electrical energy under specific lighting conditions. Photovoltaic power generation systems are designed to maximize the capture of sunlight and conversion into electrical energy, but in actual applications, many factors may affect the maximum output power of the system.

[0003] In the process of photovoltaic power generation, partial shadow is an important influencing factor. Partial shadow may be caused by surrounding buildings, trees or other obstacles, resulting in a part of the photovoltaic panel or some cells not receiving sufficient sunlight. The impact of this partial shading is far more than the reduced efficiency of a single shaded cell; because the cells in a photovoltaic panel are usually connected in series, the output drop of one cell will affect the power output of the entire photovoltaic panel, and the output drop of one photovoltaic panel will affect the power output of the entire photovoltaic system. Summary of the invention

[0004] The present invention obtains the corresponding lighting conditions by performing image analysis on each photovoltaic panel, generates an environmental feature vector corresponding to each photovoltaic panel based on the lighting conditions, constructs an environmental feature map based on the environmental feature vectors corresponding to all photovoltaic panels, predicts the maximum output power of the photovoltaic system under the current environmental conditions based on the environmental feature map, and generates a corresponding photovoltaic power generation delay control strategy, thereby taking into account the lighting conditions of each photovoltaic panel, avoiding the influence of local extreme points on the maximum output power due to local shading, reducing dependence on manual intervention, reducing operation and maintenance costs, improving the automation level of the system, and avoiding potential influence on system stability due to rapid or sudden changes through the photovoltaic power generation delay control strategy.

[0005] A photovoltaic power generation control method, comprising:

[0006] Continuously monitor the actual output power q of the photovoltaic system and determine whether the photovoltaic control restart condition is met. The photovoltaic control restart condition is "(η-q) / q≥ε", where η is the current maximum output power recorded, which is subsequently updated through the photovoltaic control restart operation, and ε is the power deviation threshold; if the photovoltaic control restart condition is met, execute the photovoltaic control restart operation; if the photovoltaic control restart condition is not met, continue to monitor the output power q of the photovoltaic system;

[0007] The PV control restart operation includes the following:

[0008] Obtaining the photovoltaic panel image to be inspected of each photovoltaic panel, where the photovoltaic panel image is an RGB image;

[0009] For each photovoltaic panel image to be detected, the image to be detected is processed by the trained image restoration model to obtain a spot-free photovoltaic panel image, and the spot-free photovoltaic panel image is subjected to illumination condition analysis to output an illumination intensity distribution matrix; the illumination intensity distribution matrix is ​​processed by a multi-layer perceptron to obtain an illumination distribution feature vector; the spot-free photovoltaic panel image is subjected to spectral analysis to output a spectral feature vector; the illumination distribution feature vector, the spectral feature vector and the ambient temperature value are concatenated to obtain an ambient feature vector;

[0010] All environmental feature vectors are concatenated to obtain an environmental feature map, in which each coordinate point corresponds to a photovoltaic panel, and each coordinate point stores the environmental feature vector corresponding to the photovoltaic panel;

[0011] The environmental feature map is sent to the trained maximum power output model for processing, and the simulated switch tube duty cycle is output. The simulated switch tube duty cycle refers to the proportion of the time that the switch tube in the photovoltaic inverter is turned on. It can adjust the output voltage and output current of the photovoltaic system, and then adjust the output power. At the same time, the maximum output power η is obtained based on the simulated switch tube duty cycle; the photovoltaic power generation delay control strategy is output based on the simulated switch tube duty cycle, and the maximum output power η is updated based on the simulated maximum output power;

[0012] The photovoltaic system is controlled based on the photovoltaic power generation delay control strategy.

[0013] Preferably, the image restoration model is established based on generating a U-net network;

[0014] The training of the image restoration model includes the following steps:

[0015] A number of photovoltaic panel images are obtained, where the photovoltaic panel images are images actually taken by the user. For each photovoltaic panel image, the photovoltaic panel image and its corresponding standard photovoltaic panel image are combined into a first training sample; all the first training samples are combined into a first training sample set, and then the first training sample set is sent to a parameter-initialized image restoration model for training. During the training, the standard photovoltaic panel image is used as a target to adjust the weight parameters to determine whether the first training condition is met. If the first training condition is met, the trained image restoration model is output; otherwise, iterative training is continued through the first training sample set.

[0016] Preferably, the spot-free photovoltaic panel image is subjected to a light condition analysis to output a light intensity distribution matrix, which specifically includes the following contents:

[0017] The spot-free photovoltaic panel image is sent to the trained light distribution detection model for processing, and the light intensity distribution matrix is ​​output;

[0018] The light distribution detection model is established based on the U-net network, including the U-net network layer, the self-attention mechanism layer and the light intensity distribution matrix output layer. The U-net network layer is used to extract the light distribution features of the spotless photovoltaic panel image; the self-attention mechanism layer is used to enhance the light distribution features; the light intensity distribution matrix output layer is used to output the light intensity distribution matrix based on the enhanced light distribution features.

[0019] Preferably, the spot-free photovoltaic panel image is sent to a trained light distribution detection model for processing, which specifically includes the following steps:

[0020] The spot-free photovoltaic panel image is sent to the U-net network layer for processing to obtain the light distribution feature T;

[0021] In the self-attention mechanism layer, the illumination distribution feature T is combined with the key-value weight matrix W k Multiply them together to get the key value matrix K; add the illumination distribution feature T and the value weight matrix W p Multiply them to get the value matrix P; convert the spot-free photovoltaic panel image to HSV space through color conversion, select the feature map corresponding to the V channel as the brightness feature map, and compare the brightness feature map with the query weight matrix W Q Multiply them to get the query matrix Q and calculate the attention weight matrix ATT = softmax((Q·K T ) / (μ) 0.5 ), where μ is the number of columns of the key-value matrix K; multiply the attention weight matrix ATT by the value matrix P to obtain the enhanced illumination distribution feature T;

[0022] In the light intensity distribution matrix output layer, the enhanced light distribution feature T is convolved to output the light intensity distribution matrix.

[0023] Preferably, the training of the illumination distribution detection model includes the following steps:

[0024] All photovoltaic panel images in the first training sample set are obtained, and all photovoltaic panel images in the first training sample set are processed by the trained image restoration model to obtain a number of spot-free photovoltaic panel images, each spot-free photovoltaic panel image and its corresponding standard light intensity distribution matrix form a second training sample, all second training samples form a second training sample set, and then the second training sample set is sent to the parameter-initialized illumination distribution detection model for training. During the training, the weight parameters are adjusted with the standard light intensity distribution matrix as the target to determine whether the second training condition is met. If the second training condition is met, the trained illumination distribution detection model is output; otherwise, iterative training is continued through the second training sample set.

[0025] Preferably, the maximum power output model is established based on a convolutional neural network;

[0026] Training the maximum power output model includes the following steps:

[0027] All photovoltaic panel images acquired in several time periods are obtained, and the environmental feature maps corresponding to all photovoltaic panel images in any time period are used as third training samples, and the third training samples are marked with the analog switch tube duty cycle output by the improved moth-to-flame optimization algorithm; all marked third training samples are combined into a third training sample set, and then the third training sample set is sent to the parameter-initialized maximum power output model for training. During the training, the loss value is calculated with the analog switch tube duty cycle corresponding to the third training sample as the target. If the loss value is within a preset range, the preset range is artificially set to constrain the accuracy of the maximum power output model, and the trained maximum power output model is output; otherwise, the iterative training is continued through the third training sample set.

[0028] Preferably, the improved moth-to-flame optimization algorithm is used to output the analog switch duty cycle, which specifically includes the following steps:

[0029] S1: Establish N moth individuals F n , n=1,2,3…N, and F n What is stored in is the duty cycle of the switch tube to be simulated;

[0030] S2: Set the maximum number of iterations M, let m = 1, and h is used to record the number of iterations;

[0031] S3: Count individual moths F n The corresponding fitness δ n , and all moth individuals F n According to the corresponding fitness δ n Arrange them from large to small, and arrange all the moth individuals F n Form a population set and select the first ζ moth individual F nThe flame individual set is composed. In the initial state, ζ=N. The moth individual F in the flame individual set is n Denoted as flame individual R i , i=1,2,3…ζ;

[0032] S4: Select moth individuals F from the population set in sequence n , and record its sequential position in the population as j, and judge whether "j>ζ" is true. If "j>ζ" is not true, directly select the flame individual R at the jth position from the flame individual set. i , and recorded as the selected moth individual F n The corresponding target flame individual F n-tar On the contrary, if "j>ζ" does not hold, then according to all moth entities F n The corresponding fitness δ n , use the roulette wheel selection algorithm to select a flame individual R from the flame individual set i The moth individual F was selected n The corresponding target flame individual F n-tar , based on the target flame individual F n-tar For the selected moth individual F n Update using the following formula:

[0033] F n = dis(F n , F n-tar )·exp(ct)·cos(2πt)+F n-tar ;

[0034] where dis(F n , F n-tar ) is the selected moth individual F n With the target flame individual F n-tar The distance between them; c is the shape constant of the logarithmic spiral; t is a random number between the interval [-1, 1];

[0035] S5: The flame individual R in the flame individual set is calculated by the formula: ζ = round[(Nm)(N-1) / M] i The quantity ζ is updated;

[0036] S6: Determine whether "m<M" is true. If "m<M" is true, go to S7; if "m<M" is not true, go to S8;

[0037] S7: Calculate individual moth F n The corresponding fitness δ n , and all moth individuals F n According to the corresponding fitness δ nArrange them from large to small, and arrange all the moth individuals F n Reorganize the population set, and all the moth individuals F in the population set n and all moth individuals F in the flame individual set n According to the corresponding fitness δ n Arrange from large to small and select the first ζ moth individual F n Form a flame individual set and return to S4;

[0038] S8: Select all moth individuals F in the population collection n and all moth individuals F in the flame individual set n Medium fitness δ n The largest moth individual F n Output as the analog switch duty cycle.

[0039] Preferably, establish N moth individuals F n , specifically including the following steps:

[0040] S1.1: For each individual moth F n , the assignment operation is performed through the following formula:

[0041]

[0042] S1.2: Repeat step S1.1 N times to create N moth individuals F n .

[0043] Preferably, the moth individual F is calculated n The corresponding fitness δ n , specifically including the following steps: based on the actual photovoltaic system, establish a corresponding photovoltaic system simulation model, then set the environmental conditions corresponding to each third training sample, and set the moth individual F n The corresponding simulated switch duty cycle is used as a parameter to set the photovoltaic system simulation model. The photovoltaic power generation simulation is performed through the photovoltaic system simulation model, and the output power is used as the fitness δ n ; and the fitness corresponding to the duty cycle of the analog switch tube is the maximum output power.

[0044] A photovoltaic power generation control system, comprising:

[0045] The actual output power monitoring module is used to continuously monitor the actual output power of the photovoltaic system and determine whether the actual output power meets the photovoltaic control restart condition;

[0046] A photovoltaic panel image acquisition module to be detected is used to acquire the photovoltaic panel image to be detected of each photovoltaic panel;

[0047] An environmental feature vector establishment vector is used to establish an environmental feature vector corresponding to each photovoltaic panel image to be detected;

[0048] The environmental feature map establishes a vector, which is used to concatenate all environmental feature vectors to obtain an environmental feature map. Each coordinate point in the environmental feature map corresponds to a photovoltaic panel, and each coordinate point stores the environmental feature vector corresponding to the photovoltaic panel.

[0049] The analog switch duty cycle output module is used to send the environmental feature map to the trained maximum power output model for processing and output the analog switch duty cycle;

[0050] Photovoltaic power generation delay control strategy establishment module, used to output photovoltaic power generation delay control strategy based on analog switch tube duty cycle;

[0051] The maximum output power update module is used to obtain the maximum output power based on the simulated switch tube duty cycle and update the maximum output power based on the simulated maximum output power.

[0052] The present invention has the following advantages:

[0053] 1. The present invention obtains the corresponding lighting conditions by performing image analysis on each photovoltaic panel, generates an environmental feature vector corresponding to each photovoltaic panel based on the lighting conditions, constructs an environmental feature map based on the environmental feature vectors corresponding to all photovoltaic panels, and then predicts the maximum output power of the photovoltaic system under the current environmental conditions based on the environmental feature map, and generates a corresponding photovoltaic power generation delay control strategy, thereby taking into account the lighting conditions of each photovoltaic panel, avoiding the influence of local extreme points on the maximum output power due to local occlusion, reducing dependence on manual intervention, reducing operation and maintenance costs, and improving the automation level of the system. The photovoltaic power generation delay control strategy also avoids the potential impact on system stability caused by rapid or sudden changes.

[0054] 2. In the present invention, the U-net network layer first extracts features from the spotless photovoltaic panel image. The U-net network is selected for its excellent performance in the field of image segmentation. It can effectively extract complex features from the image, which is particularly critical for the detection of light distribution. Subsequently, these features are enhanced through the self-attention mechanism, further improving the model's understanding and analysis capabilities of light distribution features. The self-attention mechanism can highlight important features and suppress unimportant information by calculating the association between the weight matrix and the features, thereby making the light intensity distribution detection more accurate. This mechanism shows its powerful effect, especially when dealing with the problem of uneven lighting caused by sunlight reflection. In the self-attention mechanism layer, the use of the brightness feature map is based on the characteristics of the HSV color space, which effectively utilizes the brightness information of the image and further enhances the recognition ability of the light distribution features. Through this series of steps, the light intensity distribution matrix finally obtained not only accurately reflects the lighting conditions of the photovoltaic panel, but also provides key data for subsequent photovoltaic system optimization.

[0055] 3. The present invention initializes these individuals through Fuch chaotic mapping, ensuring the wide distribution of the initial population in the search space, thus providing a good foundation for global search, and the fitness of each moth individual δ n The calculation is obtained by the photovoltaic system simulation model constructed by MATLAB / SIMULINK software. This simulation model simulates the photovoltaic power generation process according to the changes in environmental conditions such as light and temperature, and uses the output power as a measure of fitness. At the same time, the position of the moth individual, that is, the value of the duty cycle, is continuously optimized through an iterative process. In each iteration, the position of the moth individual is updated according to the distance between the moth individual and the flame individual and other parameters. This update process not only takes into account the current fitness, but also adjusts the size of the flame individual set adaptively to ensure that the algorithm can effectively explore the solution space and accurately use the current optimal solution. It can effectively find the switch duty cycle that maximizes the output power, thereby optimizing the performance of the photovoltaic system. In summary, this application not only improves the power generation efficiency of the photovoltaic system through simulation and iteration methods, but also improves the application flexibility and stability of the algorithm, which is of great significance to improving the overall performance and economy of photovoltaic power generation technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a schematic diagram of the structure of a photovoltaic power generation control system used in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0058] Example 1

[0059] A photovoltaic power generation control method, comprising:

[0060] Continuously monitor the actual output power q of the photovoltaic system and determine whether the photovoltaic control restart condition is met. The photovoltaic control restart condition is "(η-q) / q≥ε", where η is the currently recorded maximum output power, which is subsequently updated through the photovoltaic control restart operation. ε is the power deviation threshold, which is set by the configuration file and is generally 0.5; if the photovoltaic control restart condition is met, execute the photovoltaic control restart operation; if the photovoltaic control restart condition is not met, continue to monitor the output power q of the photovoltaic system; by judging the photovoltaic control restart condition, the preset power deviation threshold is used to ensure that the system always operates at close to optimal performance;

[0061] The PV control restart operation includes the following:

[0062] Obtain the photovoltaic panel image to be detected of each photovoltaic panel. The photovoltaic panel image is an RGB image. It should be noted that the entire photovoltaic power generation array includes multiple photovoltaic panels, each of which can convert light energy into electrical energy. By installing a camera on each photovoltaic panel, the image of the side of the photovoltaic panel that receives light energy can be collected to obtain the photovoltaic panel image to be detected;

[0063] For each photovoltaic panel image to be detected, the image to be detected is processed by the trained image restoration model to obtain a spot-free photovoltaic panel image. When the photovoltaic panel is photographed, the reflection of sunlight will cause glare or light spots in the image of the photovoltaic panel to be detected, which will affect the analysis of the lighting conditions of the photovoltaic panel. Therefore, the photovoltaic panel image to be detected can be processed by the image restoration model, and the glare or light spots are repaired based on the pixels around the position where the glare or light spots appear in the photovoltaic panel image to be detected. The lighting conditions of the spot-free photovoltaic panel image are analyzed, and the light intensity distribution matrix is ​​output; the light intensity distribution matrix is ​​converted into the light intensity distribution matrix. The matrix is ​​processed by a multi-layer perceptron to obtain a light distribution feature vector; the spot-free photovoltaic panel image is subjected to spectral analysis to output a spectral feature vector; the light distribution feature vector, the spectral feature vector and the ambient temperature value are concatenated to obtain an environmental feature vector, which can characterize the current environmental data of the photovoltaic panel and is used to output the subsequent maximum power point; through light intensity analysis and spectral analysis technology, environmental feature vectors reflecting the light and spectral conditions of each photovoltaic panel are generated, which not only contain light intensity information, but also combine factors such as ambient temperature, thereby comprehensively characterizing the working environment of the photovoltaic panel;

[0064] All environmental feature vectors are spliced ​​together to obtain an environmental feature map. Each coordinate point in the environmental feature map corresponds to a photovoltaic panel, and each coordinate point stores the environmental feature vector of the corresponding photovoltaic panel. Splicing all feature vectors into an environmental feature map provides the system with an intuitive and detailed representation of environmental data, which not only helps monitor the performance of each photovoltaic panel, but also provides data support for the management and optimization of the entire photovoltaic system.

[0065] The environmental feature map is sent to the trained maximum power output model for processing, and the simulated switch tube duty cycle is output. The simulated switch tube duty cycle refers to the proportion of the time that the switch tube in the photovoltaic inverter is turned on. It can adjust the output voltage and output current of the photovoltaic system, and then adjust the output power. At the same time, the maximum output power η is obtained based on the simulated switch tube duty cycle; the photovoltaic power generation delay control strategy is output based on the simulated switch tube duty cycle, and the maximum output power η is updated based on the simulated maximum output power; the maximum power output model can accurately predict the ideal output power and optimal switch tube duty cycle under current conditions. This prediction is crucial for dynamically adjusting the work of the inverter to ensure that the system always operates at the highest efficiency.

[0066] The photovoltaic system is controlled based on the photovoltaic power generation delay control strategy, specifically adjusting the conduction time ratio of the switch tube in the photovoltaic inverter in the photovoltaic system, and the photovoltaic power generation delay control strategy allows the system to make gradual adjustments, avoiding the potential impact on system stability caused by rapid or sudden changes. This gradual adjustment mechanism not only improves the system's responsiveness, but also reduces the wear on the inverter and other key components, extending the overall life of the system.

[0067] The present application performs image analysis on each photovoltaic panel to obtain the corresponding lighting conditions, generates an environmental feature vector corresponding to each photovoltaic panel based on the lighting conditions, constructs an environmental feature map based on the environmental feature vectors corresponding to all photovoltaic panels, and then predicts the maximum output power of the photovoltaic system under the current environmental conditions based on the environmental feature map, and generates a corresponding photovoltaic power generation delay control strategy, thereby taking into account the lighting conditions of each photovoltaic panel, avoiding the influence of local extreme points on the maximum output power due to local occlusion, reducing dependence on manual intervention, reducing operation and maintenance costs, and improving the automation level of the system. The photovoltaic power generation delay control strategy also avoids the potential impact on system stability caused by rapid or sudden changes.

[0068] The image restoration model is established based on the generated U-net network. The specific settings can refer to the existing U-net network model. In this embodiment, an attention mechanism is also added to the U-net network.

[0069] The training of the image restoration model includes the following steps:

[0070] A number of photovoltaic panel images are obtained, where the photovoltaic panel images are images actually taken by the user. For each photovoltaic panel image, the photovoltaic panel image and its corresponding standard photovoltaic panel image are combined into a first training sample, where the standard photovoltaic panel image is an image after the user repairs the photovoltaic panel image through software or manually; all the first training samples are combined into a first training sample set, and then the first training sample set is sent to a parameter-initialized image restoration model for training. During the training, the weight parameters are adjusted with the standard photovoltaic panel image as the target to determine whether the first training condition is met. The first training condition may be that a certain number of training times has been reached or the similarity between the image output by the model and the standard photovoltaic panel image reaches a certain threshold, which is not limited here. If the first training condition is met, the trained image restoration model is output; otherwise, iterative training continues through the first training sample set.

[0071] The spot-free photovoltaic panel image is analyzed for illumination conditions and the illumination intensity distribution matrix is ​​output, which specifically includes the following contents:

[0072] The spot-free photovoltaic panel image is sent to the trained light distribution detection model for processing, and the light intensity distribution matrix is ​​output;

[0073] The illumination distribution detection model is established based on the U-net network, including the U-net network layer, the self-attention mechanism layer and the illumination intensity distribution matrix output layer. The U-net network layer is used to extract the illumination distribution features of the spotless photovoltaic panel image, and the parameters of the U-net network layer are set with reference to the existing U-net network; the self-attention mechanism layer is used to enhance the illumination distribution features and improve the accuracy of illumination intensity distribution detection; the illumination intensity distribution matrix output layer is used to output the illumination intensity distribution matrix based on the enhanced illumination distribution features;

[0074] The spot-free photovoltaic panel image is sent to the trained light distribution detection model for processing, which specifically includes the following steps:

[0075] The spot-free photovoltaic panel image is sent to the U-net network layer for processing to obtain the light distribution feature T;

[0076] In the self-attention mechanism layer, the illumination distribution feature T is combined with the key-value weight matrix W k Multiply them together to get the key value matrix K; add the illumination distribution feature T and the value weight matrix W p Multiply them to get the value matrix P; convert the spot-free photovoltaic panel image to HSV space after color conversion. HSV space is a commonly used color model. HSV represents hue, saturation and value. Select the feature map corresponding to the V channel as the brightness feature map, and compare the brightness feature map with the query weight matrix W. QMultiply them to get the query matrix Q and calculate the attention weight matrix ATT = softmax((Q·K T ) / (μ) 0.5 ), where μ is the number of columns of the key-value matrix K; multiply the attention weight matrix ATT by the value matrix P to obtain the enhanced illumination distribution feature T;

[0077] In the light intensity distribution matrix output layer, the enhanced light distribution feature T is convolved with a convolution kernel size of 1×1, and the light intensity distribution matrix is ​​output;

[0078] In this application, the U-net network layer first extracts features from the spotless photovoltaic panel image. The U-net network is selected for its excellent performance in the field of image segmentation. It can effectively extract complex features from the image, which is particularly critical for the detection of light distribution. Subsequently, these features are enhanced through the self-attention mechanism, further improving the model's understanding and analysis capabilities of light distribution features. The self-attention mechanism can highlight important features and suppress unimportant information by calculating the association between the weight matrix and the features, thereby making the light intensity distribution detection more accurate. This mechanism shows its powerful effect, especially when dealing with uneven lighting problems caused by sunlight reflection. In the self-attention mechanism layer, the use of the brightness feature map is based on the characteristics of the HSV color space, which effectively utilizes the brightness information of the image and further enhances the recognition ability of the light distribution features. Through this series of steps, the final light intensity distribution matrix not only accurately reflects the lighting conditions of the photovoltaic panel, but also provides key data for subsequent photovoltaic system optimization.

[0079] The training of the illumination distribution detection model includes the following steps:

[0080] All photovoltaic panel images in the first training sample set are obtained, and all photovoltaic panel images in the first training sample set are processed by the trained image restoration model to obtain a number of spot-free photovoltaic panel images, and each spot-free photovoltaic panel image and its corresponding standard light intensity distribution matrix are combined into a second training sample. The standard light intensity distribution matrix stores the light intensity value obtained by the user through software or actual sensors and standardized. All second training samples are combined into a second training sample set, and then the second training sample set is sent to the parameter-initialized light distribution detection model for training. During the training, the weight parameters are adjusted with the standard light intensity distribution matrix as the target to determine whether the second training condition is met. The second training condition can be that a certain number of training times is reached or the similarity between the matrix output by the model and the standard light intensity distribution matrix reaches a certain threshold, which is not limited here. If the second training condition is met, the trained light distribution detection model is output; otherwise, iterative training is continued through the second training sample set.

[0081] The maximum power output model is established based on a convolutional neural network, and the specific settings refer to the existing convolutional neural network based on regression tasks;

[0082] Training the maximum power output model includes the following steps:

[0083] All photovoltaic panel images acquired in several time periods are obtained, and the environmental feature maps corresponding to all photovoltaic panel images in any time period are used as third training samples, and the third training samples are marked with the analog switch tube duty cycle output by the improved moth-to-flame optimization algorithm; all marked third training samples are combined into a third training sample set, and then the third training sample set is sent to the parameter-initialized maximum power output model for training. During the training, the loss value is calculated with the analog switch tube duty cycle corresponding to the third training sample as the target. If the loss value is within a preset range, the preset range is artificially set to constrain the accuracy of the maximum power output model, and the trained maximum power output model is output; otherwise, the iterative training is continued through the third training sample set.

[0084] The improved moth-to-flame optimization algorithm is used to output the analog switch duty cycle, which specifically includes the following steps:

[0085] S1: Establish N moth individuals F n , n=1,2,3…N, and F n What is stored in is the duty cycle of the switch tube to be simulated;

[0086] S2: Set the maximum number of iterations M, let m = 1, and h is used to record the number of iterations;

[0087] S3: Count individual moths F n The corresponding fitness δ n , and all moth individuals F n According to the corresponding fitness δ n Arrange them from large to small, and arrange all the moth individuals F n Form a population set and select the first ζ moth individual F n The flame individual set is composed. In the initial state, ζ=N. The moth individual F in the flame individual set is n Denoted as flame individual R i , i=1,2,3…ζ;

[0088] S4: Select moth individuals F from the population set in sequence n , and record its sequential position in the population as j, and judge whether "j>ζ" is true. If "j>ζ" is not true, directly select the flame individual R at the jth position from the flame individual set. i , and recorded as the selected moth individual F nThe corresponding target flame individual F n-tar On the contrary, if "j>ζ" does not hold, then according to all moth entities F n The corresponding fitness δ n , use the roulette wheel selection algorithm to select a flame individual R from the flame individual set i The moth individual F was selected n The corresponding target flame individual F n-tar , based on the target flame individual F n-tar For the selected moth individual F n Update using the following formula:

[0089] F n = dis(F n , F n-tar )·exp(ct)·cos(2πt)+F n-tar ;

[0090] where dis(F n , F n-tar ) is the selected moth individual F n With the target flame individual F n-tar The distance between them; c is the shape constant of the logarithmic spiral; t is a random number between the interval [-1, 1];

[0091] S5: The flame individual R in the flame individual set is calculated by the formula: ζ = round[(Nm)(N-1) / M] i The quantity ζ is updated;

[0092] S6: Determine whether "m<M" is true. If "m<M" is true, go to S7; if "m<M" is not true, go to S8;

[0093] S7: Calculate individual moth F n The corresponding fitness δ n , and all moth individuals F n According to the corresponding fitness δ n Arrange them from large to small, and arrange all the moth individuals F n Reorganize the population set, and all the moth individuals F in the population set n and all moth individuals F in the flame individual set n According to the corresponding fitness δ n Arrange from large to small and select the first ζ moth individual F n Form a flame individual set and return to S4;

[0094] S8: Select all moth individuals F in the population set n and all moth individuals F in the flame individual setn Medium fitness δ n The largest moth individual F n Output as the analog switch duty cycle.

[0095] Preferably, establish N moth individuals F n , specifically including the following steps:

[0096] S1.1: For each individual moth F n , the assignment operation is performed through the following formula:

[0097]

[0098] By using the Fuch chaotic map (the above formula) to initialize each moth individual F, it has the advantages of being insensitive to initial values, strong ergodicity, and fast convergence. n , so that the initial moth individuals are distributed more evenly in the search space, laying the foundation for the wide-area search of the initial iteration of the population;

[0099] S1.2: Repeat step S1.1 N times to create N moth individuals F n .

[0100] Preferably, the moth individual F is calculated n The corresponding fitness δ n The specific steps include: on the basis of the actual photovoltaic system, the corresponding photovoltaic system simulation model is established by using MATLAB / SIMULINK software, and then the environmental conditions corresponding to each third training sample are set, including light and temperature, etc., and the moth individual F n The corresponding simulated switch duty cycle is used as a parameter to set the photovoltaic system simulation model. The photovoltaic power generation simulation is performed through the photovoltaic system simulation model, and the output power is used as the fitness δ n ; and the fitness corresponding to the duty cycle of the analog switch tube is the maximum output power. It should be noted that the simulation of the duty cycle of the analog switch tube is in the training stage of the maximum power output model, not in the actual application stage, so there is sufficient freedom in setting the environmental conditions.

[0101] This application initializes these individuals through Fuch chaotic mapping, ensuring the wide distribution of the initial population in the search space, thus providing a good foundation for global search, and the fitness of each moth individual δ nThe calculation is obtained by the photovoltaic system simulation model constructed by MATLAB / SIMULINK software. This simulation model simulates the photovoltaic power generation process according to the changes in environmental conditions such as light and temperature, and uses the output power as a measure of fitness. At the same time, the position of the moth individual, that is, the value of the duty cycle, is continuously optimized through an iterative process. In each iteration, the position of the moth individual is updated according to the distance between the moth individual and the flame individual and other parameters. This update process not only takes into account the current fitness, but also adjusts the size of the flame individual set adaptively to ensure that the algorithm can effectively explore the solution space and accurately use the current optimal solution. It can effectively find the switch duty cycle that maximizes the output power, thereby optimizing the performance of the photovoltaic system. In summary, this application not only improves the power generation efficiency of the photovoltaic system through simulation and iteration methods, but also improves the application flexibility and stability of the algorithm, which is of great significance to improving the overall performance and economy of photovoltaic power generation technology.

[0102] Example 2

[0103] A photovoltaic power generation control system, see Figure 1 ,include:

[0104] The actual output power monitoring module is used to continuously monitor the actual output power of the photovoltaic system and determine whether the actual output power meets the photovoltaic control restart condition;

[0105] A photovoltaic panel image acquisition module to be detected is used to acquire the photovoltaic panel image to be detected of each photovoltaic panel;

[0106] An environmental feature vector establishment vector is used to establish an environmental feature vector corresponding to each photovoltaic panel image to be detected;

[0107] The environmental feature map establishes a vector, which is used to concatenate all environmental feature vectors to obtain an environmental feature map. Each coordinate point in the environmental feature map corresponds to a photovoltaic panel, and each coordinate point stores the environmental feature vector corresponding to the photovoltaic panel.

[0108] The analog switch duty cycle output module is used to send the environmental feature map to the trained maximum power output model for processing and output the analog switch duty cycle;

[0109] Photovoltaic power generation delay control strategy establishment module, used to output photovoltaic power generation delay control strategy based on analog switch tube duty cycle;

[0110] The maximum output power update module is used to obtain the maximum output power based on the simulated switch tube duty cycle and update the maximum output power based on the simulated maximum output power.

[0111] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A photovoltaic power generation control method, characterized in that: include: Continuously monitor the actual output power q of the photovoltaic system and determine whether the photovoltaic control restart condition is met. The photovoltaic control restart condition is "(η-q) / q≥ε", where η is the current maximum output power recorded, which is subsequently updated through the photovoltaic control restart operation, and ε is the power deviation threshold; if the photovoltaic control restart condition is met, execute the photovoltaic control restart operation; if the photovoltaic control restart condition is not met, continue to monitor the output power q of the photovoltaic system; The PV control restart operation includes the following: Obtaining the photovoltaic panel image to be inspected of each photovoltaic panel, where the photovoltaic panel image is an RGB image; For each photovoltaic panel image to be detected, the image to be detected is processed by the trained image restoration model to obtain a spot-free photovoltaic panel image, and the spot-free photovoltaic panel image is subjected to illumination condition analysis to output an illumination intensity distribution matrix; the illumination intensity distribution matrix is ​​processed by a multi-layer perceptron to obtain an illumination distribution feature vector; the spot-free photovoltaic panel image is subjected to spectral analysis to output a spectral feature vector; the illumination distribution feature vector, the spectral feature vector and the ambient temperature value are concatenated to obtain an ambient feature vector; All environmental feature vectors are concatenated to obtain an environmental feature map, in which each coordinate point corresponds to a photovoltaic panel, and each coordinate point stores the environmental feature vector corresponding to the photovoltaic panel; The environmental feature map is sent to the trained maximum power output model for processing, and the simulated switch tube duty cycle is output. The simulated switch tube duty cycle refers to the proportion of the time that the switch tube in the photovoltaic inverter is turned on. It can adjust the output voltage and output current of the photovoltaic system, and then adjust the output power. At the same time, the maximum output power η is obtained based on the simulated switch tube duty cycle; the photovoltaic power generation delay control strategy is output based on the simulated switch tube duty cycle, and the maximum output power η is updated based on the simulated maximum output power; The photovoltaic system is controlled based on the photovoltaic power generation delay control strategy.

2. A photovoltaic power generation control method according to claim 1, characterized in that: The image restoration model is built based on the generated U-net network; The training of the image restoration model includes the following steps: Acquire a number of photovoltaic panel images, where the photovoltaic panel images are images actually taken by a user, and for each photovoltaic panel image, combine the photovoltaic panel image and its corresponding standard photovoltaic panel image to form a first training sample; All first training samples are combined into a first training sample set, and then the first training sample set is sent to a parameter-initialized image restoration model for training. During the training, the weight parameters are adjusted with the standard photovoltaic panel image as the target to determine whether the first training condition is met. If the first training condition is met, the trained image restoration model is output; otherwise, iterative training is continued through the first training sample set.

3. A photovoltaic power generation control method according to claim 2, characterized in that: The spot-free photovoltaic panel image is analyzed for illumination conditions and the illumination intensity distribution matrix is ​​output, which specifically includes the following contents: The spot-free photovoltaic panel image is sent to the trained light distribution detection model for processing, and the light intensity distribution matrix is ​​output; The illumination distribution detection model is established based on the U-net network, including the U-net network layer, the self-attention mechanism layer and the illumination intensity distribution matrix output layer. The U-net network layer is used to extract the illumination distribution features of the spotless photovoltaic panel image; the self-attention mechanism layer is used to enhance the illumination distribution features; The light intensity distribution matrix output layer is used to output the light intensity distribution matrix based on the enhanced light distribution features.

4. A photovoltaic power generation control method according to claim 3, characterized in that: The spot-free photovoltaic panel image is sent to the trained light distribution detection model for processing, which specifically includes the following steps: The spot-free photovoltaic panel image is sent to the U-net network layer for processing to obtain the light distribution feature T; In the self-attention mechanism layer, the illumination distribution feature T is combined with the key-value weight matrix W k Multiply them together to get the key value matrix K; add the illumination distribution feature T and the value weight matrix W p Multiply them to get the value matrix P; convert the spot-free photovoltaic panel image to HSV space through color conversion, select the feature map corresponding to the V channel as the brightness feature map, and compare the brightness feature map with the query weight matrix W Q Multiply them to get the query matrix Q and calculate the attention weight matrix ATT = softmax((Q·K T ) / (μ) 0.5 ), where μ is the number of columns of the key-value matrix K; multiply the attention weight matrix ATT by the value matrix P to obtain the enhanced illumination distribution feature T; In the light intensity distribution matrix output layer, the enhanced light distribution feature T is convolved to output the light intensity distribution matrix.

5. A photovoltaic power generation control method according to claim 4, characterized in that: The training of the illumination distribution detection model includes the following steps: All photovoltaic panel images in the first training sample set are obtained, and all photovoltaic panel images in the first training sample set are processed by the trained image restoration model to obtain a number of spot-free photovoltaic panel images, each spot-free photovoltaic panel image and its corresponding standard light intensity distribution matrix form a second training sample, all second training samples form a second training sample set, and then the second training sample set is sent to the parameter-initialized illumination distribution detection model for training. During the training, the weight parameters are adjusted with the standard light intensity distribution matrix as the target to determine whether the second training condition is met. If the second training condition is met, the trained illumination distribution detection model is output; otherwise, iterative training is continued through the second training sample set.

6. A photovoltaic power generation control method according to claim 5, characterized in that: The maximum power output model is established based on convolutional neural network; Training the maximum power output model includes the following steps: All photovoltaic panel images acquired in several time periods are obtained, and the environmental feature maps corresponding to all photovoltaic panel images in any time period are used as third training samples, and the third training samples are marked with the analog switch tube duty cycle output by the improved moth-to-flame optimization algorithm; all marked third training samples are combined into a third training sample set, and then the third training sample set is sent to the parameter-initialized maximum power output model for training. During the training, the loss value is calculated with the analog switch tube duty cycle corresponding to the third training sample as the target. If the loss value is within a preset range, the preset range is artificially set to constrain the accuracy of the maximum power output model, and the trained maximum power output model is output; otherwise, the iterative training is continued through the third training sample set.

7. A photovoltaic power generation control method according to claim 6, characterized in that: The improved moth-to-flame optimization algorithm is used to output the analog switch duty cycle, which specifically includes the following steps: S1: Establish N moth individuals F n , n=1,2,3...N, and F n What is stored in is the duty cycle of the switch tube to be simulated; S2: Set the maximum number of iterations M, let m = 1, and h is used to record the number of iterations; S3: Count individual moths F n The corresponding fitness δ n , and all moth individuals F n According to the corresponding fitness δ n Arrange them from large to small, and arrange all the moth individuals F n Form a population set and select the first ζ moth individual F n The flame individual set is composed. In the initial state, ζ=N. The moth individual F in the flame individual set is n Denoted as flame individual R i , i = 1, 2, 3 ... ζ; S4: Select moth individuals F from the population set in sequence n , and record its sequential position in the population as j, and judge whether "j>ζ" is true. If "j>ζ" is not true, directly select the flame individual R at the jth position from the flame individual set. i , and recorded as the selected moth individual F n The corresponding target flame individual F n-tar On the contrary, if "j>ζ" does not hold, then according to all moth entities F n The corresponding fitness δ n , use the roulette wheel selection algorithm to select a flame individual R from the flame individual set i The moth individual F was selected n The corresponding target flame individual F n-tar , based on the target flame individual F n-tar For the selected moth individual F n Update using the following formula: F n =dis(F n ,F n-tar )·exp(ct)·cos(2πt)+F n-tar ; where dis(F n , F n-tar ) is the selected moth individual F n With the target flame individual F n-tar The distance between them; c is the shape constant of the logarithmic spiral; t is a random number between the interval [-1, 1]; S5: The flame individual R in the flame individual set is calculated by the formula: ζ = round[(Nm)(N-1) / M] i The quantity ζ is updated; S6: Determine whether "m<M" is true. If "m<M" is true, go to S7; if "m<M" is not true, go to S8; S7: Calculate individual moth F n The corresponding fitness δ n , and all moth individuals F n According to the corresponding fitness δ n Arrange them from large to small, and arrange all the moth individuals F n Reorganize the population set, and all the moth individuals F in the population set n and all moth individuals F in the flame individual set n According to the corresponding fitness δ n Arrange from large to small and select the first ζ moth individual F n Form a flame individual set and return to S4; S8: Select all moth individuals F in the population collection n and all moth individuals F in the flame individual set n Medium fitness 6 n The largest moth individual F n Output as the analog switch duty cycle.

8. A photovoltaic power generation control method according to claim 7, characterized in that: Establish N moth individuals F n , specifically including the following steps: S1.1: For each individual moth F n , the assignment operation is performed through the following formula: S1.2: Repeat step S1.1 N times to create N moth individuals F n .

9. A photovoltaic power generation control method according to claim 8, characterized in that: Calculate the moth individual F n The corresponding fitness δ n , specifically including the following steps: on the basis of the photovoltaic system actually used, a corresponding photovoltaic system simulation model is established, and then the environmental conditions corresponding to each third training sample are set, and the moth individual F n The corresponding simulated switch duty cycle is used as a parameter to set the photovoltaic system simulation model. The photovoltaic power generation simulation is performed through the photovoltaic system simulation model, and the output power is used as the fitness δ n ; and the fitness corresponding to the duty cycle of the analog switch tube is the maximum output power.

10. A photovoltaic power generation control system, characterized in that: The system applies a photovoltaic power generation control method according to any one of claims 1 to 9, including: The actual output power monitoring module is used to continuously monitor the actual output power of the photovoltaic system and determine whether the actual output power meets the photovoltaic control restart condition; A photovoltaic panel image acquisition module to be detected is used to acquire the photovoltaic panel image to be detected of each photovoltaic panel; An environmental feature vector establishment vector is used to establish an environmental feature vector corresponding to each photovoltaic panel image to be detected; The environmental feature map establishes a vector, which is used to concatenate all environmental feature vectors to obtain an environmental feature map. Each coordinate point in the environmental feature map corresponds to a photovoltaic panel, and each coordinate point stores the environmental feature vector corresponding to the photovoltaic panel. The analog switch duty cycle output module is used to send the environmental feature map to the trained maximum power output model for processing and output the analog switch duty cycle; Photovoltaic power generation delay control strategy establishment module, used to output photovoltaic power generation delay control strategy based on analog switch tube duty cycle; The maximum output power update module is used to obtain the maximum output power based on the simulated switch tube duty cycle and update the maximum output power based on the simulated maximum output power.

Images