Photovoltaic power loss prediction method and system based on multi-mode large-core interactive attention mechanism, terminal and medium

Abstract

The invention discloses a photovoltaic power loss prediction method based on a multi-mode large-kernel interactive attention mechanism, and the method comprises the steps: building a new network model LKFSolarNet for the prediction and estimation of photovoltaic power loss, and proposing a large-kernel interactive attention fusion (LKCA fusion) mechanism for the dynamic feature interaction between a photovoltaic pollution image and an environmental factor. A multi-feature aggregation downsampling (MFAD) method is put forward, a gradient flow enhancement (GFE) training mode is put forward, the above improvement improves the prediction and estimation precision of photovoltaic power loss, enhances the feature capture and fusion utilization capability of the network among different modes, enhances the extraction capability of the backbone network for photovoltaic pollution image features, and improves the prediction and estimation precision of the photovoltaic power loss. The loss of important feature information of the photovoltaic panel is avoided, the fusion effect of LKCA Fusion is further enhanced, and a shallow network is better trained.

Description

Technical Field

[0001] The present invention relates to a photovoltaic power loss prediction method, and in particular to a photovoltaic power loss prediction method based on a multimodal large-core interactive attention mechanism. Background Art

[0002] Existing research on photovoltaic power loss prediction primarily relies on sensor monitoring, data analysis, and machine learning. Sensors are often used to monitor signals using sensing devices, followed by predictions using traditional machine learning or LSTM techniques. However, these methods are susceptible to interference from the external environment on the signal data, resulting in large errors in the prediction results and affecting the accuracy of power loss predictions. In contrast, using image data for power loss prediction can better avoid these interferences. By analyzing the performance of the photovoltaic panel surface under different operating conditions, power loss can be more directly determined, helping maintenance personnel identify the cause of power drops earlier.

[0003] A search of existing technical literature reveals that relevant research primarily focuses on methods for predicting photovoltaic power loss using optical or electrical sensors for data acquisition and analysis. For example, in the Chinese patent application "Multi-time-scale Photovoltaic Power Prediction Method Based on Multiple Modal Decomposition," Hu Keyong et al. proposed a method for predicting photovoltaic system power loss. This method, based on multiple modal decomposition, uses different signal processing methods to decompose historical power data, obtaining two subsequence sets that are added as new features to the data sample set. This effectively avoids the problem of single decomposition methods potentially missing certain modal subsequences prone to aliasing, but still does not address the impact of external interference on power. Separately, in the Chinese patent application "A Distributed Photovoltaic Short-term Power Prediction Method Based on Image Recognition," Zhao Jianjun et al. proposed a power loss prediction method based on image processing. This method uses grayscale processing to simplify data and a specific denoising method to improve image quality. The Canny edge detection algorithm and the gray-level co-occurrence matrix method accurately extract edge and texture features, respectively, thereby extracting image features and integrating power and environmental factors for prediction. However, this method does not propose a deep learning algorithm for the characteristics of photovoltaic panel images, resulting in insufficient accuracy in power loss prediction and limiting the applicability and generalization of the algorithm in some cases. Summary of the Invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a photovoltaic power loss prediction method based on a multimodal large-kernel interactive attention mechanism.

[0005] Technical solution: The photovoltaic power loss prediction method based on the multimodal large-core interactive attention mechanism described in the present invention includes the following steps:

[0006] (1) Data preprocessing: Using the open source dataset Solar Panel Soiling Image Dataset, two groups of identical photovoltaic panels were used as experimental objects. One group was covered and polluted by an obstruction, while the other group was not polluted. The relative power loss of the two groups was obtained. The dataset was divided into a training set and a test set in a ratio of 8:2, and the dataset was divided into two types of datasets with 8 and 16 categories.

[0007] (2) Feature extraction:

[0008] Image feature extraction: MFAD is only added to the shallow downsampling operation of the deep network to balance the network parameter volume and improvement effect. The network completes the image feature extraction through downsampling operation and C3K2 module, and finally uses the multi-head self-attention mechanism in C2PSA to enhance the image features.

[0009] Multi-path processing and enhancement of environmental feature maps: Through encoding, fusion, dimensionality increase and multi-path processing, the environmental feature maps are enhanced and interact with the image feature maps;

[0010] (3) Modal Fusion: Modal fusion is performed on the pollution image of the photovoltaic panel and the environmental factors. The auxiliary gradient flow enhancement GFE branch is directly applied to the backbone network to enhance the image backbone network with gradient flow. In order not to increase the number of network parameters, this branch is only enabled during training and removed during model inference.

[0011] (5) A photovoltaic power loss prediction and estimation method named LKFSolarNet was developed.

[0012] Furthermore, the process of downsampling by MFAD in step (2) is as follows:

[0013] First downsampling: The first downsampling operation is performed by a traditional strided convolution.

[0014] Second downsampling: Send the image to the average pooling kernel to complete the downsampling, and then use the convolution kernel of size 1 to align the channels for subsequent splicing operations.

[0015] The third downsampling operation is completed by a dilation convolution with a dilation rate of 3.

[0016] Feature splicing and compression: The three downsampling operations are spliced ​​in the channel dimension, and the channel is compressed using a convolution kernel of size 1. After completing the BN operation, the current feature is activated to obtain more nonlinear features.

[0017] Furthermore, the multi-path of step (2) includes a large depth-wise separable convolution, a dilated convolution, and a SE attention mechanism. The feature weights generated by the three paths are multiplied together, multiplied with the image feature map Fimg, and a residual connection is constructed to obtain the output feature map F'img of the attention-enhanced image.

[0018] Furthermore, in step (4), in LKFSolarNet, the photovoltaic pollution image will be subjected to Darknet improved by MFAD to extract features, and the environmental factors will be subjected to feature fusion and dimensionality-upgrading encoding by the encoder. Subsequently, the above two features will be fully interactively fused through LKCA Fusion and finally input into the main decoder to complete the power loss prediction and estimation tasks. GFE is only enabled in the training phase to provide additional gradient flow enhancement information for the image backbone network.

[0019] Furthermore, this method uses Huber Loss as the loss function in the prediction regression task and cross entropy loss CE as the loss function in the estimation classification task.

[0020] The specific loss function is defined as shown in formula (1) and (2).

[0021]

[0022] Where y and f(x) represent the true value and predicted value of the sample respectively; δ is a hyperparameter that determines the emphasis of the loss function.

[0023]

[0024] Where C is the number of classification categories; yi is a one-hot encoding of the true label, that is, 1 when the predicted category is consistent and 0 when it is inconsistent; is the probability of the i-th class predicted by the model,

[0025] The total network loss, main loss, and GFE loss are defined as follows:

[0026] L main =Huber main +CE main (3)

[0027] L GFE =Huber GFE +CE GFE (4)

[0028] L All =L main +αL GFE (5)

[0029] Where L represents the total network loss, the main network loss, and the GFE loss, respectively; Huber represents the Huber loss; a is an adjustable parameter that determines the proportion of GFE loss in the total loss and is set to 0.2 in this paper.

[0030] The photovoltaic power loss prediction method based on the multimodal large-kernel interactive attention mechanism of the present invention establishes a new network model LKFSolarNet for the prediction and estimation of photovoltaic power loss, thereby improving the prediction and estimation accuracy of photovoltaic power loss. A large-kernel interactive attention fusion (LKCA Fusion) mechanism is proposed for the dynamic interaction of features between photovoltaic pollution images and environmental factors, which enhances the network's ability to capture and fuse features between different modalities. A multi-feature aggregation downsampling (MFAD) method is proposed to enhance the backbone network's ability to extract features from photovoltaic pollution images, avoid the loss of important feature information of photovoltaic panels, and further enhance the fusion effect of LKCA Fusion. A gradient flow enhancement (GFE) training method is proposed to better train shallow networks.

[0031] Specifically,

[0032] 1) In the field of cross-modal information interaction, the cross-attention mechanism is often used to achieve cross-modal interaction. It captures the interdependencies between different sequences, thereby dynamically adjusting the model's attention to different sequences. Although it has demonstrated excellent cross-modal interaction capabilities in large models, its high computational complexity means that in practice it requires a lot of computing power as a driver, especially in the photovoltaic industry, which significantly increases the cost of algorithm deployment. Large kernel convolution has proven in a large number of tasks that it can also have global feature capture capabilities similar to the Transformer structure, and can achieve performance similar to or even higher than the Transformer model at a relatively low computational complexity. Therefore, inspired by the above two methods, this paper proposes a multimodal information fusion structure based on large convolutional cross-attention, named LKCA Fusion. The specific structure is shown in the figure.

[0033] 2) Photovoltaic panels occupy most of the image area, which exacerbates the impact of the two aforementioned issues. When strided convolution slides on the photovoltaic feature map, the convolution kernel will pass through the photovoltaic panel area more frequently during the sliding process due to the high proportion of the photovoltaic panel area, thus increasing the information loss caused by strided sliding. Due to the high proportion of the photovoltaic panel area, the traditional convolution kernel of size 3 can only focus on the features of a small area of ​​the photovoltaic panel. However, the power loss of the photovoltaic panel depends more on the interpretation of the contextual relationship between the various regions of the photovoltaic panel. In summary, traditional strided convolution downsampling is not suitable for feature sampling of photovoltaic panel contaminated images. Based on the above analysis of the problems of traditional strided convolution in photovoltaic panel contaminated images, a Multi-Feature Aggregation Downsampling module is proposed to remedy the problems of traditional convolution.

[0034] 3) With the development of deep neural networks, network training has become one of the key factors in determining model performance. Deep neural networks often have the problem that as the number of network layers increases, the information gradient flow is increasingly lost or even disappears completely during the backpropagation process. This makes it difficult to effectively train the shallow layers of the network, affecting the model's ability to express information in the shallow layers, resulting in the model failing to perform at its original performance. In this paper's task, modal fusion of contaminated images of photovoltaic panels and environmental factors is required. The extraction of photovoltaic image features by the backbone network directly affects the results of subsequent fusion. Therefore, to further enhance the performance of the network, this paper uses an auxiliary gradient flow enhancement (GFE) branch that acts directly on the backbone network, enabling the image backbone network to be gradient-enhanced. In order not to increase the number of network parameters, this branch is only enabled during training and removed during model inference.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1) Among the existing methods for photovoltaic power loss, most use sensors to monitor signals and obtain numerical power loss data. However, this type of method has a single data source and poor robustness in more complex scenarios (especially when there is pollution on the surface of the photovoltaic panel). With the rapid increase in photovoltaic installed capacity, the demand for photovoltaic panels, which are the core components of the system, has also continued to expand. However, the power generation efficiency of photovoltaic panels is affected by many aspects of the natural environment. Such problems have brought huge challenges to the photovoltaic power generation industry. Among them, since photovoltaic panels are exposed to the atmosphere all year round, pollutants such as dust, leaves and bird droppings accumulated on their surface are one of the most direct factors affecting the efficiency of photovoltaic power generation. Therefore, it is very necessary to combine multimodal information to predict photovoltaic power.

[0037] 2) Compared with the existing technology, the present invention proposes a photovoltaic power loss prediction and estimation method based on large-kernel interactive attention fusion. Based on the characteristics of the high proportion of photovoltaic panel area in photovoltaic images, this method uses the global perception ability of large-kernel convolution and diversified downsampling methods to effectively solve the above two problems, providing new research ideas and methods for the field of photovoltaic power prediction.

[0038] 3) A new network model LKFSolarNet for the prediction and estimation of photovoltaic power loss and a novel training method were established to improve the accuracy of photovoltaic power loss prediction and estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Overall structure diagram of the model network;

[0040] Figure 2 This is the LKCA module structure diagram;

[0041] Figure 3 This is the MFAD structure diagram. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be further described below.

[0043] The overall algorithm flow chart of this embodiment is as follows Figure 1 shown.

[0044] 1) Step 1: The open-source Solar Panel SoilingImage Dataset, proposed by Mehta et al. in 2018, contains 45,754 images. Two identical photovoltaic panels were used as experimental subjects: one was contaminated by obstructions and the other was uncontaminated. The relative power loss of these two panels was then compared. The dataset includes RGB images of photovoltaic panels with varying obstructions, covering different areas and thicknesses, along with the percentage of photovoltaic power loss associated with each image at the time of acquisition. The acquisition timestamp and solar irradiance at the time of acquisition are also included. The dataset was split into a training set and a test set in an 8:2 ratio for training. Due to data acquisition limitations, the discussion in this paper is based on the feature parameters present in the current dataset. For the estimation task, the dataset was divided into two categories: 8 and 16. This means that the estimation task levels are categorized into 8 and 16, corresponding to estimated power loss errors of 12.5% ​​and 6.25%. It is important to note that both the prediction and estimation results output by the model in this paper are expressed as percentages.

[0045] First, in the image extraction backbone, considering the network's lightweight nature and the fact that the receptive field of traditional strided convolution in deep networks is sufficient to cover most areas of the photovoltaic panel, this paper only incorporates MFAD in the shallow downsampling operations of the network to balance the number of network parameters and the improvement effect. The network extracts image features through downsampling and the C3K2 module, and finally enhances these features using the multi-head self-attention mechanism in C2PSA.

[0046] The downsampling process in MFAD consists of three operations. First, a traditional strided convolution is used for the first downsampling operation. Next, the image is passed through an average pooling kernel for downsampling, followed by a convolution kernel of size 1 to align the channels for subsequent concatenation. Finally, a dilated convolution with a dilation rate of 3 is used for the third downsampling operation. These three downsampling operations are concatenated along the channel dimension, and the channels are compressed using a convolution kernel of size 1. After the batch normalization operation, the current feature is activated to obtain more nonlinear features.

[0047] MFAD utilizes an average pooling kernel to better preserve the overall features of the contaminated photovoltaic panel image during the downsampling process, effectively compensating for the information loss caused by the sliding process of traditional strided convolution. MFAD also leverages the characteristics of dilated convolution to provide the network with a larger receptive field. A dilated convolution with a dilation rate of 3 provides a receptive field 5.4 times that of a strided convolution, enabling the downsampling process to capture a larger area of ​​photovoltaic panel feature information, further increasing the proportion of overall photovoltaic panel information in the network.

[0048] 2) In step 2, after encoding and fusing the environmental features, the fused and dimensionally upgraded environmental factor features are converted into a pseudo 3D feature map through a Reshape operation, and the shape of the image feature map is matched to obtain the environmental feature map FE. The environmental feature map FE is then input into a large depth-separable convolution in the height and width directions through the first path for processing. In the second path, the feature map is input into a large depth-separable dilated convolution, thereby enhancing the contextual connection of information at a farther distance. The above two paths can be stacked multiple times as modules to enhance their information extraction capabilities. In the third path, the feature map FE will be input into the SE attention, thereby constructing the mutual dependence relationship between its original feature channels. The feature weights generated by the three paths are multiplied and then multiplied with the image feature map Fimg and a residual connection is constructed to obtain the output feature map F'img of the image after attention enhancement.

[0049] The output feature map F'E of the environmental factors also goes through the above three paths, but it is important to note that it is finally multiplied with the environmental feature map FE, and no residual connection is established. This is because the environmental factors contain a large amount of redundant information after dimensionality increase. If a residual connection is established, the originally obtained enhanced features will have certain interference characteristics.

[0050] LKCA Fusion uses large depthwise separable convolutions to establish weighted relationships between different modalities, making its parameter count far smaller than that of conventional convolution kernels and Transformer structures. Table 1 shows the parameter counts obtained after multiple stacking of LKA in LKCA Fusion. It can be seen that with only one LKA group, LKCA Fusion only requires 0.22M parameters. Adding two LKA groups increases the parameter count by an average of only 0.16M. This demonstrates LKCA Fusion's high lightweight nature, making it easier to deploy on low-computing industrial equipment.

[0051] Table 1 Parameters of different numbers of LKA

[0052]

[0053] 3) In step three, the present invention requires modal fusion of the contaminated images of photovoltaic panels and environmental factors. The extraction of photovoltaic image features by the backbone network directly affects the subsequent fusion results. Therefore, to further enhance network performance, this paper utilizes an auxiliary gradient flow enhancement (GFE) branch directly in the backbone network, thus enhancing the image backbone network with gradient flow. To minimize the increase in network parameters, this branch is only enabled during training and removed during model inference.

[0054] In this paper, Huber Loss is used as the loss function in the prediction regression task to reduce the sensitivity of the loss function to outliers. In the estimation classification task, cross entropy loss CE is used as the loss function. The specific loss function definitions are shown in formulas (1) and (2).

[0055]

[0056] Where y and f(x) represent the true value and predicted value of the sample respectively; δ is a hyperparameter that determines the emphasis of the loss function.

[0057]

[0058] Where C is the number of classification categories; yi is a one-hot encoding of the true label, that is, 1 when the predicted category is consistent and 0 when it is inconsistent; is the probability of the i-th class predicted by the model.

[0059] The total network loss, main loss, and GFE loss are defined as follows:

[0060] L main =Huber main +CE main (3)

[0061] L GFE =Huber GFE +CE GFE (4)

[0062] L All =L main +αL GFE (5)

[0063] Where L represents the total network loss, the main network loss, and the GFE loss, respectively; Huber represents the Huber loss; a is an adjustable parameter that determines the proportion of GFE loss in the total loss and is set to 0.2 in this paper.

[0064] 4) Step 4: Combining the above methods, a photovoltaic power loss prediction and estimation method named LKFSolarNet is formed. Its overall network structure is as follows: Figure 1 As shown in Figure 2, in LKFSolarNet, PV pollution images are extracted using a Darknet modified with MFAD. Environmental factors are then subjected to feature fusion and up-scaling encoding by the encoder. These two features are then interactively fused using LKCA Fusion and fed into the main decoder to complete power loss prediction and estimation. GFE is only enabled during training to provide additional gradient flow enhancement information to the image backbone network.

[0065] The experimental environment of this embodiment: system Ubuntu22.04, graphics card NVIDIA RTX 3090, deep learning framework Pytorch2.0.1. The accuracy of the model proposed in this embodiment is higher than that of any mainstream model.

[0066]

[0067] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other changes to the technical solution and technical content disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. A photovoltaic power loss prediction method based on a multimodal large-core interactive attention mechanism, characterized in that: The following steps are involved: (1) Data preprocessing: Using the open source dataset Solar Panel Soiling Image Dataset, two groups of identical photovoltaic panels were used as experimental objects. One group was covered and polluted by an obstruction, while the other group was not polluted. The relative power loss of the two groups was obtained. The dataset was divided into a training set and a test set in a ratio of 8:2, and the dataset was divided into two types of datasets with 8 and 16 categories. (2) Feature extraction: Image feature extraction: MFAD is only added to the shallow downsampling operation of the deep network to balance the network parameter volume and improvement effect. The network completes the image feature extraction through downsampling operation and C3K2 module, and finally uses the multi-head self-attention mechanism in C2PSA to enhance the image features. Multi-path processing and enhancement of environmental feature maps: Through encoding, fusion, dimensionality increase and multi-path processing, the environmental feature maps are enhanced and interact with the image feature maps; (3) Modal Fusion: Modal fusion is performed on the pollution image of the photovoltaic panel and the environmental factors. The auxiliary gradient flow enhancement GFE branch is directly applied to the backbone network to enhance the image backbone network with gradient flow. In order not to increase the number of network parameters, this branch is only enabled during training and removed during model inference. (4) A photovoltaic power loss prediction and estimation method named LKFSolarNet was developed.

2. The photovoltaic power loss prediction method based on the multimodal large-core interactive attention mechanism according to claim 1 is characterized in that: The process of downsampling by MFAD in step (2) is as follows: First downsampling: The first downsampling operation is performed by a traditional strided convolution. Second downsampling: Send the image to the average pooling kernel to complete the downsampling, and then use the convolution kernel of size 1 to align the channels for subsequent splicing operations. The third downsampling operation is completed by a dilation convolution with a dilation rate of 3. Feature splicing and compression: The three downsampling operations are spliced ​​in the channel dimension, and the channel is compressed using a convolution kernel of size 1. After completing the BN operation, the current feature is activated to obtain more nonlinear features.

3. The photovoltaic power loss prediction method based on multimodal large-core interactive attention mechanism according to claim 1 is characterized in that: The multi-path of step (2) includes large depth-wise separable convolution, dilated convolution and SE attention mechanism. The feature weights generated by the three paths are multiplied together and then multiplied with the image feature map Fimg and a residual connection is constructed to obtain the output feature map F'img of the attention-enhanced image.

4. The photovoltaic power loss prediction method based on multimodal large-core interactive attention mechanism according to claim 1 is characterized in that: In step (4), in LKFSolarNet, the photovoltaic pollution image will be subjected to Darknet improved by MFAD to extract features, and the environmental factors will be subjected to feature fusion and dimensionality up-coding by the encoder. Subsequently, the above two features will be fully interactively fused through LKCA Fusion and finally input into the main decoder to complete the power loss prediction and estimation tasks. GFE is only enabled in the training phase to provide additional gradient flow enhancement information for the image backbone network.

5. The photovoltaic power loss prediction method based on multimodal large-core interactive attention mechanism according to claim 1 is characterized in that: This method uses Huber Loss as the loss function in the prediction regression task and cross entropy loss CE as the loss function in the estimation classification task. The specific loss function is defined as shown in formula (1) and (2). Where y and f(x) represent the true value and predicted value of the sample respectively; δ is a hyperparameter that determines the emphasis of the loss function. Where C is the number of classification categories; yi is a one-hot encoding of the true label, that is, 1 when the predicted category is consistent and 0 when it is inconsistent; is the probability of the i-th class predicted by the model, The total network loss, main loss, and GFE loss are defined as follows: L main =Huber main +CE main (3) L GFE =Huber GFE +CE GFE (4) L All =L main +αL GFE (5) Where L represents the total network loss, the main network loss, and the GFE loss, respectively; Huber represents the Huber loss; a is an adjustable parameter that determines the proportion of GFE loss in the total loss and is set to 0.2 in this paper.

6. A system, terminal or medium applied to the method according to any one of claims 1 to 5.

Images