Distributed power supply capacity and position dynamic calculation method based on machine vision

Abstract

The invention discloses a distributed power supply capacity and position dynamic calculation method based on machine vision, and the method comprises the steps: S1, obtaining and fusing multi-source heterogeneous data, and constructing a multi-dimensional digital resource library; s2, the remote sensing image is finely processed, and the visual features of the distributed power supply are highlighted; s3, a Mask R-CNN deep learning model is adopted, and a pixel-level contour of each distributed power supply is accurately identified and segmented; s4, automatically converting the pixel-level contour into the capacity and the position of the distributed power supply through a multi-dimensional digital resource library and a geographic coordinate conversion technology; and S5, repeating the steps S1 to S4 to obtain a new information base of the distributed power supplies containing capacity and position information, intelligently comparing the distributed power supplies in the new information base and the old information base by using a change detection algorithm, and updating machine account information. Advanced machine vision and deep learning technologies are introduced to automatically calculate the capacity and the position of the distributed power supply, and the intelligent level of power grid operation and the overall power supply reliability are improved.

Description

Technical Field

[0001] This invention relates to the fields of power grid management and machine vision technology, and in particular to a method for dynamically calculating the capacity and location of distributed power sources based on machine vision. Background Technology

[0002] Against the backdrop of the rapid development of the distributed photovoltaic industry, power grid companies and related management departments have an increasingly urgent need for efficient management of distributed power sources. However, the existing management system still has many problems, which seriously restrict the healthy development of distributed power sources and the level of refined management of the power grid.

[0003] First, data collection and updating are inefficient. Currently, distributed power generation management relies primarily on manual on-site surveys and unstructured report compilation. This is not only costly in terms of manpower, but also requires survey personnel to verify installation locations, equipment models, and installed capacities one by one, resulting in a massive workload. Furthermore, the data collection and processing cycle is long, often with delays of several months or even half a year between project registration and data entry. This leads to a serious disconnect between grid ledger data and actual conditions, directly impacting grid load forecasting, power flow analysis, and safety verification, thus creating potential safety hazards for grid planning and operation.

[0004] Second, the accuracy and automation levels of current methods fall short of the requirements. Existing survey methods based on remote sensing imagery are technically immature, relying heavily on simple image processing techniques such as threshold segmentation and edge detection. These methods struggle to handle complex roof shapes, shading, and interference from varying lighting conditions. For example, in urban areas with numerous high-rise buildings, building shadows, tree shading, and the reflective effect of photovoltaic panels can severely interfere with traditional algorithms, leading to missed and false alarms. More importantly, current technologies lack a systematic automation solution. They cannot automatically correlate and estimate the identified roof outlines with key power grid parameters such as capacity and location, still requiring significant manual intervention for data conversion and processing. The level of automation falls far short of application requirements.

[0005] Third, the survey methods lack dynamic monitoring capabilities. Traditional survey methods are static, and the results of a one-time survey only reflect the resource situation at a specific point in time. However, the installation, dismantling, and upgrading of distributed photovoltaic power is a continuously changing process. Existing technologies cannot periodically and automatically monitor and update the distributed power sources in a region, making it difficult for the power grid to keep track of the increase or decrease in power sources in real time and to accurately assess the carrying capacity of distributed power sources. If a large number of distributed power sources are connected haphazardly, it may cause safety risks such as local voltage exceeding limits and line overload, affecting the stable operation of the power grid. Fourth, integrating multi-source heterogeneous data presents significant challenges. Distributed power generation information sources encompass imagery data, statistical reports, geographic information data, and meteorological data, which differ significantly in format, coordinate system, and time dimension, lacking a unified fusion standard. Traditional methods require separate processing and analysis of the data, failing to achieve synergistic information enhancement. This makes it difficult for grid managers to assess the potential of distributed power generation from a holistic perspective, and decision analysis lacks unified and comprehensive data support.

[0006] The statements herein provide only background information in relation to this invention and do not necessarily constitute prior art. Summary of the Invention

[0007] The purpose of this invention is to automate the calculation of the capacity and location of distributed power sources by introducing advanced machine vision and deep learning technologies. This addresses existing technical problems such as low data acquisition efficiency, insufficient identification accuracy, lack of dynamic monitoring, and difficulty in integrating multi-source heterogeneous data. The invention provides power grid companies with an efficient and intelligent management tool for distributed power sources, thereby improving the intelligence level of power grid operation and the overall reliability of power supply.

[0008] To achieve the above objectives, this invention provides a machine vision-based method for dynamically calculating the capacity and location of distributed power sources, the method comprising: Step S1: Acquire and fuse multi-source heterogeneous data to construct a multi-dimensional digital resource library; the multi-source heterogeneous data includes remote sensing images, unstructured distributed power source statistics, and geographic information data; Step S2: Refine the remote sensing images to highlight the visual features of the distributed power sources; Step S3: Use the Mask R-CNN deep learning model to accurately identify and segment the pixel-level contours of each distributed power source; Step S4: Automatically convert the pixel-level contours into the capacity and location of the distributed power sources using the multi-dimensional digital resource library and geographic coordinate transformation technology; Step S5: Repeat steps S1 to S4 to obtain a new distributed power source information library containing capacity and location information, intelligently compare the distributed power sources in the new information library and the old information library using a change detection algorithm, and update the ledger information.

[0009] Optionally, in step S1, the method for fusing multi-source heterogeneous data includes: S11, Georegistration and Calibration: Unify the multi-source heterogeneous data into the same geographic coordinate system; use georegistration technology to establish a mapping relationship between pixel coordinates and geographic coordinates by identifying ground control points on the remote sensing image; S12, Graph-data association model construction: Data construction is carried out on the roof visual features in the remote sensing image. Geocoding service is used to convert the addresses in the unstructured distributed power statistics data into latitude and longitude coordinates. Based on graph attention network, the roof visual features in the remote sensing image and the unstructured distributed power statistics data are deeply associated and fused.

[0010] Optionally, in step S2, the method for fine-tuning includes: S21, Image Correction: Perform geometric correction, radiometric correction, and atmospheric correction on the remote sensing image to ensure the accuracy and consistency of the data; S22, Image Enhancement: An adaptive histogram equalization algorithm is used to enhance the contrast of local areas, highlighting the detailed differences between the distributed power source and the surrounding environment. S23, Denoising Filtering: Using median filtering techniques to eliminate random noise.

[0011] Optionally, after step S2 and before step S3, the method further includes constructing the Mask R-CNN deep learning model, comprising: 1) A backbone network is used to extract deep features from the remote sensing images; 2) Input the deep features into the feature pyramid network to generate a feature pyramid with multi-scale features; 3) At each level of the feature pyramid, a series of candidate regions are generated using a region proposal network; 4) In the RoIAlign module, the candidate regions are mapped onto the feature map, and bilinear interpolation is used for precise alignment; 5) Output the contents of the RoIAlign module for bounding box regression, category classification, and mask generation.

[0012] Optionally, after the Mask R-CNN deep learning model is built, it needs to be optimized, including the following methods: 1) Initialize the backbone network and feature pyramid network using weights pre-trained on a large general dataset; 2) The large general dataset is expanded with various transformations to enhance its adaptability to different shooting conditions and equipment configurations.

[0013] Optionally, in step S4, the method for automatically converting the pixel-level contour into the location of the distributed power source is as follows: using the geographic coordinate transformation technology, the pixel coordinates are accurately converted into the geographic coordinates of the distributed power source using image metadata in the remote sensing image.

[0014] Optionally, the image metadata can be either affine transformation parameters or a rational function model.

[0015] Optionally, in step S4, the method for automatically converting the pixel-level contour into the capacity of the distributed power source is as follows: converting the pixel-level contour into the actual physical area of ​​the distributed power source using a geometric calculation method, and automatically estimating the capacity of the distributed power source using a parameterized model in the multi-dimensional digital resource library.

[0016] Optionally, in step S5, the method for intelligently identifying newly added, removed, or changed distributed power sources using the change detection algorithm includes: S51, if the distributed power source does not exist in the old information database but exists in the new information database, it is determined to be a newly added device; S52, if the distributed power source exists in the old information database but not in the new information database, it is determined to be a removed device; S53. If a distributed power source exists in both databases, but its capacity or area data has changed significantly, it is determined to be a changed device.

[0017] Optionally, in step S5, the change detection algorithm is either an object-based support vector machine classification algorithm or an object-based random forest classification algorithm.

[0018] Compared with the prior art, the technical solution of the present invention has at least the following beneficial effects: 1) By using machine vision and deep learning technologies to replace manual surveys, the efficiency of data collection and processing can be increased by more than 90%, significantly reducing labor costs.

[0019] 2) An optimized Mask R-CNN deep learning model is adopted to achieve pixel-level instance segmentation, effectively cope with complex roof shapes and occlusions, automatically calculate the capacity and location of distributed power sources, and significantly improve recognition accuracy.

[0020] 3) Through periodic image acquisition and change detection, dynamic and real-time monitoring of the increase, decrease and changes of distributed power sources is realized, providing timely and accurate data support for power grid load forecasting and carrying capacity analysis.

[0021] 4) By constructing a multi-dimensional digital resource library, it effectively integrates heterogeneous data from multiple sources such as images, geographic information, and statistical reports, providing a global and three-dimensional decision-making perspective for power grid management and improving the accuracy and practicality of the evaluation model. Attached Figure Description

[0022] Figure 1 This is a flowchart of the machine vision-based distributed power supply capacity and location dynamic calculation method of the present invention.

[0023] Figure 2 This is a schematic diagram illustrating the construction of the Mask R-CNN deep learning model of the present invention. Detailed Implementation

[0024] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, further illustrates the machine vision-based distributed power supply capacity and location dynamic calculation method proposed in this invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clearly illustrate the embodiments of this invention. Please refer to the accompanying drawings to make the objectives, features, and advantages of this invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for illustrative purposes to aid those skilled in the art and are not intended to limit the implementation conditions of this invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.

[0025] With the transformation of the global energy structure and the rapid development of new power systems, distributed photovoltaic (PV) power, as an important component of clean energy, is rapidly expanding in scale. However, existing distributed power generation management systems suffer from numerous problems, including low data acquisition efficiency, insufficient identification accuracy, lack of dynamic monitoring, and difficulties in integrating heterogeneous data from multiple sources. To address these issues, this invention introduces advanced machine vision and deep learning technologies to directly convert visual information into precise capacity and location data. This enables automated, high-precision identification and parameter extraction of distributed power sources in remote sensing imagery, supporting dynamic updates. This provides power grid companies with an efficient and intelligent distributed power generation management tool, thereby improving the intelligence level of power grid operation and overall power supply reliability.

[0026] like Figure 1 As shown, this invention provides a machine vision-based method for dynamically calculating the capacity and location of distributed power sources. This method includes: Step S1: Acquire and merge multi-source heterogeneous data to construct a multi-dimensional digital resource library.

[0027] Within the target area, multi-source heterogeneous data is acquired, including: remote sensing imagery, unstructured distributed power generation statistics, and geographic information data. The remote sensing imagery consists of sub-meter resolution satellite imagery, such as data from open-source or commercial sources like the Gaofen series, Google Earth, or commercial satellites. The unstructured distributed power generation statistics are reports on distributed photovoltaic projects provided by local energy departments or related agencies. These reports are typically in unstructured Excel, CSV, or PDF formats and include text fields such as project address, installed capacity, and owner information. This data also includes grid partitioning and topology data for subsequent analysis and data correlation. The geographic information data includes building vector layers, administrative boundaries, road networks, and other geographic information data for the project area, typically stored in Shapefile or GeoJSON format.

[0028] The above-acquired multi-source heterogeneous data is preprocessed and fused. Specific steps include: 1) Georegistration and Calibration: Unify all data into the same geographic coordinate system (e.g., WGS 84) to eliminate spatial inconsistencies. For remote sensing imagery, georegistration techniques are used to identify ground control points on the imagery and establish a mapping relationship between their pixel coordinates and geographic coordinates. The calibration formula is as follows: in, For geographic coordinates, For pixel coordinates, The transformation coefficients are obtained by least squares fitting.

[0029] 2) Graph-Data Correlation Model Construction: This involves deeply correlating and fusing unstructured distributed power generation statistics (“data”) with rooftop visual features in remote sensing imagery (“graph”). The core steps of model construction include: a. For the graphical data, construct the graph by using the vector data of buildings within the project area as the node set V={v1,v2,…,vn}, where each node vi represents a building. Based on the spatial topological relationships between buildings (e.g., adjacent distances less than a threshold) or power grid topology, construct the edge set E={(vi,vj)} of the graph. Each node vi contains a series of visual features Fvi, extracted from the preprocessed remote sensing imagery (e.g., texture features: gray-level co-occurrence matrix (GLCM) features of the roof surface; color features: mean values ​​of color spaces such as RGB and HSV; shape features: aspect ratio, compactness, etc.).

[0030] b. Use geocoding services to convert addresses in unstructured statistical data into latitude and longitude coordinates to achieve data association. Associate these statistical attributes (such as photovoltaic installed capacity and roof area) as node attributes Avi of the corresponding building node vi in ​​the graph. For buildings without statistical data, their attribute values ​​can be set to zero or undetermined.

[0031] c. Graph-data association is performed based on Graph Attention Network (GAT). By learning the weights of neighbor nodes, neighbor information can be aggregated more effectively, thereby finding the optimal association pattern in complex spatial topology. The update process of the GAT model can be represented as: in: It is the updated feature vector of node i; It is the feature vector of node j; It is the set of neighboring nodes of node i; is the attention coefficient of node j relative to node i, calculated through a shared attention mechanism, which reflects the importance of neighboring nodes to the central node; W is the weight matrix, learned by the model; σ is the activation function.

[0032] After the fusion of multi-source heterogeneous data, a multi-dimensional digital resource library is formed. The core fields of the multi-dimensional digital resource library include: spatial attributes (building vector outline, latitude and longitude coordinates), image features (roof color, texture, brightness, etc.) and structured attributes (installed capacity, owner type).

[0033] Step S2: Refine the remote sensing imagery to highlight the visual characteristics of distributed power sources.

[0034] To refine remote sensing data and highlight key visual features of distributed power sources, the main steps include: 1) Image correction: Perform geometric, radiometric and atmospheric corrections on the original remote sensing images to ensure the accuracy and consistency of the data.

[0035] ① Geometric Correction: This method corrects image geometric distortions caused by factors such as sensor attitude, Earth curvature, and terrain undulation. It employs a rational function model (RPC) for correction, utilizing RPC coefficients from image metadata to accurately map image pixel coordinates to the geographic coordinate system. The basic transformation formula is as follows: Where N and D are the normalized row and column coordinates, respectively, X, Y, Z are the normalized geographic coordinates, and Pi, Qi are the RPC coefficients.

[0036] ②Radiative Correction: Eliminates brightness inconsistencies caused by differences in sensor response, sunlight angle, and changes in atmospheric conditions. By converting the digital grayscale values ​​(DN) of the image into surface reflectance (ρ), brightness standardization is achieved for images from different periods and different sensors.

[0037] Where L is the radiance received by the sensor, and E sun Where θ is the solar irradiance, θ is the solar zenith angle, and T is the solar irradiance. sun and T view These represent the transmittance of sunlight and observed light in the atmosphere, respectively.

[0038] ③ Atmospheric Correction: This eliminates the scattering and absorption effects of atmospheric molecules and aerosols on solar radiation, restoring the true spectral reflectance characteristics of ground objects. The Dark Object Subtraction method can be used, with the following formula: Among them, L DN L represents the original digital grayscale value. MIN E represents the minimum grayscale value of the dark target region in the image. sun θ represents solar irradiance, and θ represents the solar zenith angle.

[0039] 2) Image enhancement and denoising: Based on image correction, further optimize the visual quality of the image and highlight the characteristics of the distributed power source.

[0040] ① Image Enhancement: The Adaptive Histogram Equalization (CLAHE) algorithm is used to enhance contrast in local areas, highlighting the subtle differences between the photovoltaic panel and its surrounding environment, and avoiding over-enhancement that may occur with global equalization. The CLAHE algorithm divides the image into multiple small regions, performs histogram equalization in each region, and then eliminates blockiness through bilinear interpolation.

[0041] ② Noise Reduction Filtering: Median filtering is used to eliminate random noise in the image, such as salt-and-pepper noise. This algorithm replaces the value of the center pixel with the median value of the pixels within the window, effectively removing isolated bright or dark spots while maintaining the sharpness of image edges. The formula is as follows: Where x is the original image, y is the filtered image, and median{} represents the median value of the pixels within the window.

[0042] The process following step S2 and before step S3 includes: constructing and optimizing the Mask R-CNN deep learning model. The construction of the deep learning model uses an optimized Mask R-CNN model as the core recognition tool. This model is a two-stage deep learning framework capable of simultaneously performing object detection and instance segmentation tasks. Its model construction is as follows... Figure 2 As shown, the main steps include: Backbone Network (ResNet): Residual networks such as ResNet-101 or ResNeXt-101 are used to extract deep features from remote sensing images.

[0043] Feature Pyramid Network (FPN): Features extracted from the backbone network are input into the FPN to generate a feature pyramid with multi-scale features, which can effectively deal with photovoltaic panels of different sizes in the image.

[0044] Region Proposal Network (RPN): At each level of the feature pyramid, the RPN generates a series of candidate regions (RoIs), which are regions that may contain photovoltaic panels.

[0045] RoIAlign module: This module maps the candidate regions generated by RPN onto the feature map and uses bilinear interpolation for precise alignment, avoiding the accuracy loss caused by quantization.

[0046] The three-branch parallel prediction head: The output of the RoIAlign module is fed into three branches, which are used for bounding box regression, class classification, and mask generation, respectively.

[0047] After the model was built, transfer learning and data augmentation techniques were used to optimize it and improve its recognition accuracy and generalization ability in complex remote sensing images. Transfer learning involved initializing the backbone network and FPN with weights pre-trained on a large, general dataset (such as COCO). This method effectively utilizes the powerful feature extraction capabilities of the pre-trained model and significantly shortens the model's convergence time. Data augmentation involved expanding the dataset with various transformations before inputting the training data into the model, enhancing its adaptability to different shooting conditions and equipment configurations. Specific methods included: random rotation, flipping, scaling, and other geometric transformations to simulate different shooting angles; random adjustments to brightness, contrast, saturation, and hue (HSV jitter) to simulate different lighting conditions; and simulating random noise in the images to improve the model's robustness against interference.

[0048] In addition, the total loss function (L) for model training is given by and The composition, formula is as follows: in, Including RPN category classification loss ( ) and RPN bounding box regression loss ( ), Including RCNN category classification loss ( ), RCNN bounding box regression loss ( ) and mask generation loss ( Mask generation loss () A binary cross-entropy loss is used to classify pixels within each candidate region to generate a pixel-level segmentation mask. In the above formula, These are hyperparameter weighting coefficients used to balance the contributions of various losses. middle, It is the predicted classification probability of anchor box i. These are the predicted regression parameters. These are the true label regression parameters. middle, yes Predicted class probability, These are real category labels. These are the predicted regression parameters. These are the true regression parameters. It is the predicted value of all pixels in the mask generation branch (i.e. ), It is all the real label values ​​(i.e. ).exist It is the side length of the mask. It is the total number of pixels in the mask. Pixels in the mask The predicted value, It is a pixel The actual label value.

[0049] By constructing and optimizing the Mask R-CNN deep learning model, the pixel-level outline of each distributed power source can be accurately identified and delineated, providing high-precision basic data for subsequent capacity and location calculations.

[0050] To support subsequent automated capacity estimation, the following key parameters can be further extracted automatically during the model identification stage: Distributed power generation (photovoltaic panel) technology types: Through a multi-category classification model, the system can identify and distinguish photovoltaic panels of different technology types, such as monocrystalline silicon, polycrystalline silicon, and thin-film batteries. This information will be directly used for accurate matching of the theoretical peak power per unit area in subsequent capacity estimation.

[0051] Shaded areas: The model can identify and mark the areas on the photovoltaic panel that are shaded by surrounding buildings, trees, or other obstacles. This information will be used for subsequent automated calculations of shadow loss.

[0052] Dirt-covered areas: By analyzing the texture and reflectivity characteristics of the photovoltaic panel surface, the model can identify and mark the areas covered by dust or dirt, providing a basis for the subsequent quantification of dirt loss.

[0053] Photovoltaic array layout: The model can identify the arrangement of photovoltaic panels, such as whether it is distributed or centralized, and the geometry of the array. This information will provide input for subsequent automated matching of cables and mismatch losses.

[0054] Step S3: Using the Mask R-CNN deep learning model, the pixel-level contours of each of the distributed power sources are accurately identified and segmented.

[0055] The optimized Mask R-CNN deep learning model described above can accurately identify and segment the pixel-level contours of each distributed power source.

[0056] Step S4: Using the multi-dimensional digital resource library and geographic coordinate transformation technology, the pixel-level contour is automatically converted into the capacity and location of the distributed power source.

[0057] The machine vision recognition results obtained in step S3 are converted into quantitative data required for power grid management. Through automated methods, the physical area and installed capacity of each distributed power source can be accurately calculated, and its precise geographical location can be determined. Specifically: 1) Pixel coordinate to geographic coordinate conversion: The pixel-level contour (mask) obtained in step S3 is automatically mapped to a precise geographic coordinate system using geographic coordinate transformation technology. Utilizing the affine transformation parameters or rational function models recorded in the image metadata of remote sensing images, a precise conversion from image pixel coordinates to geographic coordinates can be achieved. The conversion formula is as follows: Where (u,v) are pixel coordinates, (Lat,Lon) are the corresponding geographic coordinates, and a,b,c,d,e,f are affine transformation coefficients. For high-precision data, more complex rational function models can be used for transformation to eliminate errors caused by the curvature of the Earth and terrain undulations.

[0058] 2) Precise calculation of physical area: After converting the pixel-level contours obtained in step S3 into vectorized polygons, the actual physical area of ​​each distributed power source is obtained through geometric calculation methods. Detailed steps are as follows: ① The converted set of geographic coordinate points is used to form a polygon.

[0059] ② Calculate the area of ​​the polygon using Green's theorem (also known as the area formula). For a polygon consisting of n vertices (x1, y1), (x2, y2), ..., (xn, yn), its area A can be calculated using the following formula: in,( )=( This method can accurately calculate the physical area of ​​each distributed power source, providing reliable parameters for subsequent capacity estimation.

[0060] 3) Automated Estimation of Installed Capacity: Based on the calculated actual physical area, a parametric model from a multi-dimensional digital resource library is used to automatically estimate the installed capacity of each distributed power source. The core of this parametric model is to directly link the physical area with the power generation potential. Key parameters such as physical area, solar irradiance, and system efficiency are taken into consideration to ensure the accuracy of the estimation results. Detailed estimation steps are as follows: Estimation of installed capacity C. The specific formula is as follows: Where A is the precise physical area of ​​the distributed power source (unit: square meters). The theoretical peak power per unit area represents the maximum output power of a photovoltaic module per unit area under standard test conditions (unit: kW / m²). 2 For mainstream photovoltaic modules, this value is typically around 0.18 kW / m². 2 Up to 0.22 kW / m 2 The specific values ​​are matched based on the identified photovoltaic technology type (such as monocrystalline silicon or polycrystalline silicon). The comprehensive efficiency coefficient is designed to reflect power losses caused by various factors in actual operation. Its calculation formula is as follows: in, Represents temperature loss, This represents the loss of dirt. Represents shadow loss, Represents cable and inverter losses. This represents the loss due to component mismatch.

[0061] Theoretical peak power per unit area ( ), overall efficiency coefficient ( The automated quantification of the comprehensive efficiency coefficient was achieved. To make the estimation more realistic, the various losses in the comprehensive efficiency coefficient were quantified in detail. The key parameters required for the estimation were not manually input, but were achieved through close integration with the machine vision algorithm (steps S2 and S3) to realize the automated flow and acquisition of data.

[0062] In step S3, the Mask R-CNN deep learning model is trained not only to recognize pixel-level contours, but its classification branch is also trained to identify the technology type of photovoltaic modules (such as monocrystalline silicon, polycrystalline silicon, and thin-film). This model achieves automated classification by learning the unique texture, color, and shape features of different types of modules in the image. Using the precise geographic coordinates obtained in step S4, combined with the historical average annual temperature of the project area, and employing a pre-defined temperature-efficiency degradation model, the temperature loss is calculated. In this model, the temperature loss is typically linear, i.e. ,in, Let α be the average annual temperature, and α be the temperature coefficient. The quantification of shadow loss is achieved through an automated model. This automated model utilizes building height data and a solar trajectory model obtained through multi-source heterogeneous data fusion in step S1, combined with the pixel-level contours identified in step S3, to automatically calculate the shadow coverage of distributed power sources at different seasons and times throughout the year. The average annual shadow coverage is taken as the shadow loss and incorporated into the overall efficiency coefficient. Considering that dust, dirt, rain, and snow can easily reduce the power generation efficiency of photovoltaic panels, dirt loss is set according to the climate and environmental characteristics of the project area. By comparing the changes in brightness or reflectivity (which has been radiometrically corrected in step S2) of the same photovoltaic panel at different times and comparing it with historical images in a clean state, the dirt coverage is automatically estimated, thereby obtaining the cable and inverter losses. Considering the power losses caused by cable resistance and inverter conversion efficiency within the photovoltaic system, this loss is set based on a typical distributed photovoltaic system configuration, typically ranging from 3% to 5%. Within the same photovoltaic array, component mismatch losses may occur due to manufacturing differences or varying degrees of aging. Based on the identified size, shape, and layout of the photovoltaic array, a preset library of typical loss parameters is matched. For example, for smaller rooftop photovoltaic arrays, the system automatically matches a set of lower cable and inverter losses; for large, irregular photovoltaic arrays, a set of higher module mismatch losses is matched.

[0063] Step S5: Repeat steps S1 to S4 to obtain a new information database, use a change detection algorithm to intelligently compare the distributed power sources in the new information database and the old information database, and update the ledger information.

[0064] The evaluation process, from steps S1 to S4, is repeated to ensure that the new image data is fully georegistered, corrected, identified, and parameter estimated, thereby generating a completely new distributed power source information database containing capacity and location information. The new database is automatically compared with the old database based on the unique geographic identifier or location coordinates of the distributed power source. A change detection algorithm is used to intelligently identify newly added, removed, or altered distributed power sources. If a distributed power source does not exist in the old database but exists in the new database, it is determined to be a newly added device; if it exists in the old database but not in the new database, it is determined to be a removed device; if it exists in both databases but its capacity or area data has changed significantly, it is determined to be an altered device. To improve the accuracy of change identification, object-based classification algorithms such as support vector machines or random forests are used. This classification algorithm takes the characteristics of photovoltaic panel objects (such as area, shape, and spectral reflectance) at corresponding locations in the old and new images as input, and automatically determines whether their status has changed through model training, thereby filtering out false changes caused by weather or lighting differences.

[0065] Based on the change detection results, the distributed power supply ledger information in the database is automatically updated. For newly added devices, the new data record is inserted into the database. For removed devices, the corresponding device record in the database is marked as removed or archived. For changed devices, the corresponding device capacity, area, and other fields in the database are updated to the latest values.

[0066] In summary, this invention introduces advanced machine vision and deep learning technologies to automatically calculate the capacity and location of distributed power sources, solving existing technical problems such as low data acquisition efficiency, insufficient identification accuracy, lack of dynamic monitoring, and difficulty in integrating multi-source heterogeneous data. It provides power grid companies with an efficient and intelligent management tool for distributed power sources, thereby improving the intelligence level of power grid operation and overall power supply reliability.

[0067] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0068] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for dynamic calculation of distributed power supply capacity and location based on machine vision, characterized in that, The method comprises: Step S1, obtaining and fusing multi-source heterogeneous data to construct a multi-dimensional digital resource library; the multi-source heterogeneous data comprises remote sensing images, unstructured distributed power statistical data, and geographic information data; Step S2, finely processing the remote sensing images to highlight the visual features of the distributed power; Step S3, using a Mask R-CNN deep learning model to accurately identify and segment the pixel-level outline of each distributed power; Step S4, automatically converting the pixel-level outline into the capacity and location of the distributed power through the multi-dimensional digital resource library and geographic coordinate conversion technology; Step S5, repeating steps S1 to S4 to obtain a new information library of distributed power containing capacity and location information, intelligently comparing the distributed power in the new information library and the old information library by using a change detection algorithm, and updating the account information.

2. The method of claim 1, wherein, In step S1, the method for fusing multi-source heterogeneous data comprises: S11, geographic registration and calibration: unify the multi-source heterogeneous data to the same geographic coordinate system; adopt geographic registration technology to establish a mapping relationship between pixel coordinates and geographic coordinates by identifying ground control points on the remote sensing images; S12, graph-data association model construction: data construction is performed on the roof visual features in the remote sensing images, the addresses in the unstructured distributed power statistical data are converted into latitude and longitude coordinates by using geographic coding services, and the roof visual features in the remote sensing images are deeply associated and fused with the unstructured distributed power statistical data based on a graph attention network.

3. The method of claim 1, wherein, In step S2, the method for fine processing comprises: S21, image rectification: geometric correction, radiation correction, and atmospheric correction are performed on the remote sensing images to ensure the accuracy and consistency of the data; S22, image enhancement: an adaptive histogram equalization algorithm is used to enhance the contrast of local areas and highlight the detail differences between the distributed power and the surrounding environment; S23, denoising filtering: random noise is eliminated by using a median filtering technology.

4. The method of claim 1, wherein, After step S2 and before step S3, a Mask R-CNN deep learning model is also constructed, and the method comprises: 1) a backbone network is used to extract deep features from the remote sensing images; 2) the deep features are input into a feature pyramid network to generate a feature pyramid with multi-scale features; 3) a region proposal network is used to generate a series of candidate regions at each level of the feature pyramid; 4) in a RoIAlign module, the candidate regions are mapped onto a feature map and precisely aligned by using bilinear interpolation; 5) the contents in the RoIAlign module are output for bounding box regression, class classification, and mask generation.

5. The method of claim 4, wherein, After the Mask R-CNN deep learning model is constructed, the Mask R-CNN deep learning model needs to be optimized, and the method comprises: 1) the backbone network and the feature pyramid network are initialized by using weights pre-trained on a large general-purpose dataset; 2) The large general dataset is expanded by multiple transformations to enhance the adaptability to different shooting conditions and device morphologies.

6. The method of claim 1, wherein, In step S4, the method for automatically converting the pixel-level contour into the location of the distributed power supply is: using the geographic coordinate conversion technology, and utilizing the image metadata in the remote sensing image to accurately convert the pixel coordinates into the geographic coordinates of the distributed power supply.

7. The method of claim 6, wherein, The image metadata is any one of an affine transformation parameter and a rational function model.

8. The method of claim 1, wherein, In step S4, the method for automatically converting the pixel-level contour into the capacity of the distributed power supply is: converting the pixel-level contour into the actual physical area of the distributed power supply by a geometric calculation method, and automatically estimating the capacity of the distributed power supply by using the parameterized model in the multi-dimensional digital resource library.

9. The method of claim 1, wherein, In step S5, the newly added, removed or changed distributed power supply is intelligently identified by using the change detection algorithm, and the method comprises: S51, if the distributed power supply does not exist in the old information library but exists in the new information library, it is determined as a newly added device; S52, if the distributed power supply exists in the old information library but does not exist in the new information library, it is determined as a removed device; S53, if the distributed power supply exists in both information libraries, but the capacity or area data thereof has a significant change, it is determined as a changed device.

10. The method of claim 1, wherein, In step S5, the change detection algorithm is any one of an object-based support vector machine classification algorithm and an object-based random forest classification algorithm.

Images