A distributed photovoltaic cluster power prediction method and system

Abstract

The application provides a kind of distributed photovoltaic cluster power prediction method and system, and the application relates to the technical field of power system automation and new energy power generation prediction, the prediction method of the application specifically includes: obtaining the spatial hash index of the panoramic orthographic image and multi-view image covering the target area;Identify the target position of photovoltaic module on the panoramic orthographic image, solve the attitude parameter of photovoltaic array;Construct source and load characteristic standard library, determine the best matching standard user of the user to be tested;Decouple the data side photovoltaic output value of the user to be tested by using the source and load behavior mode of the best matching standard user, obtain the data side photovoltaic output value;Combine the attitude parameter to calculate the theoretical photovoltaic output value;Get the corrected photovoltaic power sequence and the corrected load sequence;Combine numerical weather prediction to carry out single-household photovoltaic and load prediction, and obtain the power prediction result of distributed photovoltaic cluster through topological aggregation.

Description

Technical Field

[0001] The invention relates to the fields of power system automation and new energy power generation prediction technology, and in particular to a method and system for predicting the power of distributed photovoltaic clusters by integrating multi-source heterogeneous data in low observability environments. Background Technology

[0002] Distributed photovoltaic (PV) power generation systems, represented by rooftop PV, have experienced explosive growth in rural areas, industrial parks, and commercial buildings. Unlike centralized ground-mounted power plants, distributed PV is characterized by its dispersed layout, small capacity per point, large scale, and local consumption. As the penetration rate of distributed PV in distribution networks continues to rise, the randomness, volatility, and intermittency of its output power pose serious challenges to the safe and stable operation of the power grid. When a large number of distributed power sources are connected, the traditional unidirectional power flow distribution network transforms into a bidirectional active power flow distribution network, which can easily lead to problems such as voltage exceeding limits, harmonic pollution, and backflow. Therefore, achieving high-precision power prediction for distributed PV clusters is of paramount importance for power grid dispatching departments to perform peak shaving and valley filling, power balance, and ensure power quality in the distribution network.

[0003] In existing technologies, methods for photovoltaic (PV) power prediction are mainly divided into two categories: physical modeling and statistical learning. Physical modeling primarily relies on numerical weather prediction data and the physical parameters of PV modules to calculate theoretical power generation using photoelectric conversion formulas. However, in actual distributed PV cluster scenarios, especially rural household PV systems, the diverse construction entities and inconsistent construction standards result in a large number of PV modules with varying orientations, tilt angles, and complex installation environments. Grid management systems often lack detailed information on these distributed sites, or the registered information deviates significantly from actual operating conditions. This makes physical modeling methods based on ideal parameters prone to significant errors in practical applications, failing to accurately reflect the true physical conditions of each distributed site.

[0004] On the other hand, statistical learning methods predict power output by establishing a mapping relationship between historical power data and meteorological factors, and widely employ artificial intelligence algorithms such as support vector machines. These methods typically rely on high-density measurement data. However, due to limitations in construction costs and communication conditions, current distributed photovoltaic (PV) clusters often face low observability issues. Most users only install meters at the grid connection point, and the dispatch center can only collect net load or net grid-connected power data after user-side load self-consumption, but cannot directly obtain the raw power generation data from the PV inverters. In this source-load coupled state, the fluctuating characteristics of PV power generation are intertwined with the random characteristics of user electricity consumption behavior, making it difficult for purely data-driven models to extract the true PV output pattern from the net load curve. For example, when net grid-connected power decreases, the model struggles to determine whether the decrease is due to cloud cover or increased self-consumption caused by users activating high-power loads. This lack of transparency in source-load data severely restricts the generalization ability and accuracy of predictive models.

[0005] In summary, existing technologies generally suffer from difficulties in obtaining physical parameters and the inability to observe the actual power output due to source-load data coupling when dealing with distributed photovoltaic clusters. How to mine low-cost incremental feature information and achieve effective decoupling of source-load data and correction of physical states under low observability conditions, thereby improving the accuracy of cluster power prediction, is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0006] This invention discloses a method for predicting the power of a distributed photovoltaic cluster, the prediction method comprising: Obtain the spatial hash index of panoramic orthophotos and multi-view images covering the target area; Identify the target location of photovoltaic modules on panoramic orthophotos, retrieve the corresponding local multi-view images using spatial hash index, and calculate the attitude parameters of the photovoltaic array; A source-load feature standard library is constructed. Based on physical attitude gating and electrical fingerprinting, cross-modal similarity matching is performed between standard users and users under test to determine the best matching standard user for the user under test. By coupling the source-load behavior patterns of the best-matching standard users with the net load data of the users under test at the gate, the photovoltaic output value on the data side is obtained. The theoretical photovoltaic output value is calculated by combining attitude parameters; based on the theoretical photovoltaic output value, the estimated capacity on the data side is iteratively corrected until convergence is obtained to obtain the corrected photovoltaic power sequence and the corrected load sequence. We combine numerical weather prediction to forecast individual photovoltaic (PV) systems and loads, and obtain power forecasts for distributed PV clusters through topology aggregation.

[0007] The spatial hash index for obtaining panoramic orthophotos and multi-view images covering the target area includes: The geographic space of the target area is divided into a regular two-dimensional grid. Establish inverted index mapping function :

[0008] in, Represents grid coordinates, For the first Zhang's original image Covered ground grid set, This is the original image set; When retrieving corresponding local multi-view images using spatial hash indexes, construct a system for the first... A subset of local multi-view images of the target location of a photovoltaic module. :

[0009] in To meet the effective observation threshold A collection of image indexes.

[0010] The calculation of the photovoltaic array's attitude parameters specifically includes: performing feature matching and multi-view triangulation using retrieved local multi-view images to generate a set of local sparse point clouds. ; The photovoltaic plane equation is obtained by fitting a local sparse point cloud set using the random sample consensus algorithm. Optimal normal vector ; Calculating the photovoltaic tilt angle based on the optimal normal vector and photovoltaic azimuth angle :

[0011]

[0012] in, These are the normalized normal vector components.

[0013] Extracting the installation thermal properties and health status features of components, specifically including: calculating the base surface point set in the local sparse point cloud set. Ventilation gap index to the fitted plane And combined with the heat loss correction coefficient of the substrate material Calculate heat dissipation factor :

[0014] in, The normalized ventilation convection score, These are the weighting coefficients; Generate standard front view texture map for photovoltaic panels Extracting the ash accumulation index and color texture entropy Calculate the health status correction factor :

[0015] in, For decay weights, This is the maximum normalized entropy value.

[0016] Based on physical posture gating and electrical fingerprinting, cross-modal similarity matching is performed between standard users and users under test: Calculate the user to be tested With standard users pose difference function Only retain Candidate standard users, among which This is the truncation threshold; Calculate the user to be tested and the candidate standard user Electrical Pearson correlation coefficient and attribute similarity :

[0017] Build the final matching score The user with the highest score is selected as the best matching standard. :

[0018] in, These are the weighting coefficients. To prevent tiny quantities with a denominator of zero.

[0019] The source-load behavior patterns of the best-matched standard users are used to decouple the net load data of the gate users under test. Specifically, this includes: constructing a decoupling objective function and solving for the optimal load scaling factor. :

[0020] Calculate the photovoltaic output value on the data side. :

[0021] in, The net power of the bidirectional meter for the user under test. The estimated capacity for the user to be tested. For standard users, the normalized photovoltaic curve, Normalized load curve for standard users.

[0022] Calculate the theoretical photovoltaic output value The specific formula is as follows: in, To identify the geometric area of ​​the photovoltaic panel, The effective radiation of the inclined plane is calculated based on the attitude angle. For component baseline conversion efficiency, As a heat dissipation factor, This is a correction factor for health status.

[0023] Based on the theoretical photovoltaic power output, the estimated capacity on the data side is iteratively corrected, specifically including: Calculate the full time window Cumulative energy deviation rate between internal data-side photovoltaic output and theoretical photovoltaic output :

[0024] Update the first based on the cumulative energy deviation rate Estimated capacity of the wheel :

[0025] Blind decoupling is re-executed using the updated estimated capacity as a constant until... Output the final corrected photovoltaic power sequence ,in This is the learning rate coefficient. This is the convergence tolerance threshold.

[0026] Until convergence is achieved, the corrected photovoltaic power series and the corrected load series are obtained, specifically including: using the converged estimated capacity. As a fixed known quantity, construct a decoupling objective function constrained by physical capacity:

[0027] Solving for the optimal load scaling factor The corrected photovoltaic power sequence was obtained. and corrected load sequence :

[0028]

[0029] in The net power of the bidirectional meter for the user under test. Normalized load curve for standard users.

[0030] This invention discloses a distributed photovoltaic cluster power prediction system, which includes: Image acquisition module: Acquires spatial hash indices of panoramic orthophotos and multi-view images covering the target area; Attitude calculation module: Identifies the target position of photovoltaic modules on the panoramic orthophoto, retrieves the corresponding local multi-view images using spatial hash index, and calculates the attitude parameters of the photovoltaic array; Matching module: Constructs a source-load feature standard library, and performs cross-modal similarity matching between standard users and users under test based on physical attitude gating and electrical fingerprinting to determine the best matching standard user for the user under test; Estimation module: Couples the source-load behavior patterns of the best-matched standard users with the net load data of the user under test to obtain the photovoltaic output value on the data side; Verification module: Calculates the theoretical photovoltaic output value by combining attitude parameters; Based on the theoretical photovoltaic output value, iteratively corrects the estimated capacity on the data side until convergence to obtain the corrected photovoltaic power sequence and the corrected load sequence; Prediction module: Combines numerical weather forecasting to predict individual photovoltaic loads and obtains power prediction results for distributed photovoltaic clusters through topology aggregation.

[0031] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for predicting the power of a distributed photovoltaic cluster.

[0032] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for predicting the power of a distributed photovoltaic cluster.

[0033] The distributed photovoltaic cluster power prediction method and system provided by this invention significantly solves the problem of prediction distortion caused by strong coupling of source and load data in low observability distribution areas by constructing a closed-loop system that deeply integrates visual perception and electrical data mining, and achieves high-precision power prediction without the need for large-scale installation of sensor hardware.

[0034] Existing technologies, when processing single-meter households with only access meter data, often face mathematical dilemmas due to the lack of measured data from the inverter side. This leads to the source-load separation algorithm encountering multiple solutions, easily misinterpreting random fluctuations in user load as changes in photovoltaic output, resulting in numerical illusions that contradict physical reality. This invention uses photovoltaic module area, orientation, and installation attributes acquired by drones to calculate the theoretical upper limit of the user's power generation capacity under current weather conditions. This physical truth serves as a standard, forcibly constraining the search space of the data-side decoupling algorithm, effectively eliminating the attribution fallacy of interpreting load troughs as power generation peaks, and ensuring that the prediction results always numerically conform to the objective load-bearing capacity of the rooftop physical infrastructure.

[0035] This invention proposes a cross-modal matching mechanism based on physical attitude gating and electrical fingerprinting, which significantly improves the accuracy of feature transfer. In rural power grids lacking full-network topology data, simply relying on geographical distance to find reference users carries significant risks, as geographical proximity does not necessarily imply electrical homogeneity or identical photovoltaic orientation. This invention first ensures the homogeneity of the photovoltaic output waveform between the reference object and the target object through strict physical attitude gating, avoiding timing misalignment caused by orientation differences. Then, it utilizes the high-frequency correlation of voltage sequences to lock in electrical neighbors, ensuring that both are within the same micro-meteorological coverage area and under the same grid voltage background. This dual-screening mechanism allows the behavioral characteristics of standard users to be accurately mapped to the target user, solving the problem of delayed perception due to micro-cloud obstruction caused by the lack of weather stations.

[0036] The dual-channel closed-loop iterative verification system constructed in this invention endows the system with adaptive identification capability for the electrical gain coefficient. The estimated capacity on the data side is a comprehensive electrical parameter that can dynamically absorb the effects of inverter efficiency, line losses, and local obstruction. Through repeated iteration and convergence of the physical and data channels, the system can automatically correct this parameter during operation, gradually bringing it closer to the actual electrical operating value from the initial theoretical value. This means that the system has extremely strong robustness, and even in the face of equipment aging or hidden losses, it can maintain the accuracy of predictions through closed-loop correction. Attached Figure Description

[0037] Figure 1 This is a flowchart of the power prediction method for distributed photovoltaic clusters according to the present invention. Detailed Implementation

[0038] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0039] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0040] Additionally, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that practice can be carried out without these specific details.

[0041] First, this embodiment provides a typical application scenario for the distributed photovoltaic cluster power prediction method, covering rural residential areas with high penetration of distributed power access. In terms of physical spatial distribution, photovoltaic modules are installed on the roofs of each user's own buildings. In rural residential areas, building forms exhibit high heterogeneity, with roof structures including traditional terracotta tile pitched roofs, cast-in-place concrete flat roofs, and later-added corrugated steel tile roofs. Limited by village layout and homestead planning, the geographical orientation of each household's photovoltaic modules is not uniformly aligned with the theoretically optimal power generation direction, i.e., due south. Instead, they are distributed discretely at multiple angles according to the house's orientation, and the tilt angle of the modules varies closely with the roof slope. Furthermore, the vegetation around the modules and the height of adjacent buildings differ among different users. Simultaneously, the module installation process is divided into flat-lay type (close to the roof) and overhead type (lifted using steel supports). Different installation methods result in different ventilation and heat dissipation conditions for the module backsheet, leading to differences in the module's operating temperature and photoelectric conversion efficiency under the same irradiance conditions.

[0042] Regarding the electrical connection topology and power flow relationship, the distributed photovoltaic system in this embodiment is connected to the low-voltage side of the distribution network and generally adopts a self-consumption and surplus power grid connection operation mode. The AC output terminal of the photovoltaic inverter is directly connected to the incoming bus of the user-side distribution box, forming a direct parallel coupling with the user's internal production and living loads (such as rural household appliances and agricultural processing equipment). In terms of power flow, the photovoltaic power generated by the photovoltaic modules is preferentially supplied to the local load. Only when the instantaneous photovoltaic power exceeds the local load power is the surplus power fed back to the public grid through the service line; conversely, when the photovoltaic output is insufficient to support the local load, the shortfall is supplemented by the public grid. In this invention, the power characteristics presented by the user to the grid side are not a single power load or power source, but rather a net power with bidirectional flow characteristics.

[0043] In terms of metering architecture and data observability, this embodiment's scenario is a typical non-full-measuring distribution transformer area. The power grid management department only installs a bidirectional smart meter at the point where the user connects to the public power grid. This meter, serving as the legal basis for settlement, can collect and upload bidirectional active power, effective voltage, and current data at the point of connection at a fixed sampling frequency. Due to limitations in construction costs and equipment maintenance difficulties, most users do not install metering devices with real-time communication capabilities separately at the output end of the photovoltaic inverter or the internal load bus end. This means that external systems can only obtain the net load or net grid-connected power data exchanged between the user and the grid, and cannot directly observe the actual photovoltaic power generation and load consumption power within the user's system. Because the meteorological sensitivity of photovoltaic output and the randomness of user load behavior physically overlap at the point of connection, single point-of-connection metering data cannot directly reflect the true operating status of the photovoltaic system. This strong coupling of source and load data and the black-box state of underlying information are the core technical characteristics of this embodiment's application scenario, and also the main physical constraints faced in subsequent accurate power prediction.

[0044] For the distributed photovoltaic (PV) scenarios in rural areas and industrial parks covered in this application, the distribution of PV modules exhibits significant spatial dispersion and local clustering, meaning they exist only in rooftop areas, while a large number of roads, farmland, and open spaces are invalid areas. If existing technologies are used to perform indiscriminate 3D point cloud reconstruction of the entire area, it will result in a huge waste of computing resources and introduce a large amount of environmental noise. This invention solves the technical problem of how to obtain a global index map at low cost and then establish a multi-view data index based on targeted attitude calculation.

[0045] The objective of this invention is not to directly generate a final 3D model of the photovoltaic cluster, but rather to acquire two types of data: first, high-resolution 2D panoramic orthophotos covering the target village, used to quickly locate the distribution of photovoltaic modules on a 2D plane; second, to establish a spatial-image index database to ensure that any physical point on the panoramic image, i.e., a potential photovoltaic location, can be indexed in the original image set with at least two original images taken from different perspectives. The acquisition of panoramic orthophotos mainly includes: adaptive flight path planning based on binocular vision constraints, constructing a spatial hash index for multi-view images, and rapid panoramic image generation based on weighted homography transformation.

[0046] S1.1: Adaptive route planning based on binocular vision constraints; In rural settings, house roofs vary in height. To ensure subsequent steps can calculate the tilt angle of the photovoltaic panels from different angles, flight path planning must meet the requirement of binocular stereo coverage, meaning any point within the target area... It must appear within the common field of view (FOV) of at least two adjacent images, and the baseline length formed by the centers of these two images must be [missing information]. The accuracy requirements for depth calculation must be met.

[0047] Set the predetermined ground sampling distance Preferably, the ground sampling distance is set. A sampling distance of 2cm / pixel ensures that the edge of the photovoltaic panel can be distinguished. This is based on the camera sensor pixel size. and lens focal length Calculate the relative flight altitude of the drone :

[0048] To meet the requirements of this application for multi-angle acquisition and avoid redundant invalid data, this invention designs a dynamic overlap rate calculation method. Unlike the traditional fixed overlap rate, this invention introduces a three-dimensional base-height ratio constraint. 3D base height ratio constraint The value is between 0.4 and 0.6. Define the heading overlap rate. and lateral overlap rate Minimum conditions to be met: Let the pixel width of the camera sensor in the heading direction be... The pixel height in the lateral direction is Then the ground coverage width .

[0049] To ensure that any point is captured by at least two viewpoints and forms effective stereo vision, the physical baseline length of adjacent exposure points is... Should meet:

[0050] At the same time, the baseline length must satisfy the geometric constraints of stereo matching:

[0051] Combining the above two equations, we can derive the minimum heading overlap rate for route planning in this scenario. The calculation formula is as follows:

[0052] Lateral overlap rate It determines the physical distance between two adjacent parallel flight paths. Addressing the maximum elevation difference of terrain features in rural scenarios. For example, the height difference from the ground to the highest roof ridge line. To prevent image mismatch between adjacent flight strips due to projection differences, i.e., blind spots or gaps, the lateral overlap rate must have sufficient redundancy. The pixel height of the camera sensor in the lateral direction, i.e., perpendicular to the flight direction, is defined as... The effective coverage width of a single image on the ground is... for: To ensure overlap is maintained even at the top of the tallest building, the lateral overlap rate is... The calculation requires the introduction of an elevation difference correction factor. Elevation correction factor The value is set between 1.5 and 2.0, and the calculation formula is as follows:

[0053] in, The base lateral overlap rate is preferably set at 60%. Based on the calculated... The planned spacing between adjacent routes is derived. :

[0054] UAVs based on minimum heading overlap , and Perform a "bow"-shaped scanning flight to collect raw image sets. And record each image simultaneously. Global Navigation Satellite System (GNSS) coordinates at the time of exposure and attitude angle .

[0055] S1.2: Construct a spatial hash index for multi-view images; To support the iterative requirements of location-based image set queries, it is not sufficient to simply save a single stitched large image; the mapping relationship between the original image and geospatial data also needs to be preserved. This invention constructs a spatial hash index table.

[0056] The geographic space of the target area is divided into a regular two-dimensional grid. For each original image Using its exterior orientation information ( and Project the image's field of view boundary onto the ground plane and calculate the set of ground grids covered by the image. .

[0057] Create inverted index mapping function :

[0058] in, This represents the grid coordinates. This index table ensures that a specific coordinate is identified on the panoramic image. When photovoltaic panels are present, the system can achieve [the following]. The time complexity is low, allowing for the rapid retrieval of all original image subsets where the location is visible. .

[0059] Step 1.3: Fast panoramic image generation based on weighted homography transformation; Based on the near-plane characteristics of the rooftops of rural houses, a homography matrix is ​​used to describe the transformation relationship between adjacent images.

[0060] For two adjacent images and Extract SIFT feature points and perform coarse matching. Let... The coordinates of the feature points on are Corresponding The coordinates of the feature points on are Define the homography matrix. Satisfies the projection transformation relationship:

[0061] in This indicates scale equivalence. To eliminate parallax artifacts caused by non-planar objects such as trees and utility poles in rural scenes, this embodiment employs a moving direct linear transformation algorithm for optimization. Unlike a globally uniform homography matrix, this algorithm applies a homography matrix to each center point in the panoramic image grid. Calculate a locally weighted homography matrix Minimize the weighted projection error function :

[0062] in, The number of matching point pairs. Weight function. Determined based on the spatial distance between feature points and the grid center:

[0063] in, For scale parameters, This represents the confidence level of the feature matching.

[0064] By solving the aforementioned weighted least squares problem, all original images are transformed to a unified geographic coordinate system, and a multi-band fusion algorithm is used to process the stitching seams, outputting a high-resolution panoramic orthophoto map covering the entire village. .

[0065] This invention addresses the logical requirement of location-first, analysis-later approach in distributed photovoltaic cluster forecasting by proposing a panoramic map construction method that combines adaptive flight path planning and spatial indexing techniques. Compared to existing technologies that directly perform undifferentiated 3D reconstruction of the entire village, this embodiment introduces a 3D baseline-to-height ratio constraint. Flight path planning physically ensures that every pixel in the panoramic image, especially potential photovoltaic panel areas, has multi-view data sources that meet the requirements for binocular vision calculation. This solves the problem of blind spots or insufficient baselines in traditional orthophotos, which can prevent tilt angle calculation. A spatial-image inverted index table was constructed, making the panoramic image not only a visual map but also a data retrieval library. This allows for high-performance attitude calculation by calling local image data for specific photovoltaic areas identified in the panoramic image, greatly reducing data processing redundancy and achieving an optimal balance between computing power and accuracy.

[0066] In rural settings, the background is extremely complex, containing distractions such as red terracotta tiles, gray concrete, vegetation, and roads. However, photovoltaic modules, as standardized industrial products, possess extremely stable optical and geometric characteristics. This invention employs a deterministic algorithm based on the fusion of spectral, texture, and geometric multidimensional features. This avoids high dependence on computational resources and, considering the fixed shape of photovoltaic panels, provides more interpretable and robust recognition results, thereby obtaining the center pixel coordinates and boundary range of each photovoltaic array in the panoramic image.

[0067] S2.1: Adaptive Spectral Threshold Segmentation Based on HSV Color Space Rural roofs come in a variety of materials, but in the visible light spectrum, crystalline silicon photovoltaic panels exhibit significantly low reflectivity and a specific color gradation (deep blue to black), which contrasts sharply with red terracotta tiles and high-gloss cement roofs. Since the RGB color space is greatly affected by light intensity, this step first involves converting the panoramic orthophoto image... Convert from RGB color space to HSV color space.

[0068] Pixels in the image The RGB components are The converted components are A photovoltaic spectral mask is constructed by utilizing the chromaticity characteristics of photovoltaic panels. The discriminant function for determining whether a pixel belongs to the photovoltaic candidate region is defined as follows:

[0069] in, This represents the hue threshold for photovoltaic modules, corresponding to the blue band. The minimum saturation threshold is used to remove interference from gray-white cement roofs; This is the brightness threshold used to remove highlight reflections and extremely dark shadow areas.

[0070] To eliminate artifacts caused by rooftop water accumulation or colored awnings, this embodiment introduces texture entropy constraints. The photovoltaic panel surface consists of an array of solar cells with a regular texture structure, while water accumulation areas are typically smooth. Local windows are defined. The gray-level co-occurrence matrix within is Calculate local energy , Represents a grayscale index;

[0071] By combining spectral masking and texture energy, a preliminary binarized segmentation map is obtained. :

[0072] in For indicator functions, This is the texture energy threshold.

[0073] S2.2: Noise Filtering Based on Morphological and Connected Component Analysis Preliminary segmentation diagram The image typically contains small noise spots caused by wires, gaps in leaves, or roof stains, as well as fractured areas due to gaps between photovoltaic panels. Morphological optimization is required to obtain the complete outline of the photovoltaic array.

[0074] Applying morphological closing operations and utilizing structuring elements Dilatation followed by erosion of the binary image. This represents the expansion operation. This represents a budgeted approach to fill the installation gaps between photovoltaic modules, integrating discrete solar panels into a connected region:

[0075] Applying morphological opening operations, utilizing structuring elements Erosion followed by dilation is applied to the image to remove isolated noise points smaller than the minimum size of the photovoltaic module, such as shadows from rooftop water heaters.

[0076] Processed image Perform the connected component labeling algorithm to extract A set of candidate connected regions .

[0077] Step 2.3: Rectangularity Filtering Model Based on Multidimensional Geometric Features Besides photovoltaic panels, there might also be dark-colored water storage tanks or corrugated steel sheds; these objects pass the S2.1 screening based on color. However, photovoltaic modules have strict industrial manufacturing standards and must be presented as regular rectangles or combinations of rectangles. This invention establishes a multi-dimensional geometric feature screening model for candidate regions. To determine whether something is genuine or counterfeit.

[0078] For each connected region Calculate its minimum circumscribed rectangle, and denote the area of ​​this circumscribed rectangle as . The length of the longer side is The length of the shorter side is Simultaneously, calculate the connected components. actual pixel area .

[0079] Define the rectangularity index :

[0080] Define aspect ratio index :

[0081] Construct a comprehensive discriminant function A region is considered a valid photovoltaic array if and only if all of the following physical constraints are met:

[0082] in, The rectangularity threshold is preferably set to 0.85, with values ​​closer to 1 indicating a closer approximation to a rectangle. The minimum area threshold is based on the ground sampling distance. The physical dimensions of a single photovoltaic panel are converted to eliminate false positives due to excessively small area. This is the upper limit of the aspect ratio, used to eliminate long and narrow road shadows or utility pole projections.

[0083] Step 2.4: Structured Output of Location Information After screening, those meeting the requirements were retained. The effective photovoltaic area. For each confirmed photovoltaic area... Calculate its centroid coordinates and the set of coordinates of the four vertices of the circumscribed rectangle. .

[0084] Mapping these two-dimensional pixel coordinates back to the geographic coordinate system of the panoramic image generates a list of photovoltaic locations. , of which The information for each photovoltaic array is represented as follows:

[0085] The list records where the photovoltaic panels are located and specifies their coverage area on the panoramic view.

[0086] The core physical variable determining power generation is the amount of radiation received, which directly depends on the spatial orientation of the photovoltaic array, namely its azimuth and tilt angles. Rural roof structures are extremely complex, ranging from ideal south-facing roofs to east-west oriented roofs, and various installation methods such as flat installation or elevated installation. A simple two-dimensional location is insufficient to derive power generation efficiency. If existing technologies are used to perform indiscriminate dense 3D reconstruction of the entire village to obtain these parameters, the rural environment, containing numerous trees, roads, and other invalid areas, would generate massive amounts of redundant data and suffer from low computational efficiency. Therefore, this embodiment employs an index-based local reconstruction strategy. Using an established spatial hash index, for each identified photovoltaic region (ROI), a subset of the original multi-view images covering that region is retrieved in reverse. By performing sparse point cloud reconstruction and plane fitting within the local region, the 3D normal vector of the photovoltaic array is calculated, thereby deriving its azimuth and tilt angles.

[0087] S3.1: ROI Multi-view Image Set Retrieval Based on Spatial Hash Mapping For the output list of photovoltaic locations The first in One photovoltaic target, of which , To identify the total number of photovoltaic cells, extract their geographic centroid coordinates. and the set of circumscribed polygon vertices .

[0088] In order to establish the original image library Find all images that can effectively observe the photovoltaic panel and use a spatial hash index table. Perform a reverse lookup.

[0089] calculate The grid index Query the index table to obtain the global image index set covering the grid. .

[0090] The drone's flight path may result in some images covering the coordinates being taken from an excessively wide angle or from behind a roof. Therefore, an effective visual set needs to be constructed. For sets Each image index in Obtain the optical center coordinates of its camera. Constructing the gaze vector Calculate the angle between the line-of-sight vector and the ground plane normal vector. Vertically upward :

[0091] Set effective observation threshold Only retain those that meet the requirements. The image index forms a target for the first A subset of local multi-view images of a photovoltaic target .

[0092] This invention ensures that all images involved in the calculation are viewed from a top-down or oblique angle, eliminating invalid viewpoints with severe texture distortion.

[0093] S3.2: Local Feature Matching and Triangulation of Sparse Point Clouds Unlike full-scene reconstruction, this step only involves... The image within is processed. To avoid matching the photovoltaic panels to trees or the ground around them, the obtained photovoltaic boundary is utilized. The feature extraction scope is strictly limited to the ROI area of ​​the photovoltaic panel.

[0094] For any two images in the subset and ,in and Extract SIFT feature points within the ROI. Use epipolar geometry constraints to eliminate mismatches, i.e., homonyms. and The following fundamental matrix constraints must be satisfied:

[0095] in This is the base matrix for the two views. After matching, multi-view triangulation is performed. For the first... For each matching feature point, a system of linear equations is constructed to solve for its three-dimensional coordinates in the world coordinate system. Finally, a local sparse point cloud set of the photovoltaic target is obtained. ,in This represents the number of valid spatial points reconstructed.

[0096] S3.3: Photovoltaic Plane Fitting Based on RANSAC The surface of a photovoltaic module is physically a perfectly flat plane, but it generates point clouds. The data may be mixed in with noise from rooftop chimneys, vents, and birds loitering. To obtain accurate planar parameters, this embodiment employs the Random Sample Consensus (RANSAC) algorithm.

[0097] Let the equation of the photovoltaic plane to be determined be:

[0098] Where the normal vector Satisfy normalization conditions .

[0099] In point cloud collection Three non-collinear points are randomly selected from the set, and the temporary plane parameters are calculated. Calculate the parameters for all points in the set. algebraic distance to the plane :

[0100] Statistical satisfaction Number of interior points ,in The threshold is set based on the flatness of the photovoltaic panel manufacturing process.

[0101] Repeat the above steps Next, the model with the most interior points is selected as the optimal solution.

[0102] Based on the optimal interior point set, the plane parameters are optimized again using the least squares method to obtain the th... The optimal normal vector of a photovoltaic target .

[0103] S3.4: Attitude Parameter Calculation and Coordinate System Transformation After obtaining the normal vector, convert it into the physical angle required for photovoltaic power generation prediction. In the geographic coordinate system (ENU: East-North-Sky), ensure the normal vector points towards the sky. .

[0104] Define photovoltaic tilt angle The angle between the normal vector and the zenith direction (Z-axis) directly determines the incident cosine loss of the photovoltaic panel for direct sunlight.

[0105] Define photovoltaic azimuth angle The angle between the projection of the normal vector onto the horizontal plane and the due north direction (Y-axis) determines the peak power generation time of the photovoltaic panel, i.e., earlier for east-facing panels and later for west-facing panels.

[0106] like Then make corrections. .

[0107] No. The physical properties of each photovoltaic target are upgraded from two-dimensional coordinates to three-dimensional attitude vectors. The system traverses all The goal is to complete the attribute calculation for the entire village's photovoltaic system.

[0108] This invention addresses the challenges of obtaining physical parameters in conventional prediction methods, given the haphazard installation and varying orientations of distributed photovoltaic (PV) systems. By employing an index retrieval-local reconstruction-fitting approach, it avoids the computational waste associated with full-domain 3D reconstruction and achieves parameter extraction. Local multi-view image retrieval utilizes only valid viewpoint images relevant to specific PV panels, eliminating interference from large-angle side and back views and ensuring high-quality feature matching. The introduction of RANSAC plane fitting effectively overcomes point cloud noise caused by protrusions such as chimneys and water heaters commonly found on rural rooftops, ensuring that the calculated tilt and azimuth angles accurately reflect the orientation of the PV module's light-receiving surface.

[0109] The actual power generation efficiency of photovoltaic (PV) modules depends not only on the amount of radiation received but is also severely constrained by the module's operating temperature and light path transmittance. Rural PV installation methods vary greatly. Modules laid flat on corrugated steel roofs experience significantly higher operating temperatures than those mounted on overhead concrete roofs due to obstructed ventilation on the back and heat absorption by the substrate. This hot spot effect leads to non-linear power output degradation. Furthermore, the environment is often dusty and contaminated with biological pollutants (bird droppings, leaves), and some modules exhibit aging and discoloration after prolonged operation, such as yellowing of the backsheet and snail-like patterns. If only the factory-standard efficiency parameters are used in the prediction, the physical losses caused by these environmental factors are ignored. Therefore, this invention, based on acquired three-dimensional attitude and local multi-view images, further extracts the installation thermal attribute characteristics of the modules to quantify heat dissipation capacity and simultaneously extracts health status characteristics to quantify light loss. Point cloud distance field analysis is used to determine the overhead / flat installation status.

[0110] S4.1: Installation thermal properties The heat dissipation performance of the component mainly depends on the ventilation gaps in the back panel and the thermal capacity characteristics of the roof substrate material. This invention utilizes the generated local sparse point cloud. and optimal plane normal vector Perform geometric analysis.

[0111] During the RANSAC fitting process, outliers that were removed included roof surface points located below the edges of the photovoltaic panels. The definition is for the... A collection of point clouds in the neighborhood of a photovoltaic target. This neighborhood point cloud set This includes points in the photovoltaic panel area and its surrounding expansion zone. The calculation set includes all points. Directed distance to the fitted plane :

[0112] Due to the normal vector Pointing towards the sky, the point on the roof surface should be below the plane, i.e. Extracting the desired result The point set as the base surface point set , This is the component thickness threshold.

[0113] Ventilation gap index Used to characterize the average distance between the component back panel and the roof:

[0114] Based on thermodynamic experience, the natural convection cooling effect tends to stabilize when the gap is greater than 0.3 meters. Therefore, a reference ventilation saturation threshold is set. , to obtain ventilation convection score :

[0115] The score A value close to 1 indicates good tiling, while a value close to 0 indicates complete flatness.

[0116] Using the selected local image set ,Will The 3D points in the image are back-projected back onto the image plane, and the RGB color mean of the corresponding pixels is extracted. The K-Means clustering algorithm was used to classify the substrate material into three typical thermal properties: for the high heat absorption type (dark-colored steel roofing sheets / asphalt), a heat loss correction coefficient was set. For medium heat absorption types (red clay tiles), a heat loss correction factor is set. For materials with low heat absorption (light-colored concrete / galvanized steel sheet), a heat loss correction factor is set. .

[0117] Output the first Heat dissipation factor of each component :

[0118] in These are the weighting coefficients.

[0119] S4.2: Extraction of ash accumulation and aging characteristics; To eliminate the influence of the shooting angle on the surface color analysis of the component, an inverse perspective transformation is first performed using the known orientation to generate a standard front-view texture map of the photovoltaic panel. The calculated normal vectors are then used... Construct rotation matrix The photovoltaic panel is rotated to be perpendicular to the camera's optical axis. Local images are resampled to generate a corrected ROI image. .

[0120] Dust accumulation can cause the component surface to appear as a grayscale whitening effect on the histogram. Convert to grayscale space and calculate the proportion of high-brightness pixels. (Dust accumulation index) :

[0121] in Total number of pixels This represents the upper limit of grayscale for clean photovoltaic panels. The larger the value, the more severe the dust accumulation and the lower the light transmittance.

[0122] Component aging typically manifests as localized unevenness in surface color. Convert to Lab color space and extract the luminance channel. and chroma channels Calculate the color texture entropy of the image. As a characteristic of aging:

[0123] in This represents the probability distribution of the color clustering histogram in the Lab color space. Severely aged components exhibit mottled color variations on their surface. Significantly higher than the new components.

[0124] Build component health status correction coefficient Used to correct photoelectric conversion efficiency:

[0125] in For decay weights, This is the maximum normalized entropy value.

[0126] In practical engineering, most distributed photovoltaic (PV) projects, especially county-wide household PV projects, are constrained by initial investment cost control. Their metering systems strictly adhere to the settlement boundary principle, meaning the power grid company only installs one bidirectional smart meter that meets settlement standards at the property boundary point (gateway). Although this meter can accurately record the energy exchange (net power) between the user and the grid, the data collected by the meter is physically an algebraic sum of PV power generation and user load power because the PV inverter's grid connection point is usually located on the internal network side behind the meter. Furthermore, although some inverters have data acquisition capabilities, in the current power data ecosystem, inverter data is usually transmitted back to the equipment manufacturer's private cloud platform. The power grid dispatch system, limited by commercial privacy agreements, inconsistent data interface standards, and network security barriers, cannot obtain real-time, penetrating raw power generation data from the inverter side of a massive number of dispersed users, thus creating a black-box characteristic of strong coupling between source and load data.

[0127] However, according to the engineering specifications for distribution network operation and management, in order to ensure the safe operation of low-voltage distribution areas and conduct necessary power quality analysis, the operation and maintenance department, while unable to bear the high cost of upgrading to full-network user monitoring, will adopt a strategy of sparse deployment and sampling observation. In actual distribution area construction, voltage-sensitive households at the end of the line, large-capacity typical households, or specific low-carbon demonstration households are usually selected to install fully-sensing intelligent terminals (TTUs) or high-precision power quality monitoring instruments, or portable waveform recording equipment is temporarily mounted during the lean survey of distribution areas. These nodes with full parameter measurement capabilities objectively constitute benchmark samples within the distribution area, i.e., standard users. Therefore, using a very small amount of standard data to infer the internal state of a massive number of blind spot users is the only low-cost and feasible technical path to solve the contradiction between low observability and the need for refined management during the current transition period of digital transformation of distribution networks. This embodiment addresses the specific technical problem of how to break down data silos and map high-dimensional information of sparse standards to a massive number of users to be measured through cross-modal feature transfer in this engineering context.

[0128] This invention introduces voltage sequence correlation as a matching fingerprint. Unlike urban cable networks, rural power grids typically have large power supply radii and high line resistance-to-reactance ratios, making node voltages extremely sensitive to changes in active power injection. According to circuit principles, adjacent users on the same feeder branch are connected to the transformer via a shared common impedance. This physical topology determines that they have a highly consistent electrical background. When a cloud passes through this branch area, the photovoltaic output of all users on the branch will drop almost simultaneously. Due to the existence of the common impedance, this synchronous change in active power is instantly converted into synchronous fluctuations in voltage amplitude. Therefore, if the voltage curves of two users show a high correlation at the minute-level granularity, it is sufficient to prove that they are in the same micro-meteorological coverage area, and their photovoltaic output time characteristics are locked. The actual output of photovoltaic inverters is often constrained by grid voltage, such as overvoltage derating strategies. If the voltage waveforms of the user under test and the standard user are highly similar, it means that they are under the same voltage constraint environment. This matching based on electrical fingerprints is essentially searching for electrical twin nodes. This means that when migrating photovoltaic curves from standard users to those of users under test, there is no need for additional complex voltage sensitivity correction and weather time difference compensation. Because both are naturally under the same voltage fluctuation background and weather transients, the confidence level of the directly transferred data is extremely high, thus cleverly avoiding the modeling difficulties caused by the lack of topological data.

[0129] Step 5.1: Construct a multidimensional source-load feature standard library Assume there are within the station area The first standard user. For the first A standard user, The system performs the following data aggregation operations: (1) Construction of physical fingerprint vector: Since the shape of the photovoltaic power output curve is mainly determined by the geometric attitude, while the amplitude attenuation is mainly determined by environmental health factors, a physical feature vector is constructed. :

[0130] in, For photovoltaic tilt angle, For photovoltaic azimuth, As a heat dissipation factor, This is a correction factor for health status.

[0131] (2) Standardize the source-load behavior pattern and obtain the standard users in the historical time window. Photovoltaic power generation sequence and load power sequence To eliminate the dimensional differences caused by varying installed capacity, a normalized photovoltaic power output sequence needs to be calculated. :

[0132] in The rated installed photovoltaic capacity for this user is used. Simultaneously, the load sequence is retained as a typical profile of the user's electricity consumption behavior.

[0133] (3) Extraction of electrical fingerprint sequence Extract the standard user's gateway electricity meter within the historical time window Voltage amplitude sequence within :

[0134] Step 5.2: Physically Gated Cascaded Similarity Matching For any user to be tested within the test area, the index is: The system is known to have its physical characteristic vectors. and electrical voltage sequence A two-stage screening mechanism is used to find the best matching criteria.

[0135] (1) First-level screening: physical attitude-based gating; The orientation of photovoltaic modules is a decisive factor in determining power generation patterns, possessing veto power. Therefore, calculating the power generation patterns of the user under test... Differences from all standard users in the gesture dimension.

[0136] Constructing the pose difference function :

[0137] in For angle weighting. Set a strict truncation threshold. Preferred It is 15 degrees.

[0138] Only those that meet the requirements are selected. The standards constitute a candidate standard set for the users to be tested. This invention forcibly eliminates all standards with different orientations, regardless of how close their electrical distance is, and they cannot be used as a reference for photovoltaic feature migration.

[0139] (2) Second-level optimization: spatiotemporal matching based on electrical fingerprint; Candidate set after physical attitude gating Further investigation is needed to identify objects with the most similar electrical environments. The correlation of voltage sequences is used to characterize micro-meteorological synchronicity and load background similarity.

[0140] Calculate the user to be tested With candidate criteria (in Electrical Pearson correlation coefficient : At the same time, the similarity between the two in terms of heat dissipation and health properties is calculated. :

[0141] The first term is calculated by dividing by the larger of the two values ​​and adding a small amount. To prevent the denominator from being zero, the difference in heat dissipation factors is forcibly mapped to the [0, 1] interval, making it present as a relative error; the second term, since the health coefficient itself is a normalized value, (3) Overall score; Build the final matching score :

[0142] in For electrical weights, For attribute weights. Select the candidate criterion with the highest score. As a user to be tested The twin agent.

[0143] Although the system has identified the optimal standard user, the source-load blind separation algorithm based directly on migration data inherently suffers from mathematical ambiguity and discriminative fuzziness. In the physical equation of net grid-connected power = photovoltaic power generation - user load, the lack of measured data from the inverter side makes it difficult for the model to accurately identify the source of power fluctuations at every moment. Specifically, when the net grid-connected power of the user under test suddenly drops, the purely data-driven decoupling algorithm often faces an attribution dilemma: this could be due to cloud cover causing a sudden decrease in photovoltaic output (i.e., a generation-side cause), or it could be due to a sudden surge in self-consumption caused by the user turning on high-power appliances (i.e., a load-side cause). If the algorithm lacks practical verification, it is highly likely to misjudge a sudden increase in load as a sudden drop in photovoltaic power, or a photovoltaic shutdown as a very low load. This algorithmic uncertainty caused by the strong coupling of source and load data leads to a situation where the decomposed photovoltaic curve, while mathematically fitting the total meter reading, is physically distorted. Therefore, a second-dimensional reference standard independent of electrical data must be introduced, namely, theoretical verification based on UAV visual perception. Using theoretically calculated photovoltaic capacity values ​​as hard constraints, boundary checks and logical errors are performed on the decoupling results of the data channels, thereby eliminating mathematically valid but physically absurd solutions, such as those that calculate photovoltaic output exceeding the limit that the physical area of ​​the roof can bear.

[0144] S6.1: Initial Source Load Decoupling Based on Feature Transfer Utilizing the best standard user of the lock users to be tested The net load data at the checkpoint is separated. The user under test is obtained. At the present moment Net power of bidirectional meter This represents a surplus of electricity generated by photovoltaic power generation.

[0145] Using standard users Normalized photovoltaic curve and normalized load curve Decoupling equations are constructed.

[0146] Let the estimated capacity to be corrected for the user under test be... The load scaling factor is . It represents the comprehensive conversion ratio between the normalized photovoltaic waveform migrated from the standard user and the absolute power actually perceived by the meter at the user under test. This coefficient not only includes the physical installed capacity of the photovoltaic modules, but also integrates multiple electrical factors such as inverter conversion efficiency, DC / AC line loss, grid connection point voltage suppression loss, and local micro-shading, making it a comprehensive fitting coefficient.

[0147] Estimated photovoltaic power output value for:

[0148] The goal of decoupling is to find the optimal... and This minimizes the error between the reconstructed net power and the measured net power.

[0149] The photovoltaic output value calculated from the electrical dimension at that moment was obtained by solving the least squares method. .

[0150] S6.2: Theoretical output calculation; The system calls the extracted physical parameters and combines them with current meteorological data (irradiance). Ambient temperature ), calculate the theoretical output value.

[0151]

[0152] in: The identified geometric area of ​​the photovoltaic panel; attitude angle Calculated effective radiation of the inclined plane; The baseline conversion efficiency for components.

[0153] This step calculates It is a theoretical photovoltaic output value based on the visual perception of drones.

[0154] S6.3: Data-side parameter correction based on physical reference In distributed photovoltaic cluster scenarios, the physical information (area, attitude) acquired by the UAV visual perception system is an objective, existing physical observation with extremely high confidence. In contrast, the data blind decoupling process in S6.1 inherently suffers from mathematical ambiguity. Specifically, the user under test... Estimated capacity This is an initial unknown variable. If this variable is not accurately estimated, the decoupling algorithm may incorrectly attribute photovoltaic power generation to load reduction or vice versa. Therefore, this invention requires data-side parameter correction based on physical benchmarks. Calculate the full time window Cumulative energy deviation rate between internal data-side output and theoretical output :

[0155] Set convergence tolerance threshold .like This indicates the estimated capacity currently used on the data side. If there is a mismatch with the physical entity, the parameter correction iteration needs to be initiated.

[0156] Based on deviation rate Update the estimated capacity on the data side. Set the iteration round index. Then the first Estimated capacity of the wheel The calculation is as follows:

[0157] in This is the learning rate coefficient, used to control the calibration step size and prevent oscillations.

[0158] when (Right now At this point, the photovoltaic power output decoupled from the data side significantly exceeds the theoretical upper limit observed by the UAV. This is a typical case of load-filling overfitting. The decoupling algorithm of this invention uses the least squares method to find the optimal solution. Mathematically, when the user under test experiences an atypical load drop during a certain period (e.g., a low electricity consumption due to equipment shutdown at noon), and the load curve of the standard user does not show this characteristic, the decoupling algorithm, in order to minimize the total residual, tends to use the shape of the photovoltaic curve to fill this load gap. The algorithm incorrectly assumes that the increase in net power is due to excessive photovoltaic power generation, rather than insufficient load consumption. This mathematically optimal fit leads to inaccurate capacity estimation. It was incorrectly raised, generating virtual photovoltaic power that exceeded the physical roof area limit.

[0159] when (Right now At this point, the photovoltaic power output decoupled from the data side is far lower than the theoretical physical value. User loads typically exhibit high-frequency, high-amplitude random fluctuations. When the fluctuation amplitude of photovoltaic output is relatively small, or when the characteristic frequency of the photovoltaic signal overlaps with the noise frequency of the load, the purely data-driven decoupling algorithm will exhibit conservatism. The algorithm struggles to accurately extract the smooth photovoltaic component from the drastically fluctuating net load curve, instead tending to incorrectly absorb the photovoltaic power generation into the random fluctuation term of the load, resulting in a significantly lower calculated effective electrical gain.

[0160] Utilizing the corrected capacity As a fixed known quantity, the least squares decomposition in step S6.1 is re-executed to update the load scaling factor. And generate new data-side output values.

[0161] Construct a decoupling objective function constrained by physical capacity:

[0162] Solving this equation yields the first... Optimal load scaling factor for wheels .

[0163] Then, the corrected photovoltaic output value on the data side is calculated. :

[0164] S6.4: Converging Output Repeat step S6.3, calculate the new deviation rate, and iterate until the convergence condition is met. Or it may reach the maximum number of iterations.

[0165] Output the final converged corrected photovoltaic power sequence :

[0166] Simultaneously output the corresponding corrected load sequence. :

[0167] This invention addresses the technical challenge of source-load aliasing in rural photovoltaic (PV) forecasting by constructing a closed-loop correction system based on physical vision. With only access meter data, mathematical models are prone to misjudging low loads at midday as high PV output. This embodiment introduces theoretical values ​​obtained through UAV vision as a hard constraint. When physical calculations indicate that the user's roof area is insufficient to generate such a large power output, the algorithm forcibly lowers the capacity parameters on the data side. This eliminates solutions from the mathematical solution space that do not conform to physical reality. The final output is... Its waveform details are strictly derived from standard user data. The electrical fingerprint migration strictly meets the energy conservation requirements of the gate meter. The measured constraints are applied. This scheme utilizes both physical means to calibrate the magnitude and electrical means to preserve fluctuations, effectively improving the robustness of distributed photovoltaic cluster predictions under low observability environments.

[0168] For each user to be tested The system not only obtained historical photovoltaic power output sequences and load sequence Through closed-loop iterative verification, the system has determined the user's true electrical capacity, which has undergone both physical and data verification. Based on this, individual household photovoltaic forecasts are made using the corrected physical parameters combined with weather forecasts, and individual household load forecasts are made using the decoupled load history. Finally, topology aggregation is performed at the transformer substation level.

[0169] S7.1: Single-Household Photovoltaic Physical Prediction Based on Corrected Parameters For each user in the cluster ( Its photovoltaic power generation capacity has been revised from the estimated capacity. The only certainty.

[0170] Obtain the future time period to be predicted Numerical weather forecast data, including total horizontal irradiance. and ambient temperature .

[0171] Using the calculated attitude angles Convert horizontal irradiance to effective irradiance on an inclined plane Constructing a single-household photovoltaic prediction model:

[0172] in, Irradiance under standard test conditions; S7.2: Single-household load trend forecasting based on decoupled history For users Using the output corrected load sequence This serves as a clean sample database of load data. Since rural users exhibit significant daily work-rest patterns (such as fixed cooking and lighting times), this embodiment employs a similar-day weighted algorithm for prediction.

[0173] Based on the predicted date The date type (weekday / holiday) and meteorological characteristics are matched from historical databases. A similar day Calculate the decay function for the weights of similar days. :

[0174] in It is a vector consisting of temperature, humidity, and light intensity.

[0175] Calculate the load forecast for a single household :

[0176] S7.3: Topology Aggregation of Cluster Power Based on the physical connection topology of the transformer area, the prediction results of all individual households are algebraically superimposed to obtain the total predicted value of the cluster's photovoltaic power. and cluster total load forecast .

[0177] Final output area gate net power prediction value :

[0178] in This represents the estimated line loss power in the low-voltage distribution area.

[0179] This invention discloses a distributed photovoltaic cluster power prediction system, which includes: Image acquisition module: Acquires spatial hash indices of panoramic orthophotos and multi-view images covering the target area; Attitude calculation module: Identifies the target position of photovoltaic modules on the panoramic orthophoto, retrieves the corresponding local multi-view images using spatial hash index, and calculates the attitude parameters of the photovoltaic array; Matching module: Constructs a source-load feature standard library, and performs cross-modal similarity matching between standard users and users under test based on physical attitude gating and electrical fingerprinting to determine the best matching standard user for the user under test; Estimation module: Couples the source-load behavior patterns of the best-matched standard users with the net load data of the user under test to obtain the photovoltaic output value on the data side; Verification module: Calculates the theoretical photovoltaic output value by combining attitude parameters; Based on the theoretical photovoltaic output value, iteratively corrects the estimated capacity on the data side until convergence to obtain the corrected photovoltaic power sequence and the corrected load sequence; Prediction module: Combines numerical weather forecasting to predict individual photovoltaic loads and obtains power prediction results for distributed photovoltaic clusters through topology aggregation.

[0180] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for predicting the power of a distributed photovoltaic cluster.

[0181] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for predicting the power of a distributed photovoltaic cluster.

[0182] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0183] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0184] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the power of a distributed photovoltaic cluster, characterized in that, The prediction method includes: Obtain the spatial hash index of panoramic orthophotos and multi-view images covering the target area; Identify the target location of photovoltaic modules on panoramic orthophotos, retrieve the corresponding local multi-view images using spatial hash index, and calculate the attitude parameters of the photovoltaic array; A source-load feature standard library is constructed. Based on physical attitude gating and electrical fingerprinting, cross-modal similarity matching is performed between standard users and users under test to determine the best matching standard user for the user under test. By coupling the source-load behavior patterns of the best-matching standard users with the net load data of the users under test at the gate, the photovoltaic output value on the data side is obtained. The theoretical photovoltaic output value is calculated by combining attitude parameters; based on the theoretical photovoltaic output value, the estimated capacity on the data side is iteratively corrected until convergence is obtained to obtain the corrected photovoltaic power sequence and the corrected load sequence. We combine numerical weather prediction to forecast individual photovoltaic (PV) systems and loads, and obtain power forecasts for distributed PV clusters through topology aggregation.

2. The method for predicting the power of a distributed photovoltaic cluster according to claim 1, characterized in that, The spatial hash index for obtaining panoramic orthophotos and multi-view images covering the target area includes: The geographic space of the target area is divided into a regular two-dimensional grid. Establish inverted index mapping function : in, Represents grid coordinates, For the first Zhang's original image Covered ground grid set, This is the original image set; When retrieving the corresponding local multi-view images using spatial hash indexing, a subset of local multi-view images targeting the nth photovoltaic module is constructed. : in To meet the effective observation threshold A collection of image indexes.

3. The method for predicting the power of a distributed photovoltaic cluster according to claim 2, characterized in that, The calculation of the photovoltaic array's attitude parameters specifically includes: performing feature matching and multi-view triangulation using retrieved local multi-view images to generate a set of local sparse point clouds. ; The photovoltaic plane equation is obtained by fitting a local sparse point cloud set to a random sample consensus algorithm. Optimal normal vector ; Calculating the photovoltaic tilt angle based on the optimal normal vector and photovoltaic azimuth angle : in, These are the normalized normal vector components.

4. The method for predicting the power of a distributed photovoltaic cluster according to claim 3, characterized in that, Extracting the installation thermal properties and health status features of components, specifically including: calculating the base surface point set in the local sparse point cloud. Ventilation gap index to the fitted plane And combined with the heat loss correction coefficient of the substrate material Calculate heat dissipation factor : in, The normalized ventilation convection score, These are the weighting coefficients; Generate standard front view texture map for photovoltaic panels Extracting the ash accumulation index and color texture entropy Calculate the health status correction factor : in, For decay weights, This is the maximum normalized entropy value.

5. The method for predicting the power of a distributed photovoltaic cluster according to claim 4, characterized in that, Based on physical posture gating and electrical fingerprinting, cross-modal similarity matching is performed between standard users and users under test: Calculate the user to be tested With standard users pose difference function Only retain Candidate standard users, among which This is the truncation threshold; Calculate the user to be tested and the candidate standard user Electrical Pearson correlation coefficient and attribute similarity : Build the final matching score The user with the highest score is selected as the best matching standard. : in, These are the weighting coefficients. To prevent tiny quantities with a denominator of zero.

6. The method for predicting the power of a distributed photovoltaic cluster according to claim 1, characterized in that, The source-load behavior patterns of the best-matched standard users are used to decouple the net load data of the gate users under test. Specifically, this includes: constructing a decoupling objective function and solving for the optimal load scaling factor. : Calculate the photovoltaic output value on the data side. : in, The net power of the bidirectional meter for the user under test. The estimated capacity for the user to be tested. For standard users, the normalized photovoltaic curve, Normalized load curve for standard users.

7. The method for predicting the power of a distributed photovoltaic cluster according to claim 4, characterized in that, Calculate the theoretical photovoltaic output value The specific formula is as follows: in, To identify the geometric area of ​​the photovoltaic panel, The effective radiation of the inclined plane is calculated based on the attitude angle. For component baseline conversion efficiency, As a heat dissipation factor, This is a correction factor for health status.

8. The method for predicting the power of a distributed photovoltaic cluster according to claim 7, characterized in that, Based on the theoretical photovoltaic power output, the estimated capacity on the data side is iteratively corrected, specifically including: Calculate the full time window Cumulative energy deviation rate between internal data-side photovoltaic output and theoretical photovoltaic output : Update the first based on the cumulative energy deviation rate Estimated capacity of the wheel : Blind decoupling is re-executed using the updated estimated capacity as a constant until... Output the final corrected photovoltaic power sequence ,in This is the learning rate coefficient. This is the convergence tolerance threshold.

9. The method for predicting the power of a distributed photovoltaic cluster according to claim 8, characterized in that, Until convergence is achieved, the corrected photovoltaic power sequence and the corrected load sequence are obtained, specifically including: using the converged estimated capacity. As a fixed known quantity, construct a decoupling objective function constrained by physical capacity: Solving for the optimal load scaling factor The corrected photovoltaic power sequence was obtained. and corrected load sequence : in The net power of the bidirectional meter for the user under test. Normalized load curve for standard users.

10. A distributed photovoltaic cluster power prediction system, characterized in that, The system includes: Image acquisition module: Acquires spatial hash indices of panoramic orthophotos and multi-view images covering the target area; Attitude calculation module: Identifies the target position of photovoltaic modules on the panoramic orthophoto, retrieves the corresponding local multi-view images using spatial hash index, and calculates the attitude parameters of the photovoltaic array; Matching module: Constructs a source-load feature standard library, and performs cross-modal similarity matching between standard users and users under test based on physical attitude gating and electrical fingerprinting to determine the best matching standard user for the user under test; Estimation module: The source-load behavior patterns of the best-matched standard users are coupled with the net load data of the user under test to obtain the photovoltaic output value on the data side; Verification module: Calculates the theoretical photovoltaic output value by combining attitude parameters; Based on the theoretical photovoltaic output value, iteratively corrects the estimated capacity on the data side until convergence to obtain the corrected photovoltaic power sequence and the corrected load sequence; Prediction module: Combines numerical weather forecasting to predict individual photovoltaic loads and obtains power prediction results for distributed photovoltaic clusters through topology aggregation.

Images