Photovoltaic support adaptive tracking control method and system

Abstract

The application discloses a photovoltaic support adaptive tracking control method and system, relates to the technical field of adaptive control, and comprises the following steps: collecting sky images, light intensity, photovoltaic panel temperature and support posture data by using multiple sensing units; the application realizes real sunlight determination through multi-dimensional feature fusion, combines irradiation field light gradient analysis, photovoltaic panel temperature synchronism verification and cloud shadow edge position matching, and can effectively identify false peak values formed by adjacent row reflected light, ground object reflected light and cloud shadow edge local strong light; when suspicious false peak values are detected, the double mechanism of historical irradiation curve comparison and solar geometric position verification is further used to ensure the accuracy of the determination of the position of the sunlight, avoid tracking deviation caused by false peak value misjudgment in the traditional single-point measurement or current feedback strategy, make the photovoltaic support always align with the real optimal light direction, and reduce the local heating and aging risk of components.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of adaptive control, in particular to a photovoltaic support adaptive tracking control method and system. BACKGROUND

[0002] In the field of photovoltaic power station support tracking control, single-axis and double-axis tracking support angle optimization has long relied on two core strategies of solar geometric position calculation and local single-point irradiance and current feedback. However, in actual complex working conditions, photovoltaic supports are faced with dynamic changes in incident solar angle, random fluctuations in cloud shadow posture, and multiple-row photovoltaic panel mirror reflection and other compound interferences.

[0003] Existing control systems are difficult to fully adapt to the above scenarios: when the sun is low in the morning and evening, photovoltaic supports are faced with dynamic changes in incident solar angle, and strong interference is formed by reflected light from adjacent rows and ground objects, leading to sensor misjudgment of the maximum power direction and deviation of the tracking angle from the optimal value; in high-density arrays, the short-term local power peak value is formed by the interlaced cloud shadow and the inter-row shading, which misleads the system to chase the false extreme value through the current feedback mechanism, causing continuous angle deviation, causing local heating of the components, aging of the packaging, and ultimately leading to a decrease in the performance ratio (PR) of the power station. The essential defect of the traditional strategy lies in the lack of systematic modeling of the space-time dimension irradiance field, and single-point measurement cannot capture spatial correlation and temporal characteristics, so eliminating systematic tracking deviation under complex working conditions is a technical problem that needs to be solved. SUMMARY

[0004] The purpose of the present application is to solve the problems of false peak value misjudgment, insufficient cloud shadow adaptation, and lack of multi-row cooperation due to random fluctuations in cloud shadow posture and strong interference formed by reflected light from adjacent rows and ground objects, and a new photovoltaic support adaptive tracking control method and system are proposed to solve the above problems.

[0005] In order to achieve the above purpose, the photovoltaic support adaptive tracking control method and system is adopted, which comprises:

[0006] The sky image, light intensity, photovoltaic panel temperature and support posture data are collected by a plurality of sensing units; the sky image, light intensity, photovoltaic panel temperature and support posture data are preprocessed, and a sky visible mask is generated;

[0007] Based on the preprocessed sky image, photovoltaic panel temperature and support posture data, an irradiance field model of the location of the photovoltaic support is constructed;

[0008] Determine whether the current wind speed is in a specific interval, if so: obtain the cloud layer contour feature and the morphological change data of the cloud layer under the action of wind force, output the dynamic position coordinate sequence of the cloud shadow on the ground, construct a cloud shadow rolling dynamic model and a cloud shadow light transmittance grading model;

[0009] The real sunlight determination is performed to determine the final angle adjustment scheme based on the real sunlight position, the irradiation field model, the cloud shadow rolling dynamic model and the cloud shadow transmittance grading model.

[0010] Further, the construction method of the irradiation field model comprises:

[0011] The solar elevation angle and the azimuth angle are calculated based on the solar geometric position, and the direct light intensity is calculated in combination with the shielding probability of the sky visible mask;

[0012] The hemispherical space scattering light intensity is calculated based on the cloud coverage and the aerosol concentration;

[0013] The reflectivity of the reflection surface material database is called, and the environmental reflected light intensity is calculated in combination with the support posture data;

[0014] The direct light intensity, the scattering light intensity and the environmental reflected light intensity are weighted and fused to generate the irradiation field model.

[0015] Further, the construction method of the cloud shadow rolling dynamic model comprises:

[0016] Step 0A: edge detection is performed on the sky image to extract a pixel coordinate point set of the cloud profile; and normal vectors of each pixel coordinate point of the cloud profile are calculated;

[0017] Step 0B: the cloud layer motion rate at the initial time is calculated based on the optical flow algorithm; the speed normal component of the cloud profile is set to 0; the vortex motion control equation is solved to obtain the speed of the cloud profile at a specific time; the speed field at the specific time is obtained; and the cloud profile position coordinates of the cloud profile at the specific time are obtained by integrating the speed field of the cloud profile at the specific time;

[0018] Step 0C: the cloud profile position coordinates of the cloud profile at the specific time are substituted into the perspective projection formula to obtain the ground cloud shadow profile coordinates; the cloud shadow profile inside is filled to generate a binary cloud shadow mask of the cloud shadow coverage area;

[0019] Step 0D: the positional relationship between the cloud shadow mask and the ground photovoltaic panel array is compared to calculate the number of photovoltaic panels covered by the cloud shadow; in combination with the irradiation sensor data, if the actual light intensity drop area deviates from the cloud shadow mask prediction area by more than a rated percentage, the vortex viscosity coefficient of the vortex motion control equation is corrected;

[0020] Steps 0B to 0D are repeated to generate the cloud shadow rolling dynamic model.

[0021] Further, the construction method of the cloud shadow transmittance grading model comprises:

[0022] Step 1A: cloud type classification of the sky image based on a semantic segmentation model, outputting a cloud type coefficient; extracting the gray value of the target cloud layer region;

[0023] Step 1B: selecting a cloud-free area in the pixel-level cloud type coefficient of the sky image, calculating the average gray value of the cloud-free area; obtaining the gray value of the target cloud layer region and calculating the cloud layer thickness of the target cloud layer region; according to the cloud type coefficient, the initial transmittance level is calculated;

[0024] Step 1C: based on the cloud layer thickness of the target cloud layer region, calculate the thickness change rate; combine the current wind speed to correct the time constant; substitute into the transmittance update formula to adjust the transmittance level in real time, the update frequency is consistent with the image acquisition frequency;

[0025] Step 1D: compare the measured value of the contrast irradiation sensor with the theoretical irradiation attenuation corresponding to the transmittance level, and generate a cloud shadow transmittance grading model.

[0026] Further, the specific steps of determining the real sunlight are as follows:

[0027] Calculate the light intensity gradient of each region in the irradiation field. If the light intensity gradient value of a certain region exceeds the set threshold value and does not match the light change trend caused by cloud movement, it is marked as a suspicious false peak;

[0028] Combined with the temperature sensor data, if the regional temperature does not rise synchronously with the light intensity or a local high temperature point appears, the false peak risk index is increased; if the suspicious false peak region coincides with the cloud shadow edge position predicted by the cloud shadow rolling dynamic model, the false peak risk index is reduced;

[0029] When the false peak risk index is lower than the set threshold value, it is determined that the currently recognized is real sunlight, that is, the region with the highest light intensity in the irradiation field is consistent with the actual position of the sun, and the light intensity change is smooth and consistent with the trend of cloud movement;

[0030] If the false peak risk index is higher than the set threshold value, the system starts a double verification mechanism.

[0031] Further, the specific steps of the double verification mechanism are as follows:

[0032] Retrieve historical irradiation data of the same period, compare the shape characteristics of the current irradiation curve, and if the trends are consistent, increase the confidence of real sunlight;

[0033] Combined with the geometric position information of the sun, verify whether the region with the strongest light is consistent with the theoretical position of the sun, and if the region with the strongest light is consistent with the theoretical position of the sun, determine the position of the real sunlight.

[0034] Further, the position data of the photovoltaic support in the same area is obtained, whether there are multiple rows of photovoltaic supports in the same area is detected, and if so:

[0035] There is a risk of row-to-row influence between the photovoltaic supports in the same area; each row of photovoltaic supports in the array is equipped with an independent sensing unit and a computing module to determine whether the photovoltaic support is in a cloud shadow coverage area, and if so, the coverage range, entry time, and estimated duration of the cloud shadow are recorded;

[0036] A topology correlation graph of the array is constructed through a graph neural network array coupling model, each row of supports is taken as a node, and the row-to-row shielding and reflection relationship is taken as an edge; the initial edge weight is calculated based on the geometric parameters of the array;

[0037] The edge weight is corrected based on a cloud shadow rolling model; through a multi-row irradiation field cross-correction model, the preliminary identification results of each row are cooperatively corrected in combination with the mutual interference coefficient matrix output by the graph neural network and the cloud shadow rolling dynamic model, the false light components caused by row-to-row reflection are deducted, and the light fluctuation of the edge row of the cloud shadow is smoothed; the real sunlight direction of each row is determined;

[0038] A plurality of angle adjustment schemes are generated, and a final angle adjustment scheme is calculated.

[0039] Further, the specific steps of correcting the edge weight based on the cloud shadow rolling model are as follows:

[0040] When the cloud shadow covers a row, the reflection light intensity of the adjacent row changes, and at this time, the corresponding edge weight increases according to the cloud shadow light transmittance level;

[0041] If the cloud shadow causes the temperature of a row of photovoltaic panels to decrease, the heat radiation exchange between the row and the adjacent row decreases, and the related edge weight is adjusted downward in proportion to the temperature drop.

[0042] Further, the specific steps of cooperatively correcting the preliminary identification results of each row through the multi-row irradiation field cross-correction model in combination with the mutual interference coefficient matrix output by the graph neural network and the cloud shadow rolling dynamic model, deducting the false light components caused by row-to-row reflection, and smoothing the light fluctuation of the edge row of the cloud shadow are as follows:

[0043] If the reflection light of the front row support is enhanced due to cloud shadow shielding, the irradiation field model of the rear row photovoltaic support is corrected based on the edge weight calculated by the graph neural network coupling estimation unit, and the false light components caused by the reflection of the front row are deducted;

[0044] If a row is at the edge of the cloud shadow, the identification result of the adjacent row is combined to perform light smoothing processing on the row, so as to eliminate the interference of local fluctuation.

[0045] Further, it comprises:

[0046] Multi-modal sensing layer: including fisheye panoramic imaging unit, irradiance sensor, temperature sensor, attitude sensor and wind speed sensor;

[0047] Data preprocessing module: for preprocessing the sky image, light intensity, temperature, attitude and wind speed collected by the multi-modal sensing layer;

[0048] Irradiance field modeling module: for receiving the preprocessed sky image, light intensity and attitude data, and constructing an irradiance field model;

[0049] Cloud shadow modeling module: for starting cloud shadow monitoring and analysis when the wind speed is in a specific interval;

[0050] Real sunlight determination module: including pseudo-peak confidence evaluation unit and double verification unit;

[0051] Execution control layer: including cooperative prediction unit, closed-loop execution unit, dynamic parameter optimization unit and back difference compensation unit;

[0052] Row-to-row cooperation module: including graph neural network coupling estimation unit, multi-row irradiance field cross correction unit, consistency checking unit and cooperative execution synchronization unit.

[0053] As described above, due to the adoption of the photovoltaic support adaptive tracking control method and system, the present application has the following advantages:

[0054] 1. The present application realizes real sunlight determination through multi-dimensional feature fusion, combines irradiance field light gradient analysis, photovoltaic panel temperature synchronization verification and cloud shadow edge position matching, and can effectively identify pseudo-peak values formed by adjacent row reflected light, ground object reflected light and cloud shadow edge local strong light; when suspicious pseudo-peak values are detected, further verification is carried out through the double mechanism of historical irradiation curve comparison and solar geometric position verification to ensure the accuracy of the determination of the position of sunlight, avoid tracking deviation caused by pseudo-peak value misjudgment in traditional single-point measurement or current feedback strategy, and make the photovoltaic support always align with the real optimal light direction, thereby reducing the risk of local heating and aging of the components;

[0055] 2. For the problem of random fluctuation of cloud shadow attitude, the present application constructs a cloud shadow rolling dynamic model and a cloud shadow transmittance grading model; the former simulates the morphological changes of the cloud layer under the action of wind force through fluid mechanics principle, predicts the projection track and coverage range of the cloud shadow on the ground in advance, and reserves sufficient time for support adjustment; the latter quantifies the light attenuation degree of different cloud shadow areas through cloud type identification, cloud layer thickness estimation and real-time transmittance updating. The system can dynamically optimize the support angle, for example: when the cloud shadow approaches quickly, the adjustment speed is accelerated to avoid obstruction, and when the cloud shadow moves slowly, the adjustment is stable to reduce mechanical loss, effectively reducing the lag of light capture caused by cloud shadow obstruction, and improving the power generation income under cloudy or intermittent obstruction conditions;

[0056] 3. In the scenario of multi-row photovoltaic array, the present application constructs an array topology correlation graph through a graph neural network, taking each row of support as a node and the inter-row shading and reflection relationship as an edge, and dynamically corrects the edge weight in combination with array geometric parameters and cloud shadow influence to accurately depict the inter-row coupling effect. At the same time, through the cross correction mechanism of multi-row irradiation field, the false components of reflected light enhanced by cloud shadow shading in the front row are deducted, and the light fluctuation of the cloud shadow edge row is smoothed to avoid the problem of intensified pseudo-peak value caused by inter-row interference. In addition, the swarm intelligence consistency verification and collaborative trajectory optimization ensure the consistency of the angle adjustment trend of each row, prevent new shading or reflected light interference between rows caused by independent adjustment of a single row, and maximize the overall power generation efficiency of the array, which is particularly significant in centralized power stations with high-density arrangement;

[0057] 4. The present application takes power generation income as the core target, quantifies the motor action cost and support fatigue damage at the same time, and sets safety constraints such as wind speed and angle limit. For example, when cloud shadow moves quickly, the system appropriately reduces the action cost weight to speed up the adjustment; when the cloud shadow is stable or the wind speed is too high, the device load is reduced by limiting the adjustment amplitude and speed. In addition, the back difference compensation unit corrects the angle error caused by mechanical clearance in real time, and the vibration spectrum monitoring avoids resonance caused by collaborative action, which not only guarantees efficient power generation, but also reduces mechanical wear and tear of the support, prolongs the service life of the equipment, and improves the economic efficiency and safety of long-term operation of the photovoltaic power station support;

[0058] 5. Through multi-modal sensing layer and hierarchical control logic, it can adapt to different installation scenes such as single support, multi-row array, and complex weather conditions such as sunny, cloudy, and moving cloud shadow; through edge computing, data preprocessing, model construction and decision execution are realized throughout the process, without manual intervention to adjust parameters, which can not only meet the flexible needs of distributed rooftop photovoltaics, but also adapt to the standardized operation and maintenance of large-scale centralized photovoltaic power station supports, reduce the complexity of on-site operation, and improve the universality of system deployment and application. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The flowchart of the present application is shown;

[0060] Figure 2 The system block diagram of the present application is shown;

[0061] Figure 3 The side view of the photovoltaic support applied in the present application is shown;

[0062] Figure 4 The overall installation diagram of the photovoltaic support applied in the present application is shown.

[0063] 1. Photovoltaic module; 2. Electrically controlled hydraulic telescopic rod; 3. Beam; 4. Inclined beam; 5. Rotation shaft; 6. Rear support. DETAILED DESCRIPTION

[0064] The technical photovoltaic support adaptive tracking control method and system in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0065] In order to more clearly and intuitively show the practical application effect and advantages of the photovoltaic support adaptive tracking control method and system proposed in the present application, and verify its feasibility and effectiveness, the present application will be further described below in conjunction with embodiments. Through specific scene simulation and data operation, how the system plays a role in actual photovoltaic support tracking is elaborated in detail, which helps readers better understand the technical details and practical value of the present application. The present application will be further described below in conjunction with embodiments;

[0066] Embodiment one:

[0067] Referring to Figure 1 , Figure 3 and Figure 4 , in the single photovoltaic support control scene, the photovoltaic support basic structure on which the system of the present application relies is: the overall framework is built with C-shaped steel standard profiles; the photovoltaic front support is an electrically controlled hydraulic telescopic rod 2; the photovoltaic rear support 6 is a triangular structure, and a rotating shaft 5 is arranged at the top end of the rear support 6; the rear support 6 is rotationally connected with a diagonal beam 4, and a beam 3 is fixedly installed on the diagonal beam 4; the photovoltaic module 1 is installed in a horizontal manner, and 5 to 8 modules are taken as a group, and through multi-level collaborative operation, accurate identification and angle dynamic adjustment of sunlight are realized, and tracking deviation caused by pseudo peak value is avoided throughout the process, and the specific process is as follows:

[0068] First, the sunlight identification process is performed; after the system is started, multiple sensing units are synchronously operated in the multi-modal sensing layer:

[0069] The fisheye panoramic imaging unit is installed at the top diagonal beam 4 of the C-shaped steel support, which is 3-4m away from the ground, and has no obstructed view. The resolution of the fisheye panoramic imaging unit needs to be ≥20 million pixels, and a 180° sky image is collected every 50ms, which completely records the cloud shape, the sun position and the sky brightness distribution;

[0070] The irradiation sensor (measurement range 0-2000W / m²) is closely attached to the surface of the middle photovoltaic panel of the horizontal module group, at 1 / 3 of the distance from the edge of the module, so as to avoid edge shielding interference. The irradiation sensor collects the light intensity of the current position in real time, and the sensing range covers the full spectrum band of natural light;

[0071] The temperature sensor has a precision of ±0.5℃, and the temperature sensors are uniformly distributed on the back of each photovoltaic panel of the horizontal assembly group. Each assembly is arranged with 2-3 temperature sensors covering the four corners and the center of the assembly to accurately capture the temperature changes of each area of the panel, especially the temperature abnormal points that may be caused by local shading;

[0072] The sampling rate of the attitude sensor is 100Hz, and the attitude sensor is fixed at the rotating shaft 5 at the top end of the photovoltaic rear support 6. The attitude sensor is coaxially installed with the rotating shaft 5 to ensure that the attitude data is synchronized with the rotation of the support and to provide real-time feedback of the current inclination angle, rotation direction and movement rate of the support.

[0073] The fisheye panoramic imaging unit can completely retain the dynamic details of the cloud shape and the position of the sun by collecting sky images at a high frequency, avoiding missing key information due to too long sampling intervals. The irradiation sensor is closely attached to the surface of the horizontal assembly and covers the full spectrum band, which can accurately capture real-time fluctuations in light intensity. The temperature sensor covers the entire domain of the horizontal assembly group, which can timely detect temperature abnormalities caused by local shading. The attitude sensor is coaxial with the rotating shaft 5 of the rear support 6, which ensures that the system can real-time master the current angle and movement trend of the support, and provides accurate attitude basis for angle adjustment decision.

[0074] The specific steps of integrating data through the multi-source data space-time alignment model to generate a multi-dimensional perception matrix are as follows:

[0075] A timestamp is added to the original data of the fisheye imaging, irradiation sensor, temperature sensor and attitude sensor, and the precision of the timestamp is ≤1ms. The time deviation of each sensor is corrected through a time deviation correction formula to ensure the consistency of the data time dimension. The time deviation correction formula is as follows:

[0076] ;

[0077] Wherein, is the synchronized time stamp; is the original time stamp of the sensor, is the time deviation of the qth sensor (obtained by pre-calibration, fixed period update).

[0078] The local collection coordinates of each sensor, such as the panel position of the temperature sensor and the installation point of the irradiation sensor, are mapped to the global coordinate system of the photovoltaic support through a space coordinate mapping formula to form a data matrix with a unified space dimension. The space coordinate mapping formula is as follows:

[0079] ;

[0080] Wherein, is the local coordinate of the e th sensor, θ is the current inclination angle of the photovoltaic support, and h is the installation height of the sensor, for the mapped global coordinates;

[0081] The time deviation correction eliminates the time difference of each sensor due to the hardware response speed and data transmission delay through timestamp and dynamic deviation compensation, ensures that the fisheye sky image, irradiation data, temperature data and attitude data are completely synchronized in the time dimension, avoids the multi-source data association error caused by different time synchronization, such as preventing the error matching of the light intensity at a certain time and the support angle at another time, and affects the accuracy of subsequent irradiation field modeling and angle adjustment decision; the space coordinate mapping unifies the local coordinates of each sensor to the global coordinate system, so that the data collected by different position sensors can correspond to the actual spatial position of the photovoltaic support, clearly present the spatial distribution rule of light intensity and temperature on the surface of the photovoltaic panel, such as the correlation of light intensity difference and temperature distribution of different regions of the photovoltaic panel, and provide spatial dimension analysis basis for identifying local occlusion and reflected light interference.

[0082] The specific steps of the fisheye lens distortion correction model to which the collected sky image first enters are as follows:

[0083] Based on the lens factory parameters and the field calibration data, a polynomial fitting formula and a distortion correction formula are established;

[0084] At least 10 groups of calibration sky images with different angles are collected through the chessboard calibration plate, the corner point coordinates are extracted, and the distortion after the pixel point to the image center distance coefficient (polynomial fitting formula) is calculated.

[0085] For sky image data, the distortion after the pixel point to the image center distance of each pixel point is corrected in reverse through the distortion correction formula, and a non-distortion image is output.

[0086] Based on the distortion after the pixel point to the image center distance coefficient and the ideal non-distortion pixel point to the image center distance obtained by fitting the calibration experiment, the distortion after the pixel point to the image center distance is calculated through the polynomial fitting formula, and an example of the polynomial fitting formula is as follows:

[0087] ;

[0088] wherein, is the distance from the distorted pixel point to the image center, is the distance from the ideal non-distortion pixel point to the image center, , , is the distance from the distorted pixel point to the image center (obtained by fitting the calibration experiment);

[0089] Input the pixel point coordinates of the distorted image, the image center coordinates and the distortion coefficient, calculate the corrected pixel coordinates through the distortion correction formula, and the distortion correction formula is as follows:

[0090]

[0091] wherein, is the pixel point coordinates of the distorted image, is the image center coordinates, is the corrected pixel coordinates; , is the distance coefficient of the pixel point after distortion to the image center; is the distance of the pixel point after distortion to the image center;

[0092] Based on the distortion correction model of the chessboard calibration and polynomial fitting, the distance coefficient of the pixel point after lens distortion to the image center can be accurately calculated, the image distortion of the edge area of the fisheye lens is eliminated through the reverse mapping algorithm, the edge cloud shape and the position of the sun of the corrected sky image are highly consistent with the actual physical position, the problems such as misjudgment of the cloud boundary and deviation of the sun position caused by lens distortion are avoided, for example, the distorted cloud profile is prevented from being misjudged as the actual shielding range, or the tracking direction deviation caused by the deviation of the sun position, which lays an accurate image foundation for subsequent cloud segmentation and sun area identification.

[0093] Subsequently, the semantic segmentation model based on deep learning automatically distinguishes the sky background, the cloud layer and the sun area: the neural network with U-Net architecture can accurately extract the pixel-level boundary of the cloud layer through the training of a large number of labeled images; combined with the optical flow tracking algorithm of continuous multiple frames of images, the displacement vector of the cloud feature points (such as edge corner points and gray value) in adjacent frames is calculated, the moving direction and speed of the cloud layer are determined, and the dynamic cloud trajectory atlas is generated; at the same time, the visible area in the sky, i.e. the part not blocked by the cloud layer, is identified, and the real-time updated sky visible mask is generated, the value of each pixel point in the sky visible mask represents the probability (0%~100%) of the position being blocked by the cloud layer, which provides a quantitative basis for the subsequent irradiation field reconstruction.

[0094] ​The irradiance field modeling module receives the preprocessed sky image, the light intensity and the attitude data, and starts to construct the irradiance field model. The module first distinguishes the direct light, the scattered light and the environmental reflected light: the direct light is the light directly irradiated by the sun to the photovoltaic panel, the intensity of which is closely related to the solar elevation angle and the atmospheric transparency; the scattered light is the diffuse light formed by the reflection of the sunlight through the atmospheric molecules and the cloud particles, which is uniformly distributed; the environmental reflected light includes the light reflected to the photovoltaic panel by the ground, the surrounding structures and other objects, the intensity of which is affected by the material and the angle of the reflecting surface. The irradiance field modeling module generates the irradiance field model containing the time evolution trend by fusing the spatial distribution characteristics of the three kinds of light, and clearly presents the distribution of the light intensity on the surface of the photovoltaic panel and the change rule with time.

[0095] The method for constructing the irradiance field model comprises:

[0096] calculating the solar elevation angle a and the azimuth angle g based on the geometric position (longitude , latitude λ and time t) of the sun, combining the shielding probability of the sky visible mask , calculating the direct light intensity by the direct light intensity formula and deducting the influence of the cloud shielding; the direct light intensity formula is as follows:

[0097] ;

[0098] wherein, is the direct light intensity; is the solar irradiance intensity at the top of the atmosphere (standard value 1367 W / m²), is the atmospheric transparency coefficient (obtained by analyzing the solar spectrum), a is the solar elevation angle, is the shielding probability of the sky visible mask (0≤ ≤1), and exp(˙) represents the natural exponential function.

[0099] calculating the hemispherical space scattered light intensity based on the solar irradiance intensity at the top of the atmosphere , the cloud coverage , the aerosol concentration value by the scattered light intensity formula; the scattered light intensity formula is as follows:

[0100] ;

[0101] wherein, is the scattered light intensity; is the cloud coverage (0≤ ≤1); is the aerosol concentration value, the aerosol concentration is standardized to the range of 0-1 to obtain the aerosol concentration value; is the atmospheric transparency coefficient, and exp(˙) represents the natural exponential function.

[0102] Call the reflectivity of the reflection surface material database , combined with the incident light intensity of each type of reflection surface s and the angle between the reflected light and the normal of the photovoltaic panel , the ambient reflected light intensity is calculated by summing the ambient reflected light intensity formula; the ambient reflected light intensity formula is as follows:

[0103] ;

[0104] Among them, is the ambient reflected light intensity, is the incident light intensity of the s-th type of reflection surface (ground, building, etc.), is the reflectivity of the s-th type of reflection surface, is the angle between the reflected light and the normal of the photovoltaic panel, and s is the s-th type of reflection surface.

[0105] By weighting and fusing the direct light intensity, the scattered light intensity, and the ambient reflected light intensity, the weight coefficient is dynamically adjusted according to the real-time weather type, and the irradiation field model is generated, and a specific example is as follows:

[0106] ;

[0107] Among them, is the irradiance; is the direct light intensity; is the scattered light intensity; is the ambient reflected light intensity; , , is the weight coefficient (dynamically adjusted according to the real-time weather type, sunny =0.7, cloudy =0.4);

[0108] The calculation of direct light intensity combines the solar geometric position and the sky visible mask, can accurately deduct the influence of cloud cover on direct light, avoid overestimation of direct light intensity due to neglecting cloud cover, for example, in cloudy weather, accurately calculate the direct light intensity penetrating the cloud layer, provide accurate basis for photovoltaic panel power estimation; The calculation of scattered light intensity integrates cloud coverage and aerosol concentration, which can dynamically adapt to the scattering light distribution characteristics under different weather conditions, for example, in foggy weather, consider the enhancement effect of high aerosol concentration on scattered light, avoid estimation deviation caused by fixed scattered light model; The calculation of environmental reflected light intensity calls the reflection surface material database, combined with the support posture data to quantify the reflection effect of different reflection surfaces, can effectively identify the false light area formed by ground and building reflected light, avoid tracking deviation caused by misjudging reflected light as direct light; The weighted fusion dynamically adjusts the weight coefficient, so that the irradiation field model can adapt to different weather types such as sunny and cloudy days, accurately reflect the actual distribution and time evolution trend of photovoltaic panel surface light intensity, for example, in sunny weather, focus on the weight of direct light, increase the weight of scattered light in cloudy weather, ensure that the irradiation field model is always highly matched with the actual lighting environment.

[0109] The specific steps of determining whether the current area wind speed is in a specific interval are as follows:

[0110] Wind speed sensors are deployed at key positions of the photovoltaic support array, such as array edges and middle high points, to collect environmental wind speed in real time. The sampling frequency is dynamically adjusted according to actual needs: reduce the frequency to save energy in a light wind environment, and increase the frequency to ensure data continuity in complex wind conditions.

[0111] The collected raw wind speed is subjected to Kalman filtering processing (Kalman filtering prediction equation and Kalman filtering update equation), to eliminate transient pulse interference (such as short-term abnormal values caused by gusts) and retain effective data reflecting the true wind conditions.

[0112] The Kalman filtering prediction equation is as follows:

[0113]

[0114]

[0115] wherein, is the predicted wind speed at time k, is the filtered wind speed at time k-1, a is the wind speed acceleration (default 0, stable wind conditions), Δt is the sampling interval, is the prediction covariance, P(k-1|k-1) is the updated covariance at time k-1, Q is the process noise covariance.

[0116] The Kalman filtering update equation is as follows:

[0117] ​​;

[0118] ;

[0119] ;

[0120] wherein K(k) is the Kalman gain, is the predicted wind speed at k, is the filtered wind speed at k-1, and R is the measurement noise covariance, is the filtered wind speed at k, is the updated covariance; P(k-1|k-1) is the updated covariance at k-1; is the wind speed observation value at k (unit: m / s); I is the unit matrix.

[0121] The correlation coefficient between the wind speed and the cloud layer movement rate and the support vibration characteristics is calculated through the data credibility correlation coefficient to determine the data effectiveness;

[0122] The covariance and variance between the wind speed, the cloud layer movement rate and the support vibration amplitude are calculated; the data credibility correlation coefficient is calculated through the data credibility correlation coefficient formula; the data credibility correlation coefficient formula is as follows:

[0123] ;

[0124] wherein, is the data credibility correlation coefficient, Cov is the covariance, and Var is the variance, is the wind speed, is the cloud layer movement rate, is the support vibration amplitude; when the data credibility correlation coefficient > 0.8, the wind speed is determined to be effective.

[0125] When determining whether the wind speed in the current region is in a specific interval, the threshold value needs to be optimized in combination with the support structure characteristics:

[0126] In combination with the wind load design value of the triangular rear support 6, the bending stiffness of the C-shaped steel overall frame, and the wind pressure resistance of the transverse component group, deformation of the triangular rear support 6 or damage of the C-shaped steel is avoided due to excessively high wind speed; in combination with the vibration frequency and amplitude data (the support vibration characteristics will change correspondingly when the wind speed increases) fed back by the support posture sensor, and the cloud layer movement rate (there is a correlation between the wind speed and the cloud layer movement speed) captured by the sky imaging unit, multi-dimensional wind speed sensing data is formed; a wind speed interval threshold value suitable for the current scene is generated;

[0127] Multi-dimensional wind speed perception combines wind speed sensor, support vibration data and cloud layer movement rate to comprehensively verify the authenticity and effectiveness of wind speed, avoid wind speed misjudgment caused by local airflow disturbance of single wind speed sensor, for example, through the correlation of support vibration frequency, amplitude and wind speed, verify whether the short-time wind speed anomaly caused by gust is real wind condition, prevent unnecessary safety protection or cloud shadow analysis caused by false wind speed; Kalman filter processing effectively eliminates transient pulse interference in wind speed, retains smooth data reflecting real wind condition, provides stable data source for wind speed interval judgment, avoids frequent fluctuation of wind speed caused by gust, affects start-stop decision of cloud shadow analysis mechanism; Dynamic wind speed interval threshold combines support structure characteristics and local climate characteristics to ensure that cloud shadow analysis is started only in suitable wind conditions, i.e. wind can push cloud layer to move quickly without damaging support structure, avoid frequent adjustment of support caused by cloud shadow analysis in strong wind weather, increase mechanical loss and safety risk, or save system computing resources in light wind weather without frequent analysis due to slow cloud shadow movement.

[0128] When the current area wind speed is in a specific interval, i.e. wind is strong enough to push cloud layer to move quickly without damaging support structure, cloud shadow modeling module will automatically start cloud shadow monitoring and analysis mechanism;

[0129] It should be noted that the specific steps of cloud shadow monitoring and analysis mechanism are as follows:

[0130] Based on the fluid dynamics of cloud shadow rolling dynamic model, the dynamic position coordinate sequence of cloud shadow projection on the ground is output; the cloud shadow rolling dynamic model is as follows:

[0131] Step 0A: adopt Canny operator to perform edge detection on sky image, extract pixel coordinate point set of cloud layer contour , wherein, is the horizontal coordinate of the nth contour pixel, is the vertical coordinate of the nth contour pixel; through B-spline curve fitting, the pixel coordinate point set is converted into continuous edge function E(u,v)=0; the normal vector of each pixel coordinate point of cloud layer contour is calculated , wherein, E is the continuous edge function, u is the horizontal coordinate parameter of continuous curve, corresponding to the image horizontal coordinate direction; v is the vertical coordinate parameter of continuous curve, corresponding to the image vertical coordinate direction; is the first order partial derivative of E to u, reflecting the change rate of E along u direction; is the first order partial derivative of E to v, reflecting the change rate of E along v direction; the normal vector represents the direction of cloud layer contour;

[0132] Based on fluid mechanics, this paper describes the change of cloud vortex vectors over time, including the diffusion term caused by viscosity and the stretching term generated by the interaction between the vortex and the velocity field. The cloud motion satisfies the simplified form of the Navier-Stokes equations in fluid mechanics (ignoring the pressure gradient term and retaining the vortex viscosity), as shown in the following example:

[0133] ;

[0134] in, is the cloud vortex vector (characterizing the intensity of cloud rotation), and ∇ represents the gradient operator; This represents the cloud's velocity vector. It is the eddy viscosity coefficient (positively correlated with cloud density and wind speed); The volume derivative (tracking the motion of cloud particles); the volume rate of change of cloud vortices (i.e., It is determined jointly by two parts. This is due to the diffusion of cloud vortices caused by viscosity. The equation represents the cloud vortex stretching caused by the interaction between the vortex and the velocity field. This equation can quantitatively describe the dynamic evolution of clouds under wind action, including rotation, stretching, and twisting.

[0135] Calculate the rate of change of the cloud shadow's position on the ground based on the projected velocity of cloud motion on the ground and the drift velocity of the cloud shadow caused by the apparent motion of the sun; an example is shown below:

[0136] ;

[0137] in, The rate of change of the cloud shadow's position over time; The projected velocity of cloud movement on the ground. The speed of cloud shadow drift caused by the apparent motion of the sun (caused by the Earth's rotation);

[0138] Step 0B: Calculate the cloud motion velocity vector at the initial moment based on the optical flow algorithm. The normal component of the velocity at the cloud contour is set to 0 (no penetration condition); the finite difference method is used to solve the vortex motion control equations to obtain the velocity of the cloud contour at a specific moment. This yields the velocity field of the cloud contour at a specific moment; by integrating the velocity field at that specific moment, the position of the cloud contour at time t+Δt is obtained. ;

[0139] Step 0C: Obtain the position coordinates of the cloud outline at a specific moment. Substituting into the perspective projection formula, we obtain the coordinates of the ground cloud shadow outline. Fill the interior of the cloud shadow outline to generate a binary cloud shadow mask covering the cloud shadow area. (1 represents covered by cloud shadow, 0 represents no coverage);

[0140] Step 0D: Contrast cloud shadow mask Calculate the number of photovoltaic panels covered by cloud shadow in relation to the position of the ground photovoltaic panel array;

[0141] Combined with the irradiance sensor data, if the actual light intensity drop area deviates from the cloud shadow mask prediction area by more than 5%, the vortex viscosity coefficient is corrected by the formula, for example:

[0142]

[0143] Wherein, is the corrected vortex viscosity coefficient; is the vortex viscosity coefficient before correction; err is the prediction deviation rate;

[0144] Repeat steps 0B-0D to achieve dynamic tracking of cloud shadow, with an update frequency consistent with the fisheye sky image acquisition frequency, i.e. 50ms / second.

[0145] Through the above formula and steps, the cloud shadow rolling dynamic model can accurately simulate the stretching, shrinking, breaking and other morphological changes of cloud shadow under different wind conditions, providing 10-30 seconds of cloud shadow position prediction for angle adjustment of photovoltaic support;

[0146] Based on the vortex motion equation of fluid mechanics, the stretching, twisting, breaking and other dynamic morphological changes of cloud layer under the action of wind can be accurately simulated, combined with perspective projection and mask generation, the projection position and coverage range of cloud shadow on the ground can be accurately predicted, and the moving track of cloud shadow can be predicted 10-30 seconds in advance, providing sufficient decision-making time for angle adjustment of support, for example, predicting that cloud shadow will cover a certain area of photovoltaic panel after 20 seconds, the system can adjust the angle in advance to avoid cloud shadow shading; Dynamic correction of vortex viscosity coefficient optimizes model parameters in real time according to the actual deviation of irradiance sensor, ensuring that the cloud shadow prediction area is highly consistent with the actual light intensity drop area, avoiding prediction deviation caused by fixed model parameters, for example, when the actual coverage range of cloud shadow deviates from the prediction by more than 5%, the parameters are corrected in time to improve the prediction accuracy.

[0147] Based on the cloud shadow light transmittance grading model, the cloud shadow light transmittance is divided into 5 levels, of which level 1 is completely opaque and level 5 is high transmittance, and the level distribution is updated in real time with the movement of cloud layer;

[0148] Cloud shadow light transmittance grading model:

[0149] ​The cloud shadow transmittance grading model quantifies the degree of cloud layer blocking sunlight, and realizes the dynamic evaluation of light transmittance in different cloud shadow areas. The model takes the cloud layer features of the fisheye sky image as input, combines the statistical rules of the historical cloud shadow database, and divides the transmittance into multiple quantitative indicators, providing accurate basis for the light attenuation of the photovoltaic support tracking strategy;

[0150] Step 1A: Based on the U-Net semantic segmentation model, the cloud type classification of the fisheye sky image is carried out, and the pixel-level cloud type coefficient C is output, where cumulus C=1.0, stratus C=0.7, cirrus C=0.4, and no cloud C=0; The gray value of the target cloud layer area is extracted; combined with the historical database, the cloud type classification result is verified by support vector machine, and the accuracy rate needs to be ≥90%;

[0151] Step 1B: Select the pixel-level cloud type coefficient of the image as the reference of the cloud-free area, calculate the average gray value of the cloud-free area, and remove the atmospheric scattering interference through the atmospheric correction algorithm, wherein the atmospheric correction algorithm is one of the core technologies in remote sensing data processing, its essence is to eliminate or weaken the interference of atmospheric absorption, scattering, reflection, etc. on solar radiation, from the apparent radiation signal obtained by the remote sensing sensor, a series of mathematical models and calculation processes are used to retrieve the real radiation and reflection information of the ground; The gray value of the target cloud layer area is counted, and the cloud layer thickness D is calculated by substituting the formula; According to the cloud type coefficient C, the initial transmittance level is obtained through the level conversion formula ;

[0152] Based on the average gray value of the cloud-free area, the average gray value of the target cloud layer area, the correction coefficient (dynamically adjusted according to the cloud type), and the preset average thickness reference value of the cloud layer, the cloud layer thickness is calculated through the cloud layer thickness estimation formula. For example, the cloud layer thickness estimation formula is as follows:

[0153] ;

[0154] Wherein: D is the cloud layer thickness, is the average gray value of the cloud-free area (0-255), is the average gray value of the target cloud layer area, is the correction coefficient (dynamically adjusted according to the cloud type, cumulus =1.2, stratus =0.9), is the preset average thickness reference value of the cloud layer, unit: m, calibrated by experiment;

[0155] Based on the cloud layer thickness, the preset reference thickness, and the average gray value of the target cloud layer area, the transmittance level is calculated by taking the integer part and limiting the range. For example, the calculation formula of the transmittance level T (1-5 levels, 1 level completely not transparent, 5 level high transmittance) is as follows:

[0156] ;

[0157] Wherein: T is the light transmittance level, when the light transmittance level conversion formula is negative, the light transmittance level T is 1 level; min is the minimum value; max is the maximum value; is the floor function; D is the cloud thickness; is the average gray value of the target cloud area; is the preset reference thickness, such as 1m.

[0158] Step 1C: Based on the cloud thickness of the target cloud area in the continuous frame image, the thickness change rate is calculated ; combined with the current wind speed The characteristic time constant τ=30 / (1+0.1 (current wind speed unit: m / s); substitute the light transmittance update formula, real-time adjust light transmittance level T(t), update frequency consistent with image acquisition frequency (50ms / time); light transmittance update formula, for example:

[0159] Considering the thickness change caused by cloud movement, the light transmittance T(t) at time t is corrected by the initial light transmittance level and the thickness change rate :

[0160] ;

[0161] Wherein, T(t) is the light transmittance at time t; is the initial light transmittance level; is the maximum change of light transmittance level (±2 levels); sgn(˙) represents the sign function; is the cloud thickness change rate, positive value of cloud thickness change rate indicates thickening, negative value of cloud thickness change rate indicates thinning; τ is the characteristic time constant, default 30 seconds, decrease when wind speed increases.

[0162] Step 1D: Compare the measured value of the contrast irradiation sensor with the theoretical irradiation attenuation corresponding to the light transmittance level (1 level attenuation 90%, 2 level attenuation 70%, 3 level attenuation 50%, 4 level attenuation 30%, 5 level attenuation 10%);

[0163] If the deviation err=|measured attenuation-theoretical attenuation|>15%, the above correction coefficient is corrected: , wherein, is the corrected correction coefficient; iterative optimization model parameters, ensure that the matching degree of light transmittance level and actual light attenuation ≥85%.

[0164] Through the above formula and steps, the cloud shadow transmittance grading model can realize dynamic quantification of the transmittance of different types and thicknesses of cloud layers, provide key light attenuation basis for angle adjustment of the photovoltaic support during the movement of the cloud shadow, and especially effectively improve the tracking accuracy under cloudy weather.

[0165] The cloud type classification combined with the thickness estimation, the gray value and the historical database verification can accurately identify different cloud types such as cumulus, stratus and cirrus, and quantify the thickness of the cloud layer, thereby providing an accurate basis for the transmittance grade division and avoiding errors in the transmittance grade caused by misjudgment of the cloud type or deviation in the thickness estimation; the transmittance grade dynamic update combined with the cloud layer thickness change rate and the wind speed correction can reflect the transmittance change caused by the movement of the cloud layer in real time, for example, timely lowering the transmittance grade when the cloud layer thickens, and raising the grade when the cloud layer thins, so as to ensure that the transmittance grade is consistent with the actual light attenuation degree; the iterative optimization correction coefficient continuously optimizes the model parameters by comparing the measured and theoretical irradiation attenuation amounts, thereby ensuring the matching degree of the transmittance grade and the actual light attenuation, providing accurate light attenuation basis for the angle adjustment of the support, for example, accurately judging the light intensity in the cloud shadow area according to the transmittance grade, and avoiding insufficient or excessive angle adjustment caused by errors in the light attenuation estimation.

[0166] The real sunlight determination module includes a false peak confidence evaluation unit and a double verification unit, and the specific steps of real sunlight determination are as follows:

[0167] The false peak confidence evaluation unit analyzes based on the generated irradiation field model and the cloud field optical flow data:

[0168] The light intensity gradient of each region in the irradiation field is calculated, if a region has a sudden increase in light intensity but the surrounding region has no obvious change, that is, the light intensity gradient exceeds the set threshold, and the light intensity gradient does not match the cloud field optical flow (the light change trend caused by the movement of the cloud layer), then it is marked as a suspicious false peak;

[0169] Combined with the temperature sensor data, if the temperature of the region does not rise synchronously with the light intensity (the temperature difference exceeds 5°C) or a local high temperature point (more than 20°C higher than the ambient temperature) appears, then the false peak risk index is increased; if the suspicious false peak region coincides with the cloud shadow edge position predicted by the cloud shadow rolling dynamic model, then the false peak risk index is reduced. The calculation formula of the false peak risk index is as follows:

[0170] ;

[0171] Wherein, the false peak risk index (0-100), the higher the value, the more likely it is a false peak; , , the weight coefficient; the light gradient anomaly coefficient (0-1); temperature synchronization abnormality coefficient (0-1), close to 1 when temperature is not synchronized with light; cloud shadow edge coincidence coefficient (0-1), close to 1 when cloud shadow edge coincides;

[0172] When the pseudo-peak risk index is lower than the set threshold, the system determines that the current identified is the real sunlight, that is, at this time, the region with the highest light intensity in the irradiation field is consistent with the actual position of the sun, and the light intensity changes smoothly (the change rate of the last 3 sampling periods is less than 10%), which is consistent with the trend of cloud movement.

[0173] If the pseudo-peak risk index is higher than the set threshold, the system starts a double verification mechanism:

[0174] On the one hand, historical irradiation data of the same period (same season, similar weather) is called, and the shape characteristics of the current irradiation curve are compared through dynamic time warping algorithm. If the trends are consistent (the similarity is more than 85%), the confidence of the real sunlight is increased;

[0175] Specifically, the method for increasing the confidence of the real sunlight includes:

[0176] Set a basic confidence range, such as 50% to 70%, as the initial judgment before comparing the trends of the two;

[0177] If the trends are consistent, the confidence is increased according to the specific degree of consistency: if only the core trend is consistent and the detailed trend is slightly deviated, the confidence can be increased by 10% to 15% in the basic range, reaching 60% to 85%, wherein the core trend is, for example, the direction of intensity change, and the detailed trend is, for example, the change rate;

[0178] If the core trend and the detailed trend are highly consistent, such as the intensity change rate and the spectral stability completely match the real sunlight law, the confidence can be increased by 20% to 30%, and the highest can reach 80% to 99%; It should be noted that the upper limit of the confidence is not set to 100% to avoid misjudgment that may be caused by extreme interference factors, such as high-precision simulated sunlight equipment, to ensure that the judgment has basis and leaves room.

[0179] On the other hand, combined with the geometric position information of the sun (based on the local latitude and longitude, time calculation of the sun position), it is verified whether the region with the strongest light is consistent with the theoretical position of the sun (deviation less than 2°), and the position of the real sunlight is finally determined through comprehensive analysis;

[0180] The calculation formula of the multi-feature fusion false peak risk index can accurately identify suspicious false peaks through multi-dimensional analysis of illumination gradient, temperature synchronicity, and cloud shadow edge coincidence degree, avoid misjudging short-term power peaks caused by reflected light and local shielding as real sunlight, for example, a sudden increase in light intensity in a certain area but no synchronous temperature rise, and located at the edge of the cloud shadow, the system can accurately mark it as a false peak to prevent the tracking direction from being misled; The double verification mechanism combines historical irradiation data and solar geometric position to further verify the position of real sunlight. The historical data comparison ensures that the current irradiation curve is consistent with the similar weather trend at the same period by using the dynamic time warping algorithm, excluding the interference of abnormal irradiation fluctuations. The solar geometric position verification ensures that the area with the strongest illumination coincides with the theoretical position of the sun, avoiding deviation of the tracking direction caused by sensor errors or external interference, and finally achieving accurate determination of real sunlight to provide the correct target direction for subsequent angle adjustment.

[0181] The execution control layer performs the photovoltaic support angle adjustment process:

[0182] When the real sunlight is identified, the cooperative prediction unit starts to work. Based on the time evolution trend of the irradiation field model and the cloud shadow movement direction and speed provided by the cloud shadow rolling dynamic model, the cooperative prediction unit predicts the irradiation state in the short time to come. During the prediction process, the system will focus on the light mutation that may be caused by the movement of the cloud shadow (if the cloud shadow is moving towards the photovoltaic panel direction and is expected to reach in a short time), and then reserve a buffer interval for angle adjustment in the predicted trajectory, i.e. the trajectory candidate will contain an angle range rather than a fixed value, in order to cope with sudden changes in light. Finally, the cooperative prediction unit generates multiple sets of optimal tilt angle and azimuth angle trajectory candidates, each trajectory corresponds to a complete angle adjustment scheme, covering the angle change rhythm in the short time to come.

[0183] It should be noted that when the cooperative prediction unit generates multiple sets of optimal tilt angle and azimuth angle trajectory candidates, the support structure constraints need to be considered: the tilt angle trajectory needs to adapt to the stroke limit of the electric-hydraulic telescopic rod 2, and the azimuth angle trajectory needs to match the rotation range of the rear support 6 top end rotating shaft 5, to avoid exceeding the mechanical limit;

[0184] The specific steps of the cooperative prediction unit generating multiple sets of optimal tilt angle and azimuth angle trajectory candidates are as follows:

[0185] The time evolution trend of the irradiation field model, the cloud shadow movement direction and speed provided by the cloud shadow rolling dynamic model are input, and the long short-term memory network (LSTM) is used to predict the irradiation state in the short time to come;

[0186] In the prediction process, if the cloud shadow is moving towards the photovoltaic panel and is expected to arrive in a short time, the cooperative prediction unit automatically reserves a buffer interval for angle adjustment in the predicted trajectory (the interval size is positively correlated with the cloud movement rate), that is, the trajectory candidate contains an angle range rather than a fixed value.

[0187] Finally, 3-5 groups of optimal tilt angle and azimuth angle trajectory candidates are generated, each trajectory corresponding to a complete angle adjustment scheme, covering parameters such as starting time, rate, and end value of angle change.

[0188] Based on the time evolution trend of the irradiation field and the cloud shadow movement parameters, the irradiation state in the near future is predicted, and the sudden change in light caused by the movement of the cloud shadow is identified in advance, for example, predicting that the cloud shadow will cause a sudden drop in light intensity in 1 minute. A buffer interval is reserved for the trajectory candidate. The buffer interval is designed to dynamically set the angle range according to the cloud movement rate, avoiding the inability of fixed angle trajectories to respond to sudden changes in light, such as when the cloud shadow moves quickly, expanding the angle adjustment range to ensure that the support can be adjusted to the optimal angle in time when the light changes suddenly. Multiple trajectory candidates cover different angle adjustment schemes, providing a rich selection space for subsequent multi-objective decision-making, ensuring that the optimal scheme can be selected from multiple dimensions such as power generation revenue, action cost, and equipment wear and tear.

[0189] The trajectory candidates are evaluated by the dynamic weight formula of the decision control layer:

[0190] The trajectory candidates are evaluated by minimizing the objective function. The weight coefficients are dynamically adjusted considering the power generation power deviation, total angle adjustment amount, fatigue damage value, and safety constraint penalty term. For example:

[0191] ;

[0192] Where: J is the objective function value (dimensionless), which needs to be minimized; 、 、 、 is the dynamic weight coefficient; is the expected power generation power (unit: W); is the predicted power generation power (unit: W); is the total angle adjustment amount (unit: °); is the fatigue damage value (dimensionless, 0-1), calculated based on the material fatigue accumulation model; is the safety constraint penalty term (dimensionless), which significantly increases when the constraint is violated; is the preset rated power; is the maximum allowed angle adjustment amount;

[0193] Expected power generation: According to the irradiance intensity and angle matching degree (calculated by the photovoltaic panel power curve), quantify the power generation benefits that each trajectory can bring;

[0194] Total angle adjustment amount: Statistically summarize the total amplitude and frequency of angle adjustment in the trajectory, combine the motor energy consumption parameters and mechanical wear coefficient, and calculate the comprehensive cost of the action;

[0195] Fatigue damage value: Combine the material fatigue characteristics of the C-shaped steel support, the piston seal life of the electrically controlled hydraulic telescopic rod 2, and the fatigue strength of the welded nodes of the triangular rear support 6 to ensure that the evaluation results fit the actual wear of the support. Based on the fatigue accumulation model of material mechanics, analyze the damage degree of the support structure caused by frequent adjustment;

[0196] Safety constraint penalty term: Combine the current wind speed, temperature, and other environmental parameters to set the safety boundary for angle adjustment (such as limiting the maximum adjustment amplitude when the wind speed exceeds the threshold).

[0197] When the cloud shadow coverage area is large and the moving speed is fast, the system reduces the weight of the action cost through the real-time weight adjustment mechanism (the weight value decreases from 0.3 to 0.1), and prioritizes the photovoltaic panel to quickly adjust the angle to capture limited light resources. For example, if it is predicted that the cloud shadow will block the photovoltaic panel in a short time, the system will choose a trajectory with a larger adjustment amplitude but can quickly align with the remaining light area to ensure more power is obtained before the cloud shadow completely blocks it. After comprehensive trade-off, the MPC module determines the final angle adjustment scheme from the candidate trajectories.

[0198] After the execution control layer receives the final angle instruction, the electrically controlled hydraulic telescopic rod 2 of the photovoltaic front support and the top shaft 5 of the rear support 6 cooperate to act: the electrically controlled hydraulic telescopic rod 2 adjusts the inclination angle of the support by telescoping, and the rear support 6 shaft 5 adjusts the azimuth angle by rotating, with the speed ratio of the reducer switching to 1:10 at high speed and 1:50 for precise positioning; the encoder collects the actual angle every 10ms, compares it with the instruction angle to form a closed loop, and combines the installation reference of the C-shaped steel support; during the adjustment process, the closed-loop control unit strictly follows the action amplitude limiting rule (the angle change does not exceed ±90° mechanical range), and according to the preset acceleration and speed profile (such as S-curve), the support moves smoothly, avoiding mechanical impact caused by sudden stop and sudden rotation.

[0199] In particular, the dynamic parameter optimization unit dynamically optimizes the motion parameters according to the cloud layer movement rate predicted by the cloud shadow rolling dynamic model, and the specific steps are as follows:

[0200] If the cloud shadow approaches quickly (the moving speed exceeds the set threshold), the dynamic parameter optimization unit automatically increases the acceleration (from 0.5° / s² to 1° / s²) and running speed (from 3° / s to 5° / s) of the motor, shortening the angle adjustment time;

[0201] If the cloud shadow moves slowly or gradually moves away, reduce the adjustment speed to minimize mechanical wear and ensure that the photovoltaic panel is aligned with the area of ​​strongest sunlight before the cloud shadow arrives.

[0202] It should be noted that when the dynamic parameter optimization unit adjusts the motor parameters according to the cloud movement rate, the response delay of the electro-hydraulic telescopic rod 2 must be considered; when the hysteresis compensation unit corrects the mechanical backlash, it must consider the installation accuracy of the C-shaped steel bracket and the fitting clearance of the rotating shaft 5 of the rear bracket 6; the hysteresis compensation unit will correct the angle error caused by the mechanical backlash in real time, and continuously fine-tune the motor output by comparing the actual angle fed back by the encoder with the command angle through the hysteresis compensation correction formula. Each angle adjustment is 0.1°-0.3°, ensuring that the deviation between the final angle and the target value is controlled within ±0.1°. An example of the hysteresis compensation correction formula is as follows:

[0203] ;

[0204] in, To compensate for the angle; The compensation coefficient is determined by the mechanical clearance characteristics; From the perspective of the target; This is the actual angle fed back by the encoder.

[0205] Example 2:

[0206] See Figure 2 - Figure 4 In the multi-row array photovoltaic support control scenario, each row of supports adopts the basic structure of Example 1 (C-shaped steel integral frame, front electro-hydraulic telescopic rod 2, rear triangular structure and pivot 5, and 5-8 horizontally placed modules as a group), and the spacing between rows must meet the array layout requirements (row spacing ≥ 2.5 times the module height to avoid inherent shading of the rear row by the front row when there is no cloud shadow). Based on the individual support control, the row-to-row coordination module strengthens the row-to-row coordination mechanism. Through graph neural network coupling estimation unit, multi-row irradiance field cross-correction unit, consistency verification unit, and collaborative execution synchronization unit, the overall angle optimization is achieved. The specific process is as follows:

[0207] Each row of photovoltaic support structures in the array is equipped with an independent sensing unit and computing module, and initial identification is performed following the identification process for each individual support structure. During this process, each row deeply integrates locally sensed cloud shadow information into the identification process. The specific steps are as follows:

[0208] Analyzing the light intensity change curve of the self-position, whether there is a sudden drop or sudden rise matching the cloud shadow movement, combined with the installation position of the C-shaped steel support (relative to the array reference line offset ≤0.3m), to exclude the light data deviation caused by the deviation of the support installation; Monitor the temperature field distribution (the temperature field presents uniform cooling characteristics when covered by cloud shadow), need to cover the whole domain of the horizontal component group (5-8 components per row, 2-3 temperature sensors per component), to ensure the accuracy of the uniform cooling characteristics of the temperature field when covered by cloud shadow; Extract the relative relationship between the cloud shadow edge in the sky image and the self-position;

[0209] Determine whether it is in the cloud shadow coverage area through the above analysis, if so, record the coverage range, entry time and estimated duration of the cloud shadow (based on the cloud layer movement rate calculation).

[0210] The distributed computing architecture ensures that each row can process data synchronously through the edge node parallel processing mechanism, quickly completes the analysis of the light situation under the cloud shadow coverage, for example: if a row detects a sudden drop in light intensity and a uniform temperature field cooling, combined with the moving track of the cloud layer in the sky image, it can be determined that it has entered the cloud shadow area, and this feature is marked in the preliminary identification result. The preliminary identification result of each row not only contains the position information of the sunlight and the state of the support structure (such as the torque value of the fastening bolt of the C-shaped steel support, the oil pressure state of the electric hydraulic telescopic rod 2, which provides the basis for the equipment health degree for subsequent cooperative correction), but also attaches the cloud shadow influence characteristics (such as whether it is affected by the cloud shadow, the degree of influence, etc.), which provides the basis for subsequent cooperative correction;

[0211] Each row can independently analyze whether it is in the cloud shadow coverage area through independent sensing and cloud shadow feature fusion, through multi-dimensional verification of light intensity change curve, temperature field distribution, and relative position of cloud shadow edge, to avoid cloud shadow misjudgment caused by single indicator, for example, a row of light intensity sudden drop but temperature field not uniform cooling, can exclude cloud shadow coverage, reduce the probability of misjudgment; Distributed parallel processing through edge node synchronous operation ensures consistent data processing efficiency of multiple rows, avoids cooperative decision lag caused by partial row calculation delay, for example, when multiple rows detect cloud shadow coverage at the same time, it can complete preliminary identification and feature marking synchronously, providing uniform initial data for subsequent cooperative correction; The preliminary identification result attaches the cloud shadow influence characteristics, so that the subsequent cooperative correction can accurately locate the degree and type of cloud shadow influence on each row, avoiding correction deviation caused by information loss, for example, it is clear that a row is in the center of the cloud shadow area and another row is in the edge area, the correction strategy can be differentiated.

[0212] The graph neural network coupling estimation module takes each row of photovoltaic racks as a network node, and the node feature vector includes the rack structure parameters of the row (C-shaped steel rack height, horizontal component group span, support angle of triangular rear rack 6), position coordinates (distance relative to the starting point of the array), preliminary identification results (solar azimuth and elevation angles), cloud shadow influence features (cloud shadow coverage probability, predicted duration), and equipment status (motor health, current angle); the edges between nodes represent the shading and reflection relationship between rows, and the edge weight quantifies the strength of this relationship through a coupling influence coefficient matrix. The initial value of the coefficient matrix is calculated based on the array geometric parameters (row spacing, C-shaped steel rack height difference), reflecting the inherent shading and reflection relationship in the absence of cloud shadows.

[0213] Taking each row of photovoltaic racks as a node (numbered i = 1, 2,..., N, N being the total number of rows), the initial value of the edge weight between nodes (i.e., the inherent shading / reflection coefficient) is calculated based on the array geometric parameters, as follows:

[0214] ;

[0215] Wherein: is the initial coupling coefficient of row i to row j (0-1, the larger the value, the stronger the influence); is the horizontal distance between row i and row j (unit: m); is the reference distance (default 5m, value according to array design); is the solar elevation angle (unit: °, calculated from latitude, longitude and time); is the vertical C-shaped steel rack height difference between row i and row j; is the reference height (default 2m); , is the weight coefficient;

[0216] The initial edge weight calculation based on array geometric parameters can accurately quantify the inherent shading and reflection relationship between rows in the absence of cloud shadows, such as calculating the shading intensity of the front row on the rear row according to the row spacing and C-shaped steel rack height difference, or the reflection influence between adjacent rows, providing a basis for subsequent dynamic correction and avoiding misjudgment of row-to-row interference due to neglecting the inherent coupling relationship, such as the significant shading influence of the front row on the rear row in high-density array, which can be quantified in advance by the initial matrix to lay the foundation for irradiance field correction.

[0217] Under the influence of cloud shadows, the graph neural network coupling estimation module updates the edge weight through a real-time edge weight correction algorithm, as follows:

[0218] When a cloud shadow covers a row of photovoltaic racks, the reflected light intensity of adjacent rows will change (e.g., cloud reflection may enhance indirect light), and at this time the corresponding edge weight will increase according to the cloud shadow transmittance level;

[0219] When the cloud shadow covers row i, the edge weight Real-time updates:

[0220] ;

[0221] in: The corrected coupling coefficient between row i and row j; The light transmittance level of the cloud shadow at location i (levels 1-5, level 1 is completely opaque, level 5 is highly transparent). The difference between the temperature of the outlet and the ambient temperature; Transmittance correction factor (0.1-0.2, dimensionless, weight increases by 10%-20% for each level increase in transmittance); This is the temperature correction factor (temperature correction factor is -0.02, unit is...). (For every 1°C drop in temperature, the weighting is reduced by 2%).

[0222] If cloud shadows cause a decrease in the temperature of a row of photovoltaic panels, the heat radiation exchange between that row and the adjacent row will decrease, and the relevant edge weights will be reduced proportionally to the temperature drop, i.e., for every 1°C drop in temperature, the weight will be reduced by 2%.

[0223] Through this dynamic update mechanism, graph neural networks can depict the changes in inter-row coupling relationships in real time, providing accurate quantitative basis for subsequent collaborative correction.

[0224] The update of edge weights related to cloud shadow transmittance and temperature can reflect the impact of cloud shadow coverage on inter-row coupling relationships in real time. For example, if a row is covered by cloud shadows, the reflected light will be enhanced, and the corresponding edge weight will be increased. The system can use this to correct the interference of reflected light in the irradiance field of the rear row, avoiding misjudging the enhanced reflected light as direct light. The reduction of edge weights due to temperature decrease can accurately reflect the actual situation of reduced heat radiation exchange between rows, avoiding misjudgment of inter-row illumination relationships due to the influence of heat radiation. For example, if cloud shadow coverage lowers the temperature of the front row, its heat radiation to the rear row will be weakened. The reduction of edge weights can eliminate the interference of this factor on the irradiance field analysis. Dynamic edge weights enable the graph neural network coupling estimation unit to adapt to the dynamic changes of cloud shadows in real time, accurately characterize the instantaneous state of inter-row coupling relationships, and provide accurate quantitative basis for multi-row collaborative correction, prediction, and decision-making, ensuring that the collaborative strategy can be dynamically adjusted with changes in cloud shadows.

[0225] Combining the inter-perturbation coefficient matrix output by the graph neural network (reflecting the inherent influence between rows) and the cloud shadow rolling dynamic model (reflecting dynamic influence), the specific steps for collaboratively correcting the preliminary identification results of each row through multi-row irradiance field cross-correction units are as follows:

[0226] If the front row support is blocked by cloud shadow, resulting in enhanced reflected light, according to the edge weight calculated by the graph neural network coupling estimation unit, the irradiance field model of the rear row photovoltaic support is corrected, and the false light component caused by the reflection of the front row is deducted (correction amount = reflected light intensity x edge weight coefficient);

[0227] Based on the reflected light intensity of the front row, the corrected coupling coefficient and the cloud shadow reflection enhancement coefficient (dynamically adjusted according to the transmittance level), the false light intensity that needs to be deducted from the rear row is calculated, that is, the false light correction amount of the rear row j caused by the cloud shadow reflection of the front row i, for example:

[0228] ;

[0229] Among them, is the false light intensity that needs to be deducted from the rear row j (unit: W / m²); is the reflected light intensity of the front row i (unit: W / m², calculated by the environmental reflection light formula in embodiment one); is the corrected coupling coefficient of row i to row j; is the cloud shadow reflection enhancement coefficient (1.2-1.5, dynamically adjusted according to the cloud shadow transmittance level , the higher, the larger);

[0230] Subtract the false light intensity from the original irradiance intensity of the rear row to obtain the corrected irradiance intensity. The corrected irradiance intensity of the rear row j is as follows:

[0231] ;

[0232] Among them: is the original irradiance intensity of the rear row j (unit: W / m²); is the corrected irradiance intensity (unit: W / m²); is the false light intensity that needs to be deducted from the rear row j (unit: W / m²);

[0233] The reflected light interference deduction based on the edge weight calculated by the graph neural network coupling estimation unit can accurately calculate the false light intensity of the rear row caused by the cloud shadow reflection of the front row, avoid the rear row photovoltaic support misjudging the reflected light as real direct light, for example, the cloud shadow coverage of the front row photovoltaic support causes the reflected light to enhance, and the multi-row irradiance field correction unit can calculate the correction amount according to the edge weight and the reflected light intensity, and eliminate the false component from the rear row irradiance data, so as to ensure that the rear row irradiance field model can reflect the real light condition, and prevent the tracking direction deviation caused by reflected light interference.

[0234] If a row is located at the edge of a cloud shadow and the light intensity fluctuates greatly, the multi-row irradiance field cross-correction unit will refer to the recognition results of the adjacent rows to perform light smoothing processing on this row. The weighted average algorithm is used, and the weights of the adjacent rows are positively correlated with the edge weights to eliminate the interference of local fluctuations.

[0235] Ranking the edge of the cloud shadow The light intensity is weighted and smoothed using its coupling coefficient with adjacent rows and the corrected irradiance. The smoothed light intensity is then calculated, as shown in the example below:

[0236] ;

[0237] in: The light intensity after smoothing (unit: W / m²). To arrange The set of adjacent rows (default 2 rows on the left and 2 rows on the right); To arrange The coupling coefficient (edge ​​weight) with the adjacent row m. To arrange The self-corrected irradiance intensity, The corrected irradiance is the intensity of radiation from the adjacent row m. Arrange the edges of the cloud shadows;

[0238] The multi-row irradiance field cross-correction unit optimizes the illumination data of the cloud shadow edge row by weighted smoothing of adjacent rows. By assigning weights to the data of adjacent rows through edge weights, the illumination data with large fluctuations tends to be stable. For example, if a row is located at the edge of the cloud shadow and the illumination intensity fluctuates frequently, smoothing by referring to the illumination data of adjacent stable rows can eliminate local fluctuation interference, avoid false peak misjudgment or frequent angle adjustment caused by illumination fluctuations, and reduce motor operation costs and bracket fatigue damage.

[0239] The pseudo-peak confidence assessment unit re-evaluates the pseudo-peak risk index for each row based on the corrected irradiance field, cloud field optical flow, and cloud shadow characteristics. Then, the consistency verification unit is initiated.

[0240] By comparing the identification results of adjacent rows, the consistency coefficient (0-1) of the angle change trend is calculated. If the coefficient exceeds 0.8, the trend is determined to be consistent.

[0241] Based on the angle adjustment amounts of row i and row j, calculate the angle trend consistency coefficient between them; the angle trend consistency coefficient between adjacent rows i and j is shown in the following example:

[0242] ;

[0243] in: The consistency coefficient (0-1, the closer the value is to 1, the more consistent the trend). Angle adjustment amount for row i; Angle adjustment amount for row j; Minimum value (0.01, to avoid denominator of 0);

[0244] Based on the consistency coefficient between all pairs of rows, the overall consistency coefficient is calculated, for example:

[0245] ;

[0246] Wherein: Overall consistency coefficient (0-1, ≥0.9); N is the total number of rows; Consistency coefficient;

[0247] If the results of a row of photovoltaic support and the surrounding 3 rows of photovoltaic support above exist significant difference (angle deviation exceeds 5°), the secondary verification is started, that is, the historical data of the row of photovoltaic support, the sky image details and the temperature distribution characteristics are called for review, and if necessary, the adjacent row will also request to share more detailed cloud shadow monitoring data (such as cloud layer contour coordinates, transmittance curve) through the inter-row data sharing agreement;

[0248] The adjacent row consistency coefficient can quantify the similarity of the angle adjustment trend of each row, avoid the angle deviation caused by single row sensor error or local interference, for example, if a row appears angle adjustment anomaly due to sensor failure, and the adjacent row consistency coefficient is lower than 0.8, the system can identify the anomaly in time; The global consistency coefficient comprehensively evaluates the overall coordination effect through multi-row data, ensures that the angle adjustment trend of all rows is highly consistent (the overall coefficient is ≥0.9), avoids the mutual shielding or reflected light interference between rows caused by partial row angle deviation, for example, the inconsistent angle trend of multiple rows of photovoltaic supports may cause the front row of photovoltaic supports to produce new shielding to the rear row of photovoltaic supports after adjustment, and the consistency verification can avoid this problem in advance.

[0249] Secondary verification deviation threshold:

[0250] If the angle deviation between a row w and the surrounding s' row exceeds the threshold, the secondary verification is started:

[0251] ;

[0252] Wherein: Angle deviation between row w and surrounding row s'; Angle deviation between row w and surrounding row s'; Angle deviation between row w and surrounding row s'; Current angle of row w; Current angle of row w; Angle of surrounding row s'.

[0253] Through iterative correction (up to 3 rounds), until the identification results of all rows reach overall consistency, that is, the overall consistency coefficient exceeds 0.9, the true direction of sunlight corresponding to each row is finally determined.

[0254] The collaborative prediction unit performs the following steps to make short-term predictions on the irradiance status of each row of photovoltaic supports by summarizing the real sunlight identification results of each row, the mutual interference coefficient matrix output by the graph neural network coupled estimation module, and the cloud shadow rolling dynamic model data:

[0255] The model input layer receives parameters such as the solar angle and cloud shadow effects of each row of photovoltaic brackets;

[0256] The consistency verification unit realizes inter-row information exchange through the message passing mechanism and calculates the quantification value of each row affected by the occlusion and reflection of other rows;

[0257] The consistency verification unit calculates the impact of occlusion and reflection on row i by other rows j through message passing. Based on the occlusion and reflection impact of row j on row i, combined with the weighting coefficient, the total impact value is calculated, as shown in the example below:

[0258] ;

[0259] in: The total influence value of row j on row i (0-1); To account for the shading effect (0-1, L0=3m is the reference length); The effect of reflection (0-1); , These are the weighting coefficients (0.6 and 0.4 respectively).

[0260] The output layer generates the irradiance variation curves and optimal angle trajectory parameters for each row;

[0261] The consistency verification unit realizes inter-row information exchange through message passing, accurately calculates the impact value of each row on the shading and reflection of other rows, and avoids trajectory prediction deviation caused by ignoring the dynamic influence between rows of photovoltaic brackets. For example, when the predicted angle of a row is adjusted, its shading impact on the photovoltaic brackets of the row behind can be quantified in advance, providing a basis for trajectory constraints. The quantified impact value enables the prediction model to clearly define the interaction strength between photovoltaic brackets of each row. For example, if the shading impact of the front row on the rear row is strong, the adjustment range of the front row photovoltaic brackets will be limited when predicting the trajectory to prevent aggravating the shading of the rear row, ensuring that the candidate trajectory of each row can not only meet its own power generation needs, but also take into account the impact on other rows.

[0262] The collaborative prediction unit will identify key affected rows on the cloud's movement path. (That is, the row that is expected to be covered by cloud shadows for the longest time and suffer the greatest loss of light), which is the key affected row when generating trajectory candidates. Reserve larger angle adjustment space to quickly capture the light gap between cloud shadows.

[0263] Calculate the cloud shadow impact score from the predicted coverage time, and calculate the extended angle adjustment range based on the basic angle adjustment range and the cloud shadow impact score; the key affected range on the cloud shadow movement path is The angle adjustment range is as follows:

[0264] ;

[0265] Wherein: is the extended angle adjustment range; is the basic angle adjustment range; is the cloud shadow impact score, which is between 0 and 1, ; is the predicted coverage time, is the total statistical duration;

[0266] At the same time, the cooperative prediction unit takes into account the cooperation between the rows through the cooperative constraint mechanism:

[0267] If the angle adjustment of a certain row of photovoltaic support leads to shielding of the rear row through geometric calculation, the adjustment amplitude will be limited in the trajectory, and the maximum will not exceed 30°;

[0268] If multiple rows of photovoltaic supports are in the same cloud shadow belt, i.e. the cloud shadow coverage range is continuous for more than 3 rows, the synchronization rhythm control algorithm will be used to ensure that their angle adjustment rhythm is consistent (the adjustment start time difference does not exceed 0.5s), avoiding new mutual shielding due to asynchronous action.

[0269] The final generated optimal inclination and azimuth angle trajectory candidates of each row of photovoltaic supports not only meet the power generation needs of a single row of photovoltaic supports, but also meet the cooperation requirements of the whole array;

[0270] For the rows of photovoltaic supports with long coverage time and large light loss on the cloud shadow movement path, the angle adjustment range is expanded so that they can quickly capture more light in the cloud shadow gap. For example, a certain row of photovoltaic supports is predicted to be covered by cloud shadow for 1 minute. After expanding the adjustment range, it can flexibly adjust the angle to track the direct light in the cloud shadow gap and reduce light loss. The cooperative constraint mechanism limits the shielding adjustment amplitude and synchronizes the adjustment rhythm of multiple rows of photovoltaic supports to avoid mutual interference between rows of photovoltaic supports. For example, when a certain row of photovoltaic supports may shield the rear row of photovoltaic supports, the maximum adjustment amplitude is limited. When multiple rows of photovoltaic supports are in the same cloud shadow belt, the adjustment rhythm is synchronized to prevent new shielding or reflected light interference due to asynchronous action, ensuring that the overall light utilization efficiency of the array is maximized.

[0271] The cloud shadow cooperation reward reinforcement learning model of the decision control layer optimizes the trajectory candidates of each row globally:

[0272] State space: current angles of each row, predicted angles of sunlight, shadow position and movement parameters, inter-row coupling coefficients;

[0273] Action space: angle adjustment of each row (discrete values in the range of -5° to +5°);

[0274] Reward function: composed of basic reward (single-row power generation income), synergy reward (difference between total income and single-row income), penalty (action cost, fatigue damage, safety constraint violation), calculate basic reward (based on single-row power generation income), synergy reward (based on the difference between total income and single-row income), minus action cost penalty, fatigue damage penalty and safety constraint violation penalty; weighted using normalization coefficients to get the reward function, for example:

[0275] ;

[0276] ;

[0277] ;

[0278] ;

[0279] ;

[0280] ;

[0281] wherein, , , , and are normalization coefficients; is the reward function; is the basic reward; is the power generation weight of row i (0-1, according to capacity allocation); is the real-time power generation of row i (unit: W); is the synergy reward; μ is the synergy coefficient (1.2-1.5, the larger the shadow coverage, the larger μ); is the actual total power of the array (unit: W); is the sum of independent power generation of each row (unit: W); is the action cost penalty; is the unit angle cost coefficient (0.1, set according to motor energy consumption); is the fatigue damage penalty (dimensionless); is the fatigue coefficient (0.5); is the fatigue damage value of row i (0-1, calculated based on Miner's rule); is a safety constraint violation value; is a safety factor (10); is the wind speed (unit: m / s) at row i; is the wind speed safety limit (unit: m / s, default 15 m / s); is the rated power of the photovoltaic array; is the maximum single angle change amount; is the preset wind speed safety limit.

[0282] It should be noted that when multiple rows of photovoltaic supports are adjusted in coordination according to cloud shadow changes, if the overall power generation gain (compared to the sum of independent adjustment of each row) exceeds a set threshold, the consistency verification unit will give an additional coordination reward (reward value = gain x coordination factor). This reward will be converted into decision weight, guiding the coordination execution synchronization unit to prefer stronger coordination in subsequent optimization. For example, when the cloud shadow moves from one side of the array to the other side, if the front row of supports adjusts the angle first for the rear row to make way (adjusting 5s in advance), enabling the rear row to capture light in advance, the overall benefit is improved, and the consistency verification unit will give positive incentive to this coordination behavior (increase decision weight). Through iterative training, the consistency verification unit can quickly converge to the global optimal strategy, determining the final angles of each row, ensuring efficient operation of individual rows, and maximizing overall power generation of the array.

[0283] The specific steps of the coordination execution and synchronization control performed by the coordination execution synchronization unit are as follows:

[0284] In the execution phase, each row of supports starts the adjustment action according to the allocated final angle; the coordination execution synchronization unit realizes interaction between each row of control units through a 5G ultra-low latency communication protocol, ensuring that the electric-hydraulic telescopic rod 2 and the rotating shaft 5 move synchronously, and the edge computing node unifies and coordinates the adjustment timing through a global clock synchronization mechanism, ensuring the synchronization of all rows:

[0285] Based on the horizontal distance between row i and row i+1 and the cloud layer movement rate, the adjustment time interval is calculated to ensure that the rear row has started adjusting when the cloud shadow reaches the front row. The adjustment time interval calculation is as follows:

[0286] The cloud layer movement rate is When the cloud shadow reaches the front row, the adjustment time interval of adjacent rows i and i+1 is:

[0287] ;

[0288] Wherein: is the time interval (unit: s, minimum 0.2s); is the horizontal distance between row i and row i+1 (unit: m); 0.5 is the safety factor (to ensure that the cloud shadow reaches the front row when the rear row has started adjusting); Cloud layer movement rate (unit: m / s);

[0289] The time stamp calibration of each row controller is realized by Network Time Protocol (NTP) (error < 1 ms);

[0290] The edge node sends timestamped adjustment instructions to each row, and each row controller starts the action at the specified time point at the same time to avoid relative motion interference due to time difference.

[0291] At the same time, the cooperative execution synchronization unit dynamically plans the adjustment sequence and time interval of each row according to the cloud shadow movement direction and speed predicted by the cloud shadow rolling dynamic model:

[0292] If the cloud shadow moves quickly from the east side of the array to the west, the cooperative execution synchronization unit will instruct the easternmost row to start adjustment first, and when it completes about half of the adjustment amount (judged by encoder feedback), the second row on the east side will start to act, forming a wave-like adjustment rhythm combined with the installation interval of the C-shaped steel support.

[0293] The time interval is dynamically calculated according to the cloud layer movement rate (the faster the movement speed, the smaller the interval, and the minimum interval is 0.2s), ensuring that each row can complete the angle calibration before the cloud shadow arrives.

[0294] During the adjustment process, the closed-loop system of the electric-hydraulic telescopic rod 2 of each row strictly follows the action amplitude limit, acceleration and speed profile requirements, and the error is corrected in real time by the back difference compensation unit. At the same time, the vibration spectrum monitoring model analyzes the vibration sensor data, and the vibration spectrum monitoring model needs to match the natural frequency of each row of C-shaped steel support. If resonance is detected, the acceleration will be adjusted immediately to avoid damage to the C-shaped steel welding node or the triangular rear support 6 welding node, ensuring smooth and reliable adjustment process and avoiding resonance or overload caused by cooperative action; the back difference compensation unit combines the installation reference of the photovoltaic support in the array to ensure the consistency of the angle correction of each row;

[0295] Based on the wave-like adjustment rhythm of the cloud shadow movement direction, the row in the front of the cloud shadow is adjusted first, and the subsequent rows follow in turn, ensuring that each row can complete the angle calibration before the cloud shadow arrives. For example, if the cloud shadow moves from east to west, the eastern row adjusts first and the western row adjusts later, avoiding the situation that the rear row has not been adjusted when the cloud shadow covers the front row, missing the opportunity to capture light. The dynamic time interval is optimized according to the cloud layer movement rate, and the faster the speed, the smaller the interval, ensuring that the adjustment rhythm matches the cloud shadow movement. For example, when the cloud shadow moves quickly, the adjustment interval between rows is shortened to prevent the rear row from being blocked by the cloud shadow due to too long interval, and when the cloud shadow moves slowly, the interval is lengthened to reduce the frequency and loss of mechanical action.

[0296] Through operation, the multi-row array photovoltaic support relies on the unified C-shaped steel support, the electrically-controlled hydraulic telescopic rod 2 and the triangular rear support 6 structure, realizes accurate cooperative adjustment, can accurately identify the real sunlight under the complex cloud shadow environment, dynamically optimizes the tracking angle, effectively avoids the tracking deviation guided by the false peak value, and significantly improves the power generation efficiency and equipment safety of the photovoltaic power station.

[0297] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the photovoltaic support adaptive tracking control method and system and the inventive concept thereof within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An adaptive tracking control method for photovoltaic brackets, characterized in that, include: Multiple sensing units are used to collect sky images, light intensity, photovoltaic panel temperature, and support posture data; The sky image, light intensity, photovoltaic panel temperature and support posture data are preprocessed to generate a sky visible mask; Based on the preprocessed sky image and support posture data, an irradiance field model of the location of the photovoltaic support is constructed. Determine whether the current wind speed is within a specific range. If the current wind speed is within a specific range: obtain cloud contour features and cloud shape change data under the action of wind, output the dynamic position coordinate sequence of cloud shadow projection on the ground, and construct a dynamic model of cloud shadow rolling and a cloud shadow transmittance grading model. To determine if the currently identified sunlight is real sunlight, the following steps are taken: Based on the real sunlight location, irradiance field model, cloud shadow rolling dynamic model, and cloud shadow transmittance grading model, multiple angle adjustment schemes are generated, and the final angle adjustment scheme is calculated. Calculate the light intensity gradient of each region in the irradiation field. If the light intensity gradient value of a certain region exceeds the set threshold and does not match the trend of light change caused by cloud movement, it is marked as a suspicious pseudo-peak. Based on temperature sensor data, if the regional temperature does not rise synchronously with the light intensity or if local high-temperature points appear, the false peak risk index is increased; if the suspected false peak area coincides with the cloud shadow edge position predicted by the cloud shadow rolling dynamic model, the false peak risk index is decreased. When the false peak risk index is lower than the set threshold, it is determined that the current identified light is real sunlight. That is, the area with the highest light intensity in the irradiance field is consistent with the actual position of the sun, and the light intensity changes steadily, which is consistent with the cloud movement trend. If the false peak risk index is higher than the set threshold, the system will activate a dual verification mechanism.

2. The photovoltaic bracket adaptive tracking control method according to claim 1, characterized in that, The method for constructing the irradiation field model includes: The solar elevation angle and azimuth angle are calculated based on the geometric position of the sun, and the intensity of direct sunlight is calculated by combining the occlusion probability of the visible sky mask. Calculate the intensity of scattered light in the hemispherical space based on cloud cover and aerosol concentration; The reflectivity of the reflective surface material database is used, combined with the bracket attitude data, to calculate the intensity of ambient reflected light. The intensity of direct light, scattered light, and ambient reflected light are weighted and fused to generate an irradiation field model.

3. The photovoltaic bracket adaptive tracking control method according to claim 1, characterized in that, The method for constructing the cloud shadow rolling dynamic model includes: Step 0A: Perform edge detection on the sky image and extract the set of pixel coordinates of the cloud contour; calculate the normal vector of each pixel coordinate of the cloud contour. Step 0B: Calculate the initial cloud motion rate based on the optical flow algorithm; set the normal component of the velocity at the cloud contour to 0; solve the vortex motion control equation to obtain the velocity of the cloud contour at a specific moment; thus obtain the velocity field at a specific moment; integrate the velocity field of the cloud contour at a specific moment to obtain the position coordinates of the cloud contour at that specific moment. Step 0C: Substitute the cloud outline position coordinates at a specific moment into the perspective projection formula to obtain the ground cloud shadow outline coordinates; fill the inside of the cloud shadow outline to generate a binary cloud shadow mask covering the cloud shadow area. Step 0D: Compare the positional relationship between the cloud shadow mask and the ground photovoltaic panel array, and calculate the photovoltaic panel number covered by the cloud shadow; combine the irradiance sensor data, and if the deviation between the actual light intensity reduction area and the cloud shadow mask prediction area exceeds the rated percentage, then correct the vortex viscosity coefficient of the vortex motion control equation. Repeat steps 0B to 0D to generate a dynamic model of cloud shadow rolling.

4. The photovoltaic bracket adaptive tracking control method according to claim 1, characterized in that, The method for constructing the cloud shadow transmittance grading model includes: Step 1A: Classify cloud types in the sky image based on the semantic segmentation model and output cloud type coefficients; extract the gray values ​​of the target cloud region; Step 1B: Select the cloudless area from the pixel-level cloud type coefficient in the sky image and calculate the average gray value of the cloudless area; obtain the gray value of the target cloud area and calculate the cloud thickness of the target cloud area; calculate the initial transmittance level based on the cloud type coefficient. Step 1C: Calculate the thickness change rate based on the cloud thickness of the target cloud area; correct the time constant by combining it with the current wind speed; substitute it into the transmittance update formula and adjust the transmittance level in real time, with the update frequency consistent with the image acquisition frequency; Step 1D: Compare the measured values ​​of the irradiance sensor with the theoretical irradiance attenuation corresponding to the transmittance level to generate a cloud shadow transmittance grading model.

5. The photovoltaic bracket adaptive tracking control method according to claim 1, characterized in that, The specific steps of the dual verification mechanism are as follows: By retrieving historical irradiance data for the same period and comparing it with the current irradiance curve, the confidence level of the actual sunlight is increased if the trends of the two are consistent. By combining the geometric position information of the sun, we can verify whether the area of ​​strongest sunlight matches the theoretical position of the sun. If the area of ​​strongest sunlight matches the theoretical position of the sun, we can determine the actual position of sunlight.

6. The photovoltaic bracket adaptive tracking control method according to claim 1, characterized in that, The system obtains the location data of photovoltaic (PV) brackets within the same area and detects whether there are multiple rows of PV brackets within the same area. If so, there is a risk of inter-row interference between the PV brackets in this area. Each row of PV brackets in the array is equipped with an independent sensing unit and computing module to determine whether the PV bracket is in the cloud shadow coverage area. If it is, the system records the cloud shadow coverage area, entry time, and expected stay time. A topological graph of the array is constructed using a graph neural network array coupling model. Each row of supports is used as a node, and the occlusion and reflection relationships between rows are used as edges. The initial edge weights are calculated based on the array's geometric parameters. The edge weights are corrected based on the cloud shadow rolling model; the preliminary identification results of each row are collaboratively corrected by a multi-row irradiance field cross correction model, combined with the mutual interference coefficient matrix output by the graph neural network and the cloud shadow rolling dynamic model, the false illumination components caused by inter-row reflection are deducted and the illumination fluctuations of the cloud shadow edge rows are smoothed; the true sunlight direction of each row is determined. Multiple angle adjustment schemes are generated, and the final angle adjustment scheme is calculated.

7. The photovoltaic bracket adaptive tracking control method according to claim 6, characterized in that, The specific steps for correcting edge weights based on the cloud shadow rolling model are as follows: When a cloud shadow covers a row, the intensity of the reflected light from that row to the adjacent rows will change, and the corresponding edge weight will increase accordingly based on the cloud shadow's transmittance level. If cloud shadows cause a decrease in the temperature of a row of photovoltaic panels, the heat radiation exchange between that row and the adjacent rows will decrease, and the relevant edge weights will be adjusted down proportionally to the temperature decrease.

8. The photovoltaic bracket adaptive tracking control method according to claim 6, characterized in that, The specific steps for collaboratively correcting the preliminary identification results of each row using a multi-row irradiance field cross-correction model, combined with the mutual perturbation coefficient matrix output by a graph neural network and a cloud shadow rolling dynamic model, to deduct false illumination components caused by inter-row reflections and smooth illumination fluctuations at the edge rows of cloud shadows are as follows: If the reflected light is enhanced due to cloud shadows on the front row of photovoltaic supports, the irradiance field model of the rear row of photovoltaic supports is corrected based on the edge weights calculated by the graph neural network coupled estimation unit, and the false light components caused by the reflection from the front row are deducted. If a row is located at the edge of a cloud shadow, the row is smoothed by combining the recognition results of adjacent rows to eliminate the interference of local fluctuations.

9. A photovoltaic bracket adaptive tracking control system, employing the photovoltaic bracket adaptive tracking control method according to any one of claims 1-8, characterized in that, include: Multimodal sensing layer: includes fisheye panoramic imaging unit, irradiance sensor, temperature sensor, attitude sensor and wind speed sensor; Data preprocessing module: used to preprocess sky images, light intensity, temperature, attitude and wind speed acquired by the multimodal sensing layer; Irradiance field modeling module: used to receive preprocessed sky images, light intensity and attitude data, and build irradiance field models; Cloud shadow modeling module: used to initiate cloud shadow monitoring and analysis when the wind speed is within a specific range; Real sunlight determination module: includes a pseudo-peak confidence assessment unit and a dual verification unit; Execution control layer: includes a collaborative prediction unit, a closed-loop execution unit, a dynamic parameter optimization unit, and a hysteresis compensation unit; The inter-row coordination module includes a graph neural network coupling estimation unit, a multi-row irradiation field cross correction unit, a consistency verification unit, and a coordination execution synchronization unit.

Images