A tower type photo-thermal power station mirror field cloud shadow identification and prediction method based on visual and timing DNI data fusion

Abstract

The application discloses a tower type photo-thermal power station mirror field cloud shadow identification and prediction method based on visual and time sequence DNI data fusion, and comprises the following steps: a plurality of overlooking cameras arranged on the edge of the top platform of the heat absorbing tower are utilized to continuously collect mirror field color images according to a predetermined frame rate, and a tensor H(x, y, t) is formed; a mirror surface space distribution mask function is adopted to automatically remove heliostat shadow and reflection area, and a purified gray matrix G(x, y, t) containing only cloud shadow information is output; short-time prediction is carried out on the cloud shadow moving track and distribution situation in the next time period; for any mirror field gray area judged as being blocked by cloud shadow, the purified gray value is input into the currently calibrated gray-DNI mapping function, short-time DNI estimation results are calculated, and high spatial resolution full mirror field DNI prediction maps are generated in combination with the area coordinates. The application can solve the problem that it is difficult to accurately obtain the DNI change of the local mirror field blocked by cloud shadow in the prior art.

Description

Technical Field

[0001] This invention relates to the field of mirror-tower coupling and coordination systems for tower-type concentrated solar power (CSP) plants, specifically to a method for identifying and predicting cloud shadows in the mirror field of tower-type CSP plants based on the fusion of visual and temporal DNI data. This invention utilizes data fusion from tower-mounted cameras and weather stations to identify cloud shadows in the mirror field and predict their impact on direct irradiance through algorithms, providing decision support for the operation control and scheduling of CSP plants. Background Technology

[0002] Tower solar thermal power plants use heliostats to concentrate sunlight onto a receiver at the top of a central tower, which then drives a generator to produce electricity via a high-temperature heat transfer medium. They have advantages such as high heat collection efficiency, strong heat storage capacity, and clean and environmentally friendly operation.

[0003] When a tower-type solar thermal power plant is in operation, the heliostat array automatically tracks the sun's position and focuses the light onto the heat carrier inside the receiver at the top of the tower. The heat exchange device generates a high-temperature, high-pressure working fluid to drive the generator set to output electrical energy. However, during operation, cloud shadows formed by cloud movement can block the reflected light path of the heliostat, causing a momentary drop in the DNI (Diverterless Noise Level), resulting in fluctuations in the solar collector power, complex system scheduling, and potentially adverse effects on the safe operation of the equipment.

[0004] Existing tower solar thermal power plant systems, in response to cloud cover scenarios, suffer from frequent defocusing of the mirror field and salt efflorescence in the receivers due to issues with the timeliness and accuracy of cloud cover prediction and cloud change forecasting. This results in energy loss, increased power consumption, and reduced power generation. Furthermore, rapid cloud changes can cause excessive thermal stress, potentially leading to receiver deformation or damage, as well as the risk of molten salt freezing. All of these factors negatively impact the safe operation of the receivers.

[0005] Existing research has proposed several methods for cloud shadow processing. One approach is cloud image analysis based on All-Sky Imaging System (ASI), which captures panoramic cloud images using fisheye lenses and predicts cloud shadow positions using image segmentation and velocity estimation. However, this method is limited by distortion correction accuracy and projection model errors, making it difficult to predict local details of the cloud field. Another approach combines stereo vision reconstruction technology using multiple overhead cameras to acquire three-dimensional cloud information from multiple perspectives to estimate cloud height and motion. However, this method is costly, has limited field of view coverage, and is difficult to deploy on a large scale. A third approach uses a coupled method of numerical weather prediction (NWP) models and irradiance simulation to calculate and predict the field DNI (Distributed National Irradiance) through atmospheric radiative transfer. However, this method involves large computational loads, poor real-time performance, and reliance on high-precision meteorological input, making it difficult to meet fine-grained operational requirements. These methods still have shortcomings in capturing short-term changes in cloud shadows and improving the accuracy of local DNI. Summary of the Invention

[0006] The purpose of this invention is to provide a method for cloud shadow identification and prediction of mirror field in tower solar thermal power plants based on the fusion of visual and temporal DNI data. This method can make full use of synchronous on-site measurement information, has advanced algorithm modeling capabilities and real-time performance, and can be quickly integrated and deployed in existing tower solar thermal power plants. It can achieve accurate prediction of local mirror field DNI changes, improve power generation scheduling efficiency and equipment safety assurance level, and has wide applicability. It solves the problem in the prior art that it is difficult to accurately obtain the DNI changes when the local mirror field is blocked by cloud shadows.

[0007] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for identifying and predicting cloud shadows in the mirror field of a tower solar thermal power plant based on the fusion of visual and temporal DNI data, the method comprising the following steps:

[0009] Step A: Using multiple overhead cameras deployed at the edge of the top platform of the heat absorption tower, continuously acquire color images of the mirror field at a predetermined frame rate to form a tensor H(x, y, t) containing three-dimensional information of space (x, y) and time t.

[0010] Step B involves performing geometric distortion correction on the acquired tensor H(x, y, t) based on camera calibration parameters, followed by denoising using high-pass filtering and spatial median filtering. The resulting three-channel color image is then converted to grayscale space to obtain a preliminary grayscale matrix. A weighted mask function is then obtained by combining the structural position mask and the reflective dynamic enhancement factor based on the brightness statistics of the local area in the current frame. A time sliding window is introduced to generate a mirror spatial distribution mask function. This mirror spatial distribution mask function is used to automatically remove the heliostat shadow and reflective areas, outputting a purified grayscale matrix G(x, y, t) containing only cloud shadow information.

[0011] Step C: Generate the purification grayscale matrix sequence for the current time period. Using the purification grayscale matrix sequence for the current time period, make a short-term prediction of the cloud shadow movement trajectory and distribution pattern for the next time period. If the prediction result shows that the entire mirror field will remain cloud-free in the next time period, return to step A to continue real-time acquisition; otherwise, proceed to step D.

[0012] Step D: For any grayscale region of the mirror field that is determined to be obscured by cloud shadows, extract the purified grayscale value of the grayscale region of the mirror field from the purified grayscale matrix G(x, y, t), input it into the currently calibrated grayscale-DNI mapping function, calculate the short-time DNI estimation result, and generate a high spatial resolution full mirror field DNI prediction map by combining the regional coordinates. The full mirror field DNI prediction map is then sent to the scheduling decision system of the solar thermal power plant in real time through the communication link.

[0013] Furthermore, in step A, the multiple overhead cameras deployed at the edge of the top platform of the heat absorption tower are evenly arranged circumferentially along the edge of the top platform of the heat absorption tower. Each overhead camera corresponds to a sub-area of ​​the field of view, and the fields of view of adjacent overhead cameras overlap.

[0014] Furthermore, in step B, the process of obtaining the preliminary grayscale matrix includes the following steps:

[0015] Assume overall layout The camera overlooks the platform, recording the first... The camera overlooks the platform at all times. The original color image captured is:

[0016]

[0017] in Represents image coordinates, , , These represent the pixel values ​​for the red, green, and blue channels, respectively; full-field image. Composited from all overhead camera images by geometric registration:

[0018]

[0019] in Indicates the first Functions for coordinate transformation and stitching of images from a top-down camera;

[0020] For images Geometric distortion correction is performed using a matrix including intrinsic parameters. and distortion coefficient The distortion correction image is obtained by performing distortion correction processing on the camera calibration parameters, including those included.

[0021] ;

[0022] The distortion-corrected image is converted to grayscale using a standard weighted transform to generate a grayscale image:

[0023] .

[0024] Furthermore, in step B, the process of automatically removing the heliostat shadow and reflective areas using a mirror spatial distribution mask function, and outputting a purified grayscale matrix G(x, y, t) containing only cloud shadow information, includes the following steps:

[0025] Introduce a dynamic illumination weighting factor to generate a weighted mask function:

[0026]

[0027] in It is a preset structural position mask, where 1 represents a mirror area and 0 represents a non-mirror area; It is a reflective dynamic enhancement factor obtained based on the brightness statistics of local areas in the current frame, used to express whether a certain area has high reflectivity anomaly characteristics;

[0028] A time-sliding window is introduced to take the union of all brightness anomaly regions in a series of consecutive frames to generate a specular spatial distribution mask function:

[0029] ;

[0030] in It is the width of the sliding window, used to determine the number of data points contained in the window; It represents the values ​​of all times within the sliding window, ensuring that no values ​​are repeated or omitted when calculating values ​​within the interval;

[0031] The weighted mask function is morphologically dilated, and then Gaussian blurring is applied at the edges to create soft boundaries. This is used to transition the masking effect of the reflection area to a grayscale reduction, thus transforming the weighted mask function from a static spatial occlusion template into a spatiotemporally adaptive artifact masking mechanism, resulting in a mirror spatial distribution mask function. ;

[0032] Using a mirror spatial distribution mask function Automatically remove heliostat shadows and reflective areas, outputting a purified grayscale matrix G(x, y, t) containing only cloud shadow information:

[0033] .

[0034] Furthermore, in step C, the process of making short-term predictions of the cloud shadow's movement trajectory and distribution pattern in the next time period includes the following steps:

[0035] Based on the past grayscale image of a frame Temporal motion estimation is performed using Kalman filtering, and future motion is predicted. The direction and shape changes of cloud shadow movement within a second; to determine the next frame If the overall brightness of the pixel grayscale distribution in the mid-field increases or there is no obvious change in the cloud shadow outline, it can be inferred that the current field of view is in a cloud-free state.

[0036] Further, in step D, the calibration process of the grayscale-DNI mapping function includes the following steps:

[0037] Step D1: Construct a grayscale-DNI mapping function based on a regularized polynomial regression function.

[0038]

[0039] in This is the DNI data for the current coordinates; It is a grayscale-DNI mapping function; This represents the current coordinate where grayscale data is collected. These are the polynomial fitting parameters that minimize the expected value of the absolute value of the error; and are obtained by solving...

[0040]

[0041] Obtain the initial model parameters; where for DNI data acquired by the weather station at that moment; for The DNI prediction data is obtained by using the grayscale values ​​at the top of the weather station through a mapping function. The second norm of the parameter; To control the complexity of the model, it is desirable to make the parameters generalizable.

[0042] Step D2: Set up several small meteorological observation points at key locations in the mirror field to monitor and obtain local DNI and meteorological parameters;

[0043] Step D3 involves performing a region intersection operation between the binarized cloud shadow region and the coordinate regions of each observation point to determine whether the current cloud shadow covers any meteorological station observation point. If no overlap is detected, proceed directly to step D5 for short-term DNI estimation; otherwise, proceed to step D4 to trigger the online calibration process.

[0044] Step D4: When the cloud shadow overlaps with the area of ​​a weather station, automatically extract the gray mean of the weather station area and simultaneously obtain the DNI measured value of the weather station at the corresponding time. Merge the newly formed gray-DNI sample pair with historical data according to preset rules, and compensate and adjust the parameters of the gray-DNI mapping function through a polynomial regularization optimization algorithm.

[0045] Step D5: For any grayscale region of the mirror field that is determined to be obscured by cloud shadows, input the purified grayscale value of the grayscale region of the mirror field into the currently calibrated grayscale-DNI mapping function, calculate the short-time DNI estimation result, and generate a high spatial resolution full-mirror field DNI prediction map by combining the regional coordinates.

[0046] Step D3 further includes:

[0047] Define four meteorological station areas If a certain area exists satisfy: If the weather station is obscured by clouds at time t, then it is considered that the weather station is obscured by clouds at time t.

[0048] Step D4 further includes:

[0049] In the meteorological station area Within, the average gray level of the grayscale image at time t is statistically analyzed.

[0050] ;

[0051] Simultaneously record the direct irradiance collected by the corresponding weather station. ;

[0052] New sample pairs It is added as training data to the model update module. ;

[0053] Parameters are updated using a polynomial regularization optimization algorithm. .

[0054] Step D5 further includes:

[0055] For any mirror field region obscured by cloud shadows at time t Obtain its average gray value :

[0056] ;

[0057] Average gray value Substitute the updated grayscale-DNI mapping function In the middle, the corresponding predicted DNI value is obtained. and the estimated value The corresponding mirror field coordinates are uploaded to the solar thermal power plant's scheduling and decision-making system to form a short-term spatially distributed DNI prediction map.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] First, the method for cloud shadow identification and prediction of the irradiance field of a tower solar thermal power plant based on the fusion of visual and temporal DNI data of the present invention, by calibrating the gray-scale temporal sequence of the tower top camera with the DNI of the weather station, can avoid prediction deviations caused by inaccurate model input or distortion correction errors, compared with the traditional method that relies solely on numerical forecasts or all-sky fisheye imaging. This method enables short-term accurate estimation of the DNI of the cloud shadow-occupied area of ​​the irradiance field, thereby improving the spatial resolution and real-time performance of irradiance prediction.

[0060] Secondly, the cloud shadow recognition and prediction method for tower solar thermal power plants based on the fusion of visual and temporal DNI data of the present invention uses Kalman filtering to smooth the cloud shadow motion and perform multi-step prediction. For the problems of unstable prediction and sensitivity to velocity changes in the previous single-frame or simple difference method, it can improve the robustness of cloud shadow position and velocity estimation, reduce abnormal noise interference, and realize reliable prediction of cloud shadow trajectory and shape in the future short time window, so as to facilitate the scheduling system to adjust the tracking angle or activate the backup energy in advance.

[0061] Third, the cloud shadow identification and prediction method for the mirror field of tower solar thermal power plants based on the fusion of visual and temporal DNI data of the present invention, through the gray-scale-DNI regression mapping model, addresses the shortcomings of existing methods in accurately quantifying the correspondence between the optical thickness of cloud shadows and the reduction of irradiance. It can achieve fine modeling of the DNI loss of cloud shadows with different optical thicknesses, thereby obtaining high-precision and short-term DNI prediction results at each point in the mirror field, and improving the accuracy and reliability of irradiance prediction under dynamic cloud shadows.

[0062] Fourth, the cloud shadow recognition and prediction method for the mirror field of tower solar thermal power plants based on the fusion of visual and temporal DNI data of the present invention improves the adaptive ability and long-term stability of the model by means of online feedback and periodic model update mechanism, which addresses the problem of accuracy decay of traditional offline training models under different environments and seasons. It enables the algorithm to self-calibrate for new samples in the actual mirror field, avoids error accumulation caused by changes in atmospheric composition or aging of reflective surfaces, and is easy to deploy in multiple power plants and multiple climate zones. Attached Figure Description

[0063] Figure 1 A schematic diagram showing the arrangement of cameras at the top of the tower and the overlap of their fields of view; the attached diagrams are labeled as follows: ① Cloud shadow camera system, ② Ventilation vent, ③ Unmanned aerial vehicle (UAV) group, ④ Weather station;

[0064] Figure 2 Comparison of distortion correction effects before and after: The left image is a schematic diagram of the shooting process, and the right image is a schematic diagram of the distortion correction.

[0065] Figure 3 Here is a flowchart of the time-series matching algorithm;

[0066] Figure 4 This is a schematic diagram showing the overlap between the meteorological station area and the binary cloud shadow map, with the area covered by the cloud shadow indicated in yellow.

[0067] Figure 5 This is a local DNI prediction distribution cloud map.

[0068] Figure 6 This is a flowchart of the method for identifying and predicting cloud shadows in a tower solar thermal power plant based on the fusion of visual and temporal DNI data, according to the present invention. Detailed Implementation

[0069] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0070] This invention discloses a method for identifying and predicting cloud shadows in the mirror field of a tower solar thermal power plant based on the fusion of visual and temporal DNI data. The method includes the following steps:

[0071] Step A: Using multiple overhead cameras deployed at the edge of the top platform of the heat absorption tower, continuously acquire color images of the mirror field at a predetermined frame rate to form a tensor H(x, y, t) containing three-dimensional information of space (x, y) and time t.

[0072] Step B involves performing geometric distortion correction on the acquired tensor H(x, y, t) based on camera calibration parameters, followed by denoising using high-pass filtering and spatial median filtering. The resulting three-channel color image is then converted to grayscale space to obtain a preliminary grayscale matrix. A weighted mask function is then obtained by combining the structural position mask and the reflective dynamic enhancement factor based on the brightness statistics of the local area in the current frame. A time sliding window is introduced to generate a mirror spatial distribution mask function. This mirror spatial distribution mask function is used to automatically remove the heliostat shadow and reflective areas, outputting a purified grayscale matrix G(x, y, t) containing only cloud shadow information.

[0073] Step C: Generate the purification grayscale matrix sequence for the current time period. Using the purification grayscale matrix sequence for the current time period, make a short-term prediction of the cloud shadow movement trajectory and distribution pattern for the next time period. If the prediction result shows that the entire mirror field will remain cloud-free in the next time period, return to step A to continue real-time acquisition; otherwise, proceed to step D.

[0074] Step D: For any grayscale region of the mirror field that is determined to be obscured by cloud shadows, extract the purified grayscale value of the grayscale region of the mirror field from the purified grayscale matrix G(x, y, t), input it into the currently calibrated grayscale-DNI mapping function, calculate the short-time DNI estimation result, and generate a high spatial resolution full mirror field DNI prediction map by combining the regional coordinates. The full mirror field DNI prediction map is then sent to the scheduling decision system of the solar thermal power plant in real time through the communication link.

[0075] See Figure 6 The method for identifying and predicting cloud shadows in the mirror field of a tower-type solar thermal power plant according to the present invention includes the following steps:

[0076] Preparatory work: Before the system is officially put into operation, offline training will be conducted using at least one month of synchronously collected data from the tower solar thermal power plant site. Specifically, the collected data will be the grayscale values ​​of clouds obscuring the weather station, and the DNI data of the weather station at that moment will be recorded synchronously. Based on regularized multinomial regression, a grayscale-DNI mapping function will be constructed.

[0077] Step 1: The system utilizes multiple overhead cameras positioned along the edge of the absorber tower's top platform to continuously acquire color images of the mirror field at a predetermined frame rate, forming a tensor H(x, y, t) containing three-dimensional information in space (x, y) and time t. The cameras are evenly arranged circumferentially with slight overlap in their fields of view to ensure full coverage of the large heliostat array. This design is suitable for tower-type systems with highly dense and regularly arranged mirrors, and clearly distinguishes itself from other image source methods.

[0078] Step 2: For the acquired H(x, y, t), geometric distortion correction is first performed based on the camera calibration parameters. Then, noise is removed using methods such as high-pass filtering and spatial median filtering. Subsequently, the three-channel color image is converted to grayscale space to obtain a preliminary grayscale matrix. To address the interference of false cloud shadows caused by the heliostat's own projection and high reflection in the tower mirror field, a mirror field shadow recognition and shielding module is specifically added to the system: based on a preset mirror spatial distribution mask function, the heliostat shadows and reflective areas are automatically removed, and the final output is a purified grayscale matrix G(x, y, t) containing only cloud shadow information.

[0079] Step 3: Using the denoised and cleaned grayscale sequence, the system calls an improved time-series matching algorithm (such as sliding window similarity measurement or Kalman filter prediction) to make short-term predictions of the cloud shadow movement trajectory and distribution pattern in the next few seconds to tens of seconds. If the prediction result shows that the entire mirror field will remain cloud-free in the next time period, it directly returns to Step 1 to continue real-time acquisition; otherwise, it proceeds to the subsequent meteorological station area judgment.

[0080] Step 4: The system pre-deploys four small meteorological station observation points P1, P2, P3, and P4 at key locations in the mirror field to monitor local DNI and meteorological parameters. By performing a region intersection operation between the binarized cloud shadow area and the coordinate areas of each observation point, it is determined whether the current cloud shadow covers any meteorological station observation point. If no overlap is detected, the system can directly proceed to step 6 for short-term DNI estimation; otherwise, it needs to proceed to step 5 to trigger the online calibration process.

[0081] Step 5: When the cloud shadow overlaps with the area of ​​a weather station, the system automatically extracts the gray-scale mean of that area and simultaneously obtains the measured DNI value of the station at the corresponding time, pushing the newly formed gray-scale-DNI sample pair to the calibration module. The offline real-time update mechanism will merge these new samples with historical data according to preset rules, and use a multinomial regularization optimization algorithm to compensate and adjust the mapping function parameters to maintain the model's high adaptability to changes in the tower mirror field environment;

[0082] Step 6: For any grayscale region of the heliostat field identified as being obscured by cloud shadows, the system inputs the purified grayscale value of that region into the currently calibrated grayscale-DNI mapping function to calculate the short-time DNI estimation result, and combines it with the region coordinates to generate a high spatial resolution full-field DNI prediction map. This prediction map and corresponding data are transmitted in real time to the scheduling and decision-making system of the solar thermal power plant via a communication link to drive control actions such as heliostat attitude optimization and backup energy scheduling.

[0083] Step 7: After completing one round of cloud shadow recognition, calibration, and DNI estimation, the system clears temporary cache data, resets its state, and automatically returns to Step 1 to continue the next round of real-time monitoring and prediction. Through the above closed-loop process, combined with offline training and real-time updates, dedicated shadow processing, and tower scene constraints, this solution ensures unique applicability and accuracy in tower-type solar thermal power plants with densely packed high-reflectivity mirrors.

[0084] In the preliminary steps, a gray-to-DNI mapping function is constructed based on regularized polynomial regression.

[0085]

[0086] By solving

[0087]

[0088] Obtain initial model parameters This will serve as the basis for the online calibration in step 5.

[0089] In step 1, multiple cameras are deployed on the top of the absorber tower of the tower-type solar thermal power plant. These cameras are evenly arranged circumferentially at the bottom of the tower's platform, with each camera corresponding to a specific sub-region of the field of view. The fields of view of different cameras overlap to ensure redundancy and image stitching stability. The intrinsic and extrinsic parameters of each camera are calibrated using appropriate geometric calibration. This configuration is as shown in the attached diagram. Figure 1 As shown. Let the total layout be... Taiwan camera, recording the first Taiwanese cameras at all times The original color image captured is:

[0090]

[0091] in Represents image coordinates, , , These represent the pixel values ​​for the red, green, and blue channels, respectively. Full-field image. Synthesized from all camera images by geometric registration:

[0092]

[0093] in Indicates the first Coordinate transformation and stitching operations of images from multiple cameras.

[0094] After the camera continuously acquires a sequence of color images of the entire mirror field, the image preprocessing stage (step 2) is initiated. In step 2, a full-field grayscale image is obtained. The process includes the following steps:

[0095] For images Geometric distortion correction is performed using camera calibration parameters (including intrinsic parameter matrices). and distortion coefficient Distortion correction processing is performed to obtain the distortion-corrected image:

[0096]

[0097] Figure 2 This is a comparison image showing the effect before and after distortion removal. The image is then converted to grayscale using a standard weighted transform:

[0098] .

[0099] In existing image recognition and target localization fields, spatial mask matrices are commonly used for background modeling and foreground extraction, primarily based on the constant projected contours of known structures in a static scene within the image coordinates. For example, in security monitoring and vehicle detection scenarios, masks are used to shield static backgrounds or known fixed obstacles. However, in these applications, mask generation typically assumes a fixed camera viewpoint, stable occlusion targets, and minimal environmental changes, thus the generated masks can remain effective for a long time without frequent adjustments.

[0100] However, when this type of masking function is directly applied to the image processing scenario of the tower-type solar thermal power plant mirror field in this invention, significant incompatibility occurs. This is because the heliostats in the mirror field are not completely static structures: although their positions are fixed, they rotate daily with the sun's movement, causing continuous changes in their mirror orientation. This results in dynamic changes in the reflected area and projection shape in the camera, making it difficult for traditional static masks to accurately cover the reflective areas. The actual image is significantly affected by strong reflections. Compared to general grayscale image analysis scenarios, the brightness of the mirror reflection areas in the images acquired by this system is much higher than that of normal ground or cloud shadow areas, easily causing false cloud shadow recognition errors. Ordinary masks only obscure structural boundaries and do not consider the dynamic changes in reflection intensity. The camera's viewing angle experiences slight drift due to platform vibrations, environmental wind, etc. Especially in scenarios where cameras are deployed at the top of a high tower, the camera may be affected by micro-vibrations or thermal expansion and contraction, causing pixel-level viewing angle shifts. Existing masking methods cannot correct these shifts in real time.

[0101] To address the above problems, this invention provides a mask function. Three structural improvements were made to make it suitable for cloud image cleanup in tower solar thermal power plant scenarios:

[0102] First, a dynamic illumination weighting factor is introduced. Traditional masks are binary functions that only mask based on whether the physical location belongs to a mirror. To address the impact of changes in the reflective range of the mirror surface with attitude, this invention rewrites the mask function as a weighted continuous-valued function:

[0103]

[0104] in: It is a preset structural position mask (1 represents the mirror area, 0 represents the non-mirror area); It is a dynamic enhancement factor for reflectivity based on the brightness statistics of a local area in the current frame, that is, whether a certain area has the characteristic of "high reflectivity anomaly". This mechanism can identify specular reflective areas that change with the solar altitude angle and effectively block new high-brightness artifacts caused by changes in the direction of sunlight.

[0105] Secondly, robust processing using multi-frame temporal windows is incorporated. Considering slight camera drift or image micro-vibration disturbances, this invention no longer uses single-frame images to construct the mask, but instead introduces a temporal sliding window:

[0106]

[0107] This involves taking the union of all brightness-abnormal areas in a series of consecutive frames, improving the mask's tolerance for temporary reflection drift, and ensuring that dynamic reflective areas are not missed due to changes between frames.

[0108] Finally, adaptive blurring of the mask boundaries is performed. To avoid "rigid mismatch" at the mask edges due to calibration errors or pose drift, the system introduces a buffer band extension based on the static mask boundaries:

[0109] First, the original mask is morphologically dilated by 3 pixels; then, a Gaussian blur operation is performed at the edges to form a "soft boundary," which transitions the masking effect of the reflective area to a grayscale reduction rather than direct removal.

[0110] This processing method ensures that the system can stably suppress mirror artifacts without damaging cloud shadow details under different lighting conditions, cloud cover, and attitude disturbances.

[0111] Through the above three improvements, the mask function The process evolved from a static spatial occlusion template to a spatiotemporally adaptive artifact masking mechanism, ultimately resulting in a cleaned grayscale image.

[0112]

[0113] It exhibits higher accuracy and stability when input into subsequent cloud extraction and DNI estimation modules.

[0114] In step 3, based on the past grayscale image of a frame Temporal motion estimation is performed using Kalman filtering, and future motion is predicted. The direction and shape changes of cloud shadow movement within a second. To determine the next frame... If the overall brightness of the pixel grayscale distribution in the middle field of view increases or there is no obvious change in the cloud shadow outline, it is inferred that the current field of view is in a cloud-free state, and the process jumps back to step 1; otherwise, proceed to step 4. Figure 3 The flowchart of the temporal matching algorithm shows that it calculates the similarity of grayscale images from several consecutive frames to determine whether cloud cover will occur in the field of view at a future time. If the result is clear and cloudless, the algorithm updates the correction status in step 1. Otherwise, it proceeds to step 4 for spatial overlap detection, and the grayscale binarization result is then processed. With the observation points of each meteorological station in the region Perform region intersection operations.

[0115] In step 4, four weather station areas are defined. If a certain area exists satisfy: If so, it is assumed that the weather station is obscured by clouds. (Attached) Figure 4 This is a schematic diagram showing the overlap between the meteorological station area and the binary cloud shadow map. Upon detection of an intersection, step 5 (model calibration) is automatically triggered; otherwise, the process proceeds to step 6 (DNI estimation).

[0116] In step 5, when the cloud shadow overlaps with a certain area of ​​the weather station, the system automatically extracts the average gray value of that area. Simultaneously record the direct irradiance collected by the weather station. Then utilize the accumulated The parameters of the regression model are periodically updated using a multinomial regularization optimization algorithm for the sample pairs. Specifically, in the weather station area Within, calculate the average gray level of the current grayscale image.

[0117]

[0118] Simultaneously, the direct irradiance collected by the weather station was recorded. .

[0119] New sample pairs It is added as training data to the model update module. ; Update parameters using a polynomial regularization optimization algorithm .

[0120] If calibration is not triggered, proceed to step 6, in which the average grayscale of the cloud-occluded area is used. Call the updated mapping function Calculate the short-term DNI prediction values ​​for this region and generate a DNI distribution map with spatial coordinate annotations (see Appendix). Figure 5 (Local DNI prediction distribution cloud map), this map and prediction data are sent to the control center to provide a basis for subsequent power generation scheduling and mirror orientation adjustment. Finally, the internal cache is reset according to step 7 and the loop returns to step 1 to start the next round of real-time monitoring. Specifically, for any mirror field area obscured by cloud shadows... Obtain its average grayscale value:

[0121]

[0122] Substitute it into the updated model function in step 5 In the middle, the corresponding predicted DNI value is obtained. and the estimated value The corresponding mirror field coordinates are uploaded to the solar thermal power plant's scheduling and decision-making system to form a short-term spatially distributed DNI prediction map, which serves as the input basis for energy scheduling and attitude control algorithms.

[0123] During implementation, all key algorithm modules and data flows must be marked in the overall system flowchart to ensure a clear correspondence between the interfaces and data structures of each submodule, thereby improving the efficiency and maintainability of system development, testing, and deployment.

[0124] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0125] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for identifying and predicting cloud shadows in the mirror field of a tower solar thermal power plant based on the fusion of visual and temporal DNI data, characterized in that, The method includes the following steps: Step A: Using multiple overhead cameras deployed at the edge of the top platform of the heat absorption tower, continuously acquire color images of the mirror field at a predetermined frame rate to form a tensor H(x, y, t) containing three-dimensional information of space (x, y) and time t. Step B involves performing geometric distortion correction on the acquired tensor H(x, y, t) based on camera calibration parameters, followed by denoising using high-pass filtering and spatial median filtering. The resulting three-channel color image is then converted to grayscale space to obtain a preliminary grayscale matrix. A weighted mask function is then obtained by combining the structural position mask and the reflective dynamic enhancement factor based on the brightness statistics of the local area in the current frame. A time sliding window is introduced to generate a mirror spatial distribution mask function. This mirror spatial distribution mask function is used to automatically remove the heliostat shadow and reflective areas, outputting a purified grayscale matrix G(x, y, t) containing only cloud shadow information. Step C: Generate the purification grayscale matrix sequence for the current time period. Using the purification grayscale matrix sequence for the current time period, make a short-term prediction of the cloud shadow movement trajectory and distribution pattern for the next time period. If the prediction result shows that the entire mirror field will remain cloud-free in the next time period, return to step A to continue real-time acquisition; otherwise, proceed to step D. Step D: For any grayscale region of the mirror field that is determined to be obscured by cloud shadows, extract the purified grayscale value of the grayscale region of the mirror field from the purified grayscale matrix G(x, y, t), input it into the currently calibrated grayscale-DNI mapping function, calculate the short-time DNI estimation result, and generate a high spatial resolution full mirror field DNI prediction map by combining the regional coordinates. The full mirror field DNI prediction map is then sent to the scheduling decision system of the solar thermal power plant in real time through the communication link.

2. The method for identifying and predicting cloud shadows in a tower solar thermal power plant based on the fusion of visual and temporal DNI data as described in claim 1, in step A, multiple overhead cameras deployed on the edge of the top platform of the heat absorption tower are evenly arranged circumferentially along the edge of the top platform of the heat absorption tower, each overhead camera corresponds to capturing a sub-region of the mirror field, and the fields of view of adjacent overhead cameras overlap.

3. The method for identifying and predicting cloud shadows in the mirror field of a tower solar thermal power plant based on the fusion of visual and temporal DNI data as described in claim 1, wherein step B, the process of obtaining the preliminary grayscale matrix, includes the following steps: Assume overall layout The camera overlooks the platform, recording the first... The camera overlooks the platform at all times. The original color image captured is: ； in Represents image coordinates, , , These represent the pixel values ​​for the red, green, and blue channels, respectively; full-field image. Composited from all overhead camera images by geometric registration: ； in Indicates the first Functions for coordinate transformation and stitching of images from a top-down camera; For images Geometric distortion correction is performed using a matrix including intrinsic parameters. and distortion coefficient The distortion correction image is obtained by performing distortion correction processing on the camera calibration parameters, including those included. ； The distortion-corrected image is converted to grayscale using a standard weighted transform to generate a grayscale image: 。 4. The method for cloud shadow recognition and prediction of a tower solar thermal power plant mirror field based on visual and temporal DNI data fusion as described in claim 1, in step B, the process of automatically removing the shadows and reflective areas of the heliostats using a mirror spatial distribution mask function and outputting a purified grayscale matrix G(x, y, t) containing only cloud shadow information includes the following steps: Introduce a dynamic illumination weighting factor to generate a weighted mask function: ； in It is a preset structural position mask, where 1 represents a mirror area and 0 represents a non-mirror area; It is a reflective dynamic enhancement factor obtained based on the brightness statistics of local areas in the current frame, used to express whether a certain area has high reflectivity anomaly characteristics; A time-sliding window is introduced to take the union of all brightness anomaly regions in a series of consecutive frames to generate a specular spatial distribution mask function: ； in It is the width of the sliding window, used to determine the number of data points contained in the window; It represents the value of all times within the sliding window; The weighted mask function is morphologically dilated, and then Gaussian blurring is applied at the edges to create soft boundaries. This is used to transition the masking effect of the reflection area to a grayscale reduction, thus transforming the weighted mask function from a static spatial occlusion template into a spatiotemporally adaptive artifact masking mechanism, resulting in a mirror spatial distribution mask function. ; Using a mirror spatial distribution mask function Automatically remove heliostat shadows and reflective areas, outputting a purified grayscale matrix G(x, y, t) containing only cloud shadow information: 。 5. The method for cloud shadow recognition and prediction of a tower solar thermal power plant based on visual and temporal DNI data fusion as described in claim 1, wherein step C, the process of short-term prediction of the cloud shadow movement trajectory and distribution pattern in the next time period includes the following steps: Based on the past grayscale image of a frame Temporal motion estimation is performed using Kalman filtering, and future motion is predicted. The direction and shape changes of cloud shadow movement within seconds; If we determine the next frame If the overall brightness of the pixel grayscale distribution in the mid-field increases or there is no obvious change in the cloud shadow outline, it can be inferred that the current field of view is in a cloud-free state.

6. The method for identifying and predicting cloud shadows in the mirror field of a tower solar thermal power plant based on the fusion of visual and temporal DNI data according to claim 1, wherein step D, the calibration process of the grayscale-DNI mapping function includes the following steps: Step D1: Construct a grayscale-DNI mapping function based on a regularized polynomial regression function. ； in This is the DNI data for the current coordinates; It is a grayscale-DNI mapping function; This represents the current coordinate where grayscale data is collected. These are the polynomial fitting parameters that minimize the expected value of the absolute value of the error; and are obtained by solving... ； Obtain the initial model parameters; where for DNI data acquired by the weather station at that moment; for The DNI prediction data is obtained by using the grayscale values ​​at the top of the weather station through a mapping function. The second norm of the parameter; To control the complexity of the model, it is desirable to make the parameters have generalizability; Step D2: Set up several small meteorological observation points at key locations in the mirror field to monitor and obtain local DNI and meteorological parameters; Step D3 involves performing a region intersection operation between the binarized cloud shadow region and the coordinate regions of each observation point to determine whether the current cloud shadow covers any meteorological station observation point. If no overlap is detected, proceed directly to step D5 for short-term DNI estimation; otherwise, proceed to step D4 to trigger the online calibration process. Step D4: When the cloud shadow overlaps with the area of ​​a weather station, automatically extract the gray mean of the weather station area and simultaneously obtain the DNI measured value of the weather station at the corresponding time. Merge the newly formed gray-DNI sample pair with historical data according to preset rules, and compensate and adjust the parameters of the gray-DNI mapping function through a polynomial regularization optimization algorithm. Step D5: For any grayscale region of the mirror field that is determined to be obscured by cloud shadows, input the purified grayscale value of the grayscale region of the mirror field into the currently calibrated grayscale-DNI mapping function, calculate the short-time DNI estimation result, and generate a high spatial resolution full-mirror field DNI prediction map by combining the regional coordinates.

7. The method for identifying and predicting cloud shadows in the mirror field of a tower solar thermal power plant based on the fusion of visual and temporal DNI data as described in claim 6, step D3 further includes: Define four meteorological station areas If a certain area exists satisfy: If the weather station is obscured by clouds at time t, then it is considered that the weather station is obscured by clouds.

8. The method for identifying and predicting cloud shadows in the mirror field of a tower solar thermal power plant based on the fusion of visual and temporal DNI data as described in claim 6, wherein step D4 further includes: In the meteorological station area Within, the average gray level of the grayscale image at time t is statistically analyzed. ； Simultaneously record the direct irradiance collected by the corresponding weather station. ; New sample pairs It is added as training data to the model update module. ; Parameters are updated using a polynomial regularization optimization algorithm. .

9. The method for identifying and predicting cloud shadows in the mirror field of a tower solar thermal power plant based on the fusion of visual and temporal DNI data as described in claim 6, step D5 further includes: For any mirror field region obscured by cloud shadows at time t Obtain its average gray value : ； Average gray value Substitute the updated grayscale-DNI mapping function In the middle, the corresponding predicted DNI value is obtained. and the estimated value The corresponding mirror field coordinates are uploaded to the solar thermal power plant's scheduling and decision-making system to form a short-term spatially distributed DNI prediction map.

Images