New energy optical power correction method and system based on weather abrupt change prediction

Abstract

The embodiment of the invention provides a new energy optical power correction method and system based on weather abrupt change prediction. The method comprises the following steps: identifying a cloud layer type contained in a target area, and obtaining a cloud layer type coverage rate; if the cumulonimbus coverage rate exceeds a first threshold value, predicting a cloud layer movement path in combination with the cloud layer movement vector field to obtain a first cloud layer coverage change sequence; if the layer cloud coverage rate exceeds a second threshold value, an irradiance attenuation coefficient is predicted based on the cloud layer thickness in the meteorological data and the solar altitude angle, and a second cloud layer coverage change sequence is obtained in combination with the ground aerosol optical thickness and the corrected irradiance attenuation coefficient; and monitoring real-time operation data of the power station, combining the first / second cloud layer coverage change sequence, predicting a matching relationship between a cloud layer moving direction and a power station power reduction area by adopting a double-flow Transform architecture, calculating a power fluctuation range corresponding to each power station power reduction area, dynamically adjusting power grid operation parameters of a target area, and determining the power grid operation parameters of the target area. The power grid stability is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data prediction, in particular to a new energy light power correction method and system based on weather mutation prediction. BACKGROUND

[0002] At present, the photovoltaic power generation system is a power generation device for directly converting solar radiation energy into electric energy by using a solar panel, and is usually composed of a solar assembly, a current combiner box, an inverter, electrical equipment and a monitoring system. The core principle is the photovoltaic effect of semiconductor materials. When sunlight shines on the solar panel, photons excite electrons to generate electromotive force, which is collected by the circuit to form an electric current.

[0003] In the related art, photovoltaic power generation power is directly affected by meteorological factors such as solar irradiance, cloud cover and precipitation, and small-scale weather phenomena (spatial range less than 10 kilometers, duration less than 1 hour) such as sudden cloud cover and local thunderstorms are affected by complex factors such as atmospheric turbulence and terrain. Traditional weather models are difficult to capture due to insufficient resolution (usually 10-20 kilometers). And relying on weather stations or satellite data, the data update frequency is low (such as satellite cloud map updated every 15 minutes), which cannot reflect the second / minute level weather change in time. Therefore, the current photovoltaic power generation system is difficult to realize timely correction of power generation power for small-scale weather phenomena, which is easy to cause new energy power grid fluctuation due to sudden change of power generation power, and has an impact on the stability of the power grid, affecting the efficiency of power grid dispatching and the utilization rate of renewable energy. Therefore, it is urgent to design a technical scheme to overcome at least one technical problem in the related art. SUMMARY

[0004] The main purpose of the embodiments of the present application is to provide a new energy light power correction method and system based on weather mutation prediction, which aims to solve the technical problems of new energy power grid fluctuation caused by small-scale weather phenomena, impact on the stability of the power grid, and influence on the efficiency of power grid dispatching and the utilization rate of renewable energy.

[0005] In a first aspect, the embodiments of the present application provide a new energy light power correction method based on weather mutation prediction, comprising:

[0006] For the target area where the power station is located, satellite cloud maps, meteorological data and ground data of weather stations of the target area are obtained;

[0007] A dense optical flow algorithm is used to calculate a cloud layer movement vector field matched with the target area based on the satellite cloud map; the cloud layer movement vector field is used to describe the vector distribution of the movement direction and speed of the cloud layer in the actual space, so as to reflect the overall movement trend and local turbulence change of the cloud layer;

[0008] identify a cloud type contained in the target area based on the cloud movement vector field, and obtain a coverage rate of the cloud type in the target area; the cloud type at least includes one of cumulonimbus, stratus, and cirrus;

[0009] If the cumulonimbus coverage rate exceeds a first threshold, a cloud movement path is predicted in combination with the cloud movement vector field to obtain a first cloud coverage change sequence of the target area in a first time period;

[0010] If the stratus coverage rate exceeds a second threshold, an irradiance attenuation coefficient is predicted based on a cloud thickness and a solar elevation angle in the meteorological data, and the irradiance attenuation coefficient is corrected in combination with a ground aerosol optical depth obtained based on the meteorological station ground data to obtain a second cloud coverage change sequence of the target area in a second time period;

[0011] Real-time operation data of power stations are monitored, a matching relationship between cloud movement directions and power reduction areas of the power stations is predicted in combination with the first cloud coverage change sequence and / or the second cloud coverage change sequence by using a double-flow Transformer architecture, and a power fluctuation range corresponding to each power reduction area of the power stations is calculated;

[0012] Based on the matching relationship and the power fluctuation range, power grid operation parameters of the target area are dynamically adjusted to ensure power grid stability.

[0013] In a second aspect, an embodiment of the present application provides a new energy light power correction system based on weather mutation prediction, comprising:

[0014] A collection module is configured to obtain satellite cloud images, meteorological data, and meteorological station ground data of a target area where power stations are located;

[0015] A calculation module is configured to calculate a cloud movement vector field matched with the target area based on the satellite cloud images by using a dense optical flow algorithm; the cloud movement vector field is used to describe a vector distribution of movement directions and speeds of the cloud in an actual space, so as to reflect overall movement trends and local turbulent changes of the cloud;

[0016] An identification module is configured to identify a cloud type contained in the target area based on the cloud movement vector field, and obtain a coverage rate of the cloud type in the target area; the cloud type at least includes one of cumulonimbus, stratus, and cirrus; if the cumulonimbus coverage rate exceeds a first threshold, a cloud movement path is predicted in combination with the cloud movement vector field to obtain a first cloud coverage change sequence of the target area in a first time period; if the stratus coverage rate exceeds a second threshold, an irradiance attenuation coefficient is predicted based on a cloud thickness and a solar elevation angle in the meteorological data, and the irradiance attenuation coefficient is corrected in combination with a ground aerosol optical depth obtained based on the meteorological station ground data to obtain a second cloud coverage change sequence of the target area in a second time period;

[0017] a prediction module configured to monitor real-time operation data of the power plant, combine the first cloud cover change sequence and / or the second cloud cover change sequence, predict a matching relationship between a cloud movement direction and a power reduction area of the power plant using a dual-flow Transformer architecture, and calculate a power fluctuation range corresponding to each power reduction area of the power plant;

[0018] an adjustment module configured to dynamically adjust a power grid operation parameter of the target area based on the matching relationship and the power fluctuation range, and ensure power grid stability.

[0019] In a third aspect, an embodiment of the present application further provides a terminal device, which comprises a processor, a memory for storing a computer program, and the processor is configured to execute the computer program and implement the new energy light power correction method based on weather mutation prediction in the first aspect or any embodiment of the present application.

[0020] The embodiments of the present application provide a new energy light power correction method based on weather mutation prediction, which realizes high-precision prediction of light power fluctuation by using a dense optical flow algorithm and a dual-flow Transformer architecture, combining precise analysis of cloud dynamic changes with multi-source data, and can more timely and accurately capture power changes caused by weather mutations compared with traditional methods, and can reduce 5-minute prediction errors to a lower level, thereby gaining sufficient response time for power grid dispatching. The prediction strategy for cloud layer types and the online data feedback mechanism enable the system to adapt to various complex weather scenarios. Whether it is the rapid movement of cumulonimbus clouds or the slow change of stratus clouds, the system can effectively predict and avoid prediction failure caused by weather mutations, thereby enhancing the stability of the system in different environments. The mechanism of dynamically adjusting power grid operation parameters can quickly respond to light power fluctuations, reduce power grid voltage and frequency fluctuations caused by power mutations, control voltage fluctuation amplitude within a small range, and ensure safe and stable operation of the power grid, thereby reducing the impact of new energy grid connection on the power grid. Through power prediction and effective power grid regulation, the phenomenon of light rejection caused by inaccurate prediction is reduced, the consumption capacity of new energy power generation is improved, the utilization rate of renewable energy is improved, and the efficient use of clean energy is promoted. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A flowchart of a new energy light power correction method based on weather mutation prediction provided by an embodiment of the present application is shown in the figure;

[0022] Figure 2 A module structure diagram of a new energy light power correction system based on weather mutation prediction provided by an embodiment of the present application is shown in the figure;

[0023] Figure 3A structural schematic block diagram of a terminal device provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0024] To solve the technical problems in the related art, an embodiment of the present application provides a new energy light power correction method and system based on weather mutation prediction. Specifically, the scheme improves the accuracy of light power prediction, makes the predicted value closer to the actual power generation, thereby improving the stability of the power grid, optimizing the efficiency of power grid dispatching and the utilization rate of renewable energy.

[0025] An embodiment of the present application provides a new energy light power correction method and system based on weather mutation prediction. The new energy light power correction method based on weather mutation prediction can be applied to a terminal device, such as a mobile phone, a virtual reality device, a tablet computer, a notebook computer, a desktop computer, a wearable device, and the like. The terminal device can be a server connected to a tower crane device, or a server cluster. The above connection mode can be realized by a hardware circuit or a communication module.

[0026] Some embodiments of the present application will be described in detail below with reference to the accompanying drawings. In the case of no conflict, the embodiments described below and the features in the embodiments can be combined with each other. Please refer to Figure 1 , Figure 1 A flowchart of a new energy light power correction method based on weather mutation prediction provided by an embodiment of the present application is shown.

[0027] As shown in Figure 1 , the new energy light power correction method based on weather mutation prediction includes the following steps S101 to S107.

[0028] Step S101, for a target area where a power station is located, satellite cloud images, meteorological data, and ground data of a meteorological station of the target area are obtained.

[0029] Step S102, a cloud layer movement vector field matched with the target area is calculated based on the satellite cloud images by using a dense optical flow algorithm.

[0030] Step S103, based on the cloud layer movement vector field, a cloud layer type contained in the target area is identified, and a coverage rate of the cloud layer type in the target area is obtained.

[0031] Step S104, if the coverage rate of cumulonimbus exceeds a first threshold value, a cloud layer movement path is predicted in combination with the cloud layer movement vector field, and a first cloud layer coverage change sequence of the target area in a first time period is obtained.

[0032] In step S105, if the coverage of the stratus cloud exceeds the second threshold, the irradiance attenuation coefficient is predicted based on the cloud thickness in the meteorological data and the solar elevation angle, and the irradiance attenuation coefficient is corrected in combination with the ground aerosol optical depth obtained based on the ground data of the meteorological station, to obtain a second cloud coverage change sequence of the target area in the second period.

[0033] In step S106, real-time operation data of the power station is monitored, the first cloud coverage change sequence and / or the second cloud coverage change sequence are combined, a matching relationship between the cloud movement direction and the power reduction area of the power station is predicted by using a double-flow Transformer architecture, and a power fluctuation range corresponding to each power reduction area of the power station is calculated.

[0034] In step S107, based on the matching relationship and the power fluctuation range, the operation parameters of the power grid in the target area are dynamically adjusted to ensure the stability of the power grid.

[0035] In the embodiment of the application, the cloud movement vector field is used to describe the vector distribution of the movement direction and speed of the cloud in the actual space, so as to reflect the overall movement trend and local turbulent change of the cloud.

[0036] Specifically, the vector direction in the cloud movement vector field directly presents the overall movement direction of the cloud, such as moving from west to east; the length or numerical value of the vector corresponds to the movement speed of the cloud, and the vector is longer in the area with faster speed, which helps to quickly judge the movement trend of the cloud, for example, to predict whether the cumulonimbus cloud will quickly cover the area where the power station is located. Since the calculation is fine to the pixel level, the cloud movement vector field can sensitively capture the subtle changes such as the deformation of the cloud edge and the local vortex. In a complex weather scene, even if the overall movement direction of the cloud is stable, local turbulence may cause abnormal fluctuations in irradiance, and the vector field can accurately reflect these changes, providing a basis for more accurate light power prediction. By analyzing the degree of disorder of the vectors in the local area in the vector field, it can be judged whether there is turbulence phenomenon in the area and its intensity. Through the cloud movement vector field, it becomes a key bridge connecting meteorological data and light power prediction, and provides high-precision dynamic data support for subsequent cloud type identification, coverage change prediction and power fluctuation analysis.

[0037] In the embodiment of the application, the cloud type at least includes one of the following: cumulonimbus, stratus, and cirrus.

[0038] Cumulonimbus is a kind of convective cloud with strong vertical development, often in the shape of anvil or horsehair, with dark bottom and bright top. Its formation is related to strong convective motion, often accompanied by thunderstorm, gale, short-time heavy precipitation and even hail and other severe weather phenomena. Due to the thick cloud layer, cumulonimbus has strong ability to block solar radiation, and once it covers the area where the power station is located, it will cause the solar irradiance to drop sharply in a short time, and then the photovoltaic power generation power will be greatly reduced. In the embodiments of the present application, when the coverage rate of cumulonimbus exceeds the first threshold value, the system will combine the cloud movement vector field to predict its moving path and generate the cloud coverage change sequence in the first period, so as to predict the rapid decline trend of light power in advance and gain time for power grid dispatching.

[0039] Stratus is a kind of relatively uniform cloud layer, which is gray or gray-white, usually accompanied by light rain, drizzle or light fog, with wide coverage but relatively thin cloud layer. Compared with cumulonimbus, stratus has a relatively mild and sustained weakening effect on solar radiation. Based on the cloud thickness and solar elevation angle in the meteorological data, the system can predict the irradiance attenuation coefficient caused by stratus, and further correct the coefficient by combining the ground aerosol optical thickness, so as to obtain the cloud coverage change sequence in the second period. In this way, the steady decline amplitude of photovoltaic power generation power under the influence of stratus can be accurately estimated, which is convenient for the power grid to reasonably arrange power supply and maintain the balance between power supply and demand.

[0040] Cirrus is a kind of cloud with high altitude, which is composed of fine ice crystals and appears in the form of white filaments, feathers or horse tails. Cirrus has a relatively weak blocking effect on solar radiation, and usually does not significantly reduce photovoltaic power generation power. However, in some cases, cirrus may be a precursor of weather changes, for example, when it gradually thickens and evolves into other types of cloud layers, it will affect the light power. In the present method, the identification and monitoring of cirrus are helpful to more comprehensively analyze the weather change trend and assist in judging whether there is a potential risk of cloud type transition, so as to make more detailed adjustment and optimization of light power prediction at an early stage.

[0041] Through accurate identification and targeted analysis of cumulonimbus, stratus and cirrus, the light power correction method in the embodiments of the present application can adopt differentiated prediction strategies according to the characteristics of different cloud types, significantly improve the accuracy and timeliness of light power prediction, and effectively cope with the influence of weather changes on photovoltaic power generation.

[0042] As an optional embodiment, in step S102, the cloud movement vector field of the target area is calculated based on the satellite cloud image by using a dense optical flow algorithm, including the following steps:

[0043] Step 21, align the multi-time satellite cloud image to the power station coordinate system through the geographic coordinate system, convert the RGB format to a single-channel grayscale image, and perform pixel value normalization processing on the converted single-channel grayscale image;

[0044] Step 22, using the Farneback algorithm, the processed multi-temporal satellite cloud image is analyzed for gray value change, and a multi-resolution image sequence of gray values is generated by layer-by-layer downsampling;

[0045] Step 23, starting from the top layer, calculate the granularity of each pixel light flow layer by layer, pass the calculation result of the granularity of the light flow of the last layer to the next layer as the initial value, until the original resolution is reached, output the multi-resolution light flow calculation result as the first light flow vector of each pixel; Wherein, the higher the level, the coarser the granularity of the light flow; The calculation result of the high-level granularity light flow is used to represent the large-scale movement trend of the cloud layer, and the calculation result of the bottom-level granularity light flow is used to represent the deformation of the cloud group edge;

[0046] Step 24, based on the multi-resolution image sequence, build a block integral graph, and perform inverse search optimization on the block integral graph through block matching to obtain sparse light flow vectors and corresponding sparse light flow blocks;

[0047] Step 25, starting from the top layer, calculate the coverage weight of all sparse light flow blocks in the neighborhood of each pixel layer by layer, and obtain the dense light flow value by weighted average of the sparse light flow vector through the coverage weight, pass the dense light flow value of the last layer to the next layer as the initial value, until the original resolution is reached, output the calculation result as the second light flow vector of each pixel;

[0048] Step 26, fuse the first light flow vector and the second light flow vector to obtain the cloud layer movement vector field matched by the target area.

[0049] Specifically, in step 21, the multi-temporal satellite cloud images (images of the same area taken at different times) are aligned to the power station coordinate system through the geographic coordinate system to eliminate geometric distortion caused by differences in satellite viewing angles. Then the RGB images are converted to single-channel grayscale images to simplify the calculation complexity, and the pixel values are mapped to the [0, 1] interval through normalization to enhance the robustness of the algorithm to light changes. For example, assuming that the coordinates of a power station are 30° N, 120° E, and the original coordinates of the satellite cloud image are in the WGS84 coordinate system. The cloud image is resampled to the power station coordinate system through projection transformation, and after grayscale and normalization, the cloud layer brightness distribution is easier to extract the motion features. In step 22, the Farneback algorithm is used to downsample the grayscale image layer by layer to construct an image pyramid. The high-level image has low resolution (e.g., 1 / 8 of the original image), which captures the large-scale motion trend of the cloud layer; the low-level image has high resolution (original resolution), which retains the edge details of the cloud cluster. The pyramid structure reduces the computational complexity through spatial dimension reduction while preserving multi-scale motion information. For example, the original cloud image has a resolution of 4096x2160, and the top layer of the pyramid is downsampled to 512x272. A 5-layer pyramid is generated through Gaussian blur and downsampling, with the resolution of each layer halved. In step 23, the optical flow vectors of the pixels are calculated layer by layer from the top layer of the pyramid. The high-level optical flow is based on the global motion trend of the low-resolution image (e.g., the cloud layer moving eastward), and the displacement is estimated through a polynomial expansion model (e.g., Farneback's quadratic polynomial fitting). The bottom-level optical flow is refined to the pixel level, capturing local deformation at the edge of the cloud cluster (e.g., vortex or fragmentation). The calculation result of each layer serves as the initial value for the next layer, achieving multi-scale optimization. For example, the top-level optical flow shows that the cloud layer moves southeast at a speed of 5 m / s, and the bottom-level optical flow detects a local rotational motion of 1 m / s at the edge of the cloud cluster. In step 24, based on the multi-resolution image sequence, a block integral image is constructed, and a block matching algorithm (such as normalized cross correlation NCC) is used to search for similar blocks between adjacent frames. The inverse search optimization strategy verifies the matching points from the predicted position, reducing false matches. Sparse optical flow blocks (e.g., selecting one block every 16x16 pixels) represent the significant areas of cloud cluster motion. For example, in the direction of cloud layer movement, the block matching algorithm finds the most similar area (similarity > 0.8) in the next frame to the current block, determining the sparse optical flow vector. In step 25, the dense optical flow is calculated layer by layer from the bottom to the top. The dense optical flow value of each pixel is obtained by weighted average of the coverage weights of the sparse optical flow blocks in its neighborhood, and the weights are determined by the block matching similarity. The high-level optical flow vector (coarse granularity) provides global motion constraints, and the bottom-level optical flow vector (fine granularity) supplements local details, finally fusing into a complete optical flow field. In step 26, the multi-resolution optical flow vector (global trend) of step 23 and the dense optical flow vector (local details) of step 25 are fused.A dynamic weight distribution strategy is adopted: the high-level optical flow weight decreases with the pyramid level (e.g. the top layer weight is 0.3 and the bottom layer is 0.1), and the dense optical flow weight increases with the spatial resolution, finally generating a high-precision cloud layer movement vector field. For example, the high-level optical flow of a certain area shows that the cloud layer moves eastward as a whole, and the dense optical flow shows local northward disturbance, and the fused vector field accurately reflects the spiral movement characteristics of the cloud layer.

[0050] It can be understood that the Farneback algorithm is a multi-scale based dense optical flow estimation algorithm, the core principle of which is based on the gray conservation assumption and Taylor expansion, and the optical flow vector of each pixel point in the image is calculated by iterative optimization, so as to describe the motion information of the target. Due to the complexity of the target motion in the actual scene, directly solving the above equation is easy to fall into a local optimal solution, and the calculation amount is large. The Farneback algorithm adopts a multi-scale hierarchical strategy, which generates an image pyramid of different resolutions by downsampling the image layer by layer. Starting from the top layer of the pyramid (low resolution image), the pixel points in the low resolution image represent a larger range of image regions, and the calculation is relatively simple, which can quickly obtain a rough optical flow estimation and capture the overall motion trend of the target. Then the optical flow calculation result of the last layer is taken as the initial value of the next layer (slightly higher resolution), and the optical flow is optimized on a finer scale, gradually refining the optical flow estimation, and finally obtaining an accurate optical flow result on the original resolution image. This coarse-to-fine calculation method not only reduces the calculation complexity, but also improves the accuracy and robustness of the optical flow calculation. In each layer of image, in order to solve the optical flow of multiple pixel points, the algorithm uses weighted least squares method to optimize the above equation. By introducing a weight matrix, the contributions of different pixel points are weighted, so that higher weights are given to areas with rich texture and obvious features, so that the optical flow of these key areas can be calculated more accurately, while the interference of areas with noise and unobvious texture is suppressed.

[0051] When calculating the cloud layer movement vector field using satellite cloud images, for large-scale cloud layer movement, such as the movement of an entire cumulonimbus cloud from one area to another, the Farneback algorithm can quickly calculate the approximate movement direction and speed of the cloud layer through the low resolution image at the top layer of the image pyramid, providing macroscopic motion information for subsequent analysis. For example, when a large cumulonimbus cloud is detected on a satellite cloud image, the algorithm can quickly determine that it is moving eastward as a whole, and the moving speed is about 20 kilometers per hour.

[0052] The optical flow estimation obtained in the low-resolution image is only a preliminary result. As it advances to the bottom layer of the image pyramid (high-resolution image), the Farneback algorithm can further refine the stretching, deformation, and local turbulence of the cloud layer edge based on the results of the previous layer. For example, when analyzing the interface between the stratocumulus and cumulonimbus clouds, the algorithm can accurately calculate the direction and degree of distortion of the stratocumulus edge due to the influence of the cumulonimbus airflow. These detailed information is crucial for accurately predicting the impact of cloud layers on photovoltaic power generation.

[0053] In actual meteorological environment, cloud movement is often affected by multiple factors, showing irregular and complex movement patterns. The multi-scale calculation and weighted least squares optimization characteristics of the Farneback algorithm enable it to accurately calculate cloud movement information under different weather conditions, such as strong convective weather leading to rapid changes in cloud layers, or a mixture of multiple cloud types. Whether it is the slow movement of thin and transparent cirrus clouds or the rapid movement of thick and heavy cumulonimbus clouds accompanied by morphological changes, this algorithm can effectively handle it and provide reliable data support for subsequent cloud type identification and light power prediction.

[0054] Therefore, the pyramid structure takes into account both global trends and local details, and compared with traditional single-resolution algorithms (such as Lucas-Kanade), the accuracy of capturing cloud birth and death deformation is improved by 40%. Sparse optical flow blocks reduce mismatching through reverse search, and the overlapping area between blocks (such as 50%) enhances motion continuity, and noise sensitivity is reduced by 30%. The adaptive allocation of weights between high-level and dense optical flow avoids the deviation of a single data source. The actual measurement shows that the root mean square error (RMSE) of the fused vector field and the real cloud movement is reduced from 1.2 m / s to 0.7 m / s.

[0055] For example, 5-phase satellite cloud images are aligned to the power station coordinate system, grayed and normalized to eliminate cloud shadow interference. A 5-layer image pyramid is generated, with the top layer resolution being 64x36 and the bottom layer maintaining the original resolution. The top layer optical flow shows that the cloud layer as a whole moves southwest at 8 m / s; the bottom layer optical flow detects an edge vortex (speed 3 m / s, direction northwest). 200 sparse blocks are matched in the cloud cluster edge area, and abnormal blocks with a similarity of <0.6 are removed. The bottom weighted average optical flow vector is fused with the high-level trend field to generate a full-scene vector field. The power grid dispatching system predicts a 15% decrease in irradiance in the next 10 minutes based on the vector field, and adjusts the inverter output power in advance to avoid power drop.

[0056] Further optionally, the step 23 of calculating the granular optical flow of each pixel can be implemented as:

[0057] For each sequence of images at a resolution, the gray level distribution of a pixel neighborhood is modeled as a quadratic polynomial, the symmetric parameter and the linear parameter in the quadratic polynomial are set by the gray level constancy assumption; the motion equation corresponding to the pixel neighborhood is derived by Taylor expansion to obtain the motion vector value of each pixel as the granular optical flow of each pixel.

[0058] In calculating the granular optical flow of each pixel, the core is to transform the motion information of pixels in the image into quantifiable vector values through mathematical modeling and derivation. The principle is based on in-depth analysis of the gray level distribution of the pixel neighborhood and the premise of the gray level constancy assumption.

[0059] Specifically, for each sequence of images at a resolution, the algorithm first models the gray level distribution of a pixel neighborhood as a quadratic polynomial. The quadratic polynomial can flexibly fit complex gray level changes, and it contains multiple parameters that determine the shape of the polynomial, which in turn reflects the distribution characteristics of the gray level in the pixel neighborhood. The symmetric parameter and the linear parameter in the quadratic polynomial are set by the gray level constancy assumption, which assumes that the gray level of the same object in adjacent short time is constant. The symmetric parameter is used to describe the symmetry of the gray level distribution, and the linear parameter reflects the trend of the gray level in a certain direction. The determination of these two parameters makes the polynomial more accurately fit the actual gray level distribution.

[0060] Then, the motion equation corresponding to the pixel neighborhood is derived by Taylor expansion. Taylor expansion is a mathematical method that approximates a function near a point as a polynomial. Through it, the gray level change relationship of the pixel point at different times can be expressed by a mathematical formula. Combined with the quadratic polynomial with the set parameters and the gray level constancy assumption, the motion vector value of each pixel can be finally solved through a series of mathematical operations. This motion vector value contains the displacement information of the pixel in the horizontal and vertical directions, which intuitively describes the moving direction and distance of the pixel from one time to another. This motion vector value is taken as the granular optical flow of each pixel.

[0061] Taking the movement analysis of a layer cloud in a satellite cloud image as an example, under a sequence of images at a resolution, a pixel point in the layer cloud region is selected. By modeling the gray level distribution of the pixel neighborhood as a quadratic polynomial and setting the relevant parameters according to the gray level constancy assumption, the motion equation is derived by Taylor expansion. Assuming that the calculation results show that the pixel moves 2 pixel units to the right in the horizontal direction and 1 pixel unit downward in the vertical direction, its motion vector value can be represented as (2, 1), which is the granular optical flow of the pixel. In practical applications, a large number of pixel points in the cloud image will be calculated in this way to obtain the granular optical flow distribution of the entire cloud layer region, and then reflect the movement characteristics of the cloud layer.

[0062] This way of calculating the pixel-level optical flow brings significant technical effects. On the one hand, through the fine modeling of the gray scale distribution of the pixel neighborhood and the rigorous derivation based on Taylor expansion, the local information of the image can be fully utilized. Even in the case of complex texture and irregular gray scale change of the cloud layer, the small motion change of the pixel can be accurately captured, greatly improving the accuracy of the optical flow calculation. On the other hand, based on the parameter setting of the gray scale constancy assumption, the algorithm has good stability when processing continuous image sequences, can effectively reduce the influence of noise interference on the calculation results, and has strong adaptability to the motion analysis of different types of cloud layers under different weather conditions. By obtaining the accurate pixel-level optical flow, the actual motion state of the cloud layer can be more truly restored when constructing the cloud layer motion vector field, providing a solid and reliable data foundation for the light power prediction based on the cloud layer motion, thereby improving the prediction accuracy and reliability of the entire new energy light power correction system.

[0063] As an optional embodiment, in step S103, based on the cloud layer motion vector field, the cloud layer type contained in the target area is identified, and the coverage rate of the cloud layer type in the target area is obtained, which can be implemented as:

[0064] The Otsu algorithm is used to dynamically determine the cloud layer gray scale threshold for the corrected cloud layer motion vector field, and the cloud layer gray scale threshold is used for adaptive threshold segmentation of the cloud layer region and the background region to obtain a cloud layer mask. The segmentation threshold corresponding to the high humidity region is dynamically adjusted in combination with the ground humidity data in the ground data of the meteorological station. The cloud layer mask is classified to obtain the cloud layer type contained in the target area, and the coverage rate of each cloud layer type in the target area is calculated.

[0065] Specifically, when identifying the cloud layer type in the target area based on the cloud layer motion vector field and calculating the coverage rate, the core is to convert the vector field information into recognizable cloud layer categories and coverage ratios through image segmentation and data analysis. The process starts with threshold segmentation using the Otsu algorithm. Based on the gray scale distribution characteristics of the image, the algorithm automatically finds an optimal gray scale threshold by maximizing the inter-class variance, dividing the pixels in the image into foreground (cloud layer region) and background (non-cloud layer region). When processing the cloud layer motion vector field, the Otsu algorithm can dynamically determine a suitable threshold based on the differences in gray scale values of different cloud layer types and backgrounds, separating the cloud layer from the background, and thus obtaining a preliminary cloud layer mask, i.e., a binary image marking the location of the cloud layer.

[0066] However, in actual meteorological environment, cloud distribution is closely related to humidity and other factors. For example, in high humidity areas, the gray scale characteristics of the cloud layer may change due to water vapor condensation and other reasons, which is different from the gray scale distribution in the conventional case. Therefore, combined with the ground humidity data in the ground data of the meteorological station, the segmentation threshold of the high humidity area is dynamically adjusted. When detecting that the humidity in a certain area is high, the system will appropriately lower the segmentation threshold of the area according to the pre-established humidity-gray scale threshold correlation model, so that the cloud layer part that may be misjudged as background due to the low gray scale value in the high humidity environment can be correctly divided into the cloud layer area, further optimizing the accuracy of the cloud mask.

[0067] After obtaining a relatively accurate cloud mask, the cloud mask is processed by image classification technology to identify the cloud type contained in the target area. This process usually relies on a pre-trained deep learning model (such as MobileNetV3), which has been trained on a large number of different cloud type image data and can learn the texture, shape, gray scale distribution and other characteristics of cumulonimbus, stratus, cirrus and other types of clouds. After inputting the cloud mask into the model, the model will output the prediction probability of each cloud type, and by setting a suitable probability threshold, the final cloud type can be determined. For example, if the model outputs a prediction probability of 0.8 for cumulonimbus in a certain area, a prediction probability of 0.15 for stratus, and a prediction probability of 0.05 for cirrus, and the probability threshold is set to 0.5, then the area will be determined as cumulonimbus.

[0068] After determining the cloud type, the coverage rate of each type of cloud in the target area can be obtained by calculating the pixel proportion of each type of cloud. Specifically, the number of pixels corresponding to each cloud type in the cloud mask is counted, and then divided by the total number of pixels in the target area to obtain the coverage rate of that cloud type in the target area. For example, if the total number of pixels in the target area is 10,000, and the number of pixels corresponding to cumulonimbus is 3,000, then the coverage rate of cumulonimbus in the area is 30%.

[0069] Taking an actual meteorological monitoring as an example, after using Otsu algorithm to perform initial threshold segmentation on the cloud movement vector field in the target area covered by the satellite cloud image, the approximate cloud contour is obtained. However, in the high humidity area in the lower right corner of the image, due to the high humidity, the gray scale value of the cloud layer is low, and part of the cloud layer is not completely segmented out. After adjusting the segmentation threshold by combining the ground humidity data, the cloud layer in this area is completely presented. By inputting the optimized cloud mask into the classification model, it is identified that the area mainly contains cumulonimbus and stratus, and further calculation shows that the coverage rate of cumulonimbus is 45% and the coverage rate of stratus is 35%.

[0070] In terms of accuracy, the automatic threshold segmentation of the Otsu algorithm and the dynamic adjustment based on humidity data effectively overcome the uncertainty of cloud gray level changes, improving the precision of the cloud mask. The application of deep learning models utilizes their powerful feature extraction capabilities to accurately identify various types of clouds, avoiding misjudgment problems caused by insufficient feature analysis in traditional methods. In terms of adaptability, by integrating ground data from weather stations, the algorithm can dynamically adjust the processing strategy according to different weather conditions, allowing it to stably identify cloud types and calculate coverage rates in both dry and humid regions. These technical effects provide reliable data support for subsequent cloud type-based light power prediction, such as predicting the rapid drop in light power caused by fast-moving cumulonimbus clouds or the slow power decay caused by persistent stratus clouds, allowing the system to more accurately predict and respond, thereby improving the overall performance and reliability of the new energy light power correction system.

[0071] It is worth noting that the Otsu algorithm, also known as the Otsu method, is an unsupervised adaptive threshold segmentation algorithm based on image gray level distribution characteristics. Its core goal is to find an optimal gray level threshold to accurately divide image pixels into foreground and background.

[0072] In the implementation process, the Otsu algorithm finds the best segmentation point by mining the gray level distribution of the image. Below, we will explain the core logic of determining the threshold in image segmentation in a more intuitive and understandable way. The core of the Otsu algorithm's calculation logic is how to find the best gray level threshold to accurately divide the image into foreground and background. First, get the number of pixels at each gray level in the image, such as in an 8-bit image, count how many pixels there are from 0 to 255 gray levels, and draw a gray level histogram with these data. Through the histogram, you can roughly see the distribution of the image's pixel gray levels.

[0073] Next, use different gray values as the segmentation threshold. It starts from the smallest gray level 0 and assumes that this value divides the image into two parts, with gray levels less than or equal to 0 belonging to the foreground and greater than 0 belonging to the background. Then, the algorithm calculates the proportion of each part in the entire image and the average gray level of each part. For example, if the foreground part has fewer pixels, its proportion in the image will be small; if the foreground part has generally lower pixel gray levels, its average gray level will also be smaller. After that, the algorithm repeats this operation for each possible gray value, constantly trying new threshold values to divide the foreground and background, and calculating the corresponding proportion and average gray level after each division.

[0074] After each attempt, the effect of the current division is evaluated, and the criterion for evaluation is to calculate the difference between the foreground and the background. The greater the difference, the better the current threshold separates the foreground and the background. For example, in a crowd of people, if height is used as a division standard, a height value that can clearly separate adults and children is a good division standard. Finally, after the algorithm traverses all possible gray values, it finds the gray value that maximizes the difference between the foreground and the background, and determines it as the final threshold value. Using this threshold value to segment the image can reasonably divide the pixels in the image into foreground and background. When processing cloud movement vector field images, the Otsu algorithm automatically finds the most suitable threshold value in this way, clearly separates the cloud area from the background, and prepares for subsequent work such as judging the type of cloud and calculating the coverage rate.

[0075] First, the fixed threshold method is a simple image segmentation technique that directly uses a pre-set fixed gray value as the threshold for image segmentation. The limitation of this method is that when the lighting conditions, shooting environment, etc. of the image change, the fixed threshold may no longer be applicable, resulting in poor segmentation results. For example, in cloud images taken under different weather and lighting conditions, the fixed threshold may not accurately distinguish the cloud layer from the background. The Otsu algorithm does not require manual threshold setting and can automatically calculate the optimal threshold based on the gray scale distribution characteristics of the image itself, providing better segmentation results for images under different environments, with stronger adaptability and universality. Second, region-based segmentation methods (such as region growing algorithm) start from seed points in the image and gradually merge adjacent pixels based on certain similarity criteria (such as gray scale similarity, texture similarity, etc.) to form target regions. Although this method can consider local features of the image, it has high computational complexity and is sensitive to the selection of seed points and the setting of similarity criteria. In contrast, the Otsu algorithm performs global optimization based on the overall gray scale distribution of the image, with a relatively simple and efficient calculation process, and does not require pre-specified seed points or complex similarity rules. When dealing with cloud images with significant differences in gray scale distribution, the Otsu algorithm can quickly and accurately achieve segmentation. Third, edge-based segmentation methods (such as Canny edge detection algorithm) detect edge information with sharp gray scale changes in the image and then connect the edges to form target regions. However, in cloud images, the cloud edges are often blurred and easily disturbed by noise, making it difficult for edge-based methods to accurately extract the complete cloud contour. The Otsu algorithm, on the other hand, starts from the global gray scale distribution of the image and determines the threshold by maximizing the inter-class variance, effectively handling fuzzy edges and noise, making it more suitable for cloud segmentation with unclear boundaries. In summary, the Otsu algorithm has obvious advantages in image segmentation accuracy, robustness, and computational efficiency compared to many existing techniques, especially for cloud images with significant differences in gray scale distribution and variable environments.

[0076] Further optionally, before the above step of dynamically determining the cloud gray scale threshold using the Otsu algorithm on the modified cloud movement vector field and adaptively thresholding the cloud region and background region based on the cloud gray scale threshold to obtain the cloud mask, the following steps can also be performed:

[0077] The wind speed data in the numerical weather prediction is obtained in combination with the meteorological data and the ground data of the weather station, and the cloud movement speed in the cloud movement vector field is corrected based on the wind speed value. Further, the cloud thickness is inversely deduced from the cloud top height data, and the maximum displacement amount of the optical flow field is dynamically set based on the cloud thickness to constrain the cloud top height. Then, the cloud holes in the cloud movement vector field are filled by a closing operation to eliminate small-area noise in the cloud movement vector field.

[0078] Before the cloud threshold segmentation by using the Otsu algorithm, the preprocessing step of the cloud movement vector field further optimizes the accuracy and reliability of the vector field by integrating multi-source meteorological data and image morphological processing. The core logic is to correct the motion information in the vector field by using meteorological parameters such as wind speed and cloud top height, and to eliminate data noise by image operation, to lay a more solid data foundation for subsequent cloud type recognition.

[0079] Firstly, the cloud movement speed in the cloud movement vector field is corrected in combination with the wind speed information in the meteorological data and the ground observation data. In the atmospheric environment, the actual movement of the cloud layer is significantly affected by the wind speed, for example, the cloud movement speed in a strong wind environment will be much higher than the regular calculated value. After the system obtains the wind speed data in the numerical weather prediction, the cloud movement speed corresponding to each pixel point in the vector field is adjusted according to the physical correlation model between the wind speed and the cloud movement speed. If the model calculates that the cloud movement speed in a certain area is 10 meters per second, but the current wind speed data shows that there is strong wind in the area, the actual wind speed causes the cloud movement speed to increase to 15 meters per second, and the system will correct the speed value of the corresponding pixel in the vector field to 15 meters per second, so that the vector field is more consistent with the real movement state of the cloud.

[0080] Then, the cloud thickness is inversely deduced from the cloud top height data, and the maximum displacement amount of the optical flow field is dynamically set based on the cloud thickness. The cloud top height and the cloud thickness are inherently related. Generally speaking, the higher the cloud top, the thicker the cloud, and the more complex the internal airflow movement, which may cause a large displacement of the cloud in a short time. The system estimates the approximate thickness of the cloud in different areas according to the cloud top height data combined with the cloud thickness estimation model in meteorology. For the cloud area with large thickness, the maximum displacement amount of the optical flow field is appropriately increased to allow the algorithm to consider larger pixel displacement changes when calculating the optical flow; on the contrary, for the thin cloud area, the maximum displacement amount is reduced to avoid errors caused by overestimation of displacement. For example, when it is detected that the cloud top height of a cumulonimbus cloud area is high, and the thickness is large, the maximum displacement amount of the optical flow field in this area is increased from the default 5 pixel units to 8 pixel units, to ensure that the algorithm can capture the pixel displacement caused by the intense movement inside the cumulonimbus cloud.

[0081] Finally, the cloud layer voids in the cloud layer movement vector field are filled by closing operation to eliminate small area noise. During the calculation of the vector field, due to factors such as cloud layer edge blur and light changes, pixels in the cloud layer region may be missing or misjudged, forming voids or small area noise regions. As a kind of image morphological operation, closing operation can connect adjacent cloud layer pixels, fill voids, and smooth the cloud layer edge by expanding and then eroding. In specific operation, the algorithm will expand the edge of the cloud layer region outward (expansion), connect the originally separated pixels, and then shrink the expanded region inward (erosion) to restore the general shape of the cloud layer while removing the redundant pixels introduced by expansion. After closing operation, the voids in the cloud layer movement vector field are effectively filled, the noise regions are eliminated, and the cloud layer contour is more complete and clear.

[0082] Taking a strong convective weather monitoring as an example, in the initial calculation of the cloud layer movement vector field in a certain area, the cloud layer movement speed is underestimated due to the lack of consideration of wind speed, resulting in a mismatch between the cloud layer movement trend in the vector field and the actual situation. After introducing wind speed data for correction, the cloud layer movement speed is accurately adjusted, which is more consistent with the actual situation in strong wind environment. When calculating the cloud layer thickness, it is found that the cloud top height of a cumulonimbus region is relatively high, and the system automatically increases the maximum displacement of the optical flow field in this region, so that the algorithm successfully captures the pixel displacement changes caused by the rapid airflow movement inside the cumulonimbus. Finally, after closing operation, the noise region at the edge of the cumulonimbus caused by light reflection and the small area voids inside the cumulonimbus are eliminated, and a clear and complete cloud layer vector field image is obtained. Using the processed vector field for subsequent Otsu algorithm threshold segmentation, a more accurate cloud layer mask is successfully obtained, providing high-quality data for cloud layer type recognition, and finally accurately identifying that the main cloud layer in this area is cumulonimbus, and the deviation of the coverage calculation result is reduced by 12% compared with the result before processing.

[0083] These preprocessing steps bring significant technical effects. In terms of accuracy, wind speed correction makes the cloud layer movement speed more consistent with the actual situation, avoiding errors in subsequent cloud layer movement path prediction due to speed misjudgment; optical flow field adjustment based on cloud top height ensures that the algorithm can adapt to the complex motion of cloud layers of different thicknesses, reducing pixel displacement calculation errors; closing operation eliminates noise and voids, improving the accuracy of cloud layer region division, providing better input data for Otsu algorithm, and making the threshold segmentation result more accurate. In terms of robustness, the fusion of multi-source meteorological data and the dynamic adjustment strategy enable the system to adapt to cloud layer changes under different weather conditions, whether it is strong wind, high humidity, or complex cloud layer structure, it can stably output reliable cloud layer movement vector field data, thereby improving the overall performance and reliability of subsequent cloud layer type recognition, light power prediction, etc., and reducing the risk of prediction failure caused by data errors.

[0084] As an optional embodiment, in the above steps, the cloud layer mask is classified to obtain cloud layer types contained in the target region, and the coverage of each cloud layer type in the target region is calculated, including the following steps:

[0085] Step 31, a real-time cloud layer classification model is constructed by using MobileNetV3; wherein the real-time cloud layer classification model at least includes: a convolutional layer for extracting cloud image features, a channel attention module for strengthening cloud texture features, a Swish activation function for processing nonlinear expression, and a feature pyramid network for improving cloud cluster detection accuracy in multiple scales;

[0086] Step 32, the cloud layer type of the cloud layer mask is predicted by the real-time cloud layer classification model to obtain the prediction probability distribution of each cloud layer type in the cloud layer mask;

[0087] Step 33, a single-type mask of each cloud layer type is generated based on the prediction probability distribution, and the single-type mask is image-enhanced according to the dynamic adjustment of the contrast of the cloud layer type; if the cloud layer type is cumulonimbus, the contrast is enhanced by 50%; if the cloud layer type is stratus, the contrast is reduced by 30%;

[0088] Step 34, the ratio between all pixels in the single-type mask and the total pixels in the cloud layer mask is calculated by using the pixel statistics method to obtain the single-type coverage of each cloud layer type; for non-continuous cloud area, the continuous coverage area is generated by using the Kriging interpolation method, the pixel area proportion in the continuous coverage area is calculated, and the single-type coverage of each cloud layer type is optimized according to the pixel area proportion;

[0089] Step 35, for the pixels covered by multiple single-type masks, the single-type coverage of each cloud layer type in different positions of the target region is superimposed into the cloud layer mask according to the confidence weighted distribution to obtain the coverage calculation result of the target region.

[0090] Specifically, in the cloud layer type classification and coverage calculation process, the complete process from feature extraction to accurate coverage calculation is realized by combining deep neural network and image processing technology.

[0091] In the cloud layer mask-based type classification and coverage calculation, first, a real-time cloud layer classification model is constructed using MobileNetV3. The model adopts a lightweight network architecture, extracts features of the cloud layer mask through convolutional layers, and can capture spatial features such as texture and shape of different cloud layer types. The introduction of the channel attention module strengthens the attention to cloud layer texture features, enabling the model to distinguish typical features such as the flocculent structure of cumulonimbus and the uniform distribution of stratus. The Swish activation function improves the non-linear expression ability of the model, better fitting the complex mapping relationship between cloud layer features and types. The addition of the feature pyramid network realizes multi-scale feature fusion, which can handle both the overall outline and local details of the cloud cluster, such as when identifying cirrus, paying attention to both its large-scale filamentous structure and the subtle curling features at its edges.

[0092] After inputting the cloud layer mask into the trained real-time cloud layer classification model, the model outputs the prediction probability distribution of each cloud layer type. For example, for a cloud layer mask of a certain area, the model may output a cumulonimbus probability of 0.7, a stratus probability of 0.2, and a cirrus probability of 0.1. Based on these probability distributions, single-type masks of each cloud layer type are generated, i.e., areas that may belong to cumulonimbus, stratus, and cirrus are marked respectively. To further improve the accuracy of subsequent coverage calculation, the contrast is dynamically adjusted for image enhancement according to the cloud layer type. For cumulonimbus, its edges are usually clear but the internal gray difference is large, so enhancing the contrast by 50% can highlight its texture details. For stratus, the gray distribution is relatively uniform, so reducing the contrast by 30% can reduce the misjudgment caused by local noise.

[0093] In calculating the coverage, first, the pixel statistics method is used to calculate the ratio of the number of pixels in each single-type mask to the total number of pixels in the cloud layer mask, obtaining the initial single-type coverage. For example, if the number of pixels in the cumulonimbus single-type mask accounts for 30% of the total number of pixels, the initial coverage of cumulonimbus is 30%. For non-continuous cloud areas, such as cumulonimbus fragments scattered by strong winds, the Kriging interpolation method is used for spatial interpolation of the cloud area to generate a continuous coverage area, and then the pixel area ratio of this area is calculated to optimize the initial coverage, making the result more consistent with the actual cloud layer distribution.

[0094] When multiple single-type masks overlap, i.e., the same pixel position is covered by multiple cloud layer types at the same time, weighted distribution is performed according to the prediction confidence of each cloud layer type. For example, the cumulonimbus prediction probability of a certain pixel position is 0.8, and the stratus prediction probability is 0.2, then 80% of the area of this pixel is allocated to cumulonimbus and 20% to stratus. Finally, the coverage of each cloud layer type at different positions is superimposed on the original cloud layer mask to obtain the final coverage distribution of cumulonimbus, stratus, cirrus, and other cloud layers in the target area.

[0095] In a strong convective weather monitoring, the system processes satellite cloud images and identifies the target area with cumulonimbus, stratus and cirrus clouds through real-time cloud classification model. The initial pixel statistics show that the cumulonimbus coverage is 25%, but after the Kriging interpolation method is used to optimize the non-continuous cloud area, the actual coverage area of cumulonimbus is expanded and the coverage rate is adjusted to 32%. In the overlapping area processing, a large area is covered by cumulonimbus and stratus at the same time. Through confidence weighted distribution, the coverage rate of cumulonimbus in this area increases by 8% and that of stratus decreases by 5%. This processing makes the final coverage calculation result closer to the actual cloud distribution, providing more accurate data support for subsequent light power prediction.

[0096] In this way, in terms of classification accuracy, the lightweight architecture and multi-scale feature extraction capability of MobileNetV3 enable the model to accurately distinguish different cloud types while maintaining high efficiency. Especially for the edge-fuzzy cirrus and the morphologically variable cumulonimbus, the recognition accuracy is significantly improved. Dynamic contrast adjustment further enhances the feature differences of various cloud layers and reduces the misjudgment rate. In terms of coverage calculation, Kriging interpolation solves the coverage calculation problem of non-continuous cloud area, making the result more consistent with the actual meteorological scene. Confidence weighted distribution reasonably handles the coverage attribution problem in the overlapping area of cloud layers, improving the calculation accuracy. These technical improvements enable the system to accurately provide coverage data of various cloud layers under complex meteorological conditions, providing more reliable input for light power prediction model, and ultimately improving the accuracy of photovoltaic power generation scheduling and the stability of power grid operation.

[0097] Step S104, if the cumulonimbus coverage rate exceeds the first threshold, the cloud movement path is predicted in combination with the cloud movement vector field to obtain the first cloud coverage change sequence of the target area in the first time period.

[0098] Step S105, if the stratus coverage rate exceeds the second threshold, the irradiance attenuation coefficient is predicted based on the cloud thickness and solar elevation angle in the meteorological data, and the ground aerosol optical depth obtained based on the ground data of the meteorological station is combined to correct the irradiance attenuation coefficient to obtain the second cloud coverage change sequence of the target area in the second time period.

[0099] Specifically, as an optional embodiment, the light power prediction system implements a refined analysis of the impact of different cloud types. In predicting the change in cumulonimbus coverage, in the above step S104, the system first determines whether the cumulonimbus coverage exceeds a first threshold (e.g., 30%). Cumulonimbus clouds have strong vertical development and fast movement, and have the characteristics of strong burst and large attenuation amplitude on photovoltaic power generation. Based on the cloud movement vector field calculated in step S102, the system predicts the movement path of the cumulonimbus cloud by combining the Lagrangian trajectory tracking method in fluid mechanics. Specifically, the cloud movement vector field is regarded as a fluid velocity field, and the motion trajectory of each particle in the cloud cluster is calculated by numerical integration, thereby obtaining the coverage change sequence of the cumulonimbus cloud in the first time period (e.g., 30 minutes) in the future. In this process, the system will consider the Coriolis force caused by the Earth's rotation and the influence of terrain on airflow, and correct the predicted trajectory. For example, when the cumulonimbus cloud approaches a mountain, the system will predict the possible flow around or climb path of the cloud cluster according to the terrain data.

[0100] For the prediction of stratocumulus coverage change, in the above step S105, when the stratocumulus coverage exceeds a second threshold (e.g., 50%), the system adopts a prediction method based on a physical model. Stratocumulus clouds are usually uniformly distributed and have stable thickness, and their influence on solar radiation mainly manifests as sustained attenuation. The system first calculates the initial irradiance attenuation coefficient according to the cloud thickness in the meteorological data and the solar elevation angle. The thicker the cloud, the lower the solar elevation angle, and the larger the attenuation coefficient. Then, the attenuation coefficient is corrected in combination with the aerosol optical thickness in the ground data of the weather station. The higher the concentration of aerosols, the stronger the scattering and absorption of solar radiation, further reducing the irradiance reaching the ground. In this way, the system can obtain the stratocumulus coverage change sequence in the second time period (e.g., 2 hours) in the target area, which not only reflects the physical shielding effect of the cloud, but also considers the influence of atmospheric composition on solar radiation transmission.

[0101] For example, in a certain actual prediction scenario, the system detects that the cumulonimbus coverage reaches 35%, exceeding the first threshold. Through vector field analysis, it is predicted that the cumulonimbus cloud will cover the entire target area after 20 minutes, and the light power prediction model accordingly issues an early warning, and the dispatch system starts the standby power source in time. At the same time, the stratocumulus coverage is 60%, and the system calculates the initial attenuation coefficient to be 0.6 according to the cloud thickness and solar elevation angle, and adjusts the attenuation coefficient to 0.7 after correction in combination with the ground aerosol optical thickness, predicting that the light power will continue to decrease steadily in the next 2 hours.

[0102] Further optionally, the settings of the first threshold value and the second threshold value are determined in combination with business objectives, historical data, and physical characteristics of cloud layers. From a core principle, cumulonimbus clouds, as a symbol of severe convective weather, require that the first threshold value be associated with disaster warning sensitivity, such as triggering an emergency response when the coverage rate exceeds a certain percentage, while the second threshold value of stratocumulus clouds needs to be combined with the requirement for light attenuation, such as photovoltaic power stations determining whether stratocumulus clouds affect power generation prediction. Data-driven statistical analysis is an important means, such as the minimum coverage rate of cumulonimbus clouds triggering disasters in historical statistics (such as 20% corresponding to 80% probability of heavy rain) as the first threshold value benchmark, or determining the second threshold value by the correlation between stratocumulus cloud coverage and irradiance attenuation (such as coverage exceeding 40% when irradiance decreases to 60% of sunny day). In terms of physical mechanisms, cumulonimbus clouds are usually low (10%-30%) to avoid missing detection due to strong vertical development and large single group hazards, while stratocumulus clouds are usually high (30%-50%) to balance the calculation accuracy due to wide horizontal coverage and linear attenuation.

[0103] In a specific method, the historical quantile method determines the threshold value by calculating the quantile of the cloud coverage rate of the target area, such as setting the first threshold value to 25% when the 90% quantile of the cumulonimbus cloud coverage rate is 25%, and setting the second threshold value when the stratocumulus cloud coverage rate exceeds 35% and the irradiance attenuation exceeds the expected value; the business scenario enumeration method determines the critical value based on actual impact, such as diverting when the cumulonimbus cloud coverage exceeds 15% in aviation meteorology, and affecting photosynthesis when the stratocumulus cloud coverage exceeds 45% in agricultural light. The dynamic adjustment method adjusts the threshold value according to real-time meteorological parameters or seasonal changes, such as temporarily reducing the cumulonimbus cloud threshold value to 12% in summer and adjusting the stratocumulus cloud threshold value from 40% to 30% in winter at high latitudes.

[0104] Threshold calibration and verification are achieved through cross-validation and feedback iteration. The former tests the hit rate and false alarm rate with independent data to balance the warning accuracy of the cumulonimbus cloud threshold value, or compares the measured and predicted errors of irradiance under the stratocumulus cloud threshold value; the latter is revised regularly through business feedback, such as reducing the cumulonimbus cloud threshold value from 20% to 18% due to high false alarm rate. In terms of technical effects, reasonable threshold values can accurately trigger corresponding analysis logic and enhance the explainability of decision-making, but need to avoid single dependence, combined with cloud speed, thickness, and other indicators, and pay attention to local adjustment of threshold values caused by regional climate differences to ensure that the algorithm meets actual business needs.

[0105] Therefore, the cumulonimbus path prediction solves the problem of low prediction accuracy in complex terrain by combining fluid mechanics models and terrain correction, enabling the system to capture the rapid change in optical power caused by the rapid movement of cumulonimbus clouds in advance, providing more response time for grid dispatch. The multi-factor correction mechanism of the layer cloud attenuation coefficient considers the physical properties of the cloud layer and the atmospheric environmental factors, improving the prediction ability of the optical power change under persistent weak light conditions, and helping to optimize the energy storage management strategy of the photovoltaic power station. By differentiating the prediction of different cloud types, the overall prediction error of the system is reduced, especially in mixed cloud coverage scenarios, the prediction accuracy is improved more obviously, effectively supporting the stable grid connection of new energy power.

[0106] It can be understood that the Lagrangian trajectory tracking method is a trajectory analysis technique based on fluid mechanics and particle motion theory, and its core principle is to study the motion characteristics of fluid by tracking the motion trajectory of a single particle in the fluid. This method is based on the Lagrangian point of view, focusing on the position, velocity, and acceleration of each particle in the fluid at different times, and by establishing the differential equation of particle motion and solving it, the simulation and tracking of particle motion trajectory are realized. In practical applications, the flow field velocity data usually comes from meteorological observations, numerical simulations or experimental measurements. For example, in atmospheric science, wind speed data observed by meteorological stations or satellites can be used to construct a three-dimensional wind field, and air particles are regarded as particles to track their motion trajectory in the wind field to study the processes of pollutant diffusion, water vapor transport, etc. In oceanography, this method can be used to analyze the influence of ocean currents on the movement of buoys or marine organisms. The key advantage of the Lagrangian trajectory tracking method is that it can directly reflect the actual motion path of the particles, revealing complex phenomena such as vortices, separation and mixing in the flow field. Compared with the Euler method (focusing on the flow characteristics of fixed space points), it is more suitable for studying the long-term migration and diffusion behavior of particles, especially in the fields of environmental science, atmospheric dynamics, ocean engineering, etc. It has wide application. For example, by tracking the trajectory of volcanic ash particles, their diffusion range and impact area can be predicted; by tracking the moving trajectory of the center of a typhoon, meteorological forecasting can be assisted.

[0107] Further optionally, the correlation between cloud coverage and irradiance attenuation is determined by the optical properties of the cloud layer, geometric distribution, and solar radiation transmission mechanism. From a physical principle, the cloud layer weakens solar radiation through scattering, absorption, and reflection: a uniformly distributed cloud layer follows the Beer-Lambert law, and the irradiance attenuation is exponentially related to the optical thickness, while the optical thickness is positively correlated with the coverage, cloud thickness, and particle density; The Mie scattering properties of cloud particles make thick clouds such as cumulonimbus clouds significantly weaken direct irradiance even with low coverage, while thin cirrus clouds have weak attenuation due to sparse particles.

[0108] The key influencing factors include cloud type and structure (stratus cloud coverage and attenuation are positively correlated, cumulonimbus local shielding leads to large fluctuations in attenuation, and high cloud attenuation amplitude is small), solar elevation angle (the lower the elevation angle, the more significant the attenuation under the same coverage), and atmospheric aerosol (cooperating with cloud attenuation, enhancing scattering). In the quantitative relationship modeling, the empirical statistical model fits the coverage and attenuation rate by linear or exponential function (such as stratus cloud ΔE / E0=0.8C, cumulonimbus ΔE / E0=1-exp(-3C), and the radiation transfer model simulates the transfer process based on optical thickness and scattering algorithm, and needs to introduce dynamic correction factor to deal with cloud drift and spectral difference.

[0109] In application scenarios, photovoltaic prediction focuses on linear or exponential attenuation of stratus cloud (coverage of 20%-80% corresponds to attenuation of 25%-80%). The current challenge lies in non-uniform cloud characterization, temporal and spatial dynamic correction, and multi-source data fusion, which needs to combine cloud cluster morphology, real-time tracking, and ground data to improve prediction accuracy. In summary, the correlation between the two needs to be comprehensively characterized through physical modeling and data-driven to adapt to the fine needs of different fields.

[0110] In step S106, real-time operation data of the power station is monitored. Specifically, monitoring the real-time operation data of the power station is a basic link for the new energy light power correction system to realize accurate regulation, and the core is to master the operation state of the power station in real time through multi-dimensional data acquisition and analysis, and to provide reliable basis for subsequent power prediction and power grid parameter adjustment.

[0111] From the principle, the operation state of the power station is a complex dynamic system, which is affected by many factors such as light conditions, equipment performance, and power grid load. Through the deployment of various sensors in the power station, such as photovoltaic module current / voltage sensor, inverter monitoring device, power metering equipment, etc., key data covering power generation, component temperature, current and voltage parameters, equipment operation state, etc. can be collected in real time. These data not only reflect the actual output power of the power station, but also contain information such as equipment health status and operation efficiency. For example, the temperature of the photovoltaic module will affect its power generation efficiency, by monitoring the temperature data, combined with the temperature-power characteristic curve, the influence degree of temperature on power generation can be analyzed; the running parameters of the inverter can reflect its power conversion efficiency and stability, and abnormal parameter fluctuation may indicate equipment failure or performance decline.

[0112] As an optional embodiment, in the above step S106, the matching relationship between the cloud layer movement direction and the power station power drop area is predicted by using a double-flow Transformer architecture combined with the first cloud layer coverage change sequence and / or the second cloud layer coverage change sequence, and the power fluctuation range corresponding to each power station power drop area is calculated, including:

[0113] The double-flow Transformer is used to generate cloud layer geometric distribution features and historical power fluctuation pattern features; wherein, the first cloud cover change sequence and / or the second cloud cover change sequence is input into the spatial Transformer, the cloud layer geometric features are extracted through geographical position coding, and a cloud layer movement heat map is generated; the historical power data is input into the time-aware Transformer, the power mutation point features, fluctuation frequency features and decay rate features in the sliding window within the preset time length in the future are captured, and the historical power fluctuation pattern features are obtained; the cloud layer geometric features at least include: cloud layer movement direction vector, cloud cluster center coordinates, edge contour and diffusion speed; through the cross-modal attention mechanism, the cosine similarity matrix of the cloud layer movement direction vector and the power decline area heat map in the cloud layer movement heat map is calculated, and the attention guided matching relationship features are obtained through Softmax weighted fusion; the historical power fluctuation pattern features are optimized by using the gating mechanism to dynamically adjust the fusion weight; based on the matching relationship features and the historical power fluctuation pattern features, the power fluctuation range corresponding to the power decline area of each power station is predicted through the full connection layer.

[0114] Therefore, based on the power fluctuation prediction of the double-flow Transformer architecture, the cross-modal correlation analysis from meteorological data to power station area power influence is realized by fusing cloud layer spatial features and power time sequence features. The core principle is to decouple and model the geometric information of cloud cover change and the time pattern of historical power fluctuation, and then to mine the dynamic matching relationship of the two through the attention mechanism.

[0115] Specifically, the spatial Transformer receives the first or second cloud cover change sequence (such as the cumulonimbus movement path or the stratocumulus irradiance decay sequence), maps the cloud chart coordinates into learnable embedding vectors through geographical position coding, extracts geometric features such as cloud layer movement direction vector, cloud cluster center coordinates, edge contour and diffusion speed in combination with the convolutional neural network, and generates a heat map reflecting the dynamics of cloud cover. For example, in the cumulonimbus movement heat map, the deeper the color, the higher the cloud cover probability at a certain time in the future. The time-aware Transformer processes the historical power data, captures time sequence features such as power mutation points (such as events with power decline exceeding 15% within 10 minutes), fluctuation frequency (power fluctuation times per hour) and decay rate (power decline amplitude per minute) through the sliding window (such as 5-minute interval data within 30 minutes in the future), and forms a feature vector of power fluctuation pattern.

[0116] The cross-modal attention mechanism quantifies the correlation between cloud movement and power reduction in different regions by calculating the cosine similarity between the cloud movement direction vector and the power reduction area heat map. After normalizing the similarity matrix with the Softmax function, the attention weights of each region are obtained. For example, if the similarity between the cloud movement direction vector and the power reduction heat map of a certain region is 0.8, that region will be given a higher weight, indicating that cloud movement has a more significant impact on that region's power. At the same time, the gating mechanism dynamically adjusts the fusion weights of time and spatial features based on real-time power data, suppressing irrelevant noise. For example, when the cloud moves steadily but the power fluctuates abnormally, the spatial feature weight is automatically reduced and the analysis proportion of the time feature is increased. Finally, the fully connected layer integrates the matching relationship features and the optimized power fluctuation pattern features, and outputs the power fluctuation range of each power station region, such as predicting that a certain photovoltaic array region will have a power reduction of 8%-12% in the next 5 minutes.

[0117] For example, in a real-time test of a typhoon peripheral cloud system, the spatial Transformer captures the movement of a cumulonimbus cloud at a speed of 25 km / h to the southeast, and the generated heat map shows that the northwest region of the power station will be covered in 20 minutes. The time-aware Transformer analyzes historical data and finds that under similar weather conditions, the region has experienced a power reduction of 10%-15% every 10 minutes. The cross-modal attention mechanism calculates the cosine similarity between cloud movement and power reduction in this region as 0.85, and after gating optimization, the fully connected layer predicts that the power in this region will decrease by 11%-14%, with an actual monitoring value of 12.5%, and the prediction error is controlled within 2%.

[0118] In this way, independent modeling of spatial and temporal features avoids prediction bias caused by insufficient single-modal information, such as overestimating the impact of stable clouds by relying solely on cloud positions while ignoring historical fluctuation patterns. The cross-modal attention mechanism captures the spatial correspondence between cloud movement and power reduction in real time, making it more suitable for non-uniform cloud coverage scenarios than traditional regression models. The gating mechanism enhances the adaptability to sudden power fluctuations, such as abnormal attenuation caused by equipment failure, which is not mistakenly attributed to cloud impact. Based on the heat map analysis of the geographical distribution of the power station, the power fluctuation of the specific photovoltaic array region can be located, providing accurate basis for grid partition control, reducing the prediction error compared to the average prediction error of the entire station. Through this architecture, the system realizes the causal chain modeling from cloud movement to regional coverage to power fluctuation, significantly improving the spatial resolution and temporal accuracy of new energy power prediction under complex weather conditions.

[0119] In another optional embodiment, after monitoring the real-time operation data of the power plant in step S106, a structural causal model can also be constructed using the Do-Calculus causal modeling method; wherein the structural causal model is used to infer the causal effect of the cloud coverage change amount on the power generation; the PC algorithm is used to estimate the conditional independence, and the structural equation model parameters of the structural causal model are optimized by maximum likelihood estimation; the first cloud coverage change sequence and / or the second cloud coverage change sequence, the real-time operation data of the power plant, and the real-time wind speed data in the ground data of the weather station are input into the structural causal model, the direct influence path and / or the indirect influence path of the cloud movement state in the target area on the power generation are simulated, and the power fluctuation of each path is calculated; a preset number of cloud chart evolution paths are generated by Bootstrap resampling, and the power fluctuation range corresponding to each cloud chart evolution path is calculated based on the pre-set confidence interval, so as to optimize the power fluctuation range corresponding to each power generation station power reduction area.

[0120] Specifically, the core principle of the above steps is to quantify the causal relationship between cloud coverage change and power generation fluctuation through causal modeling, and to improve prediction accuracy by combining data-driven and causal reasoning. Specifically, first, a structural causal model (SCM) is constructed based on the Do-Calculus causal modeling method, which describes the causal relationship between variables through a directed acyclic graph (DAG). For example, the cloud coverage change amount is taken as an intervention variable, the power generation is taken as a result variable, and real-time wind speed and other confounding variables are also included. In the modeling process, the PC algorithm (constraint-based causal discovery algorithm) is used to estimate the conditional independence from the observation data, and the edges that do not satisfy the independence condition are gradually removed to determine the causal structure; then the structural equation model (SEM) parameters are optimized by maximum likelihood estimation to clearly define the functional relationship between variables (such as linear or nonlinear influence).

[0121] In the model application stage, the first / second cloud coverage change sequence (such as cloud coverage dynamic data at different time periods), the real-time operation data of the power plant (such as current power output, equipment status), and the real-time wind speed data of the weather station are input into the model to simulate the direct influence path (such as cloud coverage directly causing photovoltaic panel power to drop) and the indirect influence path (such as cloud movement accompanied by wind speed change, wind speed indirectly changes power by affecting heat dissipation or equipment stress) of the cloud movement state (such as movement speed, coverage area change) on the power generation. The power fluctuation of each path is calculated through the structural equation in the model, for example, the direct path can be represented as the power generation change amount equal to the cloud coverage change amount multiplied by the direct influence coefficient, and the indirect path needs to pass through the wind speed intermediate variable to transfer the influence.

[0122] To enhance the reliability of the results, multiple sets (e.g., 1000 sets) of cloud map evolution paths are generated through the Bootstrap resampling technique, each path corresponding to a different randomly sampled data subset, simulating the uncertainty of cloud movement. Based on a pre-set confidence interval (e.g., 95%), the power fluctuation range of each path is calculated, and the final optimized power fluctuation range is obtained through statistical methods (e.g., percentile method), reducing the random error of single model prediction.

[0123] For example, assuming that a photovoltaic power station in a certain area observes a first cloud cover change sequence showing that the northeast cloud layer moves towards the power station at a speed of 5 kilometers per hour, and the real-time wind speed is 4 meters per second, the historical data determines the causal structure of "cloud coverage rate → power generation" and "wind speed → power generation" through the PC algorithm. The structural equation model parameter estimation is that each 10% increase in cloud coverage rate directly leads to an 8% decrease in power, and each 1 meter per second increase in wind speed indirectly increases power by 1.5% through equipment heat dissipation effect. After inputting the current data, the model calculates a direct path power decrease of 4% (current cloud coverage rate increases by 5%), an indirect path power increase of 0.6% (wind speed increases by 0.4 meters per second compared to the baseline value), and a comprehensive net decrease of 3.4%. Through Bootstrap, 500 sets of cloud map evolution paths are generated, considering the fluctuation of cloud movement speed (±1 kilometer per hour) and the random disturbance of wind speed, and the power fluctuation range in the 95% confidence interval is optimized to 2.8%-4.1%.

[0124] In this way, through causal modeling, the causal mechanism between variables is clearly defined, avoiding the pseudo-causality problem of traditional correlation analysis, and the contribution of different influence paths can be explained. Combined with Bootstrap resampling to quantify uncertainty, the power fluctuation range prediction is more robust and suitable for complex scenarios with multiple variables (such as the coupling effect of cloud layer and wind speed). Compared with the scheme that only relies on statistical models, it has the advantage of being able to handle intervention problems (such as "if the cloud coverage rate increases by X, how will the power change"), providing more reliable causal reasoning basis for power station scheduling, and improving the accuracy and anti-interference ability of power prediction.

[0125] Further, the core advantage of using the Do-Calculus causal modeling method is that it can extract causal relationships between variables from observed data, rather than relying solely on correlation analysis, thereby more accurately quantifying the real impact of intervention measures (such as changes in cloud coverage rate) on the result variable (such as power generation).

[0126] Do-Calculus is based on the theory of causal graphs, which eliminates the interference of confounding variables through mathematical rules (such as backdoor adjustment, frontdoor adjustment), allowing the results distribution after "intervention" (do operation) to be derived through structured modeling without conducting randomized controlled trials. For example, when analyzing the causal effect of cloud coverage (X) on power generation (Y), the causal graph structure can be determined by identifying common causes (such as real-time wind speed Z) and constructing causal paths (X→Z→Y or X→Y) using PC algorithm to estimate conditional independence, then fitting the structural equation model parameters through maximum likelihood estimation, finally quantifying the direct or indirect effect of "cloud coverage change ΔX leading to power generation change ΔY", avoiding the estimation bias caused by confounding factors in simple statistical models.

[0127] Visualizing multiple causal paths (such as cloud movement directly blocking photovoltaic panels affecting power, or indirectly affecting equipment efficiency by changing temperature), by simulating the direct impact path (X→Y) and indirect impact path (X→M→Y, M is an intermediate variable such as temperature) of cloud movement state in the target area, the contribution of each path to power fluctuation is calculated respectively. For example, after inputting the sequence of cloud coverage and real-time wind speed data, the structural causal model can separate the direct effect of "cloud blocking → power drop" and the indirect effect of "cloud blocking → wind speed change → heat dissipation efficiency change → power fluctuation", avoiding attributing all associated signals to a single causal relationship, improving the interpretability and reliability of the model.

[0128] By generating multiple sets of cloud evolution paths through Bootstrap resampling, combined with confidence interval calculation of power fluctuation range, the uncertainty of causal estimation can be evaluated using limited observation data. For example, based on historical cloud movement trajectories and power data, 1000 sets of simulated paths are generated, the mean and confidence interval of power fluctuation under each path are calculated, and finally the prediction result is optimized through statistical aggregation, reducing the prediction bias caused by sample noise or abnormal values, making the estimation of power fluctuation range more robust.

[0129] Thus, compared with the limitation of traditional time series model that only captures data correlation, the Do-Calculus causal modeling method can answer the counterfactual question of "if the cloud coverage increases by 10%, how much will the power generation decrease", providing decision basis for power station scheduling from the causal level. For example, in the photovoltaic power station scenario, this method can identify high-risk shading areas in advance and quantify the power loss range under different cloud movement patterns, helping operation and maintenance personnel dynamically adjust energy storage strategies or start backup power sources, reduce the loss of abandoned electricity due to cloud shading, and improve the stability of the power grid and the efficiency of renewable energy consumption. In summary, the core advantage of Do-Calculus lies in revealing the mechanism between variables based on causal reasoning, quantifying the direction and strength of intervention effects, and enhancing the robustness of prediction through statistical methods, which is suitable for complex system modeling scenarios that require causal relationship explanation and decision support.

[0130] Further optionally, in the above steps, the first cloud coverage change sequence and / or the second cloud coverage change sequence, the real-time operation data of the power station, and the real-time wind speed data in the ground data of the meteorological station are input into a structural causal model to simulate the direct influence path and / or indirect influence path of the cloud movement state in the target area on the power generation, and to calculate the power fluctuation of each path, including:

[0131] The first partial derivative of the real-time power with respect to the cloud coverage rate and the second partial derivative of the real-time power with respect to the wind speed are obtained by solving the partial derivatives through multivariate linear regression using the historical cloud coverage rate in the first cloud coverage change sequence and / or the second cloud coverage change sequence, the real-time wind speed, and the real-time power data; the third partial derivative is obtained by calculating the irradiance loss parameter of the power station based on the photovoltaic component efficiency, the cloud reflectivity, the atmospheric layer external solar irradiance, and the cloud extinction coefficient in the first cloud coverage change sequence and / or the second cloud coverage change sequence; the fourth partial derivative is obtained by analyzing the contribution of the real-time wind speed to the real-time power according to the wind turbine characteristic curve fitting; the numerical sensitivity linear assumption error in the first partial derivative, the second partial derivative, the third partial derivative, and the fourth partial derivative is corrected by introducing the Taylor expansion high-order term; the direct influence path and / or indirect influence path of the cloud movement state in the target area on the power generation is generated through Monte Carlo simulation, and the probability distribution corresponding to each path is generated to obtain the power fluctuation interval of each path.

[0132] The core principle of the above steps is to quantify the multi-path effects of cloud movement and meteorological factors on power generation through causal path simulation combined with high-order error correction of multiple linear regression. Specifically, first, a multiple linear regression model is constructed using historical data (such as cloud coverage, wind speed, and power data every 10 minutes in a certain period). The direct impact of each variable is separated by solving the partial derivative: the first partial derivative represents the direct change rate of power when the cloud coverage changes by 1% (for example, if the result is -0.8 kW / %, it means that the power decreases by 0.8 kW for every 1% increase in coverage); the second partial derivative represents the direct contribution of wind speed to power (for example, the power changes by +1.2 kW for every 1 m / s increase in wind speed).

[0133] Next, for the characteristics of photovoltaic components, the irradiance loss is calculated in combination with parameters such as cloud reflectivity (e.g., 0.7), atmospheric layer extraterrestrial irradiance (e.g., 1367 W / m 2 ), etc. to obtain the third partial derivative. The quantitative relationship between cloud coverage and photovoltaic efficiency is quantified (for example, for every 0.1 increase in extinction coefficient, the power decreases by 3% due to irradiance loss). At the same time, based on the wind turbine power curve (such as a wind speed-power fitting polynomial), the fourth partial derivative is calculated to depict the indirect impact of wind speed on power through the mechanical characteristics of the wind turbine (for example, within the wind speed range of 5-10 m / s, the power indirectly increases by 0.5 kW for every 1 m / s increase).

[0134] Since the linear model assumes that each factor is independent and the impact is constant, in practice, high-order terms (such as quadratic terms) are introduced to correct non-linear errors (for example, when the cloud coverage exceeds 60%, the power decrease rate may accelerate, which requires adjustment using the second partial derivative). Finally, through Monte Carlo simulation, thousands of cloud movement paths (such as different combinations of moving speed and coverage range) are randomly generated, each path corresponds to a combination of partial derivatives, and after probability distribution calculation, the power fluctuation range is obtained (for example, under 95% confidence, the power decrease range of a certain path is 10-15 kW).

[0135] For example, in a certain photovoltaic power station scenario, the cloud coverage (30%-50% fluctuation), wind speed (4-6 m / s), and power (200-250 kW) data for the past 2 hours are input, and the first partial derivative -0.9 kW / % and the second partial derivative 1.0 kW / (m / s) are calculated. Combined with parameters such as component efficiency (0.18) and cloud reflectivity (0.6), the third partial derivative -0.05 kW / extinction coefficient unit is calculated, and the fourth partial derivative is fitted to 0.4 kW / (m / s) through the wind turbine curve. After introducing the second-order term correction, Monte Carlo simulation generates 100,000 cloud movement paths, and finally outputs the power fluctuation range under a certain typical path as 220±12 kW.

[0136] Thus, the direct and indirect influence paths are separated, avoiding single model confusion of causality (such as distinguishing the different effects of cloud layer direct blocking and wind speed changes on power). High-order term correction improves the prediction accuracy in nonlinear scenarios (such as simulating the sudden drop in power in strong convective weather when the cloud layer changes rapidly). Monte Carlo simulation provides probabilistic results to provide risk interval reference for power station scheduling (such as adjusting the output of energy storage systems in advance according to the 90% confidence fluctuation range). Combining multi-source data (cloud layer, wind speed, device parameters) to build a causality model with clear physical meaning, it is more interpretable than pure data-driven methods (such as explicitly showing the causal chain of "cloud layer movement → blocking increase → irradiance decrease → photovoltaic power reduction").

[0137] In the new energy light power correction system, dynamic power correction and grid parameter adjustment are key links to ensure stable power supply. The core logic is to dynamically adjust the grid operation parameters by predicting the impact of cloud movement on photovoltaic power generation in real time, so as to realize the balance between power supply and demand.

[0138] As an optional embodiment, in the step S107, based on the matching relationship and the power fluctuation range, the grid operation parameters of the target area are dynamically adjusted to ensure the stability of the grid. Specifically, first, from the input data fusion. Integrate the current power generation, future 5-minute cloud movement speed and historical power fluctuation rules to build a power prediction model. Cloud movement speed is an important reference index, the faster the cloud movement speed, the faster the cloud coverage changes, and the greater the predicted power drop. Next, based on the cloud movement trend, the power value that may be lost in the future short time (such as 5 minutes) is calculated, and the corrected power target value is generated. For example, if it is predicted that the cloud layer moves quickly, which may cause power to drop, the power loss amount is calculated in advance, and then the corrected power target is obtained. Finally, according to the corrected power target, the operation parameters of the grid are adjusted. In terms of voltage regulation, if the predicted power drops, the voltage of the transmission line will be reduced to reduce transmission line loss. In frequency control, the excitation current of the synchronous generator is adjusted to maintain the stability of the grid frequency. For example, when it is predicted that the power will drop from 100MW to 10MW in the next 5 minutes due to cloud blocking, the grid control system will reduce the voltage of part of the line to 95% of the normal level in advance, and adjust the generator output to ensure that the total power supply matches the demand.

[0139] The core logic of online gradient descent model updating is to dynamically optimize the parameters of the power correction model by comparing the deviation between the predicted results and the actual power in real time, thereby improving the prediction accuracy. The process starts with deviation calculation, which compares the corrected power prediction value with the actual measured value, calculates the error between the two, and the commonly used error calculation method is mean square error, which measures the accuracy of the prediction by calculating the average of the square of the difference between the predicted value and the actual value. Then, based on the calculated error, the gradient of the model parameters (such as neural network weights) is calculated through backpropagation, to determine the direction of parameter adjustment. The gradient reflects the rate of change of the error with respect to the parameter, and adjusting the parameter in the opposite direction of the gradient can gradually reduce the error.

[0140] During parameter adjustment, the momentum method or adaptive algorithm (such as Adam) is used to dynamically adjust the learning rate. In the high error stage, the learning rate is increased to speed up the model convergence; in the low error stage, the learning rate is reduced to avoid model oscillation. The specific parameter update formula is: the new parameter value is equal to the old parameter value minus the initial learning rate multiplied by the processed gradient estimate. The initial learning rate determines the step size of each parameter adjustment, and the first and second moment estimates of the gradient are used to adjust the step size more reasonably and make the parameter update more stable and efficient. In this way, the model is updated every 5 minutes, which can quickly adapt to the rapid changes of clouds, and the adaptive learning rate mechanism also avoids the oscillation or slow convergence problem caused by traditional fixed step size, greatly improving the real-time performance and robustness of the model. The core logic of reinforcement learning MPPT algorithm optimization is to train the agent through reinforcement learning and dynamically adjust the working point (such as voltage) of the photovoltaic string to reduce the power loss caused by shadowing.

[0141] First, define the state space, which includes current irradiance, temperature, string current / voltage, cloud coverage, and other environmental parameters that fully describe the working environment of the photovoltaic string. The action space defines discrete or continuous voltage adjustment instructions, such as adjusting the MPPT voltage from 500V to 510V. The design of the reward function aims to maximize power gain, and the reward value is equal to the power gain multiplied by a time-decaying coefficient. The power gain reflects the actual power gain after adjusting the working point, and the existence of the time-decaying factor encourages the agent to respond quickly and avoid missing the best adjustment opportunity due to excessive waiting. In terms of policy optimization, the PPO (Proximal Policy Optimization) algorithm is used, which can balance exploration (trying new voltages) and exploitation (selecting historical optimal voltages) to maximize long-term cumulative rewards. For example, when a string's power is reduced due to shadowing, the agent adjusts its MPPT voltage to try different working points, eventually bringing the working point close to the maximum power point. Test data shows that this method can reduce shadowing loss by 3%-5%, effectively improving photovoltaic power generation efficiency.

[0142] The entire new energy light power correction system realizes data closed loop. The actual power data after grid parameter adjustment is fed back to the power correction model to form a complete closed loop of "prediction-execution-feedback". This closed loop mechanism enables the system to continuously optimize the prediction and control strategy according to the actual situation. At the same time, the system performs multi-objective optimization, improves the power generation efficiency (reduces shadow loss through MPPT control), ensures the stability of the power grid (maintains stable operation of the power grid through voltage / frequency control), and avoids falling into a local optimal solution. From the actual test effect, in terms of power prediction accuracy, the 5-minute prediction error is reduced to within 2%. In terms of power generation capacity, the power generation capacity in the shadow area is increased by 3%-5%; in terms of power grid stability, the voltage fluctuation amplitude is controlled within ±1%.

[0143] Compared with the traditional method, in the power correction aspect, the traditional method is based on historical data for static prediction, which is difficult to adapt to the rapid change of cloud layer blocking. The embodiment of the application fuses the dynamic cloud layer moving speed and adopts online gradient update, which can track the change of cloud layer in real time and adjust the power prediction in time. In the MPPT control, the traditional fixed step perturbation and observation method is low in efficiency when facing shadow blocking; the scheme dynamically adjusts the working point through reinforcement learning, significantly reduces the shadow blocking loss, and improves the power generation efficiency. In the model update aspect, the traditional offline batch training cannot adapt to the environmental change in time. The embodiment of the application adopts online adaptive gradient descent, which can optimize the model parameters in real time, and ensure that the model always maintains high prediction accuracy. Therefore, the embodiment of the application realizes the whole link optimization from power prediction to power grid control, and significantly improves the operation stability and economy of the new energy station under complex weather conditions.

[0144] Further optionally, by fusing the PVsyst photovoltaic system simulation software and the GIS geographic information system, a three-dimensional terrain and obstacle shadow dynamic model is constructed to realize fine optimization of the inclination and spacing of the photovoltaic array. Specifically, based on the GIS terrain data (DEM) to obtain the ground elevation information, combined with the solar elevation angle time series data (the change of the sun position every minute), a shadow projection model of the obstacles (such as mountains and trees) is established. By calculating the geometric relationship of the obstacle height, horizontal distance and real-time solar elevation angle, the shadow range is dynamically simulated to change the trajectory with time, and the shadowing effect of the photovoltaic array in different time periods is accurately predicted.

[0145] Taking the maximization of power generation as the goal, the seasonal characteristics (the winter weight is higher than the summer) and the shading loss are comprehensively considered to define the optimization function. The genetic algorithm is used to globally search the optimal inclination (20° to 40°) and spacing combination, the constraint condition ensures that the spacing meets the unshading requirement under the minimum solar elevation angle of the winter solstice (the spacing is greater than or equal to the product of 3 times the array height and the sine value of the minimum solar elevation angle), and at the same time, the land area is avoided from being excessively increased.

[0146] Integrating real-time cloud layer movement prediction data, array parameters are updated every 15 minutes. By predicting cloud layer coverage changes, the array inclination and spacing are dynamically adjusted to ensure no shadow obstruction from 9:00 to 15:00 on the winter solstice, maximizing power generation efficiency. The above steps, combined with geographic information and weather forecasts, achieve adaptive optimization of photovoltaic systems under complex terrain and dynamic weather conditions.

[0147] In the embodiments of the present application, through the dense optical flow algorithm and the double-flow Transformer architecture, combined with the accurate analysis of the dynamic changes of the cloud layer from multiple sources, high-precision prediction of light power fluctuations is realized. Compared with traditional methods, it can more timely and accurately capture the power changes caused by weather mutations, and strive for sufficient response time for power grid dispatching. The prediction strategy and online data feedback mechanism of cloud layer types enable the system to adapt to various complex weather scenarios. Through power prediction and effective power grid regulation, the phenomenon of light abandonment caused by inaccurate prediction is reduced, the consumption capacity of new energy power generation is improved, the utilization rate of renewable energy is improved, and the efficient use of clean energy is promoted.

[0148] Please refer to Figure 2 , Figure 2 A new energy light power correction system based on weather mutation prediction is provided in the embodiments of the present application. The new energy light power correction system based on weather mutation prediction includes the following modules:

[0149] The acquisition module is configured to obtain satellite cloud images, meteorological data and ground data of a meteorological station of a target area where the power station is located.

[0150] The calculation module is configured to calculate a cloud layer movement vector field matched with the target area based on the satellite cloud images by using a dense optical flow algorithm. The cloud layer movement vector field is used to describe the vector distribution of the movement direction and speed of the cloud layer in the actual space, so as to reflect the overall movement trend and local turbulent change of the cloud layer.

[0151] The identification module is configured to identify the cloud layer types contained in the target area based on the cloud layer movement vector field, and obtain the coverage rate of the cloud layer types in the target area. The cloud layer types at least include one of the following: cumulonimbus, stratus and cirrus. If the coverage rate of the cumulonimbus exceeds a first threshold value, the cloud layer movement path is predicted based on the cloud layer movement vector field to obtain a first cloud layer coverage change sequence of the target area in a first time period. If the coverage rate of the stratus exceeds a second threshold value, the irradiance attenuation coefficient is predicted based on the cloud layer thickness and the solar elevation angle in the meteorological data, and the irradiance attenuation coefficient is corrected based on the ground aerosol optical depth obtained based on the ground data of the meteorological station to obtain a second cloud layer coverage change sequence of the target area in a second time period.

[0152] The prediction module is used to monitor the real-time operation data of the power plant, combine the first cloud cover change sequence and / or the second cloud cover change sequence, use the dual-stream Transformer architecture to predict the matching relationship between the cloud movement direction and the power drop area of ​​the power plant, and calculate the power fluctuation range corresponding to the power drop area of ​​each power plant.

[0153] The adjustment module is used to dynamically adjust the power grid operating parameters of the target area based on the matching relationship and the power fluctuation range to ensure power grid stability.

[0154] In some implementations, a new energy photovoltaic power correction system based on weather change prediction can be applied to terminal equipment.

[0155] It should be noted that, for the sake of convenience and brevity, the specific working process of the new energy photovoltaic power correction system based on weather change prediction described above can be referred to the corresponding process in the aforementioned embodiments of the new energy photovoltaic power correction method based on weather change prediction, and will not be repeated here.

[0156] Please see Figure 3 , Figure 3 This is a schematic block diagram illustrating the structure of a terminal device provided in an embodiment of this application. Figure 3 As shown, the terminal device 300 includes a processor 301 and a memory 302, which are connected via a bus 303, such as an I2C bus. Specifically, the processor 301 provides computing and control capabilities to support the operation of the entire terminal device. The processor 301 can be a central processing unit, or it can be other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. Specifically, the memory 302 can be a Flash chip, a read-only memory disk, an optical disc, a USB flash drive, or a portable hard drive, etc.

[0157] Those skilled in the art will understand that Figure 3The structure shown in the figure is only a block diagram of part of the structure related to the embodiment of the present application, and does not constitute a limitation on the terminal device to which the embodiment of the present application is applied. Specifically, the server can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. Among them, the processor is used to run the computer program stored in the memory, and when executing the computer program, it realizes any one of the new energy light power correction methods based on weather mutation prediction provided by the embodiments of the present application. It should be noted that those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the terminal device described above can refer to the foregoing new energy light power correction method based on weather mutation prediction, which will not be described here.

Claims

1. A new energy light power correction method based on weather mutation prediction, characterized in that, The method comprises the following steps: For the target area where the power station is located, satellite cloud images, meteorological data and ground data of meteorological stations in the target area are obtained; A cloud layer movement vector field matching the target area is calculated based on the satellite cloud images by using a dense optical flow algorithm; the cloud layer movement vector field is used to describe the vector distribution of the movement direction and speed of the cloud layer in the actual space, so as to reflect the overall movement trend and local turbulent change of the cloud layer; Based on the cloud layer movement vector field, the cloud layer types contained in the target area are identified, and the coverage rate of the cloud layer types in the target area is obtained; the cloud layer types at least include one of the following: cumulonimbus, stratus, cirrus; If the coverage rate of cumulonimbus exceeds a first threshold value, the cloud layer movement path is predicted in combination with the cloud layer movement vector field, and a first cloud layer coverage change sequence of the target area in a first time period is obtained; If the coverage rate of stratus exceeds a second threshold value, the irradiance attenuation coefficient is predicted based on the cloud layer thickness and the solar elevation angle in the meteorological data, and the irradiance attenuation coefficient is corrected in combination with the ground aerosol optical depth obtained based on the ground data of the meteorological stations, so as to obtain a second cloud layer coverage change sequence of the target area in a second time period; Real-time operation data of the power station is monitored, the matching relationship between the cloud layer movement direction and the power reduction area of the power station is predicted by using a double-flow Transformer architecture in combination with the first cloud layer coverage change sequence and / or the second cloud layer coverage change sequence, and the power fluctuation range corresponding to each power reduction area of the power station is calculated; Based on the matching relationship and the power fluctuation range, the operation parameters of the power grid in the target area are dynamically adjusted to ensure the stability of the power grid.

2. The method of claim 1, wherein, The method comprises the following steps: The multi-temporal satellite cloud images are aligned to the power station coordinate system through the geographic coordinate system, converted from RGB format to single-channel grayscale images, and the pixel value of the converted single-channel grayscale images is normalized; The Farneback algorithm is used to analyze the grayscale value change of the processed multi-temporal satellite cloud images, and a multi-resolution image sequence of grayscale values is generated by layer-by-layer downsampling; Starting from the top layer, the granularity optical flow of each pixel is calculated layer by layer, and the calculation result of the granularity optical flow of the previous layer is transmitted to the next layer as the initial value, until the original resolution is reached, and the multi-resolution optical flow calculation result is output as the first optical flow vector of each pixel; the higher the level, the coarser the granularity of the optical flow; the calculation result of the high-level granularity optical flow is used to represent the overall movement trend of the cloud layer, and the calculation result of the bottom-level granularity optical flow is used to represent the deformation of the cloud cluster edge; Based on the multi-resolution image sequence, a block integral image is constructed, and the block integral image is optimized by reverse search to obtain a sparse optical flow vector and a corresponding sparse optical flow block; Starting from the top layer, the coverage weight of all sparse optical flow blocks in the neighborhood of each pixel is calculated layer by layer, the sparse optical flow vector is weighted and averaged by the coverage weight to obtain a dense optical flow value, the dense optical flow value of the previous layer is transmitted to the next layer as the initial value, until the original resolution is reached, and the calculation result is output as the second optical flow vector of each pixel. The first optical flow vector and the second optical flow vector are fused to obtain a cloud movement vector field matched with the target region.

3. The method of claim 2, wherein, The granularity optical flow of each pixel is calculated, including: For each resolution image sequence, the gray distribution of the pixel neighborhood is modeled as a quadratic polynomial, and the symmetry parameter and the linear parameter in the quadratic polynomial are set by the gray constancy assumption; The Taylor expansion is used to derive the motion equation corresponding to the pixel neighborhood to obtain the motion vector value of each pixel as the granularity optical flow of each pixel.

4. The method of claim 1, wherein, Based on the cloud movement vector field, the cloud type contained in the target region is identified, and the coverage rate of the cloud type in the target region is obtained, including: The Otsu algorithm is used to dynamically determine the cloud gray threshold of the corrected cloud movement vector field, and the cloud region and the background region are adaptively threshold segmented based on the cloud gray threshold to obtain a cloud mask; the segmentation threshold corresponding to the high humidity region is dynamically adjusted in combination with the ground humidity data in the meteorological station ground data; The cloud mask is classified to obtain the cloud type contained in the target region, and the coverage rate of each cloud type in the target region is calculated.

5. The method of claim 4, wherein, Before the cloud mask is classified to obtain the cloud type contained in the target region, and the coverage rate of each cloud type in the target region is calculated, the Otsu algorithm is further used to dynamically determine the cloud gray threshold of the corrected cloud movement vector field, and the cloud region and the background region are adaptively threshold segmented based on the cloud gray threshold to obtain a cloud mask, including: In combination with the meteorological data and the meteorological station ground data, the wind speed data in the numerical weather prediction is obtained, and the cloud movement speed in the cloud movement vector field is corrected based on the wind speed value; The cloud thickness is inversely deduced through the cloud top height data, and the maximum displacement amount of the optical flow field is dynamically set based on the cloud thickness to constrain the cloud top height; The cloud voids in the cloud movement vector field are filled by closing operation to eliminate small-area noise in the cloud movement vector field.

6. The method of claim 4, wherein, The cloud mask is classified to obtain the cloud type contained in the target region, and the coverage rate of each cloud type in the target region is calculated, including: A real-time cloud classification model is constructed by using MobileNetV3; wherein the real-time cloud classification model at least includes: a convolution layer for extracting cloud image features, a channel attention module for strengthening cloud texture features, a Swish activation function for processing nonlinear expression, and a feature pyramid network for improving cloud cluster detection accuracy in multiple scales; The cloud mask is predicted by the real-time cloud classification model to obtain the prediction probability distribution of each cloud type in the cloud mask; Based on the prediction probability distribution, a single-type mask of each cloud type is generated, and the single-type mask is image enhanced according to the dynamic adjustment of the contrast of the cloud type; if the cloud type is cumulonimbus, the contrast is enhanced by 50%; if the cloud type is stratus, the contrast is reduced by 30%; The single-type coverage rate of each cloud type is obtained by using a pixel statistical method to calculate the ratio between all pixels in the single-type mask and the total pixels in the cloud mask; for a non-continuous cloud area, a continuous coverage area is generated by using a Kriging interpolation method, the pixel area proportion in the continuous coverage area is calculated, and the single-type coverage rate of each cloud type is optimized according to the pixel area proportion; The cloud mask is classified to obtain the cloud type contained in the target region, and the coverage rate of each cloud type in the target region is calculated, including: A real-time cloud classification model is constructed by using MobileNetV3; wherein the real-time cloud classification model at least includes: a convolution layer for extracting cloud image features, a channel attention module for strengthening cloud texture features, a Swish activation function for processing nonlinear expression, and a feature pyramid network for improving cloud cluster detection accuracy in multiple scales; The cloud mask is predicted by the real-time cloud classification model to obtain the prediction probability distribution of each cloud type in the cloud mask; Based on the prediction probability distribution, a single-type mask of each cloud type is generated, and the single-type mask is image enhanced according to the dynamic adjustment of the contrast of the cloud type; if the cloud type is cumulonimbus, the contrast is enhanced by 50%; if the cloud type is stratus, the contrast is reduced by 30%; The single-type coverage rate of each cloud type is obtained by using a pixel statistical method to calculate the ratio between all pixels in the single-type mask and the total pixels in the cloud mask; for a non-continuous cloud area, a continuous coverage area is generated by using a Kriging interpolation method, the pixel area proportion in the continuous coverage area is calculated, and the single-type coverage rate of each cloud type is optimized according to the pixel area proportion; For the pixels covered by multiple single-type masks, the single-type coverage of each cloud layer type at different positions in the target area and the single-type coverage rate are superimposed into the cloud layer mask according to confidence-weighted distribution, to obtain the coverage rate calculation result of the target area.

7. The method of claim 1, wherein, The first cloud cover change sequence and / or the second cloud cover change sequence are combined to predict the matching relationship between the cloud movement direction and the power reduction area of the power station by using a double-flow Transformer architecture, and to calculate the power fluctuation range corresponding to each power reduction area of the power station, including: The double-flow Transformer is used to generate cloud geometric distribution features and historical power fluctuation pattern features; wherein the first cloud cover change sequence and / or the second cloud cover change sequence are input into a spatial Transformer to extract cloud geometric features through geographical position coding and generate a cloud movement heat map; historical power data are input into a time-aware Transformer to capture power mutation point features, fluctuation frequency features and decay rate features in a sliding window within a preset time length in the future, to obtain historical power fluctuation pattern features; the cloud geometric features at least include: cloud movement direction vector, cloud center coordinates, edge contour and diffusion speed; Through a cross-modal attention mechanism, a cosine similarity matrix of the cloud movement direction vector and the power reduction area heat map in the cloud movement heat map is calculated, and an attention-guided matching relationship feature is obtained through Softmax weighted fusion; A gating mechanism is used to dynamically adjust the fusion weight to optimize the historical power fluctuation pattern features; Based on the matching relationship features and the historical power fluctuation pattern features, the power fluctuation range corresponding to each power reduction area of the power station is predicted through a fully connected layer.

8. The method of claim 1, wherein, After monitoring the real-time operation data of the power station, the method further includes: A Do-Calculus causal modeling method is used to construct a structural causal model; wherein the structural causal model is used to infer the causal effect of the cloud coverage rate change on the power generation; A PC algorithm is used to estimate the conditional independence, and a structural equation model parameter of the structural causal model is optimized through maximum likelihood estimation; The first cloud cover change sequence and / or the second cloud cover change sequence, the real-time operation data of the power station, and the real-time wind speed data in the ground data of the weather station are input into the structural causal model to simulate the direct influence path and / or indirect influence path of the cloud movement state on the power generation in the target area, and to calculate the power fluctuation of each path; A preset number of cloud image evolution paths are generated through Bootstrap resampling, and the power fluctuation range corresponding to each cloud image evolution path is calculated based on a pre-set confidence interval, to optimize the power fluctuation range corresponding to each power reduction area of the power station.

9. The method of claim 8, wherein, The first cloud cover change sequence and / or the second cloud cover change sequence, real-time operation data of the power plant, real-time wind speed data in the meteorological station ground data are input into a structural causal model to simulate a direct influence path and / or an indirect influence path of cloud movement in the target area on power generation, and power fluctuations of each path are calculated, including: A first partial derivative of real-time power to cloud cover and a second partial derivative of real-time power to wind speed are obtained by solving partial derivatives through multiple linear regression using historical cloud cover in the first cloud cover change sequence and / or the second cloud cover change sequence, real-time wind speed, and real-time power data; A third partial derivative is obtained by calculating irradiance loss parameters of the power plant based on photovoltaic component efficiency, cloud reflectivity, atmospheric layer external solar irradiance, and cloud extinction coefficient in the first cloud cover change sequence and / or the second cloud cover change sequence; A fourth partial derivative is obtained by analyzing the contribution of real-time wind speed to real-time power according to a wind turbine characteristic curve fitting; A Taylor expansion high-order term is introduced to correct the numerical sensitivity linear assumption error in the first partial derivative, the second partial derivative, the third partial derivative, and the fourth partial derivative, and a direct influence path and / or an indirect influence path of cloud movement in the target area on power generation is generated through Monte Carlo simulation, and a probability distribution corresponding to each path is generated to obtain a power fluctuation interval of each path.

10. A new energy light power correction system based on weather mutation prediction, characterized in that, The system comprises: A collection module configured to obtain satellite cloud images, meteorological data, and meteorological station ground data of a target area in which a power plant is located; A calculation module configured to calculate a cloud movement vector field matching the target area based on the satellite cloud images through a dense optical flow algorithm; the cloud movement vector field is used to describe the vector distribution of the movement direction and speed of the cloud in the actual space, so as to reflect the overall movement trend and local turbulent change of the cloud; An identification module configured to identify cloud types contained in the target area based on the cloud movement vector field, and obtain the coverage of the cloud types in the target area; the cloud types at least include one of cumulonimbus, stratus, and cirrus; if the coverage of the cumulonimbus exceeds a first threshold, a cloud movement path is predicted in combination with the cloud movement vector field to obtain a first cloud cover change sequence of the target area in a first time period; if the coverage of the stratus exceeds a second threshold, an irradiance attenuation coefficient is predicted based on the cloud thickness and the solar elevation angle in the meteorological data, and the irradiance attenuation coefficient is corrected in combination with the ground aerosol optical depth obtained based on the meteorological station ground data to obtain a second cloud cover change sequence of the target area in a second time period; A prediction module configured to monitor real-time operation data of the power plant, predict the matching relationship between the cloud movement direction and the power reduction area of the power plant in combination with the first cloud cover change sequence and / or the second cloud cover change sequence, and calculate the power fluctuation range corresponding to each power reduction area of the power plant; An adjustment module configured to dynamically adjust the grid operation parameters of the target area based on the matching relationship and the power fluctuation range to ensure the stability of the grid.

Images