A photovoltaic power station ultra-short-term power prediction method and system based on solar trajectory process propagation

Abstract

This invention provides a method and system for ultra-short-term power prediction of photovoltaic power plants based on solar trajectory process propagation, relating to the field of photovoltaic power generation prediction technology. The method includes: collecting real-time meteorological data, ground-based cloud map data, and historical operating power data of the photovoltaic power plant; calculating solar trajectory process parameters, including continuously changing solar altitude angle and solar azimuth angle, based on the geographical latitude and longitude of the photovoltaic power plant and the prediction time window; generating an initial power prediction curve using a basic prediction model; executing a correction process based on solar trajectory process propagation, constructing a solar ray projection channel model, calculating the additional projection velocity caused by the apparent solar motion by combining the rate of change of the solar trajectory process parameters, superimposing the geometric projection velocity of the cloud movement vector on the ground with the additional projection velocity to obtain a slip velocity vector, using the vector to calculate the propagation time delay parameters of cloud shadows reaching each subarray, and dynamically calibrating the time axis of the initial power prediction curve by translation.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, and in particular to a method and system for ultra-short-term power prediction of photovoltaic power plants based on the propagation of solar trajectory processes. Background Technology

[0002] The output power of photovoltaic power plants is highly volatile and intermittent due to the influence of meteorological environment. Cloud cover is the main factor causing drastic changes in photovoltaic power in the very short term.

[0003] To ensure the safe and stable operation of the power grid, existing technologies widely employ ultra-short-term power prediction technology based on ground-based all-sky imagers.

[0004] Existing technical solutions typically utilize image processing algorithms to identify cloud morphology and location, combine wind speed information to extrapolate cloud movement paths, and calculate the timing of cloud shadows shading photovoltaic modules based on geometric projection relationships. In these prediction models, the sun's position within a short prediction window is usually considered fixed, or projection calculations are performed solely based on the static solar altitude angle at a given moment, directly assuming a fixed proportional relationship between the speed of cloud shadow movement on the ground and the horizontal movement speed of the cloud in the sky. However, in actual physical processes, the sun is constantly in a state of continuous apparent motion, with its altitude and azimuth angles changing over time. Especially in the morning and evening, the solar altitude angle is low and its rate of change is large. The dynamic change in the sun's position produces a significant optical leverage effect during ground projection, causing a nonlinear deviation between the actual sliding speed of the shadow and the physical movement speed of the cloud. Existing technologies ignore this dynamic process, resulting in phase errors between the estimated power surge moments and the actual occurrence moments. Furthermore, photovoltaic power plants cover vast areas, and it takes time for the shadow to sweep across the entire site; simple time-series predictions cannot accurately reflect the spatial propagation sequence of the shadow wavefront between different subarrays. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for ultra-short-term power prediction of photovoltaic power plants based on the propagation of solar trajectory processes, so as to solve the problems pointed out in the background art.

[0006] In a first aspect, the present invention provides a method for ultra-short-term power prediction of photovoltaic power plants based on the propagation of solar trajectory processes, the method comprising the following steps: Collect real-time meteorological data, ground-based cloud map data, and historical operating power data of photovoltaic power plants; Based on the geographical latitude and longitude of the photovoltaic power station and the predicted time window, the solar trajectory process parameters are calculated, including the solar altitude angle and solar azimuth angle that change continuously over time. Using the historical operating power data and the real-time meteorological data, an initial power prediction curve is generated through a basic prediction model; The method also includes performing a correction process based on propagation of the solar trajectory process: Identify cloud layer movement vectors based on the aforementioned ground-based cloud map data; A model of the sunlight projection channel is constructed, and the cloud movement vector is mapped to the ground projection plane where the photovoltaic power station is located. The sliding velocity vector of the cloud shadow propagating on the ground is calculated by combining the rate of change of the solar trajectory process parameters. The propagation time delay parameters of the cloud shadow reaching each sub-array of the photovoltaic power station are calculated using the slip velocity vector; The time axis of the initial power prediction curve is dynamically shifted and calibrated based on the propagation delay parameter, and the final ultra-short-term power prediction result is output.

[0007] Optionally, the construction of the solar radiation projection channel model and the calculation of the slip velocity vector of cloud shadows propagating on the ground specifically include: Extract the differential value of the solar altitude angle with respect to time to obtain the solar altitude angular velocity; Establish a geometric correlation function between cloud height, solar altitude angle, and ground projection distance; The cloud movement vector is decomposed into tangential and normal components. Based on the geometric correlation function and the solar altitude angular velocity, the additional projected velocity caused by the solar motion is calculated. The slip velocity vector is obtained by vector superimposing the geometric projection velocity of the cloud movement vector onto the ground with the additional projection velocity.

[0008] Optionally, the calculation of the propagation time delay parameters of the cloud shadow reaching each subarray of the photovoltaic power station using the slip velocity vector specifically includes: A virtual spatial topology map of a photovoltaic power station is constructed, wherein the center coordinates of each subarray are marked in the topology map; The subarrays on the windward side are defined as outpost nodes, and the subarrays on the leeward side are defined as follower nodes; Based on the slip velocity vector, calculate the physical transit time of the cloud shadow vanguard from the outpost node to the follower node; The physical transit time is used as the propagation delay parameter.

[0009] Optionally, the step of dynamically shifting and calibrating the time axis of the initial power prediction curve based on the propagation delay parameter specifically includes: Detect whether the measured power of the sentinel node undergoes a sudden change; If a mutation occurs, extract the mutation time and mutation slope; Using the propagation delay parameter, the predicted time when the power change of the following node occurs can be calculated; The falling edge of the predicted power corresponding to the following node in the initial power prediction curve is shifted to the prediction time.

[0010] Optionally, the calculation of the slip velocity vector of the cloud shadow propagating on the ground also includes compensation for the lensing effect at the cloud edge: Calculate the angle between the angle of incidence of sunlight and the normal to the edge of the cloud; When the included angle is within the preset lens enhancement range, the sliding velocity vector is decelerated and corrected to simulate the blurring delay of the shadow boundary caused by the refraction of light at the edge of the cloud.

[0011] Optionally, the basic prediction model is a long short-term memory neural network or a gated recurrent unit, and the time resolution of the initial power prediction curve is 1 minute to 15 minutes.

[0012] Optionally, the method further includes an abnormal operating condition rejection step: Before calculating the slip velocity vector, it is determined whether the instantaneous wind speed in the real-time meteorological data exceeds a preset threshold; If the preset threshold is exceeded, the cloud formation is determined to be in a state of rapid fragmentation, the correction process is stopped, and the initial power prediction curve is directly output.

[0013] Optionally, the construction of the virtual spatial topology map of the photovoltaic power station further includes: Obtain a digital elevation model of the site where the photovoltaic power station is located; Based on the terrain slope and aspect, the equivalent light-receiving coordinates of each subarray in the virtual space topology map are corrected.

[0014] Optionally, after outputting the final ultra-short-term power prediction result, a closed-loop feedback step is also included: Collect the measured power data at the next moment; Calculate the deviation between the measured power change time and the predicted time; The deviation value is used to reverse the cloud height parameter in the solar projection channel model.

[0015] Secondly, the present invention provides a photovoltaic power plant ultra-short-term power prediction system based on solar trajectory process propagation, the system comprising: The data acquisition module is configured to collect real-time meteorological data, ground cloud map data, and historical operating power data of the photovoltaic power station; The trajectory calculation module is configured to calculate solar trajectory process parameters based on the geographical latitude and longitude of the photovoltaic power station and the prediction time window; The basic forecasting module is configured to generate an initial power forecast curve using the historical operating power data and the real-time meteorological data; The propagation correction module is configured to construct a solar ray projection channel model, calculate the slip velocity vector of cloud shadow propagation on the ground by combining the rate of change of the solar trajectory process parameters, and calculate the propagation time delay parameters based on the slip velocity vector; The result output module is configured to dynamically shift and calibrate the time axis of the initial power prediction curve according to the propagation delay parameter, and output the final ultra-short-term power prediction result.

[0016] The present invention has achieved the following beneficial effects: This invention constructs a solar projection channel model and introduces the rate of change of solar trajectory process parameters. When calculating the ground propagation speed of cloud shadows, it not only considers the physical movement of the cloud layer itself but also quantitatively superimposes the additional projection velocity caused by the apparent motion of the sun. This corrects the velocity vector deviation caused by neglecting the motion of the light source in existing technologies, making the calculated cloud shadow slip velocity consistent with the actual atmospheric optical geometry motion. This invention utilizes the virtual spatial topology of a photovoltaic power station to establish the spatial association between outpost nodes and follower nodes. The calculated slip velocity vector is transformed into physical time delay parameters for the propagation of cloud shadows between different subarrays. Based on this, the initial power prediction curve generated by the basic prediction model is dynamically shifted and calibrated on the time axis, effectively solving the problem of lag or lead in the prediction results in time phase. Furthermore, this method integrates compensation mechanisms for cloud edge lensing effects and terrain slope, restoring the power surge and waveform distortion characteristics during cloud obstruction. This makes the final output ultra-short-term power prediction results closer to the actual operating state of the power station in terms of trend and occurrence time.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A schematic diagram of the composition structure of a photovoltaic power plant ultra-short-term power prediction system based on solar trajectory process propagation provided in an embodiment of the present invention; Figure 2A flowchart illustrating the ultra-short-term power prediction method for photovoltaic power plants based on solar trajectory process propagation provided in an embodiment of the present invention; Figure 3 A schematic diagram of the device architecture of a photovoltaic power plant ultra-short-term power system based on solar trajectory process propagation provided in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] The present invention first provides a photovoltaic power plant ultra-short-term power prediction system based on the propagation of solar trajectory processes.

[0022] Reference manual attached Figure 1 The system mainly consists of three parts: an environmental perception subsystem, a data transmission and edge preprocessing subsystem, and a core power prediction calculation platform.

[0023] The environmental perception subsystem is responsible for the full-dimensional digital mapping and real-time capture of the micro-meteorological environment of the photovoltaic power station and the area above it from 0 to 5000 meters.

[0024] First, regarding the ground-based cloud image acquisition unit: This system deploys an industrial-grade Total Sky Imager (TSI) at the geometric center of the photovoltaic power station and at the upwind edge of the prevailing wind direction (e.g., at the northwest corner boundary of the station if local historical meteorological data indicates a northwesterly wind). This imager integrates specialized meteorological observation equipment with sophisticated opto-mechanical-electronic technology. Its optical front end uses a 180-degree field-of-view fisheye lens with an extremely short focal length (typically set to 1.4mm-1.8mm) to ensure comprehensive capture of cloud distribution across the entire hemisphere. To cope with the intense ultraviolet radiation, high temperatures, and sandstorms of the photovoltaic power station, the lens surface is coated with multiple layers of anti-reflective coating and an anti-ultraviolet-infrared hardened filter. For the image sensor, a high dynamic range (HDR) industrial-grade CMOS sensor is selected, with a resolution of at least 2000×2000 pixels and a bit depth of no less than 12 bits, ensuring clear differentiation of the texture details of thin cirrus clouds and the thickness differences of stratocumulus clouds even in high-contrast areas around direct sunlight.

[0025] Furthermore, to eliminate the destructive interference of direct sunlight on image quality (such as vertical light column trailing and sensor pixel burn-out), an automatic shading device based on high-precision GPS positioning and astronomical calendar control is integrated on the top of the imager. This device incorporates a high-torque stepper motor and an absolute encoder, which, based on real-time calculated solar altitude and azimuth angles, drives a specially treated black physical shading strip to precisely block the sun between the lens's optical center and the image, thus forming a regular, precisely positioned black shading strip on the image. This not only protects the precision sensor but also provides a precise geometric reference for distortion correction and cloud segmentation in subsequent image processing. In addition, the imager integrates a constant temperature control module that automatically activates the lens heating and defogging function when the ambient humidity exceeds 80% or the temperature is below the dew point, ensuring clear imaging in all weather conditions.

[0026] Secondly, regarding the high-precision radiation observation unit: To quantify the visual characteristics of clouds into their actual physical shading effect on photovoltaic modules, accurate ground radiation data must be obtained as calibration values. This system has at least four micro-weather stations evenly deployed within the site according to the terrain. Each station is equipped with a Class I total radiation meter conforming to ISO 9060 standards for measuring horizontal total radiation (GHI), and a diffuse radiation meter with an automatic shading ring for measuring diffuse radiation (DHI). To accurately retrieve cloud optical depth (COD), the system also includes a direct radiation meter, working in conjunction with a dual-axis automatic solar tracker to collect direct normal radiation (DNI) in real time. These radiation sensors continuously sample at a frequency of 1Hz using a high-precision 24-bit A / D converter, with measurement uncertainty strictly controlled within 2%, and the equipment is periodically calibrated using a standard source.

[0027] Secondly, regarding the dynamic meteorological monitoring unit: The movement, deformation, and dissipation of clouds are controlled by the dynamic characteristics of the atmospheric wind field. To accurately capture the driving force behind cloud movement, this system has installed a wind measurement tower at the station and three-dimensional ultrasonic anemometers and wind vanes at different altitudes of 10 meters, 30 meters, and 50 meters. Unlike traditional mechanical cup anemometers, ultrasonic anemometers have no inertial components and can capture high-frequency gusts and vertical wind shear characteristics, which is crucial for subsequently determining the degree of atmospheric turbulence and correcting the movement speed of upper-level clouds.

[0028] The massive amounts of raw data collected by the perception layer need to be transmitted to the computing center through a stable, low-latency, high-speed channel. This embodiment uses an industrial-grade fiber optic ring network as the backbone communication link within the site, supplemented by 5G industrial modules as backup links, ensuring lossless data transmission even in environments with strong electromagnetic interference. Before the data enters the core prediction algorithm, a preprocessing subsystem is deployed in the edge computing gateway to perform data cleaning and standardization tasks.

[0029] For meteorological data, the system has built-in range detection algorithms based on physical extreme values ​​(e.g., logical verification that nighttime irradiance must be zero and relative humidity must not exceed 100%) and jump detection algorithms based on time consistency (e.g., temperature change within 1 second must not exceed 0.5℃).

[0030] For cloud imagery data, preprocessing is particularly crucial. The edge gateway incorporates a distortion correction matrix, pre-generated based on a polynomial calibration model of a fisheye lens. The original circular fisheye image undergoes matrix transformation, unfolding into an equidistant cylindrical or orthographic projection image that conforms to human visual perception, eliminating severe stretching distortion in edge regions. Furthermore, for the black bands left on the image by automatic shading, the system employs a neighborhood pixel-based texture synthesis and filling algorithm for repair, or generates corresponding invalid masks in subsequent algorithms to remove them, ensuring the accuracy of cloud cover calculations. Simultaneously, to ensure time synchronization of multi-source data, the edge gateway uses a combination of NTP (Network Time Protocol) and PTP (Precise Time Protocol) to align the acquisition timestamps of all data to the millisecond level.

[0031] The core power prediction computing platform is typically deployed in a dedicated server rack within the substation or runs on a high-performance computing cluster in a scheduling cloud. This platform is equipped with GPU accelerator cards optimized for vector operations (such as the NVIDIA Tesla series) for parallel processing of demanding image optical flow calculations and deep learning inference tasks. Internally, the platform runs the photovoltaic power plant ultra-short-term power prediction software based on solar trajectory process propagation, as described in this invention. This software adopts a modular microservice architecture design, including a data management module, an astronomical calculation module, an optical flow analysis module, a physical projection module, and a power inversion module. These modules communicate efficiently with each other through shared memory or message queues.

[0032] This invention provides a method for ultra-short-term power prediction of photovoltaic power plants based on the aforementioned hardware system. (See attached specification.) Figure 2 The execution flow of this method is as follows: Step 1: Full-dimensional data collection and anomaly cleaning.

[0033] The approach begins with high-quality data input. The system synchronously aligns and reads the following three core data streams with a time resolution of minutes (e.g., 1 minute or 5 minutes, depending on grid dispatch requirements): Real-time meteorological data: In addition to conventional temperature, humidity, and pressure data, the focus is on wind vector data at different altitudes. The system constructs a near-surface wind shear model by fitting logarithmic or exponential law data to wind speed data at 10m, 30m, and 50m, thereby extrapolating and estimating the wind speed vector at cloud base height (typically between 1000m and 3000m). The mathematical model for estimating wind speed at cloud height is as follows: ; in: This represents the calculated wind speed at cloud height (unit: meters per second). This indicates the measured wind speed at a reference height (e.g., 50m) (unit: meters per second). Indicates the height of the cloud base (unit: meters); Indicates reference altitude (unit: meters); This represents the wind shear index, which is selected based on the local surface roughness.

[0034] This formula provides a dynamic reference boundary for subsequent verification of cloud movement vectors.

[0035] Ground-based cloud image data: The system acquires a sequence of cloud images after geometric correction. To adapt to different lighting conditions under varying weather conditions, the system performs adaptive histogram equalization or adaptive gamma correction on the images to enhance the contrast between cloud edges and the blue sky background. Subsequently, threshold segmentation is performed using the red-blue channel ratio method (RBR = R channel pixel value / B channel pixel value). Because clouds scatter red and blue light with relatively small differences in the visible light band (appearing white or gray), while clear skies primarily scatter blue light dominated by Rayleigh scattering, the RBR value of cloud pixels is significantly higher than that of the sky background. The system automatically determines the optimal segmentation threshold using the Otsu method to generate a binarized cloud detection mask.

[0036] Historical operating power data: The system extracts power curves for the entire station and each subarray (inverter level or combiner box level) from the SCADA historical database. To eliminate the contamination of training samples by human-induced curtailment commands, the system simultaneously reads the inverter's status word and AGC (Automatic Generation Control) commands. Data from periods of power curtailment, equipment maintenance, or communication interruptions are marked as invalid samples and excluded from subsequent training and correction of the basic model. Furthermore, a sliding median filter is used to smooth electromagnetic interference noise in the data.

[0037] Step 2: Fine calculation of solar trajectory process parameters.

[0038] Traditional methods often ignore changes in the sun's position over very short periods, assuming the light source is stationary. However, when predicting power over the next four hours, significant changes in the sun's position can drastically alter the geometry of shadow projection.

[0039] In this step, the system incorporates a high-precision Solar Position Algorithm (SPA), which is based on the NREL (National Renewable Energy Laboratory) standard and takes into account Earth's nutation, phototropic aberration, and atmospheric refraction corrections, achieving an accuracy of 0.0003 degrees.

[0040] First, the system reads the precise geographical coordinates of the photovoltaic power station, defined as: Local geographical latitude (unit: radians); Local geographical longitude (unit: radians); Local altitude (unit: meters).

[0041] Secondly, determine the time window sequence for prediction. ,in, Indicates the start time of the prediction. To predict the total number of time points, To predict the step size.

[0042] For each time step in the sequence The system calculates two core parameters in the horizontal coordinate system: Solar elevation angle : The angle between the center of the sun's rays and the local ideal ground plane.

[0043] Solar Azimuth Angle : The angle between the projection of sunlight onto the ground plane and the north direction (or the south direction, depending on the coordinate system definition).

[0044] This invention not only calculates static angles, but also further calculates the rate of change of these angles over time, i.e., process parameters. The system calculates these parameters using numerical differentiation methods or by directly utilizing the derivative output of the SPA algorithm: Solar altitude angular velocity ( Defined as The unit is radians per second; where, Describes the differential operator. Indicates time, Represents an infinitesimal increment in time; Solar azimuth angular velocity ( Defined as The unit is radians per second.

[0045] These two derivative parameters have crucial physical significance in subsequent calculations of shadow slip velocity. For example, at noon, the solar altitude angle changes slowly ( However, during sunrise and sunset, The value is relatively large. Without incorporating this rate of change, the prediction of cloud shadow speed during the morning and evening hours will produce significant errors.

[0046] Step 3: Build the basic prediction model and generate the initial curve.

[0047] Before making complex physical corrections, a baseline needs to be established. This step utilizes deep learning technology to construct a basic model that reflects the power output characteristics of a photovoltaic power plant under average climatic conditions.

[0048] The model preferably employs a Long Short-Term Memory (LSTM) neural network or a gated recurrent unit (GRU) architecture. The reason for choosing such recurrent neural networks (RNNs) is that photovoltaic power data exhibits significant time-series autocorrelation (i.e., the power at the current moment is strongly correlated with the power at past moments).

[0049] The input layer feature vectors of the model include: Historical power series ; Historical irradiance sequence ;in, Indicates the current prediction time; This represents the time series backtracking step size (i.e., the historical window size) of the input model. This represents the measured power at the previous moment. Indicates pushing forward the first The measured power at each moment; and These represent the measured irradiance data at the corresponding times.

[0050] Time feature encoding (hours, seasonal index, periodic encoding through sine / cosine transform). Theoretical power in clear sky is an important physical constraint characteristic.

[0051] Among them, the theoretical power of clear sky Based on clear sky horizontal total irradiance The conversion is performed. (i.e., total horizontal irradiance under clear sky conditions) was calculated using the Ineichen-Perez model: in: This is the Earth-Sun distance correction factor; It is the solar constant; For air quality; For Rayleigh optical thickness; It is an atmospheric turbidity factor; This is the solar zenith angle.

[0052] The model outputs the initial power prediction curve within the future prediction window through nonlinear transformation of multiple layers of neurons. .

[0053] It should be particularly pointed out that... While curves can accurately fit the diurnal variation trend of power (such as a parabolic profile), neural networks, by their very nature, fit historical statistical patterns, making them ill-suited for accurately predicting sudden, rapidly moving fragmented cumulus cloud obstruction events. In cloudy or overcast conditions, the curves output by LSTMs are often too smooth, losing details of high-frequency fluctuations. Therefore, a subsequent physical correction process must be introduced to inject minute-level cloud shadow fluctuation information.

[0054] Step 4: Physical correction process based on solar trajectory propagation.

[0055] The system utilizes the ground-based cloud image sequence obtained in step one to track cloud movement using either the sparse optical flow (Lucas-Kanade) method or the dense optical flow (Farneback) method. In the algorithm implementation, strong corner points or texture-rich regions in the cloud mask image are first extracted as feature points. Then, the optimal matching positions of these feature points are found between two consecutive image frames, and their displacement vectors in the image pixel coordinate system are calculated. To convert pixel displacement into physical velocity, the cloud base height (CBH) parameter needs to be introduced. This parameter can be obtained through on-the-ground laser altimeter measurements. If no on-the-ground measuring equipment is available, the system utilizes the temperature readings from a ground-based weather station. (Celsius) and dew point temperature (Celsius), through empirical formula Make an estimate.

[0056] Using the fisheye lens projection model of the all-sky imager (e.g.) ,in, This represents the radial distance of a cloud pixel from the optical center on the image plane. Indicates the effective focal length of the fisheye lens. This represents the angle of incidence of the corresponding spatial point relative to the camera's optical axis (i.e., the zenith angle), combined with the cloud base height. The system inverts the pixel displacement vector into the horizontal physical velocity vector of the cloud layer in the sky. This vector contains two components: Physical speed of cloud movement (meters per second); The physical azimuth (in radians) of cloud movement.

[0057] Furthermore, existing technologies typically assume that the speed at which cloud shadows move on the ground is equal to the speed at which the clouds move in the sky. However, this assumption only holds true in the extremely ideal case where the sun is at its zenith (vertical incidence) and stationary. In practical applications, the oblique projection effect of sunlight and the apparent motion of the sun itself must be considered.

[0058] Therefore, this invention constructs a three-dimensional optical path geometric model. In this model, the center point of the cloud shadow... The position vector is the center point of the cloud layer. With the solar position vector The function.

[0059] Based on geometric derivation, on a flat surface, the horizontal offset distance of a cloud shadow relative to the vertical projection point of the cloud layer is... Satisfy the following formula: ; in: Indicates the horizontal distance (in meters) between the center of the cloud shadow and the vertical projection point of the cloud layer. Indicates the height of the cloud base (unit: meters); Represents the solar altitude angle (unit: radians).

[0060] This formula shows that the projected distance is proportional to the cotangent of the solar altitude angle.

[0061] To calculate the movement speed of the cloud shadow (i.e., the derivative of its position with respect to time), we need to perform a time-dependent transformation on the above positional relationship. The total differential of .

[0062] The slide velocity vector of cloud shadows propagating on the ground. It actually consists of two parts, and its calculation formula is: ; in: This represents the total sliding velocity vector of the cloud shadow on the ground; This represents the geometric projection velocity vector caused solely by the movement of the cloud body itself; This represents the additional projected velocity vector caused solely by the apparent motion of the sun.

[0063] For Part 1 Under the assumption of a flat surface, its magnitude is approximately equal to the cloud velocity. However, the direction needs to be projected and transformed according to the sun's azimuth angle.

[0064] For Part Two This is the additional projected velocity caused by the sun's motion. (This is derived from the distance formula.) By taking the derivative, we can obtain the magnitude of the radial component of this additional velocity. : ; in: Represents the radial component of the shadow velocity caused by the sun's motion (unit: meters per second); The cosecant function representing the solar altitude angle, i.e. ; It represents the rate of change of the solar altitude angle over time (angular velocity, unit: radians / second).

[0065] Please note the formula. Term (square of cosecant). When the solar altitude angle When the temperature is relatively low (e.g., below 30 degrees Celsius), The value will increase rapidly. It increases exponentially. This means that even extremely small changes in the solar altitude angle in the morning or evening ( This optical leverage factor also amplifies the shadow cast by the ground on the photovoltaic panel, causing the shadow to pass through the panel at a much faster speed than the clouds. Traditional methods miss this. The component factor causes a larger error in traditional predictions for morning and evening time periods.

[0066] To achieve accurate computational calculations, the system establishes a local two-dimensional Cartesian coordinate system with due north as the Y-axis and due east as the X-axis. The azimuth angle of cloud movement is defined. and solar azimuth All of these are angles rotated clockwise with true north as 0 degrees.

[0067] First, the physical velocity vector of the cloud layer Decomposed into coordinate components: ; ; in: This represents the component of the physical movement speed of clouds on the X-axis (due east direction), with the unit being meters per second (m / s). This represents the component of the physical movement speed of clouds along the Y-axis (due north direction), with units of meters per second (m / s). This represents the physical movement rate of the cloud layer calculated in step four, in meters per second (m / s). This indicates the azimuth angle of cloud movement, in radians (rad). These are trigonometric functions of sine and cosine, respectively.

[0068] Secondly, the additional velocity vector components caused by the sun's motion are calculated. Because... The shadow of the cloud represents the direction away from the sun (i.e., the angle). The radial scaling factor of ) has the following projection component: ; ; in: This represents the component of the additional velocity caused by the apparent motion of the sun on the X-axis, in meters per second (m / s). This represents the component of the additional velocity caused by the apparent motion of the sun on the Y-axis, in meters per second (m / s). This represents the radial component of the shadow velocity caused by the solar motion calculated by the aforementioned formula, in meters per second (m / s). This indicates the solar azimuth at the current moment, in radians (rad). Pi is a constant value, approximately 3.14159.

[0069] Finally, the total sliding velocity vector of the cloud shadow on the ground is synthesized. Components and modulus: ; ; ; in: The X-axis component represents the total sliding velocity vector of the cloud shadow on the ground, in meters per second (m / s). The Y-axis component represents the total sliding velocity vector of the cloud shadow on the ground, in meters per second (m / s). The modulus represents the total sliding velocity of the cloud shadow on the ground, used for subsequent calculations of propagation delay, and is expressed in meters per second (m / s).

[0070] This invention utilizes real-time calculation and This additional velocity vector was precisely calculated and then vector-combined with the cloud projection velocity to obtain the cloud shadow slip velocity that conforms to the laws of geometric optics. This calculation process will be executed cyclically in the system's physics engine module at a speed of milliseconds.

[0071] After calculating the slip velocity vector, the system pre-loads a virtual spatial topology map of the photovoltaic power station into memory. This map is constructed based on the power station's CAD as-built drawings and precisely marks the physical center coordinates of hundreds of photovoltaic sub-arrays (usually represented by inverters or transformer sub-arrays) throughout the station. .

[0072] Based on the slip velocity vector calculated at the current moment The system dynamically defines the roles of these subarrays based on their orientation. Outpost nodes: defined as the subarray located at the forefront (windward / sunward side) of the direction from which the slip velocity vector originates. They are the first units in the entire station to detect cloud shadow intrusion. The specific selection algorithm is as follows: Construct the set of physical center coordinates of the entire sub-array .

[0073] Define the normalized reverse vector of cloud shadow glide velocity. Its components are calculated as follows: ; ; Calculate the coordinates of each subarray in the vector Projection value in direction : ; Select The subarray with the largest value is used as the sentinel node, which is the first contact point of the cloud shadow wavefront in spatial geometry.

[0074] in: This indicates the total number of subarrays within a photovoltaic power plant; This represents the index number of the subarray, with a value ranging from 1 to N; Indicates the first The horizontal and vertical coordinates of the subarrays in the local Cartesian coordinate system are expressed in meters (m). These represent the X-axis and Y-axis components of the unit vector in the opposite direction of the cloud shadow's sliding velocity, respectively, and are dimensionless values. All of these are sliding velocity-related parameters calculated in the aforementioned steps; Indicates the first The projection distance scalar of each subarray in the direction of the cloud shadow source, in meters (m).

[0075] Follower node: Defined as a subarray located downstream of the sentinel node.

[0076] The system calculates the propagation of the cloud shadow wavefront from the outpost node to every follower node. physical distance This distance is the difference in projected distances between the position vectors of the two nodes along the direction of their sliding velocity.

[0077] Then, the propagation time delay parameter (Propagation Time Lag) is calculated using kinematic formulas. ): ; in: Indicates that the cloud shadow has reached the first Lag time of each follower node (in seconds); Represents the sentinel node and the follower node Projected distance in the direction of cloud shadow movement (unit: meters); The magnitude of the cloud shadow's sliding velocity vector (unit: meters per second).

[0078] This set of parameters The time structure of the cloud shadow sweeping across the entire power station was precisely quantified, among which, This indicates the number of following nodes in a photovoltaic power station. Indicates that the cloud shadow has reached the first The lag time of each following node. For a power plant covering several kilometers, the maximum time lag may reach 5-10 minutes, which is a huge available time window for ultra-short-term forecasting.

[0079] Finally, the system feeds back the physical calculation results onto the power curve.

[0080] The system monitors the measured power of the sentinel nodes in real time. Once a significant negative abrupt change in sentinel power is detected (e.g., the slope of the power change exceeds a preset threshold), the system will take action. (And the duration matches the characteristics of cloud obstruction), the system immediately locks the trigger time of the event. and the magnitude of the drop .

[0081] Based on the calculated time delay matrix, the system infers that the cloud shadow will... Time reaches the One follower node.

[0082] Therefore, the system generates a basic model targeting the first... The initial prediction curves of each node are reshaped and corrected. Time axis calibration: Forcefully shift and align the power drop edge in the predicted curve to the desired time axis. .

[0083] Amplitude calibration: Based on the measured attenuation rate of the leading node (reflecting the optical thickness (COD) of the cloud layer), the predicted power valley value of the following node is corrected. The system generates the first... The final corrected power curve of each follower node : ; in: Indicates the first Each follower node The final power prediction after time correction, in kilowatts (kW) or megawatts (MW). The time variable representing the prediction; This represents the output of the base model (such as LSTM) described in step three, for the first... Initial power prediction function for each node; It represents the power attenuation coefficient calculated based on the measured data of the outpost (i.e., the ratio of the measured power drop at the outpost to the clear-sky power at that moment, with a value ranging from 0 to 1). This indicates the time shift, measured in seconds (s).

[0084] The time shift The calculation formula is: ; in: This indicates the measured trigger time at which the system detects a power surge at the sentinel node; This indicates that the cloud shadow calculated above propagates to the first... The physical time delay parameter of each follower node, in seconds (s); This indicates the moment when the power valley appears in the initial prediction curve output by the base model.

[0085] Through this relay prediction mechanism that combines outpost and follow-up, the present invention effectively extrapolates the measured information of a single point in the spatiotemporal dimension, ensuring that the total power prediction results of the entire station are highly consistent in phase and waveform.

[0086] Step 5: Dynamic compensation of cloud edge lensing effect based on Mie scattering physics mechanism.

[0087] During the 30 to 90-second window before the shadow front of a thick or dense cumulus cloud officially reaches the photovoltaic module, the instantaneous irradiance (GHI) received by the photovoltaic module often experiences a steep reverse surge, and its value may even exceed the theoretical clear sky irradiance. This phenomenon is known as cloud edge over-irradiance or lensing effect.

[0088] If this effect is ignored, the prediction system will calculate that the power is about to decrease based on the geometric projection, thus sending a downsizing command to the dispatch side prematurely; in reality, the power plant is in a period of power surge at this time. This phase misalignment will cause the prediction error (RMSE) to surge, and may even trigger logic oscillations in the inverter's MPPT (maximum power point tracking).

[0089] Therefore, the slip velocity vector calculated in step four of this invention Based on this, the following photodynamic compensation sub-process is executed: The system uses computer vision algorithms (such as the Canny edge detection operator) to extract the edge contours of cloud clusters in real time from the ground cloud map, and combines cloud height parameters to construct a local three-dimensional tangent plane of the cloud edge.

[0090] The system calculates the vector of the incident sunlight in real time. With the normal vector of the cloud front The angle between .

[0091] According to atmospheric radiative transfer theory, the forward peak effect of Mie scattering produced by cloud droplet particles is most significant when sunlight passes through the sparse droplet region at the edge of a cloud at a grazing angle. This system defines an empirical lens enhancement range, typically defined as... .

[0092] When the cloud shadow leader is detected to be within the aforementioned lens enhancement range, the system determines that: although the physical movement speed of the cloud remains unchanged, the optical shadow boundary of the ground lags behind the geometric projection boundary due to multiple scattering and refraction paths bending of light as it passes through the cloud edge.

[0093] To effectively represent this physical delay in the mathematical model, this invention introduces a deceleration correction factor. (The value range is usually set to 0.85-0.95). The system corrects the slip velocity vector as follows: ; in: This represents the effective slip velocity vector after correction for the lens effect; This represents a dimensionless optical deceleration factor, the value of which depends on the optical thickness gradient at the edge of the cloud.

[0094] The physical meaning of this correction is that, due to the lensing effect, the effective leading edge of the shadow passes through the photovoltaic array at a slower speed than the physical projection of the cloud. This causes the power drop edge in the prediction curve to be precisely delayed on the time axis, eliminating prediction errors.

[0095] In addition to velocity correction, the system also superimposes a duration of [value] before the falling edge of the predicted curve. The positive pulse component. The amplitude gain of this pulse. It exhibits a non-linear positive correlation with the optical thickness (COD) of clouds and the solar zenith angle. In this embodiment, this positive pulse component... A Gaussian kernel function is used to model the physical distribution of light intensity energy. The calculation formula is as follows: ; in: Indicates in The power increment of the lens effect superimposed at each moment, in kilowatts (kW). This indicates the calculated time when the cloud shadow vane arrives at the subarray, in seconds (s). The parameter representing the width of the Gaussian pulse takes a value of ,in The preset duration window for the lens effect; Represented by natural constant An exponential function with base 0.

[0096] The amplitude gain of the pulse The following empirical photodynamic formula is used for calculation: ; in: This represents the maximum power increase caused by the lensing effect, expressed in kilowatts (kW). This represents the theoretical clear-sky power at the current moment, in kilowatts (kW). This represents the lens intensity coefficient, a dimensionless constant set based on historical data fitting, with a typical value range of 0.15 to 0.25. is the base of the natural logarithm, with a value of approximately 2.71828; This represents the optical thickness of the cloud layer obtained by inversion using a radiometer; it is a dimensionless physical quantity. The solar zenith angle (in radians) is represented as follows: Subtract the solar altitude angle; It represents the square of the sine function.

[0097] This step accurately reproduces the typical waveform of the measured data, which first surges and then drops sharply, greatly enhancing the physical authenticity of the prediction results.

[0098] Step 6: Complex terrain correction based on 3D digital elevation model (DEM).

[0099] For photovoltaic power stations built in mountainous areas widely distributed in western China, the differences in light and shadow propagation caused by terrain undulations cannot be ignored. The traditional assumption of flat ground leads to significant projection errors.

[0100] During the initialization phase, the system imported a high-precision digital elevation model (DEM) of the site area, with a grid resolution better than 5 meters. For each photovoltaic subarray node... The system extracts its altitude ,slope and slope .

[0101] When calculating cloud shadow propagation, the system constructs a three-dimensional light cone projection coordinate system.

[0102] The system calculates the cloud shadow slip velocity vector. Tangential component on the local inclined plane where the subarray is located The calculation formula is as follows: ; in: This indicates the tangential slip rate of cloud shadows on the terrain slope (unit: meters per second). This indicates the rate of cloud shadow slippage on a horizontal plane (unit: meters per second). This represents the angle between the vector of sunlight rays and the vector of the local slope normal; This represents the angle between the cloud shadow's movement direction vector and the local slope tangent vector.

[0103] When a cloud shadow is cast on a shaded slope, the path of light is lengthened, and the equivalent speed at which the shadow sweeps across the ground increases, resulting in a slower power decrease (time axis stretching). When a cloud shadow is cast on a sunlit slope, the path of light is compressed, and the equivalent speed at which the shadow sweeps across the ground decreases. However, because the slope intercepts more light flux, the shadow boundary is sharper, and the power decrease becomes steeper (time axis compression).

[0104] This system, through the aforementioned geometric transformations, adjusts the propagation time delay parameters of hundreds of subarrays across the entire station. One by one, corrections were made to eliminate the spatiotemporal distortion errors caused by the terrain.

[0105] Step 7: Automatic termination logic and robust control under extreme operating conditions.

[0106] To ensure the system's high reliability under all-weather conditions, this invention designs defensive logic for extreme weather conditions.

[0107] The system monitors the instantaneous wind speed at a height of 50 meters in real time. When the wind speed exceeds a preset threshold (e.g., 12 m / s), according to the principles of fluid dynamics, the atmosphere is in a state of strong turbulence (Reynolds number much greater than 4000). At this time, the cloud layer undergoes violent dissipation and reorganization, no longer satisfying the grayscale conservation assumption of the optical flow method.

[0108] The system calculates the texture entropy of the entire sky cloud map. If the entropy value is too high, it indicates that the sky is full of fine mackerel clouds or broken cumulus clouds, and the correspondence between light and shadow is extremely chaotic.

[0109] Once any of the above conditions are triggered, the system immediately suspends the complex physical correction process and automatically switches to conservative prediction mode. In this mode, the system only outputs a smooth curve generated by the LSTM base model and marks low confidence in the data quality bits. This strategy effectively avoids drastic oscillations in the prediction curve, ensuring the safety of power grid dispatch.

[0110] Step 8: Closed-loop feedback adaptive calibration based on measured deviation.

[0111] To address the industry problem of the difficulty in accurately measuring cloud base height (CBH), this invention designs a backpropagation calibration mechanism.

[0112] The system continuously monitors the measured power of the sentinel nodes and follower nodes. Assume the prediction model calculates that the cloud shadow should be in... The time reaches a certain following node, but the measured power drop occurs during... time.

[0113] Calculate time deviation .like (Late arrival), according to geometric formulas This suggests that the actual cloud height may be higher than the set value (or the cloud speed may be slower than the observed value).

[0114] The system utilizes PID control algorithms or Kalman filters to... As an observation residual, the cloud base height parameter is dynamically corrected: ; in: Indicates the corrected cloud base height; Indicates the cloud base height at the previous moment; This is the proportionality coefficient. The integral coefficient; This represents the deviation between the measured arrival time and the predicted arrival time.

[0115] Revised This information is immediately applied to the forecast calculations for the next minute. Through this online iteration, the system can rapidly approximate the actual atmospheric conditions, thereby improving forecast accuracy.

[0116] The present invention further provides a detailed device architecture for a photovoltaic power plant ultra-short-term power prediction system based on solar trajectory process propagation.

[0117] Reference manual attached Figure 3 The system is physically deployed in a high-performance industrial server cluster within the booster station of a photovoltaic power plant, and logically divided into the following core modules: Multi-source heterogeneous data acquisition and spatiotemporal alignment module: This module is the system's sensing front end. Physically, it connects to the all-sky imager, micro-weather station, and SCADA server via an industrial Ethernet switch. This module incorporates an FPGA (Field-Programmable Gate Array) hardware timestamp engine. It does not rely on the operating system's soft clock but directly parses PTP (IEEE 1588) protocol messages, ensuring strict alignment of image frames, radiation data, and power data with microsecond-level precision. This is the physical basis for high-precision spatiotemporal correlation analysis.

[0118] Solar Trajectory and Geometric Light Path Calculation Module: This module is the mathematical core of the system. It embeds a high-precision C++ implementation library of the NRELSPA algorithm. This module maintains a solar ray projection matrix. This matrix is ​​refreshed at a frequency of 1Hz and stores the solar altitude angle at the current moment in real time. Azimuth and its first derivative In particular, the module also includes an atmospheric camber difference correction unit, which corrects the optical path curvature for the low-angle apparent position of the sun based on measured temperature and pressure.

[0119] Vector Field Analysis and Dynamics Modeling Module: This module is responsible for cloud image processing and velocity field calculation. Physically, it utilizes NVIDIA Tesla series GPUs configured on the server for acceleration. This module runs a dense optical flow algorithm. Unlike sparse optical flow, which only tracks a few corner points, dense optical flow calculates the displacement vector of every pixel in the image, forming a smooth velocity field. Subsequently, the module uses a Kalman filter to fuse visual velocity with the dynamic wind speed from a ground-based anemometer, outputting the optimally estimated cloud movement vector.

[0120] Dynamic Spatial Topology and Power Inversion Module: This module constructs a dynamic geometric mesh object in memory. The static layer loads the 3D DEM model of the photovoltaic power plant and the subarray distribution topology. The dynamic layer renders the shadow textures overlaid on the static layer in real time based on the cloud shadow positions. To improve computational efficiency, the R-Tree spatial indexing algorithm is adopted. When calculating the impact of a cloud on the ground, the system quickly retrieves all subarrays within the cloud's projection bounding box, avoiding the unnecessary consumption of computational resources for full-site traversal. It also incorporates a diode equivalent circuit model, calculating the component mismatch loss non-linearly based on the shadow occlusion ratio, rather than a simple area ratio calculation.

[0121] Closed-loop feedback controller: This module runs an adaptive parameter optimization algorithm. It maintains a long short-term memory queue, storing predicted and measured values ​​from the past four hours. Gradient descent is used to fine-tune cloud height online. Lens gain Equal hyperparameters ensure that the system always operates in a state of minimum error.

[0122] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for ultra-short-term power prediction of photovoltaic power plants based on solar trajectory process propagation, the method comprising the following steps: Collect real-time meteorological data, ground-based cloud map data, and historical operating power data of photovoltaic power plants; Based on the geographical latitude and longitude of the photovoltaic power station and the predicted time window, the solar trajectory process parameters are calculated, including the solar altitude angle and solar azimuth angle that change continuously over time. Using the historical operating power data and the real-time meteorological data, an initial power prediction curve is generated through a basic prediction model; The method is characterized by further including performing a correction process based on the propagation of the solar trajectory process: Identify cloud layer movement vectors based on the aforementioned ground-based cloud map data; A model of the sunlight projection channel is constructed, and the cloud movement vector is mapped to the ground projection plane where the photovoltaic power station is located. The sliding velocity vector of the cloud shadow propagating on the ground is calculated by combining the rate of change of the solar trajectory process parameters. The propagation time delay parameters of the cloud shadow reaching each sub-array of the photovoltaic power station are calculated using the slip velocity vector; The time axis of the initial power prediction curve is dynamically shifted and calibrated based on the propagation delay parameter, and the final ultra-short-term power prediction result is output.

2. The method for ultra-short-term power prediction of photovoltaic power plants based on solar trajectory process propagation according to claim 1, characterized in that, The construction of the solar radiation projection channel model and the calculation of the slip velocity vector of cloud shadows propagating on the ground specifically include: Extract the differential value of the solar altitude angle with respect to time to obtain the solar altitude angular velocity; Establish a geometric correlation function between cloud height, solar altitude angle, and ground projection distance; The cloud movement vector is decomposed into tangential and normal components. Based on the geometric correlation function and the solar altitude angular velocity, the additional projected velocity caused by the solar motion is calculated. The slip velocity vector is obtained by vector superimposing the geometric projection velocity of the cloud movement vector onto the ground with the additional projection velocity.

3. The method for ultra-short-term power prediction of photovoltaic power plants based on solar trajectory process propagation according to claim 1, characterized in that, The calculation of the propagation time delay parameters of cloud shadows reaching each sub-array of the photovoltaic power station using the slip velocity vector specifically includes: A virtual spatial topology map of a photovoltaic power station is constructed, wherein the center coordinates of each subarray are marked in the topology map; The subarrays on the windward side are defined as outpost nodes, and the subarrays on the leeward side are defined as follower nodes; Based on the slip velocity vector, calculate the physical transit time of the cloud shadow vanguard from the outpost node to the follower node; The physical transit time is used as the propagation delay parameter.

4. The method for ultra-short-term power prediction of photovoltaic power plants based on solar trajectory process propagation according to claim 3, characterized in that, The step of dynamically shifting and calibrating the time axis of the initial power prediction curve based on the propagation delay parameter specifically includes: Detect whether the measured power of the sentinel node undergoes a sudden change; If a mutation occurs, extract the mutation time and mutation slope; Using the propagation delay parameter, the predicted time when the power change of the following node occurs can be calculated; The falling edge of the predicted power corresponding to the following node in the initial power prediction curve is shifted to the prediction time.

5. The method for ultra-short-term power prediction of photovoltaic power plants based on solar trajectory process propagation according to claim 2, characterized in that, The calculation of the slip velocity vector of cloud shadows propagating on the ground also includes compensation for the lensing effect at the cloud edge: Calculate the angle between the angle of incidence of sunlight and the normal to the edge of the cloud; When the included angle is within the preset lens enhancement range, the sliding velocity vector is decelerated and corrected to simulate the blurring delay of the shadow boundary caused by the refraction of light at the edge of the cloud.

6. The method for ultra-short-term power prediction of photovoltaic power plants based on solar trajectory process propagation according to claim 1, characterized in that, The basic prediction model is a long short-term memory neural network or a gated recurrent unit, and the time resolution of the initial power prediction curve is 1 minute to 15 minutes.

7. The method for ultra-short-term power prediction of photovoltaic power plants based on solar trajectory process propagation according to claim 1, characterized in that, The method also includes an abnormal operating condition elimination step: Before calculating the slip velocity vector, it is determined whether the instantaneous wind speed in the real-time meteorological data exceeds a preset threshold; If the preset threshold is exceeded, the cloud formation is determined to be in a state of rapid fragmentation, the correction process is stopped, and the initial power prediction curve is directly output.

8. The method for ultra-short-term power prediction of photovoltaic power plants based on solar trajectory process propagation according to claim 3, characterized in that, The virtual space topology map for constructing the photovoltaic power station also includes: Obtain a digital elevation model of the site where the photovoltaic power station is located; Based on the terrain slope and aspect, the equivalent light-receiving coordinates of each subarray in the virtual space topology map are corrected.

9. The method for ultra-short-term power prediction of photovoltaic power plants based on solar trajectory process propagation according to claim 4, characterized in that, After outputting the final ultra-short-term power prediction result, a closed-loop feedback step is also included: Collect the measured power data at the next moment; Calculate the deviation between the measured power change time and the predicted time; The deviation value is used to reverse the cloud height parameter in the solar projection channel model.

10. A photovoltaic power plant ultra-short-term power prediction system based on solar trajectory process propagation, characterized in that, The system includes: The data acquisition module is configured to collect real-time meteorological data, ground cloud map data, and historical operating power data of the photovoltaic power station; The trajectory calculation module is configured to calculate solar trajectory process parameters based on the geographical latitude and longitude of the photovoltaic power station and the prediction time window; The basic forecasting module is configured to generate an initial power forecast curve using the historical operating power data and the real-time meteorological data; The propagation correction module is configured to construct a solar ray projection channel model, calculate the slip velocity vector of cloud shadow propagation on the ground by combining the rate of change of the solar trajectory process parameters, and calculate the propagation time delay parameters based on the slip velocity vector; The result output module is configured to dynamically shift and calibrate the time axis of the initial power prediction curve according to the propagation delay parameter, and output the final ultra-short-term power prediction result.

Images