A photovoltaic power station photovoltaic power prediction method and system

Abstract

The application provides a photovoltaic power station photovoltaic power prediction method and system. Among them, a cloud layer movement time sequence library is constructed based on historical cloud charts, and a nonlinear model of cloud speed and output attenuation is established. The light sensor array monitors the cloud layer thickness and aerosol density in real time, combines the model to calculate the shadow area light intensity attenuation gradient, and predicts the remaining time of the cloud group reaching the core area. According to the remaining time, the power prediction confidence interval is divided, the satellite data is input into the multi-scale meteorological fusion architecture, and the meteorological feature extraction path of the attention mechanism is adjusted based on the cloud density. The ground reflectivity correction value calculated by fusing the light sensor scattering data is used to generate a short-term power prediction sequence of the sudden cloud layer scene. The sequence is input into the energy management system, combined with the reactive power compensation parameters of the grid-connected point voltage deviation and the grid rules, and the dual-objective scheduling instructions of active and reactive power are output. The technical scheme provided by the application improves the cooperative scheduling accuracy of the photovoltaic power station under the sudden change of the cloud layer.

Description

Technical Field

[0001] This application relates to the field of power prediction technology, and in particular to a method and system for predicting photovoltaic power in a photovoltaic power plant. Background Technology

[0002] With the large-scale grid connection of photovoltaic (PV) power generation, sudden power fluctuations caused by cloudy weather pose a severe challenge to the stable operation of the power grid. Rapid cloud movement can trigger sudden and drastic changes in surface irradiance, causing significant fluctuations in the output power of PV power plants within a short period. Traditional forecasting methods, limited by the update rate of meteorological data and spatial monitoring blind spots, struggle to capture the characteristics of power fluctuations caused by cloud cover in a timely manner, resulting in the grid dispatch system's inability to effectively predict power fluctuation trends. There is an urgent need for a forecasting method that integrates dynamic meteorological evolution characteristics with comprehensive power plant status awareness to achieve accurate early warning of minute-level power fluctuations.

[0003] The current mainstream approach employs a predictive framework that combines meteorological satellite cloud trajectory prediction with power plant-level irradiance monitoring. This approach uses satellite cloud images to invert cloud movement direction and combines this with real-time data from irradiance sensors deployed within the power plant to construct a deep learning-based time-series prediction model. Specifically, it predicts the temporal changes in cloud shadow coverage over the power plant based on cloud trajectory prediction, simultaneously fusing irradiance monitoring data to generate a photovoltaic power prediction curve for future periods. Under stable cloud morphology conditions, this method can provide trend prediction results, offering a fundamental reference for power grid dispatching.

[0004] However, this approach has significant limitations in practical applications under cloudy weather scenarios: First, the update rate of the meteorological satellite's cloud trajectory prediction model cannot match the actual cloud movement speed, leading to accumulated prediction errors in cloud arrival time. Second, sparsely distributed irradiance sensors within the power plant struggle to characterize spatially heterogeneous power surges caused by cloud shadow occlusion, resulting in missed detections of local occlusion events. Finally, the deep learning model's strong dependence on historical data leads to a lag in its response to sudden irradiance changes, resulting in excessively smooth prediction curves that fail to accurately reflect short-term, drastic fluctuations in real-world scenarios. These shortcomings cause a significant phase difference between the prediction results output by the existing method and the actual power curve, failing to meet the power grid's demand for refined rapid power regulation. Summary of the Invention

[0005] This application provides a photovoltaic power prediction method and system for photovoltaic power plants, which solves the problems of lag and error accumulation in the prediction of sudden changes in photovoltaic power under cloudy weather conditions in the prior art.

[0006] Firstly, this application provides a method for predicting the photovoltaic power output of a photovoltaic power plant, including:

[0007] By constructing a cloud movement time series database through historical cloud map sequences of the photovoltaic power station area, the spatiotemporal relationship between cloud trajectory and light change is extracted, and a nonlinear model of cloud movement speed and photovoltaic power output attenuation is established.

[0008] A light sensor array is deployed at the edge of the photovoltaic array to detect cloud thickness and aerosol density in real time. The light intensity attenuation gradient in the local shadow area is calculated by combining the nonlinear model to predict the remaining time for the target cloud to reach the core area of ​​the photovoltaic array.

[0009] The confidence interval for short-term power prediction is dynamically divided based on the remaining time. Meteorological satellite data is input into a multi-scale meteorological fusion architecture, and the feature extraction path for different meteorological parameters is adjusted based on the current cloud density using an attention mechanism.

[0010] By fusing the adjusted feature extraction path with the surface reflectance correction value, a short-term power prediction sequence adapted to abrupt cloud formation scenarios is generated. The surface reflectance correction value is calculated from the scattering characteristics fed back in real time by the optical sensor array.

[0011] The short-term power prediction sequence is input into the energy management system. Combined with the voltage deviation and reactive power compensation demand parameters of the photovoltaic power plant grid connection point, a dual-objective scheduling instruction that takes into account both active power regulation and reactive power control is generated. The reactive power regulation demand parameters are set by the energy management system according to the power factor requirements of the grid connection point in the power grid operation rules, and are used to constrain the priority and magnitude range of reactive power regulation.

[0012] Optionally, by fusing the adjusted feature extraction path with the surface reflectance correction value, a short-term power prediction sequence adapted to abrupt cloud formation scenarios is generated. The surface reflectance correction value is calculated from the scattering characteristics fed back in real-time by the optical sensor array, including:

[0013] The scattering characteristic parameters at multiple locations below the cloud layer are collected in real time by an array of optical sensors. The scattering characteristic parameters include the light intensity attenuation ratio at different angles and the scattering angle distribution.

[0014] Based on the light intensity attenuation ratio and scattering angle distribution in the scattering characteristic parameters, the surface reflectance correction value corresponding to each optical sensor location is calculated. The surface reflectance correction value reflects the dynamic influence of cloud thickness changes on surface reflectance.

[0015] The surface reflectance correction value is fused with the feature extraction path. The fusion process refers to weighting and superimposing the surface reflectance correction value with the cloud density distribution in meteorological satellite data to generate a fused feature vector with dynamic reflectance compensation.

[0016] The cloud density distribution and the surface reflectance correction value in the fused feature vector are coupled by time series and spatial grid to establish a prediction equation for the movement rate of abrupt cloud layers above the photovoltaic array.

[0017] The time variation parameter of cloud density distribution in the moving rate prediction equation is dynamically matched with the remaining time. By using the product relationship between cloud moving rate and remaining time in the time variation parameter, a short-term power prediction sequence adapted to sudden cloud scenarios is generated.

[0018] Optionally, the cloud density distribution and the surface reflectance correction value in the fused feature vector are parameter-coupled according to time series and spatial grid to establish a prediction equation for the movement rate of abruptly changing clouds above the photovoltaic array, including:

[0019] The cloud density distribution in the fused feature vector is divided into continuous time windows according to the time series, and the parameter change trend of cloud density distribution over time is extracted in each time window.

[0020] The surface reflectance correction value is divided into multiple grid cells according to the spatial grid, and the spatial difference parameter of the surface reflectance correction value between adjacent grid cells is calculated.

[0021] The trend of the parameter change is coupled with the spatial difference parameter, and a spatiotemporal joint change weight is generated by multiplying the rate of change of cloud density within the time window with the magnitude of the difference in surface reflectance of the spatial grid cell.

[0022] Based on the spatiotemporal joint change weights, the cloud density distribution and the surface reflectance correction value within the current time window are dynamically matched to construct a parameter change model for the cloud movement direction and speed.

[0023] By using the proportional relationship between the rate of change of cloud density and the magnitude of the difference in surface reflectivity in the parameter change model, the movement rate parameter of the target cloud above the photovoltaic array is calculated, and the movement rate prediction equation is obtained.

[0024] Optionally, a light sensor array is deployed at the edge of the photovoltaic array to detect cloud thickness and aerosol density in real time. Combined with the aforementioned nonlinear model, the light intensity attenuation gradient in the local shaded area is calculated to predict the remaining time for the target cloud cluster to reach the core area of ​​the photovoltaic array, including:

[0025] A photosensitive array is deployed at the edge of the photovoltaic array to collect scattered light intensity distribution data at multiple locations below the cloud layer in real time, and to extract the scattered light angle distribution parameters and light intensity attenuation ratio parameters from the scattered light intensity distribution data.

[0026] Based on the light intensity attenuation ratio parameter and the scattered light angle distribution parameter, the cloud thickness parameter and aerosol density parameter corresponding to each photosensitive unit in the photosensitive array are detected in real time. The cloud thickness parameter is obtained by comparing the light intensity attenuation ratio with a preset occlusion threshold.

[0027] The cloud thickness parameter and aerosol density parameter are input into a nonlinear model, and the dynamic change rate of the cloud thickness parameter and the fluctuation amplitude of the aerosol density parameter are correlated to generate the light intensity attenuation gradient of the local shadow area.

[0028] Based on the light intensity attenuation gradient, the cloud movement direction parameter in the nonlinear model is weighted and superimposed with the spatial distance of the photovoltaic array core area to generate the driving intensity factor of cloud approaching the core area.

[0029] The cloud movement speed is dynamically corrected by the driving intensity factor, and the ratio of the dynamically corrected cloud movement speed to the spatial distance of the photovoltaic array core area is calculated to predict the remaining time for the target cloud to reach the photovoltaic array core area.

[0030] Optionally, the cloud thickness parameter and aerosol density parameter are input into a nonlinear model, and the dynamic change rate of the cloud thickness parameter is correlated with the fluctuation amplitude of the aerosol density parameter to generate a light intensity attenuation gradient in the local shaded area, including:

[0031] The cloud thickness parameter is divided into continuous time segments according to a time window, and the rate of change of the cloud thickness parameter in each time segment is extracted as the dynamic rate of change.

[0032] The aerosol density parameter is calculated based on the spatial distribution differences between adjacent optical sensor locations to determine the aerosol density fluctuation amplitude at each optical sensor location.

[0033] Based on the dynamic response rules of photovoltaic power output attenuation with cloud thickness parameters and aerosol density in the nonlinear model, the dynamic change rate is multiplied by the aerosol density fluctuation amplitude to generate a superimposed weighting factor.

[0034] The spatial distribution gradient of cloud thickness parameter and the adjacent differences of aerosol density parameter within the current time segment are weighted and fused according to the superposition weighting factor to generate a combined attenuation factor;

[0035] The combined attenuation factor is input into the dynamic response rule of photovoltaic power output attenuation in the nonlinear model. The light intensity attenuation gradient of the local shaded area is output through the linear combination ratio of cloud thickness parameter and aerosol density in the dynamic response rule.

[0036] Optionally, the short-term power prediction sequence is input into the energy management system, and combined with the voltage deviation and reactive power compensation demand parameters of the photovoltaic power plant grid connection point, a dual-objective scheduling instruction that takes into account both active power regulation and reactive power control is generated, including:

[0037] The short-term power prediction sequence is input into the energy management system to extract the active power prediction parameters within the current time window. The active power prediction parameters characterize the power output fluctuation characteristics of the photovoltaic array after being affected by cloud cover.

[0038] Real-time collection of voltage deviation parameters at the grid connection point of photovoltaic power plants; analysis of voltage deviation direction and magnitude; classification of voltage compensation requirement levels.

[0039] By combining the reactive power compensation requirement parameters with the current voltage deviation parameters, the reactive power compensation adjustment intensity parameters are calculated. The adjustment intensity parameters reflect the reactive power adjustment level required for grid connection point voltage stability.

[0040] By integrating the active power prediction parameters, voltage compensation demand level, and reactive power compensation adjustment intensity parameters, a dual-objective scheduling factor is generated by superimposing and balancing the active power adjustment demand and reactive power control demand through a dynamic weight allocation rule.

[0041] By using the superposition and balance ratio of active and reactive power demand in the dual-objective scheduling factor, a dual-objective scheduling command that takes into account both active power regulation and reactive power control is output.

[0042] Optionally, the confidence interval for short-term power prediction is dynamically divided based on the remaining time. Meteorological satellite data is input into a multi-scale meteorological fusion architecture, and the feature extraction path for different meteorological parameters is adjusted based on the current cloud density using an attention mechanism, including:

[0043] Based on the remaining time for the target cloud cluster to reach the core area of ​​the photovoltaic array, the remaining time is divided into multiple consecutive time window lengths;

[0044] The confidence interval range for short-term power prediction is divided according to the relationship between the time window length and the cloud movement rate.

[0045] Meteorological satellite parameters are divided into multi-scale meteorological feature layers, and the multi-scale meteorological fusion architecture is input to extract cloud density parameters and wind speed parameters at different scales.

[0046] Based on the confidence interval range, calculate the dynamic weight values ​​of cloud density parameters and wind speed parameters, and adjust the priority of cloud density parameters and wind speed parameters in the feature extraction path through the dynamic weight values.

[0047] The cloud density and wind speed parameters in the feature extraction path after priority adjustment are assigned weight ratios according to the confidence interval range. The cloud density and wind speed parameters are then weighted and accumulated using the weight ratios to generate a short-term power prediction sequence adapted to abrupt cloud scenarios.

[0048] Secondly, this application provides a photovoltaic power prediction system for a photovoltaic power plant, comprising:

[0049] The module is used to build a cloud movement time series library by using historical cloud map sequences of the photovoltaic power station area, extract the spatiotemporal relationship between cloud trajectory and light change, and establish a nonlinear model of cloud movement speed and photovoltaic power output attenuation.

[0050] The prediction module is used to deploy a light sensor array at the edge of the photovoltaic array to detect cloud thickness and aerosol density in real time. It combines the nonlinear model to calculate the light intensity attenuation gradient in the local shadow area and predict the remaining time for the target cloud to reach the core area of ​​the photovoltaic array.

[0051] The adjustment module is used to dynamically divide the confidence interval of short-term power prediction according to the remaining time, input meteorological satellite data into the multi-scale meteorological fusion architecture, and adjust the feature extraction path of different meteorological parameters based on the current cloud density through the attention mechanism.

[0052] The generation module is used to fuse the surface reflectance correction value through the adjusted feature extraction path to generate a short-term power prediction sequence adapted to the sudden cloud layer scenario. The surface reflectance correction value is calculated from the scattering characteristics fed back in real time by the optical sensor array.

[0053] The input module is used to input the short-term power prediction sequence into the energy management system. Combined with the voltage deviation and reactive power compensation demand parameters of the photovoltaic power plant grid connection point, it generates a dual-objective scheduling instruction that takes into account both active power regulation and reactive power control. The reactive power regulation demand parameters are set by the energy management system according to the power factor requirements of the grid connection point in the power grid operation rules, and are used to constrain the priority and magnitude range of reactive power regulation.

[0054] Thirdly, embodiments of this application provide a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a photovoltaic power prediction method for a photovoltaic power plant as described in the first aspect above.

[0055] Fourthly, embodiments of this application provide a computer storage medium storing a computer program, which, when executed by a computer, implements a photovoltaic power prediction method for a photovoltaic power plant as described in the first aspect.

[0056] This application constructs a cloud movement time series database using historical cloud image sequences of the photovoltaic power station area, extracts the spatiotemporal relationship between cloud trajectories and light intensity changes, and establishes a nonlinear model of cloud movement speed and photovoltaic power output attenuation. This enables dynamic correlation modeling of cloud movement patterns and photovoltaic power output attenuation, improving the spatiotemporal coupling accuracy of the prediction data. By deploying a light sensor array at the edge of the photovoltaic array to detect cloud thickness and aerosol density in real time, and combining this with the aforementioned nonlinear model to calculate the light intensity attenuation gradient in local shadowed areas, the remaining time for the target cloud to reach the core area of ​​the photovoltaic array can be predicted. This allows for real-time capture of the microscopic changes in local shadows and accurate prediction of the time window of cloud abrupt changes affecting the core power generation area. By dynamically dividing the confidence interval for short-term power prediction based on the remaining time, inputting meteorological satellite data into a multi-scale meteorological fusion architecture, and adjusting the attention mechanism for feature extraction paths of different meteorological parameters based on the current cloud density, short-term power prediction can be enhanced in the face of abrupt changes. The system possesses the capability for dynamic segmentation of reliable intervals in cloud scenarios, optimizing the adaptability of meteorological parameter feature extraction. By fusing the adjusted feature extraction path with surface reflectance correction values, a short-term power prediction sequence adapted to abrupt cloud scenarios is generated. The surface reflectance correction values ​​are calculated from the scattering characteristics fed back in real time by the optical sensor array, which enhances the anti-interference capability of the prediction sequence in complex light scattering environments and ensures the synchronization of prediction results with actual light intensity abrupt changes. By inputting the short-term power prediction sequence into the energy management system, and combining the voltage deviation and reactive power compensation demand parameters of the photovoltaic power plant grid connection point, a dual-objective scheduling instruction that considers both active power regulation and reactive power control is generated. The reactive power regulation demand parameters are set by the energy management system according to the power factor requirements of the grid connection point in the power grid operation rules, constraining the priority and magnitude range of reactive power regulation. This enables dynamic optimization of active and reactive power coordination of the photovoltaic power plant grid connection power, meeting the multi-objective constraint requirements of the power grid operation rules for power regulation.

[0057] Furthermore, by employing dynamic reflection compensation and spatiotemporal coupling modeling with multi-source data, the accuracy and real-time performance of photovoltaic power prediction under abrupt cloud formation scenarios are significantly improved. Based on the multi-angle scattering characteristics collected in real time by a light sensor array, combined with a dynamic correction mechanism for surface reflectivity caused by changes in cloud thickness, this method overcomes the limitations of traditional prediction methods that neglect the dynamic feedback of surface light scattering during cloud movement. By weighted superposition of cloud density distribution and reflectivity correction values, coupled with spatiotemporal grid parameters, a moving rate prediction equation integrating dynamic surface reflection compensation is established. This accurately characterizes the impact of the product of cloud moving rate and remaining time on the photovoltaic array, addressing the problem of incomplete modeling of the spatiotemporal transmission process of abrupt cloud shading effects. Ultimately, this achieves a strong correlation between short-term power prediction sequences and the physical processes of cloud abrupt changes, ensuring reliable support for grid dispatch commands under complex weather conditions, effectively reducing the risk of power fluctuations caused by rapid cloud movement, and improving the grid-connected operation stability of photovoltaic power plants.

[0058] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 A flowchart of a photovoltaic power prediction method for a photovoltaic power plant provided in this application is shown;

[0061] Figure 2 This application provides a schematic diagram of the structure of a photovoltaic power prediction system for a photovoltaic power plant.

[0062] Figure 3 A schematic diagram of the structure of a computing device provided in this application is shown. Detailed Implementation

[0063] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0064] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0065] Researchers have found that existing short-term power prediction methods for photovoltaic (PV) power plants largely rely on single meteorological data sources, making it difficult to accurately capture the dynamic impact of sudden cloud changes on light intensity attenuation. Furthermore, they lack adaptive correction for real-time changes in surface reflectivity, leading to accumulated prediction errors and delayed reactive power regulation. Based on this, a power prediction and scheduling optimization method for PV power plants is proposed. This method integrates cloud movement trajectories, real-time light sensor data, and multi-scale meteorological parameters to dynamically generate short-term power prediction sequences adapted to sudden cloud changes, and coordinates dual-objective scheduling commands for both active and reactive power. The technical solution of this application is applicable to power system energy management and frequency regulation control scenarios with a high proportion of PV grid connection.

[0066] The entire R&D process embodies the technical integration of multi-source data fusion and dynamic model correction, aiming to overcome the shortcomings of existing solutions, such as reliance on single meteorological data leading to delayed response to sudden cloud changes, static models being unable to adapt to dynamic light intensity decay, and delays in reactive power regulation caused by the disconnect between forecasting and scheduling instructions.

[0067] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0068] Figure 1 This application provides a flowchart of a photovoltaic power prediction method for photovoltaic power plants, as shown in the embodiments. Figure 1 As shown, the method includes:

[0069] 101. Construct a cloud movement time series library by using historical cloud map sequences of the photovoltaic power station area, extract the spatiotemporal relationship between cloud trajectory and light change, and establish a nonlinear model of cloud movement speed and photovoltaic power output attenuation.

[0070] In this step, the cloud motion time series database refers to a database composed of historical satellite cloud image sequences, containing spatiotemporal characteristics such as cloud location, shape, and movement speed. The nonlinear model refers to a mathematical model describing the relationship between cloud movement speed and photovoltaic power output degradation; for example, an increase of 1 m / s in cloud speed leads to a 5% to 15% decrease in power output (depending on cloud thickness).

[0071] In this embodiment, firstly, historical meteorological satellite cloud image sequences of the photovoltaic power station area are collected. Image processing techniques (such as optical flow or cloud tracking algorithms) are used to extract the cloud movement trajectories, and the coverage area and movement speed over time are recorded to construct a cloud movement time series database. Secondly, combined with the power output data of the photovoltaic power station during the same period, the law of light intensity attenuation under different cloud movement speeds is analyzed (such as the lag time and attenuation magnitude of power output reduction caused by cloud shading). Finally, using a nonlinear regression model (such as a neural network or support vector machine), based on the cloud movement time series database and the law of light intensity attenuation under different cloud movement speeds, a mapping relationship between cloud movement speed and power output attenuation is established, forming a quantifiable photovoltaic power output prediction model for cloud shadow impact, providing a benchmark for real-time shading prediction.

[0072] In a scenario where a 300MW photovoltaic power plant encountered a rapidly moving cumulonimbus cloud cluster, the operation and maintenance team retrieved meteorological satellite cloud image sequences from 10:00 to 16:00 daily during the summer of the past three years, constructing a time-series database containing cloud movement trajectories, coverage areas, and corresponding photovoltaic power output data. By analyzing the correlation between cloud movement speed and power plant power output decline, it was found that when the cloud speed exceeds 15m / s, the power output attenuation rate increases non-linearly, establishing a predictive model of "cloud front arrival time - power output loss rate." For example, when the cloud cluster approaches at a speed of 18m / s, the model predicts that the core array area will experience a risk of a sudden 65% drop in power output within 12 minutes. This judgment is verified based on actual shading data at similar movement speeds in the cloud cluster's historical trajectory, providing a benchmark reference for subsequent real-time monitoring.

[0073] 102. Deploy a light sensor array at the edge of the photovoltaic array to detect cloud thickness and aerosol density in real time. Combine the nonlinear model to calculate the light intensity attenuation gradient in the local shadow area and predict the remaining time for the target cloud to reach the core area of ​​the photovoltaic array.

[0074] In this step, the light intensity attenuation gradient refers to the rate at which light intensity decreases due to cloud cover; for example, light intensity decreases by 20% for every 100-meter increase in cloud thickness. The remaining time refers to the estimated time required for the target cloud cluster to move from its current edge location to the core area of ​​the photovoltaic array.

[0075] In this embodiment, firstly, a photosensitive array deployed at the edge of the photovoltaic array collects cloud thickness data in real time (by measuring the ratio of transmitted light intensity to scattered light intensity) and aerosol density (through spectral absorption characteristic analysis). Combined with the nonlinear model from step one, the light intensity attenuation gradient of the current local shadow area is calculated (e.g., an increase in cloud thickness causes a 5% decrease in light intensity every 10 seconds). Secondly, based on the cloud's movement speed and the photovoltaic array's layout direction, the remaining time for the target cloud to reach the core power generation area from the edge is predicted (e.g., a cloud moving at a speed of 2 m / s needs 50 seconds to cover the core area). Finally, the predicted remaining time is bound to the light intensity attenuation gradient to generate a dynamic shadow coverage time axis, which is used for dividing the confidence interval for subsequent power prediction.

[0076] When the cumulonimbus cloud front was 5 kilometers from the power plant boundary, the 48 optical sensor arrays deployed on the east side of the array detected a surge in cloud optical thickness from 2.1 to 5.8, and an increase in aerosol density to three times the normal value. The system input the real-time monitoring data into the established power output attenuation model, calculating that the light intensity attenuation gradient of subarrays 5-8 in the eastern area reached an emergency state of 12% attenuation per minute. Combined with the three-dimensional cloud motion vector provided by the Doppler weather radar, it predicted that the main body of the cloud would cover the core power generation area of ​​the power plant in 8 minutes and 30 seconds. At this time, the system automatically triggered a level-two warning and sent a pre-notification to the power grid dispatch center that "power output fluctuation exceeds 50% within 15 minutes," thus gaining a critical time window for the allocation of reserve capacity in the power system.

[0077] 103. Based on the remaining time, dynamically divide the confidence interval of short-term power prediction, input meteorological satellite data into a multi-scale meteorological fusion architecture, and adjust the feature extraction path of different meteorological parameters based on the current cloud density using the attention mechanism.

[0078] In this step, the confidence interval refers to the probability distribution range of the power prediction value, for example, a prediction error of ±3% within 5 minutes of remaining time. The multi-scale meteorological fusion architecture refers to a multi-layered data processing framework that integrates satellite cloud images, ground-based optical sensors, and historical data. The attention mechanism refers to a neural network module that dynamically adjusts the weights of different meteorological parameters.

[0079] In this embodiment, firstly, based on the remaining predicted time, the confidence interval of the short-term power prediction is dynamically divided according to time granularity (e.g., predictions within the next minute are divided into 10-second intervals, with a confidence level of 80%). Secondly, multi-scale data provided by meteorological satellites (such as high-resolution cloud density maps and low-resolution wind speed fields) are input into the multi-scale meteorological fusion architecture, and the weights of different parameters are dynamically adjusted through an attention mechanism: when cloud density is high, the feature extraction of cloud thickness and movement direction is enhanced; when cloud density is sparse, the influence of wind speed on cloud shadow diffusion is emphasized. Finally, the fused meteorological feature vector is output as the core input of the power prediction model.

[0080] Based on the remaining time for the cloud cluster to reach the core area, the system dynamically adjusts the confidence interval for the power prediction over the next 30 minutes from ±10% under normal operating conditions to ±4%. Simultaneously, the multi-scale meteorological fusion architecture receives 0.5km resolution cloud data from the Fengyun-4 satellite. For the currently monitored high-density cumulonimbus cloud characteristics, the weighting coefficient of the attention mechanism for near-surface humidity parameters is increased from 0.3 to 0.6, enhancing the ability to identify turbulent motion at the cloud base. By fusing temperature and humidity correction parameters provided by ground meteorological stations, an input dataset containing an 8-dimensional meteorological feature vector is reconstructed, focusing on capturing the impact of cloud phase transitions (water vapor condensation) on the surface temperature of photovoltaic panels, improving the prediction model's response accuracy to sudden cloud formations by 22%.

[0081] 104. By fusing the adjusted feature extraction path with the surface reflectance correction value, a short-term power prediction sequence adapted to the sudden cloud layer scenario is generated. The surface reflectance correction value is calculated from the scattering characteristics fed back in real time by the optical sensor array.

[0082] In this step, the surface reflectance correction value refers to the correction for light reflection effect parameters altered by ground environment (such as snow cover, vegetation). The short-term power forecast sequence refers to the minute-level forecast of photovoltaic power output for the next 15 minutes.

[0083] In this embodiment, firstly, surface scattered light data (such as reflected light intensity at different angles) collected in real time by an optical sensor array is used to calculate a surface reflectance correction value (e.g., an increase in aerosols leading to a 3% decrease in reflectance). Secondly, the correction value is superimposed on the meteorological feature vector generated by 103, and a short-term power prediction sequence adapted to abrupt cloud formation scenarios is generated using a time series prediction model (such as LSTM or Transformer). For example, if the corrected reflectance indicates increased surface light absorption, the predicted power output decrease is reduced by 2% compared to the original model. Finally, the power prediction result, including the confidence interval and correction parameters, is output for the energy management system's scheduling decisions.

[0084] When cumulonimbus clouds completely covered the eastern area of ​​the power station, the optical sensor array detected a sudden increase in surface reflectivity from 0.21 under clear skies to 0.39. This was due to enhanced diffuse reflection of light caused by multiple scattering at the bottom of the cloud layer. The system converted real-time scattering characteristic data into reflectivity correction parameters and injected them into the time-series memory unit of the LSTM prediction model, dynamically correcting the light intensity conversion coefficient of the string-level output calculation module. By fusing the adjusted meteorological feature extraction path, an output prediction curve for the next 15 minutes was generated, and the prediction error of its maximum fluctuation range was reduced from 9.7% before correction to 3.5%. At this time, the energy storage system entered the pre-adjustment mode in advance according to the prediction curve, lowering the discharge preparation threshold of energy storage unit No. 2 from 85% to 70% to ensure rapid response to the impending sharp drop in output.

[0085] 105. Input the short-term power prediction sequence into the energy management system, and combine it with the voltage deviation and reactive power compensation demand parameters of the photovoltaic power station grid connection point to generate a dual-objective scheduling instruction that takes into account both active power regulation and reactive power control.

[0086] In this step, the reactive power regulation demand parameter is set by the energy management system according to the power factor requirement of the grid connection point in the power grid operation rules. This parameter constrains the priority and magnitude range of reactive power regulation. The dual-objective dispatch instruction refers to a coordinated control instruction that simultaneously optimizes active power adjustment and reactive power compensation. The power factor requirement refers to the requirement that the grid-connected point cosφ must be maintained between 0.95 and 1.0.

[0087] In this embodiment, firstly, the energy management system receives a short-term power forecast sequence and acquires real-time voltage deviation data (e.g., voltage fluctuation exceeding ±5%) and reactive power compensation demand parameters (e.g., power factor needs to be maintained above 0.95) at the grid connection point. Secondly, it sets reactive power adjustment priorities according to grid rules: if the power factor is not up to standard, reactive power compensation equipment (e.g., SVG) is adjusted first; if the voltage deviation exceeds the limit, active power output is adjusted to balance the load. Finally, a dual-objective scheduling instruction is generated through a multi-objective optimization algorithm (e.g., particle swarm optimization), for example, while ensuring that active power output follows the predicted value, capacitor banks are dynamically switched to meet reactive power demand, and adjustment instructions are output to the photovoltaic inverter and compensation device for execution.

[0088] Ten minutes into the period of heaviest cloud cover, the voltage at the power station's grid connection point dropped to 0.91 pu. The system immediately initiated dual-objective optimization: based on short-term power forecast sequences, active power regulation commands were distributed across 14 transformer substations, limiting the output of the damaged subarrays in the eastern zone to no more than 30% of their rated value, while simultaneously increasing the overload margin of the unshaded units in the western zone to 1.1 times; in accordance with the grid's power factor requirement of 0.9, the SVG equipment output was dynamically adjusted from -15 Mvar to +22 Mvar, and the virtual synchronous machine function of the energy storage system was activated to provide 10 Mvar of dynamic support. Through rolling optimization using model predictive control algorithms, the grid connection point voltage was stabilized within the 0.95-1.03 pu range within 150 seconds, and the power factor remained consistently between 0.91 and 0.93. Compared to traditional single-objective control methods, this reduced the number of protection device actions by 76%, and the power station cumulatively reduced output losses by 23 MWh during the entire cloud passage.

[0089] In summary, steps 101 to 105 achieve spatiotemporal collaborative sensing and dynamic optimization control for photovoltaic power surge prediction under cloudy weather conditions. By integrating historical cloud movement patterns with real-time local meteorological monitoring data from a photoelectric sensor array, a nonlinear correlation model between dynamic cloud shadow shading and power attenuation is constructed, overcoming the limitations of traditional methods in spatiotemporal representation of cloud surge events. Combined with an adaptive feature extraction mechanism based on a multi-scale meteorological fusion architecture, the prediction deviation of cloud trajectory and surface reflectivity parameters are dynamically corrected to generate a high-confidence short-term power prediction sequence. Finally, based on the real-time operating status of the power grid, active and reactive power coordinated dispatch commands are generated, improving the timeliness of power surge prediction while ensuring a dynamic balance between grid voltage stability and reactive power regulation, forming a closed-loop optimization system of meteorological sensing, power prediction, and grid dispatch.

[0090] To improve the prediction timeliness and spatial resolution of photovoltaic power surges under cloudy weather conditions, this study aims to establish a real-time compensation mechanism for dynamic cloud shading and light scattering effects. By spatiotemporally coupling ground scattering feature perception with satellite cloud density, a dynamic correlation model between cloud shadow movement rate and surface reflectivity is constructed to address the prediction phase difference and smoothing effect caused by cloud heterogeneity in traditional methods. Based on multi-source data fusion and a dynamic matching mechanism for remaining time, a power prediction sequence adapted to scenarios of rapid cloud splitting and merging is generated, enabling accurate tracking and early warning of photovoltaic power surges under meteorological disturbances.

[0091] In some embodiments, step 104 involves fusing the adjusted feature extraction path with the surface reflectance correction value to generate a short-term power prediction sequence adapted to abrupt cloud formation scenarios. The surface reflectance correction value is calculated from the scattering characteristics fed back in real-time by the optical sensor array, including:

[0092] 201. Real-time acquisition of scattering characteristic parameters at multiple locations below the cloud layer using an optical sensor array, wherein the scattering characteristic parameters include the light intensity attenuation ratio at different angles and the scattering angle distribution;

[0093] In step 201, scattering characteristic parameters refer to physical quantities that characterize the optical properties of clouds by measuring the degree of light intensity attenuation and scattering angle distribution in different directions. The light intensity attenuation ratio refers to the ratio of the intensity of incident light to transmitted light in a specific direction. The scattering angle distribution refers to the pattern of light intensity distribution in different spatial directions after passing through the cloud layer.

[0094] In this embodiment, firstly, scattering characteristic parameters are collected in real time at multiple locations below the cloud layer using an array of optical sensors. These parameters include the ratio of incident to scattered light intensity at different angles (e.g., measuring the intensity difference between incident and scattered light using a multi-angle spectrometer) and the scattering angle distribution (e.g., recording the energy proportion of scattered light in different directions using a hemispherical sensor). Secondly, environmental noise interference (such as baseline drift caused by changes in solar altitude angle) is removed using calibrated sensor parameters to generate a standardized scattering characteristic dataset. Finally, the scattering characteristic parameters of each optical sensor location are bound to its geographic coordinates to form a spatially gridded real-time observation data table, providing input for subsequent surface reflectance correction.

[0095] 202. Based on the light intensity attenuation ratio and scattering angle distribution in the scattering characteristic parameters, calculate the surface reflectance correction value corresponding to each optical sensor location. The surface reflectance correction value reflects the dynamic influence of cloud thickness changes on surface reflectance.

[0096] In step 202, the surface reflectance correction value refers to the true surface reflectance estimate after eliminating the influence of cloud scattering. Dynamic influence refers to the time-varying deviation in surface reflectance measurements caused by changes in cloud thickness.

[0097] In this embodiment of the application, firstly, the light intensity attenuation ratio is extracted from the scattering characteristic parameters of 201 (for example, the incident light intensity at a certain position is 1000W / m). 2 The intensity after scattering is 800 W / m 2 If the attenuation rate is 20%, then the attenuation ratio is calculated. Combined with the scattering angle distribution (e.g., a high proportion of wide-angle scattering indicates thicker clouds), the surface reflectance correction value for each light sensor location is calculated using empirical formulas or machine learning models (e.g., random forest regression). For example, when cloud thickness leads to a decrease in surface reflectance, the correction value can be quantified as -5%. Finally, the correction values ​​are associated with the corresponding sensor locations according to a spatial grid to generate a dynamic reflectance correction map, used to compensate for the impact of cloud thickness changes on surface light absorption characteristics.

[0098] 203. The surface reflectance correction value is fused with the feature extraction path. The fusion process refers to weighting and superimposing the surface reflectance correction value with the cloud density distribution in the meteorological satellite data to generate a fused feature vector with dynamic reflectance compensation.

[0099] In step 203, the fused feature vector refers to a multi-dimensional data vector that fuses ground reflection correction and satellite cloud features. Weighted overlay refers to dynamically adjusting the fusion weights of the ground and satellite data based on cloud conditions.

[0100] In this embodiment, firstly, a cloud density distribution map (such as cloud optical thickness retrieved from infrared cloud images) is obtained from meteorological satellite data and aligned with the generated surface reflectance correction map using the same spatial grid. Secondly, a weighted overlay algorithm is used to fuse the two types of data: areas with high cloud density are given higher weights (e.g., 70%) from satellite data, while areas with significant changes in surface reflectance are given higher weights (e.g., 30%) from the correction value. For example, if the satellite cloud density of a certain grid is 0.8 (0-1 normalized value) and the reflectance correction value is -0.05, the fused feature value is 0.8×0.7+(-0.05)×0.3=0.545. Finally, a fused feature vector matrix with dynamic reflectance compensation is generated as the core input for cloud movement rate prediction.

[0101] 204. Couple the cloud density distribution and the surface reflectance correction value in the fused feature vector with parameters according to the time series and spatial grid to establish a prediction equation for the movement rate of abrupt cloud layers above the photovoltaic array;

[0102] In step 204, the motion rate prediction equation refers to the mathematical model describing the relationship between cloud motion velocity and multi-source observation parameters. Parameter coupling refers to the process of jointly modeling spatiotemporal data under physical constraints.

[0103] In this embodiment, firstly, the cloud density distribution and surface reflectance correction value in the fused feature vector are coupled according to a time series (e.g., one frame per minute) and a spatial grid (e.g., a 100m × 100m grid) to establish the correlation between the two over time (e.g., for every 0.1 increase in cloud density, the reflectance correction value decreases by 0.02). Secondly, a spatiotemporal prediction model (e.g., ConvLSTM) is trained based on historical data, with the fused feature vectors of the current and previous time steps as input, outputting the future trend of cloud density change on the spatial grid. Finally, based on the spatial gradient of the density change trend (e.g., density increases from west to east) and the time-shifting pattern, a prediction equation for the movement rate of abruptly changing clouds above the photovoltaic array is constructed, for example, movement rate = density gradient on the west side × time decay coefficient.

[0104] 205. Dynamically match the time variation parameter of cloud density distribution in the moving rate prediction equation with the remaining time, and generate a short-term power prediction sequence adapted to the sudden cloud scenario by using the product relationship between cloud moving rate and remaining time in the time variation parameter.

[0105] In step 205, the product relationship refers to the integral relationship between the cloud's movement distance and its rate of movement and the remaining time. Dynamic matching refers to the closed-loop process of adjusting the remaining time prediction based on real-time observations.

[0106] In this embodiment, firstly, the time-varying parameters of cloud density distribution (such as the density gradient change rate) are extracted from the cloud movement rate prediction equation and dynamically matched with the remaining time of the cloud cluster predicted in step two (e.g., 120 seconds remaining before the cloud cluster reaches the core area). Secondly, based on the product relationship between the movement rate (e.g., 2 m / s) and the remaining time in the time-varying parameters, the expected dwell time of the cloud in each region of the photovoltaic array (e.g., 60 seconds in the core area) is calculated. Finally, combined with the power-light intensity response curve of the photovoltaic module (e.g., a 10% decrease in light intensity leads to an 8% reduction in output), a short-term power prediction sequence adapted to abrupt cloud formation scenarios is generated, for example, the output linearly decreases from 1MW to 0.7MW within the next 5 minutes, providing a scheduling basis for the energy management system.

[0107] Here is a specific example:

[0108] In a smart dispatch scenario of a 500MW photovoltaic power plant, at 14:23 one afternoon, the monitoring system, through an array of 128 light sensors deployed in a square area, captured a sudden change in light intensity caused by a rapidly moving cumulonimbus cloud cluster in the northwest direction. The S25 sensor in block 5 of the array was the first to detect an 85% light intensity attenuation at a 70-degree azimuth angle, with the scattering angle distribution exhibiting typical characteristics of thick clouds. Based on this, the system calculated that the surface reflectance correction value for this area should be adjusted from the baseline of 0.23 to 0.18. The system then fused the correction value with real-time cloud images from the Fengyun-4 satellite to generate a feature vector containing cloud density gradients and reflectance compensation weights, showing a 3km... 2 The cloud core advanced southeastward at a speed of 18 km / h. By coupling time-series data, a motion rate prediction equation was established, predicting that the cloud layer would cover the core power generation area of ​​the power station in 9 minutes. Based on this, the dispatch center generated a 15-minute power prediction sequence: the current output of 140MW would suddenly drop to 55MW at 14:32, and recover to 120MW at 14:47. This prediction result triggered the frequency regulation mode switch of the energy storage system in advance, switching the 50MW lithium battery pack from constant power charging to fast response standby mode, and simultaneously sending an output fluctuation warning to the grid. In actual operation, the cloud coverage time was only ±1 minute from the prediction, and the power prediction accuracy was improved to 92%, effectively mitigating the impact of sudden cloud formations on the regional power grid.

[0109] In summary, steps 201 to 205 achieve spatiotemporal collaborative sensing and dynamic compensation prediction of photovoltaic power surges under cloudy weather conditions. By analyzing cloud scattering characteristics in real time using a photosensitive sensor array and dynamically correcting surface reflectivity parameters, the limitations of traditional methods in characterizing the spatial heterogeneity of cloud shadow occlusion are overcome. Based on the spatiotemporal weighted fusion of satellite cloud density and ground scattering compensation, a cloud shadow movement rate prediction equation is constructed to address the problem of trajectory deviation accumulation in cloud splitting / merging scenarios. Through a dynamic matching mechanism for remaining time and iterative optimization of multi-source features, a high spatiotemporal resolution power prediction sequence is generated. This sequence is then coupled with grid dispatch parameters to generate active and reactive power coordinated control commands, forming a closed-loop control link of "cloud dynamic sensing - rate prediction - power compensation - grid response." This effectively eliminates prediction curve lag and smoothing effects, ensuring the grid's rapid adjustment capability to sudden power surges.

[0110] In some embodiments, step 204, which involves parametrically coupling the cloud density distribution and surface reflectance correction values ​​in the fused feature vector according to time series and spatial grid, to establish a prediction equation for the movement rate of abruptly changing clouds above the photovoltaic array, includes:

[0111] 301. Divide the cloud density distribution in the fused feature vector into continuous time windows according to the time series, and extract the parameter change trend of cloud density distribution over time in each time window;

[0112] In step 301, the time window refers to dividing a continuous time series into analysis units of fixed duration. The parameter change trend refers to the rate of increase / decrease and fluctuation characteristics of cloud density over time.

[0113] In this embodiment, firstly, the cloud density distribution data in the fused feature vector is divided into continuous time windows of fixed duration (e.g., 5 minutes), with each window containing density distribution snapshots at multiple time points. Secondly, for the density data within each time window, a moving average or linear regression algorithm is used to extract its time-varying trend parameters (e.g., density increasing by 0.1 units every 10 seconds). Finally, the trend parameters (e.g., rate of change, fluctuation amplitude) of each time window are recorded to form a dynamic cloud feature sequence in the time dimension, which serves as input for spatiotemporal coupling analysis.

[0114] 302. Divide the surface reflectance correction value into multiple grid cells according to the spatial grid, and calculate the spatial difference parameter of the surface reflectance correction value between adjacent grid cells;

[0115] In step 302, a grid cell refers to dividing the photovoltaic array area into 1km × 1km spatial analysis units. The spatial difference parameter refers to the absolute value of the difference in surface reflectance correction values ​​between adjacent grids.

[0116] In this embodiment, firstly, the surface reflectance correction value is mapped to the corresponding grid cells according to the spatial grid division of the photovoltaic array (e.g., a 100m × 100m grid), ensuring that each cell corresponds to a reflectance correction value. Secondly, the reflectance difference parameter between adjacent grid cells is calculated (e.g., if the correction value for grid A is -0.05 and the adjacent grid B is -0.03, then the difference is 0.02). Next, the maximum direction and magnitude of reflectance change are extracted using a spatial gradient algorithm (e.g., the Sobe l operator), generating a spatial difference parameter matrix between grid cells, which is used to characterize the local abrupt changes in surface reflectance.

[0117] 303. Couple the parameter change trend with the spatial difference parameter, and generate a spatiotemporal joint change weight by multiplying the cloud density change rate within the time window with the surface reflectance difference amplitude of the spatial grid unit.

[0118] In step 303, the spatiotemporal joint change weight refers to the composite influence coefficient that integrates the intensity of temporal and spatial changes. The product relationship refers to the coupling calculation of the rate of change in the time dimension and the magnitude of spatial differences.

[0119] In this embodiment, firstly, the cloud density change rate within the time window (e.g., density increases by 0.1 units every 10 seconds) is coupled with the reflectance difference parameter between adjacent grids (e.g., difference of 0.02). Secondly, a spatiotemporal joint change weight is generated through a product relationship (e.g., change rate 0.1 × difference 0.02 = 0.002). A larger weight value indicates that the cloud movement in that area is more significantly affected by abrupt changes in surface reflectance. Finally, the weight values ​​are recorded according to the time window and grid cells to form a spatiotemporal weight distribution map, which is used to dynamically adjust the parameter matching of the cloud movement model.

[0120] 304. Based on the spatiotemporal joint change weights, dynamically match the cloud density distribution and the surface reflectance correction value within the current time window to construct a parameter change model for the cloud movement direction and speed;

[0121] In step 304, the parameter variation model refers to the mathematical formula describing the relationship between the direction and velocity of cloud movement. Dynamic matching refers to the optimization process of adjusting model parameters according to real-time weights.

[0122] In this embodiment, firstly, based on spatiotemporal joint change weights, the cloud density distribution and surface reflectance correction values ​​within the current time window are dynamically matched: in areas with high weights, the causal relationship between cloud density changes and surface reflectance differences is preferentially associated. Secondly, the dominant direction of cloud movement (e.g., northeastward in areas with high weights) is extracted using vector synthesis methods (such as principal component analysis), and a parametric model of the movement speed (e.g., speed = rate of change × spatial difference × calibration coefficient) is constructed by combining the product relationship between the rate of change and spatial differences. Finally, a cloud motion model with direction and speed parameters is output, providing a computational framework for speed prediction.

[0123] 305. By using the proportional relationship between the rate of change of cloud density and the magnitude of the difference in surface reflectivity in the parameter change model, the movement rate parameter of the target cloud above the photovoltaic array is calculated, and the movement rate prediction equation is obtained.

[0124] In step 305, the proportional relationship refers to the dynamic correlation between changes in cloud density and differences in surface reflection. The movement rate prediction equation refers to the optimized formula ultimately used to calculate the movement speed of the cloud front.

[0125] In this embodiment, firstly, the proportional relationship between the rate of change of cloud density and the difference in surface reflectance is extracted from the parameter variation model (e.g., a rate of 0.1 units / second corresponds to a difference of 0.02, so the ratio is 5:1). Secondly, the movement rate parameter of the target cloud above the photovoltaic array is derived based on the proportional relationship (e.g., rate = proportional coefficient × difference value), and substituted into the actual parameters in the spatiotemporal weight distribution map to generate a movement rate prediction equation (e.g., rate = 5 × 0.02 = 0.1 m / s). Finally, the accuracy of the equation is verified through multiple sets of historical data, and a corrected movement rate prediction equation is output to calculate the remaining time and coverage area of ​​the cloud reaching the photovoltaic core area.

[0126] Here is a specific example:

[0127] In a smart operation and maintenance scenario of a 200MW photovoltaic power plant, the monitoring system detected a rapidly moving cluster of fragmented cumulus clouds in the southeast direction one afternoon. Using an array of 64 photosensitive sensors deployed in the northeast quadrant of the power plant, the system continuously tracked the cloud dynamics in 5-minute time windows: between 14:00 and 14:05, the sensors in block S17 detected a sudden increase in cloud density from 30% to 65%, corresponding to a decrease in the surface reflectance correction value from 0.25 to 0.18. Meanwhile, the reflectance correction value of the adjacent block S18 differed by 0.07, forming a significant spatial gradient. The system coupled the temporal cloud density increase rate (7% / minute) with the spatial reflectance difference parameter to generate a spatiotemporal joint weighting coefficient of 0.85 pointing southwest. Combined with the power plant's 500m × 500m spatial grid division, the reflectance difference band between grid cells S17 and S21 was calculated to extend at a speed of 12m / s, constructing a movement model of the cloud cluster deflected 15 degrees southwest. In actual observation, when the cloud cluster covered the core power generation area of ​​the power station at 14:20, the error between the system's predicted movement speed of 18 m / s based on the spatiotemporal parameter model and the actual value measured by the meteorological radar was less than 3%, accurately predicting that the edge of the cloud layer would leave the No. 7 array area 8 minutes later. Based on this, the dispatch system activated the energy storage frequency regulation unit 5 minutes in advance, smoothing out the sharp drop in power of 42MW, improving the power station's AGC control accuracy to 98%, and effectively coping with the risk of power grid frequency fluctuations caused by this sudden cloud passage.

[0128] In summary, steps 301 to 305 achieve spatiotemporal joint dynamic modeling and refined parameter calculation for cloud movement rate prediction. By segmenting time windows and dividing spatial grids, the spatiotemporal correlation characteristics between cloud density distribution and surface reflectivity are analyzed, and a coupling mechanism between the temporal dimension parameter change trend and the spatial dimension difference magnitude is constructed. Based on a dynamic matching strategy of spatiotemporal joint change weights, the product effect of cloud density change rate and surface reflectivity difference is integrated to establish a multi-parameter co-evolution model of cloud movement direction and speed. Finally, by using the nonlinear proportional relationship between cloud density change rate and surface reflectivity difference magnitude, the dynamic movement rate parameters of the target cloud cluster above the photovoltaic array are calculated. This overcomes the problem of accumulated trajectory prediction bias in traditional methods for cloud splitting / merging scenarios, providing high-precision support for the spatiotemporal evolution law of cloud shadow occlusion in power prediction under multi-cloud abrupt change scenarios.

[0129] In some embodiments, step 102, which involves deploying a light sensor array at the edge of the photovoltaic array to detect cloud thickness and aerosol density in real time, and using the nonlinear model to calculate the light intensity attenuation gradient in the local shadowed area to predict the remaining time for the target cloud to reach the core area of ​​the photovoltaic array, includes:

[0130] 401. Deploy a photosensitive array at the edge of the photovoltaic array to collect scattered light intensity distribution data at multiple locations below the cloud layer in real time, and extract the scattered light angle distribution parameters and light intensity attenuation ratio parameters from the scattered light intensity distribution data;

[0131] In step 401, the scattered light intensity distribution data refers to the information on changes in light intensity at different spatial locations obtained through an optical sensor. The scattered light angle distribution parameter describes the characteristics of the scattered intensity distribution in different directions after light passes through clouds. The light intensity attenuation ratio parameter indicates the degree of attenuation of light intensity in a specific direction due to cloud obstruction.

[0132] In this embodiment, firstly, a photosensitive array deployed at the edge of the photovoltaic array collects real-time scattered light intensity distribution data at multiple locations below the cloud layer using multi-angle sensors. A hemispherical optical probe records the scattered light intensity values ​​in different directions (e.g., relative light intensity at 0°, 45°, and 90°). Secondly, a calibrated light intensity attenuation model (e.g., the difference between theoretical clear-sky light intensity and actual measured value) is used to calculate the light intensity attenuation ratio parameter at each location (e.g., if the actual light intensity is 70% of the theoretical value, the attenuation ratio is 30%). Simultaneously, angular distribution parameters are extracted through angular distribution statistics (e.g., wide-angle scattering accounts for more than 60%). Finally, the attenuation ratio and angular distribution parameters are stored according to sensor location numbers to form a spatially labeled scattering feature dataset, providing input for cloud parameter inversion.

[0133] 402. Based on the light intensity attenuation ratio parameter and the scattered light angle distribution parameter, the cloud thickness parameter and aerosol density parameter corresponding to each photosensitive unit in the photosensitive array are detected in real time. The cloud thickness parameter is obtained by comparing the light intensity attenuation ratio with a preset occlusion threshold.

[0134] In step 402, the cloud thickness parameter reflects the estimated optical thickness of the cloud in the vertical direction. The aerosol density parameter represents the degree of influence of suspended particulate matter in the air on light scattering. The preset occlusion threshold refers to an empirically determined critical value for light intensity attenuation.

[0135] In this embodiment, firstly, the light intensity attenuation ratio parameter is compared with a preset cloud occlusion threshold (e.g., an attenuation ratio exceeding 20% ​​is considered cloud occlusion) to determine the cloud thickness parameter corresponding to each photosensitive unit (e.g., 30% attenuation corresponds to cloud thickness level 2). Secondly, the aerosol density parameter (e.g., aerosol concentration level 3) is analyzed using the spectral absorption characteristics (e.g., a decrease in scattering intensity at a specific wavelength) in the scattered light angle distribution parameter. Finally, the model is validated using historical data to generate a real-time cloud thickness and aerosol density parameter table for each sensor location, used for local shadow gradient calculation.

[0136] 403. Input the cloud thickness parameter and aerosol density parameter into a nonlinear model, and perform correlation calculation between the dynamic change rate of the cloud thickness parameter and the fluctuation amplitude of the aerosol density parameter to generate the light intensity attenuation gradient of the local shadow area.

[0137] In step 403, the nonlinear model refers to a machine learning model that describes the relationship between cloud parameters and light intensity gradient. The dynamic rate of change refers to the percentage change in cloud thickness per unit time. The fluctuation amplitude refers to the intensity of oscillations in aerosol density over time.

[0138] In this embodiment, firstly, the dynamic change rate of cloud thickness parameters (e.g., thickness increases by 0.5 levels every 10 seconds) and the fluctuation range of aerosol density parameters (e.g., concentration fluctuates within ±0.2 levels) are input into a pre-trained nonlinear model (e.g., a random forest regressor). Secondly, the light intensity attenuation gradient (e.g., gradient value is thickness increase × aerosol fluctuation range) is calculated through the association rules within the model (e.g., the product relationship between thickness increase and aerosol fluctuation). Finally, the light intensity attenuation gradient map of each local region is output, indicating the intensity and direction of shadow diffusion, providing a basis for the generation of cloud movement driving factors.

[0139] 404. Based on the light intensity attenuation gradient, the cloud movement direction parameter in the nonlinear model is weighted and superimposed with the spatial distance of the photovoltaic array core area to generate the driving intensity factor of cloud approaching the core area.

[0140] In step 404, the driving intensity factor refers to a composite parameter that integrates the effects of cloud motion direction and distance. Weighted superposition refers to adjusting the contribution weight of the direction parameter based on spatial distance.

[0141] In this embodiment, firstly, the spatial distance between the core area and the edge photosensitive units is obtained from the photovoltaic array layout model (e.g., the nearest edge unit is 50 meters from the core area). Secondly, the cloud movement direction parameters output by the nonlinear model (e.g., northeast direction weight 0.7) and the spatial distance parameters are weighted and superimposed according to preset weights (direction weight 60%, distance weight 40%) to generate a driving intensity factor (e.g., 0.7×0.6+50×0.4=20.42). Finally, the priority of cloud movement is adjusted according to the magnitude of the driving intensity factor; a higher factor indicates a greater risk of the cloud approaching the core area.

[0142] 405. The cloud movement speed is dynamically corrected by the driving intensity factor, and the ratio of the dynamically corrected cloud movement speed to the spatial distance of the photovoltaic array core area is calculated to predict the remaining time for the target cloud to reach the photovoltaic array core area.

[0143] In step 405, dynamic correction refers to the process of adjusting the kinematic model parameters based on real-time observation data. Ratio calculation refers to calculating the remaining time by dividing the distance parameter by the rate parameter.

[0144] In this embodiment, firstly, the initial cloud movement speed is dynamically corrected using a driving intensity factor (e.g., an original speed of 1 m / s corresponds to a corrected speed of 1.2 m / s with a factor of 20.42). Secondly, the remaining time for the target cloud cluster to arrive is calculated based on the corrected speed and the remaining spatial distance to the core area (e.g., 100 meters) (100 / 1.2 ≈ 83 seconds). Finally, the remaining time is linked to the light intensity attenuation gradient to generate a time-stamped cloud shadow coverage early warning sequence, providing key parameters for short-term photovoltaic power output forecasting and grid dispatching.

[0145] Here is a specific example:

[0146] In a smart forecasting scenario at a 300MW photovoltaic power station along the coast, a rapidly approaching cumulonimbus cloud was detected by the photosensitive array on the northwest edge of the power station at noon on a certain day. The 32 photosensitive units deployed in area S15 of the array boundary collected real-time data showing a 78% attenuation of scattered light intensity at a 70-degree azimuth angle. The scattering angle distribution exhibited typical characteristics of a thick cloud layer. Based on this, the system calculated that the cloud thickness parameter in this area exceeded level 6.5 (the preset threshold was level 5). Simultaneously, it detected an abnormal increase in aerosol density to 220 μg / m³ due to the rise of sea salt spray. 3 The nonlinear model correlates the dynamic change in cloud thickness (increasing by 1.2 levels per minute) with aerosol fluctuations of 30 μg / m³. 3 Correlation was performed to generate a light intensity attenuation gradient band pointing southeast, with its intensity decreasing from 0.8 kW / m² over 3 minutes. 2 Expanded to 2.3kW / m 2 Based on the spatial distance of 2.8 kilometers between the cloud front and the core area of ​​the power station, the model calculated a southeast-direction driving intensity factor of 0.78 (out of a maximum of 1.0). Using this factor, the system corrected the initial cloud movement speed provided by the weather radar from 12 m / s to 15 m / s, predicting that the core area of ​​the cumulonimbus cloud would cover the power station's No. 7 array in 11 minutes. In actual operation, this prediction-triggered energy storage system activated the 35MW frequency regulation unit 8 minutes in advance, effectively buffering the impact of the power generation suddenly dropping from 210MW to 85MW when the cloud arrived. The prediction time error was controlled within ±90 seconds, improving the power station's AGC regulation response speed by 40% and successfully maintaining the grid frequency within the safe range of 49.8-50.2Hz.

[0147] In summary, steps 401 to 405 achieve real-time perception and accurate prediction of the remaining time for the dynamic approach of cloud shadows. By analyzing the angle and attenuation characteristics of scattered light in real time through a photosensitive array, the spatial distribution of cloud thickness and aerosol density is dynamically calculated, overcoming the bottleneck of traditional methods in characterizing the microscopic physical parameters of clouds. Based on the nonlinear correlation between the rate of change of cloud thickness and the amplitude of aerosol fluctuations, a gradient evolution model of light intensity attenuation in local shadow areas is constructed. Combined with a weighted coupling mechanism of cloud movement direction and spatial distance to the core area, a driving intensity factor is generated to dynamically correct the cloud movement speed. Finally, through the dynamic matching relationship between cloud movement speed and spatial distance, a closed-loop feedback mechanism for the approach of cloud shadows to the core area is established, effectively eliminating the prediction deviation of the remaining time in cloud splitting / merging scenarios. This provides high-precision support for the spatiotemporal evolution law of dynamic cloud shadow shading for photovoltaic power surge warning, ensuring the rapid response capability of the power grid dispatch system to sudden power fluctuations.

[0148] In some embodiments, step 403, which involves inputting the cloud thickness parameter and aerosol density parameter into a nonlinear model, and performing a correlation calculation between the dynamic change rate of the cloud thickness parameter and the fluctuation amplitude of the aerosol density parameter to generate a light intensity attenuation gradient in the local shadow region, includes:

[0149] 501. Divide the cloud thickness parameter into continuous time segments according to the time window, and extract the rate of change of the cloud thickness parameter in each time segment as the dynamic rate of change;

[0150] In step 501, the dynamic rate of change refers to the percentage change in cloud thickness per unit time. A time window refers to dividing a continuous time series into analysis units of fixed duration.

[0151] In this embodiment, firstly, the real-time collected cloud thickness parameters are divided into continuous time segments of fixed duration (e.g., 5 minutes), with each time segment containing thickness data from multiple moments. Secondly, the rate of change of the cloud thickness parameters within each time segment (e.g., thickness increasing by 0.2 units every 10 seconds) is extracted using linear regression or moving average algorithms as the dynamic rate of change. Finally, the dynamic rate of change for each time segment is stored according to timestamps and sensor location numbers to form a spatiotemporal dynamic parameter table for subsequent coupling analysis with aerosol parameters.

[0152] 502. Calculate the aerosol density fluctuation amplitude at each optical sensor location based on the spatial distribution difference of the aerosol density parameter at adjacent optical sensor locations.

[0153] In step 502, the aerosol density fluctuation amplitude refers to the degree of difference in aerosol concentration between adjacent sensors. Spatial distribution difference refers to the non-uniform distribution characteristics of aerosol parameters within a region.

[0154] In this embodiment, firstly, based on the spatial distribution of the optical sensor array, the difference in aerosol density parameters between adjacent sensor locations is calculated (e.g., density at location A is 3 units, adjacent location B is 2.5 units, the difference is 0.5 units). Secondly, the aerosol density fluctuation amplitude at each location is calculated using sliding window statistics (e.g., the average difference between three adjacent groups of sensors) (e.g., fluctuation amplitude is ±0.3 units). Finally, the fluctuation amplitude is bound to the sensor location to generate an aerosol spatial fluctuation feature map, which is used to characterize the abrupt change characteristics of local aerosol concentration.

[0155] 503. Based on the dynamic response rule of photovoltaic power output attenuation with cloud thickness parameter and aerosol density in the nonlinear model, the dynamic change rate is multiplied with the aerosol density fluctuation amplitude to generate a superimposed weighting factor.

[0156] In step 503, the superposition weighting factor refers to the composite weighting parameter that integrates the effects of cloud motion and aerosols. The product operation refers to the coupled calculation of the rate of change in the time dimension and the amplitude of spatial fluctuations.

[0157] In this embodiment, firstly, based on the photovoltaic power output attenuation rules defined in a nonlinear model (such as a neural network) trained on historical data (e.g., a 5% decrease in power output for every 1 unit increase in cloud thickness, and a 2% decrease in power output for every 1 unit fluctuation in aerosol density), the dynamic change rate (e.g., 0.2 units / second) is multiplied with the aerosol density fluctuation amplitude (e.g., 0.3 units) from step two to generate a superimposed weighting factor (e.g., 0.2 × 0.3 = 0.06). Secondly, the combined influence weights of cloud layer and aerosol on power output attenuation are adjusted according to the magnitude of the weighting factor; a larger factor indicates a more significant synergistic effect between the two. Finally, a list of weighting factors for each sensor location is output for weighted fusion calculation.

[0158] 504. The spatial distribution gradient of cloud thickness parameter and the adjacent differences of aerosol density parameter within the current time segment are weighted and fused according to the superposition weighting factor to generate a combined attenuation factor;

[0159] In step 504, the combined attenuation factor refers to the combined influence parameter of the fusion of cloud gradient and aerosol differences. Weighted fusion refers to adjusting the contribution ratio of different parameters according to weighting factors.

[0160] In this embodiment, firstly, the spatial distribution gradient of the cloud thickness parameter within the current time segment is calculated using a spatial gradient algorithm (such as the Sobe operator). The gradient is calculated as 0.1 units / meter for thickness increasing from west to east. Secondly, the adjacent differences in the aerosol density parameter (e.g., 0.5 units) are weighted and fused with the superposition weighting factor (0.06) from step three (e.g., gradient × weighting factor + difference × weighting factor = 0.1 × 0.06 + 0.5 × 0.06 = 0.036). Finally, a combined attenuation factor (e.g., 0.036) is generated to quantify the local shadow attenuation intensity under the combined effect of clouds and aerosols, providing input for the calculation of the light intensity attenuation gradient.

[0161] 505. Input the combined attenuation factor into the dynamic response rule of photovoltaic power output attenuation in the nonlinear model, and output the light intensity attenuation gradient of the local shadow area through the linear combination ratio of cloud thickness parameter and aerosol density in the dynamic response rule.

[0162] In step 505, the linear combination ratio refers to the weighted ratio of the contribution of clouds and aerosols to the power output attenuation. The dynamic response rule refers to the mathematical model of how photovoltaic power output changes with environmental parameters.

[0163] In this embodiment, firstly, the combined attenuation factor is input into the photovoltaic output attenuation dynamic response rule of the nonlinear model. Based on a preset linear combination ratio of cloud thickness and aerosol density (e.g., cloud weight 70%, aerosol weight 30%), the light intensity attenuation gradient of the local shaded area is calculated (e.g., 0.036 × 70% + 0.036 × 30% = 0.036). Secondly, combined with the power-light intensity response curve of the photovoltaic module (e.g., a 10% decrease in light intensity leads to an 8% decrease in output), the attenuation gradient is converted into a predicted output decrease value (e.g., a gradient of 0.036 corresponds to a 2.88% output decrease). Finally, a light intensity attenuation gradient map with spatial location markers is output, identifying the shaded coverage areas in the photovoltaic array that need to be prioritized for scheduling, providing a decision-making basis for grid power balance.

[0164] Here is a specific example:

[0165] In a smart control scenario for a 200MW photovoltaic power plant to cope with a sandstorm, the monitoring system detected a rapidly moving mixed cloud of sand in the northwest direction one afternoon. Forty-eight sets of optical sensors deployed in area S09 at the power plant boundary continuously tracked the cloud in 3-minute time windows, detecting a sudden increase in cloud thickness from level 4 (light shading) at 14:05 to level 6 (severe shading) at 14:08, a dynamic change rate of 0.67 levels per minute. Simultaneously, adjacent sensors S10 to S12 detected an aerosol density increase from 180 μg / m³. 3 Fluctuation up to 260 μg / m 3 80 μg / m 3The system uses a photovoltaic power output attenuation model to correlate cloud thickening rate with aerosol fluctuation amplitude, generating a superposition weighting factor of 0.72 along the northwest-southeast axis. Based on this factor, the system fuses the spatial gradient of cloud thickness (1.2 levels per 100 meters) and adjacent aerosol density difference parameters in region S09 to construct an intensity level of 4.8 μg / m³. 3 The combined attenuation factor. After inputting into the model, combined with historical data showing that each increase of 1 level in cloud thickness corresponds to a 12% decrease in output, and each increase of 50 μg / m³ in aerosol concentration... 3 With an additional 3% attenuation response rule, the output is directed at 9.2 kW / m² for power plant array 5. 2 Light intensity attenuation gradient zone. Based on this, the system activates the 20MW energy storage unit in the area 6 minutes in advance, buffering the predicted 153MW power drop to 126MW. When the actual cloud and sand cover occurred, the power output fluctuation was less than 5% of the prediction error, successfully maintaining the grid frequency within the acceptable range and avoiding the risk of AGC control instability caused by a sudden drop in visibility.

[0166] In summary, steps 501 to 505 achieved dynamic coupling modeling and accurate calculation of the light intensity attenuation gradient in multi-cloud scenarios. Through spatiotemporal joint analysis of the dynamic change rate of cloud thickness and the spatial fluctuation amplitude of aerosols, a superposition weighting factor generation mechanism based on nonlinear response rules was constructed. A weighted combination of cloud thickness gradient and adjacent aerosol differences was integrated to generate a combined attenuation factor characterizing the intensity of local shadow evolution. Based on the dynamic response law of photovoltaic power output to cloud-aerosol parameters in the nonlinear model, the mapping relationship between the light intensity attenuation gradient and the linear combination of multiple parameters was analyzed. This overcomes the limitations of traditional methods in decoupling the modeling of cloud micro-physical parameters and macro-attenuation effects, forming a closed-loop optimization link of "parameter calculation - weight superposition - attenuation prediction," significantly improving the spatiotemporal resolution of light intensity abrupt gradient prediction and providing refined cloud dynamic evolution parameter support for photovoltaic power abrupt change early warning.

[0167] In some embodiments, step 103 involves inputting the short-term power prediction sequence into the energy management system, and combining it with the voltage deviation and reactive power compensation demand parameters at the grid connection point of the photovoltaic power station to generate a dual-objective scheduling instruction that balances active power regulation and reactive power control, including:

[0168] 601. Input the short-term power prediction sequence into the energy management system and extract the active power prediction parameters within the current time window. The active power prediction parameters characterize the power output fluctuation characteristics of the photovoltaic array after being affected by cloud cover.

[0169] In step 601, the active power prediction parameter refers to the predicted value of the change in power generation of the photovoltaic power station due to the influence of clouds during a specific future period. The output fluctuation characteristics describe the amplitude, frequency, and trend of the photovoltaic output power fluctuation over time.

[0170] In this embodiment, firstly, the energy management system receives a short-term power prediction sequence and extracts the active power prediction parameters for the current time window (e.g., the next 15 minutes). By using sliding window mean or peak statistics, the power output fluctuation characteristics of the photovoltaic array affected by cloud cover during this period are calculated (e.g., power output fluctuating between 80kW and 120kW). Secondly, the prediction parameters are compared with real-time power output data to verify the prediction accuracy and correct deviations. Finally, an active power prediction curve with a confidence interval is generated as the basic input for scheduling instructions.

[0171] 602. Real-time collection of voltage deviation parameters at the grid connection point of photovoltaic power plants, analysis of voltage deviation direction and amplitude, and classification of voltage compensation requirement levels;

[0172] In step 602, the voltage deviation parameter refers to the percentage difference between the actual voltage at the grid connection point and the rated value. Voltage compensation requirement levels are categorized by control priority based on the degree of deviation, including observation level, early warning level, and emergency control level.

[0173] In this embodiment, firstly, voltage deviation parameters (e.g., actual voltage 225V, rated voltage 230V, deviation -5V) are collected in real time using a voltage sensor at the grid connection point. Secondly, the direction of the deviation (negative indicates low voltage) and the absolute value of the magnitude (5V) are analyzed, and compensation demand levels are divided according to preset thresholds (e.g., a deviation of 5V corresponds to level 2). Finally, the voltage deviation parameters are bound to a compensation level mapping table to generate the current voltage compensation demand priority (e.g., level 2 requires medium-intensity compensation) for use in reactive power regulation calculations.

[0174] 603. Based on the reactive power compensation requirement parameters and the current voltage deviation parameters, calculate the reactive power compensation adjustment intensity parameters, which reflect the reactive power adjustment level required for grid connection point voltage stability;

[0175] In step 603, the reactive power compensation adjustment intensity parameter refers to the reactive power adjustment level required to maintain voltage stability. The adjustment level is dynamically calculated based on the voltage deviation, including the type of compensation equipment, switching capacity, and response speed requirements.

[0176] In this embodiment, firstly, the reactive power compensation requirement parameters (e.g., power factor ≥ 0.9) of the grid connection point are read from the power grid operation rule base. Secondly, combining the voltage compensation requirement level (level 2) from step 602 and the current power factor (e.g., 0.85), the required reactive power compensation adjustment intensity parameters (e.g., 50kVar compensation) are calculated using a lookup table method or linear interpolation. Finally, the action level of the reactive power equipment is determined based on the adjustment intensity parameters (e.g., connecting a set of 50kVar capacitors) to ensure voltage stability and power factor compliance.

[0177] 604. By integrating the active power prediction parameters, voltage compensation demand level and reactive power compensation adjustment intensity parameters, the active power adjustment demand and reactive power control demand are superimposed and balanced through dynamic weight allocation rules to generate a dual-objective scheduling factor.

[0178] In step 604, the dual-objective scheduling factor refers to the composite control coefficient that coordinates active power tracking and voltage stability. Dynamic weight allocation refers to adjusting the priority weights of the two objectives based on the real-time state of the power grid.

[0179] In this embodiment, firstly, the active power prediction parameters (e.g., fluctuation range 80-120kW) and the reactive power compensation regulation intensity parameters (50kVar) from step 603 are input into the dynamic weight allocation model. Secondly, weighting rules are set according to the current grid state: if the voltage deviation level is high (e.g., level 3), reactive power control is assigned a 70% weight; if the output fluctuation is severe (e.g., fluctuation amplitude exceeds 30%), active power regulation is assigned a 60% weight. Finally, a dual-objective scheduling factor is generated through weighted superposition (e.g., 50kVar × 70% + 120kW × 40%) to quantify the coordinated adjustment ratio of active and reactive power.

[0180] 605. By using the superposition and balance ratio of active power demand and reactive power demand in the dual-objective scheduling factor, output a dual-objective scheduling command that takes into account both active power regulation and reactive power control.

[0181] In step 605, the superimposed balance ratio refers to the weighted allocation relationship between active and reactive power regulation demands. The dual-objective dispatch command refers to a coordinated control command that simultaneously includes power adjustment and reactive power compensation, and must meet grid security constraints and economic requirements.

[0182] In this embodiment, firstly, the proportional relationship in the dual-objective scheduling factors (e.g., active power weight 40%, reactive power weight 60%) is analyzed to determine the priority. Secondly, if active power fluctuations need to be smoothed first, the photovoltaic inverter is instructed to limit its output according to the predicted curve (e.g., limit output to 100kW ± 10%); if voltage deviations need urgent correction, the SVG device is instructed to generate an additional 60kVar of reactive power. Finally, specific operation instructions are generated (e.g., "inverter limit to 110kW, add 60kVar capacitor") and simultaneously sent to the photovoltaic power station control system for execution, ensuring grid operation safety and power quality.

[0183] Here is a specific example:

[0184] In a smart control scenario for sandstorm response at a 200MW photovoltaic power station in North China, the monitoring system detected a rapidly moving mixed cloud of sand in the northwest direction one afternoon. Forty-eight sets of optical sensors deployed in area S09 at the power station boundary continuously tracked the cloud layer in 3-minute time windows, detecting a sharp increase in cloud thickness from level 4 (light shading) at 14:05 to level 6 (severe shading) at 14:08, a dynamic change rate of 0.67 levels per minute. Simultaneously, adjacent sensors S10 to S12 detected an aerosol density increase from 180 μg / m³. 3 Fluctuation up to 260 μg / m 3 80 μg / m 3 The system uses a photovoltaic power output attenuation model to correlate cloud thickening rate with aerosol fluctuation amplitude, generating a superposition weighting factor of 0.72 along the northwest-southeast axis. Based on this factor, the system fuses the spatial gradient of cloud thickness (1.2 levels per 100 meters) and adjacent aerosol density difference parameters in region S09 to construct an intensity level of 4.8 μg / m³. 3 The combined attenuation factor. After inputting into the model, combined with historical data showing that each increase of 1 level in cloud thickness corresponds to a 12% decrease in output, and each increase of 50 μg / m³ in aerosol concentration... 3 With an additional 3% attenuation response rule, the output is directed at 9.2 kW / m² for power plant array 5. 2 Light intensity attenuation gradient zone. Based on this, the system activates the 20MW energy storage unit in the area 6 minutes in advance, buffering the predicted 153MW power drop to 126MW. When the actual cloud and sand cover occurred, the power output fluctuation was less than 5% of the prediction error, successfully maintaining the grid frequency within the acceptable range and avoiding the risk of AGC control instability caused by a sudden drop in visibility.

[0185] In summary, steps 601 to 605 achieve dynamic optimization and closed-loop control of active and reactive power coordinated scheduling of photovoltaic power plants. By integrating short-term power forecast parameters and real-time voltage deviation data at the grid connection point, a multi-objective coupled model of active power output fluctuations and reactive power compensation needs is constructed. Based on dynamic weight allocation rules, power regulation needs are superimposed and balanced to generate a dual-objective scheduling factor that takes into account both voltage stability and power tracking. Combined with grid operation rule constraints, and through the proportional relationship between active and reactive power needs within the factor, dual-objective scheduling instructions with controllable priority and magnitude are output, forming a closed-loop control mechanism of "power prediction - voltage sensing - dynamic balancing - instruction generation". This overcomes the limitations of traditional scheduling systems in single-objective adjustment under sudden change scenarios, ensuring the grid's dynamic voltage stability and rapid power tracking capability under cloud-induced disturbances.

[0186] In some embodiments, step 104, which involves dynamically dividing the confidence interval for short-term power prediction based on the remaining time, inputting meteorological satellite data into a multi-scale meteorological fusion architecture, and adjusting the feature extraction path for different meteorological parameters based on the current cloud density using an attention mechanism, includes:

[0187] 701. Based on the remaining time for the target cloud cluster to reach the core area of ​​the photovoltaic array, divide the remaining time into multiple consecutive time window lengths;

[0188] In step 701, the time window length refers to dividing the remaining time into equal or variable length analysis periods. The remaining time refers to the estimated time for the target cloud to reach the core region of the photovoltaic array.

[0189] In this embodiment, firstly, based on the remaining time (e.g., 120 seconds) for the target cloud to reach the core area of ​​the photovoltaic array, it is divided into multiple consecutive time windows of fixed duration (e.g., 30 seconds) (e.g., window 1: 0-30 seconds, window 2: 30-60 seconds). If the remaining time is less than an integer window, the length of the last window is adjusted according to the actual remainder (e.g., the window is set to 20 seconds when 20 seconds remain). Secondly, combined with the cloud movement rate (e.g., 2 meters / second), the range of photovoltaic area that the cloud may cover within each time window is calculated (e.g., window 1 covers a distance of 0-60 meters). Finally, a list of time windows and their corresponding coverage distance mapping table are generated to provide a time reference for subsequent confidence interval division.

[0190] 702. Divide the confidence interval range of short-term power prediction according to the relationship between the time window length and the cloud movement rate;

[0191] In step 702, the confidence interval range refers to the credible fluctuation range of the power prediction results. The cloud movement rate relationship refers to the mathematical correlation rule between the time window length and cloud speed.

[0192] In this embodiment, firstly, based on the relationship between the time window length and the cloud movement rate (e.g., window 1, with a duration of 30 seconds, corresponds to a movement distance of 60 meters), the confidence interval range for short-term power prediction is divided (e.g., the confidence level for window 1 is 80%, and for window 2 it is 60%). Secondly, the sliding window mean method is used to statistically analyze the prediction error distribution of the same time window in historical data, and the confidence interval width is dynamically adjusted (e.g., the confidence interval is widened by ±10% when the error is large). Finally, the upper and lower limits of the confidence interval for each time window are generated (e.g., the predicted power output for window 1 is 100kW ± 8kW), forming a prediction interval list with probability distribution for multi-scale meteorological feature fusion.

[0193] 703. Divide meteorological satellite parameters into multi-scale meteorological feature layers, input the multi-scale meteorological fusion architecture, and extract cloud density parameters and wind speed parameters at different scales;

[0194] In step 703, the multi-scale meteorological feature layer refers to the hierarchical division of satellite data according to spatial resolution (e.g., 5km / 1km / 500m). The multi-scale meteorological fusion architecture refers to a neural network model that integrates meteorological data of different resolutions.

[0195] In this embodiment, firstly, meteorological satellite parameters are divided into multi-scale meteorological feature layers according to spatial resolution (e.g., high-resolution layer: cloud density with a 100-meter grid; medium-resolution layer: wind speed field with a 500-meter grid). Secondly, a multi-scale meteorological fusion architecture (e.g., a pyramid-shaped neural network) is input, and cloud density parameters (e.g., density value of 0.8 for the high-resolution layer) and wind speed parameters (e.g., wind speed of 5 m / s for the medium-resolution layer) at different scales are extracted through convolutional layers. Finally, the parameter matrices of each scale feature layer are output and aligned with the time window to ensure spatiotemporal consistency.

[0196] 704. Based on the confidence interval range, calculate the dynamic weight values ​​of cloud density parameters and wind speed parameters, and adjust the priority of cloud density parameters and wind speed parameters in the feature extraction path through the dynamic weight values;

[0197] In step 704, the dynamic weight value refers to the importance coefficient of meteorological parameters that are dynamically adjusted according to the forecast lead time. The feature extraction path priority refers to the processing order of different feature channels in the neural network.

[0198] In this embodiment, firstly, dynamic weight values ​​for cloud density and wind speed parameters are calculated based on the confidence interval range (e.g., window 1 with a confidence level of 80%) (e.g., density weight = confidence level × 0.7, wind speed weight = confidence level × 0.3). Secondly, in the feature extraction path, parameter priorities are adjusted through a gating attention mechanism (e.g., cloud density weight is increased to 0.8 within a high-confidence window). Finally, a dynamic weight table is generated (e.g., window 1: density weight 0.8, wind speed weight 0.2) to guide the contribution ratio of parameters during feature fusion.

[0199] 705. The cloud density parameter and wind speed parameter in the feature extraction path after priority adjustment are assigned weight ratios according to the confidence interval range. The cloud density parameter and wind speed parameter are weighted and accumulated by the weight value ratio to generate a short-term power prediction sequence adapted to the sudden cloud scenario.

[0200] In step 705, the weight ratio refers to the allocation of the contribution of different meteorological parameters in the final forecast. Weighted accumulation refers to the linear superposition process of proportionally integrating the influence of multiple parameters.

[0201] In this embodiment, firstly, weights are assigned according to the confidence interval range (e.g., density weight 0.8, wind speed weight 0.2 within window 1), and the cloud density parameter (0.8) and wind speed parameter (5 m / s) of the multi-scale feature layer are weighted and accumulated (e.g., 0.8 × 0.8 + 5 × 0.2 = 1.64). Secondly, the weighted value is input into the photovoltaic power output attenuation model (e.g., power-light intensity response curve), and a short-term power prediction value is calculated in conjunction with the current irradiance (e.g., a weighted value of 1.64 corresponds to a 30% decrease in output). Finally, a short-term power prediction sequence adapted to abrupt cloud formation scenarios is generated (e.g., output of 70 kW in window 1, output of 50 kW in window 2), and bound to a confidence interval for use by the energy management system.

[0202] Here is a specific example:

[0203] In a scenario involving the dispatching of a 150MW photovoltaic power plant in response to a sudden thunderstorm, the system detected a rapidly moving cumulonimbus cloud cluster 3 kilometers to the northwest, predicting it would cover the core area of ​​the power plant in 18 minutes. The system divided the remaining time into three time windows: 0-6 minutes, 6-12 minutes, and 12-18 minutes, corresponding to high, medium, and low confidence intervals, respectively. Simultaneously, it processed 1km resolution cloud density data from the Fengyun-4 satellite and 10-meter altitude wind speed data from ground meteorological stations, extracting key parameters including a cloud core density of 85% and a moving wind speed of 22 m / s. In the 0-6 minute high confidence window, the cloud density weight was set to 0.8 (wind speed 0.2), generating a short-term power prediction of 123MW. In the 6-12 minute medium confidence window, the weight was adjusted to 0.6 for density and 0.4 for wind speed, revising the prediction to 107MW. In the 12-18 minute low confidence window, focusing on the impact of wind speed fluctuations, the weight was set to 0.4 for density and 0.6 for wind speed, reducing the prediction to 89MW. When the cloud cover actually arrived 14 minutes ahead of schedule, the system, based on a dynamic weighting model and incorporating the latest radar data in real time, urgently adjusted the predicted value for the 12-18 minute window to 72MW, simultaneously triggering a 35MW energy storage unit to switch from charging to discharging mode. In actual operation, the power plant output dropped from 135MW to 70MW when the cloud cover occurred, with an error of only 2MW from the final prediction. This supported the grid dispatch system in adjusting the regional tie-line power 5 minutes in advance, controlling the frequency deviation within ±0.05Hz, and effectively avoiding the problem of inaccurate power prediction caused by sudden changes in the cloud movement rate.

[0204] In summary, steps 701 to 705 achieve spatiotemporal collaborative optimization and dynamic confidence adaptation for short-term power prediction under multi-cloud abrupt change scenarios. By dividing the remaining time window and extracting multi-scale meteorological features hierarchically, a dynamic weight allocation mechanism for cloud density and wind speed parameters is constructed. The priority of meteorological parameters is adaptively adjusted based on the confidence interval range, overcoming the single-scale limitation of traditional methods in modeling coupled features of abrupt change scenarios. Through multi-scale parameter weighted accumulation and spatiotemporal confidence matching, a power prediction sequence that takes into account both cloud movement rate and wind speed disturbances is generated, forming a closed-loop optimization link of "time window segmentation - multi-scale fusion - dynamic weight iteration." This solves the response lag and error divergence problems caused by the fixed priority of meteorological parameters in traditional prediction models, ensuring the power grid dispatch system's ability to quickly track and maintain stability during short-term power abrupt changes.

[0205] Figure 2 This application provides a schematic diagram of the structure of a photovoltaic power prediction system for a photovoltaic power plant, as shown in the embodiment of the present application. Figure 2 As shown, the system includes:

[0206] Module 21 is used to construct a cloud movement time series library by using historical cloud map sequences of the area where the photovoltaic power station is located, extract the spatiotemporal relationship between cloud trajectory and light change, and establish a nonlinear model of cloud movement speed and photovoltaic power output attenuation.

[0207] Prediction module 22 is used to deploy a light sensor array at the edge of the photovoltaic array to detect cloud thickness and aerosol density in real time, and calculate the light intensity attenuation gradient in the local shadow area in combination with the nonlinear model to predict the remaining time for the target cloud to reach the core area of ​​the photovoltaic array.

[0208] The adjustment module 23 is used to dynamically divide the confidence interval of short-term power prediction according to the remaining time, input meteorological satellite data into the multi-scale meteorological fusion architecture, and adjust the feature extraction path of different meteorological parameters based on the current cloud density through the attention mechanism.

[0209] The generation module 24 is used to fuse the surface reflectance correction value through the adjusted feature extraction path to generate a short-term power prediction sequence adapted to the sudden cloud scene. The surface reflectance correction value is calculated by the scattering characteristics fed back in real time by the optical sensor array.

[0210] Input module 25 is used to input the short-term power prediction sequence into the energy management system, and combine it with the voltage deviation and reactive power compensation demand parameters of the photovoltaic power station grid connection point to generate a dual-objective scheduling instruction that takes into account both active power regulation and reactive power control. The reactive power regulation demand parameters are set by the energy management system according to the power factor requirements of the grid connection point in the power grid operation rules, and are used to constrain the priority and magnitude range of reactive power regulation.

[0211] Figure 2The photovoltaic power prediction system for a photovoltaic power station described above can perform... Figure 1 The implementation principle and technical effects of the photovoltaic power prediction method for a photovoltaic power station described in the illustrated embodiment will not be repeated here. The specific operation methods of each module and unit in the photovoltaic power prediction system for a photovoltaic power station in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0212] In one possible design, Figure 2 The photovoltaic power prediction system for a photovoltaic power station shown in the embodiment can be implemented as a computing device, such as... Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0213] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 32.

[0214] The processing component 32 is used for the above Figure 1 The embodiment described is a method for predicting photovoltaic power in a photovoltaic power plant.

[0215] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0216] Storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0217] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0218] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0219] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0220] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0221] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The embodiment shown illustrates a method for predicting photovoltaic power generation in a photovoltaic power plant.

[0222] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0223] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0224] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0225] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting photovoltaic power output of a photovoltaic power plant, characterized in that: By constructing a cloud movement time series database through historical cloud map sequences of the photovoltaic power station area, the spatiotemporal relationship between cloud trajectory and light change is extracted, and a nonlinear model of cloud movement speed and photovoltaic power output attenuation is established. A light sensor array is deployed at the edge of the photovoltaic array to detect cloud thickness and aerosol density in real time. The light intensity attenuation gradient in the local shadow area is calculated by combining the nonlinear model to predict the remaining time for the target cloud to reach the core area of ​​the photovoltaic array. The confidence interval for short-term power prediction is dynamically divided based on the remaining time. Meteorological satellite data is input into a multi-scale meteorological fusion architecture, and the feature extraction path for different meteorological parameters is adjusted based on the current cloud density using an attention mechanism. By fusing the adjusted feature extraction path with the surface reflectance correction value, a short-term power prediction sequence adapted to abrupt cloud formation scenarios is generated. The surface reflectance correction value is calculated from the scattering characteristics fed back in real time by the optical sensor array. The short-term power prediction sequence is input into the energy management system. Combined with the voltage deviation and reactive power compensation demand parameters of the photovoltaic power plant grid connection point, a dual-objective dispatch instruction that takes into account both active power regulation and reactive power control is generated. The reactive power compensation demand parameters are set by the energy management system according to the power factor requirements of the grid connection point in the power grid operation rules, and are used to constrain the priority and magnitude range of reactive power regulation. Among them, by fusing the adjusted feature extraction path with the surface reflectance correction value, a short-term power prediction sequence adapted to abrupt cloud formation scenarios is generated, including: Scattering characteristic parameters at multiple locations below the cloud layer are collected in real time using an array of optical sensors. These scattering characteristic parameters include the light intensity attenuation ratio and scattering angle distribution at different angles. Based on the light intensity attenuation ratio and scattering angle distribution in the scattering characteristic parameters, a surface reflectivity correction value is calculated for each optical sensor location. This surface reflectivity correction value reflects the dynamic impact of cloud thickness changes on surface reflectivity. The surface reflectivity correction value is then fused with the feature extraction path. This fusion process involves weighted superposition of the surface reflectivity correction value and cloud density distribution from meteorological satellite data to generate a fused feature vector with dynamic reflection compensation. The cloud density distribution and surface reflectivity correction value in the fused feature vector are parameter-coupled according to time series and spatial grid to establish a prediction equation for the movement rate of abruptly changing clouds above the photovoltaic array. The time variation parameter of cloud density distribution in the movement rate prediction equation is dynamically matched with the remaining time. By using the product relationship between the cloud movement speed and the remaining time in the time variation parameter, a short-term power prediction sequence adapted to abruptly changing cloud scenarios is generated. Specifically, a light sensor array is deployed at the edge of the photovoltaic array to detect cloud thickness and aerosol density in real time. Combined with the aforementioned nonlinear model, the light intensity attenuation gradient in the local shaded area is calculated to predict the remaining time for the target cloud cluster to reach the core area of ​​the photovoltaic array, including: A photosensitive array is deployed at the edge of the photovoltaic array to collect real-time scattered light intensity distribution data at multiple locations below the cloud layer. The scattered light angle distribution parameter and light intensity attenuation ratio parameter are extracted from the scattered light intensity distribution data. Based on the light intensity attenuation ratio parameter and scattered light angle distribution parameter, the cloud thickness parameter and aerosol density parameter corresponding to each photosensitive unit in the photosensitive array are detected in real-time. The cloud thickness parameter is obtained by comparing the light intensity attenuation ratio with a preset shading threshold. The cloud thickness parameter and aerosol density parameter are input into a nonlinear model, and the dynamic change rate of the cloud thickness parameter is correlated with the fluctuation amplitude of the aerosol density parameter to generate a light intensity attenuation gradient in the local shaded area. Based on the light intensity attenuation gradient, the cloud movement direction parameter in the nonlinear model is weighted and superimposed with the spatial distance to the core area of ​​the photovoltaic array to generate a driving intensity factor for cloud approaching the core area. The driving intensity factor is used to dynamically correct the cloud movement speed, and the ratio of the dynamically corrected cloud movement speed to the spatial distance to the core area of ​​the photovoltaic array is calculated to predict the remaining time for the target cloud to reach the core area of ​​the photovoltaic array.

2. The method according to claim 1, characterized in that, The cloud density distribution and surface reflectance correction values ​​in the fused feature vector are parameter-coupled according to time series and spatial grid to establish a prediction equation for the movement rate of abruptly changing clouds above the photovoltaic array, including: The cloud density distribution in the fused feature vector is divided into continuous time windows according to the time series, and the parameter change trend of cloud density distribution over time is extracted in each time window. The surface reflectance correction value is divided into multiple grid cells according to the spatial grid, and the spatial difference parameter of the surface reflectance correction value between adjacent grid cells is calculated. The trend of the parameter change is coupled with the spatial difference parameter, and a spatiotemporal joint change weight is generated by multiplying the rate of change of cloud density within the time window with the magnitude of the difference in surface reflectance of the spatial grid cell. Based on the spatiotemporal joint change weights, the cloud density distribution and the surface reflectance correction value within the current time window are dynamically matched to construct a parameter change model for the cloud movement direction and speed. By using the proportional relationship between the rate of change of cloud density and the magnitude of the difference in surface reflectivity in the parameter change model, the movement rate parameter of the target cloud above the photovoltaic array is calculated, and the movement rate prediction equation is obtained.

3. The method according to claim 1, characterized in that, The cloud thickness parameter and aerosol density parameter are input into a nonlinear model. The dynamic change rate of the cloud thickness parameter and the fluctuation amplitude of the aerosol density parameter are correlated to generate a light intensity attenuation gradient in the local shadow region, including: The cloud thickness parameter is divided into continuous time segments according to a time window, and the rate of change of the cloud thickness parameter in each time segment is extracted as the dynamic rate of change. The aerosol density parameter is calculated based on the spatial distribution differences between adjacent optical sensor locations to determine the aerosol density fluctuation amplitude at each optical sensor location. Based on the dynamic response rules of photovoltaic power output attenuation with cloud thickness parameters and aerosol density in the nonlinear model, the dynamic change rate is multiplied by the aerosol density fluctuation amplitude to generate a superimposed weighting factor. The spatial distribution gradient of cloud thickness parameter and the adjacent differences of aerosol density parameter within the current time segment are weighted and fused according to the superposition weighting factor to generate a combined attenuation factor; The combined attenuation factor is input into the dynamic response rule of photovoltaic power output attenuation in the nonlinear model. The light intensity attenuation gradient of the local shaded area is output through the linear combination ratio of cloud thickness parameter and aerosol density in the dynamic response rule.

4. The method according to claim 1, characterized in that, The short-term power prediction sequence is input into the energy management system, and combined with the voltage deviation and reactive power compensation demand parameters of the photovoltaic power plant grid connection point, a dual-objective scheduling instruction that takes into account both active power regulation and reactive power control is generated, including: The short-term power prediction sequence is input into the energy management system to extract the active power prediction parameters within the current time window. The active power prediction parameters characterize the power output fluctuation characteristics of the photovoltaic array after being affected by cloud cover. Real-time collection of voltage deviation parameters at the grid connection point of photovoltaic power plants; analysis of voltage deviation direction and magnitude; classification of voltage compensation requirement levels. By combining the reactive power compensation requirement parameters with the current voltage deviation parameters, the reactive power compensation adjustment intensity parameters are calculated. The adjustment intensity parameters reflect the reactive power adjustment level required for grid connection point voltage stability. By integrating the active power prediction parameters, voltage compensation demand level, and reactive power compensation adjustment intensity parameters, a dual-objective scheduling factor is generated by superimposing and balancing the active power adjustment demand and reactive power control demand through a dynamic weight allocation rule. By using the superimposed balance ratio of active power demand and reactive power demand in the dual-objective scheduling factor, a dual-objective scheduling command that takes into account both active power regulation and reactive power control is output.

5. The method according to claim 1, characterized in that, Based on the remaining time, the confidence interval for short-term power prediction is dynamically divided. Meteorological satellite data is input into a multi-scale meteorological fusion architecture. The feature extraction path for different meteorological parameters is adjusted based on the current cloud density using an attention mechanism, including: Based on the remaining time for the target cloud cluster to reach the core area of ​​the photovoltaic array, the remaining time is divided into multiple consecutive time window lengths; The confidence interval range for short-term power prediction is divided according to the relationship between the length of the time window and the speed of cloud movement; Meteorological satellite parameters are divided into multi-scale meteorological feature layers, and the multi-scale meteorological fusion architecture is input to extract cloud density parameters and wind speed parameters at different scales. Based on the confidence interval range, calculate the dynamic weight values ​​of cloud density parameters and wind speed parameters, and adjust the priority of cloud density parameters and wind speed parameters in the feature extraction path through the dynamic weight values. The cloud density and wind speed parameters in the feature extraction path after priority adjustment are assigned weight ratios according to the confidence interval range. The cloud density and wind speed parameters are then weighted and accumulated using the weight ratios to generate a short-term power prediction sequence adapted to abrupt cloud formation scenarios.

6. A photovoltaic power plant photovoltaic power prediction system, used to execute the photovoltaic power prediction method for a photovoltaic power plant according to any one of claims 1 to 5, characterized in that, include: The module is used to build a cloud movement time series library by using historical cloud map sequences of the photovoltaic power station area, extract the spatiotemporal relationship between cloud trajectory and light change, and establish a nonlinear model of cloud movement speed and photovoltaic power output attenuation. The prediction module is used to deploy a light sensor array at the edge of the photovoltaic array to detect cloud thickness and aerosol density in real time. It combines the nonlinear model to calculate the light intensity attenuation gradient in the local shadow area and predict the remaining time for the target cloud to reach the core area of ​​the photovoltaic array. The adjustment module is used to dynamically divide the confidence interval of short-term power prediction according to the remaining time, input meteorological satellite data into the multi-scale meteorological fusion architecture, and adjust the feature extraction path of different meteorological parameters based on the current cloud density through the attention mechanism. The generation module is used to fuse the surface reflectance correction value through the adjusted feature extraction path to generate a short-term power prediction sequence adapted to the sudden cloud layer scenario. The surface reflectance correction value is calculated from the scattering characteristics fed back in real time by the optical sensor array. The input module is used to input the short-term power prediction sequence into the energy management system, and combine it with the voltage deviation and reactive power compensation demand parameters of the photovoltaic power plant grid connection point to generate a dual-objective scheduling instruction that takes into account both active power regulation and reactive power control. The reactive power compensation demand parameters are set by the energy management system according to the power factor requirements of the grid connection point in the power grid operation rules, and are used to constrain the priority and magnitude range of reactive power regulation.

7. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a photovoltaic power prediction method for a photovoltaic power plant as described in any one of claims 1 to 5.

8. A computer storage medium, characterized in that, The device contains a computer program that, when executed by a computer, implements a photovoltaic power prediction method for a photovoltaic power plant as described in any one of claims 1 to 5.

Images