Distributed photovoltaic short-time output dynamic prediction method and system

Abstract

The invention discloses a distributed photovoltaic short-time output dynamic prediction method and system. The method comprises the following steps: acquiring micro-meteorological data, a cloud picture image and historical output data of a target area; calculating a building shadow shielding coefficient, and generating shadow influence factor time sequence data; obtaining cloud image occlusion index time series data based on the cloud image and a pre-constructed cloud image occlusion prediction model; fusing the data, and constructing and training a short-time output dynamic prediction model based on a long short-term memory network; predicting photovoltaic output by using real-time data and the trained model, and adjusting the model according to a prediction error and a confidence interval; according to the method provided by the invention, various factors such as microclimate, building shadow and cloud layer shielding are comprehensively considered, and the complex nonlinear relationship between the factors and the photovoltaic output is captured by using the deep learning model, so that the accuracy and robustness of distributed photovoltaic short-time output prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of output prediction, and more specifically, to a method and system for dynamically predicting short-term output of distributed photovoltaic power. Background Art

[0002] With the transformation and upgrading of the global energy structure, distributed photovoltaic power generation has developed rapidly as a clean, efficient and flexible renewable energy generation method. However, photovoltaic output is intermittent and volatile, and its output is affected by many factors, especially cloud cover and building shadows, which poses challenges to the stable operation and dispatch of the power grid. Therefore, accurately predicting the short-term output of distributed photovoltaics is of great significance to improving the power grid's ability to absorb photovoltaic power generation and ensuring the safe and stable operation of the power system.

[0003] The Chinese patent application with the authorization announcement number CN116205377B discloses a distributed photovoltaic power station output prediction method, system, computer and storage medium. The method collects the periodic historical data of the target distributed photovoltaic power station, obtains effective output data through the preprocessing algorithm of the periodic historical data, and optimizes the relevant parameters of the two-class classification model based on the effective output data using the particle swarm algorithm to obtain a prefabricated prediction algorithm. Finally, according to the weather data of the prediction day of the target distributed photovoltaic power station, the prefabricated prediction algorithm is used to perform a short-term output prediction for the power generation on the prediction day. However, this method is mainly aimed at the overall output prediction of the distributed photovoltaic power station, without considering the differences of individual photovoltaic power generation units, especially the impact of building shadows and cloud movement on the output of individual photovoltaic power generation units, making it difficult to achieve refined output prediction.

[0004] The Chinese patent application with the authorization announcement number CN113487100B discloses a global accurate prediction method and system for photovoltaic power generation output. The method numerically accumulates the individual power prediction results of each photovoltaic power station in the regional power grid to obtain the global prediction value of the photovoltaic power generation output of the regional power grid; generates a coordinate distribution map of the photovoltaic power stations in the regional power grid; calculates the temporal and spatial correlation between the outputs of photovoltaic power stations in the regional power grid based on the difference between the predicted output and the actual output of each photovoltaic power station in the regional power grid, and establishes a prediction error distribution time series diagram; calculates the front and rear positions of the photovoltaic output error movement, and estimates the prediction error at the next moment; combines the geographic coordinate information and the meteorological information of the area where each photovoltaic power station is located, calculates the correlation coefficient of the photovoltaic power station prediction error, and corrects the prediction error at the next moment. Although this method takes into account the temporal and spatial correlation between photovoltaic power stations and attempts to improve prediction accuracy by correcting prediction errors, it is essentially still based on statistical analysis of historical data. It does not deeply explore cloud map information and is difficult to accurately capture the rapid fluctuations in photovoltaic output caused by cloud movement. In distributed photovoltaic scenarios, the output differences between photovoltaic power generation units caused by buildings and cloud cover are not considered enough, resulting in a large room for improvement in prediction accuracy. At the same time, this method also lacks a mechanism for real-time dynamic adjustment of the model based on the model prediction error.

[0005] In summary, the existing photovoltaic output prediction methods lack refined modeling at the distributed photovoltaic power generation unit level, making it difficult to accurately predict the output differences of individual photovoltaic power generation units caused by building shadows and cloud movement, resulting in insufficient prediction accuracy; the cloud map information is not fully utilized, and there is a lack of dynamic tracking of cloud movement and refined quantification of cloud map occlusion, making it difficult to accurately capture the complex impact of cloud occlusion on photovoltaic output, resulting in inaccurate predictions of photovoltaic output fluctuations caused by cloud occlusion. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for dynamic prediction of distributed photovoltaic short-term output, which realizes accurate and robust dynamic prediction of the short-term output of each unit of the distributed photovoltaic system by finely modeling the shadow shading coefficient and cloud map shading index of each photovoltaic power generation unit, integrating real-time micrometeorological data, cloud map feature information and historical output data, and using a deep learning model based on a long short-term memory network (LSTM) for prediction. The model is dynamically adjusted according to the prediction error, thereby realizing accurate and robust dynamic prediction of the short-term output of each unit of the distributed photovoltaic system.

[0007] To achieve the above object, the present invention provides the following technical solutions: A distributed photovoltaic short-term output dynamic prediction method, comprising: Obtain micro-meteorological data, sky cloud image, satellite cloud image data, and historical output data of the distributed photovoltaic system in the target area; calculate the building shadow shielding coefficient in the target area and generate time series data of the shadow impact factor; obtain the cloud shielding index time series data based on the sky cloud image, satellite cloud image data and the pre-built cloud shielding prediction model; the distributed photovoltaic system in the target area includes n3 photovoltaic power generation units; Based on the micrometeorological data of the target area, the time series data of the shadow impact factor, the time series data of the cloud map occlusion index, the historical output data of the distributed photovoltaic system in the target area and the pre-built short-term output dynamic prediction model, the predicted output values ​​of the n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the future preset time are obtained, and a predicted output sequence is generated; the actual output sequence of the n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the future preset time is obtained, and the output prediction error is calculated according to the predicted output sequence and the actual output sequence; the short-term output dynamic prediction model is adjusted according to the output prediction error and the preset confidence interval.

[0008] Furthermore, the acquisition of micro-meteorological data, sky cloud images, satellite cloud data of the target area, and historical output data of the distributed photovoltaic system in the target area includes: Deploy n1 micro-meteorological monitoring stations in the target area, and obtain micro-meteorological data of the target area through the micro-meteorological monitoring stations; the micro-meteorological data includes historical micro-meteorological data and real-time micro-meteorological data; Deploy n2 wide-angle cameras at high points in the target area to collect sky cloud images; Obtain satellite cloud image data of the target area and georeference the satellite cloud image data with the sky cloud image captured by the camera; Fusing the sky cloud image and the satellite cloud image data to obtain a fused cloud image; the fused cloud image includes a historical fused cloud image and a real-time fused cloud image; Load the historical output data of the distributed photovoltaic system in the target area.

[0009] Furthermore, the calculation of the building shadow shielding coefficient of the target area and the generation of the shadow impact factor time series data includes: The building three-dimensional model and the geographical location of the target area are obtained, and the sun position information at each moment is calculated according to the geographical location of the target area. Based on the building three-dimensional model of the target area and the sun position information at each moment, the building shadow shielding coefficients of n1 micro-meteorological monitoring stations and the building shadow shielding coefficients of n3 photovoltaic power generation units at each moment are calculated; the building shadow shielding coefficients of each micro-meteorological monitoring station and each photovoltaic power generation unit are arranged in chronological order to generate time series data of shadow influencing factors.

[0010] Furthermore, the cloud map occlusion index time series data obtained based on the sky cloud map image, satellite cloud map data and the pre-built cloud map occlusion prediction model includes: Generate a fused cloud image sequence according to the fused cloud image; the fused cloud image sequence includes n4 fused cloud images, where n4≥2; obtain a cloud layer motion vector based on the fused cloud image sequence; the cloud layer motion vector is a two-dimensional vector [Vx, Vy], where Vx represents a speed in the horizontal direction and Vy represents a speed in the vertical direction; Based on the historical fusion cloud image, a cloud image hierarchical parsing model based on convolutional neural network is constructed; Based on the real-time fusion cloud image and cloud classification analysis model, the real-time cloud feature vector and real-time cloud feature level are obtained; Generate a cloud feature sequence based on the real-time cloud feature vector; generate a cloud feature level sequence based on the real-time cloud feature level; generate a cloud motion vector sequence based on the cloud motion vector; According to the cloud feature sequence, cloud feature level sequence, cloud motion vector sequence and the pre-built cloud occlusion prediction model, the cloud occlusion index of each photovoltaic power generation unit within the preset future time is obtained; the cloud occlusion index of each photovoltaic power generation unit is arranged in chronological order to generate cloud occlusion index time series data.

[0011] Furthermore, the cloud image classification parsing model based on the convolutional neural network is constructed based on the historical fusion cloud image, including: Annotate the historical fusion cloud map images and divide each historical fusion cloud map image into four levels, namely, the first-level feature level, the second-level feature level, the third-level feature level and the fourth-level feature level; Determine the cloud layer feature vector loss function and the classification loss function, and perform weighted average of the cloud layer feature vector loss function and the classification loss function to obtain the loss function of the cloud image hierarchical analysis model; The annotated historical fusion cloud image is used to form a sample data set; according to the cloud layer feature vector loss function, sample pairs are constructed in the sample data set; The cloud graph hierarchical parsing model is trained based on the sample data set of the sample pairs and the loss function of the cloud graph hierarchical parsing model.

[0012] Furthermore, obtaining the real-time cloud image feature vector and the real-time cloud image feature level includes: The real-time fused cloud image is input into the cloud image hierarchical analysis model to obtain the real-time cloud image feature vector and cloud image feature level probability vector p=[p1,p2,p3,p4]; where p i Indicates the probability value of the real-time fusion cloud image being the i-th level feature level, 1≤i≤4; Take the cloud image feature level probability vector p i The maximum value of is used as the real-time cloud feature level of the real-time fused cloud image.

[0013] Furthermore, the cloud map shielding index of each photovoltaic power generation unit within a preset future time is obtained, which includes: The cloud feature sequence is divided into a cloud feature modeling sequence and a cloud feature occlusion prediction sequence; the cloud feature level sequence is divided into a cloud feature level modeling sequence and a cloud feature level occlusion prediction sequence; the cloud layer motion vector sequence is divided into a cloud layer motion vector modeling sequence and a cloud layer motion vector occlusion prediction sequence; The historical output data of the distributed photovoltaic system in the target area is converted into a historical cloud map occlusion index sequence; a cloud map occlusion prediction model is constructed and trained according to the cloud map feature modeling sequence, the cloud map feature level modeling sequence, the cloud layer motion vector modeling sequence and the historical cloud map occlusion index sequence to obtain a trained cloud map occlusion prediction model; According to the cloud image feature occlusion prediction sequence, the cloud image feature level occlusion prediction sequence, the cloud layer motion vector occlusion prediction sequence and the trained cloud image occlusion prediction model, the cloud image occlusion index of each photovoltaic power generation unit within the future preset time is obtained.

[0014] Furthermore, the method of obtaining the predicted output values ​​of n3 photovoltaic power generation units in the target area distributed photovoltaic system within a preset future time includes: The shadow influence factor time series data is divided into the shadow influence factor modeling time series data and the shadow influence factor output prediction time series data; the cloud map occlusion index time series data is divided into the cloud map occlusion index modeling time series data and the cloud map occlusion index output prediction time series data; Align the historical micro-meteorological data, shadow impact factor modeling time series data, cloud cover index modeling time series data and historical output data of the target area according to timestamps; The aligned historical micro-meteorological data, shadow impact factor modeling time series data, cloud map occlusion index modeling time series data and historical output data are fused to generate a fused data set; Based on the fused data set, a short-term output dynamic prediction model is constructed and trained; the short-term output dynamic prediction model adopts a long short-term memory network, including an input layer, an LSTM layer, a fully connected layer and an output layer; the output layer contains n5 neurons, each of which outputs the predicted output value of a photovoltaic power generation unit within a preset time in the future.

[0015] Furthermore, the step of obtaining the predicted output values ​​of n3 photovoltaic power generation units in the target area distributed photovoltaic system within a preset future time period also includes: Based on the real-time micrometeorological data of the target area, the output prediction time series data of the shadow influence factor, the output prediction time series data of the cloud map shading index and the short-term output dynamic prediction model, the predicted output values ​​of n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the future preset time are obtained.

[0016] A distributed photovoltaic short-term output dynamic prediction system, which is used to implement the above-mentioned distributed photovoltaic short-term output dynamic prediction method, the system comprises: Data acquisition module: used to obtain micro-meteorological data, sky cloud images, satellite cloud data of the target area, and historical output data of the distributed photovoltaic system in the target area; Calculation module: used to calculate the building shadow shielding coefficient of the target area and generate the time series data of the shadow influence factor; based on the sky cloud image, satellite cloud image data and the pre-built cloud shielding prediction model, obtain the cloud shielding index time series data; the distributed photovoltaic system in the target area includes n3 photovoltaic power generation units; Output prediction module: Based on the micro-meteorological data of the target area, the time series data of the shadow impact factor, the time series data of the cloud map shielding index, the historical output data of the distributed photovoltaic system in the target area and the pre-built short-term output dynamic prediction model, the predicted output values ​​of n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the preset time in the future are obtained, and the predicted output sequence is generated; Model adjustment module: used to obtain the actual output sequence of n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the preset future time, calculate the output prediction error based on the predicted output sequence and the actual output sequence; adjust the short-term output dynamic prediction model based on the output prediction error and the preset confidence interval.

[0017] Compared with the prior art, the present invention has the following beneficial effects: By integrating multi-source heterogeneous data such as micro-meteorological data, building shadow effects, cloud occlusion information, and historical output data, and utilizing the powerful nonlinear fitting capabilities of deep learning models, the accurate prediction of distributed photovoltaic output is achieved, effectively reducing the prediction error and improving the reliability of the prediction results. A model adjustment mechanism based on prediction error and confidence interval is introduced, so that the model can be adaptively updated according to the actual operation situation, timely capture the impact of environmental changes on photovoltaic output, and maintain the long-term stability of the model prediction performance. The shadow occlusion coefficient and cloud map occlusion index are calculated for each photovoltaic power generation unit, and independent output neurons are constructed for prediction, which realizes the refined modeling of each unit in the distributed photovoltaic system, fully considers the differences between different units, and further improves the prediction accuracy. By constructing a cloud map hierarchical analysis model and a cloud map occlusion prediction model, the rich information contained in the sky cloud map image and satellite cloud map data is fully mined, and the image information that is difficult to directly use is converted into quantifiable feature vectors, levels and occlusion indexes, which effectively improves the model's perception of cloud occlusion effects. The method of combining multi-source data fusion and deep learning models is adopted to enhance the robustness of the model, enabling it to adapt to complex and changeable weather conditions and operating environments, and reducing prediction deviations caused by quality issues of a single data source or insufficient model generalization capabilities. This application can provide more accurate and reliable distributed photovoltaic output forecast data for power grid dispatching, help optimize power grid dispatching strategies, improve the economy and safety of power grid operation, promote the consumption of renewable energy, and promote the transformation and upgrading of the energy structure. By constructing and training deep learning models, this application realizes the automation of the entire process of data processing, feature extraction, model prediction and adjustment, reduces the need for manual intervention, and improves prediction efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 It is a principle flow chart of the distributed photovoltaic short-term output dynamic prediction method in the present invention; Figure 2 A flow chart of a method for obtaining micro-meteorological data, sky cloud image, satellite cloud image data of a target area, and historical output data of a distributed photovoltaic system in a target area in a distributed photovoltaic short-term output dynamic prediction method of the present invention; Figure 3 A flow chart of a method for generating time series data of shadow impact factors in a distributed photovoltaic short-term output dynamic prediction method of the present invention; Figure 4 A flow chart of a method for constructing a cloud graph hierarchical parsing model based on a convolutional neural network in a distributed photovoltaic short-term output dynamic prediction method of the present invention; Figure 5 A flow chart of a method for obtaining a real-time cloud map feature vector and a real-time cloud map feature level in a distributed photovoltaic short-term output dynamic prediction method of the present invention; Figure 6 A flow chart of a method for constructing a short-term output dynamic prediction model in the distributed photovoltaic short-term output dynamic prediction method of the present invention; Figure 7 A flow chart of a method for adjusting a short-term output dynamic prediction model according to an output prediction error and a preset confidence interval in a distributed photovoltaic short-term output dynamic prediction method of the present invention; Figure 8 It is a functional module diagram of the distributed photovoltaic short-term output dynamic prediction system in the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Example 1

[0022] See also Figure 1 As shown, this embodiment provides a distributed photovoltaic short-term output dynamic prediction method, including: Step S1000, obtaining micro-meteorological data, sky cloud image, satellite cloud image data, and historical output data of the distributed photovoltaic system in the target area; calculating the building shadow shielding coefficient of the target area, and generating time series data of the shadow impact factor; based on the sky cloud image, satellite cloud image data and the pre-built cloud shielding prediction model, obtaining the cloud shielding index time series data; Furthermore, step S1000 includes: Step S1100, obtaining micro-meteorological data, sky cloud images, satellite cloud data of the target area, and historical output data of the distributed photovoltaic system in the target area; Furthermore, if Figure 2 As shown, step S1100 includes: Step S1110, deploying n1 micro-meteorological monitoring stations in the target area, and obtaining micro-meteorological data of the target area through the micro-meteorological monitoring stations; the micro-meteorological data includes temperature, humidity, wind speed, wind direction, air pressure, direct radiant illuminance (DNI) and diffuse radiant illuminance (DHI); the micro-meteorological data includes historical micro-meteorological data and real-time micro-meteorological data; Specifically, micro-meteorological data refers to the meteorological characteristics within a specific small area, which directly affects the actual output of the photovoltaic power generation unit. In order to accurately capture this data, it is necessary to strategically deploy n1 micro-meteorological monitoring stations in the target area. The specific value of n1 needs to be considered comprehensively based on factors such as the area, terrain, and building distribution of the target area. Generally speaking, the larger the target area, the more complex the terrain, and the more uneven the distribution of buildings, the more micro-meteorological monitoring stations are required. For example, for a community with an area of ​​1 square kilometer, flat terrain, and evenly distributed buildings, 3-5 micro-meteorological monitoring stations can be deployed; while for a community of the same area but with large terrain fluctuations and unevenly distributed buildings, 5-8 micro-meteorological monitoring stations may be deployed to fully capture the spatial differences in micro-meteorology. The final number selection is determined by simulation testing to ensure the representativeness of the micro-meteorological data, which means that the micro-meteorological data can truly reflect the micro-meteorology of the location of the photovoltaic power generation unit. The monitoring station should be deployed as close as possible to the photovoltaic power generation unit to obtain micro-meteorological data that is closest to the actual operating environment of the photovoltaic power generation unit. For example, if the photovoltaic power generation units in a community are mainly distributed on the roof, then the micro-meteorological monitoring station should also be deployed on the roof first.

[0023] Each monitoring station is equipped with a variety of sensors to measure meteorological parameters such as temperature, humidity, wind speed, wind direction, air pressure, direct irradiance (DNI) and diffuse irradiance (DHI). Temperature and humidity can affect the conversion efficiency of photovoltaic cells. For example, high temperatures can cause the output power of photovoltaic cells to drop. Wind speed can affect the heat dissipation of photovoltaic panels, thereby affecting their output power. Wind direction is related to the shading effect of buildings. Changes in air pressure can reflect changes in weather systems and indirectly affect photovoltaic output. Direct irradiance (DNI) refers to the radiation intensity of the sun directly hitting the earth's surface, which is the main energy source for photovoltaic power generation. Diffuse irradiance (DHI) refers to the radiation intensity of sunlight reaching the earth's surface after being scattered by the atmosphere. Even on cloudy days, diffuse radiation can still be used by photovoltaic panels. Different meteorological parameters have different changing frequencies. For example, irradiance is affected by the movement of clouds and changes very quickly, so it needs to be collected at a frequency of seconds; while parameters such as temperature and humidity change relatively slowly and can be collected at a frequency of minutes. The collected raw data is preliminarily screened and cleaned to remove outliers or erroneous data caused by sensor failure, communication failure or environmental interference. The data collected by each monitoring station is converted into a unified data format, including unified timestamps, unified units, and unified data formats, so as to facilitate subsequent data storage, processing and analysis. Historical data is used to train subsequent models, while real-time data is used for real-time photovoltaic output prediction.

[0024] This step can obtain accurate, reliable and representative micro-meteorological information by strategically deploying micro-meteorological monitoring stations and finely collecting and processing micro-meteorological data. This information can finely describe the meteorological conditions of the environment in which the photovoltaic power generation units are located in the target area, providing accurate input for subsequent model predictions, thereby improving prediction accuracy.

[0025] Step S1120, deploying n2 wide-angle cameras at high points in the target area to collect sky cloud images; the sky cloud images include historical sky cloud images and real-time sky cloud images; Specifically, the sky cloud image can intuitively reflect the distribution, type and movement of clouds over the target area, which is crucial for predicting short-term fluctuations in photovoltaic output. The value of n2 depends on the size and terrain of the target area. Generally speaking, at least two wide-angle cameras are required to cover the sky of the entire target area. For larger areas or areas with complex terrain, more cameras may be required. For example, for a community of 1 square kilometer, 2-4 wide-angle cameras can be deployed. Wide-angle cameras need to be deployed at high points in the target area, such as the roof of a building, to obtain a wide field of view and avoid obstruction. At the same time, the direction of the camera needs to be considered to ensure that the sky in and around the target area can be covered. For example, two cameras can be pointed east and west respectively to capture the changes in clouds in the direction of sunrise and sunset. Clouds are constantly changing. In order to capture the movement and evolution of clouds, sky cloud images need to be collected at a higher frequency. The specific acquisition frequency can be determined according to actual needs and the speed of cloud change, and is generally set to seconds or minutes. For example, an image can be collected every 30 seconds or 1 minute.

[0026] Due to factors such as camera lens distortion and lighting changes, the collected original images may have some problems and need to be preprocessed, including image dedistortion, image enhancement, and image cropping. Historical sky cloud images are used for subsequent training of cloud analysis models, and need to cover different weather conditions, different cloud types, and different time periods to ensure data diversity. For example, cloud images of different times of day in the past year can be collected, and it is ensured that they contain images under various weather conditions such as sunny, cloudy, and overcast days. Real-time sky cloud images are used to input into the trained cloud classification analysis model to obtain cloud feature vectors and feature levels, which are finally used for real-time photovoltaic output prediction.

[0027] Cloud cover is the main factor causing short-term fluctuations in photovoltaic output. The movement of clouds can cause rapid changes in the intensity of solar radiation received by photovoltaic panels, thus causing fluctuations in photovoltaic output. Traditional prediction methods based on numerical weather forecasts are difficult to accurately predict such rapid changes in clouds, resulting in low prediction accuracy. This step deploys wide-angle cameras and collects sky cloud images at a high frequency to obtain detailed information about clouds above the target area, such as the distribution, type, thickness, direction of movement, and speed of the clouds. This information can help the model accurately determine the obstruction of clouds and predict short-term fluctuations in photovoltaic output, thereby improving prediction accuracy.

[0028] Step S1130, obtaining satellite cloud image data of the target area, and geo-referencing the satellite cloud image data with the sky cloud image captured by the camera; the satellite cloud image data includes historical satellite cloud image data and real-time satellite cloud image data; Specifically, satellite cloud image data provides cloud information at a larger scale, which can make up for the limitations of the observation range of ground cameras and help the model better understand the overall movement trend of the cloud layer. Commonly used high-resolution satellite data sources include Landsat, Sentinel, etc. When selecting a satellite data source, it is necessary to consider the spatial resolution, temporal resolution, and spectral resolution of the data. According to the geographic coordinates and time range of the target area, the corresponding satellite cloud image data is downloaded from the satellite data source. The downloaded satellite cloud image data is preprocessed, including radiation correction, atmospheric correction, geometric correction, etc., to eliminate errors and noise in the data. The preprocessed satellite cloud image data is georeferenced with the sky cloud image image collected by the ground camera to align the two data in space. This is a key step. It unifies data from different sources and scales into the same geographic coordinate system, laying the foundation for subsequent data fusion. Historical satellite cloud image data is used to train cloud image parsing models, which need to cover different weather conditions, different cloud types, different time periods, etc. to ensure data diversity.

[0029] Although ground cameras can provide detailed information about clouds over the target area, their observation range is limited. This step obtains satellite cloud image data and performs georeferencing to combine large-scale cloud information with ground camera observations, allowing the model to consider both local and overall cloud changes, thereby improving the accuracy and stability of the prediction.

[0030] Step S1140, fusing the sky cloud image and the satellite cloud image data to obtain a fused cloud image; the fused cloud image includes a historical fused cloud image and a real-time fused cloud image; Specifically, step S1140 fuses the sky cloud image collected by the ground camera and the satellite cloud image data to obtain a more informative and comprehensive fused cloud image. This embodiment adopts a pixel-level fusion method to directly fuse the pixel values ​​of the two images. The corresponding pixel values ​​of the two images are weighted averaged to obtain the fused pixel values. For example, the corresponding pixel values ​​of the ground camera cloud image and the satellite cloud image can be assigned weights of 0.6 and 0.4 respectively, and then weighted averaged. The weight coefficient can be adjusted according to the quality and reliability of the two images.

[0031] For example, assuming that the grayscale value of the ground camera cloud image at the pixel point (x, y) is A, and the grayscale value of the satellite cloud image at the same position is B, and the weighted averaging method is used, with weights of 0.6 and 0.4 respectively, the grayscale value of the fused image at this point is 0.6*A+0.4*B.

[0032] The historical fused cloud image is used to train the cloud image classification analysis model. The real-time fused cloud image is used to input the trained cloud image classification analysis model to obtain the cloud image feature vector and feature level, which is finally used for real-time photovoltaic output prediction.

[0033] Ground camera cloud images and satellite cloud image data each have their own advantages and disadvantages. Ground camera cloud images can provide detailed information about the clouds above the target area, but the observation range is limited; satellite cloud images can provide cloud information over a wider range, but the resolution is low and it is difficult to capture the detailed changes in the clouds. By fusing ground camera cloud images and satellite cloud image data, a more informative and comprehensive fused cloud image is obtained. This fusion strategy makes full use of the advantages of both types of data, retaining the high resolution and detailed information of the ground camera cloud images, and integrating the large-scale background information of the satellite cloud images, providing better data input for subsequent model training and prediction. For example, the fused image can contain both the fine structure of the clouds above the target area and the distribution and movement trend of the clouds in a larger range around it, thereby more accurately reflecting the shielding effect of the clouds and improving the accuracy of photovoltaic output prediction.

[0034] Step S1150, loading the historical output data of the distributed photovoltaic system in the target area, wherein the distributed photovoltaic system in the target area includes n3 photovoltaic power generation units, and the historical output data of the distributed photovoltaic system includes the actual output, installed capacity and timestamp of the n3 photovoltaic power generation units.

[0035] Specifically, historical output data records the actual power generation of photovoltaic power generation units in the past period of time, and is the key label data for training short-term output dynamic prediction models. It is necessary to connect with the monitoring platform or data center of the distributed photovoltaic system in the target area to obtain historical output data. Historical output data should contain the actual output value, installed capacity and timestamp of each photovoltaic power generation unit. The actual output value refers to the power value actually output by the photovoltaic power generation unit at a certain moment, usually in kilowatts (kW). The installed capacity refers to the designed maximum output power of the photovoltaic power generation unit, usually also in kilowatts (kW). The installed capacity is a fixed value and can be used as the identification information of the photovoltaic power generation unit. The timestamp records the time point corresponding to each output data, which needs to be consistent with the timestamp of the micro-meteorological data and cloud map data for data alignment and fusion. It is necessary to determine the coverage period of historical data, such as data from the past year or two. Generally speaking, the longer the data period, the better the training effect of the model, but the availability and storage cost of the data also need to be considered.

[0036] This step provides the necessary label data for model training by obtaining the historical output data of the distributed photovoltaic system in the target area. These data reflect the actual operation of the photovoltaic power generation unit in the past period of time and are the basis for the model to learn the photovoltaic output law. The short-term output dynamic prediction model is essentially a supervised learning model that requires a large amount of historical data for training. The historical output data is used as label data, together with the input micro-meteorological data, cloud map data and other feature data, to train the model to learn the relationship between photovoltaic output and various influencing factors. Only by using accurate and reliable historical output data for training can a prediction model with good performance be obtained.

[0037] Step S1200, calculating the building shadow shielding coefficient of the target area and generating time series data of the shadow impact factor; Furthermore, if Figure 3 As shown, step S1200 includes: Step S1210, obtaining a three-dimensional building model of the target area, including the height, outline and location information of the building; Step S1220, obtaining the geographical location of the target area, and calculating the sun position information at each moment according to the geographical location of the target area, wherein the sun position information includes the sun altitude angle and the azimuth angle; Step S1230, based on the three-dimensional building model of the target area and the sun position information at each moment, calculate the building shadow shielding coefficients of n1 micro-meteorological monitoring stations at each moment and the building shadow shielding coefficients of n3 photovoltaic power generation units at each moment; Step S1240, arranging the building shadow shielding coefficients of each micro-meteorological monitoring station and each photovoltaic power generation unit in chronological order to generate time series data of shadow impact factors.

[0038] Specifically, the three-dimensional building model of the target area contains the height, outline and precise geographic location information of each building. The three-dimensional building model can be obtained through surveying and mapping departments, aerial photography and laser radar scanning; the height of the building refers to the height of the building relative to the ground, which is a key parameter for calculating the length of the shadow. The outline refers to the projection shape of the building on the horizontal plane, which determines the shape of the shadow. The location information refers to the precise geographic coordinates of the building, such as longitude and latitude coordinates, which are used to determine the relative positions between buildings and the orientation relationship with the sun. The formation of the shadow is the result of light propagation along a straight line. To calculate the range and intensity of the shadow, it is necessary to know the position of the light source (sun), the shape and size of the obstruction (building), and the position of the receiving surface (photovoltaic panel or micro-meteorological monitoring station). Step S1210 provides the necessary data basis for the subsequent calculation of the shadow shading coefficient by obtaining the three-dimensional building model of the target area. An accurate three-dimensional building model is the key to accurately calculate the shadow shading. Only by obtaining the height, outline and location information of the building can the projection and change of the shadow be accurately simulated.

[0039] The geographical location of the target area refers to the longitude and latitude coordinates of the target area. Usually, the longitude and latitude coordinates of the center of the target area can be used as a representative. The solar position information refers to the position and altitude of the sun relative to the target area, which is usually expressed by the solar altitude angle and azimuth angle. The solar altitude angle refers to the angle between the sun's rays and the horizon, and the value range is -90° to 90°. The solar altitude angle is the largest at noon, the solar altitude angle is 0° at sunrise and sunset, and the solar altitude angle is negative at night. The solar azimuth angle refers to the angle between the projection of the sun's rays on the horizon and the north direction, usually measured in a clockwise direction with the north as 0°, and the value range is 0° to 360°. For example, the solar azimuth angle in the east direction is 90°, in the south direction is 180°, and in the west direction is 270°. According to the geographical location, date and time of the target area, the solar altitude angle and azimuth angle at that moment can be calculated using the solar position algorithm. Common solar position algorithms include SPA (Solar Position Algorithm), PSA (PreciseSolar Position Algorithm), etc. The accuracy of these algorithms is very high, reaching the arc second level. The position of the sun is a key factor affecting the direction and length of shadow projection. At different times of the day, the altitude and azimuth of the sun change continuously, causing the direction and length of the shadow to change accordingly. Step S1220 obtains the geographical location of the target area and uses the sun position algorithm to calculate the sun altitude and azimuth at each moment, providing key sun position information for the subsequent calculation of the shadow occlusion coefficient.

[0040] Step S1230 calculates the building shadow shielding coefficient of each micro-meteorological monitoring station and each photovoltaic power generation unit at each moment based on the building three-dimensional model obtained in step S1210 and the sun position information calculated in step S1220. Common shadow shielding calculation methods include ray tracing, shadow volume algorithm and projection method; the ray tracing method emits a series of rays from the sun position to the target point (micro-meteorological monitoring station or photovoltaic power generation unit) to determine whether these rays are blocked by the building. If a ray is blocked by the building, it means that the sun's rays in this direction cannot reach the target point. According to the proportion of blocked rays, the shadow shielding coefficient can be calculated. This method has high accuracy, but the amount of calculation is large. The shadow volume algorithm constructs a shadow volume according to the building and the sun's position. If the target point is located in the shadow volume, it means that the point is blocked by the shadow. This method is efficient, but the calculation accuracy for complex scenes may not be as good as the ray tracing method. The projection method projects the building and the sun's position onto the same plane, and then calculates the shielding of the target point by the building's shadow on the plane. This method is relatively simple to implement, but the calculation accuracy for complex scenes is limited. In practical applications, the appropriate shadow shielding calculation method is selected according to specific needs. Different photovoltaic power generation units and micro-meteorological monitoring stations are blocked by building shadows in different ways due to their different locations. Even for the same photovoltaic power generation unit, the shadow shielding situation will change at different times due to the change in the position of the sun. Therefore, it is necessary to calculate the shadow shielding coefficient of each photovoltaic power generation unit and micro-meteorological monitoring station at each moment. Only in this way can the impact of building shadows on photovoltaic output be accurately reflected, thereby improving the accuracy of the prediction model.

[0041] Step S1240 arranges the shadow shielding coefficients at each moment calculated in step S1230 in chronological order to generate time series data of shadow impact factors. The shadow shielding coefficients need to be sorted according to their corresponding timestamps to ensure that the time sequence of the data is correct. For each micro-meteorological monitoring station and each photovoltaic power generation unit, a corresponding shadow impact factor time series data is generated. This data is a time series that records the shadow shielding coefficients at each moment.

[0042] The shadow shielding coefficient is a quantity that changes over time. By arranging the shadow shielding coefficients in chronological order, a time series can be formed, which can reflect the dynamic changes of the shadow shielding effect of the building. For example, during the day, as the position of the sun changes, the shadow shielding coefficient will show a certain periodic change pattern. During the year, as the seasons change, the shadow shielding coefficient will also show a certain seasonal change pattern. These patterns can be learned by the model to improve the accuracy of the prediction. The shadow influencing factor time series data is like a "historical record of shadow shielding". The model can learn the change pattern of shadow shielding from it, and use this pattern to predict future shadow shielding conditions, and then predict changes in photovoltaic output.

[0043] As the position of the sun changes throughout the day and year, the shadow cast by the building will also change, thus affecting the solar radiation received by the photovoltaic panels and thus affecting their power generation. Step S1200 quantifies the impact of building shadows on photovoltaic output by calculating the building shadow shielding coefficient and generating time series data of shadow impact factors. This quantification method converts the shadow impact that was originally difficult to measure directly into numerical data that can be used for model training and prediction, thereby improving the accuracy and practicality of the prediction model.

[0044] Step S1300, based on the historical fusion cloud image, construct a cloud image hierarchical parsing model based on a convolutional neural network (CNN); Furthermore, if Figure 4 As shown, step S1300 includes: Step S1310, annotating the historical fusion cloud image, and dividing each historical fusion cloud image into four levels, namely, the first-level feature level, the second-level feature level, the third-level feature level and the fourth-level feature level; Specifically, in order to train the convolutional neural network, it is necessary to annotate the historical fused cloud image, that is, assign a feature level label to each image. Meteorological experts or experienced photovoltaic operation and maintenance personnel can manually annotate according to the characteristics of the cloud image. This method has high annotation quality, but high cost and low efficiency. It is also possible to use existing cloud classification data sets (such as cloud classification products from satellites such as MODIS and CALIOP) for transfer learning. The output of the existing cloud classification model is used as a reference to semi-automatically annotate the cloud image data of this method. This method can improve annotation efficiency and reduce costs.

[0045] Cloud images are divided into four levels according to cloud amount and cloud layer characteristics, each with clear definitions and distinction criteria: Level 1 feature level: clear sky or few clouds, cloud cover less than 10%. The corresponding cloud feature vector should reflect that the blue color of the sky area is dominant, and may contain a small amount of white, small and scattered clouds. In the case of clear sky or few clouds, the solar radiation is strong, and the cloud feature vector should reflect that the color of the sky area is blue as the main tone and very uniform, and the corresponding photovoltaic output is large and stable.

[0046] Secondary feature level: thin clouds or a small amount of cumulus clouds, cloud cover between 10%-40%, and thin clouds. The corresponding cloud feature vector should reflect that the cloud texture is finer, the light transmittance is better, and it may appear light gray or white. In the case of thin clouds or a small amount of cumulus clouds, solar radiation is slightly affected, and the cloud feature vector should reflect that the cloud texture is finer, the light transmittance is better, and the cloud boundary is also more obvious. At this time, the photovoltaic output will drop slightly, but it will still remain at a relatively high level.

[0047] Level 3 feature level: thicker clouds or fast-moving clouds, cloud cover between 40%-80%, thicker clouds or faster moving speed. The corresponding cloud image feature vector should reflect that the cloud texture is rough, the light transmittance is poor, it may appear dark gray or white, and may contain certain dynamic features, such as blurred edges. In the case of thicker clouds or fast-moving clouds, solar radiation will be significantly blocked by the clouds, and photovoltaic output will decrease rapidly and fluctuate. The cloud image feature vector should reflect that the cloud texture is rough, the light transmittance is poor, it may appear dark gray or white, and may contain certain dynamic features, such as blurred edges.

[0048] Level 4 feature level: heavy clouds or thunderstorm clouds, cloud cover greater than 80%, very heavy clouds or accompanied by thunderstorm weather. The corresponding cloud feature vector should reflect the dark color of the cloud, almost opaque, and may appear dark gray, black, or with lightning and other features. In the case of heavy clouds or thunderstorm clouds, solar radiation is severely blocked and photovoltaic power generation is basically stopped. The cloud feature vector should reflect the dark color of the cloud, almost opaque, and may appear dark gray, black, or with lightning and other features.

[0049] The cloud image classification and analysis model is a supervised learning model that needs to be trained using labeled data. Step S1310 provides labeled training data for the training of the cloud image classification and analysis model by annotating the historical fusion cloud image.

[0050] Step S1320, determining a cloud layer feature vector loss function and a classification loss function, performing weighted average of the cloud layer feature vector loss function and the classification loss function to obtain a loss function of a cloud image hierarchical parsing model; Specifically, the loss function is used to measure the difference between the model's prediction results and the true labels, and guide the optimization of model parameters during training. The cloud feature vector loss function is used to measure the extraction error of the cloud feature vector. The goal of this loss function is to make the feature vectors of cloud images of the same level as close as possible, and the feature vectors of cloud images of different levels as far apart as possible. In this method, the cloud feature vector loss function is defined as the contrastive loss function (ContrastiveLoss).

[0051] The core idea of ​​contrast loss is that the feature vectors of cloud images of the same level should be as close as possible, and the feature vectors of cloud images of different levels should be as far apart as possible. The training sample pairs can be constructed as follows: Positive sample pair: two cloud images of the same level.

[0052] Negative sample pair: two cloud images of different levels.

[0053] Then, define the contrast loss function ,For example:

[0054] Where: Y represents the label of the sample pair, Y=0 represents a positive sample pair, and Y=1 represents a negative sample pair. D represents the distance between the feature vectors of two samples (e.g., Euclidean distance). margin is a pre-set threshold. The goal of the contrast loss function is to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs, but the maximum distance between negative sample pairs does not exceed margin.

[0055] The classification loss function is used to measure the prediction error of the cloud image feature level. The goal of this loss function is to make the cloud image feature level probability vector predicted by the model as close as possible to the actual feature level label. The cross-entropy loss function is used in this method. The cloud layer feature vector loss function and the classification loss function are assigned different weights, and then weighted summation is performed. The weight ratio can be adjusted according to the actual situation. For example, the weight can be determined according to the importance of the feature vector and feature level in the photovoltaic output prediction, and the loss function of the final cloud image classification analysis model is obtained.

[0056] The cloud image hierarchical parsing model has two outputs: the cloud image feature vector and the cloud image feature level probability vector. Therefore, two loss functions need to be defined to measure the errors of these two outputs respectively. The cloud layer feature vector loss function (contrast loss function) enables the model to learn the feature differences of cloud images of different levels, map cloud images of different levels to different regions in the feature space, and map cloud images of the same level to similar regions. The classification loss function (cross entropy loss function) enables the model to accurately predict the feature level of the cloud image based on the extracted feature vector. Combining these two loss functions allows the model to learn both discriminative feature representations and accurate classification capabilities. This joint optimization strategy can simultaneously optimize the model's feature extraction and classification capabilities, enabling the model to extract discriminative cloud image feature vectors and accurately predict the feature level of the cloud image.

[0057] Step S1330, the annotated historical fusion cloud image is used to form a sample data set; according to the cloud layer feature vector loss function, a sample pair is constructed in the sample data set; the annotated historical fusion cloud image includes a feature level label; Specifically, all the annotated historical fusion cloud map images and their corresponding feature level labels are combined together to form a sample data set. According to the requirements of the contrast loss function, positive sample pairs and negative sample pairs need to be constructed.

[0058] Positive sample pair: Randomly select two cloud images with the same feature level labels from the sample data set to form a positive sample pair.

[0059] Negative sample pair: Randomly select two cloud images with different feature level labels from the sample dataset to form a negative sample pair.

[0060] Sample pair ratio: The ratio of positive and negative sample pairs can be adjusted according to actual conditions. For example, the ratio of positive and negative sample pairs can be set to 1:1, or adjusted according to the number of samples in different levels of cloud maps.

[0061] The contrast loss function needs to be calculated based on sample pairs. By constructing positive and negative sample pairs, the cloud map classification problem can be transformed into a contrast learning problem, so that the model can learn how to distinguish cloud maps of different levels. For example, during the training process, the model will gradually learn the feature differences between the first-level cloud map and the second-level cloud map, as well as the feature differences between the second-level cloud map and the third-level cloud map, and so on. This learning method based on sample pairs is similar to the process of humans learning to distinguish different things. By comparing similar and different things, they gradually master the key features that distinguish them. Step S1330 provides qualified training data for the training of the cloud map classification parsing model by constructing sample data sets and sample pairs. The sample pair construction method enables the model to learn the differences between the features of cloud maps of different levels, thereby improving the feature extraction ability of the model.

[0062] Step S1340, training the cloud graph hierarchical parsing model according to the sample data set of the sample pairs and the loss function of the cloud graph hierarchical parsing model.

[0063] Specifically, the cloud image hierarchical parsing model selects CNN architecture, such as VGG, ResNet, Inception, etc. These architectures have achieved good results in image feature extraction and have good transfer learning capabilities. Design the hierarchical structure of the network according to the characteristics of the cloud image data and the complexity of the task. Usually, the first half of the network is used for feature extraction, which can include multiple convolutional layers, pooling layers, activation functions, batch normalization layers, etc.; the second half of the network is designed with two branches as needed, which are used to output cloud image feature vectors and cloud image feature levels respectively. The feature vector extraction branch adds several fully connected layers after the basic CNN architecture to extract cloud image feature vectors. The output of these fully connected layers is a low-dimensional vector that contains the key feature information of the cloud image. Usually, the dimension of the feature vector is much lower than the number of pixels of the original image. For example, a 128-dimensional or 256-dimensional vector can be used to represent cloud image features. The cloud image feature level prediction branch adds one or more fully connected layers after the basic CNN architecture and uses the Softmax function as the output layer to predict the feature level of the cloud image. The Softmax function can convert the output of the network into a probability distribution, indicating the probability that the cloud belongs to each level. For example, for a four-level classification, the output layer can contain 4 neurons, each of which outputs a probability value, indicating the probability that the cloud belongs to the corresponding level.

[0064] The sample data set is divided into a training set, a validation set, and a test set. For example, 70% of the data can be used as a training set, 15% of the data as a validation set, and 15% of the data as a test set. The parameters of the cloud image hierarchical parsing model are randomly initialized. A batch of sample pairs is input into the model. The model extracts features and predicts feature levels for the two cloud images in each sample pair, and obtains cloud image feature vectors and feature level probability vectors. According to the output of the model and the label of the sample pair, the cloud layer feature vector loss (contrast loss) and classification loss (cross entropy loss) are calculated, and the weighted average of the two is used to obtain the final loss value. According to the loss value, the gradient of the model parameters is calculated by the back propagation algorithm. Use an optimizer (such as Adam, SGD, etc.) to update the model parameters according to the calculated gradient. Repeat forward propagation, loss calculation, back propagation, and parameter update until the performance of the model on the validation set no longer improves or reaches the predetermined number of iterations. Use the validation set to evaluate the model during the training process, and adjust the model's hyperparameters (such as learning rate, batch size, number of iterations, network structure, etc.) based on the evaluation results. Use the test set to test the trained model to obtain the final performance of the model.

[0065] Step S1340 obtains a model that can automatically extract cloud features and perform classification by training the cloud image classification and analysis model. The model can convert the input cloud image into a structured data representation, namely, a cloud feature vector and a cloud feature level. The cloud feature vector contains the key feature information of the cloud, such as the texture, shape, color, edge and other features of the cloud, which can reflect the degree of occlusion and change trend of the cloud layer. The cloud feature level classifies the cloud image in a coarse-grained manner, which can reflect the overall occlusion of the cloud layer, and lays the foundation for the subsequent cloud occlusion index prediction and photovoltaic output prediction.

[0066] Cloud images contain rich cloud information, which is crucial for predicting photovoltaic output. However, the original cloud image data is large in size and the information is redundant, making it difficult to use directly for model training. Through the CNN model, the features of the cloud image can be automatically extracted and converted into low-dimensional feature vectors, thereby achieving data dimensionality reduction and feature abstraction. In addition, grading the cloud image can further simplify the representation of cloud information and improve the training efficiency of the model. Step By constructing a CNN-based cloud image grading parsing model, automatic feature extraction and grading of cloud images are achieved. The model is able to convert complex cloud image information into structured data (cloud feature vectors and cloud feature levels) that can be used for subsequent model training and prediction, laying the foundation for subsequent steps.

[0067] Step S1400, obtaining a real-time cloud map feature vector and a real-time cloud map feature level based on the real-time fusion cloud map image and the cloud map hierarchical analysis model; Furthermore, if Figure 5 As shown, step S1400 includes: Step S1410, input the real-time fused cloud image into the cloud classification analysis model to obtain the real-time cloud feature vector and the cloud feature level probability vector p=[p1, p2, p3, p4]; where p i Indicates the probability value of the real-time fusion cloud image being the i-th level feature level, 1≤i≤4; Step S1420, obtain the cloud image feature level probability vector p i The maximum value of is used as the real-time cloud feature level of the real-time fused cloud image.

[0068] Specifically, step S1400 uses the cloud image grading and parsing model trained in step S1300 to parse the real-time fused cloud image generated in step S1140 to obtain a real-time cloud image feature vector and a real-time cloud image feature level. The real-time fused cloud image refers to the fused image obtained by fusing the sky cloud image of the ground camera and the satellite cloud image data at the current moment. The real-time cloud image feature vector refers to the feature vector extracted by the cloud image grading and parsing model based on the real-time fused cloud image at the current moment. The vector contains the key feature information of the current cloud image, such as the texture, shape, color, edge and other features of the cloud. The real-time cloud image feature level refers to the cloud image feature level predicted by the cloud image grading and parsing model based on the real-time fused cloud image at the current moment. The level is obtained based on the cloud image feature level probability vector. Specifically, the level with the highest probability is taken as the final level.

[0069] The cloud feature level probability vector p is a probability vector containing 4 elements, where p i Indicates the probability value of the current cloud image predicted by the model to belong to the i-th level feature level. For example, a possible output is [0.1, 0.2, 0.6, 0.1], which means that the model believes that the probability of the current cloud image belonging to the first level feature level is 0.1, the probability of belonging to the second level feature level is 0.2, the probability of belonging to the third level feature level is 0.6, and the probability of belonging to the fourth level feature level is 0.1.

[0070] The cloud feature level probability vector represents the model's estimate of the probability that the current cloud belongs to each feature level. Selecting the level corresponding to the maximum probability value as the final feature level is a common decision-making method based on the maximum a posteriori probability criterion. That is, selecting the category with the highest probability as the prediction result can minimize the risk of prediction error. Although this method is simple, it is effective in most cases, especially when the model's prediction confidence is high (for example, the maximum probability value is much larger than other probability values).

[0071] Step S1500, obtain a cloud map feature sequence, a cloud map feature level sequence and a cloud layer motion vector sequence, and obtain a cloud map occlusion index for each photovoltaic power generation unit within a preset time in the future according to the cloud map feature sequence, the cloud map feature level sequence, the cloud layer motion vector sequence and a pre-built cloud map occlusion prediction model; arrange the cloud map occlusion index of each photovoltaic power generation unit in chronological order to generate cloud map occlusion index time series data.

[0072] Further, step S1500 includes: Step S1510, generating a fused cloud image sequence according to the fused cloud image; the fused cloud image sequence includes n4 fused cloud images, n4≥2; based on the fused cloud image sequence, obtaining a cloud layer motion vector; the cloud layer motion vector is a two-dimensional vector [Vx, Vy], wherein Vx represents the speed in the horizontal direction, and Vy represents the speed in the vertical direction; Specifically, the purpose of step S1500 is to predict the cloud map occlusion index of each photovoltaic power generation unit in the future period of time based on the real-time cloud map features and cloud map levels obtained in step S1400, combined with cloud movement information, and using a pre-built cloud map occlusion prediction model, and generate cloud map occlusion index time series data.

[0073] The fused cloud image sequence consists of a series of fused cloud image images arranged in time order, which are acquired through step S1140. In order to capture the movement of the clouds, at least two images are required (n4≥2). Usually, in order to more accurately estimate the movement of the clouds, more images are used, such as 10 images taken every 1 minute in the past 10 minutes. The cloud motion vector is a two-dimensional vector [Vx, Vy], which is used to describe the speed and direction of the movement of the clouds on the image plane. Vx represents the velocity component of the clouds in the horizontal direction (x-axis), and the unit can be pixels / second or meters / second (needs to be converted according to the actual spatial resolution of the image). Positive values ​​indicate movement to the right, and negative values ​​indicate movement to the left. Vy represents the velocity component of the clouds in the vertical direction (y-axis), and the unit is the same as Vx. Positive values ​​indicate downward movement, and negative values ​​indicate upward movement.

[0074] Common methods for calculating cloud motion vectors include optical flow and block matching. The optical flow method estimates the motion of pixels based on the temporal change of the grayscale value of the image. The optical flow method assumes that the motion of objects in the image sequence remains unchanged in a short period of time, that is, the grayscale value of the pixels of the same object between adjacent frames remains constant. By calculating the grayscale gradient of the pixel, the motion vector of the pixel can be estimated. Common optical flow algorithms include the Lucas-Kanade algorithm and the Horn-Schunck algorithm. The block matching method divides the image into several blocks (for example, 16x16 pixels), and then searches for the block that is most similar to the current block in the next frame of the image (for example, using the mean square error or the sum of absolute differences as a similarity measure) to obtain the motion vector of the block. The block matching method assumes that all pixels in the block have the same motion vector.

[0075] The movement of clouds is the main factor causing short-term fluctuations in photovoltaic output. By extracting the cloud motion vector, we can understand the direction and speed of cloud movement, and thus predict the position and obstruction of clouds in the future. For example, if the cloud motion vector shows that the clouds are moving towards a photovoltaic power generation unit, then it can be predicted that the probability of the photovoltaic power generation unit being obstructed by clouds will increase in the future. This prediction based on cloud motion information is more accurate than prediction based solely on cloud image information at the current moment, because it takes into account the dynamic changes of clouds.

[0076] Step S1520, generating a cloud feature sequence according to the real-time cloud feature vector; generating a cloud feature level sequence according to the real-time cloud feature level; generating a cloud motion vector sequence according to the cloud motion vector; Specifically, the cloud feature sequence is to arrange the real-time cloud feature vectors in a period of time in chronological order to form a cloud feature sequence. For example, if a real-time cloud feature vector is obtained every 1 minute, then the cloud feature vectors of the past 10 minutes form a cloud feature sequence of length 10. The cloud feature level sequence is to arrange the real-time cloud feature levels in a period of time in chronological order to form a cloud feature level sequence. For example, if a real-time cloud feature level is obtained every 1 minute, then the cloud feature levels of the past 10 minutes form a cloud feature level sequence of length 10. The cloud motion vector sequence is to arrange the cloud motion vectors in a period of time in chronological order to form a cloud motion vector sequence. For example, if a cloud motion vector is calculated every 1 minute, then the cloud motion vectors of the past 10 minutes form a cloud motion vector sequence of length 10. The cloud occlusion prediction model needs to predict the future cloud occlusion index based on the cloud features, cloud levels and cloud motion information in the past period of time. Organizing this information into a time series can facilitate the model to learn and utilize the temporal correlation between this information. For example, the model can learn how the cloud feature vector usually changes when the cloud feature level changes from level 2 to level 3, and the relationship between this change and the speed of cloud movement, etc. This time series-based modeling method can better capture the dynamic characteristics of cloud changes, thereby improving the accuracy of predictions.

[0077] Step S1530, dividing the cloud feature sequence into a cloud feature modeling sequence and a cloud feature occlusion prediction sequence; dividing the cloud feature level sequence into a cloud feature level modeling sequence and a cloud feature level occlusion prediction sequence; dividing the cloud layer motion vector sequence into a cloud layer motion vector modeling sequence and a cloud layer motion vector occlusion prediction sequence; Specifically, the purpose of step S1530 is to divide the cloud feature sequence, cloud feature level sequence and cloud layer motion vector sequence generated in step S1520 into a modeling sequence and an occlusion prediction sequence. The modeling sequence is used to train the cloud occlusion prediction model, while the occlusion prediction sequence is used to input into the trained model during actual prediction to obtain the future cloud occlusion index.

[0078] Usually, the first part of each sequence is used as the modeling sequence, and the second part is used as the occlusion prediction sequence. For example, for a cloud feature sequence of length 30, the first 20 data points can be used as the cloud feature modeling sequence, and the last 10 data points can be used as the cloud feature occlusion prediction sequence. It is necessary to ensure that the divided modeling sequence and occlusion prediction sequence are continuous in timestamp. In other words, the start timestamp of the occlusion prediction sequence should be immediately after the end timestamp of the modeling sequence. The modeling sequence is used to train the cloud occlusion prediction model. The model will learn the relationship between the cloud feature modeling sequence, the cloud feature level modeling sequence, the cloud layer motion vector modeling sequence and the historical cloud occlusion index sequence. The occlusion prediction sequence is used to input into the trained cloud occlusion prediction model during actual prediction to obtain the future cloud occlusion index.

[0079] The training and prediction of the cloud occlusion prediction model require different data. In the training phase, the model needs to learn the rules in the historical data; in the prediction phase, the model needs to predict the future situation based on the current information. Therefore, the data needs to be divided into a modeling sequence and an occlusion prediction sequence. This division method provides a clear definition of data input and output for the training and prediction of the cloud occlusion prediction model, ensures the logical consistency of model training and prediction, and avoids the model "seeing" future data during the training phase, thereby ensuring the reliability of the prediction.

[0080] Step S1540, converting the historical output data of the distributed photovoltaic system in the target area into a historical cloud map occlusion index sequence; constructing and training a cloud map occlusion prediction model according to the cloud map feature modeling sequence, the cloud map feature level modeling sequence, the cloud layer motion vector modeling sequence and the historical cloud map occlusion index sequence, and obtaining a trained cloud map occlusion prediction model; Specifically, the core of step S1540 is to build and train a cloud occlusion prediction model. The model takes cloud feature modeling sequence, cloud feature level modeling sequence, and cloud layer motion vector modeling sequence as input, and uses historical cloud occlusion index sequence as training target (label), and learns the relationship between input data and cloud occlusion index through training.

[0081] Since the output of the cloud occlusion prediction model is the cloud occlusion index, it is necessary to convert the historical output data into a historical cloud occlusion index sequence. A commonly used conversion method is based on the sunny day index model: 1. Calculate the Clear Sky Index: The Clear Sky Index is a dimensionless parameter that represents the ratio of actual solar radiation to solar radiation under clear sky conditions. The Byrd Clear Sky Model or other similar models are usually used to calculate solar radiation under clear sky conditions based on parameters such as geographic location, time, and solar altitude angle.

[0082] 2. Calculate the cloud obstruction index: Using the sunny index and actual PV output data, the cloud obstruction index can be estimated. For example, the cloud obstruction index can be defined as 1 minus the ratio of the actual PV output to the PV output under sunny conditions. The PV output under sunny conditions can be estimated based on parameters such as sunny solar radiation and the installed capacity of the PV power generation unit.

[0083] A long short-term memory network is selected to construct a cloud map occlusion prediction model. When training the cloud map occlusion prediction model, the cloud map feature modeling sequence, the cloud map feature level modeling sequence, and the cloud layer motion vector modeling sequence are used as inputs, and the historical cloud map occlusion index sequence is used as a label and input into the cloud map occlusion prediction model. The model performs forward propagation calculation based on the input sequence data to obtain a predicted cloud map occlusion index sequence. The predicted cloud map occlusion index sequence is compared with the historical cloud map occlusion index sequence to calculate the value of the loss function. Commonly used loss functions include mean square error (MSE), mean absolute error (MAE), etc. According to the value of the loss function, the gradient of the model parameters is calculated by the back propagation algorithm. Use an optimizer (such as Adam, SGD, etc.) to update the model parameters according to the calculated gradients. Repeat forward propagation, loss calculation, and back propagation until the model converges or reaches a predetermined number of iterations.

[0084] Step S1540 constructs and trains a cloud occlusion prediction model to obtain a model that can predict the future cloud occlusion index based on cloud features, cloud levels, and cloud motion information. This model is a key component for realizing short-term dynamic prediction of photovoltaic output, and its prediction accuracy directly affects the final photovoltaic output prediction accuracy.

[0085] Step S1550, obtaining the cloud map occlusion index of each photovoltaic power generation unit within a preset time in the future according to the cloud map feature occlusion prediction sequence, the cloud map feature level occlusion prediction sequence, the cloud layer motion vector occlusion prediction sequence and the trained cloud map occlusion prediction model.

[0086] Specifically, the cloud map feature occlusion prediction sequence, cloud map feature level occlusion prediction sequence and cloud motion vector occlusion prediction sequence obtained in step S1530 are input into the trained cloud map occlusion prediction model to obtain the cloud map occlusion index of each photovoltaic power generation unit within a preset time in the future (for example, the next 15 minutes). The cloud map occlusion index is a bridge connecting cloud map information and photovoltaic output prediction. The trained cloud map occlusion prediction model is used to predict the cloud map occlusion index of each photovoltaic power generation unit within a period of time in the future based on the input cloud map feature occlusion prediction sequence, cloud map feature level occlusion prediction sequence and cloud motion vector occlusion prediction sequence. This step is a key link in realizing short-term dynamic prediction of photovoltaic output, and the accuracy of the prediction result directly affects the final photovoltaic output prediction accuracy.

[0087] Step S2000, based on the micrometeorological data of the target area, the time series data of the shadow impact factor, the time series data of the cloud map occlusion index, the historical output data of the distributed photovoltaic system in the target area and the pre-built short-term output dynamic prediction model, obtain the predicted output values ​​of the n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the future preset time, and generate a predicted output sequence; obtain the actual output sequence of the n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the future preset time, and calculate the output prediction error according to the predicted output sequence and the actual output sequence; adjust the short-term output dynamic prediction model according to the output prediction error and the preset confidence interval.

[0088] Furthermore, step S2000 includes: Step S2100, constructing a short-term output dynamic prediction model based on the historical micro-meteorological data of the target area, the time series data of the shadow impact factor, the time series data of the cloud map shielding index, and the historical output data of the distributed photovoltaic system in the target area; Furthermore, if Figure 6 As shown, step S2100 includes: Step S2110, dividing the shadow influence factor time series data into shadow influence factor modeling time series data and shadow influence factor output prediction time series data; dividing the cloud map occlusion index time series data into cloud map occlusion index modeling time series data and cloud map occlusion index output prediction time series data; Specifically, the purpose of step S2110 is to divide the shadow impact factor time series data generated in step S1200 and the cloud map shielding index time series data generated in step S1500 into two parts, namely, modeling time series data and output prediction time series data. The modeling time series data is used to train the short-term output dynamic prediction model, and the output prediction time series data is used to input into the trained model during actual prediction to obtain the future photovoltaic output. Usually, the first part of each time series data is used as the modeling time series data, and the latter part is used as the output prediction time series data. For example, for a shadow impact factor time series data with a length of 100, the first 80 data points can be used as the shadow impact factor modeling time series data, and the latter 20 data points can be used as the shadow impact factor output prediction time series data. The modeling time series data and the output prediction time series data are continuous in timestamp. The starting timestamp of the output prediction time series data is immediately after the ending timestamp of the modeling time series data. In addition, the timestamp of the output prediction time series data needs to be aligned with the timestamp of the real-time micrometeorological data input in step S2200. This step divides the shadow impact factor time series data and cloud map occlusion index time series data into modeling time series data and output prediction time series data, providing a clear definition of data input and output for the training and prediction of the short-term output dynamic prediction model. This division method ensures the logical consistency of model training and prediction, and aligns with the timestamp of real-time micro-meteorological data, which is convenient for practical application.

[0089] Step S2120, aligning the historical micro-meteorological data of the target area, the shadow impact factor modeling time series data, the cloud map occlusion index modeling time series data and the historical output data according to the timestamps; Specifically, the data from different sources (historical micro-meteorological data, shadow impact factor modeling time series data, cloud map occlusion index modeling time series data, and historical output data) are aligned according to their respective timestamps. After alignment, each row of data corresponds to the same time point and contains all relevant information at that moment. Data from different sources may have different sampling frequencies. For example, the sampling frequency of micro-meteorological data may be 1 minute, while the sampling frequency of cloud map occlusion index may be 5 minutes, and the sampling frequency of output data may be 15 minutes. In this case, the data needs to be resampled so that they have the same sampling frequency. The short-term output dynamic prediction model needs to consider the influence of multiple factors at the same time, including micro-meteorological data, shadow impact factor, cloud map occlusion index, and historical output data. These data come from different sources and have different sampling frequencies and timestamps. In order to fuse these data into a unified model, they need to be aligned according to the timestamp to ensure the consistency of the data in time. The model can receive all relevant information at each time step, laying the foundation for subsequent data fusion and model training.

[0090] Step S2130, fusing the aligned historical micrometeorological data, shadow impact factor modeling time series data, cloud image occlusion index modeling time series data and historical output data to generate a fused data set; Specifically, the aligned data from different sources (historical micro-meteorological data, shadow impact factor modeling time series data, cloud map occlusion index modeling time series data and historical output data) are merged according to timestamps to form a data set containing all information. Each row of the fused data set corresponds to a time point, and each column corresponds to a feature (such as temperature, humidity, wind speed, shadow impact factor, cloud map occlusion index, output, etc.). The short-term output dynamic prediction model needs to consider the influence of multiple factors at the same time, including micro-meteorological data, shadow impact factor, cloud map occlusion index and historical output data. By fusing these data into a unified data set, it is convenient for the model to learn the relationship between these factors at the same time, thereby improving the accuracy of the prediction. Each row of the fused data set contains all the relevant information at that moment. The model can directly extract features from the data set and learn the mapping relationship between features and target variables (PV output).

[0091] Step S2140, based on the fused data set, construct and train a short-term output dynamic prediction model; the short-term output dynamic prediction model adopts a long short-term memory network, including an input layer, an LSTM layer, a fully connected layer and an output layer; the output layer contains n5 neurons, each neuron outputs a predicted output value of a photovoltaic power generation unit within a preset time in the future.

[0092] Specifically, the short-term output dynamic prediction model adopts a long short-term memory network, including an input layer, an LSTM layer, a fully connected layer, and an output layer; the input layer receives feature data in the fusion data set, including historical micro-meteorological data, shadow impact factor modeling time series data, and cloud map occlusion index modeling time series data. The specific time span of the input data needs to be determined according to the actual situation. For example, the data of the past 1 hour or 2 hours can be selected as input. The LSTM layer is the core of the model and is used to extract the time features in the input data. The number of LSTM layers and the number of neurons in each LSTM unit are hyperparameters of the model and need to be adjusted according to the actual situation. Usually, 1-3 LSTM layers can be used, each containing dozens to hundreds of neurons. The fully connected layer connects the output of the LSTM layer to one or more fully connected layers to map the time series features extracted by the LSTM layer to the final output space. The output layer contains n5 neurons, corresponding to n3 photovoltaic power generation units, n5=n3. Each neuron outputs the predicted output value of a photovoltaic power generation unit within a preset time in the future. For example, if the preset time is 15 minutes in the future and a value is predicted every 5 minutes, each neuron will output 3 predicted values.

[0093] During the training process of the short-term output dynamic prediction model, the fusion data set is divided into a training set, a validation set, and a test set. The data in the training set is input into the model, and the model performs forward propagation calculation based on the input sequence data to obtain the predicted output value. The predicted output value is compared with the actual output value (historical output data) to calculate the value of the loss function. This method can select the mean absolute error (MAE) as the loss function. According to the value of the loss function, the gradient of the model parameters is calculated by the back propagation algorithm. Use an optimizer (such as Adam, SGD, etc.) to update the parameters of the model according to the calculated gradient. Repeat forward propagation, loss calculation, and back propagation until the model converges or reaches a predetermined number of iterations. Use the validation set to evaluate the model in the training process and monitor the performance of the model on the validation set, such as MAE, RMSE and other indicators. According to the performance of the validation set, adjust the hyperparameters of the model, such as the learning rate, the number of LSTM layers, the number of neurons, etc., to prevent overfitting and improve the generalization ability of the model. Use the test set to test the trained model and evaluate the final performance of the model.

[0094] Step S2100 builds a short-term dynamic output prediction model that takes multiple factors into consideration, which can more comprehensively reflect the changing pattern of photovoltaic output and thus improve the prediction accuracy. The model integrates micro-meteorological data, shadow impact factors, cloud map occlusion index and historical output data, and learns the complex relationship between these data through the LSTM network to achieve accurate prediction of future photovoltaic output. This prediction method that integrates multiple factors is similar to a comprehensive analysis of multiple factors, and can more accurately grasp the changing trend of photovoltaic output.

[0095] Step S2200, based on the real-time micro-meteorological data of the target area, the output prediction time series data of the shadow influence factor, the output prediction time series data of the cloud map shielding index and the short-term output dynamic prediction model, obtain the predicted output values ​​of n3 photovoltaic power generation units in the distributed photovoltaic system of the target area within a preset time in the future, and generate a predicted output sequence; Step S2300, obtaining the actual output sequence of n3 photovoltaic power generation units in the target area distributed photovoltaic system within a preset future time, and calculating the output prediction error according to the predicted output sequence and the actual output sequence; The calculation of the output prediction error includes:

[0096] in: : Output prediction error.

[0097] : The total number of time steps to predict (for example, to predict the next 15 minutes, with one data point every 5 minutes, then .

[0098] : time step index, .

[0099] : Dynamic adjustment factor of the absolute error term, determined according to the average irradiance of the previous few time steps. If the average irradiance of the previous few time steps is large, it means that the lighting conditions are good. At this time, the predicted absolute error may be large, so α(t) can be appropriately increased. The formula is designed as:

[0100] in represents the average DNI value of the first k time steps, is a preset DNI threshold.

[0101] : Dynamic adjustment factor of the relative error term; determined according to the predicted output value of the current time step. If the predicted output value of the current time step is small, it means that the photovoltaic system may be in a low-power operation state. At this time, the relative error may be large, so β(t) can be appropriately increased. The formula can be designed as:

[0102] in represents the predicted output value at time t, is the maximum installed capacity of the photovoltaic system.

[0103] : Dynamic adjustment factor of the time lag error term; determined according to the cloud movement speed in the previous few time steps. If the cloud movement speed in the previous few time steps is fast, it means that the cloud occlusion changes quickly. At this time, the time lag error may be large, so γ(t) can be appropriately increased. The formula can be designed as:

[0104] in represents the average cloud motion speed in the first k time steps, is a preset speed threshold. The value range is Between and satisfy .

[0105] : The absolute error term at time t, which calculates the average absolute error of all photovoltaic units at time t. The calculation formula is as follows:

[0106] in: : The predicted output value of the j-th photovoltaic power generation unit at time t, obtained through the short-term output dynamic prediction model, 1≤j≤n3; : The actual output value of the j-th photovoltaic power generation unit at time t is obtained in real time through the monitoring platform of the photovoltaic system.

[0107] : The relative error term at time t, which calculates the average relative error of all photovoltaic units at time t, taking into account the situation where the actual output is zero or close to zero. The calculation formula is as follows:

[0108] in, is a very small positive number (such as 0.001) used to prevent the denominator from being zero and avoid numerical calculation problems.

[0109] : The time lag error term at time t is used to penalize the time offset between the predicted curve and the actual curve. The calculation formula is as follows:

[0110] Where Δt is the preset time offset step, for example, Δt=1 means considering an offset of one time step (e.g. 5 minutes). This item calculates the minimum absolute error between the predicted output value and the actual output value of each PV unit within a time offset of one step, and then takes the average of all units. The purpose is to evaluate whether the shape of the predicted curve is similar to the actual curve, even if there is a time deviation.

[0111] along with and The difference between them increases as the along with When the actual lighting conditions are good, the predicted absolute error has a greater impact on the total error. Will cause Increase. along with When the PV system is in low power operation, the relative error of the prediction has a greater impact on the total error. Will cause Increase. along with When the cloud moves faster, the predicted time lag error has a greater impact on the total error. At this time, a larger γ(t) will lead to Increase. As the forecast time step T increases, more complex trends may appear. The forecast errors of a single time step will be averaged, but the cumulative effect of the errors may also be more significant. The specific trend of change depends on the performance of the forecast model at each time step.

[0112] This formula comprehensively considers the absolute error, relative error and time lag error, and can more comprehensively evaluate the performance of the prediction model. The introduction of a dynamic adjustment factor allows the error evaluation to be adaptively adjusted according to different working conditions, thereby improving the accuracy of the evaluation. In view of the characteristics of distributed photovoltaic systems, the error is calculated for each photovoltaic power generation unit and averaged, which can more accurately reflect the prediction error of the entire system. This formula is a more refined and dynamic output prediction error evaluation tool, which can provide more accurate guidance for the development and optimization of distributed photovoltaic short-term output prediction models. By analyzing As well as the values ​​of the individual components, you can gain insight into the performance of the model and identify areas for improvement, thereby increasing the accuracy and reliability of your predictions.

[0113] Step S2400: adjusting the short-term output dynamic prediction model according to the output prediction error and the preset confidence interval.

[0114] Furthermore, if Figure 7 As shown, step S2400 includes: Step S2410: if the output prediction error is within a preset confidence interval, the short-term output dynamic prediction model is kept unchanged; Step S2420: If the output prediction error exceeds a preset confidence interval, the short-term output dynamic prediction model is retrained.

[0115] Specifically, by inputting real-time micro-meteorological data, shadow influence factor output prediction time series data and cloud map shielding index output prediction time series data, the short-term output dynamic prediction model can predict the output value of each photovoltaic power generation unit in the target area distributed photovoltaic system within the preset time in the future, and arrange them in chronological order to form a predicted output sequence. The preset confidence interval refers to an error range pre-set according to historical prediction error data or other methods. For example, the confidence interval can be set to a 95% confidence level, that is, it is believed that 95% of the prediction errors should fall within this interval. When the output prediction error of the model is within the preset confidence interval, it is considered that the prediction performance of the model is good, no adjustment is required, and the model remains unchanged. When the output prediction error of the model exceeds the preset confidence interval, it is considered that the prediction performance of the model has decreased, and the actual output sequence obtained in step S2300 and the corresponding real-time micro-meteorological data, shadow influence factor time series data, and cloud map shielding index time series data are added to the training data set of the model, and the model is retrained using the updated training data set so that the model learns the latest data rules. By adding the latest actual output data and corresponding environmental data to the training data set and retraining the model, the model can learn the latest data rules, thereby improving the prediction accuracy of the model. Step S2400 adjusts the model according to the output prediction error and confidence interval to achieve adaptive updating of the model and improve the prediction accuracy and robustness of the model.

[0116] Example 2

[0117] This embodiment provides a distributed photovoltaic short-term output dynamic prediction system based on the first embodiment. Figure 8 As shown, including: Data acquisition module: used to obtain micro-meteorological data, sky cloud images, satellite cloud data of the target area, and historical output data of the distributed photovoltaic system in the target area; Calculation module: used to calculate the building shadow shielding coefficient of the target area and generate the time series data of the shadow influence factor; based on the sky cloud image, satellite cloud image data and the pre-built cloud shielding prediction model, obtain the cloud shielding index time series data; the distributed photovoltaic system in the target area includes n3 photovoltaic power generation units; Output prediction module: Based on the micro-meteorological data of the target area, the time series data of the shadow impact factor, the time series data of the cloud map shielding index, the historical output data of the distributed photovoltaic system in the target area and the pre-built short-term output dynamic prediction model, the predicted output values ​​of n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the preset time in the future are obtained, and the predicted output sequence is generated; Model adjustment module: used to obtain the actual output sequence of n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the preset future time, calculate the output prediction error based on the predicted output sequence and the actual output sequence; adjust the short-term output dynamic prediction model based on the output prediction error and the preset confidence interval.

[0118] In the data acquisition module, the acquisition of micro-meteorological data, sky cloud image, satellite cloud image data of the target area, and historical output data of the distributed photovoltaic system in the target area includes: Step S1110, deploying n1 micro-meteorological monitoring stations in the target area, and obtaining micro-meteorological data of the target area through the micro-meteorological monitoring stations; the micro-meteorological data includes temperature, humidity, wind speed, wind direction, air pressure, direct radiant illuminance (DNI) and diffuse radiant illuminance (DHI); the micro-meteorological data includes historical micro-meteorological data and real-time micro-meteorological data; Step S1120, deploying n2 wide-angle cameras at high points in the target area to collect sky cloud images; the sky cloud images include historical sky cloud images and real-time sky cloud images; Step S1130, obtaining satellite cloud image data of the target area, and geo-referencing the satellite cloud image data with the sky cloud image captured by the camera; the satellite cloud image data includes historical satellite cloud image data and real-time satellite cloud image data; Step S1140, fusing the sky cloud image and the satellite cloud image data to obtain a fused cloud image; the fused cloud image includes a historical fused cloud image and a real-time fused cloud image; Step S1150, loading the historical output data of the distributed photovoltaic system in the target area, wherein the distributed photovoltaic system in the target area includes n3 photovoltaic power generation units, and the historical output data of the distributed photovoltaic system includes the actual output, installed capacity and timestamp of the n3 photovoltaic power generation units.

[0119] In the calculation module, the calculation of the building shadow shielding coefficient of the target area and the generation of the shadow impact factor time series data include: Step S1210, obtaining a three-dimensional building model of the target area, including the height, outline and location information of the building; Step S1220, obtaining the geographical location of the target area, and calculating the sun position information at each moment according to the geographical location of the target area, wherein the sun position information includes the sun altitude angle and the azimuth angle; Step S1230, based on the three-dimensional building model of the target area and the sun position information at each moment, calculate the building shadow shielding coefficients of n1 micro-meteorological monitoring stations at each moment and the building shadow shielding coefficients of n3 photovoltaic power generation units at each moment; Step S1240, arranging the building shadow shielding coefficients of each micro-meteorological monitoring station and each photovoltaic power generation unit in chronological order to generate time series data of shadow impact factors.

[0120] In the calculation module, the cloud map occlusion index time series data is obtained based on the sky cloud map image, satellite cloud map data and the pre-built cloud map occlusion prediction model, including: Based on the historical fused cloud image, a cloud image hierarchical parsing model based on convolutional neural network (CNN) is constructed; based on the real-time fused cloud image and the cloud image hierarchical parsing model, the real-time cloud image feature vector and real-time cloud image feature level are obtained; Generate a fused cloud image sequence according to the fused cloud image; the fused cloud image sequence includes n4 fused cloud images, where n4≥2; obtain a cloud layer motion vector based on the fused cloud image sequence; the cloud layer motion vector is a two-dimensional vector [Vx, Vy], where Vx represents a speed in the horizontal direction and Vy represents a speed in the vertical direction; Generate a cloud feature sequence based on the real-time cloud feature vector; generate a cloud feature level sequence based on the real-time cloud feature level; generate a cloud motion vector sequence based on the cloud motion vector; The cloud feature sequence is divided into a cloud feature modeling sequence and a cloud feature occlusion prediction sequence; the cloud feature level sequence is divided into a cloud feature level modeling sequence and a cloud feature level occlusion prediction sequence; the cloud layer motion vector sequence is divided into a cloud layer motion vector modeling sequence and a cloud layer motion vector occlusion prediction sequence; The historical output data of the distributed photovoltaic system in the target area is converted into a historical cloud map occlusion index sequence; a cloud map occlusion prediction model is constructed and trained according to the cloud map feature modeling sequence, the cloud map feature level modeling sequence, the cloud layer motion vector modeling sequence and the historical cloud map occlusion index sequence to obtain a trained cloud map occlusion prediction model; According to the cloud image feature occlusion prediction sequence, the cloud image feature level occlusion prediction sequence, the cloud layer motion vector occlusion prediction sequence and the trained cloud image occlusion prediction model, the cloud image occlusion index of each photovoltaic power generation unit within the future preset time is obtained.

[0121] The cloud image classification parsing model based on the convolutional neural network (CNN) is constructed based on the historical fusion cloud image, including: Step S1310, annotating the historical fusion cloud image, and dividing each historical fusion cloud image into four levels, namely, the first-level feature level, the second-level feature level, the third-level feature level and the fourth-level feature level; Step S1320, determining a cloud layer feature vector loss function and a classification loss function, performing weighted average of the cloud layer feature vector loss function and the classification loss function to obtain a loss function of a cloud image hierarchical parsing model; Step S1330, the annotated historical fusion cloud image is used to form a sample data set; according to the cloud layer feature vector loss function, a sample pair is constructed in the sample data set; the annotated historical fusion cloud image includes a feature level label; Step S1340, training the cloud graph hierarchical parsing model according to the sample data set of the sample pairs and the loss function of the cloud graph hierarchical parsing model.

[0122] The real-time cloud image feature vector and the real-time cloud image feature level are obtained based on the real-time fusion cloud image and the cloud image hierarchical analysis model, including: Step S1410, input the real-time fused cloud image into the cloud classification analysis model to obtain the real-time cloud feature vector and the cloud feature level probability vector p=[p1, p2, p3, p4]; where p i Indicates the probability value of the real-time fusion cloud image being the i-th level feature level, 1≤i≤4; Step S1420, obtain the cloud image feature level probability vector p i The maximum value of is used as the real-time cloud feature level of the real-time fused cloud image.

[0123] In the output prediction module, the predicted output values ​​of n3 photovoltaic power generation units in the target area distributed photovoltaic system within a preset future time are obtained, and the predicted output sequence is generated, including: Step S2100, constructing a short-term output dynamic prediction model based on the historical micro-meteorological data of the target area, the time series data of the shadow impact factor, the time series data of the cloud map shielding index, and the historical output data of the distributed photovoltaic system in the target area; Step S2200, based on the real-time micrometeorological data of the target area, the output prediction time series data of the shadow influence factor, the output prediction time series data of the cloud map occlusion index and the short-term output dynamic prediction model, obtain the predicted output values ​​of n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the future preset time, and generate a predicted output sequence.

[0124] The step S2100 includes: Step S2110, dividing the shadow influence factor time series data into shadow influence factor modeling time series data and shadow influence factor output prediction time series data; dividing the cloud map occlusion index time series data into cloud map occlusion index modeling time series data and cloud map occlusion index output prediction time series data; Step S2120, aligning the historical micro-meteorological data of the target area, the shadow impact factor modeling time series data, the cloud map occlusion index modeling time series data and the historical output data according to the timestamps; Step S2130, fusing the aligned historical micrometeorological data, shadow impact factor modeling time series data, cloud image occlusion index modeling time series data and historical output data to generate a fused data set; Step S2140, based on the fused data set, construct and train a short-term output dynamic prediction model; the short-term output dynamic prediction model adopts a long short-term memory network, including an input layer, an LSTM layer, a fully connected layer and an output layer; the output layer contains n5 neurons, each neuron outputs a predicted output value of a photovoltaic power generation unit within a preset time in the future.

[0125] In the model adjustment module, the adjustment of the short-term output dynamic prediction model according to the output prediction error and the preset confidence interval includes: Step S2410: if the output prediction error is within a preset confidence interval, the short-term output dynamic prediction model is kept unchanged; Step S2420: If the output prediction error exceeds a preset confidence interval, the short-term output dynamic prediction model is retrained.

[0126] The methods, systems, and devices of the present application may be implemented in many ways. For example, the methods, systems, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers recording media storing programs for executing the method according to the present application.

[0127] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0128] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A distributed photovoltaic short-term output dynamic prediction method, characterized in that: The method comprises: Obtain micro-meteorological data, sky cloud image, satellite cloud image data, and historical output data of the distributed photovoltaic system in the target area; calculate the building shadow shielding coefficient in the target area and generate time series data of the shadow impact factor; obtain the cloud shielding index time series data based on the sky cloud image, satellite cloud image data and the pre-built cloud shielding prediction model; the distributed photovoltaic system in the target area includes n3 photovoltaic power generation units; Based on the micrometeorological data of the target area, the time series data of the shadow impact factor, the time series data of the cloud map occlusion index, the historical output data of the distributed photovoltaic system in the target area and the pre-built short-term output dynamic prediction model, the predicted output values ​​of the n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the future preset time are obtained, and a predicted output sequence is generated; the actual output sequence of the n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the future preset time is obtained, and the output prediction error is calculated according to the predicted output sequence and the actual output sequence; the short-term output dynamic prediction model is adjusted according to the output prediction error and the preset confidence interval.

2. The distributed photovoltaic short-term output dynamic prediction method according to claim 1 is characterized in that: The acquisition of micro-meteorological data, sky cloud images, satellite cloud data of the target area, and historical output data of the distributed photovoltaic system in the target area includes: Deploy n1 micro-meteorological monitoring stations in the target area, and obtain micro-meteorological data of the target area through the micro-meteorological monitoring stations; the micro-meteorological data includes historical micro-meteorological data and real-time micro-meteorological data; Deploy n2 wide-angle cameras at high points in the target area to collect sky cloud images; Obtain satellite cloud image data of the target area and georeference the satellite cloud image data with the sky cloud image captured by the camera; Fusing the sky cloud image and the satellite cloud image data to obtain a fused cloud image; the fused cloud image includes a historical fused cloud image and a real-time fused cloud image; Load the historical output data of the distributed photovoltaic system in the target area.

3. The distributed photovoltaic short-term output dynamic prediction method according to claim 1 is characterized in that: The step of calculating the building shadow shielding coefficient of the target area and generating the shadow impact factor time series data comprises: The building three-dimensional model and the geographical location of the target area are obtained, and the sun position information at each moment is calculated according to the geographical location of the target area. Based on the building three-dimensional model of the target area and the sun position information at each moment, the building shadow shielding coefficients of n1 micro-meteorological monitoring stations and the building shadow shielding coefficients of n3 photovoltaic power generation units at each moment are calculated; the building shadow shielding coefficients of each micro-meteorological monitoring station and each photovoltaic power generation unit are arranged in chronological order to generate time series data of shadow influencing factors.

4. The distributed photovoltaic short-term output dynamic prediction method according to claim 1 is characterized in that: The cloud map occlusion index time series data obtained based on the sky cloud map image, satellite cloud map data and the pre-built cloud map occlusion prediction model includes: Generate a fused cloud image sequence according to the fused cloud image; the fused cloud image sequence includes n4 fused cloud images, where n4≥2; obtain a cloud layer motion vector based on the fused cloud image sequence; the cloud layer motion vector is a two-dimensional vector [Vx, Vy], where Vx represents a speed in the horizontal direction and Vy represents a speed in the vertical direction; Based on the historical fusion cloud image, a cloud image hierarchical parsing model based on convolutional neural network is constructed; Based on the real-time fusion cloud image and cloud classification analysis model, the real-time cloud feature vector and real-time cloud feature level are obtained; Generate a cloud feature sequence based on the real-time cloud feature vector; generate a cloud feature level sequence based on the real-time cloud feature level; generate a cloud motion vector sequence based on the cloud motion vector; According to the cloud feature sequence, cloud feature level sequence, cloud motion vector sequence and the pre-built cloud occlusion prediction model, the cloud occlusion index of each photovoltaic power generation unit within the preset future time is obtained; the cloud occlusion index of each photovoltaic power generation unit is arranged in chronological order to generate cloud occlusion index time series data.

5. The distributed photovoltaic short-term output dynamic prediction method according to claim 4 is characterized in that: The obtaining of the real-time cloud image feature vector and the real-time cloud image feature level includes: The real-time fused cloud image is input into the cloud image hierarchical analysis model to obtain the real-time cloud image feature vector and cloud image feature level probability vector p=[p1,p2,p3,p4]; where p i Indicates the probability value of the real-time fusion cloud image being the i-th level feature level, 1≤i≤4; Take the cloud image feature level probability vector p i The maximum value of is used as the real-time cloud feature level of the real-time fused cloud image.

6. The distributed photovoltaic short-term output dynamic prediction method according to claim 4 is characterized in that: The cloud image shielding index of each photovoltaic power generation unit within a preset time in the future is obtained by: The cloud feature sequence is divided into a cloud feature modeling sequence and a cloud feature occlusion prediction sequence; the cloud feature level sequence is divided into a cloud feature level modeling sequence and a cloud feature level occlusion prediction sequence; the cloud layer motion vector sequence is divided into a cloud layer motion vector modeling sequence and a cloud layer motion vector occlusion prediction sequence; The historical output data of the distributed photovoltaic system in the target area is converted into a historical cloud map occlusion index sequence; a cloud map occlusion prediction model is constructed and trained according to the cloud map feature modeling sequence, the cloud map feature level modeling sequence, the cloud layer motion vector modeling sequence and the historical cloud map occlusion index sequence to obtain a trained cloud map occlusion prediction model; According to the cloud image feature occlusion prediction sequence, the cloud image feature level occlusion prediction sequence, the cloud layer motion vector occlusion prediction sequence and the trained cloud image occlusion prediction model, the cloud image occlusion index of each photovoltaic power generation unit within the future preset time is obtained.

7. The distributed photovoltaic short-term output dynamic prediction method according to claim 4 is characterized in that: The cloud image classification parsing model based on the convolutional neural network is constructed based on the historical fusion cloud image, including: Annotate the historical fusion cloud map images and divide each historical fusion cloud map image into four levels, namely, the first-level feature level, the second-level feature level, the third-level feature level and the fourth-level feature level; Determine the cloud layer feature vector loss function and the classification loss function, and perform weighted average of the cloud layer feature vector loss function and the classification loss function to obtain the loss function of the cloud image hierarchical analysis model; The annotated historical fusion cloud image is used to form a sample data set; according to the cloud layer feature vector loss function, sample pairs are constructed in the sample data set; The cloud graph hierarchical parsing model is trained based on the sample data set of the sample pairs and the loss function of the cloud graph hierarchical parsing model.

8. The distributed photovoltaic short-term output dynamic prediction method according to claim 1 is characterized in that: The method of obtaining the predicted output values ​​of n3 photovoltaic power generation units in the target area distributed photovoltaic system within a preset future time includes: The shadow influence factor time series data is divided into the shadow influence factor modeling time series data and the shadow influence factor output prediction time series data; the cloud map occlusion index time series data is divided into the cloud map occlusion index modeling time series data and the cloud map occlusion index output prediction time series data; Align the historical micro-meteorological data, shadow impact factor modeling time series data, cloud cover index modeling time series data and historical output data of the target area according to timestamps; The aligned historical micro-meteorological data, shadow impact factor modeling time series data, cloud map occlusion index modeling time series data and historical output data are fused to generate a fused data set; Based on the fused data set, a short-term output dynamic prediction model is constructed and trained; the short-term output dynamic prediction model adopts a long short-term memory network, including an input layer, an LSTM layer, a fully connected layer and an output layer; the output layer contains n5 neurons, each of which outputs the predicted output value of a photovoltaic power generation unit within a preset time in the future.

9. The distributed photovoltaic short-term output dynamic prediction method according to claim 8 is characterized in that: The step of obtaining the predicted output values ​​of n3 photovoltaic power generation units in the target area distributed photovoltaic system within a preset future time also includes: Based on the real-time micrometeorological data of the target area, the output prediction time series data of the shadow influence factor, the output prediction time series data of the cloud map shading index and the short-term output dynamic prediction model, the predicted output values ​​of n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the future preset time are obtained.

10. A distributed photovoltaic short-term output dynamic prediction system, which is used to implement the distributed photovoltaic short-term output dynamic prediction method according to any one of claims 1 to 9, characterized in that: The system comprises: Data acquisition module: used to obtain micro-meteorological data, sky cloud images, satellite cloud data of the target area, and historical output data of the distributed photovoltaic system in the target area; Calculation module: used to calculate the building shadow shielding coefficient of the target area and generate the time series data of the shadow influence factor; based on the sky cloud image, satellite cloud image data and the pre-built cloud shielding prediction model, obtain the cloud shielding index time series data; the distributed photovoltaic system in the target area includes n3 photovoltaic power generation units; Output prediction module: Based on the micro-meteorological data of the target area, the time series data of the shadow impact factor, the time series data of the cloud map shielding index, the historical output data of the distributed photovoltaic system in the target area and the pre-built short-term output dynamic prediction model, the predicted output values ​​of n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the preset time in the future are obtained, and the predicted output sequence is generated; Model adjustment module: used to obtain the actual output sequence of n3 photovoltaic power generation units in the distributed photovoltaic system in the target area within the preset future time, calculate the output prediction error based on the predicted output sequence and the actual output sequence; adjust the short-term output dynamic prediction model based on the output prediction error and the preset confidence interval.

Images