Device and method for improving accuracy of ultra-short-term prediction of distributed photovoltaic power generation

Abstract

The application discloses a device and method for improving the accuracy of distributed photovoltaic power ultra-short-term prediction, and relates to the technical field of distributed photovoltaic power prediction; the device comprises a data acquisition module, a data preprocessing module, a space-time feature fusion module, a prediction output module and a power supply module; the method collects unified data space-time reference through space-time matching of multiple data sources, divides weather scenes based on meteorological characteristics and performs scene adaptive preprocessing, dynamically updates power station space correlation weight combined with cloud movement law and fuses space-time features, and finally realizes 0-4 hour ultra-short-term power prediction through prediction deviation closed loop correction; the application solves the defects of the prior art, such as insufficient adaptability of the preprocessing mode and inability to capture dynamic spatial correlation, effectively improves the ultra-short-term prediction accuracy and stability, can support power grid dispatching, photovoltaic consumption and collaborative control, and is suitable for regional cluster type distributed photovoltaic scenes.

Description

Technical Field

[0001] This invention relates to the field of distributed photovoltaic power generation prediction technology, specifically to an apparatus and method for improving the accuracy of ultra-short-term prediction of distributed photovoltaic power generation. Background Technology

[0002] With the rapid popularization of distributed photovoltaic power generation, its power output is significantly affected by environmental factors such as irradiance, temperature, cloud cover, and wind speed, and has the characteristics of randomness and strong fluctuation. 0-4 hour ultra-short-term power forecasting is the core support for grid dispatch, photovoltaic consumption, and energy storage coordinated control, and is directly related to the grid connection stability and utilization efficiency of distributed photovoltaic systems. At present, distributed photovoltaic power stations are mostly scattered in areas such as building rooftops and industrial plants.

[0003] Existing distributed photovoltaic (PV) ultra-short-term power prediction technologies are mostly based on a single PV power plant. They rely on historical operating data and numerical weather prediction data to complete power prediction through conventional time-series prediction models. Some optimization schemes incorporate multi-source meteorological data such as satellite cloud images and combine them with spatial correlation analysis of regional power plants to improve the prediction effect. However, existing technologies still have obvious shortcomings in practical applications.

[0004] First, using a fixed mode to complete data preprocessing cannot adapt to the fluctuation characteristics of photovoltaic power output under different weather scenarios, and is prone to problems such as loss of key fluctuation features or noise residue.

[0005] Secondly, spatial feature mining uses fixed site relationships and does not consider the power fluctuation transmission law caused by cloud movement, thus failing to capture the dynamic spatial relationship between sites in the region.

[0006] Therefore, it is essential to design devices and methods to improve the accuracy of ultra-short-term forecasts of distributed photovoltaic power generation. Summary of the Invention

[0007] The purpose of this invention is to provide an apparatus and method for improving the accuracy of ultra-short-term power prediction for distributed photovoltaic power generation, in order to solve the problems mentioned in the background art, such as insufficient adaptability of existing prediction technology data preprocessing mode, inability to capture dynamic spatial correlation between power stations in the region, resulting in insufficient accuracy and stability of ultra-short-term power prediction, making it difficult to support grid dispatch, photovoltaic consumption and coordinated control.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] The first aspect is to provide a device for improving the accuracy of ultra-short-term prediction of distributed photovoltaic power generation, including a data acquisition module, a data preprocessing module, a spatiotemporal feature fusion module, and a prediction output module;

[0010] The data acquisition module is used to collect multi-source photovoltaic data from multiple distributed photovoltaic power stations within the target area, and to perform spatiotemporal benchmark unified processing on the data to output multi-source raw datasets.

[0011] The data preprocessing module receives multi-source raw datasets, preprocesses the corresponding photovoltaic data for different weather scenarios to obtain standardized time-series datasets, and assigns corresponding credibility weights to different photovoltaic data in the standardized time-series datasets according to the reliability level of the data sources.

[0012] The spatiotemporal feature fusion module receives a standardized time-series dataset. It incorporates a graph convolutional unit and a temporal convolutional unit. The graph convolutional unit dynamically updates the spatial correlation weights between photovoltaic power stations based on cloud movement features extracted from satellite cloud imagery. Using these updated spatial correlation weights, it extracts spatial correlation features corresponding to each type of photovoltaic data from the standardized time-series dataset. The temporal convolutional unit extracts temporal fluctuation features corresponding to each type of photovoltaic data from the standardized time-series dataset. The spatiotemporal feature fusion module weights the spatial correlation features and temporal fluctuation features of different photovoltaic data according to the confidence weights, then fuses the weighted features to output a comprehensive feature vector.

[0013] The prediction output module receives the comprehensive feature vector, generates the initial prediction value for the current period through the built-in prediction model, and uses the deviation between the measured power of the previous acquisition period and the initial prediction value at the corresponding time to perform online correction on the feature mapping weights of the prediction model. The corrected feature mapping weights are then substituted into the prediction model to generate the final prediction value.

[0014] As a further technical solution of the present invention, the data acquisition module is a multi-interface acquisition module, including a first acquisition interface, a second acquisition interface, and a third acquisition interface, which are respectively used to communicate with the distributed photovoltaic power station monitoring terminal, the satellite cloud image data platform, and the numerical weather forecast data platform, and a spare data acquisition interface is also reserved.

[0015] The unified spatiotemporal reference processing includes latitude and longitude spatial matching and power grid dispatch clock reference calibration. The latitude and longitude spatial matching is based on the latitude and longitude coordinates of each photovoltaic power station, with the spatial distance between the power station and surrounding meteorological grid points as the core basis. The closer the meteorological grid point is to the power station, the higher its data influence weight, thereby completing the spatial anchoring of satellite cloud image data, numerical weather forecast data and the location of a single power station. The power grid dispatch clock reference calibration is based on the millisecond-level Coordinated Universal Time (UTC) clock of the power grid dispatch system, correcting the original acquisition time deviation of different data sources, and unifying the acquisition timestamps of all data to the reference time of the current acquisition cycle.

[0016] The collected multi-source photovoltaic data includes measured data from photovoltaic power plants, satellite cloud image data, and numerical weather prediction data.

[0017] As a further technical solution of the present invention, the data preprocessing module includes a time synchronization unit, a scene adaptive noise cleaning unit, and an interpolation completion unit that are sequentially cascaded; the preprocessing includes time synchronization, scene adaptive noise cleaning, and missing value completion.

[0018] The timing synchronization refers to uniformly organizing all data to a time scale corresponding to the preset acquisition frequency.

[0019] The scene-adaptive noise cleaning method uses a sliding smoothing process to remove abnormal fluctuation noise in the data. The window size of the sliding smoothing process is bound to the weather scene. Specifically, the smallest window is used for the sunny steady-state scene, the medium window is used for the cloudy weak change scene and the cloudy fluctuation scene, and the largest window is used for the sudden transit scene.

[0020] The missing value completion refers to filling in a small number of discontinuous missing data based on the valid values ​​of the data at adjacent time points.

[0021] The weather scene segmentation is based on meteorological features extracted from multi-source raw datasets; the meteorological features include the discrete fluctuation of measured irradiance of all power stations in the target area and the average cloud cover.

[0022] The reliability level of the data sources decreases in the following order: on-site measured data from the power plant, satellite cloud image data, and numerical weather forecast data.

[0023] As a further technical solution of the present invention, the spatial association weight update between power stations is based on cloud cluster movement features extracted from satellite cloud image data. The cloud cluster movement features include cloud cluster movement orientation and cloud cluster movement speed. The graph convolution unit dynamically adjusts the spatial association weight between power stations based on the physical spatial adjacency relationship between power stations and the cloud cluster movement orientation and cloud cluster movement speed, so that power stations located upstream of the cloud cluster movement path are given higher association weights, thereby forming a dynamically updated power station spatial association system. For the standardized time series datasets of three types of photovoltaic data, the spatial association features corresponding to each type of photovoltaic data are mined by combining the dynamic spatial association system.

[0024] The temporal convolutional unit processes the temporal data of each type of photovoltaic data. First, it calculates the rate of change of the data over time, accurately identifies the key inflection points of sudden changes in irradiance and step changes in power, enhances the features corresponding to the key inflection points, and then extracts the temporal fluctuation features corresponding to each type of photovoltaic data from the enhanced temporal data.

[0025] The prediction output module has a built-in prediction model and a closed-loop correction unit. The closed-loop correction unit interacts with the measured power output terminal of the data acquisition module to extract the measured power generation of each power station in the previous complete acquisition cycle and the initial predicted power at the corresponding time. It calculates the deviation between the measured value and the predicted value of each power station and summarizes them to obtain the average prediction deviation of the entire target area. The preset threshold forms a fixed proportional relationship with the average measured power of the area in the previous cycle. If the average prediction deviation of the area does not exceed the preset threshold, the feature mapping weight of the prediction model remains unchanged. If the average prediction deviation of the area exceeds the preset threshold, the gradient descent algorithm is used to correct the feature mapping weight of the prediction model online.

[0026] The final prediction value is the predicted power generation value of each power station in the target area for the next 0-4 hours at the preset acquisition frequency and corresponding resolution.

[0027] Secondly, a method is provided to improve the accuracy of ultra-short-term forecasts of distributed photovoltaic power generation, including the following steps:

[0028] S1. Multi-source data spatiotemporal matching acquisition: The data acquisition module synchronously collects measured data, satellite cloud image data, and numerical weather forecast data from multiple distributed photovoltaic power stations in the region. The collected data is then spatially matched with latitude and longitude and calibrated with the grid dispatch clock reference time to unify the spatiotemporal reference of the data and form a multi-source raw dataset.

[0029] S2. Scene Adaptive Data Preprocessing: The data preprocessing module receives the multi-source raw dataset, divides the data into different weather scenarios based on meteorological characteristics, preprocesses the corresponding data for different weather scenarios, outputs a standardized time-series dataset, and assigns a credibility weight based on the data source to each group of photovoltaic data.

[0030] S3. Dynamic weighted spatiotemporal feature fusion: The spatiotemporal feature fusion module receives the standardized time-series dataset, dynamically updates the spatial correlation weights between photovoltaic power stations in the region based on the cloud movement features retrieved from satellite cloud images, and then mines the spatial correlation features between power stations in the region corresponding to different photovoltaic data by combining the spatial correlation weights. At the same time, it captures the temporal fluctuation features of photovoltaic power and environmental parameters corresponding to different photovoltaic data, and performs weighted splicing and fusion of the dynamic spatial correlation features and temporal fluctuation features corresponding to different photovoltaic data based on the credibility weights to output a comprehensive feature vector.

[0031] S4. Closed-loop correction power prediction output: The prediction output module receives the comprehensive feature vector, generates an initial prediction value through the built-in prediction model, and then, based on the deviation between the measured power of the power station in the previous acquisition cycle and the corresponding initial prediction value, when the deviation exceeds a preset threshold, the gradient descent algorithm is used to correct the feature mapping weights of the prediction model online, and outputs the final prediction value of the ultra-short-term photovoltaic power generation of each power station in the region for the next 0-4 hours.

[0032] As a further technical solution of the present invention, S1 specifically includes the following:

[0033] S1.1 Multi-source data synchronous acquisition: The data acquisition module synchronously acquires three types of core data according to the preset acquisition frequency: on-site measured data of all distributed photovoltaic power stations in the target area, including real-time power generation of the power stations, on-site irradiance, and ambient temperature; satellite cloud image data of the target area, including cloud coverage, cloud movement direction, and cloud movement speed; and numerical weather forecast data of the target area, including short-term irradiance forecast, temperature, wind speed, and air pressure fusion data.

[0034] S1.2 Latitude and Longitude Spatial Matching: For the latitude and longitude coordinates of each photovoltaic power station, the spatial distance between the power station and the surrounding meteorological grid points is used as the core basis. The closer the meteorological grid point is to the power station, the higher the weight of its corresponding data on the power station. In this way, the precise spatial anchoring of satellite cloud image data, numerical weather forecast data and the location of a single power station is completed, so that each power station is matched with the exclusive meteorological data of the corresponding location, and the spatial benchmark of all data is unified.

[0035] S1.3, Power Grid Clock Reference Time Calibration: Using the UTC millisecond-level clock of the power grid dispatching system as the time reference, correct the original acquisition time deviation of different data sources, and uniformly align the acquisition timestamps of all data to the reference time of the current acquisition cycle, which is completely matched with the preset acquisition step size, and unify the time reference of all data.

[0036] S1.4 Output of multi-source raw dataset: After unifying the spatial and temporal benchmarks, a multi-source raw dataset with completely consistent spatiotemporal benchmarks is formed.

[0037] As a further technical solution of the present invention, S2 specifically includes the following:

[0038] S2.1 Extraction of core meteorological features: From the multi-source raw dataset output by S1, two core meteorological features are extracted: the degree of discrete fluctuation of the measured irradiance of all power stations in the target area, which reflects the degree of drastic change in irradiance in the area; and the average cloud cover in the target area, which reflects the overall cloud cover in the area.

[0039] S2.2 Fixed Classification of Weather Scenarios: Based on the two extracted core meteorological features, the dataset is divided into four types of weather scenarios according to pre-set fixed judgment rules: those with extremely low irradiance fluctuation and extremely low cloud cover are classified as sunny steady-state scenarios; those with low irradiance fluctuation and moderate cloud cover are classified as cloudy weakly changing scenarios; those with moderate irradiance fluctuation and moderate cloud cover are classified as cloudy fluctuating scenarios; and those with extremely high irradiance fluctuation or extremely high cloud cover are classified as sudden transit scenarios.

[0040] S2.3 Scene Adaptive Preprocessing Operations: For the segmented weather scenes, perform scene-matched preprocessing operations, including time synchronization, scene adaptive noise removal, and missing value completion, specifically:

[0041] Time synchronization: Unify and standardize all data to a time scale corresponding to the preset acquisition frequency, and fully align with the fixed acquisition frequency and prediction resolution;

[0042] Scene-adaptive noise cleaning: A sliding smoothing method is used to remove abnormal fluctuation noise in the data. The window size of the smoothing process is bound to the weather scene. Specifically, the smallest window is used for sunny steady-state scenes, the medium window is used for cloudy weak-change and multi-cloud fluctuation scenes, and the largest window is used for sudden transition scenes to avoid over-smoothing that loses key features or under-smoothing that leaves residual noise.

[0043] Missing value completion: For a small number of discontinuous missing data, the missing data is completed based on the valid values ​​of the adjacent time points before and after the missing data to ensure the integrity and continuity of the time series data;

[0044] S2.4 Standardized Time Series Data Output and Credibility Weight Assignment: After preprocessing the three types of photovoltaic data, a standardized time series dataset is output; and a fixed credibility weight is assigned to each type of photovoltaic data according to the reliability level of the data source, with the highest weight for on-site measured data from power plants, followed by satellite cloud image data, and the lowest weight for numerical weather prediction data. The weights of all photovoltaic data form a complete fixed ratio system.

[0045] As a further technical solution of the present invention, S3 specifically includes the following:

[0046] S3.1 Dynamic Spatial Association Weight Update and Spatial Feature Mining: Based on the standardized time-series dataset of S2, the spatial association weights between power stations are dynamically adjusted using the cloud movement direction and speed in satellite cloud imagery data as the core basis and the physical spatial adjacency between power stations as the foundation. Power stations located upstream of the cloud movement path are assigned higher association weights, thus forming a dynamically updated spatial association system for power stations. For the standardized time-series datasets of the three types of photovoltaic data, the spatial association features corresponding to each type of photovoltaic data are mined by combining the dynamic spatial association system with the graph convolution unit built into the input module.

[0047] S3.2, Time-series fluctuation feature capture: For the standardized time-series dataset in S2, the time-series data of each type of photovoltaic data is processed separately: First, the rate of change of the data over time is calculated to accurately identify the key inflection points of sudden changes in irradiance and step changes in power. The features corresponding to these inflection points are enhanced to strengthen the identification of key fluctuation information. Then, the enhanced time-series data is input into the built-in time-series convolution unit of the module to extract the time-series fluctuation features corresponding to each type of photovoltaic data.

[0048] S3.3 Weighted Feature Fusion and Comprehensive Feature Vector Output: According to the exclusive credibility weight bound to each type of photovoltaic data, the spatial correlation features and temporal fluctuation features corresponding to the data source are weighted respectively; then the weighted spatial features and temporal features of all data sources are spliced ​​and fused to form a comprehensive feature vector that simultaneously contains the credibility differences of different data sources, spatial correlation information between power plants, and power temporal fluctuation information.

[0049] As a further technical solution of the present invention, S4 specifically includes the following:

[0050] S4.1 Initial Prediction Value Generation: Input the comprehensive feature vector output from S3 into the prediction model built into the module. Through the feature mapping and fitting rules built into the model, generate the initial prediction value of the power generation of each photovoltaic power station in the target area for the next 0-4 hours at the resolution corresponding to the preset acquisition frequency.

[0051] S4.2 Calculation of prediction deviation in the previous acquisition cycle: Extract the measured power generation of each power station and the initial predicted power at the corresponding time in the previous complete acquisition cycle, calculate the degree of deviation between the measured value and the predicted value of each power station, and then summarize to obtain the average prediction deviation of the entire target area.

[0052] S4.3 Online Weight Correction: A fixed deviation trigger threshold is preset, which forms a fixed proportional relationship with the average measured power of the region in the previous period; if the average prediction deviation of the region does not exceed the threshold, the feature mapping weights of the prediction model remain unchanged; if the average prediction deviation of the region exceeds the threshold, the gradient descent algorithm is used to correct the feature mapping weights of the prediction model online.

[0053] S4.4 Final Prediction Output and Closed-Loop Link Completion: Substitute the corrected feature mapping weights into the prediction model to regenerate the final prediction value of ultra-short-term photovoltaic power generation for each power station in the target area for the next 0-4 hours; the final prediction value is synchronously transmitted to the grid dispatch terminal and the photovoltaic power station monitoring platform, and the measured power and the final prediction value of this cycle are fed back to the deviation calculation link of the next acquisition cycle to form a complete closed-loop correction link.

[0054] As a further technical solution of the present invention, the initial prediction value generation is to generate the comprehensive feature vector output by S3. Input the built-in prediction model, and map it through the model's generator function. Generate the initial predicted power of each power station in the region for the next 0-4 hours:

[0055]

[0056] in: The feature mapping weight matrix of the prediction model; For the first The future of the power station Initial predicted power generation at time [time] Predict the step size for the future.

[0057] Compared with existing technologies, the beneficial effects of this device and method for improving the accuracy of ultra-short-term forecasts of distributed photovoltaic power generation are:

[0058] By performing spatial matching and temporal benchmark calibration on multi-source collected data, the spatiotemporal benchmarks of data from different sources are unified, solving the problems of spatial misalignment of data sources and insufficient temporal synchronization caused by the dispersed layout of distributed photovoltaic sites. This enables the basic data used for forecasting to accurately match the actual geographical location and operating sequence of each photovoltaic site, and to conform to the actual meteorological conditions and power generation conditions at the site. At the same time, corresponding data preprocessing methods are matched based on different meteorological operating scenarios, avoiding the adaptability defects of fixed processing modes under different meteorological conditions. Combined with the credibility weight assignment according to the reliability classification of data sources, the integrity, rationality and accuracy of credibility classification of the basic forecasting data are guaranteed.

[0059] By fully exploring the spatial correlation characteristics between photovoltaic sites in the region and the temporal fluctuation characteristics of photovoltaic power changes, and adjusting the spatial correlation weights between sites in combination with the actual movement patterns of cloud clusters, the core correlation factors and key fluctuation characteristics affecting photovoltaic power changes can be accurately captured. In addition, by extracting features independently from different data sources and weighting and fusing them according to their corresponding credibility weights, the effective information loss caused by single-dimensional feature extraction and indiscriminate mixing of multi-source data is avoided. The extracted fused features can comprehensively and accurately reflect the actual trend of photovoltaic power changes, providing reliable feature support for subsequent power prediction.

[0060] By achieving dynamic correction of the prediction model through closed-loop feedback of prediction deviation, the prediction output effect can be continuously optimized, alleviating the problem of prediction deviation accumulating and amplifying over prediction time, and ensuring the stability and reliability of prediction results under different operating conditions. At the same time, the implementation of the entire solution is based entirely on the existing infrastructure of the device, without requiring any changes to the hardware composition, module connection relationship and core function positioning of the device, making it easy to be directly applied in the existing distributed photovoltaic power prediction system.

[0061] The output prediction results can be adapted to the needs of grid dispatch and photovoltaic power plant operation and maintenance. It can provide accurate reference for the power regulation and grid connection of distributed photovoltaic power, which helps to improve the utilization efficiency of distributed photovoltaic power, and at the same time provides effective technical support for the power balance dispatch and safe and stable operation of regional power grid. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the device architecture of the present invention;

[0063] Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Please see Figure 1 The present invention provides an embodiment of an apparatus for improving the accuracy of ultra-short-term prediction of distributed photovoltaic power generation, comprising a data acquisition module, a data preprocessing module, a spatiotemporal feature fusion module, a prediction output module, and a power supply module;

[0066] The data acquisition module is used to collect multi-source photovoltaic data from multiple distributed photovoltaic power stations within the target area, and to perform spatiotemporal benchmark unified processing on the data to output multi-source raw datasets.

[0067] The data preprocessing module receives multi-source raw datasets, preprocesses the corresponding photovoltaic data for different weather scenarios to obtain standardized time-series datasets, and assigns corresponding credibility weights to different photovoltaic data in the standardized time-series datasets according to the reliability level of the data sources.

[0068] The spatiotemporal feature fusion module receives a standardized time-series dataset. It incorporates graph convolutional units and temporal convolutional units. The graph convolutional unit dynamically updates the spatial correlation weights between photovoltaic power stations based on cloud movement features extracted from satellite cloud imagery. Using these updated weights, it extracts spatial correlation features corresponding to each type of photovoltaic data from the standardized time-series dataset. The temporal convolutional unit extracts temporal fluctuation features corresponding to each type of photovoltaic data from the standardized time-series dataset. The spatiotemporal feature fusion module weights the spatial correlation features and temporal fluctuation features of different photovoltaic data according to their reliability weights, then fuses the weighted features to output a comprehensive feature vector.

[0069] The prediction output module receives the comprehensive feature vector, generates the initial prediction value for the current period through the built-in prediction model, and uses the deviation between the measured power of the previous acquisition period and the initial prediction value at the corresponding time to perform online correction on the feature mapping weights of the prediction model. The corrected feature mapping weights are then substituted into the prediction model to generate the final prediction value.

[0070] The data acquisition module is a multi-interface acquisition module, including a first acquisition interface, a second acquisition interface, and a third acquisition interface, which are used to communicate with the distributed photovoltaic power station monitoring terminal, the satellite cloud image data platform, and the numerical weather prediction data platform, respectively. A spare data acquisition interface is also reserved.

[0071] The unified spatiotemporal reference processing includes latitude and longitude spatial matching and power grid dispatch clock reference calibration. Latitude and longitude spatial matching is based on the latitude and longitude coordinates of each photovoltaic power station, with the spatial distance between the power station and surrounding meteorological grid points as the core basis. The closer the meteorological grid point is to the power station, the higher its data influence weight. This completes the spatial anchoring of satellite cloud image data, numerical weather prediction data and the location of individual power stations. Power grid dispatch clock reference calibration is based on the millisecond-level Coordinated Universal Time (UTC) clock of the power grid dispatch system. It corrects the original acquisition time deviation of different data sources and aligns the acquisition timestamps of all data to the reference time of the current acquisition cycle.

[0072] The collected multi-source photovoltaic data includes measured data from photovoltaic power plants, satellite cloud image data, and numerical weather prediction data.

[0073] The data preprocessing module includes a timing synchronization unit, a scene adaptive noise cleaning unit, and an interpolation completion unit that are cascaded in series; preprocessing includes timing synchronization, scene adaptive noise cleaning, and missing value completion.

[0074] Time synchronization involves uniformly organizing all data to a time scale corresponding to the preset acquisition frequency.

[0075] Scene-adaptive noise cleaning uses a sliding smoothing process to remove abnormal fluctuation noise from the data. The window size of the sliding smoothing process is bound to the weather scene. The smallest window is used for sunny steady-state scenes, the medium window is used for cloudy weak-change scenes and cloudy fluctuating scenes, and the largest window is used for sudden transit scenes.

[0076] Missing value completion is used to fill in a small number of discontinuous missing data based on the valid values ​​of the data before and after the data.

[0077] The weather scene segmentation is based on meteorological features extracted from multi-source raw datasets; the meteorological features include the discrete fluctuation of measured irradiance of all power stations in the target area and the average cloud cover.

[0078] The reliability level of data sources decreases in the following order: on-site measured data from power plants, satellite cloud image data, and numerical weather prediction data.

[0079] The spatial association weight update between power plants is based on cloud movement features extracted from satellite cloud image data. The cloud movement features include cloud movement orientation and cloud movement speed. The graph convolutional unit dynamically adjusts the spatial association weight between power plants based on the physical spatial adjacency relationship between them, combined with the cloud movement orientation and cloud movement speed. This assigns higher association weights to power plants located upstream of the cloud movement path, thereby forming a dynamically updated spatial association system for power plants. For the standardized time-series datasets of the three types of photovoltaic data, the spatial association features corresponding to each type of photovoltaic data are mined by combining the dynamic spatial association system.

[0080] The temporal convolutional unit processes the temporal data of each type of photovoltaic data. First, it calculates the rate of change of the data over time, accurately identifies the key inflection points of sudden changes in irradiance and step changes in power, enhances the features corresponding to the key inflection points, and then extracts the temporal fluctuation features corresponding to each type of photovoltaic data from the enhanced temporal data.

[0081] The prediction output module has a built-in prediction model and a closed-loop correction unit. The closed-loop correction unit interacts with the measured power output of the data acquisition module to extract the measured power generation of each power station in the previous complete acquisition cycle and the initial predicted power at the corresponding time. It calculates the deviation between the measured value and the predicted value of each power station and summarizes the average prediction deviation of the entire target area. The preset threshold forms a fixed proportional relationship with the average measured power of the area in the previous cycle. If the average prediction deviation of the area does not exceed the preset threshold, the feature mapping weight of the prediction model remains unchanged. If the average prediction deviation of the area exceeds the preset threshold, the gradient descent algorithm is used to correct the feature mapping weight of the prediction model online.

[0082] The final prediction value is the predicted power generation value of each power station in the target area for the next 0-4 hours at the preset acquisition frequency and corresponding resolution.

[0083] Please see Figure 2 The present invention provides an embodiment of a method for improving the accuracy of ultra-short-term forecasts of distributed photovoltaic power generation, comprising the following steps:

[0084] S1. Multi-source data spatiotemporal matching acquisition: The data acquisition module synchronously collects measured data, satellite cloud image data, and numerical weather forecast data from multiple distributed photovoltaic power stations in the region. The collected data is then spatially matched with latitude and longitude and calibrated with the grid dispatch clock reference time to unify the spatiotemporal reference of the data and form a multi-source raw dataset.

[0085] S2, Scene Adaptive Data Preprocessing: The data preprocessing module receives multi-source raw datasets, divides the data into different weather scenarios based on meteorological characteristics, preprocesses the corresponding data for different weather scenarios, outputs standardized time-series datasets, and assigns a credibility weight based on the data source to each group of photovoltaic data.

[0086] S3. Dynamic weighted spatiotemporal feature fusion: The spatiotemporal feature fusion module receives a standardized time-series dataset, dynamically updates the spatial correlation weights between photovoltaic power stations in the region based on the cloud movement features retrieved from satellite cloud images, and then mines the spatial correlation features between power stations in the region corresponding to different photovoltaic data by combining the spatial correlation weights. At the same time, it captures the temporal fluctuation features of photovoltaic power and environmental parameters corresponding to different photovoltaic data, and performs weighted splicing and fusion of the dynamic spatial correlation features and temporal fluctuation features corresponding to different photovoltaic data based on the credibility weights to output a comprehensive feature vector.

[0087] S4. Closed-loop correction power prediction output: The prediction output module receives the comprehensive feature vector, generates the initial prediction value through the built-in prediction model, and then, based on the deviation between the measured power of the power station in the previous acquisition cycle and the corresponding initial prediction value, when the deviation exceeds the preset threshold, the gradient descent algorithm is used to correct the feature mapping weight of the prediction model online, and outputs the final prediction value of the ultra-short-term photovoltaic power generation of each power station in the region for the next 0-4 hours.

[0088] Furthermore, in another embodiment, S1 specifically includes the following:

[0089] S1.1 Multi-source data synchronous acquisition: The data acquisition module synchronously acquires three types of core data according to the preset acquisition frequency: on-site measured data of all distributed photovoltaic power stations in the target area, including real-time power generation of the power stations, on-site irradiance, and ambient temperature; satellite cloud image data of the target area, including cloud coverage, cloud movement direction, and cloud movement speed; and numerical weather forecast data of the target area, including short-term irradiance forecast, temperature, wind speed, and air pressure fusion data.

[0090] S1.2 Latitude and Longitude Spatial Matching: For the latitude and longitude coordinates of each photovoltaic power station, the spatial distance between the power station and the surrounding meteorological grid points is used as the core basis. The closer the meteorological grid point is to the power station, the higher the weight of its corresponding data on the power station. In this way, the precise spatial anchoring of satellite cloud image data, numerical weather forecast data and the location of a single power station is completed, so that each power station is matched with the exclusive meteorological data of the corresponding location, and the spatial benchmark of all data is unified.

[0091] S1.3, Power Grid Clock Reference Time Calibration: Using the UTC millisecond-level clock of the power grid dispatching system as the time reference, correct the original acquisition time deviation of different data sources, and uniformly align the acquisition timestamps of all data to the reference time of the current acquisition cycle, which is completely matched with the preset acquisition step size, and unify the time reference of all data.

[0092] S1.4 Output of multi-source raw dataset: After unifying the spatial and temporal benchmarks, a multi-source raw dataset with completely consistent spatiotemporal benchmarks is formed.

[0093] Furthermore, in another embodiment, S2 specifically includes the following:

[0094] S2.1 Extraction of core meteorological features: From the multi-source raw dataset output by S1, two core meteorological features are extracted: the degree of discrete fluctuation of the measured irradiance of all power stations in the target area, which reflects the degree of drastic change in irradiance in the area; and the average cloud cover in the target area, which reflects the overall cloud cover in the area.

[0095] S2.2 Fixed Classification of Weather Scenarios: Based on the two extracted core meteorological features, the dataset is divided into four types of weather scenarios according to pre-set fixed judgment rules: those with extremely low irradiance fluctuation and extremely low cloud cover are classified as sunny steady-state scenarios; those with low irradiance fluctuation and moderate cloud cover are classified as cloudy weakly changing scenarios; those with moderate irradiance fluctuation and moderate cloud cover are classified as cloudy fluctuating scenarios; and those with extremely high irradiance fluctuation or extremely high cloud cover are classified as sudden transit scenarios.

[0096] S2.3 Scene Adaptive Preprocessing Operations: For the segmented weather scenes, perform scene-matched preprocessing operations, including time synchronization, scene adaptive noise removal, and missing value completion, specifically:

[0097] Time synchronization: Unify and standardize all data to a time scale corresponding to the preset acquisition frequency, and fully align with the fixed acquisition frequency and prediction resolution;

[0098] Scene-adaptive noise cleaning: A sliding smoothing method is used to remove abnormal fluctuation noise in the data. The window size of the smoothing process is bound to the weather scene. Specifically, the smallest window is used for sunny steady-state scenes, the medium window is used for cloudy weak-change and multi-cloud fluctuation scenes, and the largest window is used for sudden transition scenes to avoid over-smoothing that loses key features or under-smoothing that leaves residual noise.

[0099] Missing value completion: For a small number of discontinuous missing data, the missing data is completed based on the valid values ​​of the adjacent time points before and after the missing data to ensure the integrity and continuity of the time series data;

[0100] S2.4 Standardized Time Series Data Output and Credibility Weight Assignment: After preprocessing the three types of photovoltaic data, a standardized time series dataset is output; and a fixed credibility weight is assigned to each type of photovoltaic data according to the reliability level of the data source, with the highest weight for on-site measured data from power plants, followed by satellite cloud image data, and the lowest weight for numerical weather prediction data. The weights of all photovoltaic data form a complete fixed ratio system.

[0101] Furthermore, in another embodiment, S3 specifically includes the following:

[0102] S3.1 Dynamic Spatial Association Weight Update and Spatial Feature Mining: Based on the standardized time-series dataset of S2, the spatial association weights between power stations are dynamically adjusted using the cloud movement direction and speed in satellite cloud imagery data as the core basis and the physical spatial adjacency between power stations as the foundation. Power stations located upstream of the cloud movement path are assigned higher association weights, thus forming a dynamically updated spatial association system for power stations. For the standardized time-series datasets of the three types of photovoltaic data, the spatial association features corresponding to each type of photovoltaic data are mined by combining the dynamic spatial association system with the graph convolution unit built into the input module.

[0103] S3.2, Time-series fluctuation feature capture: For the standardized time-series dataset in S2, the time-series data of each type of photovoltaic data is processed separately: First, the rate of change of the data over time is calculated to accurately identify the key inflection points of sudden changes in irradiance and step changes in power. The features corresponding to these inflection points are enhanced to strengthen the identification of key fluctuation information. Then, the enhanced time-series data is input into the built-in time-series convolution unit of the module to extract the time-series fluctuation features corresponding to each type of photovoltaic data.

[0104] S3.3 Weighted Feature Fusion and Comprehensive Feature Vector Output: According to the exclusive credibility weight bound to each type of photovoltaic data, the spatial correlation features and temporal fluctuation features corresponding to the data source are weighted respectively; then the weighted spatial features and temporal features of all data sources are spliced ​​and fused to form a comprehensive feature vector that simultaneously contains the credibility differences of different data sources, spatial correlation information between power plants, and power temporal fluctuation information.

[0105] Furthermore, in another embodiment, S4 specifically includes the following:

[0106] S4.1 Initial Prediction Value Generation: Input the comprehensive feature vector output from S3 into the prediction model built into the module. Through the feature mapping and fitting rules built into the model, generate the initial prediction value of the power generation of each photovoltaic power station in the target area for the next 0-4 hours at the resolution corresponding to the preset acquisition frequency.

[0107] S4.2 Calculation of prediction deviation in the previous acquisition cycle: Extract the measured power generation of each power station and the initial predicted power at the corresponding time in the previous complete acquisition cycle, calculate the degree of deviation between the measured value and the predicted value of each power station, and then summarize to obtain the average prediction deviation of the entire target area.

[0108] S4.3 Online Weight Correction: A fixed deviation trigger threshold is preset, which forms a fixed proportional relationship with the average measured power of the region in the previous period; if the average prediction deviation of the region does not exceed the threshold, the feature mapping weights of the prediction model remain unchanged; if the average prediction deviation of the region exceeds the threshold, the gradient descent algorithm is used to correct the feature mapping weights of the prediction model online.

[0109] S4.4 Final Prediction Output and Closed-Loop Link Completion: Substitute the corrected feature mapping weights into the prediction model to regenerate the final prediction value of ultra-short-term photovoltaic power generation for each power station in the target area for the next 0-4 hours; the final prediction value is synchronously transmitted to the grid dispatch terminal and the photovoltaic power station monitoring platform, and the measured power and the final prediction value of this cycle are fed back to the deviation calculation link of the next acquisition cycle to form a complete closed-loop correction link.

[0110] Application example:

[0111] Step 1: Multi-source data spatiotemporal matching acquisition:

[0112] S1.1, Synchronous Acquisition of Multi-Source Data:

[0113] The data acquisition module collects three types of core data synchronously at a fixed frequency. The collected data includes the following:

[0114] 1. Within the area Measured data from distributed photovoltaic power stations: real-time power generation of each power station On-site irradiance Ambient temperature ,in This refers to the serial number of the photovoltaic power station. , This serves as the reference time for the current acquisition cycle, corresponding to a fixed acquisition step size.

[0115] 2. Satellite cloud imagery data for the target area: cloud coverage, cloud movement azimuth. Cloud movement speed ;

[0116] 3. Numerical weather forecast data for the target area: fused data of short-term irradiance forecast, temperature, wind speed, and air pressure. ;

[0117] S1.2, Latitude and Longitude Spatial Matching:

[0118] For the latitude and longitude coordinates of each photovoltaic power station An inverse distance weighted interpolation algorithm is used to achieve spatial anchoring of satellite cloud images, numerical weather predictions, and power station locations, unifying the spatial reference. The spatial matching formula is as follows:

[0119]

[0120]

[0121] in: For the first The power station and the first The spherical distance of a meteorological grid point; The number of fixed, effective meteorological grid points around the power station is set according to the meteorological grid layout in the area where the power station is located. is the distance attenuation coefficient, a fixed constant used to adjust the weight of the distance's influence on the interpolation result; For the first Meteorological grid Raw satellite cloud image data at any given time; For the first Meteorological grid Raw data for numerical weather forecasts at any given time; For spatial matching of power stations Cloud coverage data for the corresponding location; For spatial matching of power stations Meteorological forecast fusion data for the corresponding location;

[0122] S1.3, Power Grid Clock Reference Time Calibration:

[0123] Using the UTC millisecond-level clock of the power grid dispatching system as a reference, all collected data are timestamped to eliminate time synchronization deviations from different data sources and unify the time reference. The calibration formula is as follows:

[0124]

[0125] in: For the first The original collection timestamp of the data source; This refers to the time deviation between the data source and the power grid reference clock; To achieve a calibrated unified timestamp, the timestamps of all data are ultimately unified to [the desired value]. The timing should be aligned with the data acquisition step size;

[0126] S1.4, Output of multi-source raw datasets:

[0127] After completing spatial matching and temporal calibration, a multi-source original dataset with unified spatiotemporal reference is formed. This serves as input data for subsequent steps;

[0128] Step 2: Scene Adaptive Data Preprocessing

[0129] S2.1 Extraction of core meteorological features:

[0130] The multi-source raw dataset output from step S1 In the process, the regional irradiance fluctuation rate is extracted. and regional average cloud cover The two meteorological characteristics are calculated using the following formula:

[0131]

[0132]

[0133] in: for The average measured irradiance of all power stations within the time period is from [data source missing]. ; For power station Measured irradiance; This represents the total number of photovoltaic power stations in the region. Cloud coverage data after matching a single power station;

[0134] S2.2 Fixed division of weather scenes:

[0135] Based on the two extracted meteorological features, a fixed threshold method is used to divide the dataset into four weather scenarios, with the following determination rules:

[0136] 1. Steady-state scenario under clear skies: and

[0137] 2. Cloudy weather with slight changes: and

[0138] 3. Cloudy and fluctuating scenario: and

[0139] 4. Sudden Transition Scenario: or

[0140] in: Thresholds are set for scene segmentation based on the operational characteristics of regional photovoltaic power plants;

[0141] S2.3 Scene Adaptive Preprocessing Operation:

[0142] For the segmented weather scene, the following preprocessing operations are performed sequentially: time-series synchronization, scene-adaptive sliding filtering, and missing value completion:

[0143] 1. Time synchronization: Interpolate all data to a fixed acquisition step size time scale, aligning with the acquisition frequency and prediction resolution;

[0144] 2. Scene-Adaptive Noise Removal: A moving average algorithm is used, with the sliding window length adaptively adjusted according to the weather scene. The filtering formula is as follows:

[0145]

[0146] in: After timing synchronization, the power station exist The original single-source data at any given moment; The length of the sliding window is adaptively selected based on the weather scenario. For the power station after filtering and noise reduction exist Single-source data at any given moment; The time sequence traversal within the sliding window represents the time step covered by the window.

[0147] 3. Missing value imputation: For missing data segments, linear interpolation is used for imputation. The imputation formula is as follows:

[0148]

[0149] in: , These are the valid measured data for the time intervals immediately before and after the missing time interval. This is the complete data after completion;

[0150] S2.4, Standardized Time Series Dataset Output and Confidence Weight Assignment:

[0151] The preprocessed output is a standardized time series dataset. ;

[0152] Based on the reliability of the data source, each group of photovoltaic data is assigned a corresponding credibility weight. The weights satisfy The measured data has the highest weight, followed by satellite cloud image data, and then numerical forecast data has the lowest weight.

[0153] Step 3: Dynamic weighted spatiotemporal feature fusion:

[0154] S3.1 Dynamic Spatial Association Weight Update and Spatial Feature Mining:

[0155] Based on the cloud movement features collected in step S1 and preprocessed in step S2 ( , The spatial correlation weights between power stations are dynamically updated, with the input data being a standardized dataset. The updated formula is:

[0156]

[0157] in: For fixed power station spatial adjacency matrix elements, These are the serial numbers of the corresponding photovoltaic power stations. Adjacent power stations are assigned a value of 1, and non-adjacent power stations are assigned a value of 0. These are the dynamically updated adjacency matrix elements; This is a weighting adjustment coefficient used to control the degree of influence of cloud movement on spatial correlation weights. The upper limit is used in open plains areas, and the lower limit is used in severely obstructed mountainous areas. Azimuth angle of cloud movement With power station to the power station The azimuth angle; The maximum historical movement speed of cloud clusters within the region was determined by extracting 15-minute satellite cloud imagery data from the target area over the past 12 months, calculating the linear movement speed of all valid cloud clusters, and taking the 95th percentile of all statistical values ​​as the maximum velocity. ; Updated once per calendar month, incorporating the latest satellite cloud imagery data from the previous month; The linear velocity of the cloud cluster at the current moment;

[0158] The updated dynamic adjacency matrix With standardized time series datasets By inputting the data into graph convolutional units, spatial correlation feature vectors corresponding to different photovoltaic data are mined. ;

[0159] S3.2, Capturing Temporal Fluctuation Characteristics:

[0160] For the standardized time series dataset output in step S2 The gradient-enhanced temporal convolution algorithm is used to capture the temporal fluctuation characteristics of abrupt changes in irradiance and step changes in power. The specific steps are as follows:

[0161] 1. Calculate the first-order gradient value of time series data to locate the inflection point of data fluctuation. The calculation formula is as follows:

[0162]

[0163] in: To fix the data acquisition step size; for The gradient value of the standardized data at any given time is used to characterize the rate of change of the data; , Standardized single-source data for adjacent time points;

[0164] 2. Set the gradient value Concatenated with the corresponding standardized time-series dataset, and input into a time-series convolutional unit, the time-series fluctuation feature vectors corresponding to different photovoltaic data are extracted. ;

[0165] S3.3 Weighted Feature Fusion and Output of Comprehensive Feature Vector:

[0166] Based on the single-source data credibility weight assigned in step S2 Spatial feature vectors corresponding to different photovoltaic data and time series feature vectors The features are weighted separately, and then the weighted features are concatenated and fused. The fusion formula is as follows:

[0167]

[0168] in: For feature concatenation operations, single-source spatial features and temporal features are concatenated according to dimensions; For the first Specific credibility weights for data sources; the final output is a comprehensive feature vector. This serves as input data for subsequent steps;

[0169] Step 4: Closed-loop correction power prediction output:

[0170] S4.1 Initial Prediction Value Generation:

[0171] The comprehensive feature vector output in step S3 Input the built-in prediction model, and map it through the model's generator function. The initial predicted power of each power station in the region for the next 0-4 hours is generated using the following formula:

[0172]

[0173] in: The feature mapping weight matrix of the prediction model; For the first The future of the power station Initial predicted power generation at time [time] The future prediction step size corresponds to a prediction duration of 0-4 hours.

[0174] S4.2 Calculation of prediction deviation in the previous acquisition cycle:

[0175] Extract the measured power from the previous acquisition cycle and the corresponding initial predicted value, then calculate the prediction deviation using the following formula:

[0176]

[0177]

[0178] in: For the previous cycle power station The measured power; This represents the initial predicted power at the corresponding moment in the previous cycle; For single power station prediction absolute deviation; The average prediction bias for the region; This represents the total number of photovoltaic power stations in the region.

[0179] S4.3 Online Weight Correction:

[0180] The feature mapping weights of the model are corrected online in a lightweight manner based on the prediction bias, and the correction rules are as follows:

[0181] 1. Preset deviation threshold The threshold is determined based on the average measured power of the region in the previous period;

[0182] 2. If No correction is triggered; feature mapping weights are maintained. constant;

[0183] 3. If This triggers weight correction, which updates the weights using the gradient descent algorithm. The update formula is as follows: in: The model learning rate; This is the corrected feature map weight matrix; the correction only fine-tunes the weight values.

[0184] S4.4 Final Predicted Value Output and Closed-Loop Link Completion:

[0185] Corrected weights Substitute the values ​​into the prediction model to regenerate the final ultra-short-term photovoltaic power generation forecasts for each power station in the region for the next 0-4 hours at a fixed resolution. The formula is:

[0186]

[0187] The final predicted value is synchronized to the power grid dispatch terminal and the photovoltaic power station monitoring platform through the device's signal output interface. At the same time, the predicted data and measured power are fed back to the deviation calculation stage of the next acquisition cycle, forming a complete closed-loop correction link to ensure a continuous closed-loop data flow.

[0188] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A device for improving the accuracy of ultra-short-term forecasts of distributed photovoltaic power generation, characterized in that, It includes a data acquisition module, a data preprocessing module, a spatiotemporal feature fusion module, and a prediction output module; The data acquisition module is used to collect multi-source photovoltaic data from multiple distributed photovoltaic power stations within the target area, and to perform spatiotemporal benchmark unified processing on the data to output multi-source raw datasets. The data preprocessing module receives multi-source raw datasets, preprocesses the corresponding photovoltaic data for different weather scenarios to obtain standardized time-series datasets, and assigns corresponding credibility weights to different photovoltaic data in the standardized time-series datasets according to the reliability level of the data sources. The spatiotemporal feature fusion module receives a standardized time-series dataset. It incorporates a graph convolutional unit and a temporal convolutional unit. The graph convolutional unit dynamically updates the spatial correlation weights between photovoltaic power stations based on cloud movement features extracted from satellite cloud imagery. Using these updated spatial correlation weights, it extracts spatial correlation features corresponding to each type of photovoltaic data from the standardized time-series dataset. The temporal convolutional unit extracts temporal fluctuation features corresponding to each type of photovoltaic data from the standardized time-series dataset. The spatiotemporal feature fusion module weights the spatial correlation features and temporal fluctuation features of different photovoltaic data according to the confidence weights, then fuses the weighted features to output a comprehensive feature vector. The prediction output module receives the comprehensive feature vector, generates the initial prediction value for the current period through the built-in prediction model, and uses the deviation between the measured power of the previous acquisition period and the initial prediction value at the corresponding time to perform online correction on the feature mapping weights of the prediction model. The corrected feature mapping weights are then substituted into the prediction model to generate the final prediction value.

2. The device for improving the accuracy of ultra-short-term forecasting of distributed photovoltaic power generation according to claim 1, characterized in that: The data acquisition module is a multi-interface acquisition module, including a first acquisition interface, a second acquisition interface, and a third acquisition interface, which are used to communicate with the distributed photovoltaic power station monitoring terminal, the satellite cloud image data platform, and the numerical weather prediction data platform, respectively. A spare data acquisition interface is also reserved. The unified spatiotemporal reference processing includes latitude and longitude spatial matching and power grid dispatch clock reference calibration. The latitude and longitude spatial matching is based on the latitude and longitude coordinates of each photovoltaic power station, with the spatial distance between the power station and surrounding meteorological grid points as the core basis. The closer the meteorological grid point is to the power station, the higher its data influence weight, thereby completing the spatial anchoring of satellite cloud image data, numerical weather forecast data and the location of a single power station. The power grid dispatch clock reference calibration is based on the millisecond-level Coordinated Universal Time (UTC) clock of the power grid dispatch system, correcting the original acquisition time deviation of different data sources, and unifying the acquisition timestamps of all data to the reference time of the current acquisition cycle. The collected multi-source photovoltaic data includes measured data from photovoltaic power plants, satellite cloud image data, and numerical weather prediction data.

3. The device for improving the accuracy of ultra-short-term forecasting of distributed photovoltaic power generation according to claim 1, characterized in that: The data preprocessing module includes a time synchronization unit, a scene adaptive noise cleaning unit, and an interpolation completion unit that are cascaded in series; the preprocessing includes time synchronization, scene adaptive noise cleaning, and missing value completion. The timing synchronization refers to uniformly organizing all data to a time scale corresponding to the preset acquisition frequency. The scene-adaptive noise cleaning method uses a sliding smoothing process to remove abnormal fluctuation noise in the data. The window size of the sliding smoothing process is bound to the weather scene. Specifically, the smallest window is used for the sunny steady-state scene, the medium window is used for the cloudy weak change scene and the cloudy fluctuation scene, and the largest window is used for the sudden transit scene. The missing value completion refers to filling in a small number of discontinuous missing data based on the valid values ​​of the data at adjacent time points. The weather scene segmentation is based on meteorological features extracted from multi-source raw datasets; the meteorological features include the discrete fluctuation of measured irradiance of all power stations in the target area and the average cloud cover. The reliability level of the data sources decreases in the following order: on-site measured data from the power plant, satellite cloud image data, and numerical weather forecast data.

4. The device for improving the accuracy of ultra-short-term forecasting of distributed photovoltaic power generation according to claim 1, characterized in that: The spatial association weight update between power stations is based on cloud movement features extracted from satellite cloud image data. The cloud movement features include cloud movement orientation and cloud movement speed. The graph convolutional unit dynamically adjusts the spatial association weight between power stations based on the physical spatial adjacency relationship between them, combined with the cloud movement orientation and cloud movement speed. This assigns higher association weights to power stations located upstream of the cloud movement path, thereby forming a dynamically updated power station spatial association system. For the standardized time-series datasets of the three types of photovoltaic data, the spatial association features corresponding to each type of photovoltaic data are mined by combining the dynamic spatial association system. The temporal convolutional unit processes the temporal data of each type of photovoltaic data. First, it calculates the rate of change of the data over time, accurately identifies the key inflection points of sudden changes in irradiance and step changes in power, enhances the features corresponding to the key inflection points, and then extracts the temporal fluctuation features corresponding to each type of photovoltaic data from the enhanced temporal data. The prediction output module has a built-in prediction model and a closed-loop correction unit. The closed-loop correction unit interacts with the measured power output terminal of the data acquisition module to extract the measured power generation of each power station in the previous complete acquisition cycle and the initial predicted power at the corresponding time. It calculates the deviation between the measured value and the predicted value of each power station and summarizes them to obtain the average prediction deviation of the entire target area. The preset threshold forms a fixed proportional relationship with the average measured power of the area in the previous cycle. If the average prediction deviation of the area does not exceed the preset threshold, the feature mapping weight of the prediction model remains unchanged. If the average prediction deviation of the area exceeds the preset threshold, the gradient descent algorithm is used to correct the feature mapping weight of the prediction model online. The final prediction value is the predicted power generation value of each power station in the target area for the next 0-4 hours at the preset acquisition frequency and corresponding resolution.

5. A method for improving the accuracy of ultra-short-term forecasts of distributed photovoltaic power generation, characterized in that: Includes the following steps: S1. Multi-source data spatiotemporal matching acquisition: The data acquisition module synchronously collects measured data, satellite cloud image data, and numerical weather forecast data from multiple distributed photovoltaic power stations in the region. The collected data is then spatially matched with latitude and longitude and calibrated with the grid dispatch clock reference time to unify the spatiotemporal reference of the data and form a multi-source raw dataset. S2. Scene Adaptive Data Preprocessing: The data preprocessing module receives the multi-source raw dataset, divides the data into different weather scenarios based on meteorological characteristics, preprocesses the corresponding data for different weather scenarios, outputs a standardized time-series dataset, and assigns a credibility weight based on the data source to each group of photovoltaic data. S3. Dynamic weighted spatiotemporal feature fusion: The spatiotemporal feature fusion module receives the standardized time-series dataset, dynamically updates the spatial correlation weights between photovoltaic power stations in the region based on the cloud movement features retrieved from satellite cloud images, and then mines the spatial correlation features between power stations in the region corresponding to different photovoltaic data by combining the spatial correlation weights. At the same time, it captures the temporal fluctuation features of photovoltaic power and environmental parameters corresponding to different photovoltaic data, and performs weighted splicing and fusion of the dynamic spatial correlation features and temporal fluctuation features corresponding to different photovoltaic data based on the credibility weights to output a comprehensive feature vector. S4. Closed-loop correction power prediction output: The prediction output module receives the comprehensive feature vector, generates an initial prediction value through the built-in prediction model, and then, based on the deviation between the measured power of the power station in the previous acquisition cycle and the corresponding initial prediction value, when the deviation exceeds a preset threshold, the gradient descent algorithm is used to correct the feature mapping weights of the prediction model online, and outputs the final prediction value of the ultra-short-term photovoltaic power generation of each power station in the region for the next 0-4 hours.

6. The method for improving the accuracy of ultra-short-term forecasting of distributed photovoltaic power generation according to claim 5, characterized in that: S1 specifically includes the following: S1.1 Multi-source data synchronous acquisition: The data acquisition module synchronously acquires three types of core data according to the preset acquisition frequency: on-site measured data of all distributed photovoltaic power stations in the target area, including real-time power generation of the power stations, on-site irradiance, and ambient temperature; satellite cloud image data of the target area, including cloud coverage, cloud movement direction, and cloud movement speed; and numerical weather forecast data of the target area, including short-term irradiance forecast, temperature, wind speed, and air pressure fusion data. S1.2 Latitude and Longitude Spatial Matching: For the latitude and longitude coordinates of each photovoltaic power station, the spatial distance between the power station and the surrounding meteorological grid points is used as the core basis. The closer the meteorological grid point is to the power station, the higher the weight of its corresponding data on the power station. In this way, the precise spatial anchoring of satellite cloud image data, numerical weather forecast data and the location of a single power station is completed, so that each power station is matched with the exclusive meteorological data of the corresponding location, and the spatial benchmark of all data is unified. S1.3, Power Grid Clock Reference Time Calibration: Using the UTC millisecond-level clock of the power grid dispatching system as the time reference, correct the original acquisition time deviation of different data sources, and uniformly align the acquisition timestamps of all data to the reference time of the current acquisition cycle, which is completely matched with the preset acquisition step size, and unify the time reference of all data. S1.4 Output of multi-source raw dataset: After unifying the spatial and temporal benchmarks, a multi-source raw dataset with completely consistent spatiotemporal benchmarks is formed.

7. The method for improving the accuracy of ultra-short-term forecasting of distributed photovoltaic power generation according to claim 5, characterized in that: S2 specifically includes the following: S2.1 Extraction of core meteorological features: From the multi-source raw dataset output by S1, two core meteorological features are extracted: the degree of discrete fluctuation of the measured irradiance of all power stations in the target area, which reflects the degree of drastic change in irradiance in the area; and the average cloud cover in the target area, which reflects the overall cloud cover in the area. S2.2 Fixed Classification of Weather Scenarios: Based on the two extracted core meteorological features, the dataset is divided into four types of weather scenarios according to the pre-set fixed judgment rules: those with extremely low irradiance fluctuation and extremely low cloud coverage are classified as sunny steady-state scenarios. Low irradiance fluctuations and moderate cloud cover are classified as cloudy with slight changes; moderate irradiance fluctuations and moderate cloud cover are classified as cloudy with fluctuations; and extremely high irradiance fluctuations or extremely high cloud cover are classified as sudden change transit scenarios. S2.3 Scene Adaptive Preprocessing Operations: For the segmented weather scenes, perform scene-matched preprocessing operations, including time synchronization, scene adaptive noise removal, and missing value completion, specifically: Time synchronization: Unify and standardize all data to a time scale corresponding to the preset acquisition frequency, and fully align with the fixed acquisition frequency and prediction resolution; Scene-adaptive noise cleaning: A sliding smoothing method is used to remove abnormal fluctuation noise in the data. The window size of the smoothing process is bound to the weather scene. Specifically, the smallest window is used for sunny steady-state scenes, the medium window is used for cloudy weak-change and multi-cloud fluctuation scenes, and the largest window is used for sudden transition scenes to avoid over-smoothing that loses key features or under-smoothing that leaves residual noise. Missing value completion: For a small number of discontinuous missing data, the missing data is completed based on the valid values ​​of the adjacent time points before and after the missing data to ensure the integrity and continuity of the time series data; S2.4 Standardized Time Series Data Output and Credibility Weight Assignment: After preprocessing the three types of photovoltaic data, a standardized time series dataset is output; and a fixed credibility weight is assigned to each type of photovoltaic data according to the reliability level of the data source, with the highest weight for on-site measured data from power plants, followed by satellite cloud image data, and the lowest weight for numerical weather prediction data. The weights of all photovoltaic data form a complete fixed ratio system.

8. The method for improving the accuracy of ultra-short-term forecasting of distributed photovoltaic power generation according to claim 5, characterized in that: S3 specifically includes the following: S3.1 Dynamic Spatial Association Weight Update and Spatial Feature Mining: Based on the standardized time-series dataset of S2, the spatial association weights between power stations are dynamically adjusted using the cloud movement direction and speed in satellite cloud imagery data as the core basis and the physical spatial adjacency between power stations as the foundation. Power stations located upstream of the cloud movement path are assigned higher association weights, thus forming a dynamically updated spatial association system for power stations. For the standardized time-series datasets of the three types of photovoltaic data, the spatial association features corresponding to each type of photovoltaic data are mined by combining the dynamic spatial association system with the graph convolution unit built into the input module. S3.2, Time-series fluctuation feature capture: For the standardized time-series dataset in S2, the time-series data of each type of photovoltaic data is processed separately: First, the rate of change of the data over time is calculated to accurately identify the key inflection points of sudden changes in irradiance and step changes in power. The features corresponding to these inflection points are enhanced to strengthen the identification of key fluctuation information. Then, the enhanced time-series data is input into the built-in time-series convolution unit of the module to extract the time-series fluctuation features corresponding to each type of photovoltaic data. S3.3 Weighted Feature Fusion and Comprehensive Feature Vector Output: According to the exclusive credibility weight bound to each type of photovoltaic data, the spatial correlation features and temporal fluctuation features corresponding to the data source are weighted respectively; then the weighted spatial features and temporal features of all data sources are spliced ​​and fused to form a comprehensive feature vector that simultaneously contains the credibility differences of different data sources, spatial correlation information between power plants, and power temporal fluctuation information.

9. The method for improving the accuracy of ultra-short-term forecasting of distributed photovoltaic power generation according to claim 5, characterized in that: S4 specifically includes the following: S4.1 Initial Prediction Value Generation: Input the comprehensive feature vector output from S3 into the prediction model built into the module. Through the feature mapping and fitting rules built into the model, generate the initial prediction value of the power generation of each photovoltaic power station in the target area for the next 0-4 hours at the resolution corresponding to the preset acquisition frequency. S4.2 Calculation of prediction deviation in the previous acquisition cycle: Extract the measured power generation of each power station and the initial predicted power at the corresponding time in the previous complete acquisition cycle, calculate the degree of deviation between the measured value and the predicted value of each power station, and then summarize to obtain the average prediction deviation of the entire target area. S4.3 Online Weight Correction: A fixed deviation trigger threshold is preset, which forms a fixed proportional relationship with the average measured power of the region in the previous period; if the average prediction deviation of the region does not exceed the threshold, the feature mapping weights of the prediction model remain unchanged; if the average prediction deviation of the region exceeds the threshold, the gradient descent algorithm is used to correct the feature mapping weights of the prediction model online. S4.4 Final Prediction Output and Closed-Loop Link Completion: Substitute the corrected feature mapping weights into the prediction model to regenerate the final prediction value of ultra-short-term photovoltaic power generation for each power station in the target area for the next 0-4 hours; the final prediction value is synchronously transmitted to the grid dispatch terminal and the photovoltaic power station monitoring platform, and the measured power and the final prediction value of this cycle are fed back to the deviation calculation link of the next acquisition cycle to form a complete closed-loop correction link.

10. The method for improving the accuracy of ultra-short-term forecasting of distributed photovoltaic power generation according to claim 5, characterized in that: The initial predicted value generation is achieved by taking the comprehensive feature vector output by S3. Input the built-in prediction model, and map it through the model's generator function. Generate the initial predicted power of each power station in the region for the next 0-4 hours: in: The feature mapping weight matrix of the prediction model; For the first The future of the power station Initial predicted power generation at time [time] Predict the step size for the future.

Images