Photovoltaic power generation prediction method and related equipment

Abstract

The present application provides a photovoltaic power generation prediction method and related equipment. The method includes: obtaining power output information of a photovoltaic power station; obtaining key meteorological data based on a large meteorological model; fusing the key meteorological data and the power output information to obtain a data feature set; the key meteorological data and the power output information in the data feature set have a mapping relationship; training a first model based on the data feature set to obtain a second model; correcting the second model based on the real-time push data of the photovoltaic power station to obtain a third model; performing prediction error analysis on the third model, and in response to the prediction error being greater than a preset threshold, adjusting the parameters in the third model to obtain a fourth model; and using the fourth model to predict photovoltaic power generation. The present application improves the accuracy and reliability of photovoltaic power generation prediction through model optimization and correction, improves system adaptability, and ensures stable and efficient operation of power generation prediction.

Description

Technical Field

[0001] The present application relates to the field of photovoltaic power generation technology, and in particular to a photovoltaic power generation power prediction method and related equipment. Background Art

[0002] Current photovoltaic power generation prediction technology still faces significant challenges in dealing with complex meteorological conditions. Relying on meteorological data provided by numerical weather forecasts, the spatial resolution and temporal accuracy are limited, leading to large errors in forecast data, especially under extreme weather conditions. Furthermore, traditional machine learning models struggle to deeply exploit the spatiotemporal correlations in meteorological data and are insufficiently able to capture nonlinear characteristics. They are unable to effectively respond to sudden meteorological changes (such as cloud cover and severe convective weather), resulting in highly volatile and unstable forecast results. Furthermore, centralized computing platforms struggle to meet real-time forecasting requirements, limited by network load and transmission delays, further impacting the responsiveness and real-time nature of short-term power forecasting.

[0003] At the operational level of photovoltaic power stations, existing technologies fail to fully incorporate non-meteorological conditions such as equipment operating status and obstruction factors, which have a significant impact on power output. Traditional models often rely on simple feature engineering to model the relationship between meteorology and power, but lack consideration of the differentiated characteristics of equipment and non-meteorological factors. The model's generalization ability is insufficient, making it difficult to adapt to the individual needs of different power stations. In particular, in scenarios involving obstruction or low equipment efficiency, the accuracy of the prediction model decreases significantly. In addition, with the increasing scale and distribution of photovoltaic power stations, traditional equipment management methods are difficult to match the operating status of rapidly deployed equipment. Coupled with the topology identification problem of low-voltage distribution network access, the complexity of equipment communication and real-time data transmission is further increased.

[0004] In summary, the limitations of existing technologies primarily lie in the lack of accuracy and timeliness of meteorological data, the model's inability to express complex characteristics, and difficulties in communication and equipment management. Given these issues, PV power forecasting systems urgently require technological breakthroughs and in-depth optimization to ensure stability under complex meteorological conditions, real-time responsiveness, and adaptability to individual device characteristics. This is crucial to achieving precise support for renewable energy generation scheduling and widespread adoption. Summary of the Invention

[0005] In view of this, the purpose of this application is to propose a photovoltaic power generation power prediction method and related equipment.

[0006] Based on the above objectives, this application provides a photovoltaic power generation power prediction method, including:

[0007] Obtain power output information of photovoltaic power plants; obtain key meteorological data based on large meteorological models;

[0008] The meteorological key data and the power output information are fused to obtain a data feature set; the meteorological key data and the power output information in the data feature set have a mapping relationship;

[0009] Training the first model based on the data feature set to obtain a second model;

[0010] Based on the real-time push data of the photovoltaic power station, the second model is corrected to obtain a third model;

[0011] performing a prediction error analysis on the third model, and in response to a prediction error being greater than a preset threshold, adjusting parameters in the third model to obtain a fourth model;

[0012] The fourth model is used to predict photovoltaic power generation power.

[0013] In a possible implementation, fusing the key meteorological data and the power output information to obtain a data feature set includes:

[0014] Using sliding window technology, a time series analysis method is used to dynamically capture the temporal correlation between the key meteorological data and the power output information;

[0015] Using a data reconstruction algorithm, obtaining a spatial correlation relationship between the key meteorological data and the power output information;

[0016] Using a piecewise fitting algorithm, extracting power drop sequence data corresponding to power changes caused by cloud density parameters in the key meteorological data;

[0017] Taking the geographical location of the photovoltaic power station as a node, converting the data with spatiotemporal correlation in the key meteorological data into a graph structure, and using a graph neural network to adjust the temporal correlation relationship, the spatial correlation relationship, and the power drop sequence data;

[0018] The data feature set is obtained based on the adjusted time correlation relationship, the spatial correlation relationship and the power drop sequence data.

[0019] In one possible implementation, training the first model based on the data feature set to obtain the second model includes:

[0020] Processing the data feature set using a cross-validation algorithm, inputting the processed data feature set into the first model, and dynamically adjusting the learning rate of the first model using an adaptive learning rate optimization algorithm to obtain a post-training learning rate;

[0021] Adjusting the hyperparameters of the first model using a grid search algorithm or a random search algorithm to obtain trained hyperparameters;

[0022] The second model is obtained by training based on the post-training learning rate and the post-training hyperparameters.

[0023] In a possible implementation, the correcting the second model based on the real-time push data of the photovoltaic power station to obtain the third model includes:

[0024] Using the second model to predict the photovoltaic power generation power in a short time scale to obtain first prediction data;

[0025] Using the second model to predict photovoltaic power generation over a long time scale to obtain second prediction data;

[0026] Calculate a first error sequence based on at least two of the first prediction data and the corresponding at least two of the real-time pushed data;

[0027] Calculate a second error sequence based on at least two of the second prediction data and the corresponding at least two of the real-time pushed data;

[0028] Analyzing an error source of each error data in the first error sequence and the second error sequence to obtain a first result;

[0029] The second model is corrected according to the first result to obtain the third model.

[0030] In a possible implementation, the correcting the second model according to the first result includes:

[0031] In response to the first result being a sudden change in at least one of the key meteorological data, dynamically adjusting the weight of the corresponding key meteorological data to perform correction processing on the second model;

[0032] In response to the first result being that the equipment status of the photovoltaic power station is abnormal, parameters corresponding to the abnormal equipment are dynamically modified or added to perform correction processing on the second model.

[0033] In one possible implementation, performing prediction error analysis on the third model, and adjusting parameters in the third model in response to a prediction error being greater than a preset threshold to obtain a fourth model, includes:

[0034] Using the third model to predict the photovoltaic power generation power in a short time scale to obtain third prediction data;

[0035] Using the third model to predict photovoltaic power generation over a long time scale to obtain fourth prediction data;

[0036] Calculating a third error sequence based on at least two of the third prediction data and corresponding at least two historical push data;

[0037] Calculating a fourth error sequence based on at least two of the fourth prediction data and the corresponding at least two of the historical push data;

[0038] Analyzing an error source of each error data in the third error sequence and the fourth error sequence to obtain a second result;

[0039] Parameters in the third model are adjusted based on the second result to obtain the fourth model.

[0040] Based on the same inventive concept, the embodiment of the present application further provides a photovoltaic power generation power prediction device, comprising:

[0041] An acquisition module is configured to acquire power output information of a photovoltaic power station; and acquire key meteorological data based on a large meteorological model;

[0042] a fusion module configured to fuse the key meteorological data and the power output information to obtain a data feature set; wherein the key meteorological data and the power output information in the data feature set have a mapping relationship;

[0043] a training module, configured to train the first model based on the data feature set to obtain a second model;

[0044] a correction module configured to correct the second model based on the real-time push data of the photovoltaic power station to obtain a third model;

[0045] an adjustment module configured to perform a prediction error analysis on the third model, and in response to a prediction error being greater than a preset threshold, adjust parameters in the third model to obtain a fourth model;

[0046] The prediction module is configured to predict the photovoltaic power generation power using the fourth model.

[0047] Based on the same inventive concept, an embodiment of the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, it implements the photovoltaic power prediction method as described in any one of the above items.

[0048] Based on the same inventive concept, an embodiment of the present application further provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable the computer to execute any of the above-mentioned photovoltaic power generation prediction methods.

[0049] Based on the same inventive concept, an embodiment of the present application further provides a computer program product, which includes computer program instructions, and the computer instructions are used to enable the computer program product to execute any of the above-mentioned photovoltaic power generation power prediction methods.

[0050] As can be seen from the above, the photovoltaic power generation prediction method and related equipment provided by this application obtain power output information of a photovoltaic power station; obtain key meteorological data based on a large meteorological model; fuse the key meteorological data and the power output information to obtain a data feature set; the key meteorological data in the data feature set and the power output information have a mapping relationship; train a first model based on the data feature set to obtain a second model; calibrate the second model based on real-time push data from the photovoltaic power station to obtain a third model; analyze the prediction error of the third model, and in response to the prediction error being greater than a preset threshold, adjust the parameters in the third model to obtain a fourth model; and use the fourth model to predict photovoltaic power generation power. The embodiments of the present application demonstrate superiority in improving prediction accuracy, enhancing adaptability, and optimizing real-time performance. By collecting power output information from a photovoltaic power station and obtaining key meteorological data based on a large meteorological model, the two are fused to generate a high-quality feature set, fully exploring the spatiotemporal correlation between power output and meteorological characteristics. Through layer-by-layer optimization during model training, a dynamic adjustment mechanism is formed from the first model to the fourth model, achieving systematic optimization from data preparation to predictive analysis. This provides strong support for solving the problems of large errors, poor adaptability, and difficulty in accurately predicting under complex weather conditions in traditional photovoltaic power generation forecasts. It can also significantly improve the real-time and accuracy of photovoltaic power forecasts. In short-term forecasts, a sliding time window is used to dynamically capture the temporal correlation between weather and power, giving the forecast a dynamic response capability of seconds. In medium- and long-term forecasts, by integrating key influencing factors such as cloud density, the failure of the forecast model under variable weather conditions is avoided, providing a reliable basis for power grid scheduling and power generation planning. At the same time, through error analysis and parameter adjustment, the models at each stage can adapt to the impact of complex weather conditions and abnormal equipment status. For example, dynamic correction of parameters such as sand and dust index, cleanliness, and equipment operating efficiency further reduces the error of power forecasts. Further through the design of dynamic optimization, this application also enhances the adaptability and robustness of photovoltaic power forecasts. In error analysis, different sources are classified and processed, and model weights and parameters are adjusted as needed. This allows for a gradual optimization from the initial data feature set to the final prediction results, significantly improving the model's performance under extreme climate conditions. In summary, this application effectively improves the accuracy, stability, and applicability of predictions through scientific data fusion, model optimization, and a multi-layer correction mechanism, providing technical support for the efficient operation of photovoltaic power generation systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in this application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are merely embodiments of this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0052] Figure 1 This is a schematic diagram of the photovoltaic power generation power prediction method according to an embodiment of the present application;

[0053] Figure 2 This is a schematic diagram of the structure of a photovoltaic power generation prediction device according to an embodiment of the present application;

[0054] Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to make the objectives, technical solutions and advantages of this application more clear, this application is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0056] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should have the usual meanings understood by people with ordinary skills in the field to which this application belongs. The "first", "second" and similar words used in the embodiments of the present application do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0057] It is understandable that before using the technical solutions of each embodiment of the present disclosure, the type, scope of use, usage scenarios, etc. of the personal information involved will be informed to the user in an appropriate manner, and the user's authorization will be obtained.

[0058] For example, in response to a user's active request, a prompt message is sent to the user to clearly inform the user that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the electronic device, application, server, storage medium, or other software or hardware that performs the operation of the disclosed technical solution based on the prompt message.

[0059] As an optional but non-limiting implementation, in response to a user's active request, the prompt information may be sent to the user in the form of a pop-up window, in which the prompt information may be presented in text form. Furthermore, the pop-up window may also contain a selection control for the user to select "agree" or "disagree" to provide personal information to the electronic device.

[0060] It is understandable that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0061] As mentioned in the technical background section, current photovoltaic power generation prediction technology still faces significant challenges in dealing with complex meteorological conditions. Relying on meteorological data provided by numerical weather forecasts, the spatial resolution and temporal accuracy are limited, leading to large errors in the predicted data, especially under extreme weather conditions. Furthermore, traditional machine learning models struggle to deeply explore the spatiotemporal correlations of meteorological data and are insufficiently able to capture nonlinear features. They are unable to effectively respond to sudden meteorological changes (such as cloud cover and severe convective weather), resulting in high volatility and a lack of stability in the prediction results. Furthermore, centralized computing platforms struggle to meet real-time prediction requirements, limited by network load and transmission delays, further impacting the response speed and real-time performance of short-term power predictions.

[0062] At the operational level of photovoltaic power stations, existing technologies fail to fully incorporate non-meteorological conditions such as equipment operating status and obstruction factors, which have a significant impact on power output. Traditional models often rely on simple feature engineering to model the relationship between meteorology and power, but lack consideration of the differentiated characteristics of equipment and non-meteorological factors. The model's generalization ability is insufficient, making it difficult to adapt to the individual needs of different power stations. In particular, in scenarios involving obstruction or low equipment efficiency, the accuracy of the prediction model decreases significantly. In addition, with the increasing scale and distribution of photovoltaic power stations, traditional equipment management methods are difficult to match the operating status of rapidly deployed equipment. Coupled with the topology identification problem of low-voltage distribution network access, the complexity of equipment communication and real-time data transmission is further increased.

[0063] In summary, the limitations of existing technologies primarily lie in the lack of accuracy and timeliness of meteorological data, the model's inability to express complex characteristics, and difficulties in communication and equipment management. Given these issues, PV power forecasting systems urgently require technological breakthroughs and in-depth optimization to ensure stability under complex meteorological conditions, real-time responsiveness, and adaptability to individual device characteristics. This is crucial to achieving precise support for renewable energy generation scheduling and widespread adoption.

[0064] Taking the above into consideration, the present invention proposes a photovoltaic power generation prediction method and related equipment, which obtains power output information of a photovoltaic power station; obtains key meteorological data based on a large meteorological model; fuses the key meteorological data with the power output information to obtain a data feature set; the key meteorological data in the data feature set and the power output information have a mapping relationship; trains a first model based on the data feature set to obtain a second model; calibrates the second model based on real-time push data from the photovoltaic power station to obtain a third model; analyzes the prediction error of the third model, and in response to the prediction error being greater than a preset threshold, adjusts the parameters in the third model to obtain a fourth model; and predicts photovoltaic power generation using the fourth model. The present invention demonstrates advantages in multiple aspects, including improving prediction accuracy, enhancing adaptability, and optimizing real-time performance. Through the synergistic effect of multiple claims, the present invention overcomes the shortcomings of traditional prediction methods in terms of real-time performance, accuracy, and adaptability to complex environments, providing more efficient, accurate, and stable technical support for photovoltaic power generation prediction. First, a systematic optimization process is formed from data acquisition to model prediction, which changes the current situation where traditional prediction methods rely heavily on data quality. By introducing a large meteorological model to obtain key meteorological data, combined with the real-time power output information of photovoltaic power stations, the two are deeply integrated and processed to generate a high-quality data feature set. The changing patterns of photovoltaic power are explored from multiple dimensions such as time correlation, spatial correlation and complex weather influences, effectively solving the problems of single data input and insufficient accuracy. At the same time, the data feature set is processed using methods such as sliding window technology, segmented fitting algorithm and graph neural network, which improves the ability to express spatiotemporal features and solves the problem of insufficient extraction of correlation information between data in complex environments. Secondly, through a series of optimization algorithms, the model is trained and dynamically adjusted layer by layer, so that the present application has excellent expressiveness in both short-term and medium- and long-term power forecasting. During the training process, cross-validation, adaptive learning rate optimization algorithm and hyperparameter search technology are used to dynamically adjust the first model to the second model, so that the model can quickly adapt to the power change characteristics of different scenarios. Through real-time correction of the third model, the model parameters are analyzed based on the two sequences of short-term prediction error and long-term prediction error, and dynamic weight adjustment and device parameter adaptation are performed in combination with real-time push data, providing an accurate correction basis for generating the fourth model. This multi-stage correction mechanism effectively addresses the model's lack of prediction accuracy under sudden meteorological conditions (such as dust storms and changes in cloud density) and the difficulty in compensating for the impact of equipment state changes. Furthermore, by continuously optimizing the adaptability and robustness of the prediction model, this application is able to accurately predict power output in complex weather conditions and multi-region equipment scenarios. In practical applications, by dynamically optimizing the weights of equipment-related factors such as the dust index and cleanliness, the model is equipped to handle complex environments, reducing the margin of error caused by environmental changes or equipment anomalies.The dynamic error correction mechanism enables the model to have high stability in long-term operation, especially in high wind speed and frequent obstruction weather conditions, where the error is significantly reduced, enhancing the reliability and adaptability of the prediction results.

[0065] In summary, this application can comprehensively improve the accuracy, real-time performance and environmental adaptability of photovoltaic power generation prediction, provide important technical support for grid dispatching, energy storage optimization and intelligent management of photovoltaic stations, and effectively promote the application value of photovoltaic power generation systems in smart grids.

[0066] The technical solutions of the embodiments of the present application are described in detail below through specific examples.

[0067] refer to Figure 1 The photovoltaic power generation power prediction method of the embodiment of the present application includes the following steps:

[0068] Step S101, obtaining power output information of a photovoltaic power station; obtaining key meteorological data based on a large meteorological model;

[0069] Step S102: fusing the key meteorological data and the power output information to obtain a data feature set; the key meteorological data and the power output information in the data feature set have a mapping relationship;

[0070] Step S103: training the first model based on the data feature set to obtain a second model;

[0071] Step S104: calibrating the second model based on the real-time push data of the photovoltaic power station to obtain a third model;

[0072] Step S105, performing prediction error analysis on the third model, and in response to the prediction error being greater than a preset threshold, adjusting parameters in the third model to obtain a fourth model;

[0073] Step S106: predicting photovoltaic power generation power using the fourth model.

[0074] Regarding step S101, this application uses a large meteorological model and edge computing technology to push key meteorological data affecting photovoltaic power generation in real time, including solar irradiance, temperature, humidity, wind speed, etc. At the same time, power output information is extracted from the historical operating data of the photovoltaic power station. Mathematical morphology denoising methods are used to clean the data, image denoising methods are used to remove outliers, and nonlinear normalization methods are used to convert the data scale. These three methods are used in combination to improve the availability and accuracy of historical data.

[0075] In this embodiment, the photovoltaic power station is located in the northwest plateau area of my country, and the environmental characteristics are as follows:

[0076] Meteorological conditions: large temperature difference between day and night, large fluctuations in wind speed, frequent sandstorms and sudden cloud cover, and dramatic changes in irradiance.

[0077] Equipment characteristics: Photovoltaic power generation components and sensors are exposed to dusty environments. Significant changes in equipment cleanliness during long-term operation will affect power performance.

[0078] To ensure that the input data of subsequent models are of high quality and reliability, the initial input data needs to be comprehensively collected and rigorously preprocessed.

[0079] The details are as follows: This photovoltaic power station needs to collect two types of core data: meteorological data and photovoltaic equipment operation data.

[0080] Meteorological data collection: This photovoltaic power station is equipped with edge meteorological monitoring equipment (such as solar irradiance sensors and wind meters) to collect data in real time through IoT devices.

[0081] Access the National Meteorological Model Application Programming Interface (API) to obtain regional weather data to ensure comprehensive coverage of the data's temporal and spatial range.

[0082] The collected meteorological characteristics are as follows:

[0083] Solar irradiance: Updated every minute, directly from on-site sensors and compared to regional average irradiance through a large model.

[0084] Temperature: Collects near-ground temperature change data (real-time change frequency is 1 minute).

[0085] Humidity: used to analyze irradiance penetration and air clarity conditions.

[0086] Wind speed and direction: The wind speed and direction per minute are obtained from an anemometer to assess the impact of wind and sand on the shading and heat dissipation of photovoltaic equipment.

[0087] Cloud density and dust index: Cloud density is provided by a combination of satellite observation images and regional meteorological models; the dust index is generated by a combination of local monitoring stations and the national real-time dust warning platform.

[0088] Furthermore, edge computing technology is used to compress and optimize the data initially processed by meteorological sensors and send it to the central data server.

[0089] Transmission frequency: 1 minute update, real-time meteorological data is saved to the database to support short-term forecasting needs.

[0090] Collection of photovoltaic power station operation data: Historical power data: Power output curve data for the past year is extracted from the power station inverter and power station management system at minute granularity.

[0091] Electrical indicators such as power, current, and voltage recorded every minute are stored in the power station's historical operation database.

[0092] Equipment status data: Cleanliness status of photovoltaic panels: Use cleanliness monitoring equipment to record the pollution and dust accumulation of each photovoltaic panel.

[0093] Photovoltaic module operating efficiency: Real-time collection of module operating temperature rise data and efficiency parameters to evaluate equipment status.

[0094] The purpose of the collection is to ensure that real-time power data and historical performance data can be fully combined with meteorological data input models to provide effective data support for prediction feature extraction.

[0095] After data collection is completed, the collected raw data may contain noise, missing values, inconsistencies and other problems, so it needs to be fully preprocessed to ensure data quality.

[0096] Abnormal spikes were detected in the wind speed data during a certain period. Records showed that the normal wind speed was 57 m / s, but within a minute the wind speed jumped to 25 m / s before returning to normal. This was likely noise caused by a sensor malfunction. Mathematical morphological filtering was then used to denoise the data, removing these extreme spikes and smoothing the wind speed curve.

[0097] Satellite imagery data often contains random noise (such as clustered, point-like shadows) when clouds obscure the image, which can easily interfere with information. A convolutional neural network (CNN) denoising algorithm is used to process the image and filter out noise that is not part of natural cloud cover.

[0098] If a 10-minute period of photovoltaic power data is missing due to a sensor failure, causing that period to be mismatched with the meteorological data, interpolation is used to fill in the missing power values, and contextual data trend correction is performed to ensure data continuity.

[0099] Data standardization involves normalizing irradiance data (sampling range 200W / m² to 1300W / m²) to a range of 0.1 to ensure uniform input data magnitude and prevent large-value features from dominating model parameter training. Features such as temperature and humidity are converted to a standard distribution using a standardization method (Z-score normalization).

[0100] The power data and meteorological data are then aligned at a minute-level granularity to ensure that the data input to the model is consistent after alignment.

[0101] Environmental and regional characteristics may cause inconsistencies in data scale (such as time intervals and geographical deviations). This application ensures the consistency of data from different sources by processing time and space scales.

[0102] For example, some historical power record data is once an hour, while real-time meteorological data is once a minute: the hourly power data is interpolated to the minute level using interpolation method, and the weighted smoothing method is used to improve the fine-grained adaptability of short-term predictions.

[0103] Spatial normalization is also necessary. Because the geographic location of PV power stations in desert plateaus differs from that in plains, they are affected by the ambient albedo and the angle of sunlight incidence. Irradiance data values are corrected based on the station's altitude and latitude, combined with ground-based measurements. For areas with strong cloud cover, regional meteorological models are used to make distribution corrections.

[0104] The cleaned data is stored in a distributed database, which is divided into the following submodules: A historical data table stores historical power and meteorological data. A real-time data table stores real-time data, updated every minute. Regular verification of the cleaned data is required. For example, solar irradiance data can be compared with the values monitored by the National Meteorological Center to ensure the accuracy of local sensor data.

[0105] This application significantly improves the spatiotemporal resolution of meteorological data by introducing a large-scale meteorological model. Real-time, high-frequency updates (data collection every minute) ensure data timeliness. Integrating edge computing technology, local computing and real-time data push significantly reduce data transmission latency and improve data responsiveness. The comprehensive collection scope of key meteorological parameters affecting photovoltaic power (such as wind speed, wind direction, and cloud density) is expanded, and real-time access to monitoring equipment is provided, further addressing the blind spots of numerical weather forecasting in complex geographic environments. A combination of mathematical morphology denoising and image denoising techniques is employed to remove spikes, jitter, and abnormal deviations from the collected data. A nonlinear data normalization method is added to standardize data collected from different geographic regions or sensor devices, ensuring uniformity and comparability of input data across PV power stations. For real-time power data from PV power stations, efficient noise removal and data calibration methods, such as smoothing data fluctuations and removing extreme data points, ensure accuracy and consistency of input data. Specifically designed image denoising and cloud image analysis technologies clearly capture the impact of cloud cover on power generation, addressing data distortion in complex weather conditions. By dynamically comparing meteorological model prediction data with actual sensor data, this technology can automatically adjust data quality under complex weather conditions and feedback data anomalies, thereby improving the model's adaptability to complex climate conditions.

[0106] For step S102, in some embodiments, the meteorological key data and the power output information are fused to obtain a data feature set, including: using a sliding window technology to dynamically capture the time correlation relationship between the meteorological key data and the power output information through a time series analysis method; using a data reconstruction algorithm to obtain the spatial correlation relationship between the meteorological key data and the power output information; using a piecewise fitting algorithm to extract power drop sequence data corresponding to power changes caused by cloud density parameters in the meteorological key data; using the geographical location of the photovoltaic power station as a node, converting the data with time and space correlation in the meteorological key data into a graph structure, and using a graph neural network to adjust the time correlation relationship, the spatial correlation relationship and the power drop sequence data; based on the adjusted time correlation relationship, the spatial correlation relationship and the power drop sequence data, the data feature set is obtained.

[0107] The above steps have completed the collection and preprocessing of meteorological and PV operation data for the desert plateau PV power station, providing accurate basic data input. The goal of this step is to deeply integrate real-time meteorological data with historical power generation data, exploring the nonlinear spatiotemporal correlations between meteorological factors and power output, and providing high-quality input for subsequent model training.

[0108] In this embodiment, a large-scale meteorological model is integrated with photovoltaic power generation data. Based on collected meteorological data, the large-scale meteorological model is used to make short-term and medium-term forecasts of future weather conditions. Combined with historical power generation data from photovoltaic power plants, the mapping relationship between meteorological factors and photovoltaic power generation is analyzed. A photovoltaic power plant power data reconstruction algorithm is proposed that considers wind direction and spatiotemporal correlations. The relationship between wind direction and speed and power plant delay time is established, extracting the power drop sequence caused by cloud cover in the photovoltaic curve. A fitting model is then constructed using a graph neural network to effectively reconstruct the photovoltaic power plant power curve, improving the accuracy of the prediction input feature set data.

[0109] This photovoltaic power station has completed the collection and preprocessing of the following data: meteorological data: solar irradiance (updated every minute), temperature, humidity, wind speed, wind direction, cloud density, sand and dust index and other complex meteorological characteristics.

[0110] Photovoltaic power generation data: historical power data (recorded at 1-minute granularity for the past year), real-time power output, and photovoltaic power station equipment status (such as cleanliness and operating efficiency).

[0111] These data have been normalized according to time and space scales and stored in a distributed database.

[0112] By analyzing meteorological data and power data, we explore the impact of weather changes on power output and establish a global mapping relationship between time and space.

[0113] Use historical data to analyze the time series relationship between meteorological and power changes.

[0114] Utilize sliding time window technology: Design a sliding window centered on the current time data with a time span of 3 hours to dynamically capture the correlation between historical meteorological data and power output.

[0115] In this example, analysis revealed that solar irradiance has a 30-minute lag in power output, and that temperature changes have a more significant impact on power efficiency on an hourly scale. Therefore, the data extraction rules are set in the sliding window:

[0116] Irradiance characteristics: Use data from the past hour and focus on capturing changes in the past 30 minutes.

[0117] Temperature and humidity: The window length is set to the past 3 hours to extract trend features of gradual changes.

[0118] Then, a time series feature matrix containing multi-period features is dynamically constructed for subsequent model training.

[0119] Given the unique environment of desert photovoltaic power plants, this study analyzes the temporal and spatial impacts of geographical features on their operation. Measured data shows that periods of high wind speed create a self-cleaning effect on the surface of photovoltaic panels, improving equipment efficiency. However, the combination of low wind speeds and sandstorms significantly impacts power output.

[0120] In this example, at 2:00 PM on a particular day, the wind speed dropped to 2 m / s, while the dust index increased, causing the power output to be significantly lower than expected. The model extracts the "wind-speed-dust coupling effect" feature to describe the power obstruction caused by dust.

[0121] In some cases, power data from 10:00 AM to 10:10 AM on a particular day is missing, hindering analysis of hysteresis characteristics. A fitting algorithm (such as spline interpolation) is used to complete the missing power sequence: a segmented fit is performed on the historical irradiance and power relationship over a 10-minute period, reconstructing the power curve. This completed sequence effectively fills the gaps and provides coherence for the model input.

[0122] Furthermore, a regional feature reconstruction model based on Graph Neural Network (GNN) is constructed to further improve data quality through high-dimensional feature coupling.

[0123] Each node represents a data collection point for a specific PV power plant (based on geographic location). Node characteristics include: time series meteorological characteristics (irradiance, temperature, humidity, cloud density, etc.) and power history curve characteristics (10-minute average and rate of change).

[0124] Edges between nodes are established based on geographic proximity. The edge weight between two nodes is calculated based on the associated characteristics of the meteorological data. For example, if cloud cover characteristics are highly similar, the edge weight will be greater than 0.8. If high wind areas are connected to low wind areas, the edge weight will decrease.

[0125] The constructed graph neural network is used to complete and reconstruct the photovoltaic power curve. Specifically, if dust obstruction causes power to drop on a certain day, the irradiance-power relationship between nodes is extracted to draw a characteristic curve of the obstruction impact. The resulting optimized reconstructed power curve is more closely aligned with the actual fluctuations.

[0126] After the above processing, the optimized data input feature set is generated:

[0127] Meteorological characteristics: dynamically changing solar irradiance, cloud density, and temperature series (including sliding window features).

[0128] Power characteristics: historical power data, complete power drop sequence.

[0129] Spatiotemporal features: Reconstructing data based on features of graph neural networks.

[0130] The above high-quality input features will be directly used in the next model training stage.

[0131] This embodiment of the present application introduces a sliding time window and time series feature analysis method to dynamically extract the temporal correlation of meteorological features (such as the time period when solar irradiance lags and affects power changes), thereby improving the model's dynamic response to complex meteorological changes. Through spatial feature analysis, a reconstruction algorithm based on wind speed, wind direction, and cloud density is designed to deeply explore regional specificities (for example, the relationship between wind speed and irradiance scatter in mountainous areas), addressing the limitations of traditional models in handling unique meteorological conditions. A graph neural network is used to establish the spatiotemporal correlations of meteorological data, modeling the regional dynamics of multiple photovoltaic power stations as a graph structure. This not only analyzes the connections between nodes but also handles nonlinear feature changes under specific spatial conditions. A nonlinear fitting algorithm for power data is constructed, and by segmented reconstruction of photovoltaic power station power curves, it can complete the power drop sequence caused by cloud obstruction, thereby increasing data robustness. A graph neural network is used to model the complex relationships between meteorological factors, such as the joint impact of wind direction and cloud density on power changes, which is incorporated into the network architecture, enabling nonlinear and in-depth mining of meteorological factors. Meteorological characteristics across different regions are calibrated based on their geographical distribution, eliminating the interference of heterogeneous meteorological data on forecast results and optimizing data input accuracy. A graph neural network is used to transform meteorological data into a high-dimensional graph structure. Dynamic changes in wind speed, irradiance, and cloud cover sequences are correlated through node-edge relationships, upgrading the traditional "single-metric feature extraction" approach to "spatiotemporal integrated feature mining," improving the quality of the power prediction feature set. Data completion and correction techniques are enhanced to handle missing points in collected data, incorporating power drop sequences as part of the simulation input, addressing the issue of incomplete data under complex meteorological conditions. Power drop data caused by cloud cover is extracted as a piecewise fitting sequence. A regional dynamic data completion algorithm is designed to effectively reconstruct the power output curve, making forecast results more stable under complex conditions. Dynamic reconstruction of the effects of wind speed and direction is added to the meteorological factor analysis to provide more complete input data for power characteristics when predicting sudden weather changes, thereby improving the model's responsiveness to unstable weather conditions.

[0132] For step S103, the first model is trained based on the data feature set to obtain the second model, including: processing the data feature set using a cross-validation algorithm, inputting the processed data feature set into the first model, and dynamically adjusting the learning rate of the first model using an adaptive learning rate optimization algorithm to obtain a post-training learning rate; adjusting the hyperparameters of the first model using a grid search algorithm or a random search algorithm to obtain post-training hyperparameters; and training the second model based on the post-training learning rate and the post-training hyperparameters.

[0133] In this example, a deep learning approach was first used to model meteorological data and historical photovoltaic power generation data to construct a photovoltaic power generation prediction model. Subsequently, temporal neural networks, such as long-short-term memory networks or gated recurrent units, were used to model and test the temporal correlation between meteorological variables and power generation. The network structure parameters were adjusted during the training process to optimize model performance. Finally, cross-validation and grid search were used to systematically optimize hyperparameters, improving prediction accuracy and robustness.

[0134] The above steps have generated a high-quality feature dataset. This data is ready for deep learning model training, with the goal of building a high-accuracy photovoltaic power prediction model.

[0135] In this embodiment, the input is an optimized set of meteorological features, including minute-by-minute irradiance, humidity, temperature, and other time series data combined with a sliding window. Historical power curves are used as auxiliary input to capture power generation patterns.

[0136] The LSTM network consists of two layers: the first layer, consisting of 128 hidden units, is used to extract the relationship between short-term weather characteristics and power; the second layer, consisting of 64 hidden units, is used to model the relationship between power and long-term meteorological changes.

[0137] Data segmentation rules:

[0138] From the optimized feature set generated in the second step, it is divided into the following according to the time series:

[0139] Training set: 70% of the dataset is used for model parameter learning. It includes power data and corresponding meteorological characteristics from the past year.

[0140] Validation set: 20% of the dataset, used for model validation. This includes data from the past two months and is used to assess whether the model can effectively avoid overfitting.

[0141] Test set: 10% of the dataset, used for performance evaluation. This includes data from the most recent week and is used for the final test of predictive power.

[0142] A 5-fold cross-validation scheme is used: the training set is divided into five groups, four of which are used for training each time, and the remaining one is used for validation to ensure that the model is universal for different data subsets.

[0143] Each training session inputs a feature sequence spanning a time window. For example, a meteorological time series feature matrix covering the past three hours (12 time steps, each 15 minutes apart) is input. A power change sequence is also input as an auxiliary feature. Power and meteorological features are further aligned by time period, for example, by pairing the 30-minute lag effect of irradiance features with power changes. Through the model training process, the model gradually captures the complex nonlinear relationship between time series features and power output.

[0144] The Adaptive Learning Rate (Adam) optimizer was used during the optimization process. Adaptive learning rate adjustment accelerates convergence and improves model training efficiency. A learning rate decay and cosine annealing strategy were employed. Initially, the learning rate was set to 0.001. A high learning rate was maintained until 20 epochs of training to facilitate rapid convergence. After the 20th epoch, the learning rate was gradually reduced to avoid overfitting in the later stages of training.

[0145] The model's key hyperparameters were set as search ranges, and the optimal combinations were tested one by one: number of hidden units in the first layer: 64, 128, and 256. Time window lengths: 2 hours, 3 hours, and 4 hours. Batch sizes: 16, 32, and 64. A grid search revealed the optimal model configuration: 128 hidden units in the first layer and 64 hidden units in the second layer. The time window length was set to 3 hours, and the batch size was 32. Random search was performed to supplement optimization: hyperparameter combinations were randomly selected and further verified using the grid search results.

[0146] For step S104, in some embodiments, the second model is corrected based on the real-time push data of the photovoltaic power station to obtain a third model, including: using the second model to predict the photovoltaic power generation power on a short time scale to obtain first prediction data; using the second model to predict the photovoltaic power generation power on a long time scale to obtain second prediction data; calculating a first error sequence based on at least two of the first prediction data and the corresponding at least two real-time push data; calculating a second error sequence based on at least two of the second prediction data and the corresponding at least two real-time push data; analyzing the error source of each error data in the first error sequence and the second error sequence to obtain a first result; and correcting the second model according to the first result to obtain the third model.

[0147] In some embodiments, the correction processing of the second model based on the first result includes: in response to the first result being a sudden change in at least one of the meteorological key data, dynamically adjusting the weight of the corresponding meteorological key data to correct the second model; in response to the first result being an abnormal equipment status of the photovoltaic power station, dynamically modifying or adding parameters corresponding to the abnormal equipment to correct the second model.

[0148] In this embodiment, a classification algorithm for daily rolling weather forecast data is designed. After screening and completion, the data is input into the model to calculate the photovoltaic power generation power at different time scales (15 minutes to the next 10 days). Then, based on the actual power data, the prediction error is analyzed, and an adaptive optimization algorithm is used to perform real-time error correction to improve the prediction accuracy.

[0149] Specifically, for short-term power forecasting, the system inputs meteorological feature sequences and historical power data from the past three hours, using a sliding window to capture real-time changes in the current time period. Features include: solar irradiance change sequence (updated every minute), wind speed and temperature series, and a historical power change curve from the past three hours. A second model is used to perform short-term forecasts (15 minutes to two hours). In some feasible embodiments, under conditions of intensified sandstorms, the second model predicts power of 16.8 MW in the next 15 minutes and 17.5 MW in the next two hours.

[0150] For long-term power forecasting, we input power curves and meteorological forecast data from the past two days, combined with the changing trends of the dust index and cloud density. Input features include the forecast results of a large-scale meteorological model covering the next 10 days, including the diurnal variation of solar irradiance and the dynamic trends of temperature and humidity. A second model was used to predict hourly power changes over the next three days. The second model predicted power for the next three days: 12 MW in the morning on the first day, 18 MW at noon, and a total of 180 MW for the entire day. On the second day, power gradually recovered as the dust storm weakened, reaching a total of 200 MW.

[0151] Real-time power data received from the PV plant is used to compare with short-term forecasts. The error between the short-term forecast and real-time power is calculated by time period. For example, the predicted power on the first day is 16.8 MW, while the real-time measured power is 16.6 MW, with an error of 1.2%.

[0152] The power forecast for the next three days was compared with actual power generation to generate a long-term error series. This error series was generated by comparing hourly forecasts with actual values. The total power difference was calculated daily to evaluate the model's long-term forecast performance. The second model predicted a power of 200 MW for the next day, but the actual power generation was 195 MW, for an error rate of 2.56%.

[0153] The impact of sudden changes in the dust index on power forecasts was analyzed. The dust index rose rapidly to 90 that morning, but only an index of 70 was considered in the input features, resulting in a larger deviation in power forecasts.

[0154] Check whether discrepancies between the cleanliness data and actual conditions of PV modules cause deviations in power predictions. Cleanliness monitoring showed that the surface contamination of the modules reached 40%, and the model did not fully compensate for the contamination effect, resulting in an overestimation of power.

[0155] Analyze the forecast bias caused by regional weather differences. The actual shaded area of PV panels in the north was underestimated, while the model defaulted to a high shade rate.

[0156] Based on meteorological error results, the dust index weight is dynamically optimized. The dust index weight is adjusted from 0.7 to 0.9 to more accurately reflect obstruction effects. To compensate for power characteristics, a cleanliness factor is added as a new input feature. If the error result significantly exceeds the preset threshold (short-term error >3%, long-term error >5%), model retraining is triggered. For error sequences in the all-day power forecast, high dust environment data from the past two days is added as training samples to update the model structure.

[0157] Real-time analysis of key meteorological data can address forecast deviations caused by sudden changes in meteorological data, such as a sudden drop in solar irradiance, a significant increase in the dust index, and a sudden change in cloud density. At 10:00 a.m. one day, the dust index rose sharply from 60 to 95 within a short period of time (15 minutes), causing the power forecast error to increase from 3% to 6%.

[0158] Based on the correlation between the characteristics of sudden changes in key meteorological data, the weights of related features were increased or decreased: the weight of the dust index was increased from an initial value of 0.7 to 0.9 (increasing the impact of obstruction). The weight of temperature was decreased to reduce the impact of temperature changes on power output when dust concentration exceeded 90. After adjusting the dust index weight, the model readjusted its prediction of obstruction range, reducing the error in power prediction from 6% to 3.5%.

[0159] For abnormal equipment status, we analyze the real-time status of PV modules and compare it with historical data and monitoring system records. Common anomalies include: increased module contamination (dust accumulation); equipment maintenance status failure (reduced PV panel connection efficiency); and PV panel cleanliness monitoring showing 40% module surface coverage (dust accumulation), resulting in low power output. The model does not fully compensate for this change in cleanliness.

[0160] In the second model, add or dynamically adjust equipment-related input parameters, such as the component cleanliness index and the temperature rise efficiency compensation factor. Dynamically set the initial cleanliness index to 0.8, and introduce a maintenance status compensation parameter to simulate the power reduction caused by dust accumulation.

[0161] Modify abnormal equipment status parameters: Based on real-time data, adjust old parameter values or add new features. Reduce the weight of equipment with poor cleanliness (reflecting low efficiency). Enhance the power prediction relevance of well-maintained equipment. By setting additional parameters in the second model for components with high dust accumulation, the power decline trend was accurately corrected, and the prediction error was reduced from 5% to 3%.

[0162] The present embodiment categorizes prediction errors (such as meteorological data errors and equipment sensor errors) and designs targeted correction strategies based on the source of the deviation. For example, to address power fluctuations caused by complex meteorological conditions, model feature weights are adjusted in real time (for example, the dynamic impact of irradiance changes is weighted higher). A lightweight error correction model was developed, incorporating sliding window technology to dynamically compensate for and adjust the error between the predicted and true values in real time. For sudden cloud cover or sudden wind speed changes, real-time correction based on segmented optimization is implemented to improve the model's adaptability to severe power fluctuations. For sudden power drops caused by cloud cover, historical patterns of irradiance drops are extracted in real time to compensate for future power trends. Multiple maintenance positive factors (such as the coupled effects of wind speed, temperature, and humidity on power output) are introduced to dynamically compare actual power variation characteristics to correct prediction errors under complex weather conditions. Real-time updates of meteorological forecast data (such as monitoring of cloud cover dynamics) ensure more accurate model input data under complex conditions by supplementing incomplete or out-of-date meteorological data. An error feedback optimization system was designed to record and track deviations from each forecast correction. Weighted optimization and dynamic parameter adjustment mechanisms were used to enhance the system's overall forecasting capabilities. Daily dynamic comparisons were performed between forecast results and actual power generation data to automatically calibrate the weights of key features in the model (e.g., real-time updates of the influence coefficients of features like cloud density and irradiance). The sliding window range was adjusted based on the error sequence, refining the time span of the input features. The operating status, maintenance records, and current efficiency of photovoltaic equipment (e.g., panel cleanliness) were collected, and a forecast correction process was introduced to compensate for deviations caused by differences in equipment status. For different types of photovoltaic stations, the model inputs were calibrated based on their historical power generation patterns, and adaptive weight settings tailored to specific regions were used to enhance the model's generalization capabilities.

[0163] For step S105, in some embodiments, the prediction error analysis is performed on the third model, and in response to the prediction error being greater than a preset threshold, the parameters in the third model are adjusted to obtain a fourth model, including: using the third model to predict the photovoltaic power generation power on a short time scale to obtain third prediction data; using the third model to predict the photovoltaic power generation power on a long time scale to obtain fourth prediction data; calculating a third error sequence based on at least two of the third prediction data and corresponding at least two historical push data; calculating a fourth error sequence based on at least two of the fourth prediction data and corresponding at least two historical push data; analyzing the error source of each error data in the third error sequence and the fourth error sequence to obtain a second result; and adjusting the parameters in the third model based on the second result to obtain the fourth model.

[0164] In this embodiment, a dynamic deviation-driven optimization method is designed to compare and analyze operational data with historical grid-side power consumption data to determine the model's forecast deviation. If the forecast error exceeds a set threshold, the model optimization mechanism is triggered, dynamically adjusting meteorological feature selection and data weighting to ensure continuous optimization of the forecast system.

[0165] In this example, for short-term predictions, the past two hours of meteorological data and recent power output data are input, combined with real-time environmental change parameters. This includes meteorological data (such as solar irradiance and dust index) and historical power data. A third model is used to predict power changes for the next 15 minutes, 30 minutes, one hour, and two hours. Given a current dust index of 90 and irradiance of 800 W / m², the predicted power for the next 15 minutes is 16.6 MW, and the next hour is 17.2 MW. Real-time data is received from the photovoltaic power plant and compared with the predictions of the third model. A short-term error sequence is generated, recording the error for each time period (e.g., 15-minute intervals). The predicted power for the first day was 16.8 MW, while the measured power was 16.5 MW, resulting in an error of 1.82%. Similarly, errors are recorded for multiple time periods, resulting in the following error sequence: Time period 1: 1.82%, Time period 2: 1.5%, and Time period 3: 2.8%.

[0166] For long-term predictions, the power change data for the last three days and the meteorological forecast series were input, combined with the daily dust index and cloud density trends. A third model was used to predict power trends for the next three days (hourly distribution). As the dust storm gradually weakened, the third model predicted a full-day power of 180 MW on the first day. On the second day, power rebounded, reaching a total of 195 MW. On the third day, power further recovered to 210 MW. The predicted data for the next three days was compared with the accumulated actual power values to generate a long-term error series. The predicted power for the first day was 180 MW, while the actual power was 175 MW, with an error rate of 2.86%. The error rates for the second day were 1.56%, and the third day were 2.1%.

[0167] Analysis of the error sequence identified the main sources of deviation: sudden meteorological changes (such as unusually high dust concentrations); decreased module operating efficiency (low cleanliness); and regional weather variations leading to uneven power forecasts. The impact of a dust index rising to 95 on power obstruction was underestimated, with model prediction errors concentrated between 10:00 and 11:00 a.m. Modules in the southern region had 80% clean coverage, and the actual power reduction trend was lower than the model prediction. Tracing the source of the high power forecast in the northern region revealed an overestimation of the irradiance factor.

[0168] Based on error source analysis, model parameters were adjusted: the weight of the sudden meteorological parameter, the dust index, was increased from 0.8 to 0.9. A compensation factor for decreased irradiance in the northern region was added to mitigate the impact of environmental obstruction. The model was retrained using error sequence feedback: data on dust increases and regional obstruction over the past three days was extracted and incorporated into new training samples. The number of hidden layer neurons in the LSTM model was increased from 128 to 256. After adjusting the parameters and retraining the model, the power forecast error on the third day was reduced from 5.5% to 3.2%.

[0169] The embodiments of the present application significantly reduce the errors in short-term and long-term predictions, solve the problems of sudden weather changes, abnormal equipment status and regional differences, make the model more adaptable in complex environments, and provide guarantees for the accuracy and robustness of photovoltaic power generation predictions.

[0170] With respect to step S107 , the photovoltaic power generation power is predicted using the fourth model.

[0171] In this embodiment, PV power forecast results are provided to the grid dispatch center and energy management system to assist with decisions such as grid load scheduling and energy storage system optimization. Error trends during the forecast process are also recorded, and machine learning methods are used to continuously optimize the forecast model, improving its adaptability to complex weather conditions and ensuring the stability and efficiency of the PV power generation system.

[0172] Through the above embodiments, it can be seen that the photovoltaic power generation prediction method described in the embodiment of the present application obtains the power output information of the photovoltaic power station; obtains meteorological key data based on the meteorological large model; fuses the meteorological key data and the power output information to obtain a data feature set; the meteorological key data and the power output information in the data feature set have a mapping relationship; trains the first model based on the data feature set to obtain the second model; corrects the second model based on the real-time push data of the photovoltaic power station to obtain the third model; analyzes the prediction error of the third model, and in response to the prediction error being greater than the preset threshold, adjusts the parameters in the third model to obtain the fourth model; and uses the fourth model to predict the photovoltaic power generation power. The embodiment of the present application provides a comprehensive solution for photovoltaic power generation power prediction that combines accuracy, real-time and adaptability to complex environments. In response to the problem of insufficient prediction accuracy caused by the existing technology relying on traditional meteorological data, the present application introduces a meteorological large model to realize real-time collection and prediction of high-resolution meteorological data, and constructs a deeply integrated feature set by combining it with the real-time power generation power and equipment operation status data of the photovoltaic power station. This feature set can reveal the changing patterns of photovoltaic power from multiple dimensions, including temporal and spatial correlations, as well as complex meteorological dynamics. This significantly improves the problems of low data input quality and the difficulty of adapting single features to complex scenarios. Furthermore, through dynamic feature selection and interactive processing, it addresses the inability of traditional models to express the temporal and spatial correlations and nonlinear characteristics of data, providing a precise basis for analyzing and predicting power fluctuations caused by sudden weather changes (such as sandstorms and severe convective weather).

[0173] In terms of model construction and optimization, this application adopts a series of innovative algorithms and dynamic adjustment mechanisms to effectively improve the adaptability and accuracy of the power prediction model. In multi-stage layer-by-layer training, through cross-validation, adaptive learning rate optimization, dynamic weight adjustment and other technologies, the model can quickly adapt to the dynamic characteristics of photovoltaic power changing with equipment status and complex weather. The model combines short-term to medium- and long-term forecasts with a real-time correction mechanism to gradually optimize short-term forecast errors and long-term power change laws, providing efficient real-time correction capabilities for power fluctuations caused by sudden meteorological conditions. In terms of processing non-meteorological factors, the model fully considers equipment characteristics such as the operating status, cleanliness, and shading effects of photovoltaic equipment, and accurately compensates for local differences in equipment and environmental impacts through dynamic weight adjustment, thereby solving the problem that traditional models fail to fully consider changes in equipment status and lead to prediction deviations.

[0174] In addition, in response to the deficiencies in the data collection and real-time transmission links of photovoltaic power stations, this application realizes the rapid acquisition and synchronous processing of real-time data from multiple sites through embedded data collection and edge computing technology, effectively alleviating the transmission delay problem caused by network load on traditional centralized computing platforms and improving the short-term prediction response speed. In distributed photovoltaic scenarios, this method can quickly adapt to the real-time observation and data management requirements under the background of frequent equipment access and exit, and significantly enhance the stability and real-time performance of data transmission in photovoltaic power stations by continuously optimizing equipment ledgers and communication systems. In addition, through the synchronous matching of equipment and data, the contradiction between the traditional ledger update speed and the demand for rapid equipment access is resolved, providing high-quality data support for power grid dispatching.

[0175] Based on the above-mentioned innovative technologies, this application has demonstrated excellent environmental adaptability and prediction performance under complex meteorological conditions and multi-regional scenarios. Regardless of whether it is strong wind speed, cloudy weather, or large-scale access to distributed photovoltaics, the prediction results can meet the requirements of grid load scheduling and energy storage optimization in terms of accuracy and stability. At the same time, combined with the dynamic error correction mechanism and continuous model optimization capabilities, this application has strong robustness for long-term application scenarios of photovoltaic power prediction, so that equipment status changes and complex meteorological influences are no longer important factors that hinder prediction accuracy. Overall, this application comprehensively improves the functionality and adaptability of the photovoltaic power generation prediction system, truly solves the multiple problems of existing technologies in data quality, model capabilities and actual application scenarios, and provides reliable support for the intelligent management of photovoltaic power generation systems, grid scheduling optimization and improvement of new energy absorption efficiency.

[0176] It should be noted that the method of the embodiment of the present application can be performed by a single device, such as a computer or server. The method of this embodiment can also be applied in a distributed scenario and performed by multiple devices working together. In such a distributed scenario, one of the multiple devices may only perform one or more steps of the method of the embodiment of the present application, and the multiple devices will interact with each other to complete the method.

[0177] It should be noted that the above description is limited to some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in an order different from that described in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0178] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides a photovoltaic power generation power prediction device.

[0179] refer to Figure 2 , the photovoltaic power generation power prediction device comprises:

[0180] The acquisition module 21 is configured to obtain power output information of the photovoltaic power station; obtain key meteorological data based on the meteorological model;

[0181] A fusion module 22 is configured to fuse the key meteorological data and the power output information to obtain a data feature set; the key meteorological data and the power output information in the data feature set have a mapping relationship;

[0182] A training module 23 is configured to train the first model based on the data feature set to obtain a second model;

[0183] a correction module 24 configured to correct the second model based on the real-time push data of the photovoltaic power station to obtain a third model;

[0184] an adjustment module 25 configured to perform a prediction error analysis on the third model, and in response to a prediction error being greater than a preset threshold, adjust parameters in the third model to obtain a fourth model;

[0185] The prediction module 26 is configured to predict the photovoltaic power generation power using the fourth model.

[0186] For the convenience of description, the above devices are described as being divided into various modules according to their functions. Of course, when implementing this application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0187] The device of the above embodiment is used to implement the corresponding photovoltaic power prediction method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0188] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein when the processor executes the program, the photovoltaic power generation power prediction method described in any of the above embodiments is implemented.

[0189] Figure 310 is a schematic diagram showing a more specific hardware structure of an electronic device provided in this embodiment. The device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are communicatively connected to each other within the device via the bus 1050.

[0190] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0191] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage devices, dynamic storage devices, etc. The memory 1020 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0192] The input / output interface 1030 is used to connect to input / output modules to enable information input and output. The input / output modules can be configured as components within the device (not shown) or externally connected to the device to provide corresponding functions. Input devices may include a keyboard, mouse, touch screen, microphone, and various sensors. Output devices may include a display, speaker, vibrator, indicator light, and the like.

[0193] The communication interface 1040 is used to connect to a communication module (not shown) to enable communication between the device and other devices. The communication module can communicate via wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, Wi-Fi, Bluetooth, etc.).

[0194] The bus 1050 comprises a path for transmitting information between the various components of the device (eg, the processor 1010 , the memory 1020 , the input / output interface 1030 , and the communication interface 1040 ).

[0195] It should be noted that although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040, and the bus 1050, in a specific implementation, the device may also include other components necessary for normal operation. In addition, it will be understood by those skilled in the art that the above device may only include the components necessary to implement the embodiments of this specification, and does not necessarily include all the components shown in the figure.

[0196] The electronic device of the above embodiment is used to implement the corresponding photovoltaic power prediction method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0197] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the photovoltaic power generation power prediction method described in any of the above embodiments.

[0198] The computer-readable media of this embodiment includes permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0199] The computer instructions stored in the storage medium of the above embodiment are used to enable the computer to execute the photovoltaic power prediction method described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0200] Based on the same inventive concept, corresponding to the photovoltaic power prediction method described in any of the above embodiments, the present disclosure also provides a computer program product comprising computer program instructions. In some embodiments, the computer program instructions can be executed by one or more processors of a computer to cause the computer and / or the processor to perform the photovoltaic power prediction method. For each step in each embodiment of the photovoltaic power prediction method, the processor executing the corresponding step may belong to the corresponding execution entity.

[0201] The computer program product of the above embodiment is used to enable the computer and / or the processor to execute the photovoltaic power prediction method described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0202] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present application (including the claims) is limited to these examples. Within the scope of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0203] In addition, to simplify the description and discussion, and to avoid obscuring the understanding of the embodiments of the present application, well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided figures. Furthermore, devices may be shown in block diagram form to avoid obscuring the understanding of the embodiments of the present application, and this also takes into account the fact that the implementation details of these block diagram devices are highly dependent on the platform on which the embodiments of the present application will be implemented (i.e., these details should be fully understood by those skilled in the art). Where specific details (e.g., circuits) are set forth to describe the exemplary embodiments of the present application, it will be apparent to those skilled in the art that the embodiments of the present application can be implemented without these specific details or with variations therefrom. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0204] Although the present invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may utilize the discussed embodiments.

[0205] The embodiments of the present application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of this application.

Claims

1. A photovoltaic power generation power prediction method, characterized in that: include: Obtain power output information of photovoltaic power plants; obtain key meteorological data based on large meteorological models; The meteorological key data and the power output information are fused to obtain a data feature set; the meteorological key data and the power output information in the data feature set have a mapping relationship; Training the first model based on the data feature set to obtain a second model; Based on the real-time push data of the photovoltaic power station, the second model is corrected to obtain a third model; performing a prediction error analysis on the third model, and in response to a prediction error being greater than a preset threshold, adjusting parameters in the third model to obtain a fourth model; Predicting photovoltaic power generation using the fourth model; The step of fusing the key meteorological data and the power output information to obtain a data feature set includes: Using sliding window technology, a time series analysis method is used to dynamically capture the temporal correlation between the key meteorological data and the power output information; Using a data reconstruction algorithm, obtaining a spatial correlation relationship between the key meteorological data and the power output information; Using a piecewise fitting algorithm, extracting power drop sequence data corresponding to power changes caused by cloud density parameters in the key meteorological data; Taking the geographical location of the photovoltaic power station as a node, converting the data with spatiotemporal correlation in the key meteorological data into a graph structure, and using a graph neural network to adjust the temporal correlation relationship, the spatial correlation relationship, and the power drop sequence data; The data feature set is obtained based on the adjusted time correlation relationship, the spatial correlation relationship and the power drop sequence data.

2. The method according to claim 1, characterized in that The training of the first model based on the data feature set to obtain the second model includes: Processing the data feature set using a cross-validation algorithm, inputting the processed data feature set into the first model, and dynamically adjusting the learning rate of the first model using an adaptive learning rate optimization algorithm to obtain a post-training learning rate; Adjusting the hyperparameters of the first model using a grid search algorithm or a random search algorithm to obtain trained hyperparameters; The second model is obtained by training based on the post-training learning rate and the post-training hyperparameters.

3. The method according to claim 1, characterized in that The method of correcting the second model based on the real-time push data of the photovoltaic power station to obtain a third model includes: Using the second model to predict the photovoltaic power generation power in a short time scale to obtain first prediction data; Using the second model to predict photovoltaic power generation over a long time scale to obtain second prediction data; Calculate a first error sequence based on at least two of the first prediction data and the corresponding at least two of the real-time pushed data; Calculate a second error sequence based on at least two of the second prediction data and the corresponding at least two of the real-time pushed data; Analyzing an error source of each error data in the first error sequence and the second error sequence to obtain a first result; The second model is corrected according to the first result to obtain the third model.

4. The method according to claim 3, characterized in that The correcting the second model according to the first result includes: In response to the first result being a sudden change in at least one of the key meteorological data, dynamically adjusting the weight of the corresponding key meteorological data to perform correction processing on the second model; In response to the first result being that the equipment status of the photovoltaic power station is abnormal, parameters corresponding to the abnormal equipment are dynamically modified or added to perform correction processing on the second model.

5. The method according to claim 1, wherein The performing prediction error analysis on the third model, and in response to the prediction error being greater than a preset threshold, adjusting parameters in the third model to obtain a fourth model, comprises: Using the third model to predict the photovoltaic power generation power in a short time scale to obtain third prediction data; Using the third model to predict photovoltaic power generation over a long time scale to obtain fourth prediction data; Calculating a third error sequence based on at least two of the third prediction data and corresponding at least two historical push data; Calculating a fourth error sequence based on at least two of the fourth prediction data and the corresponding at least two of the historical push data; Analyzing an error source of each error data in the third error sequence and the fourth error sequence to obtain a second result; Parameters in the third model are adjusted based on the second result to obtain the fourth model.

6. A photovoltaic power generation power prediction device, characterized in that: include: an acquisition module, configured to acquire power output information of the photovoltaic power station; Obtain key meteorological data based on large meteorological models; a fusion module configured to fuse the key meteorological data and the power output information to obtain a data feature set; wherein the key meteorological data and the power output information in the data feature set have a mapping relationship; a training module, configured to train the first model based on the data feature set to obtain a second model; a correction module configured to correct the second model based on the real-time push data of the photovoltaic power station to obtain a third model; an adjustment module configured to perform a prediction error analysis on the third model, and in response to a prediction error being greater than a preset threshold, adjust parameters in the third model to obtain a fourth model; a prediction module, configured to predict photovoltaic power generation power using the fourth model; Wherein, the fusion module is further configured to: Using sliding window technology, a time series analysis method is used to dynamically capture the temporal correlation between the key meteorological data and the power output information; Using a data reconstruction algorithm, obtaining a spatial correlation relationship between the key meteorological data and the power output information; Using a piecewise fitting algorithm, extracting power drop sequence data corresponding to power changes caused by cloud density parameters in the key meteorological data; Taking the geographical location of the photovoltaic power station as a node, converting the data with spatiotemporal correlation in the key meteorological data into a graph structure, and using a graph neural network to adjust the temporal correlation relationship, the spatial correlation relationship, and the power drop sequence data; The data feature set is obtained based on the adjusted time correlation relationship, the spatial correlation relationship and the power drop sequence data.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 5 is implemented.

8. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to enable a computer to execute the method according to any one of claims 1 to 5.

9. A computer program product comprising computer program instructions, which, when executed on a computer, cause the computer to perform the method according to any one of claims 1 to 5.

Images