Photovoltaic power station intelligent control method based on large model and deep learning

Abstract

The invention relates to a photovoltaic power station intelligent control method based on a large model and deep learning, which belongs to the field of photovoltaic solar energy intelligent control, and is characterized in that through crawling news policies and 9km * 9km grid weather information, meteorological policy interpretation and influence analysis based on the large model, and missing value prediction and complementation of weather data, the intelligent control of a photovoltaic power station is realized. The accurate acquisition of the weather data is realized; meanwhile, by means of an illumination intensity prediction method for predicting and complementing unknown data through known data and a sunrise and sunset information and sunshine information coding method, weather factors can be better analyzed and controlled, and accurate photovoltaic solar panel power generation control is achieved; besides, according to a weather time sequence prediction method and an intelligent decision-making method based on deep learning, a model is learned in an unsupervised manner by using historical data, and intelligent decision-making is carried out by using a cost function, so that intelligent control of the photovoltaic solar panel is realized; and finally, a result display and alarm method can accurately store and display a predicted value, and early warning is carried out according to an alarm condition.

Description

Technical Field

[0001] This invention relates to an intelligent control method for photovoltaic power plants based on large models and deep learning, belonging to the field of intelligent control of photovoltaic solar energy. Background Technology

[0002] As a representative of new energy sources, photovoltaic solar panels are increasingly important and have broad application prospects. However, due to the influence of various factors such as weather, season, and climate, the power generation efficiency of photovoltaic solar panels exhibits certain fluctuations and instability. To address this issue, real-time monitoring and analysis of multiple factors, including weather and sunlight, are necessary for precise power generation control.

[0003] While some methods for controlling the power generation of photovoltaic solar panels exist in the current technology, these methods still have some problems. For example, due to the low accuracy of weather forecasts, improving their accuracy is particularly important; currently, most meteorological data is based on ground-based observations and cannot provide global meteorological data; at the same time, how to accurately control the power generation of photovoltaic solar panels based on weather data is also an urgent problem to be solved. Summary of the Invention

[0004] Based on the problems described in the background, the problem to be solved by the present invention is to provide an intelligent control method for photovoltaic power plants based on large models and deep learning. The method utilizes deep learning models to process and analyze data from photovoltaic power plants, and to achieve intelligent control and optimization to a certain extent. This solves the problems of low accuracy in weather forecasts, inability to provide global weather data, and inability to perform precise control of photovoltaic solar panel power generation based on weather data.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a smart control method for photovoltaic power plants based on large models and deep learning, comprising the following steps:

[0006] (1) Adaptive real-time stream weather data crawling: Using an adaptive real-time stream crawling strategy, relevant weather information is crawled stably and efficiently.

[0007] (2) Intelligent completion of missing values ​​in weather data based on LSTM network and matrix factorization, using deep learning to achieve intelligent completion of missing values ​​in weather data;

[0008] (3) Based on support vector regression, a regression model adapted to nonlinear relationships is established to achieve accurate prediction of light intensity.

[0009] (4) Use sine and cosine functions to encode sunrise and sunset time information, convert the time information into continuous values, and input them into the model for training and prediction;

[0010] (5) Use a multi-factor modeling cue learning method to generate weather codes with enhanced information. Model weather conditions using multiple factors and generate weather codes with enhanced information through a cue learning method.

[0011] (6) Interpretation and impact analysis of meteorological policies based on large-scale models;

[0012] (7) Use a 9km*9km grid to make time-series predictions of weather information and other relevant data. Through feature engineering and model training, accurate predictions of future weather changes can be achieved.

[0013] (8) Based on time series prediction and optimized allocation, the coordinated control and optimization of distributed photovoltaic power generation are achieved by comprehensively using time series prediction and optimization algorithms.

[0014] (9) Display of wind power prediction results based on time series database: The prediction results are visualized and displayed using time series database.

[0015] (10) Based on the relationship between consumption, usage and power generation, smart power generation is regulated. By analyzing the relationship between consumption, usage and power generation, smart power generation regulation is achieved.

[0016] Step (1) is further divided into the following steps:

[0017] (1.1) Determine the data source for acquiring real-time data and specify the target data, including historical weather data such as past weather patterns, seasonal changes and other meteorological conditions, and real-time weather data such as current temperature, humidity, wind speed, wind direction and cloud cover;

[0018] (1.2) Use incremental crawling to avoid crawling duplicate weather information and only obtain updated or newly added weather information;

[0019] (1.3) Avoid overloading the crawling website. Use the exponential backoff algorithm. If the response time of a certain request is long, the waiting time of the next request will be increased exponentially. Adjust the waiting time. If it is still unsuccessful after the maximum waiting time is reached, record the failed request, error type and error frequency, and switch to the backup data source. After all the backup data sources have finished polling, restart the program and send a notification.

[0020] Step (1.2) is further divided into:

[0021] (1.2.1) For the weather data items to be crawled, record their status information using timestamps;

[0022] (1.2.2) Perform regular checks to determine the weather data items that need to be crawled incrementally;

[0023] (1.2.3) By comparing the status of the recorded weather data with the status of the latest weather data in the target data source, the newly added and updated weather data are determined;

[0024] (1.2.4) Based on the identification results in step (1.2.3), only crawl the newly added and updated weather data;

[0025] (1.2.5) After successfully crawling weather data, update the status of the recorded weather data to reflect the latest status.

[0026] Step (2) is further divided into the following steps:

[0027] (2.1) Collect historical weather time series data and detect the location of missing data;

[0028] (2.2) For missing data in continuous time periods, an LSTM network model is used for sequence prediction;

[0029] (2.3) For discrete missing points, the weather dataset is decomposed into low-rank using a matrix factorization mechanism;

[0030] (2.4) Repeat steps (2.2) and (2.3) until all missing data is filled in;

[0031] (2.5) Continuously collect new data to conduct incremental training of the model and evaluate and optimize the prediction effect.

[0032] Step (3) is further divided into the following steps:

[0033] (3.1) Collect historical weather data that may affect sunlight at different times, including cloud cover, rainfall, and temperature;

[0034] (3.2) Perform data cleaning and imputation operations, handle missing values ​​and outliers, and standardize the data;

[0035] (3.3) Training using the SVR model;

[0036] (3.4) Use cross-validation to adjust the hyperparameters of SVR to improve model performance and test the performance on the partitioned test set, and calculate the root mean square error (RMSE), R2 score or other appropriate regression performance metrics.

[0037] Step (4) is further divided into the following steps:

[0038] (4.1) Obtain the longitude and Earth tilt angle of the corresponding location to calculate the corresponding sunrise and sunset times, and sunrise and sunset occur once a day with a cycle of 24 hours;

[0039] (4.2) Use sine and cosine functions to calculate sunrise and sunset codes.

[0040] Step (5) is further divided into the following steps:

[0041] (5.1) Collect large-scale weather data related to weather conditions, temperature, humidity, wind speed, etc., and preprocess the data;

[0042] (5.2) Organize the data into a prompt template and mask the weather information.

[0043] (5.3) Construct a prompt learning model. In order to associate various factors such as temperature, average temperature, humidity, wind speed and weather information, use the natural language model BERT with self-attention mechanism to train and capture the relationship between features, and finally output the prediction of weather information.

[0044] (5.4) Input factors such as temperature, average temperature, humidity, wind speed and weather information into the trained cue learning model to generate weather codes with sufficient information.

[0045] Step (6) is further divided into the following steps:

[0046] (6.1) Establish a text collection using the artificial intervention meteorological policies or other relevant news obtained in the previous steps;

[0047] (6.2) Use a large model to interpret the content of the text set to obtain the purpose set, content set, and impact set of artificial meteorological intervention;

[0048] (6.3) Based on the influence set, establish the function of the impact of artificial meteorological intervention on weather;

[0049] (6.4) Integrating functions into meteorological forecasting models can improve forecast accuracy;

[0050] (6.5) The large model has more than 100 million parameters and a training dataset of more than 10GB. It is based on natural language processing technology and has the ability to understand, analyze and predict text.

[0051] Step (7) is further divided into the following steps:

[0052] (7.1) Collect weather information and other relevant data such as sunrise and sunset coding data, cloudy / sunny coding data and news and policy interpretation data for a 9km*9km grid;

[0053] (7.2) Preprocess the weather information and other relevant data, including data cleaning, formatting and standardization, to obtain the preprocessed dataset;

[0054] (7.3) Train a time series forecasting model based on the dataset. This model can predict future weather.

[0055] (7.4) Evaluate the prediction results and calculate the accuracy and error;

[0056] (7.5) Iteratively optimize the model based on the evaluation results.

[0057] Step (8) is further divided into the following steps:

[0058] (8.1) Establish a recurrent neural network model, input historical power generation and load time series data, and use it to predict power generation and load within a certain period of time in the future after training;

[0059] (8.2) Predict the recent electricity load based on model prediction results, daily periodicity, holidays and other factors;

[0060] (8.3) Set the optimization objective function and solve it using a greedy algorithm or a multi-objective optimization algorithm such as dynamic programming. Distribute the predicted power generation to different power grid areas to meet both load demand and power generation cost.

[0061] (8.4) Based on the power distribution results, control commands are sent to the intelligent devices in each area using digital communication technology to achieve precise PID control of photovoltaic power generation.

[0062] Step (9) is further divided into the following steps:

[0063] (9.1) Real-time display of predicted values, including:

[0064] (9.1.1) Query the database to obtain the most recent predicted power data and perform exponential weighted smoothing;

[0065] (9.1.2) Design an interface to dynamically display the smoothed predicted power curve and provide a time window selection function;

[0066] (9.2) Predicted value storage, including:

[0067] (9.2.1) Use a time series database to store time series data;

[0068] (9.2.2) Provides SQL interface for querying;

[0069] (9.2.3) Regularly reduce costs;

[0070] (9.3) Viewing historical data, including:

[0071] (9.3.1) Selecting any historical time period will load the historical predicted values ​​for that time period;

[0072] (9.3.2) Calculate the accuracy indicators RMSE and MAPE;

[0073] (9.3.3) Present the prediction results in the form of statistical charts.

[0074] Step (10) is further divided into the following steps:

[0075] (10.1) Read the newly added storage capacity, usage, and power generation from the database in real time;

[0076] (10.2) Compare the relationship between each group of consumption, usage and power generation. If the sum of consumption and usage is less than power generation, an alarm message is generated, displayed in the front-end alarm bar, and the alarm time and alarm message are saved to the database in real time.

[0077] The beneficial effects of this invention are:

[0078] 1. By crawling news policies and 9km*9km grid weather information, based on large-scale model interpretation and impact analysis of meteorological policies, as well as prediction and completion of missing values ​​in weather data, accurate acquisition of weather data was achieved.

[0079] 2. The light intensity prediction method that uses known data to predict and complete unknown data, as well as the encoding method for sunrise and sunset information and weather information, can better analyze and control weather factors and achieve precise control of photovoltaic solar panel power generation.

[0080] 3. A weather time series forecasting method and intelligent decision-making method based on deep learning are used to learn the model in unsupervised manner using historical data and make intelligent decisions using a cost function to achieve intelligent control of photovoltaic solar panels.

[0081] 4. It provides result display and alarm methods, which can accurately store and display predicted values ​​and issue warnings based on alarm conditions.

[0082] 5. This invention can improve the energy output of photovoltaic solar energy systems, optimize energy consumption, detect faults and anomalies in advance, and save energy and reduce emissions, thereby solving practical problems in the operation of photovoltaic solar energy systems and improving the performance and reliability of the systems. Attached Figure Description

[0083] Figure 1 This is a flowchart of the steps of the present invention; Detailed Implementation

[0084] The embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0085] Example 1

[0086] like Figure 1As shown, the specific method of adaptive real-time weather data crawling in step (1) of the present invention is as follows:

[0087] Adaptive real-time streaming weather data crawling: This feature utilizes an adaptive real-time streaming crawling strategy to stably and efficiently crawl relevant weather information.

[0088] (1.1) Determine the data source for acquiring real-time data and specify the target data, including historical weather data such as past weather patterns, seasonal changes and other meteorological conditions, and real-time weather data such as current temperature, humidity, wind speed, wind direction and cloud cover;

[0089] (1.2) Use incremental crawling to avoid crawling duplicate weather information, and only obtain updated or newly added weather information, including:

[0090] (1.2.1) For the weather data items to be crawled, record their status information using timestamps;

[0091] (1.2.2) Perform regular checks to determine the weather data items that need to be crawled incrementally;

[0092] (1.2.3) By comparing the status of the recorded weather data with the status of the latest weather data in the target data source, the newly added and updated weather data are determined;

[0093] (1.2.4) Based on the identification results in step (1.2.3), only crawl the newly added and updated weather data;

[0094] (1.2.5) After successfully crawling weather data, update the status of the recorded weather data to reflect the latest status;

[0095] (1.3) To avoid overloading the crawling website, use the exponential backoff algorithm. If the response time of a certain request is long, the waiting time of the next request will be increased exponentially. Adjust the waiting time by setting the initial waiting time T0, the multiplication factor K, and the maximum waiting time T. max The other times are T1, T2, ..., T n The calculation formula is as follows:

[0096] T1 = T0

[0097] T2 = T1 * K ...

[0099] T n =T n-1 *K

[0100] When the maximum waiting time T is reached maxIf the process still fails, record the failed requests, error types, and error frequencies, and switch to a backup data source. Once all backup data sources have finished polling, restart the program and send a notification.

[0101] like Figure 1 As shown, the specific method of intelligent completion of missing values ​​in weather data based on LSTM network and matrix factorization in step (2) of the present invention is as follows:

[0102] Intelligent completion of missing values ​​in weather data based on LSTM network and matrix factorization: This method uses deep learning to intelligently complete missing values ​​in weather data.

[0103] (2.1) Collect historical weather time series data {X1,X2,...,X...} n}, detect the set M of missing data locations;

[0104] (2.2) For missing data in a continuous time period, let the missing time period be {X}. m+1 ,...,X m+k Sequence prediction is performed using an LSTM network model.

[0105] LSTM({X1,...,X m},{X m+k+1 ,...,X n})=>{X′ m+1 ,...,X′ m+k}

[0106] (2.3) For discrete missing points, let the missing point be {X}. j Using matrix factorization, the weather dataset is decomposed into a low-rank representation:

[0107] X≈U·V

[0108] Using matrix factorization, the weather dataset is decomposed into a low-rank representation:

[0109] X≈U·V j

[0110] (2.4) Repeat steps (2.2) and (2.3) until all missing data are filled in, resulting in the complete dataset {X1,X2,...,X...}. n};

[0111] (2.5) Continuously collect new data to conduct incremental training of the model and evaluate and optimize the prediction effect.

[0112] like Figure 1 As shown, the specific method of intelligent completion of missing weather data based on LSTM network and matrix factorization in step (3) of the present invention is as follows:

[0113] Light intensity prediction based on support vector regression is achieved by establishing a regression model that adapts to nonlinear relationships, thus enabling accurate prediction of light intensity.

[0114] (3.1) Collect historical weather data that may affect sunlight at different times, including cloud cover, rainfall and temperature. Let cloud cover be X1, rainfall be X2 and temperature be X3.

[0115] (3.2) Perform data cleaning and imputation operations, handle missing values ​​and outliers, and standardize the data;

[0116] (3.3) The SVR model is trained to satisfy the following formula:

[0117]

[0118] Y represents the predicted value of light intensity, α and α * It is the Lagrange multiplier corresponding to the support vectors (a portion of the training samples), K(x) i (x) is a kernel function used to compute the kernel value of the training sample x. i The similarity between the predicted sample x and the kernel function. The kernel function can be a linear kernel, a polynomial kernel, a radial basis function (RBF), etc., used to handle nonlinear relationships. b is the intercept term of the model.

[0119] (3.4) Use cross-validation to adjust the hyperparameters of SVR to improve model performance and test the performance on the partitioned test set, and calculate the root mean square error (RMSE), R2 score or other appropriate regression performance metrics.

[0120] like Figure 1 As shown, the specific method of intelligent completion of missing weather data based on LSTM network and matrix factorization in step (4) of the present invention is as follows:

[0121] The sunrise and sunset time information is encoded using sine and cosine functions, converting the time information into continuous numerical values, which are then input into the model for training and prediction.

[0122] (4.1) Given the longitude x and Earth's tilt angle y of the corresponding location, the sunrise time is:

[0123]

[0124] Sunset time is:

[0125]

[0126] In addition, sunrise and sunset occur once a day, with a period T of 24 hours;

[0127] (4.2) Calculate the sunrise and sunset codes TE(m) using sine and cosine functions as follows:

[0128]

[0129] like Figure 1 As shown, the specific method of intelligent completion of missing weather data based on LSTM network and matrix factorization in step (5) of the present invention is as follows:

[0130] A multi-factor modeling-based cue learning method is used to generate weather codes with enhanced information. This method models weather conditions using multiple factors and generates weather codes with enhanced information through cue learning.

[0131] (5.1) Collect large-scale weather data related to weather conditions such as sunshine / cloudy / sunset, temperature, humidity, and wind speed, and preprocess the data. Let the temperature be T, and the average temperature be... Humidity is H, weather condition is S, collect relevant data. Perform data preprocessing P():

[0132]

[0133] (5.2) Organize dataset D into a prompt template and mask the weather information. The specific formula is as follows:

[0134] D′=mask(D,S)

[0135] (5.3) Construct a prompt learning model. In order to associate various factors such as temperature, average temperature, humidity, wind speed and weather information, use the natural language model BERT with self-attention mechanism to train and capture the relationship between features, and finally output the prediction of weather information.

[0136] Let the predicted weather state be

[0137]

[0138] Will Encoding as a probability distribution Input factors and actual state generate codes:

[0139]

[0140] (5.4) Input factors such as temperature, average temperature, humidity, wind speed and weather information into the trained cue learning model to generate weather codes with sufficient information.

[0141] like Figure 1 As shown, the specific method of intelligent completion of missing weather data based on LSTM network and matrix factorization in step (6) of the present invention is as follows:

[0142] Interpretation and impact analysis of meteorological policies based on large-scale models;

[0143] (6.1) Using the artificial intervention meteorological policies or other relevant news obtained in the previous steps, let it be a text set T = {t1, t2, ..., t3}. n};

[0144] (6.2) Using a large model f to interpret the text set T, we obtain the objective set G, content set C, and impact set I of artificial meteorological intervention:

[0145] f(T)=>{G,C,I}

[0146] (6.3) Based on the influence set I, establish the weather influence function h of artificial meteorological intervention:

[0147] h(I)=>E

[0148] Where E represents the set of effects of human intervention in meteorology on weather;

[0149] (6.4) Integrating the function h into the meteorological forecasting model m improves forecast accuracy:

[0150] m′(X,h(I))=>Y

[0151] Where X is the original input and Y is the new prediction result;

[0152] (6.5) The large model f has more than 100 million parameters, the training dataset is more than 10GB, and it is based on natural language processing technology, with the ability to understand, analyze and predict text.

[0153] like Figure 1 As shown, the specific method of intelligent completion of missing weather data based on LSTM network and matrix factorization in step (7) of the present invention is as follows:

[0154] Using a 9km*9km grid, time-series forecasts of weather information and other relevant data are performed. Through feature engineering and model training, accurate predictions of future weather changes are achieved.

[0155] (7.1) Collect weather information for a 9km*9km grid. weather Other relevant data D other , where: D other Includes sunrise and sunset encoded data D sun Sunny / Yellow Coding Data D cloud And news policy interpretation data D policy ;

[0156] (7.2) Regarding D weather and D otherPreprocessing, including cleaning the data (C), formatting (F), and standardizing (S), yields the preprocessed dataset D′.

[0157] D′=C(F(S(D weather ∪D other )))

[0158] (7.3) Train a time-series forecasting model M based on dataset D′. This model can predict future weather Y′:

[0159] M = Train(D′)

[0160] Y' = Predict(M)

[0161] (7.4) Evaluate the prediction result Y′ and calculate the accuracy P and error E:

[0162] P = Evaluate(Y′)

[0163] E = Error(Y′)

[0164] (7.5) Iteratively optimize model M based on the evaluation results:

[0165] M′=O(M,P,E)

[0166] like Figure 1 As shown, the specific method of intelligent completion of missing weather data based on LSTM network and matrix factorization in step (8) of the present invention is as follows:

[0167] Based on time series forecasting and optimized allocation, the coordinated control and optimization of distributed photovoltaic power generation are achieved by comprehensively utilizing time series forecasting and optimization algorithms.

[0168] (8.1) Establish a recurrent neural network model (such as LSTM or GRU), input historical power generation and load time series data, and after training, use it to predict power generation P′ and load L′ within a certain period of the future:

[0169] P′,L′=f(P t-1 ,L t-1 ,...,ω)

[0170] (8.2) Based on the model prediction results, daily periodicity, holidays and other factors, predict the recent electricity load L”;

[0171] (8.3) Set the optimization objective function and solve it using a greedy algorithm or a multi-objective optimization algorithm such as dynamic programming. Allocate the predicted power generation P” to different power grid areas to meet both load demand and power generation costs.

[0172]

[0173] (8.4) Based on the power distribution results, control commands are sent to intelligent devices in each area using digital communication technology to achieve precise PID control of photovoltaic power generation:

[0174]

[0175] like Figure 1 As shown, the specific method of intelligent completion of missing weather data based on LSTM network and matrix factorization in step (9) of the present invention is as follows:

[0176] The wind power prediction results based on the time series database are displayed by using the time series database to visualize and display the prediction results.

[0177] (9.1) Real-time display of predicted values:

[0178] (9.1.1) Query the database to obtain the most recent predicted power data P, and perform exponential weighted smoothing on P:

[0179] P t ′=αP t +(1-α)P′ t-1

[0180] (9.1.2) Design an interface to dynamically display the smoothed predicted power curve P′ and provide a time window selection function;

[0181] (9.2) Predicted value storage:

[0182] (9.2.1) Use a time series database to store time series data such as P and t;

[0183] (9.2.2) Provides SQL interface for querying;

[0184] (9.2.3) Regularly reduce costs;

[0185] (9.3) Viewing historical data:

[0186] (9.3.1) Selecting any historical time period will load the historical predicted value P” for that time period;

[0187] (9.3.2) Calculation accuracy indicators RMSE and MAPE:

[0188]

[0189] (9.3.3) Present the prediction results in the form of statistical charts.

[0190] like Figure 1As shown, the specific method of intelligent completion of missing weather data based on LSTM network and matrix factorization in step (9) of the present invention is as follows:

[0191] Smart power generation is regulated based on the relationship between consumption, usage, and power generation. By analyzing the relationship between consumption, usage, and power generation, smart power generation regulation is achieved.

[0192] (10.1) Read the newly added storage capacity, usage, and power generation from the database in real time.

[0193] (10.2) Let the absorption capacity be C, the usage capacity be U, and the power generation capacity be G. Compare the relationship between C, U, and G for each group. If C + U

Claims

1. A smart control method for photovoltaic power plants based on large models and deep learning, characterized in that, Includes the following steps: (1) Adaptive real-time stream weather data crawling: Using an adaptive real-time stream crawling strategy, relevant weather information is crawled stably and efficiently. (2) Intelligent completion of missing values ​​in weather data based on LSTM network and matrix factorization, using deep learning to achieve intelligent completion of missing values ​​in weather data; (3) Based on support vector regression, a regression model adapted to nonlinear relationships is established to achieve accurate prediction of light intensity. (4) Use sine and cosine functions to encode sunrise and sunset time information, convert the time information into continuous values, and input them into the model for training and prediction; (5) Use a multi-factor modeling cue learning method to generate weather codes with enhanced information. Model weather conditions using multiple factors and generate weather codes with enhanced information through a cue learning method. (6) Interpretation and impact analysis of meteorological policies based on large-scale models; (7) Use a 9km*9km grid to make time-series predictions of weather information and other related data. Through feature engineering and model training, accurate predictions of future weather changes can be achieved. (8) Based on time series prediction and optimized allocation, the coordinated control and optimization of distributed photovoltaic power generation are achieved by comprehensively using time series prediction and optimization algorithms. (9) Display of wind power prediction results based on time series database: The prediction results are visualized and displayed using time series database. (10) Smart power generation is regulated based on the relationship between consumption, usage and power generation. By analyzing the relationship between consumption, usage and power generation, smart power generation regulation is achieved. Step (1) is further divided into the following steps: (1.1) Determine the data source for acquiring real-time data and specify the target data, including historical weather data such as past weather patterns, seasonal changes and other meteorological conditions, and real-time weather data such as current temperature, humidity, wind speed, wind direction and cloud cover; (1.2) Use incremental crawling to avoid crawling duplicate weather information and only obtain updated or newly added weather information; (1.3) Avoid overloading the crawling website. Use the exponential backoff algorithm. If the response time of a certain request is long, the waiting time of the next request will be increased exponentially. Adjust the waiting time. If it is still unsuccessful after the maximum waiting time is reached, record the failed request, error type and error frequency, and switch to the backup data source. After all the backup data sources have finished polling, restart the program and send a notification. Step (1.2) is further divided into: (1.2.1) For the weather data items to be crawled, record their status information using timestamps; (1.2.2) Perform regular checks to determine the weather data items that need to be crawled incrementally; (1.2.3) By comparing the status of the recorded weather data with the status of the latest weather data in the target data source, the newly added and updated weather data are determined; (1.2.4) Based on the identification results in step (1.2.3), only crawl the newly added and updated weather data; (1.2.5) After successfully crawling weather data, update the status of the recorded weather data to reflect the latest status.

2. The intelligent control method for photovoltaic power plants based on large models and deep learning according to claim 1, characterized in that, Step (2) is further divided into the following steps: (2.1) Collect historical weather time series data and detect the location of missing data; (2.2) For missing data in continuous time periods, an LSTM network model is used for sequence prediction; (2.3) For discrete missing points, the weather dataset is decomposed into low-rank using a matrix factorization mechanism; (2.4) Repeat steps (2.2) and (2.3) until all missing data is filled in; (2.5) Continuously collect new data to conduct incremental training of the model and evaluate and optimize the prediction effect.

3. The intelligent control method for photovoltaic power plants based on large models and deep learning according to claim 1, characterized in that, Step (3) is further divided into the following steps: (3.1) Collect historical weather data that may affect sunlight at different times, including cloud cover, rainfall, and temperature; (3.2) Perform data cleaning and imputation operations, handle missing values ​​and outliers, and standardize the data; (3.3) Training using the SVR model; (3.4) Use cross-validation to adjust the hyperparameters of SVR to improve model performance and test the performance on the partitioned test set, and calculate the root mean square error (RMSE), R2 score or other appropriate regression performance metrics.

4. The intelligent control method for photovoltaic power plants based on large models and deep learning according to claim 1, characterized in that, Step (4) is further divided into the following steps: (4.1) Obtain the longitude and Earth tilt angle of the corresponding location to calculate the corresponding sunrise and sunset times, and sunrise and sunset occur once a day with a cycle of 24 hours; (4.2) Use sine and cosine functions to calculate sunrise and sunset codes.

5. The intelligent control method for photovoltaic power plants based on large models and deep learning according to claim 1, characterized in that, Step (5) is further divided into the following steps: (5.1) Collect large-scale weather data related to weather conditions, temperature, humidity, wind speed, etc., and preprocess the data; (5.2) Organize the data into a prompt template and mask the weather information. (5.3) Construct a prompt learning model. In order to associate various factors such as temperature, average temperature, humidity, wind speed and weather information, use the natural language model BERT with self-attention mechanism to train and capture the relationship between features, and finally output the prediction of weather information. (5.4) Input factors such as temperature, average temperature, humidity, wind speed and weather information into the trained cue learning model to generate weather codes with sufficient information.

6. The intelligent control method for photovoltaic power plants based on large models and deep learning according to claim 1, characterized in that, Step (6) is further divided into the following steps: (6.1) Establish a text collection using information obtained in previous steps regarding human intervention weather policies or other relevant news; (6.2) Use a large model to interpret the content of the text set to obtain the purpose set, content set, and impact set of artificial meteorological intervention; (6.3) Based on the influence set, establish the function of the impact of artificial meteorological intervention on weather; (6.4) Integrating functions into meteorological forecasting models can improve forecast accuracy; (6.5) The large model has more than 100 million parameters and a training dataset of more than 10GB. It is based on natural language processing technology and has the ability to understand, analyze and predict text.

7. The intelligent control method for photovoltaic power plants based on large models and deep learning according to claim 1, characterized in that, Step (7) is further divided into the following steps: (7.1) Collect weather information and other relevant data such as sunrise and sunset coding data, cloudy / sunny coding data and news and policy interpretation data for a 9km*9km grid; (7.2) Preprocess the weather information and other relevant data, including cleaning, formatting and standardizing the data, to obtain the preprocessed dataset; (7.3) Train a time series forecasting model based on the dataset. This model can predict future weather. (7.4) Evaluate the prediction results and calculate the accuracy and error; (7.5) Iteratively optimize the model based on the evaluation results.

8. The intelligent control method for photovoltaic power plants based on large models and deep learning according to claim 1, characterized in that, Step (8) is further divided into the following steps: (8.1) Establish a recurrent neural network model, input historical power generation and load time series data, and use it to predict power generation and load within a certain period of time in the future after training; (8.2) Predict the recent electricity load based on model prediction results, daily periodicity, holidays and other factors; (8.3) Set the optimization objective function and solve it using a greedy algorithm or a multi-objective optimization algorithm such as dynamic programming. Distribute the predicted power generation to different power grid areas to meet both load demand and power generation cost. (8.4) Based on the power distribution results, control commands are sent to the intelligent devices in each area using digital communication technology to achieve precise PID control of photovoltaic power generation.

9. The intelligent control method for photovoltaic power plants based on large models and deep learning according to claim 1, characterized in that, Step (9) is further divided into the following steps: (9.1) Real-time display of predicted values, including: (9.1.1) Query the database to obtain the most recent predicted power data and perform exponential weighted smoothing; (9.1.2) Design an interface to dynamically display the smoothed predicted power curve and provide a time window selection function; (9.2) Predicted value storage, including: (9.2.1) Use a time series database to store time series data; (9.2.2) Provides SQL interface for querying; (9.2.3) Regularly reduce costs; (9.3) Viewing historical data, including: (9.3.1) Selecting any historical time period will load the historical predicted values ​​for that time period; (9.3.2) Calculate the accuracy indicators RMSE and MAPE; (9.3.3) Present the prediction results in the form of statistical charts.

10. The intelligent control method for photovoltaic power plants based on large models and deep learning according to claim 1, characterized in that, Step (10) is further divided into the following steps: (10.1) Read the newly added storage capacity, usage, and power generation from the database in real time; (10.2) Compare the relationship between each group of consumption, usage and power generation. If the sum of consumption and usage is less than power generation, an alarm message is generated, displayed in the front-end alarm bar, and the alarm time and alarm message are saved to the database in real time.

Images