Cfsv2 ground solar radiation revision method and system based on quantile regression

Abstract

A CFSv2 ground solar radiation revision method and system based on quantile regression, the revision method comprises the following steps: data acquisition and pretreatment step, selecting typical sites, collecting hourly solar radiation observation data for quality control, and carrying out data cleaning; the space-time matching step includes time alignment, space alignment; the quantile regression revision model establishment and training step includes model establishment, model training and parameter optimization; the revision process step, through the quantile regression model to the mode data, to obtain the corrected irradiance prediction value; the model effect evaluation step, using the root mean square error to compare and analyze the mode data before and after revision, and evaluate the revision effect. The present application can revise the ground solar radiation data, reduce the daily radiation prediction error, significantly improve the accuracy and reliability of the long-term trend prediction of solar energy resources, especially improve the prediction accuracy under extreme weather conditions.

Description

TECHNICAL FIELD

[0001] The application relates to a method for correcting ground solar radiation data, in particular to a CFSv2 ground solar radiation correction method based on quantile regression. BACKGROUND

[0002] Solar radiation is one of the key meteorological factors affecting photovoltaic power generation performance. Currently, numerical models such as CFSv2 and reanalysis data are widely used for long-term prediction of solar radiation. However, the CFSv2 model still has large errors at the regional scale and daily time resolution, especially under cloudy weather and complex climate conditions, the prediction error is particularly obvious. Traditional correction methods, such as linear regression, cannot effectively capture the nonlinear error characteristics that change with the lead, resulting in inaccurate photovoltaic power generation prediction and affecting the efficiency and stability of power dispatching.

[0003] Quantile regression is a method that can handle nonlinear errors and capture changes in error distribution.

[0004] Currently, numerical models such as CFSv2 have large deviations in daily solar radiation prediction, and the deviation changes nonlinearly with the lead. Existing linear correction methods cannot capture this nonlinear change, resulting in low accuracy of solar radiation prediction, which in turn affects photovoltaic power generation and power dispatching. Especially in extreme weather conditions, traditional methods cannot provide accurate radiation prediction.

[0005] Related patent document: CN118426079A discloses a solar radiation correction method and system based on multi-step deviation iteration and random forest, which includes obtaining predicted solar radiation and measured solar radiation, calculating normalized data, and obtaining other predicted meteorological elements and hourly values, using a time window rolling modeling to construct a training data set; taking the predicted solar radiation, hourly values and other predicted meteorological elements as input, and taking the normalized data of the measured solar radiation as output, training the first random forest regression model; performing weather typing on the normalized data set of the predicted solar radiation, taking the weather typing as input, and taking the deviation of the first predicted normalized solar radiation and the measured solar radiation normalized data as output, training the second random forest regression model; obtaining the data of the day to be corrected and the corresponding weather typing, obtaining the first predicted normalized solar radiation and the second predicted normalized solar radiation deviation, and calculating the corrected solar radiation.

[0006] The above technology does not completely solve the problem of correcting ground solar radiation data, reducing daily radiation prediction error, and significantly improving the accuracy and reliability of long-term trend prediction of solar energy resources in the prior art. SUMMARY

[0007] The purpose of the present application is to provide a quantile regression-based CFSv2 ground solar radiation correction method, which can correct ground solar radiation data, reduce daily radiation prediction error, and significantly improve the accuracy and reliability of long-term trend prediction of solar energy resources, especially the prediction accuracy under extreme weather conditions.

[0008] To this end, another purpose of the present application is to provide a quantile regression-based CFSv2 ground solar radiation correction system.

[0009] To solve the above technical problems, the technical solution adopted by the present application is as follows:

[0010] A quantile regression-based CFSv2 ground solar radiation correction method, the technical solution of which comprises the following steps:

[0011] S1: data acquisition and preprocessing, which comprises:

[0012] Observation data quality control: select typical sites (such as 6 sites in a province), collect hourly solar radiation observation data for quality control, and carry out data cleaning.

[0013] S2: time and space matching, which comprises:

[0014] Time alignment: align the observation data and CFSv2 model data by date to ensure that they are completely consistent in time;

[0015] Spatial alignment: spatially interpolate the model data, and interpolate the grid data of the CFSv2 model to the latitude and longitude coordinates of the actual observation site.

[0016] S3: quantile regression correction model establishment and training, which comprises:

[0017] Model establishment: use quantile regression (Quantile Regression) method to establish the nonlinear regression relationship between observation and model, and the regression equation is:

[0018] ,

[0019] Wherein: represents the observed irradiance, represents the predicted irradiance of the model, lead represents the lead time, is the square term of the lead time, , 、 、 is a regression coefficient, and ε is a residual error;

[0020] Model training and parameter tuning: using training data (e.g. observation data and model data from May to October 2022) to train the regression model; then adjusting and optimizing the parameters.

[0021] S4: Correction process, using quantile regression model to correct the model data to obtain the corrected irradiance prediction value.

[0022] S5: Model effect evaluation, using root mean square error (RMSE) to compare and analyze the model data before and after correction, and evaluate the correction effect. The evaluation results show that through quantile regression correction, the RMSE value of the corrected model is reduced by an average of 20%-35%, the bias (Bias) tends to zero, and under extreme weather conditions (cloudy and complex weather conditions), the prediction error is significantly reduced.

[0023] In the above technical solution, the preferred technical solution can be that in step S1, the quality control process includes:

[0024] Eliminate irradiance outliers (i.e. eliminate physically unreasonable irradiance values), and the irradiance outliers are irradiance values greater than 1500 W / m²;

[0025] Handle missing values by interpolation or before and after time value filling to ensure data integrity and ensure data quality;

[0026] After the quality control is completed, the observation data is grouped by date, the daily average irradiance is calculated, the missing values are filled with -999, and the processed data is stored in the format of date x site matrix, and finally a complete data matrix of date x site is formed.

[0027] In the above technical solution, the preferred technical solution can also be that step S1 further includes model data preprocessing, and the method of model data preprocessing is to download shortwave radiation data from CFSv2 and merge the data into daily data to obtain a model data set. Each model data set will be converted to NetCDF format to ensure that the model data is aligned with the observation data for comparison and correction.

[0028] For the mode data preprocessing step, the mode file is read, and the GRIB2 files of the CFSv2 mode are read in batches through an automated script. The shortwave radiation data contained in the file is first converted to NetCDF format through the built-in decoding tool, and the data of each time is extracted. Every 4 time data will be merged into daily data to ensure the time sequence of the data. Specifically, 6-hourly shortwave radiation data is downloaded from CFSv2, and every 4 time data is merged into daily data. Each mode data set will be converted to NetCDF format and retain necessary fields such as initial forecast time, lead, and valid time, ensuring that the mode data can be aligned with the observation data for comparison and correction.

[0029] In the above technical solution, preferably, in step S3, in the step of model training and parameter tuning, the median τ=0.5 is used as the target quantile, and the quantile regression algorithm is used to estimate the parameters of the model, and τ is the quantile value.

[0030] In step S3, the method of parameter adjustment and optimization is to adjust and optimize the parameters of the regression model by using grid search (Grid Search) and cross validation (Cross validation) methods. Grid search is used to traverse different lead ranges and quantile values (τ) to find the best regression parameters. Cross validation is used to evaluate the generalization ability of the model to prevent overfitting.

[0031] Regularization and model selection: In order to improve the robustness of the model, L2 regularization method is used to optimize the regression model, reduce the influence of high order terms on the model, avoid overfitting caused by high complexity of the model, and select the optimal model after regularization.

[0032] Parameter optimization and model convergence: The model is optimized by maximizing the log-likelihood function and minimizing the error residual. In the optimization process, the stochastic gradient descent method (Stochastic Gradient Descent, SGD) and Adam optimizer are used to ensure that the model parameters can effectively converge. Further application of advanced techniques such as Bayesian optimization to adjust the model hyperparameters ensures the optimal configuration of the model.

[0033] This step performs quantile regression analysis on the observation data and the mode data, selects the median (τ=0.5) as the target quantile, and estimates the regression coefficient (β) 、 、 、 ). The Quantile Regression algorithm is used to fit the model by minimizing the error, capturing the non-linear characteristics of the pattern data. The error correction in this step is to revise the pattern forecast value using the Quantile Regression model to generate the corrected solar radiation forecast value. The lead time and its square term in the regression equation ensure that the non-linear error caused by the change of lead time can be handled.

[0034] In the above technical solution, the preferred technical solution can also be that in step S5, the step of model effect evaluation further includes revised effect visualization, by drawing the root mean square error (RMSE) and the change curve of lead time (lead), comparing the effects before and after revision, and further verifying the effectiveness of the Quantile Regression method.

[0035] In the above technical solution, the preferred technical solution can also be that the CFSv2 ground solar radiation revision method based on Quantile Regression further includes:

[0036] S6: Medium and long term solar energy resource trend prediction, the prediction result is applied to the trend prediction product, and the revised irradiance data is input into the medium and long term wind energy solar resource trend prediction product, which can significantly improve the prediction accuracy. Especially in the field of photovoltaic power generation, accurate solar radiation prediction can effectively improve the accuracy of new energy power generation and help power dispatching decision.

[0037] A CFSv2 ground solar radiation revision system based on Quantile Regression, the revision system uses the above-mentioned CFSv2 ground solar radiation revision method based on Quantile Regression, and the revision system comprises:

[0038] A data acquisition and preprocessing module for observation data quality control and pattern data preprocessing;

[0039] A space-time matching module for time alignment and space alignment;

[0040] A Quantile Regression revision model establishment and training module for model establishment, model training and parameter optimization;

[0041] A revision process module for revising the pattern data by the Quantile Regression model to obtain the corrected irradiance forecast value;

[0042] A model effect evaluation module for model effect evaluation;

[0043] A medium and long term solar energy resource trend prediction module for trend prediction product, and the revised irradiance data is input into the medium and long term wind energy solar resource trend prediction product.

[0044] System integration and application: The correction method of the present application has been integrated into the "wind and solar energy resource trend prediction product production management system V1.0" as part of the system. The system can real-time correct the solar radiation data from CFSv2 and other models, output more accurate solar energy resource medium and long term trend prediction results, and support further energy management and power grid scheduling.

[0045] Application scenario: The correction method of the present application can be widely applied in new energy resource evaluation systems such as photovoltaic power stations and wind farms, combined with power dispatching system, to provide more accurate photovoltaic power generation prediction data, thereby helping to improve the efficiency and stability of power grid scheduling.

[0046] The present application provides a CFSv2 ground solar radiation correction method based on quantile regression, belonging to the technical field of energy meteorology and climate prediction, especially for the field of solar energy resource evaluation, specifically relates to a method for correcting CFSv2 ground solar radiation data based on quantile regression method. The method aims to improve the accuracy of medium and long term trend prediction of solar energy resources, solve the deviation problem of existing prediction system under complex weather conditions, and especially eliminate the inconvenience of application in photovoltaic power generation prediction. The solar radiation correction method based on quantile regression of the present application aims to effectively correct the solar radiation forecast by introducing the quantile regression method. Quantile regression is a method that can handle nonlinear errors and capture error distribution changes. By introducing the lead time and its square term, the quantile regression method can better adapt to the nonlinear characteristics of solar radiation forecast and provide more accurate correction results. The present application uses this method to correct the solar radiation data of CFSv2 model, combines observation data and model data, and uses quantile regression to correct the ground solar radiation data of CFSv2, reduces the daily radiation forecast error, and significantly improves the accuracy and reliability of medium and long term trend prediction of solar energy resources, especially improves the prediction accuracy under extreme weather conditions, solves the problem that the ground solar radiation data cannot be corrected to reduce the daily radiation forecast error and significantly improve the accuracy and reliability of medium and long term trend prediction of solar energy resources. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The flow chart of the first embodiment of the CFSv2 ground solar radiation correction method based on quantile regression of the present application.

[0048] Figure 2 The structure diagram of the second embodiment of the CFSv2 ground solar radiation correction system based on quantile regression of the present application.

[0049] Figure 3The first embodiment of the present invention is a visualization result graph, which is a curve showing the changes in root mean square error (RMSE) and lead time, displaying the RMSE values ​​before and after correction; the curve is obtained based on the changes in the 7-day average forecast and observed RMSE values ​​obtained by station A in a certain province in May 2022.

[0050] Figure 4 The first embodiment of the present invention is a visualization result graph, which is a curve showing the changes in root mean square error (RMSE) and lead time, displaying the RMSE values ​​before and after correction; the curve is obtained based on the changes in the 7-day average forecast and observed RMSE values ​​obtained by station A in a certain province in June 2022.

[0051] Figure 5 The first embodiment of the present invention is a visualization result graph, which is a curve showing the changes in root mean square error (RMSE) and lead time, displaying the RMSE values ​​before and after correction; the curve is obtained based on the changes in the 7-day average forecast and observed RMSE values ​​obtained from station A in a certain province in July 2022.

[0052] Figure 6 The visualization result of the first embodiment of the present invention is a graph showing the changes in root mean square error (RMSE) and lead time, displaying the RMSE values ​​before and after correction. This graph is based on the changes in the 7-day average forecast and observed RMSE values ​​obtained from station A in a certain province in August 2022.

[0053] Figure 7 The visualization result of the first embodiment of the present invention is a graph showing the changes in root mean square error (RMSE) and lead, displaying the RMSE values ​​before and after correction. This graph is based on the changes in the 7-day average forecast and observed RMSE values ​​obtained from station A in a certain province in September 2022.

[0054] Figure 8 The first embodiment of the present invention is a visualization result graph, which is a curve showing the changes in root mean square error (RMSE) and lead time, displaying the RMSE values ​​before and after correction; the curve is obtained based on the changes in the 7-day average forecast and observed RMSE values ​​obtained from station A in a certain province in October 2022.

[0055] Figure 9The visualization result of the first embodiment of the present invention is a graph of root mean square error (RMSE) versus lead time (after correction - before correction, negative values ​​represent improvement, and negative values ​​indicate improved correction effect). This graph is based on the changes in RMSE (root mean square error) improvement for different lead times obtained from Station A in a certain province in May 2022.

[0056] Figure 10 The visualization result of the first embodiment of the present invention is a graph of root mean square error (RMSE) versus lead time (after correction - before correction, negative values ​​indicate improved correction effect). This graph is based on the changes in RMSE (root mean square error) improvement for different lead times obtained from station A in a certain province in June 2022.

[0057] Figure 11 The visualization result of the first embodiment of the present invention is a graph of root mean square error (RMSE) versus lead time (after correction - before correction, negative values ​​indicate improved correction effect). This graph is based on the changes in RMSE (root mean square error) improvement for different lead times obtained from Station A in a certain province in July 2022.

[0058] Figure 12 The visualization result of the first embodiment of the present invention is a graph of root mean square error (RMSE) versus lead time (after correction - before correction, negative values ​​indicate improved correction effect). This graph is based on the changes in RMSE (root mean square error) improvement for different lead times obtained from Station A in a certain province in August 2022.

[0059] Figure 13 The visualization result of the first embodiment of the present invention is a graph of root mean square error (RMSE) versus lead time (after correction - before correction, negative values ​​indicate improved correction effect). This graph is based on the changes in RMSE (root mean square error) improvement for different lead times obtained from Station A in a certain province in September 2022.

[0060] Figure 14 The visualization result of the first embodiment of the present invention is a graph of root mean square error (RMSE) versus lead time (after correction - before correction, negative values ​​indicate improved correction effect). This graph is based on the changes in RMSE (root mean square error) improvement for different lead times obtained from station A in a certain province in October 2022. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of this invention.

[0062] Example 1: As Figures 1 to 14 As shown, the CFSv2 ground solar radiation correction method based on quantile regression of the present invention includes the following steps:

[0063] S1: Data Acquisition and Preprocessing. This step includes:

[0064] Observational data quality control: Select typical stations [e.g., 6 stations in a certain province, Figures 3 to 14 The correction status of one of the stations (Station A) is given. Hourly solar radiation observation data were collected and quality control was performed, including data cleaning. Step S1, the quality control process includes:

[0065] Remove abnormal irradiance values ​​(i.e., remove physically unreasonable irradiance values). Abnormal irradiance values ​​are those greater than 1500 W / m².

[0066] Missing values ​​are handled by interpolation or filling in values ​​before and after the time interval, thereby ensuring data integrity and data quality.

[0067] After quality control is completed, the data is corrected using date information, the observation data is grouped by day, the daily average irradiance is calculated, missing values ​​are filled with -999, and the processed data is stored in a date × site matrix format, ultimately forming a complete date × site data matrix.

[0068] Step S1 also includes model data preprocessing. The model data preprocessing method is to download shortwave radiation data from CFSv2 and merge the data into daily data to obtain model datasets. Each model dataset will be converted into NetCDF format to ensure that the model data is aligned with the observation data for comparison and correction.

[0069] For the model data preprocessing step, the model files are read, and the GRIB2 files of the CFSv2 model are read in batches using an automated script. The shortwave radiation data contained in the files are first converted to NetCDF format using a built-in decoding tool, and data for each time period is extracted. Every four time periods of data are merged into daily data to ensure the temporal sequence of the data. Specifically, shortwave radiation data for six hours is downloaded from CFSv2, and every four time periods of data are merged into daily data. Each model dataset is converted to NetCDF format, retaining necessary fields such as initial forecast time, lead time, and valid time to ensure that the model data can be aligned with observational data for comparison and correction.

[0070] S2: Spatiotemporal matching, this step includes:

[0071] Time alignment: Register the observation data with the CFSv2 model data by date to ensure that the two are completely consistent in time;

[0072] Spatial alignment: Spatial interpolation is performed on the model data to interpolate the grid data of the CFSv2 model onto the latitude and longitude coordinates of the actual observation station.

[0073] S3: Quantile regression correction model establishment and training. This step includes:

[0074] Model Establishment: A nonlinear regression relationship between observations and the model is established using the quantile regression method. The regression equation is as follows:

[0075] ,

[0076] in: Indicates the observed irradiance. This represents the irradiance predicted by the model, and lead represents the lead time. For the term of the lead time squared, , , , Here, ε represents the regression coefficients, and ε represents the residuals.

[0077] Model training and parameter tuning: Train the regression model using training data (e.g., observational and pattern data from May to October 2022); then adjust and optimize the parameters.

[0078] In step S3, during the model training and parameter tuning process, the median τ = 0.5 is used as the target quantile, and the quantile regression algorithm is used to estimate the model parameters, where τ is the quantile value.

[0079] In step S3, the parameter adjustment and optimization method is to use grid search and cross validation to adjust and optimize the parameters of the regression model. Grid search is used to traverse different lead ranges and quantile values ​​(τ) to find the best regression parameters, and cross validation is used to evaluate the generalization ability of the model and prevent overfitting.

[0080] Regularization and model selection: In order to improve the robustness of the model, L2 regularization is used to optimize the regression model, reduce the influence of higher-order terms on the model, avoid overfitting caused by excessive model complexity, and select the optimal model after regularization.

[0081] Parameter optimization and model convergence: The model is optimized by maximizing the log-likelihood function and minimizing the error residuals. During the optimization process, stochastic gradient descent (SGD) and the Adam optimizer are used to ensure effective convergence of model parameters. Furthermore, advanced techniques such as Bayesian optimization are applied to adjust the model hyperparameters to ensure optimal model configuration.

[0082] This step performs quantile regression analysis on the observed and model data, selecting the median (τ=0.5) as the target quantile, and estimating the regression coefficients ( , , , The Quantile Regression algorithm is used to fit the model by minimizing the error, capturing the nonlinear characteristics of the model data. Error correction in this step involves using a quantile regression model to correct the model forecast values, generating revised solar radiation forecasts. The lead time and its squared term in the regression equation ensure that nonlinear errors caused by variations in lead time can be handled.

[0083] S4: The correction process involves using a quantile regression model to analyze the model data and obtain the corrected irradiance forecast values.

[0084] S5: Model performance evaluation. The root mean square error (RMSE) was used to compare and analyze the model data before and after the correction to evaluate the correction effect. The evaluation results show that, through quantile regression correction, the RMSE value of the corrected model decreased by an average of 20%-35%, and the bias approached zero. Furthermore, the forecast error was significantly reduced under extreme weather conditions (cloudy and complex weather conditions).

[0085] Step S5, the model performance evaluation step, also includes visualizing the correction effect. This involves plotting the changes in root mean square error (RMSE) versus lead time to compare the effects before and after correction, further validating the effectiveness of the quantile regression method. For example... Figures 3 to 14 As shown.

[0086] S6: Medium- and Long-Term Solar Resource Trend Forecast. Applying this forecast result to the trend forecast product, and inputting the corrected irradiance data into the medium- and long-term wind and solar resource trend forecast product, can significantly improve forecast accuracy. Especially in the field of photovoltaic power generation, accurate solar radiation forecasts can effectively improve the accuracy of new energy power generation and help power dispatch decisions.

[0087] from Figure 3 It can be seen that the original root mean square error (RMSE) value reached a peak of about 90 on the 8th day, and the corrected root mean square error (RMSE) value was 25-50. The corrected root mean square error (RMSE) value on the entire curve segment was better than the original root mean square error (RMSE) value.

[0088] from Figure 4 It can be seen that the original root mean square error (RMSE) value reached a peak of about 75 on day 25 (with a small peak of about 72 on days 8-9). The corrected root mean square error (RMSE) value was 30-53. The corrected root mean square error (RMSE) value across the entire curve segment was almost better than the original root mean square error (RMSE) value, with a more significant decrease at the end of the curve segment.

[0089] from Figure 5 It can be seen that the original root mean square error (RMSE) value peaked at about 70 on day 34. After correction, the RMSE value dropped to a trough of about 16 on day 18. It briefly approached the original value around day 30, and then the RMSE value fell back to about 35. The RMSE value of the entire corrected curve segment was better than the original RMSE value.

[0090] from Figure 6 It can be seen that the original root mean square error (RMSE) values ​​show double peaks on days 19-21 and 38-40, with the double peaks being approximately 72 and 82, respectively. The corrected RMSE values ​​range from 36 to 52, showing slight fluctuations. The corrected RMSE values ​​across the entire curve segment are lower than the original RMSE values, with the peak segment showing the most significant improvement.

[0091] from Figure 7It can be seen that the original root mean square error (RMSE) value reached its peak on days 24-26, with a peak value of 79-80; the corrected root mean square error (RMSE) value was 41-52, which was relatively stable. The corrected root mean square error (RMSE) value across the entire curve segment was lower than the original root mean square error (RMSE) value, with the peak segment of the curve segment showing the most significant improvement.

[0092] from Figure 8 It can be seen that the original root mean square error (RMSE) value was close to 82 on day 9, and soared to 105-110 on days 21-23. After 10 days of correction, the RMSE value continued to decrease, and on day 21 after correction, the RMSE value dropped to about 5. The RMSE value of the entire corrected curve segment was significantly better than the original RMSE value, and the RMSE value of the last segment of the curve segment rose back to about 25.

[0093] Summary: From Figures 3 to 8 It can be seen that the root mean square error (RMSE) values ​​on the corrected curve segment from May to October 2022 were lower than the original RMSE values ​​in most lead periods. The largest changes (improvements) in the original RMSE values ​​generally occurred at the peaks of the curve segment, such as days 8-10 of May 2022, and the peak periods in August and October 2022. Short and medium lead periods yielded the greatest benefits, while long lead periods maintained positive benefits, indicating that introducing lead periods can effectively characterize and correct systematic errors that evolve with lead periods.

[0094] from Figure 9 It can be seen that the original root mean square error (RMSE) value is optimal from day 7 to 12, ranging from -65 to -70. It is second optimal from day 16 to 20, ranging from -25 to -35. It is also second optimal from day 32 to 36, ranging from -25 to -35. At the end of the curve segment, from day 40 to 45, the RMSE value is -5 to -10. The RMSE value for the entire curve segment is negative.

[0095] from Figure 10 It can be seen that the original root mean square error (RMSE) value is strongest on days 8 and 9, approximately -35. It is next strongest on days 24-26, ranging from -25 to -30. It is weakest on days 32-34 and 40-42, ranging from -3 to -6. In other time periods, it is mostly between -10 and -20. The entire RMSE value along the curve is negative.

[0096] from Figure 11It can be seen that during the initial time period on the curve segment, the root mean square error (RMSE) turns positive from day 1 to day 10, ranging from +5 to +10, and is approximately zero on day 10. The original RMSE value is optimal from day 16 to day 20, with a trough of -38 to -41. It is approximately zero on day 30. From day 35 to day 42, the corrected RMSE value is -15 to -20.

[0097] from Figure 12 It can be seen that the root mean square error (RMSE) briefly turns positive on days 2-5, reaching a maximum of +20. On days 6-7, it turns negative after reaching zero. The main improvements occur on days 9-10 (RMSE approximately -22), days 18-20 (RMSE approximately -27), and days 39-41 (RMSE approximately -33). In the middle section of the curve, the RMSE value is mostly between -10 and -20, while the overall RMSE value across the entire curve is predominantly negative.

[0098] from Figure 13 It can be seen that in the initial period of the curve segment, the root mean square error (RMSE) value is -2 to -8. It becomes a stable negative value and is strongest on days 15-18 and 20-26, with RMSE values ​​ranging from -28 to -30. In the final segment of the curve segment, the value is mostly -9 to -15. The RMSE value does not turn positive across the entire curve segment.

[0099] from Figure 14 It can be seen that the root mean square error (RMSE) briefly turns positive on days 2-4, ranging from +5 to +10. On day 5, it turns negative after reaching zero. Significant improvement is observed on days 8-12, with RMSE values ​​ranging from -20 to -35. The strongest improvement occurs on days 18-22, with the lowest RMSE value at -100. The RMSE value remains negative at the end of the curve segment, ranging from -20 to -30.

[0100] Summary: From Figures 9 to 14It can be seen that the improvements during the six months from May to October 2022 were mainly characterized by negative root mean square error (RMSE) values. Peak improvements typically aligned with the high error bands of the original RMSE value (e.g., around days 8-12 in May, days 24-26 in September, and around day 20 in October). Positive (or nearly positive) RMSE values ​​were only observed in the initial period of the July 2022 curve segment and in the extremely short lead periods of August and October; improvements were achieved in all other lead periods. Prioritizing windows such as days 7-15 in May, days 20-30 in June and September, days 16-22 in July, days 18-20 in August, days 39-41 in August, and days 18-22 in October offers both robustness and maximum returns.

[0101] This invention is based on multiple sites in a certain province ( Figures 3 to 14 (Only the situation before and after correction for station A in a certain province is given.) Solar radiation observations and CFSv2 model data were used. Quantile regression was employed with explicit introduction of lead time for correction. Only the 7-day mean was calculated for the model sequence, while the observations were retained at the daily scale. The evaluation results are shown in [link to evaluation]. Figures 3 to 14 The results show that the RMSE of most lead periods decreased significantly, and the root mean square error (RMSE) values ​​were mainly negative, which verifies the effectiveness and robustness of the proposed method in capturing the nonlinear evolution of the lead error.

[0102] Example 2: As Figure 2 As shown, a CFSv2 terrestrial solar radiation correction system based on quantile regression is disclosed. The correction system uses the aforementioned CFSv2 terrestrial solar radiation correction method based on quantile regression. The correction system includes:

[0103] Data acquisition and preprocessing module 1 is used for observation data quality control and model data preprocessing;

[0104] Spatiotemporal matching module 2 is used for time alignment and spatial alignment;

[0105] Quantile regression correction model building and training module 3 is used for model building, model training and parameter tuning;

[0106] Correction process module 4 uses a quantile regression model to analyze the model data to obtain corrected irradiance forecast values;

[0107] Model performance evaluation module 5 is used for model performance evaluation;

[0108] The medium- and long-term solar energy resource trend prediction module 6 is used for trend prediction products, and inputs the corrected irradiance data into the medium- and long-term wind and solar energy resource trend prediction products.

[0109] System Integration and Application: The correction method of this invention has been integrated into the "Wind and Solar Energy Resource Trend Forecasting Product Management System V1.0" and has become part of the system. This system can correct solar radiation data from models such as CFSv2 in real time, output more accurate medium- and long-term trend forecasts of solar energy resources, and support further energy management and grid dispatch.

[0110] Application Scenarios: The correction method of this invention can be widely applied to new energy resource assessment systems such as photovoltaic power plants and wind farms. By combining with power dispatching systems, it can provide more accurate photovoltaic power generation prediction data, thereby helping to improve the dispatching efficiency and stability of the power grid.

[0111] In summary, this invention provides a CFSv2 ground-based solar radiation correction method based on quantile regression. This method corrects ground-based solar radiation data, reduces daily radiation forecast errors, and significantly improves the accuracy and reliability of medium- and long-term solar energy resource trend predictions, particularly under extreme weather conditions. It solves the problems of not being able to reduce daily radiation forecast errors and significantly improve the accuracy and reliability of medium- and long-term solar energy resource trend predictions when correcting ground-based solar radiation data.

Claims

1. A CFSv2 ground-based solar radiation correction method based on quantile regression, characterized in that, It includes the following steps: S1: Data Acquisition and Preprocessing. This step includes: Observational data quality control: Select typical stations, collect hourly solar radiation observation data for quality control, and carry out data cleaning; S2: Spatiotemporal matching, this step includes: Time alignment: Register the observation data with the CFSv2 model data by date to ensure that the two are completely consistent in time; Spatial alignment: Spatial interpolation is performed on the model data to interpolate the grid data of the CFSv2 model onto the latitude and longitude coordinates of the actual observation station; S3: Quantile regression correction model establishment and training. This step includes: Model Establishment: A nonlinear regression relationship between observations and the model is established using quantile regression. The regression equation is as follows: ， in: Indicates the observed irradiance, This represents the predicted irradiance by the model, and lead represents the lead time. For the term of the lead time squared, , , , Here, ε represents the regression coefficients, and ε represents the residuals. Model training and parameter tuning: Train the regression model using training data; then adjust and optimize the parameters. In step S3, the parameter adjustment and optimization method is to use grid search and cross-validation to adjust and optimize the parameters of the regression model. Grid search is used to traverse different lead periods and quantile values ​​τ to find the optimal regression parameters. Regularization and model selection: The regression model is optimized using L2 regularization, and the optimal model after regularization is selected; Parameter optimization and model convergence: The model is optimized by maximizing the log-likelihood function and minimizing the error residual. During the optimization process, stochastic gradient descent and Adam optimizer are used to ensure that the model parameters can converge effectively. S4: Correction process, using quantile regression models on model data to obtain corrected irradiance forecast values; S5: Model performance evaluation. The root mean square error is used to compare and analyze the model data before and after correction to evaluate the correction effect.

2. The CFSv2 ground solar radiation correction method based on quantile regression according to claim 1, characterized in that, In step S1, the quality control process includes: Exclusion of irradiance values: Irradiance values ​​greater than 1500 W / m² are excluded. Missing values ​​are handled by filling them in using interpolation or preceding / following time values; After quality control is completed, the data is corrected using date information, the observation data is grouped by day, the daily average irradiance is calculated, missing values ​​are filled with -999, and the processed data is stored in a date × site matrix format, ultimately forming a complete date × site data matrix.

3. The CFSv2 ground solar radiation correction method based on quantile regression according to claim 1, characterized in that, Step S1 also includes model data preprocessing. The model data preprocessing method is to download shortwave radiation data from CFSv2 and merge the data into daily data to obtain model datasets. Each model dataset will be converted into NetCDF format to ensure that the model data is aligned with the observation data for comparison and correction.

4. The CFSv2 ground solar radiation correction method based on quantile regression according to claim 3, characterized in that, In the model data preprocessing step, shortwave radiation data for 6 hours is downloaded from CFSv2, and data from every 4 hours are merged into daily data. Each model dataset is converted into NetCDF format, and the start time, lead time, and effective time are retained to ensure that the model data is aligned with the observation data for comparison and correction.

5. The CFSv2 ground solar radiation correction method based on quantile regression according to claim 1, characterized in that, In step S3, during the model training and parameter tuning process, the median τ = 0.5 is used as the target quantile, and the quantile regression algorithm is used to estimate the model parameters, where τ is the quantile value.

6. The CFSv2 ground solar radiation correction method based on quantile regression according to claim 1, characterized in that, In step S5, the model performance evaluation step also includes visualizing the correction effect. By plotting the change curves of root mean square error and lead time, the effects before and after correction are compared, further verifying the effectiveness of the quantile regression method.

7. The CFSv2 ground solar radiation correction method based on quantile regression according to claim 1, characterized in that... It also includes: S6: Medium- and Long-Term Solar Energy Resource Trend Forecast The prediction results were applied to the trend forecasting product, and the corrected irradiance data were input into the medium- and long-term wind and solar energy resource trend forecasting product.

8. A CFSv2 terrestrial solar radiation correction system based on quantile regression, wherein the correction system uses the CFSv2 terrestrial solar radiation correction method based on quantile regression as described in any one of claims 1 to 7, characterized in that, The correction system includes: The data acquisition and preprocessing module (1) is used for observation data quality control and model data preprocessing; The spatiotemporal matching module (2) is used for time alignment and spatial alignment; The quantile regression correction model building and training module (3) is used for model building, model training and parameter tuning. The correction process module (4) uses a quantile regression model to process the model data to obtain the corrected irradiance forecast value; Model performance evaluation module (5) is used for model performance evaluation; The medium- and long-term solar energy resource trend prediction module (6) is used for trend prediction products. It inputs the corrected irradiance data into the medium- and long-term wind and solar energy resource trend prediction products.

Images