Solar power generation power prediction method based on improved elastic network regression

Abstract

The invention discloses a solar power generation power prediction method based on improved elastic network regression, and belongs to the field of solar power generation system output prediction. The problem that an existing solar power generation potential prediction method is poor in accuracy is solved. The method comprises the following steps: collecting meteorological influence factor data and network access power real-time power data of a place where a target solar power generation system is located, establishing an output-meteorological characteristic data set, determining a tradeoff coefficient r and a regularization strength parameter lambda by using a grid search algorithm in combination with k-fold cross validation, and constructing an elastic network regression model; and training the regression model, optimizing a trade-off coefficient r and a regularization strength parameter lambda of the elastic network regression model by using a TPE optimization algorithm to obtain a solar power generation prediction model, testing the power generation prediction model by using a test set, and finally performing power generation power prediction by using the tested power generation prediction model. The method is suitable for solar power generation power prediction.

Description

Technical Field

[0001] The invention belongs to the field of solar power generation system output prediction. Background Art

[0002] At present, existing technical solutions mostly use prediction methods based on neural networks such as LSTM, which is a recurrent neural network composed of a forget gate, an input gate, and an output gate. LSTM solves the problems of gradient disappearance and gradient explosion by introducing a gating mechanism, thereby realizing the training and optimization of the model.

[0003] LSTM and other methods are mainly used to establish the correlation between relevant factors and output. However, due to the increasing complexity, variability, randomness and uncertainty of output data in recent years, the existing solar power generation potential prediction methods have poor accuracy. Summary of the invention

[0004] The present invention aims to solve the problem of poor accuracy of existing solar power generation potential prediction methods, and now provides a solar power generation power prediction method based on improved elastic net regression.

[0005] The solar power generation power prediction method based on improved elastic net regression of the present invention comprises:

[0006] Step 1: Collect meteorological influencing factor data and real-time power data of grid-connected power at the location of the target solar power generation system;

[0007] Step 2: pre-process the meteorological influencing factor data and the real-time power data of the grid-connected power, screen the meteorological influencing factors using the Spearman correlation coefficient analysis method, obtain strong influencing factors, correspond the pre-processed data corresponding to the strong influencing factors with the processed real-time power data of the grid-connected power according to the same principle of time points, and merge them to obtain an output-meteorological characteristic data set, and divide the data set into a training set and a test set;

[0008] Step 3: Determine the trade-off coefficient r and the regularization strength parameter λ by using a grid search algorithm combined with k-fold cross validation, and construct an elastic net regression model by using the trade-off coefficient r and the regularization strength parameter λ;

[0009] Step 4: train the elastic net regression model using the training set. After the training is completed, optimize the trade-off coefficient r and the regularization intensity parameter λ of the elastic net regression model using the TPE optimization algorithm to obtain a solar power generation prediction model.

[0010] Step 5: Use the test set to test the solar power generation prediction model to determine whether the prediction result of the solar power generation prediction model meets the accuracy requirement. If so, use the solar power generation prediction model to predict the grid-connected power of the target solar power generation system. Otherwise, execute step 6.

[0011] Step 6: Increase the number of training samples, increase the training rounds of the TPE optimization algorithm, return to step 4, until the prediction result of the solar power generation prediction model meets the accuracy requirement, and use the solar power generation prediction model to predict the power generation power of the target solar power generation system.

[0012] Furthermore, in the present invention, in step one, the meteorological influencing factors include: temperature, relative humidity, dew point, perceived temperature, rainfall, snowfall, air pressure, shortwave radiation, global oblique irradiance, ground radiation, direct radiation and diffuse reflection.

[0013] Furthermore, in the present invention, in step 2, the method for preprocessing the meteorological influencing factors and the real-time power data of the grid-connected electric power is:

[0014] Detect missing values ​​and outliers in the load characteristic data set, and use the level average interpolation method to fill or replace the missing values ​​and outliers;

[0015] The calculation formula of the level mean interpolation method is:

[0016]

[0017] Among them, y d,t is the photovoltaic grid-connected output value at time t on day d, y d,t-1 is the photovoltaic grid-connected output value at time t-1 on day d, y d,t+1 It is the photovoltaic grid-connected output value at time t+1 on the d day.

[0018] Furthermore, in the present invention, abnormal values ​​include outliers, noise, repeated values ​​and extreme values.

[0019] Furthermore, in the present invention, in step 2, the Spearman correlation coefficient analysis method is used to screen the meteorological influencing factors, and the formula for obtaining the strong influencing factors is:

[0020]

[0021] Where n represents the number of samples, d i Represents the data of meteorological influencing factors, ρ represents the correlation coefficient, and the value range of ρ is between -1 and 1, indicating the strength of the correlation between variables; a value of 1 indicates a completely positive correlation, a value of -1 indicates a completely negative correlation, and a value of 0 indicates no correlation.

[0022] Furthermore, in the present invention, in step 4, the method of using the TPE algorithm to optimize the parameters of the grid search algorithm is:

[0023] The number of training rounds of the TPE algorithm was set to 1000, the number of candidates for expected EI improvement was set to 24, the number of initial verifications was 10, each adjustment increased the total number of rounds by 1%, and the prior weight was set to 1.0;

[0024] The search range of the hyperparameters is used as the input of the TPE algorithm, and the minimization of the model performance index is used as the objective function to search for the regularization strength parameter λ and the trade-off coefficient r. The search interval of the regularization strength parameter λ is 10 -4 ~1.0, the search range of the trade-off coefficient r is 0.00~1.00.

[0025] Furthermore, in the present invention, the model performance indicators include mean absolute error MAE and daily cumulative mean absolute error DSMAE.

[0026] Furthermore, in the present invention, the mean absolute error MAE is:

[0027]

[0028] Where n is the number of samples; y i is the i-th observation value; is the i-th prediction value; the legal interval of MAE is [0, +∞]. The smaller the value, the smaller the deviation between the model's prediction result and the actual value, which means that the model's prediction accuracy is higher.

[0029] Furthermore, in the present invention, the daily cumulative mean absolute error DSMAE is:

[0030]

[0031] In the formula, n is the number of days, y is true , i is the actual grid-connected power value on the ith day (in megawatt-hours), y pred , i is the predicted value for the i-th day. The legal range of DSMAE is [0, +∞]. Like MAE, the smaller the value of DSMAE, the better the model fitting effect.

[0032] The method of the present invention is based on the elastic net regression model, and uses the TPE optimization algorithm based on Bayesian prior to improve the elastic net regression model by adjusting the regularization term strength and ratio hyperparameters, and the prediction effect is further improved. The regression model after parameter optimization by TPE takes the meteorological factors based on NWP data as input and the photovoltaic power prediction value at the corresponding time as output. Experiments are conducted on real photovoltaic data sets. The results show that the average absolute error of the prediction method is reduced by 31.96% compared with the original elastic net regression method, and has an improvement of 31.84-45.97% compared with other machine learning models, which verifies that this method can effectively improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the method of the present invention;

[0034] Figure 2 It is a correlation matrix diagram between output and meteorological factors obtained by Spearman correlation coefficient analysis method.

[0035] Figure 3 This is the result graph of the hyperparameter search K-fold cross validation;

[0036] Figure 4 This is the evaluation result graph of the hyperparameter search L1_ratio;

[0037] Figure 5 This is a graph comparing the overall prediction and the actual effect in the test set. DETAILED DESCRIPTION

[0038] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0039] Specific implementation method 1: refer to Figure 1 Specifically describing this embodiment, the solar power generation power prediction method based on improved elastic net regression described in this embodiment includes:

[0040] Step 1: Collect meteorological influencing factor data and real-time power data of grid-connected power at the location of the target solar power generation system;

[0041] Step 2: pre-process the meteorological influencing factor data and the real-time power data of the grid-connected power, screen the meteorological influencing factors using the Spearman correlation coefficient analysis method, obtain strong influencing factors, correspond the pre-processed data corresponding to the strong influencing factors with the processed real-time power data of the grid-connected power according to the same principle of time points, and merge them to obtain an output-meteorological characteristic data set, and divide the data set into a training set and a test set;

[0042] Step 3: Determine the trade-off coefficient r and the regularization strength parameter λ by using a grid search algorithm combined with k-fold cross validation, and construct an elastic net regression model by using the trade-off coefficient r and the regularization strength parameter λ;

[0043] Step 4: train the elastic net regression model using the training set. After the training is completed, optimize the trade-off coefficient r and the regularization intensity parameter λ of the elastic net regression model using the TPE optimization algorithm to obtain a solar power generation prediction model.

[0044] Step 5: Use the test set to test the solar power generation prediction model to determine whether the prediction result of the solar power generation prediction model meets the accuracy requirement. If so, use the solar power generation prediction model to predict the grid-connected power of the target solar power generation system. Otherwise, execute step 6.

[0045] Step 6: Increase the number of training samples, increase the training rounds of the TPE optimization algorithm, and return to execute step 4 until the prediction result of the solar power generation prediction model meets the accuracy requirement.

[0046] Furthermore, in the present invention, in step one, the meteorological influencing factors include: temperature, relative humidity, dew point, perceived temperature, rainfall, snowfall, air pressure, shortwave radiation, global oblique irradiance, ground radiation, direct radiation and diffuse reflection.

[0047] Furthermore, in the present invention, in step 2, the method for preprocessing the meteorological influencing factors and the real-time power data of the grid-connected electric power is:

[0048] Detect missing values ​​and outliers in the load characteristic data set, and use the level average interpolation method to fill or replace the missing values ​​and outliers;

[0049] The calculation formula of the level mean interpolation method is:

[0050]

[0051] Among them, y d,t is the photovoltaic grid-connected output value at time t on day d, y d,t-1 is the photovoltaic grid-connected output value at time t-1 on day d, y d,t+1 It is the photovoltaic grid-connected output value at time t+1 on the d day.

[0052] Furthermore, in the present invention, abnormal values ​​include outliers, noise, repeated values ​​and extreme values.

[0053] Furthermore, in the present invention, in step 2, the Spearman correlation coefficient analysis method is used to screen the meteorological influencing factors, and the formula for obtaining the strong influencing factors is:

[0054]

[0055] Where n represents the number of samples, d i Represents the data of meteorological influencing factors, ρ represents the correlation coefficient, and the value range of ρ is between -1 and 1, indicating the strength of the correlation between variables; a value of 1 indicates a completely positive correlation, a value of -1 indicates a completely negative correlation, and a value of 0 indicates no correlation.

[0056] Furthermore, in the present invention, in step 4, the method of using the TPE algorithm to optimize the parameters of the grid search algorithm is:

[0057] The number of training rounds of the TPE algorithm was set to 1000, the number of candidates for expected EI improvement was set to 24, the number of initial verifications was 10, each adjustment increased the total number of rounds by 1%, and the prior weight was set to 1.0;

[0058] The search range of the hyperparameters is used as the input of the TPE algorithm, and the minimization of the model performance index is used as the objective function to search for the regularization strength parameter λ and the trade-off coefficient r. The search interval of the regularization strength parameter λ is 10 -4 ~1.0, the search range of the trade-off coefficient r is 0.00~1.00.

[0059] Furthermore, in the present invention, the model performance indicators include mean absolute error MAE and daily cumulative mean absolute error DSMAE.

[0060] Furthermore, in the present invention, the mean absolute error MAE is:

[0061]

[0062] Where n is the number of samples; y i is the i-th observation value; is the i-th prediction value; the legal range of MAE is [0, +∞], and the smaller its value is, the smaller the deviation between the model's prediction result and the actual value is, indicating that the accuracy of the model prediction is higher. Further, in the present invention, the daily cumulative mean absolute error DSMAE is:

[0063]

[0064] In the formula, n is the number of days, y istrue , i is the actual grid-connected power value on the ith day (in megawatt-hours), y pred , i is the predicted value for the i-th day. The legal range of DSMAE is [0, +∞]. Like MAE, the smaller the value of DSMAE, the better the model fitting effect. Specific embodiment:

[0066] Step 1: Determine the time span of the output forecast target area, complete the collection of the output data and meteorological influencing factor data of the area, and construct the output characteristic data set and meteorological characteristic data set and select the model evaluation indicators widely used in the field of short-term solar power generation potential forecasting to analyze the forecast results and evaluate the model performance. The input of the data set is the meteorological influencing factor data, and the output is the output data, which is expressed as (x1, y1), (x2, y2), ···, (xN, yN);

[0067] Among them, xi is a column vector containing meteorological influencing factor data, and yi is the output corresponding to xi.

[0068] Specifically, the data of meteorological influencing factors include: temperature, relative humidity, dew point, perceived temperature, precipitation, rainfall, snowfall, snow thickness, weather code, sea level pressure, surface pressure, cloud cover, low cloud cover, medium cloud cover, high cloud cover, total soil moisture evaporation loss, vapor pressure difference, 10m wind speed, 100m wind speed, 10m wind direction, 100m wind direction, 10m gust, date code, sunshine hours, shortwave radiation, direct radiation, diffuse radiation, direct irradiance, global tilted irradiance, ground radiation, shortwave radiation instantaneous value, direct radiation instantaneous value, diffuse reflection instantaneous value, normal plane solar direct radiation global tilted irradiance instantaneous value and ground radiation instantaneous value.

[0069] In practical applications, before constructing the data set, data preprocessing (including filling of missing values ​​and abnormal values, division and standardization of data sets, etc.) is performed on the collected meteorological influencing factor data and the corresponding output data.

[0070] Step 2: Correlation analysis. The spearman correlation coefficient analysis method is used to screen meteorological elements that show a strong correlation with output, including shortwave radiation, global tilt irradiance, and ground radiation. Figure 1 shown.

[0071] besides:

[0072] Other categories of information also show certain data correlation, such as temperature and relative humidity. Therefore, in the actual modeling process, considering the accuracy of the model and the forecast demand of photovoltaic power station areas under abnormal climate, other NWP data are still introduced as modeling features. Although the above features have significant correlation with solar photovoltaic power generation (MW), they also have multicollinearity problems. For example, global tilted irradiance (W / m 2 ) and shortwave radiation (W / m 2 ), which will lead to the instability of the coefficients in the regression model, thus affecting the explanatory and predictive capabilities of the model.

[0073] When selecting features and dealing with multicollinearity issues:

[0074] It is not appropriate to use Lasso regression or Ridge regression alone. Lasso regression tends to select a small number of features and may ignore some important features. Although Ridge regression can handle multicollinearity, it retains all features and cannot effectively remove redundant information. Elastic net regression can automatically select the features most relevant to the output while maintaining the stability of the model, thereby improving the interpretability and prediction performance of the model. For example, in elastic net regression, shortwave radiation (W / m 2 ) will be adjusted due to their high correlation rather than simply being kept or deleted. Therefore, elastic net regression is the most appropriate choice in the modeling process.

[0075] Step 3: Using the data set, a short-term solar power generation prediction model is constructed through an elastic net regression model and the prediction results are compared with those of other linear models through the evaluation index.

[0076] Specifically, in the linear model with regularization term, the regularization term coefficient alpha is controlled to be 0.1, and r in the elastic net model is configured to be 0.5, which means that the elastic net treats the two regularization penalty terms equally during the fitting process. The comparison results of each model are shown in the following table:

[0077] Table 1 Linear model comparison table

[0078]

[0079]

[0080] Among the four linear models, the elastic net model has the lowest moment power prediction error of 0.5838 and the second lowest daily power prediction error of 5.4494, which has more comprehensive performance than other linear models. The Lasso model with L1 penalty has the lowest daily power total error, but the MAE value is the largest at each moment, reaching 0.6103; for the LASSO model, compared with the PLS without regularization term and the Ridge regression model with L2 penalty, the MAE value is larger, indicating that it is too targeted at the feature selection part. Although it guarantees the prediction effect of the total moment of the model, in fact, because of the large amount of key information discarded, there are only five NWP features that are effectively used in the modeling process. Too few NWP features can achieve better results in the sample space of the training set, but the large amount of discarding other information will increase the potential risk of overfitting. Although Ridge can still maintain a certain accuracy in the feature space with multicollinearity, the lack of feature selection has affected the prediction accuracy of the model. The ESN model of elastic net regression combines two regularization terms. On the one hand, the model assigns greater weights to the features with significant correlation in NWP information, such as shortwave radiation, instantaneous oblique radiation and instantaneous ground radiation. It also tries to combine other potential indirect factors, such as temperature and humidity, so that the model can have more accurate numerical results in regression and has higher robustness to potential mutation models.

[0081] Step 4: Optimization of elastic net regression hyperparameters based on the TPE optimization algorithm with Bayesian prior.

[0082] The regularization strength parameter λ and the trade-off coefficient r in the elastic net model are further combined with the TPE algorithm to perform hyperparameter search. The hyperparameter search process uses the average MAE of the 4-fold cross-validation in the training set as the search target, and the search range of the regularization strength is 10 -4 ~10 -2 , the parameter search range of the ratio trade-off coefficient r is 0.2 to 0.8.

[0083] The results of 1000 hyperparameter searches are as follows Figure 2 , 3 As shown, Figure 2 The different results obtained by the model in the cross-validation with the number of iterations, where the red line represents the optimal parameters in the search process; Figure 3 This is the distribution of A-L1. The color depth of the points represents the evaluation value of the search process. In the figure, best represents the optimal solution.

[0084] Depend on Figure 2It can be seen that in 1000 searches, the model has obtained the historical optimal parameter combination after about 50 iterations. After 20 iterations, most of the obtained parameter combinations are close to the optimal results. This shows that the search process can determine the optimal range of the model in about half of the search history, and the optimal value under this range is confirmed in 50 iteration cycles.

[0085] Combination Figure 3 It can be seen that the TPE algorithm based on Bayesian prior has an alpha parameter of 10 in the optimal parameter range finally determined. -2 The choice of the trade-off coefficient r is concentrated around the smaller 0.2 to 0.4.

[0086] The search for a smaller trade-off coefficient r represents the model's emphasis on the L1 regularization term, which means that a stronger L1 regularization penalty can improve the model's prediction effect. According to the aforementioned Lasso results, the model adjusted by TPE is more inclined to use the L1 regularization term to complete feature selection, and in this process, the Ridge's L2 regularization term is maintained to a limited extent to cope with multicollinearity features.

[0087] Step 5: Establish a solar power generation potential prediction model based on the improved elastic net regression model and compare the prediction results with other linear models through the evaluation indicators.

[0088] The model was brought back for re-modeling. Compared with the original equal-weighted elastic net model, the model after parameter adjustment, on the one hand, increased the weights of ground radiation, instantaneous shortwave radiation, instantaneous direct normal irradiance and instantaneous earth radiation characteristics, and on the other hand, suppressed the weights of shortwave radiation, ground radiation, instantaneous direct radiation and instantaneous diffuse radiation characteristics, and no longer considered the influence of perceived temperature and precipitation.

[0089] In practical applications, the numerical weather forecast (NWP) data provided by the European Centre for Medium-Term Weather Forecasts (ECMWF) from January to December 2022 is used as input, and the grid-connected output of a photovoltaic power plant in a prefecture-level city in Heilongjiang Province located in the buffer zone between Lesser Khingan Mountains and Sanjiang Plain is obtained through the SCADA system. The prediction model is used to predict the electricity consumption data from April to July 2023.

[0090] The overall effect of the optimized model in predicting data in the test set is as follows: Figure 4As shown in the figure, params_l1_ratio represents the trade-off coefficient, value represents the numerical value, and params_alpha represents the regularization strength; the prediction data error is between 0.0040 and 8.3407, and the error range of the power prediction at each moment is between 0 and 2.10623. According to the 1,350 valid prediction points in the test set, the power prediction MAE index of the TPE-EN model is 0.3972, the variance is 0.5888, and the maximum error is 2.1063MW, which occurred at 11:00 on June 17; within 90 working days, the average daily prediction error is 1.8216MWh, the variance is 1.5584, and the maximum error is 8.3407MWh on June 17. The overall prediction and actual comparison effect diagram in the test set is shown below Figure 5 As shown in the figure, act represents the actual value and better represents the predicted value. The comparison and verification results of each model are shown in Table 2 below.

[0091] Table 2 Comparison of machine learning models

[0092] Model MAE DSMAE ours 0.397224868 1.821645102 svm 0.582814226 5.631304686 mlp 0.735226855 7.897078043 knn 0.639987018 5.872658648 XGB 0.706508705 7.625700191 XGBF 0.626217758 5.795215247

[0093] Combining Table 1 and Table 2, it can be seen that the mean absolute error of the present invention is reduced by 31.96% compared with the original elastic net regression method, and is reduced by 34.91%, 33.95% and 34.05% respectively compared with the same type of regression models such as Lasso, Ridge and pls; compared with other machine learning models such as SVM, MLP, KNN, XGB and XGBRF, it has 31.84%, 45.97%, 37.93%, 43.78% and 36.57% improvement respectively. The present invention shows excellent daily cumulative power generation prediction performance while maintaining the instantaneous power generation prediction accuracy, which has significant advantages over other models.

[0094] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that the features described in conjunction with a single embodiment may be used in other described embodiments.

Claims

1. A solar power generation prediction method based on improved elastic net regression, characterized in that: The method includes: Step 1: Collect meteorological influencing factor data and real-time power data of grid-connected power at the location of the target solar power generation system; Step 2: pre-process the meteorological influencing factor data and the real-time power data of the grid-connected power, screen the meteorological influencing factors using the Spearman correlation coefficient analysis method, obtain strong influencing factors, and correspond the pre-processed data corresponding to the strong influencing factors with the pre-processed real-time power data of the grid-connected power according to the same principle of time points to obtain an output-meteorological characteristic data set, and divide the data set into a training set and a test set; Step 3: Determine the trade-off coefficient r and the regularization strength parameter λ by using a grid search algorithm combined with k-fold cross validation, and construct an elastic net regression model by using the trade-off coefficient r and the regularization strength parameter λ; Step 4: train the elastic net regression model using the training set. After the training is completed, optimize the trade-off coefficient r and the regularization intensity parameter λ of the elastic net regression model using the TPE optimization algorithm to obtain a solar power generation prediction model. Step 5: Use the test set to test the solar power generation prediction model to determine whether the prediction result of the solar power generation prediction model meets the accuracy requirement. If so, use the solar power generation prediction model to predict the grid-connected power of the target solar power generation system. Otherwise, execute step 6. Step 6: Increase the number of training samples, increase the training rounds of the TPE optimization algorithm, and return to execute step 4 until the prediction result of the solar power generation prediction model meets the accuracy requirement.

2. The solar power generation power prediction method based on improved elastic net regression according to claim 1 is characterized in that: In step one, meteorological factors include: temperature, relative humidity, dew point, perceived temperature, rainfall, snowfall, air pressure, shortwave radiation, global slant irradiance, ground radiation, direct radiation and diffuse reflection.

3. The solar power generation power prediction method based on improved elastic net regression according to claim 1 is characterized in that: In step 2, the method for preprocessing the meteorological influencing factors and the real-time power data of the grid-connected power is: Detect missing values ​​and outliers in the load characteristic data set, and use the level average interpolation method to fill or replace the missing values ​​and outliers; The calculation formula of the level mean interpolation method is: Among them, y d,t is the photovoltaic grid-connected output value at time t on day d, y d,t-1 is the photovoltaic grid-connected output value at time t-1 on day d, y d,t+1 It is the photovoltaic grid-connected output value at time t+1 on the d day.

4. The solar power generation power prediction method based on improved elastic net regression according to claim 1, characterized in that: Outliers include outliers, noise, duplicate values, and extreme values.

5. The solar power generation power prediction method based on improved elastic net regression according to claim 1, characterized in that: In step 2, the Spearman correlation coefficient analysis method is used to screen the meteorological influencing factors, and the formula for obtaining the strong influencing factors is: Where n represents the number of samples, d i Represents output-meteorological characteristic data, ρ represents the strength of the correlation between meteorological factors and output, and the value range of ρ is between -1 and 1, indicating the strength of the correlation between variables; a value of 1 indicates a perfect positive correlation, a value of -1 indicates a perfect negative correlation, and a value of 0 indicates no correlation.

6. The solar power generation power prediction method based on improved elastic net regression according to claim 1, characterized in that: In step 4, the method of using the TPE algorithm to optimize the parameters of the grid search algorithm is: The total number of training rounds of the TPE algorithm is set to 1000, the number of candidates for expected EI improvement is 24, the number of initial verifications is 10, each adjustment increases the total number of rounds by 1%, and the prior weight is set to 1.0; The search range of the hyperparameters is used as the input of the TPE algorithm, and the minimization of the model performance index is used as the objective function to search for the regularization strength parameter λ and the trade-off coefficient r. The search interval of the regularization strength parameter λ is 10 -4 ~1.0, the search range of the trade-off coefficient r is 0.00~1.

00.

7. The solar power generation power prediction method based on improved elastic net regression according to claim 1, characterized in that: Model performance indicators include mean absolute error (MAE) and daily cumulative mean absolute error (DSMAE).

8. The solar power generation power prediction method based on improved elastic net regression according to claim 1, characterized in that: The mean absolute error MAE is: Where n is the number of samples; y i is the i-th observation value; is the i-th predicted value; the interval of MAE is [0, +∞], and the smaller the value is, the smaller the deviation between the model's predicted result and the actual value is.

9. The solar power generation power prediction method based on improved elastic net regression according to claim 1, characterized in that: The daily cumulative mean absolute error DSMAE is: In the formula, n is the number of days, y is true，i is the actual grid-connected power value on the ith day, y pred，i is the predicted value for the i-th day.

Images