Hybrid model establishment method for predicting ultra-short-term photovoltaic power generation output

By combining the hybrid model of LightGBM, XGBoost and CatBoost algorithms, the limitations of ultra-short-term photovoltaic power output prediction in the existing technology are solved, the prediction accuracy and generalization capabilities are improved, and the stable operation of the power grid and photovoltaic power stations are supported.

CN118920445APending Publication Date: 2024-11-08KUNYUE INTERNET ENVIRONMENTAL TECH (JIANGSU) CO LTD

Patent Information

Application Number
CN202410836921.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-26
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The prior art has limitations in predicting the output of ultra-short-term photovoltaic power generation. Statistical methods cannot capture nonlinear relationships. Physical methods are complex in calculations and are sensitive to input parameters, resulting in insufficient prediction accuracy.

Method used

A hybrid model construction method is adopted, combining three gradient enhancement algorithms: LightGBM, XGBoost and CatBoost, and hybrid models are built through data preprocessing and feature engineering, and prediction accuracy is improved through model fusion technology.

Benefits of technology

It improves the accuracy of prediction, reduces prediction errors, enhances the generalization ability of the model, reduces the risk of overfitting, supports grid scheduling and photovoltaic power station management, reduces operating costs, and promotes the use of renewable energy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118920445A_ABST
    Figure CN118920445A_ABST
Patent Text Reader

Abstract

The invention provides a hybrid model establishment method for predicting ultra-short-term photovoltaic power generation output, and the method is based on a hybrid model of LightGBM, XGBoost and CatBoost, so as to achieve higher prediction precision and stronger generalization capability. The prediction module is used for predicting ultra-short-term photovoltaic power generation output; the three models are all machine learning algorithms based on a gradient lifting framework, and the prediction performance is improved by constructing and combining a plurality of weak learners. According to the hybrid model, the prediction accuracy can be improved, the risk of overfitting can be reduced through a model fusion technology, and the generalization ability of the model under different conditions is improved. In addition, by optimizing the training process and parameter selection of the model, the calculation efficiency of prediction can be improved, and the actual demand of ultra-short-term photovoltaic power generation output prediction is met.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic prediction of new energy power generation, and in particular to a method for establishing a hybrid model for predicting ultra-short-term photovoltaic power generation output. Background Art

[0002] With the continuous growth of global energy demand, the development and utilization of renewable energy has received widespread attention. Among many renewable energy sources, solar photovoltaic power generation has the advantages of being clean, renewable, and environmentally friendly, and has become an important direction for future energy development. In recent years, photovoltaic power generation systems have been widely used in my country, and their installed capacity has increased year by year. However, photovoltaic power generation output is affected by many factors such as weather, temperature, and humidity, and has great uncertainty and volatility, which brings certain challenges to grid dispatching and operation.

[0003] In order to improve the prediction accuracy of photovoltaic power generation output, researchers have proposed a variety of prediction methods, including statistical methods and physical methods. Statistical methods such as time series analysis and linear regression rely on historical data for modeling; physical methods are based on solar radiation transmission models and consider the impact of factors such as the atmosphere and clouds on solar radiation. However, these methods have certain limitations in predicting ultra-short-term photovoltaic power generation output. Statistical methods may not be able to capture nonlinear relationships in the data, while physical methods are complex to calculate and highly sensitive to input parameters. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a hybrid model establishment method for predicting ultra-short-term photovoltaic power generation output.

[0005] To achieve the above-mentioned purpose, the present invention adopts the following technical solution: a hybrid model establishment method for predicting ultra-short-term photovoltaic power generation output, the hybrid model establishment method steps are as follows:

[0006] S1: Dataset collection raw data:

[0007] The original data includes: basic information of photovoltaic power generation users, meteorological variable data and actual power data;

[0008] The basic information of photovoltaic power generation users includes the location and installed capacity information of 9 photovoltaic power generation users, namely users f1-f9;

[0009] The meteorological variable data are meteorological data of 9 photovoltaic power generation users from January 2022 to January 2024, every 15 minutes, including light intensity, temperature, humidity, and wind speed;

[0010] The actual power data is the power output data of 9 photovoltaic power generation users from January 2022 to January 2024, every 15 minutes;

[0011] S2: Data preprocessing:

[0012] Performing data preprocessing on the raw data collected in step S1;

[0013] Data preprocessing includes merging data sets; dividing training sets, validation sets, and test sets; resetting the index of actual power data;

[0014] Through the above data preprocessing, the original data is converted into a format suitable for use by the prediction model and contains the necessary features and target variables;

[0015] S3: feature engineering processing;

[0016] The raw data was subjected to in-depth feature engineering;

[0017] Including: meteorological feature processing, historical value feature extraction, abnormal value processing, feature extraction of the ratio of light to the strongest light of the day, feature extraction of the average light of the day, and feature extraction of the difference between temperature and the highest and lowest temperatures of the day;

[0018] Through these feature engineering processing steps, the original data is converted into a format that contains rich information and is suitable for use in prediction models, providing a data basis for the construction and training of photovoltaic power generation output prediction models;

[0019] S4: Constructing a hybrid model for predicting ultra-short-term photovoltaic power generation output:

[0020] By combining three advanced gradient boosting algorithms, LightGBM, XGBoost and CatBoost, accurate prediction of photovoltaic power generation output can be achieved;

[0021] S41: Model parameter configuration:

[0022] Parameters were configured for LightGBM, XGBoost, and CatBoost algorithms, including learning rate, number of trees, and tree depth;

[0023] S42: Training-Cross Validation:

[0024] Use the five-fold cross-validation method to divide the dataset into five folds, each of which contains a portion of training data and validation data;

[0025] For each fold, the training dataset is used to train the models of three algorithms: LightGBM, XGBoost, and CatBoost;

[0026] Evaluate model performance on the validation dataset for each fold;

[0027] The root mean square error (RMSE) of each fold is calculated as the performance evaluation indicator, and the final evaluation result is the average RMSE of multiple folds;

[0028] S43: Model Fusion:

[0029] The model fusion technique was used to average the prediction results of the three models;

[0030] S5: Model performance evaluation:

[0031] After the hybrid model is constructed in step S4, the hybrid model is evaluated:

[0032] The root mean square error (Ermse) can be obtained from formula 1:

[0033]

[0034] Accuracy (C R )From formula 2, we can get:

[0035] C R =1-E rmse Formula 2;

[0036] Qualified rate (Q R ) From formula 3 and formula 4, we can get:

[0037]

[0038] S6: Model feature importance analysis:

[0039] During the model construction and performance evaluation process, feature importance analysis is also performed to determine the features that have the greatest impact on PV power output prediction.

[0040] Preferably, in step S2, data preprocessing includes:

[0041] S21: Read meteorological variable data, actual power data and basic information of photovoltaic power generation users from the specified file path;

[0042] S22: merging the meteorological variable data with the basic information of the photovoltaic power generation user so as to include the number, installed capacity, longitude and latitude information of the photovoltaic user;

[0043] S23: Convert the timestamp into the date and time format of Shanghai time zone;

[0044] S24: Reset the index of actual power data:

[0045] Convert the actual power data from a multi-layer index to a single-layer index and rename the power column to target; extract the level of the actual power data and convert it to an integer to match it with the timestamp in step S23;

[0046] S25: The processed meteorological variable data is merged with the actual power data, and divided into a training set, a validation set, and a test set; the data of users f1-f7 are used as the training set and the validation set, and the data of users f8 and f9 are used for the final test evaluation of the model.

[0047] Preferably, in step S3,

[0048] The meteorological feature processing: converting wind speed and wind direction features to capture the impact of wind on photovoltaic power generation;

[0049] The historical value feature extraction: by performing differential processing on the radiation intensity and the temperature, the historical value features of the radiation intensity and the temperature are extracted;

[0050] The outlier processing: detecting and processing outliers in the data to reduce the sensitivity of the model to abnormal data;

[0051] Extraction of the ratio of the light intensity to the strongest light intensity of the day: Calculate the ratio of the irradiance intensity at each time point in the photovoltaic power generation output data to the strongest light intensity of the day;

[0052] The average light feature extraction of the day: the average irradiation intensity of each photovoltaic user in the daytime / nighttime state is calculated by grouping and aggregation, so as to obtain the average light condition of each photovoltaic user in a specific time period;

[0053] The difference feature between the temperature and the highest and lowest temperatures of the day is extracted: the difference between the temperature at each time point in the photovoltaic power generation output data and the highest and lowest temperatures of the day is calculated.

[0054] Preferably, in step S5, an ablation experiment is performed on the data of users f8 and f9 to verify the importance of each component in the fusion model; the ablation experiment is to evaluate the impact on the overall performance by removing or replacing key components in the model.

[0055] Preferably, the data set for photovoltaic power generation output prediction in step S1 is provided by the Digital China Innovation Competition.

[0056] Preferably, the evaluation index in the model performance evaluation in step S5 refers to the "Technical Requirements for Wind Power or Photovoltaic Power Forecasting System on the Dispatching Side".

[0057] In a second aspect, the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; the processor executes the aforementioned method for establishing a hybrid model for predicting ultra-short-term photovoltaic power generation output.

[0058] In a third aspect, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, a method for establishing a hybrid model based on predicting ultra-short-term photovoltaic power generation output is provided.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] (1) The present invention improves the prediction accuracy and reduces the prediction error through model fusion technology. Secondly, by training and testing on the data of users f8 and f9, the fusion model shows good generalization ability, which means that the model can adapt to the prediction of photovoltaic power output under different times, locations and conditions, reducing the risk of overfitting.

[0061] (2) From the perspective of economic and environmental benefits, the present invention can reduce the assessment costs caused by low prediction accuracy and improve the economic benefits of photovoltaic power plants. At the same time, the model helps promote the large-scale application of photovoltaic power generation, reduce dependence on fossil fuels, promote the use of renewable energy, and has positive significance for environmental protection.

[0062] (3) In practical applications, the model proposed in this invention can support grid dispatching and photovoltaic power station management, and improve the stability and reliability of the grid. Photovoltaic power station operators can use this model to formulate effective operation and maintenance strategies, reduce operating costs, and improve economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a technical flow chart of the present invention;

[0064] Figure 2 This is the distribution diagram of the top 5 important features of the fusion model. DETAILED DESCRIPTION

[0065] In order to provide a further understanding of the purpose, structure, features, and functions of the present invention, the following detailed description is given in conjunction with the embodiments.

[0066] Please refer to Figure 1 The present invention provides a hybrid model establishment method for predicting ultra-short-term photovoltaic power generation output, and the steps of the hybrid model establishment method are as follows:

[0067] S1: Dataset collection raw data:

[0068] The original data includes: basic information of photovoltaic power generation users, meteorological variable data and actual power data;

[0069] The basic information of photovoltaic power generation users includes the location and installed capacity information of 9 photovoltaic power generation users, namely users f1-f9;

[0070] The meteorological variable data are meteorological data of 9 photovoltaic power generation users from January 2022 to January 2024, every 15 minutes, including light intensity, temperature, humidity, and wind speed;

[0071] The actual power data is the power output data of 9 photovoltaic power generation users from January 2022 to January 2024, every 15 minutes;

[0072] S2: Data preprocessing:

[0073] Performing data preprocessing on the raw data collected in step S1;

[0074] Data preprocessing includes merging data sets; dividing training sets, validation sets, and test sets; resetting the index of actual power data;

[0075] Through the above data preprocessing, the original data is converted into a format suitable for use by the prediction model and contains the necessary features and target variables;

[0076] S3: feature engineering processing;

[0077] The raw data was subjected to in-depth feature engineering;

[0078] Including: meteorological feature processing, historical value feature extraction, abnormal value processing, feature extraction of the ratio of light to the strongest light of the day, feature extraction of the average light of the day, and feature extraction of the difference between temperature and the highest and lowest temperatures of the day;

[0079] Through these feature engineering processing steps, the original data is converted into a format that contains rich information and is suitable for use in prediction models, providing a data basis for the construction and training of photovoltaic power generation output prediction models;

[0080] S4: Constructing a hybrid model for predicting ultra-short-term photovoltaic power generation output:

[0081] By combining three advanced gradient boosting algorithms, LightGBM, XGBoost and CatBoost, accurate prediction of photovoltaic power generation output can be achieved;

[0082] LightGBM reduces computation and memory usage by using gradient-based unilateral sampling and histogram algorithms; XGBoost introduces regularization terms to control the complexity of the model; and CatBoost improves prediction accuracy through improved ranking boosting and symmetric binary tree structure.

[0083] S41: Model parameter configuration:

[0084] Parameters were configured for LightGBM, XGBoost, and CatBoost algorithms, including learning rate, number of trees, and tree depth;

[0085] By configuring parameters for three advanced gradient boosting algorithms, the performance of the model during training is ensured.

[0086] LightGBM: LightGBM is an efficient machine learning algorithm based on the gradient boosting framework. It reduces the amount of computation and memory usage by using gradient-based one-side sampling and histogram algorithms. LightGBM performs well in processing large data sets, converges quickly and provides high prediction accuracy.

[0087] XGBoost: XGBoost is another popular gradient boosting algorithm that introduces a regularization term to control the complexity of the model, thereby improving the generalization ability of the model and is widely used in various prediction tasks.

[0088] CatBoost: CatBoost is a relatively new gradient boosting algorithm that improves prediction accuracy through improved sorting boosting and a symmetric binary tree structure; CatBoost is particularly suitable for processing categorical features and is robust to noise and outliers in the data.

[0089] The key parameters of the three algorithms are shown in Table 1:

[0090]

[0091]

[0092] S42: Training-Cross Validation:

[0093] The performance of the photovoltaic power generation output prediction model can be further improved through training-cross validation.

[0094] Use the five-fold cross-validation method to divide the dataset into five folds, each of which contains a portion of training data and validation data;

[0095] For each fold, the training dataset is used to train the models of three algorithms: LightGBM, XGBoost, and CatBoost;

[0096] Evaluate model performance on the validation dataset for each fold;

[0097] The root mean square error (RMSE) of each fold is calculated as the performance evaluation indicator, and the final evaluation result is the average RMSE of multiple folds.

[0098] S43: Model Fusion:

[0099] The model fusion technique was used to average the prediction results of the three models;

[0100] The present invention adopts model fusion technology to average the prediction results of the three models. Through model fusion technology, the photovoltaic power generation output prediction model proposed by the present invention can integrate the advantages of the three models. This method can not only improve the accuracy of the prediction, but also reduce the risk of overfitting and improve the generalization ability of the model under different conditions.

[0101] S5: Model performance evaluation:

[0102] The performance of the photovoltaic power generation output prediction model proposed in this invention is verified by performance evaluation. The data of users f8 and f9 are mainly used to evaluate the model. The evaluation indicators refer to the requirements in the Technical Requirements for Wind Power or Photovoltaic Power Prediction System on the Dispatching Side, and mainly include the following indicators:

[0103] After the hybrid model is constructed in step S4, the hybrid model is evaluated:

[0104] The root mean square error (Ermse) can be obtained from formula 1:

[0105]

[0106] Where: n: the number of all samples; P Pi : actual power at time i;

[0107] P Mi : predicted power at time i; C i : The boot capacity at time i.

[0108] Accuracy (C R )From formula 2, we can get:

[0109] C R =1-E rmse Formula 2.

[0110] Qualified rate (Q R ) From formula 3 and formula 4, we can get:

[0111]

[0112] Among them: Bi : The prediction qualification rate judgment result at time i; n: The number of all samples;

[0113] P Pi : actual power at time i; P Mi : predicted power at time i;

[0114] C i : The boot capacity at time i.

[0115] According to the provisions of the Technical Requirements for Wind Power or Photovoltaic Power Forecasting System on the Dispatching Side, the present invention strictly evaluates the average accuracy of ultra-short-term power forecasting of photovoltaic power stations.

[0116] According to the requirements in Table 2, the average accuracy of ultra-short-term power forecasting must reach more than 90%. In order to verify the performance of the photovoltaic power generation output prediction model proposed in this invention, we used the 16-month photovoltaic power generation data of users f8 and f9 to conduct a detailed model evaluation.

[0117] Table 2 Power prediction performance index requirements of photovoltaic power stations

[0118]

[0119]

[0120] The evaluation results show that the model accuracy of user f8 reached 95.16%, and the pass rate was as high as 99.81%, indicating that the model performed excellently on user f8's data and was able to accurately predict photovoltaic power generation output.

[0121] Similarly, the model accuracy of user f9 is 93.27% and the pass rate is 98.76%, which also shows that the model has high prediction accuracy on the data of user f9.

[0122] S6: Model feature importance analysis:

[0123] Reference Figure 2 ,During the process of model building and performance evaluation, feature ,importance analysis is also conducted to determine the features that have the greatest ,impact on the prediction of PV power output.

[0124] Figure 2 These are the top five important features obtained after analyzing the fusion model. Through feature importance analysis, the present invention determines the key features that affect photovoltaic power generation output prediction, providing a basis for subsequent model optimization and feature selection.

[0125] Preferably, in step S2, data preprocessing includes:

[0126] S21: Read meteorological variable data, actual power data and basic information of photovoltaic power generation users from the specified file path;

[0127] S22: merging the meteorological variable data with the basic information of the photovoltaic power generation user so as to include the number, installed capacity, longitude and latitude information of the photovoltaic user;

[0128] S23: Convert the timestamp into the date and time format of Shanghai time zone;

[0129] To ensure the consistency and accuracy of time;

[0130] S24: Reset the index of actual power data:

[0131] Convert the actual power data from a multi-layer index to a single-layer index and rename the power column to target; extract the level of the actual power data and convert it to an integer to match it with the timestamp in step S23;

[0132] By adjusting the timestamp to match the sampling interval of the actual power data, data consistency and accuracy are ensured.

[0133] S25: The processed meteorological variable data is merged with the actual power data, and divided into a training set, a validation set, and a test set; the data of users f1-f7 are used as the training set and the validation set, and the data of users f8 and f9 are used for the final test evaluation of the model.

[0134] Preferably, in step S3,

[0135] The meteorological feature processing: converting wind speed and wind direction features to capture the impact of wind on photovoltaic power generation;

[0136] The ratio of light to temperature is used as a feature to reflect the impact of light conditions on photovoltaic power generation.

[0137] The historical value feature extraction: by performing differential processing on the radiation intensity and the temperature, the historical value features of the radiation intensity and the temperature are extracted;

[0138] The impact of historical data on photovoltaic power generation output can be captured through historical value feature extraction.

[0139] The outlier processing: detects and processes outliers in the data to reduce the sensitivity of the model to abnormal data.

[0140] Extraction of the ratio of the light intensity to the strongest light intensity of the day: Calculate the ratio of the irradiance intensity at each time point in the photovoltaic power generation output data to the strongest light intensity of the day;

[0141] This ratio feature reflects the degree of change in lighting conditions at the current time point relative to the optimal lighting conditions of the day, and is used as an input feature of the model to enhance the model's sensitivity to changes in lighting conditions.

[0142] The average light feature extraction of the day: the average irradiation intensity of each photovoltaic user in the daytime / nighttime state is calculated by grouping and aggregation, so as to obtain the average light condition of each photovoltaic user in a specific time period;

[0143] This feature reflects the average light conditions of PV users in different time periods and is used as an input feature of the model to enhance the model’s sensitivity to changes in light conditions.

[0144] The difference feature between the temperature and the highest and lowest temperatures of the day is extracted: the difference between the temperature at each time point in the photovoltaic power generation output data and the highest and lowest temperatures of the day is calculated.

[0145] This difference feature reflects the degree of change of the temperature conditions at the current time point relative to the temperature change range of the day, and is used as an input feature of the model to enhance the model's sensitivity to changes in temperature conditions.

[0146] Preferably, in step S5, an ablation experiment is performed on the data of users f8 and f9 to verify the importance of each component in the fusion model; the ablation experiment is to evaluate the impact on the overall performance by removing or replacing key components in the model; these experimental results support the effectiveness of model fusion technology in improving prediction accuracy and model generalization ability.

[0147] Table 3 Ablation experiments and results

[0148]

[0149] In summary, the fusion model proposed in the present invention has met the forecast index requirements in the Technical Requirements for Wind Power or Photovoltaic Power Forecasting System on the Dispatching Side for both user f8 and f9 data, and the average accuracy of ultra-short-term power forecasting exceeds 90%. This further proves that the hybrid model proposed in the present invention has high accuracy and reliability in predicting photovoltaic power generation output, and can provide strong support for power grid dispatching and photovoltaic power station management.

[0150] Ablation experiments of single models and combinations of multiple models were conducted to explore the performance of different machine learning models and their fusion models.

[0151] First, three different machine learning models - LightGBM, XGBoost and CatBoost - were used separately, trained and tested under the same conditions, and the average accuracy and average pass rate of each model on the data set were obtained; then, these models were combined in pairs to observe whether model fusion can improve the overall performance. According to the table data, there are three model combinations: LightGBM+XGBoost, LightGBM+CatBoost and XGBoost+CatBoost. The average accuracy and average pass rate of each combined model exceeded those of a single model; finally, all the models were fused together, namely LightGBM+XGBoost+CatBoost, and the highest average accuracy and average pass rate were obtained, indicating that this model fusion does bring performance improvements.

[0152] The LightGBM, XGBoost and CatBoost models used in the present invention have made significant progress in photovoltaic power generation output prediction, but they each have limitations. For example, LightGBM may not be accurate enough when dealing with predictions under extreme weather conditions; XGBoost requires a lot of expertise in feature selection and model parameter adjustment; CatBoost may encounter computational efficiency issues when dealing with high-dimensional data. In addition, a single model may not be able to capture all the complex factors that affect photovoltaic output, resulting in unsatisfactory prediction performance in some cases.

[0153] Therefore, to overcome these limitations, this study proposes a hybrid model that combines the advantages of LightGBM, XGBoost and CatBoost. By integrating these models, the present invention aims to provide a more accurate and robust photovoltaic power output forecasting solution, thereby providing more effective support for grid dispatching and photovoltaic power station management.

[0154] The hybrid model of the present invention can not only improve the accuracy of prediction, but also reduce the risk of overfitting through model fusion technology and improve the generalization ability of the model under different conditions. In addition, by optimizing the training process and parameter selection of the model, the present invention can also improve the calculation efficiency of the prediction and meet the actual needs of ultra-short-term photovoltaic power generation output prediction.

[0155] In summary, the hybrid model of the present invention represents an innovative advancement in photovoltaic power generation output prediction technology. It combines the advantages of multiple machine learning algorithms and is expected to provide strong technical support for the development of the photovoltaic industry and the stable operation of the power grid.

[0156] The present invention has been described by the above-mentioned relevant embodiments, however, the above-mentioned embodiments are only examples for implementing the present invention. It must be pointed out that the disclosed embodiments do not limit the scope of the present invention. On the contrary, changes and modifications made without departing from the spirit and scope of the present invention are all within the scope of patent protection of the present invention.

Claims

1. A hybrid modeling method for predicting ultra-short-term photovoltaic power generation output, characterized by: The steps of the hybrid model establishment method are as follows: S1: Dataset collection raw data: The original data includes: basic information of photovoltaic power generation users, meteorological variable data and actual power data; The basic information of photovoltaic power generation users includes the location and installed capacity information of 9 photovoltaic power generation users, namely users f1-f9; The meteorological variable data are meteorological data of 9 photovoltaic power generation users from January 2022 to January 2024, every 15 minutes, including light intensity, temperature, humidity, and wind speed; The actual power data is the power output data of 9 photovoltaic power generation users from January 2022 to January 2024, every 15 minutes; S2: Data preprocessing: Performing data preprocessing on the raw data collected in step S1; Data preprocessing includes merging data sets; dividing training sets, validation sets, and test sets; resetting the index of actual power data; Through the above data preprocessing, the original data is converted into a format suitable for use by the prediction model and contains the necessary features and target variables; S3: feature engineering processing; The raw data was subjected to in-depth feature engineering; Including: meteorological feature processing, historical value feature extraction, abnormal value processing, feature extraction of the ratio of light to the strongest light of the day, feature extraction of the average light of the day, and feature extraction of the difference between temperature and the highest and lowest temperatures of the day; Through these feature engineering processing steps, the original data is converted into a format that contains rich information and is suitable for use in prediction models, providing a data basis for the construction and training of photovoltaic power generation output prediction models; S4: Constructing a hybrid model for predicting ultra-short-term photovoltaic power generation output: By combining three advanced gradient boosting algorithms, LightGBM, XGBoost and CatBoost, accurate prediction of photovoltaic power generation output can be achieved; S41: Model parameter configuration: Parameters were configured for LightGBM, XGBoost, and CatBoost algorithms, including learning rate, number of trees, and tree depth; S42: Training-Cross Validation: Use the five-fold cross-validation method to divide the dataset into five folds, each of which contains a portion of training data and validation data; For each fold, the training dataset is used to train the models of three algorithms: LightGBM, XGBoost, and CatBoost; Evaluate model performance on the validation dataset for each fold; The root mean square error (RMSE) of each fold is calculated as the performance evaluation indicator, and the final evaluation result is the average RMSE of multiple folds; S43: Model Fusion: The model fusion technique was used to average the prediction results of the three models; S5: Model performance evaluation: After the hybrid model is constructed in step S4, the hybrid model is evaluated: The root mean square error (Ermse) can be obtained from formula 1: Accuracy (C R )From formula 2, we can get: C R =1-E rmse Formula 2; Qualified rate (Q R ) From formula 3 and formula 4, we can get: S6: Model feature importance analysis: During the model construction and performance evaluation process, feature importance analysis is also performed to determine the features that have the greatest impact on PV power output prediction.

2. A hybrid model establishment method for predicting ultra-short-term photovoltaic power generation output according to claim 1, characterized in that: In step S2, data preprocessing includes: S21: Read meteorological variable data, actual power data and basic information of photovoltaic power generation users from the specified file path; S22: merging the meteorological variable data with the basic information of the photovoltaic power generation user so as to include the number, installed capacity, longitude and latitude information of the photovoltaic user; S23: Convert the timestamp into the date and time format of Shanghai time zone; S24: Reset the index of actual power data: Convert the actual power data from a multi-layer index to a single-layer index and rename the power column to target; extract the level of the actual power data and convert it to an integer to match it with the timestamp in step S23; S25: The processed meteorological variable data is merged with the actual power data, and divided into a training set, a validation set, and a test set; the data of users f1-f7 are used as the training set and the validation set, and the data of users f8 and f9 are used for the final test evaluation of the model.

3. A hybrid model establishment method for predicting ultra-short-term photovoltaic power generation output according to claim 1, characterized in that: In the step S3, The meteorological feature processing: converting wind speed and wind direction features to capture the impact of wind on photovoltaic power generation; The historical value feature extraction: by performing differential processing on the radiation intensity and the temperature, the historical value features of the radiation intensity and the temperature are extracted; The outlier processing: detecting and processing outliers in the data to reduce the sensitivity of the model to abnormal data; Extraction of the ratio of the light intensity to the strongest light intensity of the day: Calculate the ratio of the irradiance intensity at each time point in the photovoltaic power generation output data to the strongest light intensity of the day; The average light feature extraction of the day: the average irradiation intensity of each photovoltaic user in the daytime / nighttime state is calculated by grouping and aggregation, so as to obtain the average light condition of each photovoltaic user in a specific time period; The difference feature between the temperature and the highest and lowest temperatures of the day is extracted: the difference between the temperature at each time point in the photovoltaic power generation output data and the highest and lowest temperatures of the day is calculated.

4. A hybrid model establishment method for predicting ultra-short-term photovoltaic power generation output according to claim 1, characterized in that: In step S5, an ablation experiment is performed on the data of users f8 and f9 to verify the importance of each component in the fusion model; the ablation experiment is to evaluate the impact on the overall performance by removing or replacing key components in the model.

5. A hybrid model establishment method for predicting ultra-short-term photovoltaic power generation output according to claim 1, characterized in that: The data set for photovoltaic power generation output prediction in step S1 is provided by the Digital China Innovation Competition.

6. A hybrid model establishment method for predicting ultra-short-term photovoltaic power generation output according to claim 1, characterized in that: The evaluation indicators in the model performance evaluation in step S5 refer to the "Technical Requirements for Wind Power or Photovoltaic Power Forecasting System on the Dispatching Side".

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; characterized in that: The processor executes the hybrid model establishment method based on predicting ultra-short-term photovoltaic power generation output as described in any one of claims 1-6.

8. A readable storage medium storing a computer program; characterized in that: When the computer program is executed by a processor, it is used to implement the hybrid model establishment method based on predicting ultra-short-term photovoltaic power generation output as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Multi-element regression prediction method and device for ultra-short-term power load

    CN117374917A

  • Photovoltaic system power prediction model construction method and device and photovoltaic system power prediction method and device

    CN117913819A

Cited By

  • Method and device for improving robustness of photovoltaic prediction of multi-task space-time diagram fusion model

    CN120449932A