Short-term wind power prediction method based on Frandsen wake flow model

Abstract

The invention provides a short-term wind power prediction method based on a Frandsen wake flow model. The short-term wind power prediction method comprises the following steps: S1, collecting data; s2, making a wake effect database of the wind power plant based on a Frandsen wake model algorithm; s3, the predicted wind speed and wind direction are introduced into a wake effect database, and downscaling processing of the predicted wind speed in consideration of the wake effect is achieved; s4, forming a data set for a plurality of newly generated downscaling wind speed features, and performing recursive feature elimination; s5, based on the remaining wind power features after elimination, model training and model stacking are carried out through multiple deep learning algorithms; and S6, performing linear combination on the trained multi-class models to form a final short-term wind power prediction model. The wake flow model and the deep learning technology are creatively fused, the predication precision is improved, meanwhile, the interpretability and robustness of the model are enhanced, dependence on historical data and artificial features is reduced, and the method has remarkable technical advantages and application prospects.

Description

Technical Field

[0001] The invention relates to the technical field of short-term wind power prediction, and in particular to a short-term wind power prediction method based on a Frandsen wake model. Background Art

[0002] With the rapid development of the wind power industry, wind power prediction technology plays a vital role in wind farm planning, operation and maintenance management, and power system scheduling. At present, wind power prediction technology mainly adopts methods based on machine learning and deep learning. For example, Chinese patent CN104700156A proposes a wind power prediction method based on support vector machine screening variables, which has high accuracy. However, this method relies on a large amount of high-quality historical data, and data noise and missing data may significantly affect the prediction effect. In addition, Chinese patent CN114819369A mentions a short-term wind power prediction method based on two-stage feature selection and random forest improved model, which is suitable for nonlinear features, but the computational overhead is high when the data scale is large, and the model training time is long. These existing technologies have improved the accuracy of wind power prediction to a certain extent, but they rely heavily on historical data and feature engineering, and the generalization ability of the model still needs to be improved. In addition, in the face of the changing wind farm environment, the adaptability of existing methods is poor and cannot effectively deal with problems such as data noise and missing data. Based on this, the present invention proposes an improved wind power prediction model, which aims to reduce the dependence on historical data, improve prediction accuracy and generalization ability by optimizing the data processing method and model architecture, so as to better adapt to the environment of different wind farms. Summary of the invention

[0003] The present invention provides a short-term wind power forecasting method based on the Frandsen wake model, which comprehensively utilizes the wake model and deep learning technology to improve the forecasting accuracy while reducing the dependence on historical data.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: A short-term wind power forecasting method based on the Frandsen wake model includes the following methods: S1. Collect the actual power of the wind farm and the single-point meteorological forecast data of the wind farm area, including the predicted wind speed, predicted wind direction and other meteorological forecast data; collect the technical parameters of each wind turbine in the wind farm, including blade diameter, hub height, thrust coefficient curve and relative position; S2. Based on the Frandsen wake model algorithm, using the thrust coefficient curve of the wind turbine, and based on different wind speeds and wind directions, a database of the wake effect of the wind farm is created; S3, using the predicted wind speed and wind direction to import the wake effect database, to achieve downscaling of the predicted wind speed taking into account the wake effect; S4. Combining the newly generated downscaled wind speed features with the existing forecast meteorological data and actual wind power generation to form a data set, and performing recursive feature elimination; S5. Based on the remaining wind power characteristics after elimination, model training and model stacking are performed through a variety of deep learning algorithms; S6. Linearly combine the trained multi-class models to form the final short-term wind power prediction model. The above-mentioned S2 specifically includes: S21. The Frandsen wake model of the wind farm is established. The wake model assumes that the flow has self-similarity, sets the wind speed to have a fixed profile, applies the momentum equation to the control volume, and obtains the wake attenuation equation related to the axial distance x. The velocity attenuation coefficient of the Frandsen wake model can be calculated as: ; The diameter of the downstream wake influence area of ​​the wind turbine can be calculated as: ; In the formula, is the blade diameter of the upstream fan; is the swept area of ​​the impeller of the upstream fan; is the area of ​​the downstream wake impact area of ​​the wind turbine; is the thrust coefficient of the fan; is the wake attenuation factor of the Frandsen wake model; is the wake expansion parameter; k is the empirical coefficient; parameter It can be calculated by the following formula: ; ; In the formula, is the wake attenuation constant, which is used to reflect the lateral diffusion speed of the wake; S22. Input the collected wind speed and wind direction data into the wake effect model, calculate the wake effect of each wind turbine under different wind speeds and wind directions, and organize the wake effect data obtained by simulation calculation into a database format, which contains wind speed, wind direction, thrust coefficient, relative position and wake effect information.

[0005] The above-mentioned S4 specifically includes: S41. Integrate the newly generated downscaled wind speed characteristics, forecasted meteorological data and wind power actual power data to construct a data set; S42. Perform feature engineering on the data, including missing value processing and data normalization, to ensure data quality and consistency; S43. Perform recursive feature elimination on the data to select the best feature subset by recursively training the model and eliminating features that have little impact on model performance.

[0006] The above-mentioned S43 specifically includes: S431, using all features to train a random forest model, and calculating the importance of each feature to the model performance; S432, removing a feature with the lowest feature importance from the current feature subset to obtain a new feature subset, inputting it into the random forest model again, calculating the importance of each feature in the new feature subset, and using a cross-validation method to obtain the prediction accuracy of the new feature subset; S433. When the prediction accuracy of the random forest model no longer improves, the recursive feature elimination process is stopped, and the feature subset finally obtained is the best feature set selected by the recursive feature elimination method.

[0007] The above-mentioned S5 specifically includes: S51, dividing the data set into a training set, a validation set and a test set; S52, training multiple basic models on the training set respectively, and optimizing the hyperparameters of each model through grid search; S53. Use the validation set to evaluate the performance of each basic model and fine-tune the model based on the evaluation results; S54, predicting multiple trained basic models on the validation set, and using the prediction results of each model as a new feature; S55. Train a meta-model based on the new features. The meta-model can learn the weights of each basic model to achieve weighted integration. S56. Use the test set to evaluate the performance of the stacked ensemble model and use it for subsequent online predictions.

[0008] The present invention provides a short-term wind power prediction method based on the Frandsen wake model, which has the following beneficial effects: (1) Downscaling of predicted wind speeds is achieved through the Frandsen wake model, which can more finely characterize the wake effect inside the wind farm. Based on physical prior knowledge, the robustness of the prediction model is enhanced and the prediction accuracy is improved.

[0009] (2) Recursive feature elimination is used to automatically screen out key wind power features, reducing the reliance on manual feature engineering.

[0010] (3) Utilize deep learning algorithms and model integration technology to fully explore the nonlinear relationship of wind power data and enhance the generalization ability of the model.

[0011] (4) Linear combination of multi-class models can adaptively adjust weights according to the predictive capabilities of different models to further improve the predictive performance.

[0012] In summary, the present invention provides a novel wind power prediction algorithm, which, by innovatively integrating the wake model and deep learning technology, improves the prediction accuracy while enhancing the interpretability and robustness of the model, reduces the dependence on historical data and artificial features, and has significant technical advantages and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 This is the overall flow chart of short-term wind power forecasting based on the Frandsen wake model; Figure 2 It is a flowchart of recursive feature elimination based on random forest model; Figure 3 It is a flowchart for establishing an integrated model based on the stacking of multiple deep learning models. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solutions and advantages of the present invention clearer, the following content will systematically and completely describe the specific technical solutions of the present invention in combination with the drawings provided according to the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0015] Embodiment 1: A short-term wind power forecasting method based on the Frandsen wake model includes the following methods: S1. Collect the actual power of the wind farm and the single-point meteorological forecast data of the wind farm area, including the predicted wind speed, predicted wind direction and other meteorological forecast data; collect the technical parameters of each wind turbine in the wind farm, including blade diameter, hub height, thrust coefficient curve and relative position; Meteorological data such as wind speed and wind direction are the core factors affecting the power output of wind farms. Accurate meteorological data collection can provide accurate input data for the calculation of wake models and subsequent predictions. This step ensures the quality of input data for wind power prediction models, reduces the impact of data errors on prediction results, and improves the reliability of prediction models. S2. Based on the Frandsen wake model algorithm, using the thrust coefficient curve of the wind turbine, and based on different wind speeds and wind directions, a database of the wake effect of the wind farm is created; The Frandsen wake model assumes that the wind speed profile is fixed and calculates the wake attenuation through the momentum equation. The wake effect database is created based on the thrust coefficient, wind speed and wind direction to capture the influence of wake on wind speed under different conditions. The generated wake effect database provides a basis for downscaling processing in subsequent steps. By considering the wake effect, the prediction is made more consistent with the flow characteristics of the actual wind field, thereby enhancing the accuracy of the prediction. S3, using the predicted wind speed and wind direction to import the wake effect database, to achieve downscaling of the predicted wind speed taking into account the wake effect; By downscaling the original wind speed, the wake effect is incorporated into the wind speed change, making the predicted wind speed closer to the actual wind conditions in the wind farm; the downscaled wind speed characteristics can better reflect the actual wind turbine operating environment, providing more representative wind speed characteristics for the input of subsequent prediction models; S4. Combining the newly generated downscaled wind speed features with the existing forecast meteorological data and actual wind power generation to form a data set, and performing recursive feature elimination; Recursive feature elimination can filter out the features that are most critical to the model's prediction effect, thereby reducing the noise and computational overhead caused by redundant features; it can improve the efficiency of model training, enhance the generalization ability of the model, avoid overfitting, and improve prediction accuracy; S5. Based on the remaining wind power characteristics after elimination, model training and model stacking are performed through a variety of deep learning algorithms; Through the training and stacking integration of multiple deep learning models, the advantages of each model in different data features are utilized to optimize the overall prediction effect; the prediction accuracy of the integrated model is better than that of a single model, which helps to improve the accuracy and robustness of wind power prediction; S6, linearly combining the trained multi-class models to form a final short-term wind power prediction model; The linear combination model integrates the prediction results of each basic model in a weighted manner to obtain a more robust prediction output; the integrated model after linear combination further improves the prediction effect, making the short-term wind power prediction model more accurate and reliable in practical applications. The above-mentioned S2 specifically includes: S21. The Frandsen wake model of the wind farm is established. The wake model assumes that the flow has self-similarity, sets the wind speed to have a fixed profile, applies the momentum equation to the control volume, and obtains the wake attenuation equation related to the axial distance x. The velocity attenuation coefficient of the Frandsen wake model can be calculated as: ; The diameter of the downstream wake influence area of ​​the wind turbine can be calculated as: ; In the formula, is the blade diameter of the upstream fan; is the swept area of ​​the impeller of the upstream fan; is the area of ​​the downstream wake impact area of ​​the wind turbine; is the thrust coefficient of the fan; is the wake attenuation factor of the Frandsen wake model; is the wake expansion parameter; k is the empirical coefficient, which is 2; parameter It can be calculated by the following formula: ; ; In the formula, is the wake attenuation constant, which is used to reflect the lateral diffusion speed of the wake. Its value is generally set to 0.075 in onshore wind farms; The determination of wake attenuation and wake influence area enables the model to reasonably reflect the diffusion and attenuation of wind turbine wake; the determination of the scope of the wake influence area can accurately simulate the mutual influence between wind turbines in the wind farm, providing reliable theoretical support for subsequent model calculations; S22, inputting the collected wind speed and wind direction data into the wake effect model, calculating the wake effect of each wind turbine under different wind speeds and wind directions, and arranging the wake effect data obtained by simulation calculation into a database format, wherein the database includes wind speed, wind direction, thrust coefficient, relative position and wake effect information; The wake effect data are summarized to build a database containing information such as wind speed, wind direction, thrust coefficient, relative position and wake effect; the database serves as a quick search tool for wake effects and provides efficient and accurate wake impact data support for subsequent steps.

[0016] The above-mentioned S4 specifically includes: S41. Integrate the newly generated downscaled wind speed characteristics, forecasted meteorological data and wind power actual power data to construct a data set; S42. Perform feature engineering on the data, including missing value processing and data normalization, to ensure data quality and consistency; S43. Perform recursive feature elimination on the data to select the best feature subset by recursively training the model and eliminating features that have little impact on model performance.

[0017] The above-mentioned S43 specifically includes: S431, using all features to train a random forest model, and calculating the importance of each feature to the model performance; S432, removing a feature with the lowest feature importance from the current feature subset to obtain a new feature subset, inputting it into the random forest model again, calculating the importance of each feature in the new feature subset, and using a cross-validation method to obtain the prediction accuracy of the new feature subset; S433. When the prediction accuracy of the random forest model is no longer improved, the recursive feature elimination process is stopped, and the feature subset finally obtained is the best feature set selected by the recursive feature elimination method; Optimize data quality through feature engineering and recursive feature elimination to ensure that the feature input of the prediction model is reasonable; the selection of the optimal feature subset improves the computational efficiency of the model and makes the final prediction more accurate.

[0018] The above-mentioned S5 specifically includes: S51, dividing the data set into a training set, a validation set and a test set; S52. Train multiple basic models on the training set, such as LSTM, CNN, MLP, etc., and optimize the hyperparameters of each model through grid search; S53. Use the validation set to evaluate the performance of each basic model and fine-tune the model based on the evaluation results; S54, predicting multiple trained basic models on the validation set, and using the prediction results of each model as a new feature; S55. Train a meta-model based on the new features. Usually, a linear model such as Logistic regression or Ridge regression is selected. The meta-model can learn the weights of each basic model to achieve weighted integration. S56. Use the test set to evaluate the performance of the stacked ensemble model and use it for subsequent online predictions; Through multi-model training, stacking integration and weighted combination, the model can fully utilize the advantages of multiple algorithms; the model prediction performance is improved, so that the accuracy and robustness of the wind power prediction model proposed in the present invention in short-term prediction are significantly enhanced.

Claims

1. A short-term wind power forecasting method based on the Frandsen wake model, characterized in that: Includes the following methods: S1. Collect the actual power of the wind farm and the single-point meteorological forecast data of the wind farm area, including the forecast data of wind speed and wind direction; collect the technical parameters of each wind turbine in the wind farm, including blade diameter, hub height, thrust coefficient curve and relative position; S2. Based on the Frandsen wake model algorithm, using the thrust coefficient curve of the wind turbine, and based on different wind speeds and wind directions, a database of the wake effect of the wind farm is created; S3, using the predicted wind speed and wind direction to import the wake effect database, to achieve downscaling of the predicted wind speed taking into account the wake effect; S4. Combining the newly generated downscaled wind speed features with the existing forecast meteorological data and actual wind power generation to form a data set, and performing recursive feature elimination; S5. Based on the remaining wind power characteristics after elimination, model training and model stacking are performed through a variety of deep learning algorithms; S6. Linearly combine the trained multi-class models to form the final short-term wind power prediction model.

2. A short-term wind power prediction method based on the Frandsen wake model according to claim 1, characterized in that: The S2 specifically includes: S21. The Frandsen wake model of the wind farm is established. The wake model assumes that the flow has self-similarity, sets the wind speed to have a fixed profile, applies the momentum equation to the control volume, and obtains the wake attenuation equation related to the axial distance x. The velocity attenuation coefficient of the Frandsen wake model can be calculated as: ； The diameter of the downstream wake influence area of ​​the wind turbine can be calculated as: ； In the formula, is the blade diameter of the upstream fan; is the swept area of ​​the impeller of the upstream fan; is the area of ​​the downstream wake impact area of ​​the wind turbine; is the thrust coefficient of the fan; is the wake attenuation factor of the Frandsen wake model; is the wake expansion parameter; k is the empirical coefficient; parameter It can be calculated by the following formula: ； ； In the formula, is the wake attenuation constant, which is used to reflect the lateral diffusion speed of the wake; S22. Input the collected wind speed and wind direction data into the wake effect model, calculate the wake effect of each wind turbine under different wind speeds and wind directions, and organize the wake effect data obtained by simulation calculation into a database format, which contains wind speed, wind direction, thrust coefficient, relative position and wake effect information.

3. A short-term wind power prediction method based on the Frandsen wake model according to claim 2, characterized in that: The S4 specifically includes: S41. Integrate the newly generated downscaled wind speed characteristics, forecasted meteorological data and wind power actual power data to construct a data set; S42. Perform feature engineering on the data, including missing value processing and data normalization, to ensure data quality and consistency; S43. Perform recursive feature elimination on the data to select the best feature subset by recursively training the model and eliminating features that have little impact on model performance.

4. A short-term wind power prediction method based on the Frandsen wake model according to claim 3, characterized in that: The S43 specifically includes: S431, using all features to train a random forest model, and calculating the importance of each feature to the model performance; S432, removing a feature with the lowest feature importance from the current feature subset to obtain a new feature subset, inputting it into the random forest model again, calculating the importance of each feature in the new feature subset, and using a cross-validation method to obtain the prediction accuracy of the new feature subset; S433. When the prediction accuracy of the random forest model no longer improves, the recursive feature elimination process is stopped, and the feature subset finally obtained is the best feature set selected by the recursive feature elimination method.

5. A short-term wind power prediction method based on the Frandsen wake model according to claim 4, characterized in that: The S5 specifically includes: S51, dividing the data set into a training set, a validation set and a test set; S52, training multiple basic models on the training set respectively, and optimizing the hyperparameters of each model through grid search; S53. Use the validation set to evaluate the performance of each basic model and fine-tune the model based on the evaluation results; S54, predicting multiple trained basic models on the validation set, and using the prediction results of each model as a new feature; S55. Train a meta-model based on the new features. The meta-model can learn the weights of each basic model to achieve weighted integration. S56. Use the test set to evaluate the performance of the stacked ensemble model and use it for subsequent online predictions.

Images