Ultra-short-term wind power forecasting method based on time domain characteristics and XGBoost

Abstract

The present invention discloses an ultra-short-term wind power prediction method based on time domain features and XGBoost. According to the influencing factors such as wind speed, wind speed change rate, wind direction, air pressure, temperature, humidity, etc. in the wind field environment, the features are classified and marked by machine learning methods, and normalized and vectorized to be represented so that they have domain feature information in the wind environment space. Then, the machine learning algorithm is used to perform model training on the represented features to find out the wind environment feature set with the greatest impact on wind power and its impact coefficient on wind power. Then, the integrated learning method is used to integrate each feature into the model to predict the ultra-short-term power of wind power. According to the k-fold cross-validation method, a model with the smallest root mean square error and mean absolute error of the prediction result is found, and finally a prediction model is formed. The factors affecting wind power are decomposed from five dimensions and four directions, and the characteristics of the influencing factors are refined and fully expanded and mined.

Description

Technical Field

[0001] The present invention relates to the field of renewable energy power generation, and more specifically to an ultra-short-term wind power prediction method based on time domain characteristics and XGBoost. Background Art

[0002] Wind power has become the third largest power source in China and is one of the most mature power generation methods in the field of new energy power generation, with the most scale development conditions and commercial development prospects. However, wind power is strongly correlated with the wind field environment, especially the wind speed. Since wind speed is volatile and uncertain, the output power of wind turbines is also highly volatile and uncertain. Once the volatile wind power is connected to the grid, it will have a greater impact on the power grid and bring severe challenges to the reliability of the power system.

[0003] As an unstable power source of the power system, wind power has the characteristics of low energy density, uneven distribution, randomness, intermittence and uncontrollability. In order to reduce the risk of low wind power reliability, cope with the peak load regulation and frequency regulation pressure of the power system, achieve more efficient wind power consumption, and alleviate the problems of low power quality, frequent voltage control, difficult active power dispatch and low system stability caused by large-scale wind power grid connection, it is becoming increasingly important to effectively improve the load forecasting accuracy of wind power generation. Accurate wind power forecasting, especially short-term and ultra-short-term power forecasting, is of great significance to the sustainable development and utilization of wind power. It is an important prerequisite and guarantee for the development of new energy industry and the reliable operation and orderly upgrading of large-scale power grids. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention proposes an integrated learning algorithm based on XBboost and wind farm environment time domain mining according to the actual need of increasingly improved wind power prediction accuracy and according to the influencing factors such as wind speed, wind speed change rate, wind direction, temperature, humidity, air pressure, etc. in the wind farm environment. By classifying and screening the time domain data, the environmental factors with the greatest impact on wind power are obtained, and then ultra-short-term wind power prediction is carried out in combination with historical wind power characteristics.

[0005] An ultra-short-term wind power prediction method based on time domain features and XGBoost, the steps are as follows:

[0006] Step 1: Raw data preprocessing

[0007] According to the influencing factors such as wind speed, wind speed change rate, wind direction, air pressure, temperature, humidity, etc. in the wind field environment, the features are classified, marked, normalized and vectorized through machine learning methods, so that they have domain feature information in the wind environment space;

[0008] The specific steps of the raw data preprocessing include: filling in the missing values ​​in the raw data, which can be done by the average value method and the fixed value method; identifying and deleting duplicate values ​​and redundant values ​​in the data; identifying and correcting erroneous values ​​in the data, which can be done by the outlier recognition method, and then inserting the corrected new values ​​by the missing value filling method; finally, normalizing the data of different dimensions. The commonly used method is to calculate the maximum and minimum values ​​of the data of each dimension and perform 0-1 mapping on the data to obtain the data representation in the 0-1 space.

[0009] Step 2: Data Analysis and Feature Mining

[0010] The machine learning algorithm is used to train the model of the represented features to find the wind environment feature set that has the greatest impact on wind power and its impact coefficient on wind power.

[0011] The feature engineering of data includes: feature identification, feature extraction, feature representation and feature classification of data. The data collected by the wind farm includes meteorological data, historical power data, status data of wind turbines and other types of data, which are multi-dimensional, large-scale and highly time-series. Different types of data have different characteristics. The identification and extraction of data features need to be based on the specific information represented by the data for feature identification and extraction, and then the machine learning algorithm is used to realize the automatic classification of features. For example, the meteorological data is characterized and extended from wind speed at different heights and wind direction at different heights to form features of multiple dimensions such as wind speed at different heights, average wind speed at different heights within different time ranges, wind speed change rate at different heights, and the influence of wind direction at different heights on wind speed at different heights. Finally, five dimensions of historical meteorological data, wind turbine status data, historical power data, wind farm environment data and geographical environment data are formed, and feature mining is extended from four directions of time series, variability, internal interactivity and external interactivity to characterize the original data and form a feature representation matrix of different types.

[0012] Step 3: Construction of wind power prediction model.

[0013] The ensemble learning method is used to integrate various features into the model to predict the ultra-short-term power of wind power, and the model with the smallest root mean square error and mean absolute error of the prediction results is found according to the k-fold cross-validation method;

[0014] There are many factors that affect wind power. How to identify and select the factors that have the greatest impact on wind power is the key to building a wind power prediction model. This part is to select and determine the main features in the model. The main contents include:

[0015] (1) Calculate the correlation coefficients between the features determined in the second step to obtain the Pearson correlation coefficient matrix between the features. The Pearson correlation coefficient is used to analyze the correlation between the input features to determine whether there are redundant features. The size of the correlation coefficient represents the degree of linear correlation between the two variables C1 and C2, and the calculation formula is:

[0016]

[0017] Where: are the standard deviations of C1 and C2 respectively; r∈[-1, 1], the closer r is to 1, the stronger the correlation between the two input features, and the closer r is to 0, the weaker the correlation between the two input features.

[0018] (2) The influence coefficient of each feature on wind power is calculated from five dimensions, the dimension with the greatest impact on wind power is selected, and the influence coefficient distribution of each feature under each dimension is obtained.

[0019] (3) By comparing the data from different wind farms, the characteristics of the final five dimensions are obtained and a wind power prediction model is constructed.

[0020] Step 4: Wind power prediction model training. The important content of this part is to determine the weights of the main features in the model. This part uses the XGBoost algorithm developed based on the ensemble learning method to train the parameters of the prediction model, and uses the K-fold cross-validation method to split the training set and the test set. The target parameters are screened by calculating the parameter model with the smallest RMSE and MAE, thereby constructing the wind power prediction algorithm of the present invention: TS_XGB.

[0021] The XGBoost algorithm is optimized and improved on the basis of GBDT, and the objective function is regularized and second-order Taylor expansion is performed. The mathematical model of XGBoost can be regarded as an additive model composed of K classification and regression trees (CART).

[0022]

[0023] Where: K is the number of trees; F is all possible CARTs; f k is a specific CART tree. In the regression process, the parameters θ={f1,f2,…,f k}, then the objective function of XGBoost becomes:

[0024]

[0025] Where: the first part is the loss function; the second part is the regularization term, which is obtained by adding the regularization terms of K trees. For the regularization term of the decision tree, each decision tree is improved through vector mapping, and the regularization term of XGBoost is obtained as:

[0026]

[0027] In the formula: γ, λ are the penalty coefficients of the model; T is the number of leaf nodes; ω is the fraction of leaves. The optimization objective function is gradually approached step by step. At the tth step, an optimized CART tree, namely f, is added to the existing t-1 trees. t , then the objective function becomes:

[0028]

[0029] Carry out Taylor second-order expansion of equation (6) to obtain:

[0030]

[0031] Where: gi is the first-order derivative of the loss function; hi is the second-order derivative of the loss function.

[0032]

[0033]

[0034] Step 5: Comparative test of wind power prediction model. After the data preparation, feature mining and model construction in the first four steps, the TS_XGBoost algorithm developed based on XGBoot was obtained. This part further verifies the accuracy of the algorithm by using data from different wind farms, and conducts comparative tests with similar algorithms, such as decision tree, SVM, etc., on the same data set to verify the effectiveness of this algorithm.

[0035] Technical effects and advantages of the present invention:

[0036] 1) The factors affecting wind power are decomposed from five dimensions and four directions, and the characteristics of the influencing factors are refined, fully expanded and explored.

[0037] 2) Starting from the correlation and membership of the features that affect wind power, effective identification and representation of the features are achieved.

[0038] 3) This method fully combines the time-space characteristics of the influencing factors and achieves a high accuracy rate for predicting ultra-short-term wind power, which can greatly improve the effectiveness of the original prediction method.

[0039] 4) This method can provide a reference for wind farm managers to identify wind power, improve the work efficiency of staff, and further improve the utilization rate of wind power equipment and wind power output rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a diagram of the improved feature mining structure used in the embodiment of the present invention.

[0041] Figure 2 This is the flow chart of the TS_XGB of the present invention: an ultra-short-term wind power prediction method based on time domain characteristics and XGBoost. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Embodiment 1

[0044] The basic idea of ​​the present invention is to classify and mark the features according to the influencing factors such as wind speed, wind speed change rate, wind direction, air pressure, temperature, humidity, etc. in the wind farm environment through machine learning methods, and normalize and vectorize them so that they have domain feature information in the wind environment space. Then, the machine learning algorithm is used to train the model of the represented features to find out the set of wind environment features that have the greatest impact on wind power and their influence coefficients on wind power. Then, the integrated learning method is used to integrate each feature into the model to predict the ultra-short-term power of wind power. According to the k-fold cross-validation method, the model that minimizes the root mean square error and the mean absolute error of the prediction results is found, and finally a prediction model is formed.

[0045] Embodiment 2

[0046] like Figure 1 and Figure 2 As shown in the figure, a super short-term wind power prediction method based on time domain features and XGBoost is presented. The specific steps are as follows:

[0047] Step 1: Preprocessing of raw data. Since the data collected by the wind farm is collected and transmitted by various wind turbine signal sensors, the raw data will inevitably cause data missing, duplication, errors and redundancy during the formation and collection process. Data preprocessing is the prerequisite for data analysis. The specific steps include: filling in the missing values ​​in the raw data, which can be done by the average value method and the fixed value method; identifying and deleting duplicate values ​​and redundant values ​​in the data; identifying and correcting erroneous values ​​in the data, which can be done by using the outlier identification method, and then using the missing value filling method to insert the corrected new values; finally, normalizing the data of different dimensions. The commonly used method is to calculate the maximum and minimum values ​​of the data of each dimension and map the data from 0 to 1 to obtain the data representation in the 0-1 space.

[0048] Step 2: Data feature engineering. Data feature engineering includes: data feature identification, feature extraction, feature representation and feature classification. The data collected by the wind farm includes meteorological data, historical power data, wind turbine status data and other types of data, which are multi-dimensional, large-scale and highly time-series. Different types of data have different characteristics. The identification and extraction of data features need to be based on the specific information represented by the data for feature identification and extraction, and then the machine learning algorithm is used to realize the automatic classification of features. For example, the meteorological data is characterized and extended from wind speed at different heights and wind direction at different heights to form features of multiple dimensions such as wind speed at different heights, average wind speed at different heights within different time ranges, wind speed change rate at different heights, and the influence of wind direction at different heights on wind speed at different heights. Finally, five dimensions are formed, including historical meteorological data, wind turbine status data, historical power data, wind farm environment data and geographical environment data, and feature mining is extended from four directions of time series, variability, internal interactivity and external interactivity to characterize the original data and form a categorized feature representation matrix.

[0049] Step 3: Construction of wind power prediction model. There are many factors that affect wind power. How to identify and select the factors that have the greatest impact on wind power is the key to building a wind power prediction model. This part is to select and determine the main features in the model. The main contents include:

[0050] (1) Calculate the correlation coefficients between the features determined in the second step to obtain the Pearson correlation coefficient matrix between the features. The Pearson correlation coefficient is used to analyze the correlation between the input features to determine whether there are redundant features. The size of the correlation coefficient represents the degree of linear correlation between the two variables C1 and C2, and the calculation formula is:

[0051]

[0052] Where: are the standard deviations of C1 and C2 respectively; r∈[-1, 1], the closer r is to 1, the stronger the correlation between the two input features, and the closer r is to 0, the weaker the correlation between the two input features.

[0053] (2) The influence coefficient of each feature on wind power is calculated from five dimensions, the dimension with the greatest impact on wind power is selected, and the influence coefficient distribution of each feature under each dimension is obtained.

[0054] (3) By comparing the data from different wind farms, the characteristics of the final five dimensions are obtained and a wind power prediction model is constructed.

[0055] Step 4: Wind power prediction model training. The important content of this part is to determine the weights of the main features in the model. This part uses the XGBoost algorithm developed based on the ensemble learning method to train the parameters of the prediction model, and uses the K-fold cross-validation method to split the training set and the test set. The target parameters are screened by calculating the parameter model with the smallest RMSE and MAE, thereby constructing the wind power prediction algorithm of the present invention: TS XGB.

[0056] The XGBoost algorithm is optimized and improved on the basis of GBDT, and the objective function is regularized and second-order Taylor expansion is performed. The mathematical model of XGBoost can be regarded as an additive model composed of K classification and regression trees (CART).

[0057]

[0058] Where: K is the number of trees; F is all possible CARTs; f k is a specific CART tree. In the regression process, the parameters θ={f1,f2,…,f k}, then the objective function of XGBoost becomes:

[0059]

[0060] Where: the first part is the loss function; the second part is the regularization term, which is obtained by adding the regularization terms of K trees. For the regularization term of the decision tree, each decision tree is improved through vector mapping, and the regularization term of XGBoost is obtained as:

[0061]

[0062] In the formula: γ, λ are the penalty coefficients of the model; T is the number of leaf nodes; ω is the fraction of leaves. The optimization objective function is gradually approached step by step. At the tth step, an optimized CART tree, namely f, is added to the existing t-1 trees. t, then the objective function becomes:

[0063]

[0064] Carry out Taylor second-order expansion of equation (6) to obtain:

[0065]

[0066] Where: gi is the first-order derivative of the loss function; hi is the second-order derivative of the loss function.

[0067]

[0068]

[0069] Step 5: Comparative test of wind power prediction model. After the data preparation, feature mining and model construction in the first four steps, the TS_XGBoost algorithm developed based on XGBoot was obtained. This part further verifies the accuracy of the algorithm by using data from different wind farms, and conducts comparative tests with similar algorithms, such as decision tree, SVM, etc., on the same data set to verify the effectiveness of this algorithm.

[0070] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An ultra-short-term wind power prediction method based on time domain characteristics and XGBoost, characterized in that: Here are the steps: Step 1: Raw data preprocessing According to the influencing factors in the wind farm environment, the features are classified, marked, normalized and vectorized by machine learning methods, so that they have domain feature information in the wind environment space; Step 2: Data Analysis and Feature Mining Use machine learning algorithms to train the model of the represented features to find out the wind environment feature set that has the greatest impact on wind power and its impact coefficient on wind power; Different types of data have different characteristics. The identification and extraction of data features need to be based on the specific information represented by the data, and then the machine learning algorithm is used to realize the automatic classification of features. Finally, five dimensions are formed: historical meteorological data, wind turbine status data, historical power data, wind farm environment data, and geographic environment data. Feature mining is expanded from four directions: time series, variability, internal interactivity, and external interactivity. The original data is characterized and represented to form a type-based feature representation matrix. Step 3: Construction of wind power prediction model The ensemble learning method is used to integrate various features into the model to predict the ultra-short-term power of wind power, and the model with the smallest root mean square error and mean absolute error of the prediction results is found according to the k-fold cross-validation method; There are many factors that affect wind power. How to identify and select the factors that have the greatest impact on wind power is the key to building a wind power prediction model. This step is to select and determine the main features in the model. The main contents include: (1) Calculate the correlation coefficients between the features determined in the second step to obtain the Pearson correlation coefficient matrix between the features. Use the Pearson correlation coefficient to analyze the correlation between the input features to determine whether there are redundant features. The size of the correlation coefficient indicates the degree of linear correlation between the two variables C1 and C2. The calculation formula is: Where: are the standard deviations of C1 and C2 respectively; r∈[-1, 1], the closer r is to 1, the stronger the correlation between the two input features, and the closer r is to 0, the weaker the correlation between the two input features; (2) Calculate the influence coefficient of each feature on wind power from five dimensions, select the dimension with the greatest impact on wind power, and obtain the influence coefficient distribution of each feature under each dimension; (3) By comparing the data from different wind farms, the characteristics of the final five dimensions are obtained and a wind power prediction model is constructed; Step 4: Wind power prediction model training This part is to determine the weights of the main features in the model; the XGBoost algorithm developed based on the ensemble learning method is used to train the parameters of the prediction model, and the K-fold cross-validation method is used to split the training set and the test set. The target parameters are screened out by calculating the parameter model with the smallest RMSE and MAE, thereby constructing the wind power prediction algorithm of the present invention: TS_XGB; The XGBoost algorithm is optimized and improved on the basis of GBDT, and the objective function is regularized and second-order Taylor expansion is performed; the mathematical model of XGBoost can be regarded as an additive model composed of K classification and regression trees (CART); Where: K is the number of trees; F is all possible CARTs; f k is a specific CART tree; in the regression process, the parameters θ={f1,f2,…,f k }, then the objective function of XGBoost becomes: Where: the first part is the loss function; the second part is the regularization term, which is obtained by adding the regularization terms of K trees; for the regularization term of the decision tree, each decision tree is improved by vector mapping, and the regularization term of XGBoost is obtained as: Where: γ, λ are the penalty coefficients of the model; T is the number of leaf nodes; ω is the score of leaves; the optimization objective function is gradually approached step by step. At the tth step, an optimized CART tree, i.e., f, is added to the existing t-1 trees. t , then the objective function becomes: Carry out Taylor second-order expansion of equation (6) to obtain: Where: gi is the first-order derivative of the loss function; hi is the second-order derivative of the loss function; Step 5: Comparative test of wind power prediction model After the data preparation, feature mining and model construction in the first four steps, the TS_XGBoost algorithm developed based on XGBoot was obtained. In this step, the accuracy of the algorithm was further verified by using data from different wind fields, and a comparative test was conducted with the same type of algorithms on the same data set to verify the effectiveness of this algorithm.

2. The ultra-short-term wind power prediction method based on time domain features and XGBoost according to claim 1 is characterized by: In the first step, the influencing factors in the wind farm environment include but are not limited to: wind speed, wind speed change rate, wind direction, air pressure, temperature, and humidity.

3. The ultra-short-term wind power prediction method based on time domain characteristics and XGBoost according to claim 1 is characterized by: In the first step, the specific steps of raw data preprocessing include: filling in missing values ​​in the raw data, which can be done by using the average value method and the fixed value method; identifying and deleting duplicate values ​​and redundant values ​​in the data; identifying and correcting erroneous values ​​in the data, which can be done by using the outlier recognition method, and then using the missing value filling method to insert the corrected new values; finally, normalizing data of different dimensions. The commonly used method is to calculate the maximum and minimum values ​​of the data of each dimension and perform 0-1 mapping on the data to obtain the data representation in the 0-1 space.

4. The ultra-short-term wind power prediction method based on time domain characteristics and XGBoost according to claim 1 is characterized by: In the second step, the feature engineering of data includes: feature identification, feature extraction, feature representation and feature classification of data; the data collected by wind farms include meteorological data, historical power data, wind turbine status data and other types of data, which are multi-dimensional, large-scale and highly temporal.

5. The ultra-short-term wind power prediction method based on time domain characteristics and XGBoost according to claim 1 is characterized by: In the fourth step, the accuracy of the algorithm is further verified and compared with similar algorithms, including but not limited to: decision tree, SVM.

Images