A method for ultra-short-term wind power prediction based on the 24 solar terms adaptive mode decomposition framework

Abstract

A method for ultra-short-term wind power prediction based on an adaptive mode decomposition framework of the 24 solar terms, belonging to the field of wind power prediction for wind power generation, includes the following steps: historical wind power data from wind farms is divided into subsets according to the solar terms; for each subset, a sample entropy difference adaptation function is designed, and the key parameters of the fully adaptive noise set empirical mode decomposition are optimized using a particle swarm optimization algorithm; after optimization, the subset is decomposed to obtain intrinsic mode components reflecting different frequency characteristics; meteorological variables highly correlated with wind power are selected; and a neural network model is constructed using the selected meteorological variables and intrinsic mode components as inputs for ultra-short-term wind power prediction. This disclosure utilizes the differences in wind resource characteristics across different solar terms, dynamically optimizes mode decomposition parameters, and integrates key meteorological features to effectively improve the accuracy of ultra-short-term wind power prediction.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction for wind power generation, and in particular to an ultra-short-term wind power prediction method based on the 24 solar terms adaptive mode decomposition framework. Background Technology

[0002] Due to the highly random, fluctuating, and intermittent nature of wind power, wind farms can easily impact grid stability when connected to the grid, increasing the difficulty of power system dispatch. Therefore, improving the accuracy of ultra-short-term wind power forecasting has become a key technical requirement for ensuring the efficient operation of wind farms and achieving safe grid dispatch.

[0003] Currently, one of the most common prediction models in the field of wind power forecasting is the machine learning model. Machine learning models can effectively extract complex nonlinear relationships from multiple feature inputs and capture effective information from wind power sequences, thereby improving the model's prediction accuracy. However, when faced with complex wind power data, the aforementioned single deep learning model cannot meet the requirements.

[0004] To improve the accuracy and efficiency of deep learning models in wind power prediction, researchers have proposed introducing signal processing methods, such as empirical mode decomposition and variational mode decomposition, into wind power prediction. After processing, the noise in the data is reduced, and the fluctuation patterns of the data are more obvious. Then, the processed data is input into the deep learning model, making it easier to extract key information from the data and improve the prediction accuracy.

[0005] Current traditional time-series forecasting models do not fully explore the intrinsic relationship between climate change and wind power characteristics, resulting in poor adaptability to wind resource characteristics at different times. Furthermore, the fully adaptive noise ensemble empirical mode decomposition algorithm uses fixed parameters, which makes it difficult to adapt to diverse wind power data and is prone to problems such as insufficient decomposition or mode aliasing. Summary of the Invention

[0006] This disclosure utilizes the close relationship between wind resource characteristics and climate change, and the fact that the Twenty-Four Solar Terms, as a traditional Chinese method for dividing phenology and climate, can accurately reflect the characteristics of meteorological patterns in different periods. It provides an ultra-short-term wind power prediction method based on the Twenty-Four Solar Terms adaptive mode decomposition framework. By combining intelligent optimization algorithms with sample entropy evaluation, it achieves dynamic parameter optimization and integrates key meteorological characteristics to improve the accuracy and stability of ultra-short-term wind power prediction.

[0007] The ultra-short-term wind power prediction method based on the 24 solar terms adaptive mode decomposition framework disclosed herein mainly includes the following steps:

[0008] 1. Collect historical wind power data from wind farms to construct an original annual wind power dataset;

[0009] 2. Based on the precise division of the annual time series by the 24 solar terms, the annual wind power dataset is divided into 24 subsets.

[0010] 3. For each solar term subset, design a "sample entropy difference" fitness function. Using the particle swarm optimization algorithm, use this fitness function as the optimization target to optimize the key parameters of the fully adaptive noise set empirical mode decomposition, and obtain the optimized parameters of the fully adaptive noise set empirical mode decomposition.

[0011] 4. The optimized fully adaptive noise set empirical mode decomposition algorithm is used to decompose the corresponding solar term subset to obtain the intrinsic mode components that reflect different frequency characteristics.

[0012] 5. Quantitative analysis was conducted on the correlation between wind power and various meteorological variables (such as wind speed, wind direction, temperature, air pressure, humidity, etc.). By calculating the correlation coefficient, the main meteorological variables that are significantly correlated with wind power were selected as inputs for subsequent models.

[0013] 6. Construct a long short-term memory network model, using the main meteorological variables and the intrinsic mode components obtained by the fully adaptive noise set empirical mode decomposition algorithm as inputs, train the model and realize ultra-short-term wind power prediction.

[0014] Furthermore, in step 2, the annual wind power data is divided into 24 consecutive time periods, with the starting time of the 24 solar terms as the dividing point. Each subset corresponds to a solar term cycle, and each subset contains historical wind power data and related meteorological variables within that solar term cycle.

[0015] Furthermore, in step 3, the formula for calculating the "sample entropy difference" fit function is:

[0016]

[0017] in, The standard deviation of the sample entropy of each intrinsic mode component obtained from the fully adaptive noise set empirical mode decomposition. For the first i The sample entropy of each intrinsic mode component. The penalty coefficient is... This represents the total number of intrinsic modal components.

[0018] The method for calculating sample entropy is as follows:

[0019] For by N A time series composed of data Arrange them into a group with the following dimensions according to their serial numbers. m , where These vectors represent the vectors from the first...i Starting from a point m A continuous x The value; define the vector and Maximum distance between It is the absolute value of the maximum difference among the corresponding elements of the two, that is:

[0020]

[0021] Therefore, the sample entropy of the time series for:

[0022]

[0023]

[0024]

[0025] in, for and Distance less than or equal to r The number of for and The distance between them is less than or equal to r The number;

[0026] Furthermore, the optimization process of the particle swarm optimization algorithm described in step 3 includes:

[0027] Initialize the particle swarm, where each particle represents a set of fully adaptive noise ensemble empirical mode decomposition parameters (noise amplitude coefficient and number of noise additions).

[0028] Calculate the sample entropy of each intrinsic mode component (IMF) after empirical mode decomposition of a fully adaptive noise ensemble. ;

[0029] The fitness value of each particle is calculated using the sample entropy difference function. The greater the sample entropy difference, the greater the fitness, and the better the final prediction result.

[0030] Update the individual optimal position of the particle ( pbest ) and global optimal position ( gbest );

[0031] Update particle velocity and position using the following formula:

[0032]

[0033]

[0034] in, It is a particlei exist t The velocity at time +1; It is inertial weight; , It is a learning factor; , It is a random number within the interval [0,1]. It is a particle i Its own historical best position; It is a particle i Current position; It is the globally optimal position of the entire particle swarm.

[0035] Furthermore, in step 5, the Pearson correlation coefficient method is preferably used to quantitatively analyze the correlation between wind power and various meteorological variables (such as wind speed, wind direction, temperature, air pressure, humidity, etc.). By calculating the correlation coefficient, meteorological variables with an absolute correlation coefficient greater than a threshold are selected as the dominant influencing factors. The formula for the Pearson correlation coefficient is as follows:

[0036]

[0037] in, It is the first j The first meteorological variable i One sample; n The number of samples; The first of the wind power data i One sample; , Meteorological variables The mean of the wind power series;

[0038] Furthermore, the input layer of the long short-term memory network model described in step 6 contains all intrinsic mode components obtained from fully adaptive noise set empirical mode decomposition and the dominant influencing factors screened in step 5.

[0039] Compared with the prior art, the beneficial effects of this disclosure are:

[0040] ① Wind power has significant nonlinear characteristics and a wide range of variation, making it difficult for traditional machine learning methods to accurately characterize its complex patterns. This disclosure introduces the Twenty-Four Solar Terms, a time and climate division system based on long-term observation, summary, and practical verification by predecessors. The annual wind power data is divided into twenty-four subset datasets according to the solar term time points. Through this division, the impact of meteorological characteristics around wind turbines on wind power generation can be considered more closely, and historical data with better adaptability to the trend of change in the period to be predicted can be selected to assist in predicting the trend of wind power change and thus improve the prediction accuracy.

[0041] ② By decomposing the wind power data using the optimized fully adaptive noise set empirical mode decomposition algorithm, the prediction results can be made more accurate;

[0042] ③ By using the Pearson correlation test, meteorological variables that are significantly correlated with wind power are obtained as inputs to the prediction network, reducing redundant information interference and effectively improving the accuracy and efficiency of prediction.

[0043] ④ Significantly improved the accuracy of ultra-short-term wind power prediction, providing a new technical approach for efficient scheduling of wind farms and stable grid operation. Attached Figure Description

[0044] The above and other objects, features and advantages of this disclosure will become more apparent from the more detailed description of exemplary embodiments of this disclosure taken in conjunction with the accompanying drawings, in which the same reference numerals generally represent the same components.

[0045] Figure 1 The flowchart is shown below for the ultra-short-term wind power prediction method based on the 24 solar terms and parameter adaptive fully adaptive noise set empirical mode decomposition according to this disclosure.

[0046] Figure 2 This is an example of a raw wind power dataset. Detailed Implementation

[0047] Preferred embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0048] To address the shortcomings of existing methods in terms of adaptability to complex wind power data decomposition, parameter optimization efficiency, and multi-feature fusion, this disclosure proposes an ultra-short-term wind power prediction method based on the 24 solar terms adaptive modal decomposition framework. Through time-segmented adaptive processing, intelligent parameter optimization, and multi-feature fusion, it achieves high-precision prediction, providing technical support for efficient wind farm scheduling and stable grid operation.

[0049] The main steps of this method include:

[0050] 1. Construct a raw dataset for the entire year based on historical wind power data from wind farms.

[0051] 2. A time division mechanism based on the 24 solar terms is introduced to split the dataset into 24 subsets;

[0052] 3. For each solar term subset, construct a "sample entropy difference" fitness function. Using the particle swarm optimization algorithm, use this fitness function as the optimization target to intelligently optimize the key parameters of the fully adaptive noise set empirical mode decomposition.

[0053] 4. The subset is decomposed using the optimized fully adaptive noise set empirical mode decomposition algorithm to obtain the eigenmode components that can reflect the characteristics of different frequencies.

[0054] 5. Quantitatively analyze the correlation between meteorological variables and wind power, and select highly correlated meteorological variables as inputs for subsequent models to reduce redundant information interference;

[0055] 6. Using the selected meteorological variables and the intrinsic mode components obtained from decomposition as inputs, a long short-term memory neural network model is constructed to achieve ultra-short-term wind power prediction.

[0056] In one exemplary implementation, historical wind power data from January 1, 2024 to December 31, 2024, of a wind farm (located approximately 11 km from the local county seat, with an altitude ranging from approximately 2200 m to 3400 m) are used as the research object. The main steps of the ultra-short-term wind power prediction method based on the 24 solar terms adaptive mode decomposition framework disclosed herein are as follows: Figure 1 As shown.

[0057] (1) Collect local wind power data at a 15-minute scale from January 1, 2024 to December 31, 2024, totaling 35,136 records, to construct the original annual wind power dataset. Simultaneously, integrate the corresponding meteorological data, selecting one or more of wind speed, wind direction, temperature, and humidity for combination. In this embodiment, the selected meteorological data includes wind speed, wind direction, temperature, and humidity, with the same time scale as the wind power data, forming a complete dataset table. For missing values ​​in the data, cubic spline interpolation is used to fill in the missing data to ensure data continuity.

[0058] To consider the intrinsic link between seasonality and climate change and its impact on wind power generation, and to conduct more accurate analysis and ultra-short-term wind power forecasting, the starting point of the 24 solar terms is used as the time point, and the annual wind power data is divided into 24 subset wind power datasets, such as... Figure 2 As shown. For example, the wind power data between the beginning of spring and the rainy season each year are grouped into a subset, with each subset containing approximately 1440 data points (the duration varies slightly for different solar terms).

[0059] (2) For each solar term subset, the key parameters of the fully adaptive noise set empirical mode decomposition are optimized by using the particle swarm algorithm.

[0060] First, clarify the parameter definitions and optimization range, including the noise amplitude coefficient. The white noise intensity added to the original signal is controlled by an optimization range of [0.05, 0.5]. The number of noise additions, N, controls the average number of times the set is averaged, and the optimization range is [20, 100] (rounded to the nearest integer).

[0061] Next, we construct the "sample entropy difference" fitness function: first, we calculate the sample entropy of each intrinsic mode component obtained from the empirical mode decomposition of the fully adaptive noise set. The embedding dimension of the sample entropy is m=2, and the tolerance is r=0.2. The sample entropy is calculated as follows:

[0062] For by N A time series composed of data Arrange them into a group with the following dimensions according to their serial numbers. m , where These vectors represent the vectors from the first... i Starting from a point m A continuous x The value; define the vector and Maximum distance between It is the absolute value of the maximum difference among the corresponding elements of the two, that is:

[0063]

[0064] Therefore, the sample entropy of the time series for:

[0065]

[0066]

[0067]

[0068] in, for and Distance less than or equal to r The number of for and The distance between them is less than or equal to r The number;

[0069] The fitness function formula is:

[0070]

[0071] in, The standard deviation of the sample entropy of each intrinsic mode component obtained by fully adaptive noise set empirical mode decomposition; For the firsti The sample entropy of each intrinsic mode component; The penalty coefficient is set to 0.5 to ensure that the intrinsic mode components are arranged in descending order of frequency. This represents the total number of intrinsic modal components.

[0072] Then, using this fitness function as the optimization objective, the key parameters of the fully adaptive noise set empirical mode decomposition are optimized. In this embodiment, this is achieved through a particle swarm optimization algorithm: initializing 30 particles, iterating 50 times, and using inertia weights. The learning factor is linearly decreased from 0.9 to 0.4, with c1=c2=2. Parameter optimization is achieved by updating the velocity and displacement formulas of the particle swarm, ultimately yielding the optimized fully adaptive noise ensemble empirical mode decomposition parameters. The particle velocity and position are updated according to the following formulas:

[0073]

[0074]

[0075] in, For particles i exist t The velocity at time +1; Inertial weight; , For learning factors; , A random number within the interval [0,1]. For particles i Its own historical best position; For particles i Current position; This is the globally optimal position for the entire particle swarm.

[0076] By decomposing wind power data using the optimized fully adaptive noise ensemble empirical mode decomposition algorithm, the prediction results can be made more accurate.

[0077] Table 1. Optimization results of the particle swarm optimization algorithm

[0078]

[0079] (3) The optimized fully adaptive ensemble empirical mode decomposition algorithm is used to decompose the wind power dataset of each subset. Taking the "Lichun-Yushui" solar term subset as an example, the wind power data of this subset is decomposed using the fully adaptive ensemble empirical mode decomposition method. After the fully adaptive ensemble empirical mode decomposition process, the subset data is decomposed into 8 intrinsic mode components and 1 residual component. These intrinsic mode components represent the wind power fluctuation characteristics at different frequency scales, such as high-frequency short-term fluctuations and low-frequency long-term trends.

[0080] (4) In order to clearly show the relationship between meteorological variables and wind power, all data are linearly scaled between 0 and 1, and then the correlation between meteorological variables and wind power data is calculated according to the Pearson correlation coefficient method. The formula is as follows:

[0081]

[0082] in, It is the first j The first meteorological variable i One sample; n The number of samples; The first of the wind power data i One sample; , meteorological variables The mean of the wind power series; Indicates the first j The degree of correlation between meteorological variables and wind power.

[0083] Pearson correlation tests showed that all meteorological variables were positively correlated with wind power. The Pearson correlation coefficients between 90m wind speed, 70m wind speed, 90m wind direction, 70m wind direction, and wind power were as high as 0.753, 0.723, 0.765, and 0.730, respectively, indicating significant correlation. Therefore, the preferred set of characteristic variables {Ws} was obtained. _90 m, Ws _70 m, Wd _90 m, Wd _70 m, IMF}.

[0084] (5) The intrinsic mode components obtained by the fully adaptive noise set empirical mode decomposition algorithm and the meteorological variables selected by the Pearson correlation coefficient method are used as model inputs. The ratio of training set, validation set and test set is 8:1:1. A long short time memory network deep neural network model is constructed, and a multi-step prediction framework is adopted with a step size of 15 min and a number of steps of 3, that is, predicting the wind power in the next 45 minutes.

[0085] Based on this, the model was trained on the training set. During training, mean squared error (MSE) was used as the loss function, Adam was used as the optimizer, the learning rate was set to 0.001, the batch size was 64, and the number of training epochs was 200. To prevent overfitting during training, model performance was monitored using the validation set, and model hyperparameters were adjusted. The optimal parameters were retained during training, and predictions were made on the test set using the optimal parameters to obtain the model's multi-step prediction results.

[0086] To verify the prediction accuracy of the method described in this embodiment, multi-step prediction of ultra-short-term wind speed was performed using reported wind power data from a wind farm from January 1 to December 31, 2024. In addition, mean absolute error and root mean square error were selected as evaluation indicators. The results show that the proposed method improves both mean absolute error and root mean square error, thus proving the effectiveness of the method.

[0087] Tables 2 and 3 show the root mean square error and mean absolute error of the model, respectively.

[0088] Table 2 Evaluation results of the root mean square error index of the model

[0089]

[0090] Table 3 Evaluation Results of Model Mean Absolute Error Index

[0091]

[0092] In this embodiment, considering the differences in wind resource characteristics across different solar terms, a particle swarm optimization algorithm is used to optimize the fully adaptive empirical mode decomposition parameters of the noise ensemble. This is combined with sample entropy evaluation to achieve adaptive adjustment of the decomposition process, ultimately forming an ultra-short-term wind power prediction method based on adaptive mode decomposition parameters. This method significantly improves the accuracy of ultra-short-term wind power prediction, providing a novel technical approach for efficient wind farm scheduling and stable grid operation.

[0093] The above technical solutions are merely exemplary embodiments of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the specific embodiments of the present invention. Therefore, the methods described above are only preferred and not restrictive.

Claims

1. A method for predicting ultra-short-term wind power based on an adaptive mode decomposition framework of the 24 solar terms, characterized in that, Includes the following steps: S1, Construct the original dataset for the whole year based on the historical wind power data of the wind farm; S2 introduces a time division mechanism based on the 24 solar terms, splitting the dataset into 24 subsets; S3. For each solar term subset, a sample entropy difference fitness function is constructed. The fitness function is used as the optimization target to optimize the key parameters of the fully adaptive noise set empirical mode decomposition. S4. Using the optimized fully adaptive noise set empirical mode decomposition algorithm, the corresponding solar term subset is decomposed to obtain several intrinsic mode components. These intrinsic mode components represent the wind power fluctuation characteristics at different frequency scales. S5. Analyze the correlation between various meteorological variables and wind power, and screen out meteorological variables with high correlation as the dominant influencing factors. S6. Construct a long short-term memory network model, using all intrinsic mode components obtained from fully adaptive noise set empirical mode decomposition and the selected dominant influencing factors as inputs to train the model and achieve ultra-short-term wind power prediction. In step S3, the specific steps for constructing the sample entropy difference fitness function and using the particle swarm optimization algorithm to optimize the key parameters of the fully adaptive noise set empirical mode decomposition include: S31, Define the key parameters and their optimization range for the fully adaptive noise set empirical mode decomposition, wherein the key parameters include: noise amplitude coefficient. , used to control the intensity of white noise added to the original signal; the number of times noise is added, N0, is used to control the number of times the set is averaged; S32, Construct the sample entropy difference fit function, the calculation formula is as follows: Where F represents the fit, The standard deviation of the sample entropy of each intrinsic mode component obtained from the fully adaptive noise set empirical mode decomposition. For the first i The sample entropy of each intrinsic mode component. The penalty coefficient is... This represents the total number of intrinsic modal components. The method for calculating sample entropy is as follows: For by N A time series composed of data Arrange them into a group with the following dimensions according to their serial numbers. m A vector sequence, where These vectors represent the vectors from the first... i Starting from a point m A continuous x The value; define the vector and Maximum distance between It is the absolute value of the maximum difference among the corresponding elements of the two, that is: Then the sample entropy of the time series for: in, for and Distance less than or equal to r The number of for and The distance between them is less than or equal to r The number; S33, using the particle swarm optimization algorithm, the fitness function is used as the optimization target to optimize the key parameters, specifically including: Initialize the particle swarm, where each particle represents a set of fully adaptive noise ensemble Empirical Mode Decomposition (EMD) key parameters, including: noise amplitude coefficient and the number of noise additions. The fitness value of each particle is calculated using the sample entropy difference fitness function. The greater the sample entropy difference, i.e. the greater the fitness, the better the final wind power prediction result. Update the individual optimal position pbest and the global optimal position gbest of the particles to complete the parameter optimization. The particle velocity and position are updated according to the following formula: in, For particles i exist t The velocity at time +1; Inertial weight; , For learning factors; , A random number within the interval [0,1]. For particles i Its own historical best position; For particles i Current position; This is the globally optimal position for the entire particle swarm.

2. The method according to claim 1, characterized in that, The specific method of step S1 includes: Obtain wind power data for a specified region at a 15-minute scale for the entire year to construct the original annual wind power dataset; at the same time, integrate the corresponding meteorological data, including one or more of wind speed, wind direction, temperature, and humidity, with the same time scale as the wind power data, to form a complete dataset table; For missing values ​​in the data, cubic spline interpolation is used to fill in the missing data to ensure the continuity of the data.

3. The method according to claim 1 or 2, characterized in that, The specific method for step S2 is as follows: Using the start time of the 24 solar terms as the dividing point, the annual wind power data is divided into 24 consecutive time period subsets. Each subset corresponds to a solar term cycle, and each subset contains historical wind power data and related meteorological variable data within that solar term cycle.

4. The method according to claim 1, characterized in that, In step S31, the optimization range of the key parameters is set as follows: :[0.05,0.5], N:[20, 100].

5. The method according to claim 1, characterized in that, The specific method of step S5 includes: Scaling all data linearly between 0 and 1; The correlation coefficient between each meteorological variable and the wind power data was calculated using the Pearson correlation coefficient method. The meteorological variables include one or more of the following: wind speed, wind direction, temperature, air pressure, and humidity. Meteorological variables with absolute correlation coefficients greater than a set threshold were selected as the dominant influencing factors. The formula for calculating the Pearson correlation coefficient is as follows: in, It is the first j The first meteorological variable i One sample; n′ The number of samples; The first of the wind power data i One sample; , meteorological variables The mean of the wind power series; Indicates the first j The degree of correlation between meteorological variables and wind power.

6. The method according to claim 1, characterized in that, Step S6 specifically includes: The intrinsic mode components obtained by the fully adaptive noise set empirical mode decomposition algorithm and the meteorological variables selected by the Pearson correlation coefficient method are used as model inputs to construct a long short time memory deep neural network model. A multi-step prediction framework is adopted with a step size of 15 minutes and a number of steps of 3, that is, to predict the wind power in the next 45 minutes. The model is trained based on the training set. During the training process, mean squared error is used as the loss function, Adam is used as the optimizer, the learning rate is set to 0.001, the batch size is 64, and the number of training rounds is 200. During training, to prevent overfitting, model performance is monitored using a validation set, model hyperparameters are adjusted, and the optimal parameters are retained during training. The optimal parameters are then used to make predictions on the test set to obtain multi-step prediction results from the model.

Images