An ultra-short-term wind power prediction method and system

Abstract

The application discloses a kind of ultra-short-term wind power prediction method and system, the method includes: the time lag characteristic analysis of original wind power sequence time series data, according to the analysis result, select several second historical wind power time series data corresponding wind power sequence, then the wind power sequence is decomposed, different subsequence is obtained, for each subsequence, construct initial DBO-GRNN prediction model, and the hyperparameter optimization of initial DBO-GRNN prediction model is carried out, different subsequence is predicted using the DBO-GRNN prediction model after hyperparameter optimization, the prediction result of each subsequence is superimposed, and wind power prediction result is obtained, improve the precision and efficiency of ultra-short-term wind power prediction.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and in particular to an ultra-short-term wind power prediction method and system. Background Technology

[0002] Wind energy, as a low-carbon, pollution-free renewable energy source that can be built both on land and at sea with low operation and maintenance costs, is currently widely used worldwide, and wind power generation has achieved unprecedented development. However, wind power generation has characteristics such as randomness, intermittency, and nonlinearity. Large-scale wind power grid connection seriously threatens the safety and stability of the power system. Therefore, improving wind power prediction performance is an effective measure to improve the stability and economy of wind power grid connection.

[0003] Currently, wind power forecasting is classified according to the time scale of the prediction, into ultra-short-term forecasting, short-term forecasting, and medium-to-long-term forecasting. Since most wind power in China is connected to the grid through centralized grid connection, its power fluctuations have a significant impact on grid dispatching. Wind power output over several days will affect the formulation of grid dispatching plans, while wind power output within several hours will directly affect the dispatching decisions of grid dispatchers. The "Wind Power Forecasting Functional Specification" stipulates that ultra-short-term wind power forecasting is the wind power output within 0-4 hours in the future, with a time resolution of not less than 15 minutes. It can be seen that ultra-short-term forecasting can realize real-time adjustment of the power system to alleviate problems such as power supply frequency regulation. However, there are few technical means for ultra-short-term wind power forecasting in the current technology. Most of them rely on collecting a large amount of historical data and making predictions directly based on the historical data through a single prediction model. This not only has low accuracy, but also the collection of historical data is not only difficult, but the amount of data to be processed is also increasing, resulting in low prediction efficiency. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention discloses an ultra-short-term wind power prediction method and system, which improves the accuracy and efficiency of ultra-short-term wind power prediction.

[0005] To achieve the above objectives, in a first aspect, the present invention discloses an ultra-short-term wind power prediction method, comprising:

[0006] Each of the acquired first historical wind power time series data is subjected to time delay characteristic analysis with a preset prediction time. The correlation coefficient between each first historical wind power time series data and the prediction time is obtained. Based on the correlation coefficient, the first historical wind power time series data are filtered to obtain a number of second historical wind power time series data.

[0007] Obtain the historical wind power sequence corresponding to each of the plurality of second historical wind power time series data, and perform mode decomposition on each historical wind power sequence to obtain the subsequence corresponding to each historical wind power sequence; the subsequence includes mode components and residual signal components;

[0008] An initial DBO-GRNN prediction model is constructed for each subsequence based on the modal components and the remaining signal components. The initial DBO-GRNN prediction model is then optimized using a preset hyperparameter optimization algorithm to obtain the DBO-GRNN prediction model.

[0009] Power prediction is performed on the subsequences according to the DBO-GRNN prediction model corresponding to each subsequence, and the prediction components corresponding to each subsequence are superimposed to obtain the wind power prediction result corresponding to the prediction time.

[0010] This invention discloses an ultra-short-term wind power prediction method. First, time-lag characteristic analysis is performed on historical wind power sequences. The historical wind power sequence with the strongest correlation to the prediction time is selected as the reference data for prediction, reducing data processing volume and improving prediction efficiency. Simultaneously, the strong correlation between the reference historical wind power sequence and the prediction time effectively improves the accuracy of the final prediction result. To this end, the correlation coefficient between the time-series data corresponding to different power sequences and the prediction time is calculated. Finally, the wind power sequence corresponding to the time-series data with the highest correlation coefficient is selected, i.e., the several second historical wind power time-series numbers are obtained. Based on the historical wind power sequence corresponding to each second historical wind power time series data, after determining the parameter data for prediction, in order to improve the accuracy of the prediction model, this invention performs mode decomposition on the historical wind power sequence, constructs a prediction model based on the mode decomposition results, reduces power fluctuations, and enables the prediction model to better uncover the intrinsic relationship between past and future moments. At the same time, in order to further improve the accuracy of the prediction model, this invention performs hyperparameter optimization on the constructed prediction model, uses the hyperparameter-optimized model for prediction, and superimposes the prediction results to achieve a more accurate power prediction result.

[0011] As a preferred example, the process of obtaining the correlation coefficient between each first historical wind power time series data and the predicted time, and filtering the plurality of first historical wind power time series data based on the correlation coefficient, includes:

[0012] The Pearson correlation coefficient between each historical wind power time series data and the predicted time is calculated according to the preset Pearson coefficient calculation function.

[0013] The Pearson correlation coefficient is compared with a preset coefficient threshold, and the first historical wind power time series data corresponding to the Pearson correlation coefficient being greater than the coefficient threshold is selected as the second historical wind power time series data.

[0014] This invention calculates the correlation between each time series data point in the historical time and the predicted time. The historical power sequence corresponding to the time series data with strong correlation is more representative of the wind power pattern, which further improves the accuracy of the wind power prediction in the later stage. At the same time, selecting time series data with strong correlation helps to reduce the power sequence that needs to be processed and improves the prediction efficiency.

[0015] As a preferred example, the modal decomposition of each historical wind power sequence to obtain the subsequences corresponding to each historical wind power sequence includes:

[0016] Based on the preset mode decomposition algorithm and the number of iterations, each historical wind power sequence is subjected to several mode decompositions to obtain several first mode components and several first residual signal components.

[0017] In each of the several mode decompositions, a preprocessed power sequence is obtained by adding several times of white noise to each historical wind power sequence.

[0018] In each mode decomposition, the preprocessed power sequence is decomposed several times using a preset EMD to obtain a first residual signal component and several initial mode components. The mean of the several initial mode components is then calculated to obtain the first mode component.

[0019] To improve the accuracy of the prediction model, this invention employs a mode decomposition method on the power sequence to reduce power volatility, thereby improving the stability of the model during subsequent model construction. This allows the prediction model to better learn the patterns of past and future moments, thus enhancing the model's accuracy.

[0020] As a preferred example, the step of performing several mode decompositions on each historical wind power sequence according to a preset mode decomposition algorithm and number of iterations to obtain several first mode components and several first residual signal components includes:

[0021] According to the preset CEEMDAN algorithm, each historical wind power sequence is subjected to iterative mode decomposition to obtain the first residual signal component and several first mode components.

[0022] White noise is added to the first residual signal component obtained after several modal decompositions, and EMD decomposition is performed. It is then determined whether the second residual signal component obtained after the EMD decomposition is a monotonic signal. If the second residual signal component is a monotonic signal, the iteration is stopped.

[0023] This invention uses whether the remaining signal component is a monotonic signal as the criterion for stopping the iteration. When the remaining signal component is a monotonic signal, it indicates that the mode decomposition at this time has eliminated the fluctuation of the power sequence and improved the accuracy of the subsequent model learning.

[0024] As a preferred example, the modal decomposition of each historical wind power sequence to obtain the subsequences corresponding to each historical wind power sequence includes:

[0025] By adding white noise several times to each historical wind power sequence, preprocessing is performed on each historical wind power sequence to obtain a preprocessed power sequence; wherein, the calculation formula for the preprocessing is:

[0026] y i(t) =y(t)+εn i (t), i = 1, 2, ..., T

[0027] Wherein, y(t) represents the original wind power sequence, y i(t) Let n represent the sequence after the i-th addition of white noise, ε represent the noise standard deviation, and n i (t) represents Gaussian white noise that follows the N(0,1) distribution, and T represents the number of times the noise is added;

[0028] The preprocessed power sequence is decomposed several times using a preset EMD; wherein the EMD is expressed as follows:

[0029]

[0030] r1(t)=y (t) -I1(t)

[0031] Wherein, I1(t) is the first modal component obtained from the decomposition, and I1(i) is y i(t) The modal components obtained after EMD decomposition, r1(t) are the remaining signal components after one mode decomposition.

[0032] This invention uses white noise to reduce data errors. At the same time, based on EMD, it has data-driven adaptability, can analyze nonlinear and non-stationary signals, and can greatly ignore the instability of power sequences, providing stable data for subsequent model construction, thereby improving the accuracy of the model prediction.

[0033] As a preferred example, the hyperparameter optimization of the initial DBO-GRNN prediction model using a preset hyperparameter optimization algorithm includes:

[0034] The smoothing factor parameters in the initial DBO-GRNN prediction model are optimized using a preset dung beetle optimization algorithm.

[0035] The loss value between the predicted value and the true value of the model after the smoothing factor parameter adjustment is continuously calculated based on the preset mean squared error loss function and mean absolute percentage error loss function. When the loss value is less than the preset loss value threshold, the DBO-GRNN prediction model is obtained by combining the current smoothing factor parameter with the initial DBO-GRNN prediction model.

[0036] This invention improves the model's ability to capture patterns and thus enhances its prediction accuracy by optimizing its hyperparameters. Furthermore, leveraging the fast convergence and high accuracy of the dung beetle optimization algorithm, this invention employs the dung beetle optimization algorithm for hyperparameter optimization, thereby improving prediction accuracy and efficiency. During the optimization process, a loss function is used to prevent overfitting and further enhance model accuracy.

[0037] Secondly, the present invention discloses an ultra-short-term wind power prediction system, the system comprising a time series analysis module, a mode decomposition module, a model optimization module and a power prediction module;

[0038] The time series analysis module is used to perform time delay characteristic analysis on each of the acquired first historical wind power time series data and a preset prediction time, obtain the correlation coefficient between each first historical wind power time series data and the prediction time, and filter the first historical wind power time series data according to the correlation coefficient to obtain a number of second historical wind power time series data.

[0039] The mode decomposition module is used to obtain the historical wind power sequence corresponding to each of the several second historical wind power time series data, and to perform mode decomposition on each historical wind power sequence to obtain the subsequence corresponding to each historical wind power sequence; the subsequence includes mode components and residual signal components;

[0040] The model optimization module is used to construct an initial DBO-GRNN prediction model corresponding to each subsequence based on the modal components and the remaining signal components, and to optimize the initial DBO-GRNN prediction model using a preset hyperparameter optimization algorithm to obtain the DBO-GRNN prediction model.

[0041] The power prediction module is used to predict the power of the subsequence according to the DBO-GRNN prediction model corresponding to each subsequence, obtain the prediction component corresponding to each subsequence, and superimpose the prediction components corresponding to each subsequence to obtain the wind power prediction result corresponding to the prediction time.

[0042] This invention discloses an ultra-short-term wind power prediction system. First, it performs time-lag characteristic analysis on historical wind power sequences, selecting the historical wind power sequence with the strongest correlation to the prediction time as the prediction reference data. This reduces data processing volume and improves prediction efficiency. Simultaneously, the strong correlation between the reference historical wind power sequence and the prediction time effectively improves the accuracy of the final prediction result. To this end, by calculating the correlation coefficient between the time-series data corresponding to different power sequences and the prediction time, the wind power sequence corresponding to the time-series data with the highest correlation coefficient is finally selected, i.e., the several second historical wind power time-series numbers are obtained. Based on the historical wind power sequence corresponding to each second historical wind power time series data, after determining the parameter data for prediction, in order to improve the accuracy of the prediction model, this invention performs mode decomposition on the historical wind power sequence, constructs a prediction model based on the mode decomposition results, reduces power fluctuations, and enables the prediction model to better uncover the intrinsic relationship between past and future moments. At the same time, in order to further improve the accuracy of the prediction model, this invention performs hyperparameter optimization on the constructed prediction model, uses the hyperparameter-optimized model for prediction, and superimposes the prediction results to achieve a more accurate power prediction result.

[0043] As a preferred example, the time series analysis module includes a coefficient calculation unit and a filtering unit;

[0044] The coefficient calculation unit is used to calculate the Pearson correlation coefficient between each historical wind power time series data and the prediction time according to the preset Pearson coefficient calculation function.

[0045] The filtering unit is used to compare the Pearson correlation coefficient with a preset coefficient threshold, and filter the first historical wind power time series data corresponding to the Pearson correlation coefficient being greater than the coefficient threshold as the second historical wind power time series data.

[0046] This invention calculates the correlation between each time series data point in the historical time and the predicted time. The historical power sequence corresponding to the time series data with strong correlation is more representative of the wind power pattern, which further improves the accuracy of the wind power prediction in the later stage. At the same time, selecting time series data with strong correlation helps to reduce the power sequence that needs to be processed and improves the prediction efficiency.

[0047] As a preferred example, the mode decomposition module includes an iterative decomposition unit, a preprocessing unit, and a decomposition unit;

[0048] The iterative decomposition unit is used to perform several mode decompositions on each historical wind power sequence according to a preset mode decomposition algorithm and the number of iterations, to obtain several first mode components and several first residual signal components.

[0049] The preprocessing unit is used to preprocess each historical wind power sequence by adding white noise several times in each of the several mode decompositions to obtain a preprocessed power sequence.

[0050] The decomposition unit is used to decompose the preprocessed power sequence several times in each mode decomposition by a preset EMD to obtain a first residual signal component and several initial mode components, and to calculate the mean of the several initial mode components to obtain the first mode component.

[0051] To improve the accuracy of the prediction model, this invention employs a mode decomposition method on the power sequence to reduce power volatility, thereby improving the stability of the model during subsequent model construction. This allows the prediction model to better learn the patterns of past and future moments, thus enhancing the model's accuracy.

[0052] Thirdly, the present invention discloses an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0053] Memory, used to store computer programs;

[0054] The processor, when executing a program stored in memory, implements an ultra-short-term wind power prediction method as described in the first aspect. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating an ultra-short-term wind power prediction method disclosed in an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of the structure of an ultra-short-term wind power prediction system disclosed in an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of a wind power sequence disclosed in an embodiment of the present invention;

[0058] Figure 4 This is a schematic diagram of the correlation coefficient between time and Pearson coefficient disclosed in an embodiment of the present invention;

[0059] Figure 5 This is a schematic diagram of a CEEMDAN decomposition result disclosed in an embodiment of the present invention.

[0060] Figure 6 This is a schematic diagram of an ablation experiment disclosed in an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example 1

[0063] This invention discloses an ultra-short-term wind power prediction method. The specific implementation process of the prediction method can be referred to... Figure 1 It mainly includes steps 101 to 104, the steps of which include:

[0064] Step 101: Perform time delay characteristic analysis on each of the acquired first historical wind power time series data and the preset prediction time to obtain the correlation coefficient between each first historical wind power time series data and the prediction time, and filter the first historical wind power time series data according to the correlation coefficient to obtain a number of second historical wind power time series data.

[0065] In this embodiment, the step includes: calculating the Pearson correlation coefficient between each historical wind power time series data and the predicted time according to a preset Pearson coefficient calculation function; comparing the Pearson correlation coefficient with a preset coefficient threshold, and selecting the first historical wind power time series data corresponding to the Pearson correlation coefficient being greater than the coefficient threshold as the second historical wind power time series data.

[0066] Specifically, in this embodiment, time-lag characteristic analysis is performed on the original wind power sequence, that is, the time series data in the original wind power sequence is obtained, and the correlation coefficient between the time series data and the predicted time is calculated. Further, in one feasible method, actual power data of a wind farm cluster is collected for experimental analysis. This dataset records the power data of the wind farm from March 1, 2019 to March 31, 2019. The original dataset records data every 10 minutes. The original data is sampled every hour, and 720 sample points are selected as the original wind power sequence. Figure 3As shown, time-lag characteristic analysis is then performed on the time-series data corresponding to the 720 sample points. In this implementation, the Pearson correlation coefficient ρ between the input characteristics and the predicted power at the 8 time lags is calculated. The Pearson correlation coefficient at the 8 time lags can be found by referring to... Figure 4 ,like Figure 4 As shown, the Pearson correlation coefficient gradually decreases with the forward time, indicating a weakening correlation with the predicted time. Therefore, to improve prediction efficiency and accuracy, a time series with |ρ| > 0.85 is considered to have a strong correlation. Figure 4 Given that |ρ|>0.85, there are 5 sets (t-1, t-2, t-3, t-4, t-5). Therefore, the first 5 time points are selected for time series modeling, and multi-channel time series joint training is performed. Analysis of the time lag characteristics of wind power sequences can better understand the lag effect in time series data, thereby improving the accuracy of prediction. Time series with strong correlation to the prediction time are selected for multi-channel time series modeling (multiple time series, where each time series corresponds to one power sequence).

[0067] This step calculates the correlation between each time series data point in the historical time and the predicted time. The historical power sequence corresponding to the time series data with strong correlation is more representative of the wind power pattern, which further improves the accuracy of the wind power prediction in the later stage. At the same time, selecting time series data with strong correlation helps to reduce the power sequence that needs to be processed and improves the prediction efficiency.

[0068] Step 102: Obtain the historical wind power sequence corresponding to each of the several second historical wind power time series data, and perform modal decomposition on each historical wind power sequence to obtain the subsequence corresponding to each historical wind power sequence; the subsequence includes modal components and residual signal components.

[0069] In this embodiment, the step includes: performing mode decomposition on each historical wind power sequence several times according to a preset mode decomposition algorithm and number of iterations to obtain several first mode components and several first residual signal components; in each of the several mode decompositions, preprocessing each historical wind power sequence by adding several times of white noise to each historical wind power sequence to obtain a preprocessed power sequence; in each mode decomposition, performing several decompositions on the preprocessed power sequence using a preset EMD algorithm to obtain first residual signal components and several initial mode components, and averaging the several initial mode components to obtain the first mode components.

[0070] Furthermore, according to the preset CEEMDAN algorithm, each historical wind power sequence is subjected to iterative mode decomposition to obtain a first residual signal component and several first mode components; white noise is added to the first residual signal component obtained after several mode decompositions, and EMD decomposition is performed to determine whether the second residual signal component obtained after the EMD decomposition is a monotonic signal; if the second residual signal component is a monotonic signal, the iteration is stopped.

[0071] Meanwhile, in this embodiment, each historical wind power sequence is preprocessed by adding white noise several times to it to obtain a preprocessed power sequence; wherein, the calculation formula for the preprocessing is:

[0072] y i(t) =y(t)+εn i (t), i = 1, 2, ..., T

[0073] Wherein, y(t) represents the original wind power sequence, y i(t) Let n represent the sequence after the i-th addition of white noise, ε represent the noise standard deviation, and n i (t) represents Gaussian white noise that follows the N(0,1) distribution, and T represents the number of times the noise is added;

[0074] The preprocessed power sequence is decomposed several times using a preset EMD; wherein the EMD is expressed as follows:

[0075]

[0076] r1(t)=y (t) -I1(t)

[0077] Wherein, I1(t) is the first modal component obtained from the decomposition, and I1(i) is y i(t) The modal components obtained after EMD decomposition, r1(t) are the remaining signal components after one mode decomposition.

[0078] Specifically, in this embodiment, after selecting multiple time-series data with |ρ|>0.85 according to the above steps, a wind power sequence corresponding to each time-series data is obtained, thereby obtaining multiple wind power sequences. The wind power sequences can be referenced... Figure 4 Then, mode decomposition is performed on each wind power sequence. In this embodiment, the CEEMDAN method is used to perform mode decomposition on the wind power sequence to obtain the subsequence corresponding to each wind power sequence. The subsequence includes mode components and residual signal components.

[0079] Furthermore, the specific steps for decomposing the wind power sequence using the CEEMDAN method are as follows:

[0080] Step 1: Add T times of white noise with a mean of 0 to the wind power sequence to obtain the preprocessed sequence y. i (t), where y i The expression for (t) is as follows:

[0081] y i(t) =y(t)+εn i (t), i = 1, 2, ..., T

[0082] In the formula, y(t) is the original wind power sequence, y i(t) Let n be the sequence after adding white noise for the i-th time, ε be the noise standard deviation, and n be the number of subsequences. i (t) is Gaussian white noise that follows the N(0,1) distribution, and T is the number of times the noise is added.

[0083] Step 2: Decompose the preprocessed sequence T times using EMD, and calculate the mean of the T modal components to obtain the first subsequence I1(t), wherein the subsequence I1(t) is expressed as follows:

[0084]

[0085] r1(t)=y (t) -I1(t)

[0086] In the formula, I1(t) is the first modal component obtained from the CEEMDAN decomposition. For y i(t) The modal components obtained after EMD decomposition, r1(t) are the remaining signal components after one CEEMDAN decomposition.

[0087] Step 3: Repeat steps 1 and 2, adding white noise to the remaining signal components after CEEMDAN decomposition K times and performing EMD decomposition again. Then determine whether the remaining signal components meet the stopping condition (the remaining signal component r after the nth decomposition). n (t) is a monotonic signal. If the condition is met, the iteration stops and the CEEMDAN algorithm decomposition ends. The specific expression of the CEEMDAN decomposition algorithm is as follows:

[0088]

[0089] r k (t)=r k-1 (t)-I k (t)

[0090] In the formula I k(t) represents the k-th modal component E obtained from the CEEMDAN decomposition. k-1 (·) represents y i(t) The (k-1)th modal component obtained after EMD decomposition, ε k-1 To add noise weighting coefficients to the remaining signal components after CEEMDAN decomposition k-1 times, r k (t) represents the remaining signal components after k decompositions of CEEMDAN.

[0091] Regarding the CEEMDAN decomposition method in the implementation and the selected time and corresponding power sequence in the above steps, in this implementation, the signal-to-noise ratio of the CEEMDAN decomposition method is set to 0.1, and 50 noise additions are performed. The maximum number of envelope additions is set to 5000. Modal decomposition is performed on the selected power sequence according to the settings of the CEEMDAN decomposition method described above to obtain the CEEMDAN decomposition result. The result can be referred to... Figure 5 ,like Figure 5 As shown, CEEMDAN decomposition produces 10 distinct IMF components and 1 residual component. The IMF1 to IMP10 components are displayed sequentially from high frequency to low frequency. Figure 5 (Only the first 6 components are displayed). The first IMF components selected represent high-frequency or noise components in the wind power sequence sample, while the later selected IMF components represent the long-term variation characteristics of the wind power sequence sample. The IMF components generated by CEEMDAN decomposition contain lower instantaneous frequencies, which can more effectively weaken the mutual influence between different subsequences, better characterize the overall trend of the original wind power sequence, and are conducive to further exploring sequence characteristics and improving prediction performance.

[0092] In order to improve the accuracy of the prediction model, this step uses mode decomposition of the power sequence to reduce power volatility, thereby improving the stability of the model during subsequent model construction. This allows the prediction model to better learn the patterns of past and future moments, thus improving the accuracy of the model.

[0093] Step 103: Construct an initial DBO-GRNN prediction model for each subsequence based on the modal components and the remaining signal components, and optimize the hyperparameters of the initial DBO-GRNN prediction model using a preset hyperparameter optimization algorithm to obtain the DBO-GRNN prediction model.

[0094] In this embodiment, the step includes: optimizing the smoothing factor parameters in the initial DBO-GRNN prediction model according to a preset dung beetle optimization algorithm; continuously calculating the loss value between the predicted value and the true value of the model after the smoothing factor parameter adjustment according to a preset mean squared error loss function and mean absolute percentage error loss function; when the loss value is less than a preset loss value threshold, obtaining the DBO-GRNN prediction model by combining the current smoothing factor parameters with the initial DBO-GRNN prediction model.

[0095] Specifically, in this embodiment, the initial DBO-GRNN prediction model is constructed using the subsequences obtained from the CEEMDAN decomposition. Then, RMSE (Root Mean Square Error) and MAPE (Mean Absolute Percentage Error) are used as evaluation metrics for experimental testing to determine the prediction accuracy of the model. The power values ​​at each time step obtained after decomposition are used as the training dataset to train the initial DBO-GRNN prediction model. During the training process, the dung beetle optimization algorithm is used to optimize the smoothing factor σ of the initial DBO-GRNN prediction model. Specifically, the optimization process is to define the mean square error of each subsequence as the objective function for optimization, and use the smoothing factor corresponding to the minimum mean square error of each subsequence as the hyperparameter of the prediction model for that subsequence.

[0096] Furthermore, in this embodiment, the specific meaning of the dung beetle optimization algorithm is as follows:

[0097] (1) First, establish a dung beetle population. A population consisting of n dung beetles is represented as:

[0098]

[0099] In the formula, d is the dimension of the variable to be optimized, and n represents the number of dung beetle individuals.

[0100] The fitness values ​​of different dung beetle individuals are expressed as follows:

[0101]

[0102] Next, the initial dung beetle population is updated in position according to formulas for rolling behavior, dancing behavior, reproductive behavior, foraging behavior, and stealing behavior.

[0103] (2) Rolling ball behavior: The update formula for rolling ball behavior is as follows:

[0104]

[0105]

[0106] Where α is a number of -1 or 1; k is a random number ∈ (0, 0.2]; b is a constant taking the value (0, 1); x worst It represents the worst position of the dung beetle globally; Δx represents the change in simulated light intensity.

[0107] (2) Dancing behavior: The update formula for dancing behavior is as follows:

[0108]

[0109] (3) Reproductive behavior: The formula for updating reproductive behavior is as follows:

[0110] Lb * =max(x * ×(1-R),Lb)

[0111] Ub * =max(x * ×(1+R),Ub)

[0112] Where, x * Lb represents the current local optimum. * and Ub * Let Lb and Ub represent the lower and upper bounds of the egg-laying process, respectively, and let Lb and Ub represent the upper and lower bounds of the optimization problem.

[0113]

[0114] in, b1 and b2 are the position information of the i-th egg in the t-th iteration, and b1 and b2 are two independent random vectors of size 1×dim.

[0115] (4) Foraging behavior: The update formula for foraging behavior is as follows:

[0116] Lb b =max(x b ×(1-R),Lb)

[0117] Ub b =max(x b ×(1+R),Ub)

[0118] Where, x b Lb represents the globally optimal position. b and Ub b These represent the lower and upper bounds of the optimal foraging area, respectively. Therefore, the dung beetle's location is updated as follows:

[0119]

[0120] in, Let C1 represent the position information of the i-th dung beetle in the t-th iteration, C2 represent a random number that follows a normal distribution, and C3 represent a random vector in the range (0,1).

[0121] (5) Theft: The update formula for theft is as follows:

[0122]

[0123] in, Let g represent the position information of the i-th dung beetle thief in the t-th iteration, g is a 1×d im-dimensional random vector following a normal distribution, and S is a constant.

[0124] In this embodiment, the initial DBO-GRNN prediction model for each subsequence is constructed using the 10 modal components and 1 residual signal component obtained after decomposing each subsequence according to the above steps. The GRNN network structure of the initial DBO-GRNN prediction model includes an input layer, a mode layer, a summation layer, and an output layer. The time step of the input layer is used to determine the length of the variable time series for wind power prediction. The determination of this parameter must consider both the completeness of the prediction knowledge and the effectiveness of model training. On the one hand, an excessively short historical sequence length will lead to a lack of prediction knowledge, thus limiting the improvement of prediction accuracy; on the other hand, an excessively long historical sequence length will increase the difficulty of model training and reduce model prediction performance.

[0125] Because the prediction results are strongly correlated with the hyperparameters selected by the prediction model, the hyperparameters chosen by the prediction model are often different when predicting different historical power sequences. These parameters are mainly selected based on the researchers' experience, and have a high degree of subjectivity. To solve this problem, we consider using an optimization algorithm to optimize the smoothness factor of the prediction model. The mean square error (MSE) of the prediction results is selected as the fitness function, and the optimization objective is to minimize the MSE. The optimization steps for the initial DBO-GRNN prediction model are as follows:

[0126] Step 1: Initialize the model parameters of the initial DBO-GRNN prediction model and calculate the current MSE value;

[0127] Step 2: Update the position of the initial DBO-GRNN prediction model using the dung beetle optimization algorithm, and calculate the MSE value after the position update;

[0128] Step 3: Determine whether the updated MSE value is the minimum value. If it is, output the optimal smoothing factor parameter corresponding to the minimum MSE value. If not, repeat step 2 until the MSE value is the minimum.

[0129] Step 4: Assign the optimal smoothing factor parameters to the GRNN network of the updated model, and train the GRNN network using the training dataset until the trained DBO-GRNN prediction model is obtained.

[0130] This step optimizes the hyperparameters of the model, which improves the model's ability to capture patterns and thus enhances the accuracy of its predictions. Furthermore, based on the fast convergence and high accuracy of the dung beetle optimization algorithm, this algorithm is used for hyperparameter optimization to improve both the accuracy and efficiency of predictions. During the optimization process, a loss function is employed to prevent overfitting and further improve the model's accuracy.

[0131] Step 104: Perform power prediction on the subsequence according to the DBO-GRNN prediction model corresponding to each subsequence, obtain the prediction component corresponding to each subsequence, and superimpose the prediction components corresponding to each subsequence to obtain the wind power prediction result corresponding to the prediction time.

[0132] In this embodiment, the step includes: using the DBO-GRNN prediction model to predict the power of its corresponding subsequences respectively, obtaining the prediction component corresponding to each historical wind power sequence, and then superimposing the multiple prediction components corresponding to multiple historical wind power sequences to obtain the wind power prediction result corresponding to the prediction time.

[0133] Furthermore, in this embodiment, to verify the beneficial effects of the method provided in this embodiment, the wind power sequence obtained according to the present invention is used as the prediction parameter. Ablation experiments are conducted on three models: GRNN, DBO-GRNN, and CEEMDAN-DBO-GRNN, to obtain the predicted and actual values ​​corresponding to each model. The ablation experiments are as follows: Figure 6 ,Depend on Figure 6 As can be seen, compared with GRNN and DBO-GRNN, the prediction results of the DBO-GRNN prediction model proposed in this embodiment are closer to the true values.

[0134] On the other hand, embodiments of the present invention also disclose an ultra-short-term wind power prediction system, the specific structural composition of which can be referred to Figure 2 The system includes a time series analysis module 201, a mode decomposition module 202, a model optimization module 203, and a power prediction module 204.

[0135] The time series analysis module 201 is used to perform time delay characteristic analysis on each of the acquired first historical wind power time series data and a preset prediction time, obtain the correlation coefficient between each first historical wind power time series data and the prediction time, and filter the first historical wind power time series data according to the correlation coefficient to obtain a number of second historical wind power time series data.

[0136] The modal decomposition module 202 is used to obtain the historical wind power sequence corresponding to each of the several second historical wind power time series data, and to perform modal decomposition on each historical wind power sequence to obtain the subsequence corresponding to each historical wind power sequence; the subsequence includes modal components and residual signal components.

[0137] The model optimization module 203 is used to construct an initial DBO-GRNN prediction model corresponding to each subsequence based on the modal components and the remaining signal components, and to optimize the initial DBO-GRNN prediction model using a preset hyperparameter optimization algorithm to obtain the DBO-GRNN prediction model.

[0138] The power prediction module 204 is used to predict the power of the subsequence according to the DBO-GRNN prediction model corresponding to each subsequence, obtain the prediction component corresponding to each subsequence, and superimpose the prediction components corresponding to each subsequence to obtain the wind power prediction result corresponding to the prediction time.

[0139] In this embodiment, the time series analysis module 201 includes a coefficient calculation unit and a filtering unit.

[0140] The coefficient calculation unit is used to calculate the Pearson correlation coefficient between each historical wind power time series data and the prediction time according to the preset Pearson coefficient calculation function.

[0141] The filtering unit is used to compare the Pearson correlation coefficient with a preset coefficient threshold, and filter the first historical wind power time series data corresponding to the Pearson correlation coefficient being greater than the coefficient threshold as the second historical wind power time series data.

[0142] In this embodiment, the modal decomposition module 202 includes an iterative decomposition unit, a preprocessing unit, and a decomposition unit.

[0143] The iterative decomposition unit is used to perform several mode decompositions on each historical wind power sequence according to a preset mode decomposition algorithm and the number of iterations, to obtain several first mode components and several first residual signal components.

[0144] The preprocessing unit is used to preprocess each historical wind power sequence by adding white noise several times in each of the several mode decompositions to obtain a preprocessed power sequence.

[0145] The decomposition unit is used to decompose the preprocessed power sequence several times in each mode decomposition by a preset EMD to obtain a first residual signal component and several initial mode components, and to calculate the mean of the several initial mode components to obtain the first mode component.

[0146] In addition to the methods and systems described above, this invention also discloses an electronic device, including a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The memory is used to store computer programs. When the processor executes the program stored in the memory, it implements the ultra-short-term wind power prediction method described in this invention.

[0147] This invention provides an ultra-short-term wind power prediction method. The method involves analyzing the time-delay characteristics of the original wind power sequence, selecting time series with strong correlation to the prediction time for multi-series modeling, and using the CEEMDAN method to perform mode decomposition on the power sequence to obtain different sub-sequences. RMSE and MAPE are used as evaluation indicators for experimental testing to determine the prediction accuracy of the model. A DBO-GRNN prediction model is constructed from the mode components and residual components obtained from CEEMDAN decomposition. The smoothing factor σ of the GRNN model is optimized using the dung beetle optimization algorithm. The hyperparameter-optimized GRNN prediction model is used to predict different sub-sequences, and the prediction components are superimposed to obtain the final prediction result, thereby improving the accuracy and efficiency of wind power prediction.

[0148] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting ultra-short-term wind power, characterized in that, include: Each of the acquired first historical wind power time series data is subjected to time delay characteristic analysis with a preset prediction time. The correlation coefficient between each first historical wind power time series data and the prediction time is obtained. Based on the correlation coefficient, the first historical wind power time series data are filtered to obtain a number of second historical wind power time series data. Obtain the historical wind power sequence corresponding to each of the plurality of second historical wind power time series data, and perform mode decomposition on each historical wind power sequence to obtain the subsequence corresponding to each historical wind power sequence; the subsequence includes mode components and residual signal components; An initial DBO-GRNN prediction model is constructed for each subsequence based on the modal components and residual signal components. The initial DBO-GRNN prediction model is then optimized using a preset hyperparameter optimization algorithm to obtain the final DBO-GRNN prediction model. Specifically, the smoothing factor parameter in the initial DBO-GRNN prediction model is optimized using a preset dung beetle optimization algorithm. The loss between the predicted and actual values ​​of the model after smoothing factor parameter adjustment is continuously calculated based on preset mean squared error loss functions and mean absolute percentage error loss functions. When the loss value is less than a preset loss threshold, the final DBO-GRNN prediction model is obtained by combining the current smoothing factor parameter with the initial DBO-GRNN prediction model. Power prediction is performed on the subsequences according to the DBO-GRNN prediction model corresponding to each subsequence, and the prediction components corresponding to each subsequence are superimposed to obtain the wind power prediction result corresponding to the prediction time.

2. The ultra-short-term wind power prediction method as described in claim 1, characterized in that, The step of obtaining the correlation coefficient between each first historical wind power time series data and the predicted time, and filtering the plurality of first historical wind power time series data according to the correlation coefficient, includes: The Pearson correlation coefficient between each historical wind power time series data and the predicted time is calculated according to the preset Pearson coefficient calculation function. The Pearson correlation coefficient is compared with a preset coefficient threshold, and the first historical wind power time series data corresponding to the Pearson correlation coefficient being greater than the coefficient threshold is selected as the second historical wind power time series data.

3. The ultra-short-term wind power prediction method as described in claim 1, characterized in that, The step of performing mode decomposition on each historical wind power sequence to obtain the subsequences corresponding to each historical wind power sequence includes: Based on the preset mode decomposition algorithm and the number of iterations, each historical wind power sequence is subjected to several mode decompositions to obtain several first mode components and several first residual signal components. In each of the several mode decompositions, a preprocessed power sequence is obtained by adding several times of white noise to each historical wind power sequence. In each mode decomposition, the preprocessed power sequence is decomposed several times using a preset EMD to obtain a first residual signal component and several initial mode components. The mean of the several initial mode components is then calculated to obtain the first mode component.

4. The ultra-short-term wind power prediction method as described in claim 3, characterized in that, The step of performing mode decomposition on each historical wind power sequence several times according to a preset mode decomposition algorithm and number of iterations to obtain several first mode components and several first residual signal components includes: According to the preset CEEMDAN algorithm, each historical wind power sequence is subjected to iterative mode decomposition to obtain the first residual signal component and several first mode components. White noise is added to the first residual signal component obtained after several modal decompositions, and EMD decomposition is performed. It is then determined whether the second residual signal component obtained after the EMD decomposition is a monotonic signal. If the second residual signal component is a monotonic signal, the iteration is stopped.

5. The ultra-short-term wind power prediction method as described in claim 1, characterized in that, The step of performing mode decomposition on each historical wind power sequence to obtain the subsequences corresponding to each historical wind power sequence includes: By adding white noise several times to each historical wind power sequence, preprocessing is performed on each historical wind power sequence to obtain a preprocessed power sequence; wherein, the calculation formula for the preprocessing is: Among them, the Represents the original wind power sequence, This represents the sequence after the i-th addition of white noise. Represents the standard deviation of noise. This represents Gaussian white noise that follows an N(0,1) distribution, where T represents the number of times the noise is added. The preprocessed power sequence is decomposed several times using a preset EMD; wherein the EMD is expressed as follows: Among them, the The first modal component obtained from the decomposition, (i) is Modal components obtained by EMD decomposition The remaining signal components after one mode decomposition.

6. A short-term wind power prediction system, characterized in that, The system includes a time series analysis module, a mode decomposition module, a model optimization module, and a power prediction module; The time series analysis module is used to perform time delay characteristic analysis on each of the acquired first historical wind power time series data and a preset prediction time, obtain the correlation coefficient between each first historical wind power time series data and the prediction time, and filter the first historical wind power time series data according to the correlation coefficient to obtain a number of second historical wind power time series data. The mode decomposition module is used to obtain the historical wind power sequence corresponding to each of the several second historical wind power time series data, and to perform mode decomposition on each historical wind power sequence to obtain the subsequence corresponding to each historical wind power sequence; the subsequence includes mode components and residual signal components; The model optimization module is used to construct an initial DBO-GRNN prediction model corresponding to each subsequence based on the modal components and residual signal components, and to optimize the initial DBO-GRNN prediction model using a preset hyperparameter optimization algorithm to obtain a DBO-GRNN prediction model. Specifically, the smoothing factor parameter in the initial DBO-GRNN prediction model is optimized using a preset dung beetle optimization algorithm. The loss value between the predicted value and the true value of the model after adjusting the smoothing factor parameter is continuously calculated based on preset mean squared error loss function and mean absolute percentage error loss function. When the loss value is less than a preset loss value threshold, the DBO-GRNN prediction model is obtained by combining the current smoothing factor parameter with the initial DBO-GRNN prediction model. The power prediction module is used to predict the power of the subsequence according to the DBO-GRNN prediction model corresponding to each subsequence, obtain the prediction component corresponding to each subsequence, and superimpose the prediction components corresponding to each subsequence to obtain the wind power prediction result corresponding to the prediction time.

7. The ultra-short-term wind power prediction system as described in claim 6, characterized in that, The time series analysis module includes a coefficient calculation unit and a filtering unit; The coefficient calculation unit is used to calculate the Pearson correlation coefficient between each historical wind power time series data and the prediction time according to the preset Pearson coefficient calculation function. The filtering unit is used to compare the Pearson correlation coefficient with a preset coefficient threshold, and filter the first historical wind power time series data corresponding to the Pearson correlation coefficient being greater than the coefficient threshold as the second historical wind power time series data.

8. The ultra-short-term wind power prediction system as described in claim 6, characterized in that, The mode decomposition module includes an iterative decomposition unit, a preprocessing unit, and a decomposition unit; The iterative decomposition unit is used to perform several mode decompositions on each historical wind power sequence according to a preset mode decomposition algorithm and the number of iterations, to obtain several first mode components and several first residual signal components. The preprocessing unit is used to preprocess each historical wind power sequence by adding white noise several times in each of the several mode decompositions to obtain a preprocessed power sequence. The decomposition unit is used to decompose the preprocessed power sequence several times in each mode decomposition by a preset EMD to obtain a first residual signal component and several initial mode components, and to calculate the mean of the several initial mode components to obtain the first mode component.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the ultra-short-term wind power prediction method according to any one of claims 1-5.

Images