Ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention

Through the primary decomposition and WOA-CNN-BIGRU-Attention model, the problem of insufficient noise interference processing and ultra-short-term model capabilities in traditional wind power prediction is solved, and the accurate prediction of wind power is achieved, which improves the accuracy and stability of the prediction. .

CN118939998AInactive Publication Date: 2024-11-12殷得承
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410946093.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

When traditional wind power power prediction methods deal with signals with large noise interference, their accuracy and stability are poor, and the learning and expression capabilities of ultra-short-term wind power power prediction models are insufficient, which affects the prediction accuracy.

Method used

The wind speed signal is decomposed into subsequences of different scales by using the WOA-CNN-BIGRU-Attention model, using ICEEMDAN for signal decomposition and denoising, whale optimization algorithm for parameter optimization, CNN extracting features, and BIGRU learning time series dependency Relationship, add attention mechanisms to focus on key information, and build a more accurate ultra-short-term wind power power prediction model.

Benefits of technology

By extracting the spatiotemporal features at different time scales, effectively learning and representing the spatiotemporal information in the wind speed sequence, accurate prediction of wind power is achieved, and the accuracy and stability of the model are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939998A_ABST
    Figure CN118939998A_ABST
Patent Text Reader

Abstract

The invention relates to an ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention, and the method comprises the following steps: 1, inputting original wind power time sequence data, firstly carrying out the data standardization processing of an original sequence, so as to eliminate the data dimension, prevent the numerical explosion of a subsequent model in the training process, and finally, carrying out the data standardization processing of the original sequence; the method comprises the steps that firstly, wind power data is acquired, then noise components existing in the wind power data are considered, an ICEEMDAN method is used for conducting decomposition and noise reduction on the original wind power data, a plurality of subsequences are obtained, wind speed signals are decomposed into the subsequences of different scales through the one-time decomposition technology, and therefore spatial-temporal characteristics under different time scales are extracted; according to the method, a WOA-CNN-BIGRU-Attention model is introduced, and a whale optimization algorithm (WOA) and an attention mechanism are combined to effectively learn and represent space-time information in a wind speed sequence, so that accurate prediction of wind power is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of wind power prediction, and in particular to an ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention. Background Art

[0002] In recent years, the development of deep learning technology has provided new ideas and methods for wind power forecasting. In particular, deep learning models such as convolutional neural networks (CNN) and recurrent neural networks (RNN) can effectively process time series data and have strong feature extraction and modeling capabilities.

[0003] The existing wind power forecasting has the following problems: (1) In the traditional signal decomposition process, the accuracy and stability are not very good when processing signals with large noise interference; (2) In the process of constructing an ultra-short-term wind power forecasting model, the learning ability and expression ability of the ultra-short-term wind power forecasting model are general, which affects the accuracy of the ultra-short-term wind power forecasting model.

[0004] Therefore, we propose an ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention. We use ICEEMDAN for signal decomposition and denoising, and then use WOA for parameter optimization. Next, we combine CNN to extract the features of time series data, BIGRU to learn the dependencies between time series, and add an attention mechanism to focus on key information. Through this comprehensive approach, we build a more accurate ultra-short-term wind power prediction model. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art, adapt to actual needs, and provide an ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention to solve the problem that in the traditional signal decomposition process, the accuracy and stability are not very good when processing signals with large noise interference; in the process of constructing an ultra-short-term wind power prediction model, the learning ability and expression ability of the ultra-short-term wind power prediction model are general, which leads to technical problems affecting the accuracy of the ultra-short-term wind power prediction model.

[0006] In order to achieve the purpose of the present invention, the technical solution adopted by the present invention is: designing an ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention, comprising the following steps:

[0007] Step 1: Input the original wind power time series data. First, standardize the original sequence to eliminate the data dimension and prevent numerical explosion in the subsequent model training process. Then, considering the noise components in the wind power data, use the ICEEMDAN method to decompose and reduce the noise of the original wind power data to obtain several subsequences.

[0008] Step 2: Use fuzzy entropy to evaluate the sequence complexity of each subsequence, and then group and reconstruct the subsequences according to the zero-crossing rate to form three types of components: high frequency, medium frequency, and low frequency, which are represented as x1(t), x2(t), and x3(t) respectively. Finally, use the sliding window mechanism to construct training sample data using the three components respectively, which will serve as training data for subsequent single prediction model building.

[0009] Step 3: According to the different frequency characteristics of each component after reconstruction, different types of models are selected to construct a single prediction model, and the prediction results of the single prediction model are obtained. In the study, for high-frequency components, due to their strong volatility, CNN-BIGRU-Attention with strong feature extraction capabilities is used to model them, and a single prediction model for high-frequency components is obtained, and its prediction results for high-frequency components are recorded as For the mid-frequency component, time CNN-BIGRU-Attention is also used to model and predict it, forming a single prediction model for the mid-frequency component and obtaining its prediction results for the mid-frequency component. For the low-frequency component, since its fluctuation is relatively stable, MLR is used to obtain the prediction results of the low-frequency component.

[0010] Step 4: Use the whale optimization algorithm WOA to determine the optimal combination weights w1, w2, and w3 of the three single prediction models, and then perform weighted summation on the prediction results of the three single prediction models to obtain the final prediction result. It can be expressed as:

[0011]

[0012] Step 5: Analyze the effectiveness and stability of the model, mainly using three evaluation indicators: root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) to analyze the effectiveness of the model.

[0013] Preferably, the whale optimization algorithm WOA includes shrinking surround shrinking search, spiral predation search and individual independent search. By simulating the behavior of whales searching for prey, the WOA algorithm can effectively find the global optimal solution or a solution close to the optimal solution in the search space.

[0014] Preferably, the ICEEMDAN is used to decompose non-stationary and nonlinear signals, which more effectively handles noise in the data and provides more accurate and robust signal components.

[0015] Preferably, the CNN is used to extract features from time series data to help understand complex patterns in the data.

[0016] Preferably, the BIGRU is a recurrent neural network structure that can capture long-term dependencies in sequence data. Through a bidirectional design, the BIGRU considers not only past information but also future information, thereby improving the expressiveness of the model.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. The present invention decomposes the wind speed signal into subsequences of different scales through a one-step decomposition technique, thereby extracting the spatiotemporal features at different time scales. Then, the WOA-CNN-BIGRU-Attention model is introduced, combining the whale optimization algorithm (WOA) and the attention mechanism to effectively learn and represent the spatiotemporal information in the wind speed sequence, thereby achieving accurate prediction of wind power. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a data preprocessing flow chart of the present invention;

[0020] Figure 2 This is a schematic diagram of the ICEEMDAN decomposition of the 10m wind tower data of the present invention;

[0021] Figure 3 It is a schematic diagram of the prediction results and true values ​​of the GRU model of the present invention;

[0022] Figure 4 Schematic diagram of the prediction results and true values ​​of the BiGRU model of the present invention;

[0023] Figure 5 It is a schematic diagram of the prediction results and true values ​​of the CNN-BIGRU model of the present invention;

[0024] Figure 6 It is a schematic diagram of the prediction results and true values ​​of the CNN-BIGRU-Attention model of the present invention;

[0025] Figure 7 This is the prediction result and real schematic diagram of the WOA-CNN-BIGRU-Attention model of the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described below in conjunction with the accompanying drawings and embodiments:

[0027] Design an ultra-short-term wind power forecasting method based on primary decomposition and WOA-CNN-BIGRU-Attention, see Figures 1 to 7 , including the following steps:

[0028] Step 1: Input the original wind power time series data. First, standardize the original sequence to eliminate the data dimension and prevent numerical explosion in the subsequent model training process. Then, considering the noise components in the wind power data, use the ICEEMDAN method to decompose and reduce the noise of the original wind power data to obtain several subsequences.

[0029] Step 2: Use fuzzy entropy to evaluate the sequence complexity of each subsequence, and then group and reconstruct the subsequences according to the zero-crossing rate to form three types of components: high frequency, medium frequency, and low frequency, which are represented as x1(t), x2(t), and x3(t) respectively. Finally, use the sliding window mechanism to construct training sample data using the three components respectively, which will serve as training data for subsequent single prediction model building.

[0030] Step 3: According to the different frequency characteristics of each component after reconstruction, different types of models are selected to construct a single prediction model, and the prediction results of the single prediction model are obtained. In the study, for high-frequency components, due to their strong volatility, CNN-BIGRU-Attention with strong feature extraction capabilities is used to model them, and a single prediction model for high-frequency components is obtained, and its prediction results for high-frequency components are recorded as For the mid-frequency component, time CNN-BIGRU-Attention is also used to model and predict it, forming a single prediction model for the mid-frequency component and obtaining its prediction results for the mid-frequency component. For the low-frequency component, since its fluctuation is relatively stable, MLR is used to obtain the prediction results of the low-frequency component.

[0031] Step 4: Use the whale optimization algorithm WOA to determine the optimal combination weights w1, w2, and w3 of the three single prediction models, and then perform weighted summation on the prediction results of the three single prediction models to obtain the final prediction result. It can be expressed as:

[0032]

[0033] Step 5: Analyze the effectiveness and stability of the model, mainly using three evaluation indicators: root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) to analyze the effectiveness of the model.

[0034] For details, see Figures 1 to 7,The whale optimization algorithm WOA includes shrinking surround shrinking search, spiral predation search and individual independent search. By simulating the behavior of whales searching for prey, the WOA algorithm can effectively find the global optimal solution or a solution close to the optimal solution in the search space.

[0035] For further information, see Figures 1 to 7 ,ICEEMDAN is used to decompose non-stationary and nonlinear signals, which handles the noise in the data more effectively and provides more accurate and robust signal components.

[0036] It is worth noting that see Figures 1 to 7 ,CNN is used to extract features from time series data and help understand complex patterns in the data.

[0037] It is worth noting that see Figures 1 to 7 ,BIGRU is a recurrent neural network structure that can capture long-term dependencies in sequence data.,Through a bidirectional design, BIGRU considers not only past information, but also future information,,thus improving the expressive power of the model.

[0038] Embodiment 1:

[0039] Improved adaptive noise empirical mode decomposition principle

[0040] ICEEMDAN is a method for signal processing and is an improved version of the classic EMD. The principle of ICEEMDAN is as follows:

[0041] ICEEMDAN provides better noise control by estimating the noise level in the signal. This adaptability allows ICEEMDAN to handle noise in different signals more effectively. ICEEMDAN introduces an ensemble strategy to get more stable and accurate results by running EMD multiple times. This ensemble method helps reduce the impact of random errors and improves the accuracy of signal decomposition. ICEEMDAN uses a weight calculation method that dynamically assigns weights to each intrinsic mode function (IMF) according to the characteristics of the signal to reduce the impact of noise on signal decomposition. ICEEMDAN uses symmetric extension boundary processing technology to effectively reduce the impact of boundary effects on signal decomposition and improve the accuracy of decomposition results. After decomposing the signal, ICEEMDAN can restore the original signal by reconstructing the extracted IMF while removing noise and unwanted components. ICEEMDAN provides a more accurate and stable signal decomposition method through the combination of noise adaptive control, integration ideas, weight calculation, symmetric extension boundary processing and other technologies, which has certain advantages in processing signals with large noise interference.

[0042] Embodiment 2:

[0043] Whale Optimization Algorithm

[0044] The Whale Optimization Algorithm (WOA) is a nature-inspired algorithm that simulates the foraging behavior of whales and aims to find the optimal solution by simulating the behavior of whales. The basic principles of this algorithm are as follows:

[0045] For the WOA algorithm, there are three search methods in the algorithm, namely, contraction and encirclement search, spiral predation search and individual independent search. Among these three search methods, contraction and encirclement search simulates the humpback whale group to contract and encircle the prey target in a certain area, spiral predation search simulates the humpback whale group to optimize the search through the spiral predation strategy, and the individual independent search mechanism allows each individual in the humpback whale group to independently search and update the optimal position. The algorithm assumes that the population size is N, and the position of the i-th whale in the d-dimensional space is the optimal solution to the problem of the prey position, see formula (3-1)

[0046]

[0047] The first search method of WOA simulates the shrinking and encircling mechanism of whales, that is, the encirclement of the population to the optimal position, and updates the population position according to formula (3-2):

[0048] X(t+1)=X * (t)+A|C*X * (t)-X(t)|(3-2)

[0049] Where: t is the current iteration number, X*(t) is the target position, X*(t) is the current optimal individual during the iteration process, A*|C*X*(t)-X(t)| is the convergence amplitude, and A and C are defined as shown in equations (3-3) and (3-4):

[0050] A=2a*ra

[0051] (3-3)

[0052] C=2r

[0053] (3-4)

[0054] T is a random vector on [0, 1], and the surrounding of the prey is simulated by updating the r population individuals in any position in the space around the current optimal point. a is the convergence factor, t is the number of iterations, and the relationship between a and t is shown in formula (3-5)

[0055]

[0056] Where: tmax is the maximum number of iterations.

[0057] The second search method of WOA simulates the spiral foraging mechanism of whales, which updates the position in a spiral. Its mathematical model is as follows:

[0058] X(t+1)=X * (t)+D*e br *cos(2πr)

[0059] (3-6)

[0060] Where: D = |X * (t)-X(t)| represents the distance between the i-th whale and the target, the constant that limits the shape of the logarithmic spiral is represented by b, and r is a random number on [-1,1].

[0061] When a whale is hunting, the process of encircling its prey and spiraling up is simultaneous, so it is assumed that both mechanisms have a 50% chance of occurring. Therefore, given p∈[0,1], the whale's position update is expressed as formula (3-7):

[0062]

[0063] In addition, in addition to the traditional search and prey hunting methods, the algorithm also incorporates the behavior of individual whales randomly searching for prey based on their own positions, thereby increasing the diversity of the search space, as shown in formula (3-8):

[0064] X(t+1)=X random (t)-A*|C*C random -X(t)|(3-8)

[0065] Where: X random (t) The individual whale positions are randomized.

[0066] By simulating the behavior of whales searching for prey, the WOA algorithm can effectively find the global optimal solution or a solution close to the optimal solution in the search space. This makes WOA a powerful optimization algorithm that can be applied to solve a variety of problems.

[0067] Embodiment 3:

[0068] Ultra-short-term wind power forecasting method based on ICEEMDAN-WOA-CNN-BIGRU-Attention

[0069] The ultra-short-term wind power forecasting method based on ICEEMDAN-WOA-CNN-BIGRU-Attention involves the combination of multiple technologies and models. ICEEMDAN is used to decompose non-stationary and nonlinear signals. It can more effectively handle noise in the data and provide more accurate and robust signal components.

[0070] WOA is a heuristic optimization algorithm inspired by the foraging behavior of whales. This algorithm can be applied to the parameter optimization process to help adjust model parameters to improve prediction accuracy. Convolutional neural network is a deep learning model suitable for processing data with spatial correlation. In wind power forecasting, CNN can be used to extract features from time series data to help understand complex patterns in the data. BIGRU is a recurrent neural network structure that can capture long-term dependencies in sequence data. Through a bidirectional design, BIGRU considers not only past information but also future information, thereby improving the expressiveness of the model. The attention mechanism allows the model to focus on key parts when processing sequence data, improving the model's ability to distinguish the importance of different time steps. In wind power forecasting, the attention mechanism helps the model better understand the influencing factors at different time points.

[0071] The basic idea of ​​this method is to first use ICEEMDAN for signal decomposition and denoising, and then use WOA for parameter optimization. Next, CNN is combined to extract the features of time series data, BIGRU is used to learn the dependencies between time series, and an attention mechanism is added to focus on key information. Through this comprehensive approach, a more accurate ultra-short-term wind power prediction model can be constructed.

[0072] Embodiment 4:

[0073] Regression model based on WOA to optimize CNN-BIGRU-Attention parameters

[0074] Optimizing CNN-BIGRU-Attention Parameters Based on WOA is a method that combines evolutionary algorithms and deep learning for regression tasks. ,WOA is used to adjust the parameters of CNN, BIGRU, and Attention models to maximize the performance indicators of the regression model, such as prediction accuracy or minimization of loss function.

[0075] Convolutional neural networks are used to process input data, such as time series data or image data, to extract features. In this regression model, CNN may be used to extract features from the input data, which can be used in subsequent regression tasks.

[0076] BIGRU is a recurrent neural network structure suitable for sequence data, which can effectively capture temporal dependencies. In this model, BIGRU may be used to learn and model temporal patterns in time series data and complex relationships between features.

[0077] The attention mechanism is used to strengthen the model's attention to different parts of the input data, so that the model can more centrally process information that contributes significantly to the regression task. In this model, the attention mechanism may be used to strengthen the focus on certain features or time points, thereby improving the performance of the regression model.

[0078] The workflow of the entire model: First, CNN is used to extract feature representations of the input data. Then, these feature representations are input into BIGRU to capture the temporal patterns in time series data. Next, the attention mechanism is used to strengthen the focus on specific features or time points. Finally, the parameters of CNN, BIGRU and Attention models are optimized through the WOA algorithm to maximize the performance indicators of the regression model.

[0079] This WOA-optimized CNN-BIGRU-Attention regression model can provide better performance in regression tasks because it combines the advantages of deep learning and evolutionary algorithms to better capture complex patterns and relationships in the data.

[0080] Embodiment 5:

[0081] Ultra-short-term wind power forecasting process based on ICEEMDAN-WOA-CNN-BIGRU-Attention

[0082] (1) Data preprocessing

[0083] Wind power time series is characterized by volatility and randomness. Directly using the original data for modeling and prediction will result in low model prediction accuracy. Therefore, in order to better obtain effective information in the original data and remove the impact of noise in the original data on the prediction performance of the combined model, data decomposition and reconstruction are used to preprocess the original data to provide data support for subsequent modeling. The data preprocessing flow chart is shown in the figure. Figure 1 shown.

[0084] (2) Data decomposition and noise reduction

[0085] When modeling, the ICEEMDAN algorithm is first used to decompose and denoise the original wind power data. The experimental data is selected from the Ali Tianchi dataset, which provides a variety of datasets in different fields and difficulty levels for data scientists, machine learning engineers, and researchers. The wind power dataset from Ali Tianchi in January 2019 is used as an example to illustrate the data preprocessing process in this article. The original data uses 15 minutes as the wind power data collection interval. 96 wind power data are collected every day, and a total of 3,648 data from January 1 to February 7.

[0086] (3) Data reconstruction

[0087] After ICEEMDAN decomposition, the various modes of the original data are displayed. However, if modeling and prediction are performed for each subsequence separately, the accumulation of prediction errors of each model will lead to a larger final prediction error due to the use of multiple models. In order to better utilize the frequency characteristics of each subsequence obtained by decomposition and reduce the noise of wind power data, this paper uses zero-crossing rate as a time series complexity calculation method.

[0088] Embodiment 6:

[0089] Simulation Analysis

[0090] The selected data set simulates the wind farm data from January 1, 2019 to February 7, 2019, and records them every 15 minutes. According to the "Wind Energy Resource Assessment Method for Wind Farms", the wind power values ​​that do not meet the requirements are eliminated. The data has the characteristics of strong randomness and non-stationarity, and directly using it as the input data of the model will affect its prediction accuracy. Therefore, the ICEEMDAN algorithm is used to decompose the wind power sequence to improve the prediction accuracy of the model. The selected data set contains the wind speed and wind direction of the wind tower at 10m, 30m, 50m, and 70m, and includes the temperature, air pressure and humidity during measurement. Finally, the accurate power generation for each period is given, and the total training set has 3648 samples.

[0091] The original wind power sequence is decomposed by ICEEMDAN to generate a series of IMF components with different complexities. Taking the wind speed of a 10m wind tower as an example, the original wind power sequence is decomposed by ICEEMDAN to generate 11 IMF subsequences IMF1 to IMF11, as shown in Figure 2 shown.

[0092] In order to verify the effectiveness of the ultra-short-term wind power prediction model based on WOA-CNN-BIGRU-Attention in this chapter, comparative experiments were conducted using GRU, BIGRU, CNN-BIGRU, CNN-BIGRU-Attention, and WOA-CNN-BIGRU-Attention models to verify the optimization capabilities of model parameters in different optimization algorithms after the data has been decomposed once. Figure 3 , 4 , 5, 6, and 7 as shown.

[0093] In order to better reflect the prediction accuracy of each prediction model, the evaluation index value of each prediction model is calculated, as shown in Table 3-1:

[0094]

[0095] Table 3-1 Evaluation index values ​​of the prediction model

[0096] The ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention first uses the ICEEMDAN algorithm to decompose and denoise the original wind power data, and then groups and reconstructs it according to the zero-crossing rate subsequence to form three types of components: high frequency, medium frequency and low frequency. According to the different frequency characteristics of each component after reconstruction, different types of models are selected to construct a single prediction model, and the prediction results of the single prediction model are obtained. Finally, a short-term wind power prediction model based on WOA-CNN-BIGRU-Attention is established, and the wind power series is predicted. The simulation results show that the model can improve the prediction accuracy.

[0097] In addition, the components designed in the present invention are all universal standard parts or components known to technical personnel in this field. Their structures and principles can be known to technical personnel through technical manuals or through conventional experimental methods. They can be fully implemented by technical personnel in this field. Needless to say, the content protected by the present invention does not involve improvements to internal structures and methods.

[0098] The embodiments of the present invention disclose preferred embodiments, but are not limited thereto. A person skilled in the art can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. However, as long as they do not deviate from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. An ultra-short-term wind power forecasting method based on primary decomposition and WOA-CNN-BIGRU-Attention, characterized in that: The following steps are involved: Step 1: Input the original wind power time series data. First, standardize the original sequence to eliminate the data dimension and prevent numerical explosion in the subsequent model training process. Then, considering the noise components in the wind power data, use the ICEEMDAN method to decompose and reduce the noise of the original wind power data to obtain several subsequences. Step 2: Use fuzzy entropy to evaluate the sequence complexity of each subsequence, and then group and reconstruct the subsequences according to the zero-crossing rate to form three types of components: high frequency, medium frequency, and low frequency, which are represented as x1(t), x2(t), and x3(t) respectively. Finally, use the sliding window mechanism to construct training sample data using the three components respectively, which will serve as training data for subsequent single prediction model building. Step 3: According to the different frequency characteristics of each component after reconstruction, different types of models are selected to construct a single prediction model, and the prediction results of the single prediction model are obtained. In the study, for high-frequency components, due to their strong volatility, CNN-BIGRU-Attention with strong feature extraction capabilities is used to model them, and a single prediction model for high-frequency components is obtained, and its prediction results for high-frequency components are recorded as For the mid-frequency component, time CNN-BIGRU-Attention is also used to model and predict it, forming a single prediction model for the mid-frequency component and obtaining its prediction results for the mid-frequency component. For the low-frequency component, since its fluctuation is relatively stable, MLR is used to obtain the prediction results of the low-frequency component. Step 4: Use the whale optimization algorithm WOA to determine the optimal combination weights w1, w2, and w3 of the three single prediction models, and then perform weighted summation on the prediction results of the three single prediction models to obtain the final prediction result. It can be expressed as: Step 5: Analyze the effectiveness and stability of the model, mainly using three evaluation indicators: root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) to analyze the effectiveness of the model.

2. The ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention according to claim 1, characterized in that: The Whale Optimization Algorithm (WOA) includes shrinking, surrounding, shrinking search, spiral predation search, and individual independent search. By simulating the behavior of whales searching for prey, the WOA algorithm can effectively find the global optimal solution or a solution close to the optimal solution in the search space.

3. The ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention according to claim 1, characterized in that: The ICEEMDAN is used to decompose non-stationary and nonlinear signals, which handles the noise in the data more effectively and provides more accurate and robust signal components.

4. The ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention according to claim 1, characterized in that: The CNN is used to extract features from time series data, helping to understand complex patterns in the data.

5. The ultra-short-term wind power prediction method based on primary decomposition and WOA-CNN-BIGRU-Attention according to claim 1, characterized in that: The BIGRU is a recurrent neural network structure that can capture long-term dependencies in sequence data. Through a bidirectional design, the BIGRU considers not only past information but also future information, thereby improving the expressiveness of the model.

Citation Information

Cited By

  • Short-term wind power prediction method and device, electronic equipment and storage medium

    CN119809138A

  • Short-term wind power prediction method, device, electronic equipment and storage medium

    CN119809138B