A method for ultra-short-term wind speed forecasting based on clustering of wind speed fluctuation patterns

Abstract

The present invention provides an ultra-short-term wind speed forecasting method based on wind speed fluctuation pattern clustering, comprising the following steps: 1. Collecting historical measured wind speed data recorded by a wind farm data acquisition and monitoring control system; 2. Decomposing the original sequence into low-frequency and high-frequency sequences using discrete wavelet transform, and further decomposing the high-frequency sequence using an improved fully adaptive noise ensemble empirical mode decomposition; 3. Calculating wind speed fluctuation intensity and wind speed fluctuation rate characteristics to quantify the fluctuation of wind speed data; 4. Constructing an adaptive wind speed interval partitioning model, extracting wind speed fluctuation characteristics, performing wind speed classification, and labeling wind speed segments; 5. Constructing two basic prediction models using an improved Beluga optimization algorithm; and 6. Outputting test results and verifying the feasibility of the model. The prediction method proposed in the present invention can effectively improve the accuracy of wind speed prediction values, significantly enhancing the reliability of wind speed prediction values.

Description

Technical Field

[0001] The present invention belongs to the field of wind power prediction and wind speed prediction, and in particular relates to an ultra-short-term wind speed prediction method based on clustering division of wind speed fluctuation patterns. Background Art

[0002] The biggest challenges facing global energy supply include security, price volatility, and greenhouse gas emissions, all of which are primarily due to the dominance of fossil fuels in the energy supply structure. The world is undergoing a gradual energy transition, with countries striving to reduce their reliance on fossil fuels and shift their energy mix toward renewable energy. Wind power is considered crucial to the world's future energy supply. However, due to the randomness and intermittency of wind power, the poor stability of wind power generation poses significant challenges to grid integration. High-precision wind speed forecasting can reduce wind power failures and ensure the safety and stability of the power system.

[0003] Current wind speed forecasting research primarily focuses on optimizing and improving models or refining modeling schemes. This results in high modeling costs and a cumbersome updating process. Furthermore, current wind speed feature extraction methods, typically using filtering and attribute reduction, have significant limitations. These methods make it difficult to effectively address the prediction errors caused by wind speed characteristics, thus failing to fundamentally resolve the error problem. Summary of the Invention

[0004] Purpose of the Invention: This invention addresses the shortcomings of existing technologies by providing an ultra-short-term wind speed forecasting method based on clustering of wind speed fluctuation patterns. This method combines wind speed volatility, fluctuation intensity, a sliding window method, a two-stage decomposition method based on an information leakage prevention mechanism, and a multi-model combination approach. Furthermore, the method utilizes a GRU and BiLSTM algorithm optimized using an improved Beluga optimization algorithm to achieve ultra-short-term wind speed forecasting, effectively improving the accuracy of wind speed predictions.

[0005] The present invention comprises the following steps:

[0006] Step 1: Collect historical measured wind speed data recorded by the wind farm data acquisition and monitoring control system;

[0007] Step 2: An information leakage prevention mechanism-based two-stage decomposition (ILPMTSD) is used to decompose the original sequence into low-frequency and high-frequency sequences using the Discrete Wavelet Transformation (DWT). The high-frequency sequence is further decomposed using the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN). An information leakage prevention mechanism is also used to prevent information leakage during the decomposition process.

[0008] Step 3: Calculate the wind speed fluctuation intensity and wind speed fluctuation rate characteristics to quantify the fluctuation of wind speed data;

[0009] Step 4: Based on wind speed characteristic statistics, a sliding window method, and a K-means cluster analysis algorithm, an adaptive wind speed interval segmentation model (AISM) is constructed. The AISM processes the low-frequency components (approximate components) after discrete wavelet transform processing, extracts wind speed fluctuation characteristics, performs wind speed classification, and labels wind speed segments.

[0010] Step 5: Use the Improved Beluga Whale Optimization algorithm to build two basic prediction models: a bidirectional long short-term memory network (BiLSTM) and a gated recurrent unit (GRU). For wind speed segments with different degrees of fluctuation, the corresponding basic prediction models are used for prediction.

[0011] Step 6: For the high-frequency component (detail component) processed by the improved fully adaptive noise ensemble empirical mode decomposition ICEEMDAN, a bidirectional long short-term memory network BiLSTM is used to predict it, and the prediction results are merged with the low-frequency component (approximate component). The test results are output and the evaluation indicators MAE, RMSE, SMAPE and , verify the feasibility of the model.

[0012] Step 2 includes:

[0013] Step 2-1, in the first stage of the two-stage decomposition, the original wind speed data is decomposed into low-frequency components (approximate components) and high-frequency components (detail components) using discrete wavelet transform. The formula is:

[0014] ,

[0015] ,

[0016] in, and Represent the low-frequency component (approximate component) and the high-frequency component (detail component), k represents the index variable of the summation, is the original wind speed data, represents a low-pass filter, represents a high-pass filter, Indicates the decomposition level;

[0017] Step 2-2: In the second stage, the improved fully adaptive noise ensemble empirical mode decomposition is used to further decompose the high-frequency components (detail components) generated in the first stage and convert them into subsequences. The specific steps include:

[0018] Step 2-2-1, add white noise to the signal to be decomposed x:

[0019] ,

[0020] in, represents the white noise added to group i, is the first empirical mode component EMD of white noise, represents the signal-to-noise ratio of the first decomposition, represents the signal after adding white noise;

[0021] Step 2-2-2, the first residual The calculation formula is:

[0022] ,

[0023] in Represents the local average envelope, which is used to determine the intrinsic mode function IMF;

[0024] Step 2-2-3, determine the initial eigenmode function of the improved fully adaptive noise ensemble empirical mode decomposition ICEEMDAN by decomposing N signals as follows: :

[0025] ;

[0026] Step 2-2-4, determine the second residual and the second modality :

[0027] ,

[0028] ;

[0029] in, represents the signal-to-noise ratio of the second decomposition, is the second empirical mode component of white noise;

[0030] Step 2-2-5, kth residual and the kth mode Determined as:

[0031] ,

[0032] ;

[0033] in, represents the signal-to-noise ratio of the kth decomposition, represents the k-1th residual, represents the kth empirical mode component of white noise;

[0034] Step 2-2-6, continue to perform step 2-2-5 for the next k stages until the residual is no longer decomposed;

[0035] In steps 2-3, to address data leakage during the decomposition process, we used a sliding window method. After multiple experiments, we implemented a fixed-width window sliding over the original sequence. Each time new data is added, the old data in the window is removed. This ensures that the number of subsequences generated after decomposition remains constant, and no data information is leaked. The window width is set to 10% of the original data length.

[0036] Step 3 includes:

[0037] Step 3-1, calculate wind speed fluctuation rate :

[0038] ,

[0039] Where N represents the number of wind speed data. Represents the e-th wind speed data;

[0040] Step 3-2, calculate the wind speed fluctuation intensity :

[0041] ,

[0042] in Indicates the mean wind speed.

[0043] Step 4 includes:

[0044] In step 4-1, the low-frequency component is segmented using the adaptive wind speed segmentation model (AISM) with a window size of 4 hours (16 data points) and a sliding step size of 1 step (15 minutes). The specific process includes:

[0045] Step 4-1-1, construct a set of wind speed data with a length of M :

[0046] ,

[0047] Step 4-1-2, define each sliding window:

[0048] ,

[0049] in, Represents the mth sliding window, each window length is 16;

[0050] Step 4-2, each sliding window As a group of wind speed sequence segments, the wind speed fluctuation rate of each wind speed sequence segment is extracted and recorded as and wind speed fluctuation intensity is recorded as , merge the processed features to form a feature matrix :

[0051] ,

[0052] in, and They represent the wind speed fluctuation rate and wind speed fluctuation intensity of the mth wind speed sequence segment respectively;

[0053] Step 4-3: Combine the two fluctuation characteristics of each wind speed data segment and As input, a K-means cluster analysis algorithm is used to obtain the classification criteria for fluctuating wind speed intervals and stable wind speed intervals, and to mark the wind speed sequence segments. The K-means cluster analysis algorithm includes the following steps:

[0054] Step 4-3-1, for an unlabeled data set , select Cluster centroids ,form clusters; Indicates the unlabeled data, Indicates the cluster centroids, is the total number of unlabeled data, is the total number of cluster centroids;

[0055] Step 4-3-2, for each unlabeled data , we need to calculate the centroid of each cluster distance, Belong to The nearest cluster centroid to itself Cluster :

[0056] ,

[0057] in, Indicates that the The function with the smallest corresponding function value of clusters, For each data point Assign to the nearest cluster centroid The corresponding cluster middle, Indicates the data points and The square of the Euclidean distance between clusters;

[0058] Step 4-3-3, update the cluster centroid of each cluster:

[0059] ,

[0060] in, is an indicator function, when When the indicator function value is 1, otherwise it is 0; this function is used to select data points, Indicates that all clusters The sum of the coordinates of the data points, Indicates belonging to a cluster The total number of data points;

[0061] Step 4-3-4, repeat steps 4-3-2 to 4-3-3 until the center of mass position no longer changes or the preset number of iterations is reached;

[0062] In step 4-3-5, the goal of K-means is to minimize the sum of squared errors within the cluster, which is expressed as:

[0063] ,

[0064] Where J represents each data point to the corresponding cluster centroid By minimizing the sum of squared distances, the K-means algorithm attempts to find a way to partition the data points so that the data points in each cluster are as tightly packed around its centroid as possible, while the clusters are as separated as possible.

[0065] Step 4-4, after the K-means cluster analysis algorithm, the wind speed fluctuation label matrix is ​​obtained :

[0066] ,

[0067] in, represents the mth wind speed sequence segment, Indicates the wind speed fluctuation mark of the corresponding wind speed sequence segment. According to the wind speed fluctuation mark, the original wind speed sequence segment is divided into two sets:

[0068] ,

[0069] ,

[0070] in, and are respectively the set of stable wind speed data segments and the set of unstable wind speed data segments, and They represent the subscript set of stable sequences and the subscript set of unstable sequences respectively.

[0071] In step 5, in the unstable wind speed data segment, the bidirectional long short-term memory network BiLSTM is used as the benchmark model for fitting; in the stable wind speed data segment, the gated recurrent unit GRU is used as the benchmark model for fitting.

[0072] In step 5, the improved Beluga optimization algorithm includes the following steps:

[0073] Step 5-1, the White Whale Optimization Algorithm includes three stages: exploration, development, and whale fall; the White Whale Optimization Algorithm is based on the balance factor To achieve the transition from exploration to development of the population, when the balance factor When the balance factor When , it is the development stage, and the calculation formula is:

[0074] ,

[0075] Where T is the current iteration number, is the total number of iterations, parameter is a random number in the range (0, 1) that changes in each iteration;

[0076] Step 5-2: Initialize the beluga population. In the beluga optimization algorithm, a single beluga is considered a search example. The initial population contains n beluga whales, which will be updated during the optimization process and is expressed as follows:

[0077] ,

[0078] Among them, X represents the position matrix of the beluga whale, represents the location information of a specific beluga whale in the corresponding dimension, n is the population size of the beluga whale, and d represents the dimension of the variable;

[0079] For all beluga whales, the corresponding fitness values ​​are stored as:

[0080] ,

[0081] in, is the fitness function, which takes the position of the beluga whale in each dimension as input. is a vector representing the fitness of all beluga whales;

[0082] Step 5-3, the exploration phase of beluga optimization simulates two beluga whales swimming closely in a synchronous or mirrored manner. The position update formula of the beluga whale is:

[0083] ,

[0084] Among them, even and odd represent the case when the variable dimension index j is an even number and the case when j is an odd number, respectively. Indicates that in the next iteration, Individuals in Position on the dimension; p is a random integer. Suppose the problem dimension is D, and p is a random integer in the range [1, D]. Indicates the position of the i-th individual on the random dimension p in the current iteration; set the population size to N, and r to be a random integer in the range [1, N]; therefore, Indicates the position of the rth random individual on the random dimension p in the current iteration; , All are random numbers between (0,1);

[0085] Step 5-4, the development phase of Beluga optimization, is inspired by the hunting behavior of Beluga whales and is expressed as the following equation:

[0086] ,

[0087] in, is the current position of the i-th beluga whale, is the current position of a random beluga whale, is the new position of the i-th beluga whale, Best location in Moby Dick; and is a random number between (0,1), is the random jump intensity; is a Levy flight function, calculated using the following formula:

[0088] ,

[0089] ,

[0090] in, is an intermediate calculation parameter, and is a normally distributed random number, is a default constant equal to 1.5, represents the gamma function;

[0091] Step 5-5, the whale fall phase of the white whale optimization algorithm occurs when a few white whales will fall to the bottom of the sea while hunting and cruising. The updated position is established using the position of the white whale and the step length of the whale's descent, which can be expressed as:

[0092] ,

[0093] in, , and Represents a random number between (0, 1), is the step size of the white whale, and the calculation formula is:

[0094] ,

[0095] in, and Represent the upper and lower limits respectively, exp is the natural exponential function, represents the step factor, where The probability of death of the beluga whale is expressed as follows:

[0096] ;

[0097] Steps 5-6: In order to improve the problem that the White Whale optimization algorithm is prone to falling into local optimality, an elite reverse learning strategy is introduced to expand the search range and increase population diversity. The expression is:

[0098] ,

[0099] ,

[0100] in, Indicates the position of the business entity in the dth dimension, K is a dynamic coefficient in the range of (0, 1), Represents the position vector of the current elite individual, represents the inverse solution of the position vector of the elite individual, and are the minimum and maximum values ​​of the dynamic boundary, respectively;

[0101] Steps 5-7, in the White Whale optimization algorithm, the balance factor It is the key factor between the exploration and development phases, and has an important impact on the global and local search of the balance algorithm. In order to further improve the search capability of the White Whale optimization algorithm, the balance factor Make adjustments;

[0102] In steps 5-8, in order to further enhance the development phase performance of the beluga whale optimization, the spiral foraging strategy from the manta ray foraging optimization is introduced.

[0103] In steps 5-7, the balance factor Make the following adjustments:

[0104] ,

[0105] Adjusted balance factor The linear change pattern of the original balance factor is broken; in the early stage of algorithm iteration, Small changes, Its decline is relatively gentle, which is conducive to the algorithm to carry out global search; at the end of the iteration, The value increases sharply, The reduction rate has increased significantly, thereby strengthening the local search capability.

[0106] In steps 5-8, the improved development phase position update formula is:

[0107] ,

[0108] ,

[0109] in, and represents a random number between (0,1), is the inertia weight, is the random jump intensity coefficient that controls the Levy flight intensity, is the Levy flight function, which is calculated in the same way as the original Beluga optimization algorithm.

[0110] The present invention also provides an electronic device, comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the method.

[0111] The present invention also provides a storage medium storing a computer program or instruction, which executes the steps of the method when the computer program or instruction is run on a computer.

[0112] The present invention has the following beneficial effects: The proposed ultra-short-term wind speed forecasting method (ILPMTSD-AISM-IBWO-BiLSTM / GRU) based on wind speed fluctuation pattern clustering uses the AISM model to effectively extract wind speed fluctuation characteristics, identify different wind speed fluctuation patterns, and further segment wind speed sequences using K-means. Furthermore, the improved Beluga optimization algorithm is used to optimize the BiLSTM and GRU neural networks to improve the reliability of prediction results. Both the AISM model and the improved Beluga optimization algorithm effectively improve the accuracy of wind speed predictions, significantly enhancing the reliability of wind speed predictions directly input into wind power prediction models and increasing the accuracy of wind speed predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 It is the overall flow chart of the present invention.

[0114] Figure 2 It is a principle diagram of the information leakage prevention mechanism of the present invention.

[0115] Figure 3 It is a wind speed segment classification result diagram during the implementation of the present invention.

[0116] Figure 4 It is an architecture diagram of the GRU network of the present invention.

[0117] Figure 5 This is an architecture diagram of the BiLSTM network of the present invention.

[0118] Figure 6 This is a diagram of the prediction effect of the present invention on the Yandun dataset.

[0119] Figure 7 This is a diagram of the prediction effect of the present invention on the Jiangsu Xiayang dataset.

[0120] Figure 8 This is a diagram of the prediction effect of the present invention on the Jiangxi Baijiashe dataset.

[0121] Figure 9 This is the result of the BiLSTM module ablation experiment of the present invention.

[0122] Figure 10This is the result of the IBWO-BiLSTM module ablation experiment of the present invention.

[0123] Figure 11 This is the result of the AISM-BiLSTM / GRU module ablation experiment of the present invention.

[0124] Figure 12 This is the result of the ablation experiment of the ILPMTSD-BiLSTM module of the present invention.

[0125] Figure 13 This is the result of the ablation experiment of the ILPMTSD-AISM-IBWO-BiLSTM / GRU module of the present invention. DETAILED DESCRIPTION

[0126] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more apparent.

[0127] This paper proposes an ultra-short-term wind speed prediction method based on wind speed fluctuation pattern clustering (ILPMTSD-AISM-IBWO-BiLSTM / GRU), such as Figure 1 and Figure 2 As shown, the following steps are included:

[0128] Step 1: Collect historical measured wind speed data recorded by the wind farm data acquisition and monitoring control system;

[0129] Step 2: A two-stage decomposition based on information leakage prevention mechanism (ILPM) is used. The original sequence is decomposed into low-frequency sequence and high-frequency sequence using discrete wavelet transform, and the high-frequency sequence is further decomposed using ICEEMDAN. The information leakage prevention mechanism is used to prevent information leakage during the decomposition process. The specific principle of the information leakage prevention mechanism ILPM is as follows: Figure 2 As shown in the figure, a window with a length of 10% of the original sequence is slid across the original sequence, and only the data within the window is decomposed. Whenever new data is added, the oldest data in the window is removed. This way, the number of subsequences generated after decomposition remains unchanged, and no data is leaked.

[0130] Step 3: Use wind speed fluctuation intensity and wind speed fluctuation rate features to quantify the fluctuation of wind speed data;

[0131] Step 3-1, calculate wind speed fluctuation rate :

[0132] ,

[0133] Where N represents the number of wind speed data. Represents the e-th wind speed data;

[0134] Step 3-2, calculate the wind speed fluctuation intensity :

[0135] ,

[0136] in Indicates the mean wind speed.

[0137] Step 4: Based on the wind speed characteristic statistical method, sliding window method, and K-means cluster analysis algorithm, an adaptive wind speed interval division model AISM is constructed; the adaptive wind speed interval division model AISM processes the low-frequency components (approximate components) after discrete wavelet transform processing, extracts wind speed fluctuation characteristics, performs wind speed classification, and marks wind speed segments. The specific steps are:

[0138] Step 4-1: Using a window size of 4 hours (16 data points) and a sliding step size of 1 step (15 minutes) as the sampling standard, the AISM model is used to segment the historical measured wind speed data. The specific process is as follows:

[0139] Step 4-1-1, construct a set of wind speed data with a length of M :

[0140] ,

[0141] Step 4-3-2, define each sliding window:

[0142] ,

[0143] in, Represents the mth sliding window, each window length is 16;

[0144] Step 4-2, each sliding window As a group of wind speed sequence segments, the wind speed fluctuation rate of each wind speed sequence segment is extracted and recorded as and wind speed fluctuation intensity is recorded as , merge the processed features to form a feature matrix :

[0145] ,

[0146] in, and They represent the wind speed fluctuation rate and wind speed fluctuation intensity of the mth wind speed sequence segment respectively;

[0147] Step 4-3, using K-means cluster analysis algorithm, according to the two fluctuation characteristics of each wind speed data segment and Perform clustering processing to obtain the classification criteria for fluctuating wind speed intervals and stable wind speed intervals, mark the wind speed sequence segments, and obtain the wind speed fluctuation marking matrix :

[0148] ,

[0149] in, Indicates the Wind speed sequence segments, Indicates the wind speed fluctuation mark of the corresponding wind speed sequence segment. According to the wind speed fluctuation mark, the original wind speed sequence segment is divided into two sets:

[0150] ,

[0151] ,

[0152] in, and are respectively the set of stable wind speed data segments and the set of unstable wind speed data segments, and They represent the subscript set of stable sequences and the subscript set of unstable sequences respectively.

[0153] The K-means clustering wind speed interval division results are as follows Figure 3 As shown in the figure, the horizontal axis is the wind speed fluctuation rate, the vertical axis is the wind speed fluctuation intensity, the wind speed interval below the decision boundary is classified as a stable wind speed interval, and the wind speed interval on the decision boundary is classified as a fluctuating wind speed interval.

[0154] In step 5, two highly accurate prediction basic models, namely the BISTM and GRU network models, were constructed by improving the Beluga optimization algorithm. Appropriate models were used for prediction of wind speed segments with different degrees of fluctuation.

[0155] The network structure of BiLSTM is as follows Figure 4 As shown, Represents the traditional operation process of the LSTM network. , , represent , , Input at the moment, , , represent , , The forward hidden state at any moment, , , represent , , Output at any moment, , , represent , , The backward hidden state at the moment;

[0156] The network structure of GRU is as follows Figure 5 As shown, and Represent the output at time P-1 and P respectively, represents the input at time P, is the sigmoid activation function, tanh is the hyperbolic tangent function, is the hidden state at time P, and To update the gate and reset the gate.

[0157] Step 6: For the high-frequency component (detail component) processed by ICEEMDAN, use BiLSTM to predict it and merge it with the prediction result of the low-frequency component (approximate component). Output the test result and calculate the evaluation indicators MAE, RMSE, SMAPE and R 2 , verify the feasibility of the model.

[0158] In this example, the data comes from the wind speed measurement data of a wind farm in Yandun for three months from February to April 2024 at 15-minute intervals. In order to test the robustness of the model proposed by this method and the effectiveness of each module, the following experiments are conducted. The specific example verification is as follows:

[0159] First, to verify the versatility of the proposed ILPMTSD-AISM-IBWO-BiLSTM / GRU ultra-short-term wind speed forecasting model in practical applications across diverse environments, this experiment selected two wind farms located at different locations at the same time: Baijiashe and Xiayang. This experiment evaluated its effectiveness under varying environmental conditions. The Baijiashe wind farm is located in a mountainous area with a complex geographical location, resulting in large and irregular fluctuations in wind speed data. The Xiayang wind speed data exhibited a generally mild trend. The model's predictions were then compared with the ground truth and a traditional machine learning model.

[0160] like Figure 6 、 Figure 7 、 Figure 8As shown, the method proposed in this paper can accurately predict wind speed in different regions and terrains. Due to the randomness and volatility of wind speed, traditional models cannot capture sudden changes in wind speed, resulting in low prediction accuracy. Compared with traditional wind speed prediction methods, the method proposed in this paper has stronger fitting capabilities. Especially when the wind speed changes suddenly, it can still capture the wind speed change, ensuring the real-time and dynamic nature of wind speed prediction.

[0161] Second, verify the effectiveness of each module in the ILPMTSD-AISM-IBWO-BiLSTM / GRU ultra-short-term wind speed prediction model proposed in this invention;

[0162] In order to verify the effectiveness of each module of the proposed model, the prediction errors of the bidirectional long short-term memory network model (BiLSTM), the improved Beluga optimization algorithm-bidirectional long short-term memory network model (IBWO-BiLSTM), the adaptive interval partitioning model-bidirectional long short-term memory network model / gated recurrent unit (AISM-BiLSTM / GRU), the two-stage decomposition based on information leakage prevention mechanism-bidirectional long short-term memory network model (ILPMTSD-BiLSTM), and the two-stage decomposition based on information leakage prevention mechanism-improved Beluga optimization algorithm-adaptive interval partitioning model-bidirectional long short-term memory network model / gated recurrent unit (ILPMTSD-AISM-IBWO-BiLSTM / GRU) were statistically compared. The mean absolute error (MAE), root mean square error (RMSE), symmetric mean absolute percentage error (SMAPE) and coefficient of determination (R 2 ) are used to evaluate the prediction accuracy of different models for each set of data.

[0163] like Figure 9 、 Figure 10 、 Figure 11 、 Figure 12 and Figure 13 The following figure shows the prediction results and prediction error histograms of each module of ILPMTSD-AISM-IBWO-BiLSTM / GRU. The prediction error histogram represents the absolute value of the difference between the actual value and the predicted value, where Big Error represents a large error greater than 0.5 and Small Error represents a small error less than 0.5. Figure 9 、 Figure 10 、 Figure 11 、 Figure 12 and Figure 13As can be seen, the wind speed prediction error of the ILPMTSD-AISM-IBWO-BiLSTM / GRU model is much lower than that of the other models; most errors are kept between 0 and 1, and the number of large errors is significantly reduced. This analysis shows that the ILPMTSD-AISM-IBWO-BiLSTM / GRU model effectively improves wind speed prediction accuracy and significantly reduces wind speed prediction error, thereby increasing the reliability of wind speed prediction.

[0164] Table 1 shows the mean absolute error (MAE), root mean square error (RMSE), symmetric mean absolute percentage error (SMAPE), and determination coefficient (R) of the BiLSTM model, IBWO-BiLSTM, AISM-BiLSTM / GRU, ILPMTSD-BiLSTM, and ILPMTSD-AISM-IBWO-BiLSTM / GRU models under experimental data. 2 ) results comparison.

[0165] Table 1

[0166] MAE (m / s) RMSE (m / s) SMAPE (%) <![CDATA[R 2 (m / s)]]> BiLSTM 0.804 1.045 20.365 0.864 IBWO-BiLSTM 0.658 0.875 16.767 0.904 AISM-BiLSTM / GRU 0.633 0.846 16.436 0.911 ILPMTSD-BiLSTM 0.343 0.445 8.561 0.975 ILPMTSD-AISM-IBWO-BiLSTM / GRU 0.246 0.324 6.467 0.986

[0167] As can be seen from Table 1, the ILPMTSD-AISM-IBWO-BiLSTM / GRU model proposed in this invention has the following advantages: MAE, RMSE, SMAPE and R 2 The values ​​of the four error evaluation indicators are all minimized. The addition of the AISM and ILPMTSD modules significantly improves the accuracy compared to the original model, and the method proposed in this invention has better prediction performance.

[0168] An embodiment of the present invention further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that each module in the aforementioned system corresponds to the specific steps of the method provided in the embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0169] The present invention also provides a computer-readable storage medium storing a computer program. It should be noted that each module in the above system corresponds to the specific steps of the method provided in the present invention, and has corresponding functional modules and beneficial effects for executing the method. For technical details not fully described in this embodiment, please refer to the method provided in the present invention.

[0170] This paper uses an ultra-short-term wind speed forecasting method based on clustering of wind speed fluctuation patterns to extract wind speed fluctuation characteristics and effectively mine potential characteristic information. A sliding window method is then used to extract wind speed fluctuation indicators for each wind speed interval, identifying different wind speed fluctuation patterns. The optimal forecasting model is dynamically selected for each wind speed fluctuation interval, further improving the stability, accuracy, and computational efficiency of the model's predictions.

[0171] The present invention provides an ultra-short-term wind speed forecasting method based on clustering of wind speed fluctuation patterns. There are numerous methods and approaches for implementing this technical solution. The above is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Any components not specified in this embodiment may be implemented using existing technologies.

Claims

1. A method for ultra-short-term wind speed prediction based on clustering of wind speed fluctuation patterns, characterized in that: The following steps are involved: Step 1: Collect historical measured wind speed data recorded by the wind farm data acquisition and monitoring control system; Step 2: A two-stage decomposition based on an information leakage protection mechanism is adopted. The original sequence is decomposed into a low-frequency sequence and a high-frequency sequence using discrete wavelet transform. The high-frequency sequence is further decomposed using an improved fully adaptive noise ensemble empirical mode decomposition, while an information leakage prevention mechanism is also used. Step 3: Calculate the wind speed fluctuation intensity and wind speed fluctuation rate characteristics to quantify the fluctuation of wind speed data; Step 4: Based on the wind speed characteristic statistical method, sliding window method, and K-means cluster analysis algorithm, an adaptive wind speed interval partitioning model AISM is constructed; The adaptive wind speed interval segmentation model AISM processes the low-frequency components after discrete wavelet transform, extracts wind speed fluctuation characteristics, performs wind speed classification, and marks wind speed segments; Step 5: Use the improved Beluga optimization algorithm to build two basic prediction models, namely a bidirectional long short-term memory network and a gated recurrent unit. For wind speed segments with different degrees of fluctuation, the corresponding basic prediction models are used for prediction; Step 6: For the high-frequency components processed by the improved fully adaptive noise set empirical mode decomposition, use the bidirectional long short-term memory network BiLSTM to predict them, merge them with the prediction results of the low-frequency components, output the test results, and calculate the evaluation indicators to verify the feasibility of the model.

2. The method according to claim 1, characterized in that Step 2 includes: Step 2-1, in the first stage of the two-stage decomposition, the original wind speed data is decomposed into low-frequency components and high-frequency components using discrete wavelet transform. The formula is: ， ， in, and Represent low-frequency components and high-frequency components respectively, k represents the index variable of summation, is the original wind speed data, represents a low-pass filter, represents a high-pass filter, Indicates the decomposition level; Step 2-2: In the second stage, the improved fully adaptive noise ensemble empirical mode decomposition is used to further decompose the high-frequency components generated in the first stage and convert them into subsequences. The specific steps include: Step 2-2-1, add white noise to the signal to be decomposed x: ， in, represents the white noise added to group i, is the first empirical mode component EMD of white noise, represents the signal-to-noise ratio of the first decomposition, represents the signal after adding white noise; Step 2-2-2, the first residual The calculation formula is: ， in represents the local average envelope; Step 2-2-3, determine the initial eigenmode function of the improved fully adaptive noise ensemble empirical mode decomposition by decomposing N signals as follows: : ； Step 2-2-4, determine the second residual and the second modality : ， ； in, represents the signal-to-noise ratio of the second decomposition, is the second empirical mode component of white noise; Step 2-2-5, kth residual and the kth mode Determined as: ， ； in, represents the signal-to-noise ratio of the k-th decomposition, represents the k-1th residual, represents the kth empirical mode component of white noise; Step 2-2-6, continue to perform step 2-2-5 for the next k stages until the residual is no longer decomposed; In step 2-3, based on the sliding window method, a window with a fixed window width is slid on the original sequence. Every time new data is added, the old data in the window will be removed.

3. The method according to claim 2, characterized in that Step 3 includes: Step 3-1, calculate wind speed fluctuation rate : ， Where N represents the number of wind speed data. Represents the e-th wind speed data; Step 3-2, calculate the wind speed fluctuation intensity : ， in Indicates the mean wind speed.

4. The method according to claim 3, characterized in that Step 4 includes: In step 4-1, the low-frequency components are segmented using the adaptive wind speed interval partitioning model with a window size of 4 hours and a sliding step size of 1 step as the sampling standard. The specific process includes: Step 4-1-1, construct a set of wind speed data with a length of M : ， Step 4-1-2, define each sliding window: ， in, Represents the mth sliding window, each window length is 16; Step 4-2, each sliding window As a group of wind speed sequence segments, the wind speed fluctuation rate of each wind speed sequence segment is extracted and recorded as and wind speed fluctuation intensity is recorded as , merge the processed features to form a feature matrix : ， in, and They represent the wind speed fluctuation rate and wind speed fluctuation intensity of the mth wind speed sequence segment respectively; Step 4-3: Combine the two fluctuation characteristics of each wind speed data segment and As input, a K-means cluster analysis algorithm is used to obtain the classification criteria for fluctuating wind speed intervals and stable wind speed intervals, and to mark the wind speed sequence segments. The K-means cluster analysis algorithm includes the following steps: Step 4-3-1, for an unlabeled data set , select Cluster centroids ,form clusters; Indicates the unlabeled data, Indicates the cluster centroids, is the total number of unlabeled data, is the total number of cluster centroids; Step 4-3-2, for each unlabeled data , we need to calculate the centroid of each cluster distance, Belong to The nearest cluster centroid to itself Cluster : ， in, Indicates that the The function with the smallest corresponding function value of clusters, For each data point Assign to the nearest cluster centroid The corresponding cluster middle, Indicates the data points and The square of the Euclidean distance between clusters; Step 4-3-3, update the cluster centroid of each cluster: ， in, is an indicator function, when When , the indicator function value is 1, otherwise it is 0; Indicates that all clusters The sum of the coordinates of the data points, Indicates belonging to a cluster The total number of data points; Step 4-3-4, repeat steps 4-3-2 to 4-3-3 until the center of mass position no longer changes or the preset number of iterations is reached; In step 4-3-5, the goal of K-means is to minimize the sum of squared errors within the cluster, which is expressed as: ， Where J represents each data point to the corresponding cluster centroid Sum of squares of distances; Step 4-4, after the K-means cluster analysis algorithm, the wind speed fluctuation label matrix is ​​obtained : ， in, represents the mth wind speed sequence segment, Indicates the wind speed fluctuation mark of the corresponding wind speed sequence segment. According to the wind speed fluctuation mark, the original wind speed sequence segment is divided into two sets: ， ， in, and are respectively the set of stable wind speed data segments and the set of unstable wind speed data segments, and They represent the subscript set of stable sequences and the subscript set of unstable sequences respectively.

5. The method according to claim 4, characterized in that In step 5, in the unstable wind speed data segment, the bidirectional long short-term memory network BiLSTM is used as the benchmark model for fitting; in the stable wind speed data segment, the gated recurrent unit GRU is used as the benchmark model for fitting.

6. The method according to claim 5, characterized in that In step 5, the improved Beluga optimization algorithm includes the following steps: Step 5-1, the White Whale Optimization Algorithm includes three stages: exploration, development, and whale fall; the White Whale Optimization Algorithm is based on the balance factor To achieve the transition from exploration to development of the population, when the balance factor When the balance factor When , it is the development stage, and the calculation formula is: ， Where T is the current iteration number, is the total number of iterations, parameter is a random number in the range of (0, 1); Step 5-2: Initialize the beluga population. In the beluga optimization algorithm, a single beluga is considered a search example. The initial population contains n beluga whales, which will be updated during the optimization process and is expressed as follows: ， Among them, X represents the position matrix of the beluga whale, represents the location information of a specific beluga whale on the corresponding dimension, n is the population size of the beluga whale, and d represents the dimension of the variable; For all beluga whales, the corresponding fitness values ​​are stored as: ， in, is the fitness function, is a vector representing the fitness of all beluga whales; Step 5-3, the exploration phase of beluga optimization simulates two beluga whales swimming closely in a synchronous or mirrored manner. The position update formula of the beluga whale is: ， Among them, even and odd represent the case when the variable dimension index j is an even number and the case when j is an odd number, respectively. Indicates that in the next iteration, the i-th individual Position on the dimension; p is a random integer. Suppose the problem dimension is D, and p is a random integer in the range [1, D]. Indicates the position of the i-th individual on the random dimension p in the current iteration; set the population size to N, and r to be a random integer in the range [1, N]; Indicates the position of the rth random individual on the random dimension p in the current iteration; , All are random numbers between (0,1); Step 5-4, the development phase of Beluga optimization, is inspired by the hunting behavior of Beluga whales and is expressed as the following equation: ， in, is the current position of the i-th beluga whale, is the current position of a random beluga whale, is the new position of the i-th beluga whale, Best location in Moby Dick; and is a random number between (0,1), is the random jump intensity; is a Levy flight function, calculated using the following formula: ， ， in, is an intermediate calculation parameter, and is a normally distributed random number, is a constant, represents the gamma function; Step 5-5, the whale fall phase of the white whale optimization algorithm occurs when hunting and cruising in the sea. The updated position is established using the position of the white whale and the step size of the whale's descent, which is expressed as: ， in, , and Represents a random number between (0, 1), is the step size of the white whale, and the calculation formula is: ， in, and Represent the upper and lower limits respectively, exp is the natural exponential function, represents the step factor, where The probability of death of the beluga whale is expressed as follows: ； Steps 5-6: In order to improve the problem that the White Whale optimization algorithm is prone to falling into local optimality, an elite reverse learning strategy is introduced to expand the search range and increase population diversity. The expression is: ， ， in, Indicates the position of the business entity on the dth dimension, is a dynamic coefficient in the range of (0, 1), Represents the position vector of the current elite individual, represents the inverse solution of the position vector of the elite individual, and are the minimum and maximum values ​​of the dynamic boundary, respectively; Steps 5-7, in the White Whale optimization algorithm, the balance factor Make adjustments; Steps 5-8 introduce the spiral foraging strategy used in manta ray foraging optimization.

7. The method according to claim 6, characterized in that In steps 5-7, the balance factor Make the following adjustments: 。 8. The method according to claim 7, characterized in that In steps 5-8, the improved development phase position update formula is: ， ， in, and represents a random number between (0,1), is the inertia weight, is the random jump intensity coefficient that controls the Levy flight intensity, is the Levy flight function.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 8.

10. A storage medium, characterized in that: A computer program or instruction is stored, and when the computer program or instruction is run on a computer, the steps of the method according to any one of claims 1 to 8 are executed.

Images