Wind speed prediction method and system based on entropy clustering

Abstract

The invention relates to the technical field of wind speed prediction, in particular to a wind speed prediction method and system based on entropy clustering, and the method comprises the following steps: carrying out the multi-scale decomposition of a wind speed sequence through a permutation entropy algorithm, generating subsequences, carrying out the sliding window probability distribution calculation of a plurality of subsequences, setting a dynamic classification threshold value based on a normalized entropy value, and carrying out the calculation of the sliding window probability distribution. And dividing the sub-sequence set into an entropy classification result set. According to the method, a classification threshold value is dynamically set through sliding window probability distribution, the sensitivity of fixed entropy division to data complexity change is overcome, nonlinear mapping between an original sequence and entropy components is established through GRU training, the feature extraction depth and low component error transmission are enhanced, high, medium and low entropy components are modeled through Bi-RNN, LSTM and SVR, feature extraction with different complexities is matched, and the accuracy of feature extraction is improved. Chaotic mapping optimization fusion weight, multi-target optimization balance component complementarity and dynamic error reference are combined with feedback to adjust a confidence interval, error distribution adaptive tracking is realized, and prediction reliability is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of wind speed prediction, and in particular to a wind speed prediction method and system based on entropy clustering. Background Art

[0002] The technical field of wind speed prediction includes means and methods for monitoring, evaluating, and predicting wind energy resources. Its core content is to use various historical wind speed data to predict the wind speed in a future time period through time series analysis, mathematical statistics modeling, machine learning modeling, etc. This technical field is widely applied in engineering scenarios such as wind farm site selection, wind power output scheduling, and system stability analysis. The wind speed prediction technology usually includes links such as data collection, feature construction, model establishment, and error correction. The key lies in improving the accuracy and adaptability of wind speed prediction to meet the prediction requirements of different spatio-temporal scales.

[0003] Among them, the wind speed prediction method based on entropy clustering refers to a method of dividing wind speed data into multiple sub-class data groups with different statistical characteristics according to the information entropy theory, and then separately modeling and predicting each sub-class data. Aiming at the characteristics of wind speed data being diverse and non-stationary in different time periods, this method first uses the entropy value as the division basis to cluster the overall wind speed data into multiple subsets that are relatively consistent internally but significantly different from each other, and then uses specific prediction models such as support vector regression and random forest regression to model each subset, and combines the output results of each model to form the final prediction value. This method mainly processes according to the information entropy division mechanism and the independent modeling process based on sub-class models.

[0004] There are three defects in the prior art. The fixed information entropy threshold leads to inaccurate sub-class division when the data distribution changes. For example, the wind speed pattern switch caused by the day-night temperature difference will cause abnormal data merging in traditional clustering. There is a problem of structural mismatch when homogeneous models process heterogeneous sub-class data. Support vector regression is prone to losing detailed features when processing high-frequency components, and random forests produce redundant calculations for stationary sequences. The static error benchmark cannot capture real-time error offsets. When the prediction environment suddenly changes, the historical error statistics deviate from the actual distribution, resulting in the failure of the confidence evaluation of interval prediction. These defects make the existing methods face the risks of prediction lag and error accumulation under complex meteorological conditions, affecting the reliability of wind power grid connection scheduling decisions. Summary of the Invention

[0005] The purpose of the present invention is to solve the drawbacks existing in the prior art, and to propose a wind speed prediction method and system based on entropy clustering.

[0006] In order to achieve the above purpose, the present invention adopts the following technical scheme: A wind speed prediction method based on entropy clustering, including the following steps: S1: Generate a subsequence set by performing multi-scale decomposition on the training set wind speed sequence using the permutation entropy algorithm, calculate the probability distribution of the multi-subsequences through a sliding window, set a dynamic classification threshold based on the normalized entropy value, and divide the subsequence set into an entropy classification result set; S2: Input the original wind speed sequence of the training set and the entropy classification result set into the GRU network for training the gated state parameters, generate a component mapping model, and input the original wind speed sequence of the test set into the component mapping model to output the entropy component set of the test set; S3: Input the high-entropy components in the entropy component set of the test set into a bidirectional gated recurrent unit to generate a high-entropy prediction result, input the medium-entropy components into a long short-term memory model to generate a medium-entropy prediction result, input the low-entropy components into a support vector regression model to generate a low-entropy prediction result, allocate initial weights based on the entropy value ratio, use the chaotic mapping algorithm to iteratively optimize the weight parameters, and perform weighted superposition on the three types of prediction results to generate a wind speed point prediction result.

[0007] As a further solution of the present invention, the entropy classification result set is specifically a high-entropy subsequence category, a medium-entropy subsequence category, and a low-entropy subsequence category. The entropy component set of the test set includes high-entropy components, medium-entropy components, and low-entropy components. The wind speed point prediction result specifically refers to the superposition value of the high-entropy prediction result, the medium-entropy prediction result, the low-entropy prediction result, and the optimized weight parameters.

[0008] As a further solution of the present invention, the normalized entropy value is calculated through the linear mapping formula where, represents the normalized value of the permutation entropy of the current window, represents the permutation entropy value of the i-th sliding window, represents the maximum value of the permutation entropy values in all sliding windows, represents the minimum value of the permutation entropy values in all sliding windows. The dynamic classification threshold is the coordinate value corresponding to the extreme point of the second derivative of the density curve of the mapped entropy value distribution. Among them, this extreme point represents the position where the density changes most significantly and is used as the classification boundary for category division; The dynamic classification threshold is the coordinate value corresponding to the extreme point of the second derivative of the density curve of the mapped entropy value distribution; The GRU network includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is equal to the time step of the wind speed sequence. The hidden layer has 2 layers and each layer contains 128 neurons. The number of neurons in the output layer is 32; The chaotic mapping algorithm uses the Logistic mapping function to generate a random sequence, where, is the chaotic sequence value of the n-th iteration, is the chaotic sequence value of the (n + 1)-th iteration. The fitness function adopts the mean square error form, that is , where represents the mean square error of the prediction result, is the number of samples, is the j-th true wind speed value, is the j-th predicted wind speed value, the perturbation amplitude of the weight updated by the gradient descent method. When optimizing the weight, the initial weight is allocated according to the entropy value ratio, and finally, through multiple iterations of optimization, the converged optimal weight parameters are obtained.

[0009] As a further solution of the present invention, the steps for obtaining the entropy classification result set are specifically as follows: S101: Obtain the wind speed sequence of the training set, call the permutation entropy algorithm to set the embedding dimension parameter and the delay parameter, construct a multi-dimensional phase space according to the embedding dimension parameter, perform time-delay slicing on the sequence through the delay parameter, reconstruct the phase space trajectories of multiple time scales, and decompose the original sequence based on the similarity difference between the trajectories to generate multi-scale decomposition subsequences; S102: Based on the multi-scale decomposition subsequences, set the covering length parameter and the moving step parameter of the sliding window, intercept the local data segments of the subsequences successively according to the step parameter, count the frequencies of the different wind speed values appearing in each window, divide the frequency by the window length parameter, calculate the probability distribution of the wind speed values corresponding to multiple windows, and generate a probability distribution matrix; The window length parameter is uniformly in seconds, and the moving step parameter is 1 second; S103: Extract the permutation entropy value of each window from the probability distribution matrix, locate the maximum and minimum values of the entropy value sequence, map the entropy value to the 0-1 interval, analyze the distribution density curve of the mapped entropy value, calculate the second derivative of the curvature with respect to the entropy value change, locate the entropy value coordinates corresponding to the derivative extreme points, and set the coordinate value as the classification threshold to divide the high-entropy and low-entropy categories, generating a dynamic classification threshold result; The second derivative is calculated by the numerical difference method, and the extreme point determination criterion is that the derivative sign changes from positive to negative or from negative to positive.

[0010] As a further solution of the present invention, the steps for obtaining the entropy component set of the test set are specifically as follows: S201: Obtain the original wind speed sequence of the training set and the entropy classification result set, construct the input layer of the GRU network based on the time step characteristics of the wind speed sequence, and at the same time normalize and initialize the weights of the input layer according to the multi-category probability distribution in the entropy classification result set. Use the gradient descent method to calculate the weight errors of the forget gate and the input gate step by step in time, and update the gate control unit parameters through chain derivation iteration to generate a set of gate control state parameters; The normalization initialization adopts the Xavier method, and the learning rate of the gradient descent method is set to ; S202: Call the set of gating state parameters, divide the training set wind speed sequence into subsequence segments, and at the same time establish a mapping index between the subsequences and the hidden layer states based on the entropy classification labels. Calculate the error gradient between the hidden layer states and the labels through backpropagation, update the output gate weights and the fully connected layer parameters, and generate the component mapping model parameters; S203: Based on the component mapping model parameters, divide the test set wind speed sequence into subsequence segments, input the GRU network to calculate the hidden layer state vectors, extract the neuron activation values corresponding to the entropy classification labels, splice the activation value sequences according to the time steps, and generate the test set entropy component set.

[0011] As a further solution of the present invention, the steps for obtaining the wind speed point prediction result are specifically as follows: S301: Call the test set entropy component set, input the high-entropy components into the bidirectional gated recurrent unit based on the entropy classification labels, expand the forward and backward hidden state transmissions of the time steps, calculate the output values of the multi-time step gated units, input the medium-entropy components into the long short-term memory model, screen the time-dependent features according to the forgetting gate and input gate states, input the low-entropy components into the support vector regression model, map the feature space using the radial basis kernel function and solve the regression hyperplane, and generate the component prediction result set; The bandwidth parameter of the radial basis kernel function , and the sequence minimal optimization algorithm is used to solve the hyperplane; S302: Based on the initial entropy value ratio of the high-entropy, medium-entropy, and low-entropy components in the component prediction result set, calculate the proportion of the high-entropy component prediction result in the total entropy value, and use the proportion as the initial weight distribution coefficient. At the same time, call the Logistic mapping function in the chaos mapping algorithm to generate a random sequence, use the mean square error of the prediction results as the fitness function, and iteratively update the perturbation amplitude of the weight parameters to generate the optimized weight parameters; The update step size of the perturbation amplitude is ; S303: Call the component prediction result set and the optimized weight parameters, perform a scalar product of the high-entropy prediction result and the high-entropy weight coefficient in the optimized weights, synchronously calculate the products of the medium-entropy and low-entropy prediction results and the corresponding weights, and align and superimpose the three types of weighted results according to the time steps to generate the wind speed point prediction result.

[0012] As a further solution of the present invention, the method further includes: S4: Based on the error standard deviations of multiple entropy classification intervals in the training set, establish a dynamic error benchmark set, input the wind speed point prediction result and the real-time error of the test set into the error distribution model, adjust the confidence interval width according to the relative relationship between the error and the benchmark value, and output the interval prediction result set.

[0013] As a further solution of the present invention, the interval prediction result set includes a dynamic error benchmark, real-time error distribution data, a confidence interval width adjustment parameter, and a probability interval range; The dynamic error benchmark set is a sequence of error standard deviations arranged in ascending order of entropy classification labels, and the confidence interval width adjustment formula is , where is the width of the current confidence interval, is the proportionality coefficient, is the wind speed prediction error value at the t-th moment, is the error benchmark value of the k-th entropy classification interval, The difference between reflects the prediction uncertainty. Based on this difference, the confidence interval width is dynamically adjusted, and the probability interval range is jointly determined by the error distribution and the width parameter.

[0014] As a further solution of the present invention, the specific steps for obtaining the interval prediction result set are as follows: S401: Obtain the prediction error sequences corresponding to multiple entropy classification intervals in the training set, divide the error data according to the intervals, calculate the standard deviation of the error values within each entropy classification interval, arrange the standard deviations in ascending order of entropy classification labels, and generate a dynamic error benchmark set; S402: Call the dynamic error benchmark set, align the wind speed point prediction result with the real-time error of the test set according to the time step, and at the same time, based on the normal distribution setting of the error distribution model, calculate the absolute value of the difference between the error value at each time step and the corresponding entropy classification benchmark value, and use the maximum likelihood estimation method to update the mean and variance parameters of the distribution model to generate error distribution parameters; The initial mean of the normal distribution is , and the variance is ; S403: Based on the error distribution parameters, linearly expand or shrink the confidence interval width according to the proportion of the error value at the current time step exceeding the benchmark value, use the mean of the distribution model as the center of the interval, and calculate the upper and lower bounds in combination with the adjusted width to generate an interval prediction result.

[0015] An entropy clustering-based wind speed prediction system, which is used to execute the above-mentioned entropy clustering-based wind speed prediction method. The system includes: A decomposition clustering module, which is used to decompose the wind speed sequence in the training set through the permutation entropy algorithm to generate a subsequence set, divide the subsequences by using a sliding window, calculate the probability distribution of the wind speed values within the window, generate a dynamic classification threshold based on the normalized entropy value function, divide the subsequences into a high-entropy component set, a medium-entropy component set, and a low-entropy component set, integrate them into an entropy classification result set, and transfer the entropy classification result set to the component mapping module; The medium entropy component set is a subsequence set with the normalized entropy value in the middle 30% interval of the dynamic classification threshold; The component mapping module is used to obtain the original wind speed sequence of the training set and the entropy classification result set, train and generate a component mapping model through a GRU network, input the wind speed sequence of the test set into the model to output the entropy component set of the test set, and transfer the entropy component set of the test set to the weight prediction module; The weight prediction module is used to call the entropy component set of the test set, input the high entropy component set into a bidirectional gated recurrent unit to generate a high entropy prediction value, input the medium entropy component set into a long short-term memory model to generate a medium entropy prediction value, input the low entropy component set into a support vector regression model to generate a low entropy prediction value, allocate initial weights based on the entropy value ratio, optimize the weight parameters by using a chaotic mapping algorithm, superimpose to generate a wind speed point prediction result, and transfer the wind speed point prediction result to the interval correction module; The interval correction module is used to establish a dynamic error benchmark set according to the standard deviation of the prediction error of the entropy classification interval of the training set, call the wind speed point prediction result and the real-time error of the test set, calculate the relative difference between the real-time error and the benchmark value through an error distribution model, adjust the width of the confidence interval, and output an interval prediction result set.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In the present invention, by combining the calculation of the sliding window probability distribution to construct a dynamic classification threshold, the sensitivity problem of the fixed entropy value division to the change of data complexity is overcome, and the subsequence division is more in line with the actual fluctuation mode of the wind speed. The GRU network gate parameter training establishes a non-linear mapping relationship between the original sequence and the entropy components, enhances the depth of feature extraction, and reduces the transmission of component classification errors. The high, medium, and low entropy components are respectively modeled by using a bidirectional recurrent unit, a long short-term memory, and a support vector regression to match the time series feature extraction requirements of data with different complexities, and avoid the structural limitations of a single model. The chaotic mapping algorithm optimizes the weight parameters, balances the complementarity of the prediction results of each component through multi-objective optimization, and improves the stability of the fusion prediction. The dynamic error benchmark set combines real-time error feedback to adjust the confidence interval, realizes the adaptive tracking of the prediction error distribution, and enhances the time effectiveness and reliability of the interval prediction. Description of the Drawings

[0017] Figure 1 It is a schematic diagram of the working process of the present invention; Figure 2 It is a flow chart of the steps for obtaining the entropy classification result set of the present invention; Figure 3 It is a flow chart of the steps for obtaining the entropy component set of the test set of the present invention; Figure 4 It is a flow chart of the steps for obtaining the wind speed point prediction result of the present invention; Figure 5 This is the flowchart of the acquisition steps for the interval prediction result set of the present invention. Detailed implementation manners

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0019] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention. In addition, in the description of the present invention, the meaning of "a plurality of" is two or more, unless otherwise specifically defined.

[0020] Embodiment 1: Please refer to Figure 1 , the present invention provides a technical solution: a wind speed prediction method based on entropy clustering, including the following steps: S1: Perform multi-scale decomposition on the training set wind speed sequence through the permutation entropy algorithm to generate a subsequence set, calculate the sliding window probability distribution for the multiple subsequences, set a dynamic classification threshold based on the normalized entropy value, and divide the subsequence set into an entropy classification result set; S2: Input the original wind speed sequence of the training set and the entropy classification result set into the GRU network for training the gated state parameters to generate a component mapping model, and input the original wind speed sequence of the test set into the component mapping model to output the entropy component set of the test set; S3: Input the high-entropy components in the entropy component set of the test set into a bidirectional gated recurrent unit to generate high-entropy prediction results, input the medium-entropy components into a long short-term memory model to generate medium-entropy prediction results, input the low-entropy components into a support vector regression model to generate low-entropy prediction results, allocate initial weights based on the entropy value ratio, use the chaotic mapping algorithm to iteratively optimize the weight parameters, and perform weighted superposition on the three types of prediction results to generate the wind speed point prediction result; S4: Establish a dynamic error benchmark set based on the error standard deviation of multiple entropy classification intervals in the training set, input the wind speed point prediction result and the real-time error of the test set into the error distribution model, adjust the confidence interval width according to the relative relationship between the error and the benchmark value, and output the interval prediction result set.

[0021] The entropy classification result set specifically includes high-entropy subsequence categories, medium-entropy subsequence categories, and low-entropy subsequence categories. The test set entropy component set includes high-entropy components, medium-entropy components, and low-entropy components. The wind speed point prediction result specifically refers to the superposition value of high-entropy prediction results, medium-entropy prediction results, low-entropy prediction results, and optimized weight parameters. The interval prediction result set includes dynamic error benchmarks, real-time error distribution data, confidence interval width adjustment parameters, and probability interval ranges.

[0022] The normalized entropy value is calculated through the linear mapping formula where, represents the normalized value of the permutation entropy of the current window, represents the permutation entropy value of the i-th sliding window, represents the maximum value of the permutation entropy values among all sliding windows, represents the minimum value of the permutation entropy values among all sliding windows. The dynamic classification threshold is the coordinate value corresponding to the extreme point of the second derivative of the density curve of the mapped entropy value distribution. Among them, this extreme point represents the position where the density changes most significantly and is used as the classification boundary for category division; The dynamic classification threshold is the coordinate value corresponding to the extreme point of the second derivative of the density curve of the mapped entropy value distribution; The GRU network includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is equal to the time step of the wind speed sequence. The hidden layer has 2 layers and each layer contains 128 neurons. The number of neurons in the output layer is 32; The chaotic mapping algorithm uses the Logistic mapping function to generate a random sequence, where, is the chaotic sequence value of the n-th iteration, is the chaotic sequence value of the (n + 1)-th iteration. The fitness function adopts the mean square error form, that is where, represents the mean square error of the prediction result, is the number of samples, is the j-th true wind speed value, is the j-th predicted wind speed value. The gradient descent method is used to update the weight perturbation amplitude. When optimizing the weights, the initial weights are allocated according to the entropy value ratio, and finally, through multiple iterations of optimization, the convergent optimal weight parameters are obtained; The dynamic error benchmark set is the error standard deviation sequence arranged in ascending order according to the entropy classification labels. The confidence interval width adjustment formula is where, is the width of the current confidence interval, is the proportionality coefficient, is the wind speed prediction error value at the t-th moment, is the error benchmark value of the k-th entropy classification interval, and The difference reflects the prediction uncertainty, and based on this difference, the width of the confidence interval is dynamically adjusted. The range of the probability interval is jointly determined by the error distribution and the width parameter.

[0023] Please refer to Figure 2 , and the specific steps for obtaining the entropy classification result set are as follows: S101: Obtain the training set wind speed sequence, call the permutation entropy algorithm to set the embedding dimension parameter and the delay parameter, construct a multi-dimensional phase space according to the embedding dimension parameter, perform time-delay slicing on the sequence through the delay parameter, reconstruct the phase space trajectories of multiple time scales, and decompose the original sequence based on the similarity difference between the trajectories to generate multi-scale decomposition subsequences; The obtained training set wind speed sequence is a set of wind speed observation values arranged in chronological order. The wind speed data used here comes from the 10-minute data continuously collected by the No. 1 wind turbine in a certain wind farm during a specific period. The sampling frequency is 1 hertz (Hz), that is, one wind speed value is recorded per second, for a total of 600 data points. The numerical unit is uniformly meters per second (m / s). Some examples are shown in Table 1.

[0024] Table 1: Training set wind speed sequence ; As shown in Table 1, the specific values of the first 10 sampling points of the training set wind speed sequence are shown.

[0025] When calling the permutation entropy algorithm, first set the embedding dimension parameter , and its value is determined with reference to the sequence length and the need to capture details. Here, is set. This value can capture the main fluctuation patterns of the wind speed while avoiding overfitting noise. Then, set the delay parameter , and its value should match the characteristic time scale of the sequence. For wind speed data with a sampling frequency of 1 Hz, seconds is set, that is, adjacent consecutive data points are used for reconstruction.

[0026] According to the embedding dimension parameter , construct a three-dimensional phase space. The specific execution process is as follows: For the wind speed sequence , construct a vector sequence , where ranges from 1 to . Since seconds and , the vector is . Taking the data in Table 1 as an example, the first vector is , the second vector is , and so on until the end of the sequence. Through the delay parameter The time-delay slicing of the sequence has been reflected in the above vector construction, that is, the elements of each vector are data points in the original sequence that are delayed by 1 second. In this way, multiple groups of phase-space trajectories on different time scales are reconstructed. Actually, three-dimensional vectors are obtained. The sequence of points formed by these vectors in three-dimensional space is the phase-space trajectory.

[0027] Based on the similarity differences between the magnitude orders (i.e., permutation patterns) of the internal element values of the vectors in these phase-space trajectories, the original sequence is decomposed. This step uses the Variational Mode Decomposition (VMD) method to decompose the original wind speed sequence into modal components and a residual component. Set the decomposition layer number . These modal components represent fluctuations with different frequency and amplitude characteristics, that is, the multi-scale decomposition subsequences.

[0028] S102: Based on the multi-scale decomposition subsequences, set the covering length parameter and moving step parameter of the sliding window. Intercept the local data segments of the subsequences successively according to the step parameter, count the frequencies of the different wind speed values appearing in each window, divide the frequency by the window length parameter, calculate the probability distribution of the wind speed values corresponding to multiple windows, and generate a probability distribution matrix; The window length parameter is uniformly in seconds, and the moving step parameter is 1 second; Based on one of the multi-scale decomposition subsequences obtained in step S101, select the first modal component for processing. This subsequence numerically shows a set of fluctuating data, and its unit is still meters per second (m / s).

[0029] Set the covering length parameter of the sliding window. Its covered physical duration is uniformly 10 seconds. Since has the same sampling frequency as the original sequence, which is 1 Hz, so corresponds to 10 consecutive data points. Set the moving step parameter . Its physical duration is 1 second, corresponding to moving 1 data point. Intercept the local data segments of the subsequence successively according to the step parameter . The data intercepted by the first window is . The data intercepted by the second window is . Proceed in this way until the end of the subsequence.

[0030] Count the frequencies of the different wind speed values appearing in each window. Examine the data segment of the first window, and its specific values are (These values are Example values), where the differential wind speed values are -0.2 m / s, 0.1 m / s, 0.3 m / s, 0.4 m / s, and their occurrence frequencies are: -0.2 appears 4 times, 0.1 appears 3 times, 0.3 appears 2 times, 0.4 appears 1 time. Divide these frequencies by the window length parameter (number of data points) to calculate the probabilities of each wind speed value within the window. Specifically: the probability of the value -0.2 is , the probability of the value 0.1 is , the probability of the value 0.3 is , the probability of the value 0.4 is . For all windows, perform this statistics and calculation to generate the probability distribution of wind speed values corresponding to each window. These distributions together constitute a probability distribution matrix, where the rows represent windows and the columns represent the differential wind speed values that appear within the window, and the matrix elements are the corresponding probabilities.

[0031] S103: Extract the permutation entropy value of each window from the probability distribution matrix, locate the maximum and minimum values of the entropy value sequence, map the entropy value to the 0 - 1 interval, analyze the distribution density curve of the mapped entropy value, calculate the second derivative of the curvature with respect to the entropy value, locate the entropy value coordinates corresponding to the derivative extreme points, and set the coordinate values as classification thresholds to divide high - entropy and low - entropy categories, generating the dynamic classification threshold result; The second derivative is calculated by the numerical difference method, and the extreme point determination criterion is that the derivative sign changes from positive to negative or from negative to positive.

[0032] For the local data segments intercepted from each window determined in step S102, extract their corresponding permutation entropy values. The specific calculation is as follows: Take the subsequence data segment within the window, for example, a data segment of length , and follow the embedding dimension and delay parameter set in S101 and seconds, perform phase - space reconstruction on this 10 - point data segment to generate three - dimensional vectors: . Determine the permutation pattern of the elements for each three - dimensional vector. There are possible permutation patterns (such as , , etc.). Statistically count the number of times each permutation pattern appears in these 8 vectors ( ), calculate its frequency , then the permutation entropy value of this window is calculated as (when , ). Repeat this process for all windows to obtain a sequence of permutation entropy values , where is the total number of windows.

[0033] Locate the maximum and minimum of this entropy value sequence. Suppose the calculated minimum permutation entropy value of the entire sequence is 0.15 and the maximum permutation entropy value is 0.95, then , , map each entropy value to the 0 - 1 interval through the linear normalization formula . For a point with an original entropy value of , its normalized value is . Analyze the distribution density curve of these mapped normalized entropy values and obtain a smooth probability density function using the kernel density estimation method . Calculate the second derivative of this density function with respect to by the central difference method: , where is a selected very small step size, taking .

[0034] Locate the entropy value coordinates corresponding to the extreme points of the second derivative . The determination criterion for the extreme points is that the sign of the second derivative changes and its first derivative (i.e., ) is zero or close to zero (in actual operation, it is often simplified to finding the zero point of the sign change, or directly finding the local maximum and minimum points of ). These points correspond to the inflection points of the original density function . Set the values corresponding to these inflection points as the classification thresholds. The specific setting of the thresholds refers to the positions of the identified inflection points. If, after calculation and analysis, the first significant inflection point (the second derivative changes from positive to negative, corresponding to the point where the slope of the density function increases fastest and then starts to slow down) is identified at the normalized entropy value of 0.33, set it as the first classification threshold . If the second significant inflection point (the second derivative changes from negative to positive, corresponding to the point where the slope of the density function decreases fastest and then starts to increase) is identified at the normalized entropy value of 0.67, set it as the second classification threshold . Then the rule for dividing high - entropy and low - entropy categories is: the sequence segment with a normalized entropy value is defined as low - entropy; is defined as medium - entropy; The sequence segment is defined as high entropy, from which the dynamic classification threshold result is generated. The "dynamic" here is reflected in that the threshold is automatically determined according to the entropy distribution characteristics of the current training data.

[0035] Please refer to Figure 3 , and the specific steps for obtaining the entropy component set of the test set are as follows: S201: Obtain the original wind speed sequence of the training set and the entropy classification result set. Based on the time step characteristics of the wind speed sequence, construct the input layer of the GRU network. At the same time, normalize and initialize the weights of the input layer according to the multi-class probability distribution in the entropy classification result set. Use the gradient descent method to calculate the weight errors of the forget gate and the input gate step by step in time, and update the gating unit parameters through chain derivation to generate the gating state parameter set; The normalization initialization uses the Xavier method, and the learning rate of the gradient descent method is set to ; Obtain the original wind speed sequence of the training set (the complete 600 data points shown in Table 1), and the entropy classification result set generated in S103. This set is the entropy category (low, medium, high) corresponding to each 10-second data segment (window) in the training set. Based on the time step characteristics of the wind speed sequence (i.e., the wind speed value at each time point ), construct the input layer of the gated recurrent unit (GRU) network, and the input dimension is 1 (only the wind speed value). )

[0036] At the same time, according to the entropy classification result set obtained in step S103, count the proportion of each category (low, medium, high entropy) in the entire training set. Let the proportion of low-entropy data segments be , the proportion of medium-entropy data segments be , and the proportion of high-entropy data segments be , where are the numbers of low, medium, and high entropy data segments respectively, is the total number of data segments. Initialize the weights from the input layer to the hidden layer of the GRU network and the weights of the recurrent connection using the Xavier normalization method. This method makes the initial weight values adapt to the size of the layer and helps the stable propagation of gradients. The specific weights are drawn from the uniform distribution , where is the number of input neurons, is the number of output neurons. The calculated category proportions are used for sample weighting or loss function adjustment in the subsequent model training, rather than directly used in the Xavier initialization process here.

[0037] Optimize the GRU network parameters using the gradient descent method, and set the learning rate , which is an empirical value, usually between and Select among them. The specific task is to use the GRU network to perform a certain prediction or representational learning on the wind speed sequence (such as predicting the wind speed at the next moment). By comparing the difference (loss) between the network output and the true target value, calculate the internal reset gate of the GRU step by step and the update gate parameters (weight matrix and bias vector ), as well as the parameters of the candidate hidden state (weight matrix and bias vector ), with respect to the gradient of the loss function. Through the Backpropagation Through Time (BPTT) algorithm and the chain rule of differentiation, iteratively update the parameters of these gated units to generate a set of gated state parameters, that is, the weights and biases of the trained GRU network.

[0038] S202: Call the set of gated state parameters, segment the wind speed sequence of the training set into subsequence segments, and at the same time establish a mapping index between the subsequences and the hidden layer states based on the entropy classification labels. Calculate the error gradient between the hidden layer state and the label through backpropagation, and update the output gate weights and the parameters of the fully connected layer to generate the component mapping model parameters; Call the set of gated state parameters trained in S201, that is, the optimized GRU network. Segment the original wind speed sequence of the training set according to the window length seconds (10 data points) and the step size seconds (1 data point) defined in S102 to obtain subsequence segments consistent with those in S103 for entropy classification.

[0039] Based on the entropy classification labels (low, medium, high) of each training subsequence segment obtained in S103, establish a mapping index between the subsequence segments and the hidden layer states obtained after being processed by the current GRU network. The specific operation is as follows: Input each subsequence segment into the GRU network and extract the hidden state vector at the last time step of this subsequence segment. This vector is used as the feature representation of this subsequence segment and establish a corresponding relationship with its known entropy label (such as "low entropy").

[0040] On this basis, construct a classification layer, usually a fully connected layer. Its input is the hidden state output by the GRU, the output dimension is 3 (corresponding to the three entropy categories of low, medium, and high), and a Softmax activation function is connected at the back to convert the output into the probabilities of each category. Calculate the error gradient between the predicted entropy label output by this classification layer and the true entropy label through the backpropagation algorithm. The loss function uses cross-entropy loss. This error gradient is not only used to update the weights and biases of this fully connected classification layer It is also further backpropagated to fine-tune the parameters of the GRU network (specifically, the part related to the output, or the parameters of the entire GRU network are jointly tuned at this stage), generating the final component mapping model parameters. This model can receive a subsequence segment of wind speed and output its corresponding entropy classification.

[0041] S203: Segment the wind speed sequence of the test set into subsequence segments based on the component mapping model parameters, input them into the GRU network to calculate the hidden layer state vectors, extract the neuron activation values corresponding to the entropy classification labels, and splice the activation value sequences according to time steps to generate the test set entropy component set.

[0042] Based on the component mapping model parameters (including GRU parameters and the subsequent classification layer parameters) trained in S202, the wind speed sequence of the test set, which is a new segment of wind speed data not used in training, is also segmented according to the same window length in seconds and step size in seconds to form test subsequence segments.

[0043] Input each test subsequence segment into the GRU network part of the component mapping model trained in S202, and use the learned parameters to calculate the hidden layer state vector of this subsequence segment. Usually, the hidden state at the last time step is taken as a representative. Then, input this hidden state into the subsequent fully connected classification layer in the model. This classification layer outputs the activation values of three neurons. After passing through the Softmax function, the probability distributions corresponding to the low, medium, and high entropy categories are obtained , where is the index of the test subsequence segment. For example, for a certain test subsequence segment, its output probability may be , indicating that the probability of this segment being judged as high entropy is 0.60.

[0044] Splice these probability vectors according to time steps (or the order of subsequence segments), or determine the entropy classification label of each subsequence segment according to the maximum probability principle (such as being judged as "high entropy" in the above example). If the method of splicing probability vectors is adopted, the test set entropy component set is a sequence composed of multiple three-dimensional probability vectors; if the method of determining labels is adopted, it is an entropy label sequence, generating the test set entropy component set for use in subsequent steps.

[0045] Please refer to Figure 4 , and the specific steps for obtaining the wind speed point prediction result are as follows: S301: Call the entropy component set of the test set. Input the high-entropy components into a bidirectional gated recurrent unit based on the entropy classification labels, expand the forward and backward hidden state transmissions of the time steps, calculate the output values of the gated units at multiple time steps, input the medium-entropy components into a long short-term memory model, screen the time-dependent features according to the forget gate and input gate states, input the low-entropy components into a support vector regression model, map the feature space using a radial basis kernel function and solve the regression hyperplane to generate a set of component prediction results; Bandwidth parameter of the radial basis kernel function , and the sequential minimal optimization algorithm is used to solve the hyperplane; Call the entropy component set of the test set obtained in step S203. This set is the predicted entropy classification labels (low, medium, high) of each 10-second subsequence segment in the test set.

[0046] For the wind speed subsequence segments in the test set determined to be high-entropy by the component mapping model, input their original wind speed data (these 10 data points) into a bidirectional gated recurrent unit (BiGRU) network specifically trained for high-entropy sequence prediction. The BiGRU network processes the sequence through forward and backward GRU layers, and at each time step , the forward GRU layer outputs a hidden state , the backward GRU layer outputs a hidden state , and these two states are usually concatenated or added ( or ) and then, through an output layer (such as a fully connected layer), the predicted value at this time step or the overall prediction result of the sequence is calculated to generate a prediction result sequence of the high-entropy components . High-entropy sequences usually fluctuate violently, and BiGRU can better capture their bidirectional dependence features.

[0047] For the wind speed subsequence segments determined to be medium-entropy, input their original wind speed data into a long short-term memory model (LSTM) network. LSTM effectively screens and memorizes the long-range time-dependent features in the sequence through its internal input gate, forget gate, and output gate structures, and is suitable for medium-complexity wind speed changes with a certain degree of persistence and trend to generate a prediction result sequence of the medium-entropy components .

[0048] For the wind speed subsequence segments determined to be low-entropy, input their original wind speed data (or statistical features extracted therefrom, such as mean, variance, etc.) into a support vector regression (SVR) model. This SVR model uses a radial basis kernel function (RBFKernel), and its expression is , where is the bandwidth parameter of the kernel function. The setting of this parameter affects the complexity and smoothness of the model and is determined by methods such as cross-validation. Set , is the sample and the square of the Euclidean distance between. The SVR implicitly maps the input features to a high-dimensional feature space through this kernel function and finds a regression hyperplane in this space such that the deviations of most sample points from the hyperplane are within a certain tolerance while minimizing the structural risk. The solution of the hyperplane adopts the sequential minimal optimization (SMO) algorithm, which is an efficient method for solving quadratic programming problems and generates a sequence of prediction results with low entropy components . The low-entropy sequence is usually relatively stable, and the SVR can better fit its trend.

[0049] Finally, the predicted values of the test subsequence segments corresponding to their respective entropy categories of these three models are collected to form a set of component prediction results containing the prediction results of all test points.

[0050] S302: Based on the initial entropy value ratios of the high-entropy, medium-entropy, and low-entropy components in the set of component prediction results, calculate the proportion of the prediction result of the high-entropy component in the total entropy value, and use this proportion as the initial weight distribution coefficient. At the same time, call the Logistic mapping function in the chaotic mapping algorithm to generate a random sequence, use the mean square error of the prediction results as the fitness function, and iteratively update the perturbation amplitude of the weight parameters to generate optimized weight parameters; The update step size of the perturbation amplitude is ; Based on the set of component prediction results obtained in step S301 (i.e., the sequence , , ), and the initial proportions of different entropy components appearing in the overall wind speed sequence obtained during the training phase or through the analysis of historical data. These proportions can be used as initial weights. Assume that the proportion of the high-entropy component is 0.20, the proportion of the medium-entropy component is 0.45, and the proportion of the low-entropy component is 0.35 (consistent with the statistics in the training set in S201), then the initial weights are set as , and the sum of these weights is 1.

[0051] Call the Logistic mapping function in the chaotic mapping algorithm, and its iterative formula is , set the control parameter (this value makes the mapping in a chaotic state), the initial value is randomly selected from the interval (0,1), but does not include 0 and 1, take , and generate a pseudo-random chaotic sequence by iterating this function.

[0052] Use the mean square error (MSE) between the combined prediction result and the true wind speed value as the fitness function, and the calculation method of MSE is , where is the number of prediction points participating in the evaluation, is the true wind speed value of the th point. The weight parameters are updated through an iterative optimization algorithm (such as particle swarm optimization or genetic algorithm, simplified here as search based on chaotic perturbation), and constrained with each item being non - negative. In each iteration, the current weights are slightly perturbed using the chaotic values generated by the Logistic map, and the update step size of the perturbation amplitude is set to . The specific perturbation method can be: starting from the current weights, a perturbation direction and magnitude are generated using the chaotic sequence to produce a new set of candidate weights. If the MSE generated by the new weight set is lower than the current optimal MSE, then this new set of weights is accepted. The iteration process continues until the MSE converges or reaches a preset maximum number of iterations (such as 200 times) to generate the optimized weight parameters

[0053] S303: Call the set of component prediction results and the optimized weight parameters, perform a scalar product of the high - entropy prediction results with the high - entropy weight coefficients in the optimized weights, synchronously calculate the products of the medium - entropy and low - entropy prediction results with their corresponding weights, and align and superimpose the three types of weighted results according to the time steps to generate the wind speed point prediction results.

[0054] Call the set of component prediction results generated in step S301, that is, the prediction sequence of the high - entropy part , the prediction sequence of the medium - entropy part , and the prediction sequence of the low - entropy part , as well as the optimized weight parameters generated in step S302 .

[0055] Perform a scalar product operation on the prediction results of the high - entropy component and the optimized high - entropy weight coefficients to obtain the weighted high - entropy prediction contribution . Synchronously, perform a scalar product on the prediction results of the medium - entropy component and the corresponding medium - entropy weight coefficients to obtain the weighted medium - entropy prediction contribution . Similarly, perform a scalar product on the prediction results of the low - entropy component and the corresponding low - entropy weight coefficients to obtain the weighted low - entropy prediction contribution .

[0056] At each time step , the wind speed data for this time step has been classified by S203 as one of high, medium, or low entropy. Therefore, during combination, only the output of the prediction model corresponding to the category at this time step is used. One implementation is that for time step , if it belongs to high entropy, the prediction for this point is , if it belongs to medium entropy, it is , if it belongs to low entropy, it is . Then, the final wind speed point prediction result is obtained by weighting the category-specific prediction value of this point with the corresponding category weight optimized overall (here the weight is for the contribution of the entire component, rather than being selected point by point). More reasonably, the final prediction is based on the entropy category to which this time point belongs. Then is directly taken from the prediction of the model of this category . Then, the weights in S302 are used to evaluate the contribution of different models in the overall prediction or their fusion in a certain integration strategy, rather than directly used to generate the final prediction value for each point. Revised understanding: The weights in S302 are applied to combination. Assume that each model gives predictions for all points (or gives a baseline prediction or zero prediction for points not dominated by it). Then at each time point , the wind speed point prediction result is calculated as: . If a certain time point is clearly classified into a certain entropy category, such as high entropy, then for this time point and may be 0 or non-dominant predictions output by their respective models. More realistically, at time point , if this point is classified by S203 as , then takes , while the weight is used at a higher level (such as when determining the model combination strategy or evaluating the overall performance of the model). According to the description in the original text "align and superimpose the weighted results of the three categories by time step", adopt the first interpretation, that is, each model has a (possibly dominant or non-dominant) prediction value for all time points, or more precisely, each time point corresponds to only one entropy category. Therefore, it is not very reasonable that the prediction value of this time point only comes from the model corresponding to this entropy category and is multiplied by the global weight of this model. The correct understanding should be that for each time point , first determine its entropy category , and then use the prediction of the corresponding model . The combination in S303 is more likely to be: use for points belonging to the high entropy interval, and use , points belonging to the low-entropy interval use . And the optimized weights are used to evaluate the overall performance of the combined model in S302, or these weights represent the "degree of trust" or "degree of contribution" in different entropy states and are used to adjust the final combination.

[0057] Here, strictly according to the literal meaning of "aligning and superimposing the weighted results of the three categories by time step": the high-entropy prediction results are scalar-multiplied with the high-entropy weight coefficients in the optimized weights , synchronously calculating the product of the medium-entropy prediction results and the corresponding weights , and the product of the low-entropy prediction results and the corresponding weights . Align and sum the weighted results of these three time series at each time step to obtain the final wind speed point prediction result , generating a sequence of wind speed point prediction results. Here it is assumed that , , is a prediction sequence defined for all time points , even if the point does not belong to the entropy category dominated by this model, the model still gives a prediction (possibly of low quality).

[0058] Please refer to Figure 5 , and the steps to obtain the interval prediction result set are specifically as follows: S401: Obtain the prediction error sequences corresponding to multiple entropy classification intervals of the training set, divide the error data by interval, calculate the standard deviation of the error values within each entropy classification interval, and arrange the standard deviations in ascending order according to the entropy classification labels to generate a dynamic error benchmark set; Obtain the final prediction sequence of the training set after the prediction process described in S303 (that is, first perform entropy classification on the training set data, then apply the corresponding BiGRU, LSTM, SVR models for prediction, and finally use the optimized weights for weighted combination) . Calculate its corresponding prediction error sequence , where is the true wind speed value in the training set.

[0059] According to the entropy classification labels (low, medium, high) obtained from the training set data in S103, divide this prediction error sequence according to the entropy classification intervals to which the corresponding time points belong, forming three independent subsets of error data: , , Among them, the data segment at time point belongs to low entropy, and the data segment at time point belongs to medium entropy, and the data segment at time point belongs to high entropy.

[0060] For all error values within each entropy classification interval, calculate their standard deviation. The error standard deviation of the low entropy interval , where is the mean of the errors in the low entropy interval, and is the number of error points in the low entropy interval. The error standard deviation of the medium entropy interval . The error standard deviation of the high entropy interval . It is calculated that: the error standard deviation of the low entropy interval m / s, the error standard deviation of the medium entropy interval m / s, and the error standard deviation of the high entropy interval m / s. Store these calculated standard deviations corresponding to their respective entropy classification labels (low, medium, high) to form a dynamic error benchmark set. This benchmark set reflects the typical prediction uncertainty levels of the model under different complexities (entropy states).

[0061] Table 2: Dynamic Error Benchmark Set ; As shown in Table 2, this table lists the error standard deviations calculated based on the final prediction errors of the training set in different entropy classification intervals, and these values will be used as the benchmarks for adjusting the prediction width of subsequent intervals.

[0062] S402: Call the dynamic error benchmark set, align the wind speed point prediction results with the real-time errors of the test set according to time steps, and at the same time, based on the normal distribution setting of the error distribution model, calculate the absolute value of the difference between the error value at each time step and the corresponding entropy classification benchmark value, and update the mean and variance parameters of the distribution model using the maximum likelihood estimation method to generate error distribution parameters; The initial mean of the normal distribution is , and the variance is ; Call the dynamic error benchmark set generated in step S401 (i.e., the standard deviations of each entropy classification interval m / s, m / s, m / s), and the wind speed point prediction results for the test set generated in S303 . At the same time, obtain the real-time (or actual) wind speed observation values corresponding to the test set to calculate the real-time prediction error of the test set. Align the point prediction results of the test set with the real-time errors according to time steps.

[0063] For the error distribution model, it is assumed to follow a normal distribution, and a set of distribution parameters is maintained for each entropy category (low, medium, high). Initially, the error mean for each category can be set m / s, and the error variance can use the square of the corresponding category error standard deviation obtained by S401, i.e., , , .

[0064] When processing the test set data, for each time step , first determine the entropy classification to which this time step (low, medium, or high) belongs through the component mapping model of S203. Then, collect all the real-time errors of the test set belonging to this classification (where is the time point index belonging to category ). Use the maximum likelihood estimation method (MLE) to update the mean and variance parameters of the normal distribution model of the errors of this category . Specifically, for the low entropy category, its updated mean , and the updated variance , where is the number of low entropy points in the test set. The same calculations are performed for the medium and high entropy categories to generate the updated error distribution parameters for each entropy classification interval .

[0065] S403: Based on the error distribution parameters, according to the proportion of the current time step error value exceeding the benchmark value, linearly expand or shrink the confidence interval width by a proportionality coefficient. Use the mean of the distribution model as the center of the interval, and calculate the upper and lower bounds in combination with the adjusted width to generate a set of interval prediction results

[0066] Based on the updated error distribution parameters (i.e., the mean and the standard deviation ) for each entropy classification interval calculated in step S402, and the corresponding category error benchmark standard deviation obtained in S401 (i.e., the values in Table 2).

[0067] When generating an interval prediction for a certain time point in the test set, first determine the entropy category to which this point belongs. The center of the prediction interval for this point is set to the point prediction value obtained by S303 plus the updated error mean of this category for bias correction: .

[0068] Basic prediction interval half-width by the corresponding confidence level (e.g., 95%) of the normal distribution quantile (for 95%, ) and the updated error standard deviation of this category The product is determined as follows: .

[0069] Then, according to certain characteristics of the current time step (such as the average absolute value of recent errors or the uncertainty indication of the predicted value at this point) and the benchmark error standard deviation of the entropy classification to which this point belongs Compare, and adjust the interval width. Let the scaling factor be used for adjustment. For example, if the recent error is significantly greater than the benchmark, , otherwise . Define , where is an adjustment coefficient, taking , is the actual prediction error of the previous time step, and the denominator comes from Table 2. If , then expand the interval, otherwise it may be shrunk (ensuring that is positive). A specific adjusted half-width is calculated as , that is, the adjustment factor is restricted between 0.5 and 2.0 to prevent excessive scaling.

[0070] Finally, the upper and lower bounds of the prediction interval at this time point are respectively: upper bound lower bound Repeat this process for all time points in the test set to generate a complete set of interval prediction results.

[0071] An entropy clustering-based wind speed prediction system, the entropy clustering-based wind speed prediction system is used to execute the above entropy clustering-based wind speed prediction method, and the system includes: A decomposition and clustering module, configured to generate a set of subsequences by decomposing the wind speed sequence of the training set through the permutation entropy algorithm, divide the subsequences by using a sliding window, calculate the probability distribution of the wind speed values within the window, generate a dynamic classification threshold based on the normalized entropy value function, divide the subsequences into a high-entropy component set, a medium-entropy component set, and a low-entropy component set, integrate them into an entropy classification result set, and transmit the entropy classification result set to the component mapping module; The medium-entropy component set is a set of subsequences whose normalized entropy values are in the middle 30% interval of the dynamic classification threshold; The component mapping module is used to obtain the original wind speed sequence of the training set and the entropy classification result set, train and generate a component mapping model through a GRU network, input the wind speed sequence of the test set into the model to output the entropy component set of the test set, and transfer the entropy component set of the test set to the weight prediction module; The weight prediction module is used to call the entropy component set of the test set, input the high-entropy component set into a bidirectional gated recurrent unit to generate a high-entropy prediction value, input the medium-entropy component set into a long short-term memory model to generate a medium-entropy prediction value, input the low-entropy component set into a support vector regression model to generate a low-entropy prediction value, allocate initial weights based on the entropy value ratio, optimize the weight parameters using a chaotic mapping algorithm, superimpose to generate a wind speed point prediction result, and transfer the wind speed point prediction result to the interval correction module; The interval correction module is used to establish a dynamic error benchmark set according to the standard deviation of the prediction error of the entropy classification interval of the training set, call the wind speed point prediction result and the real-time error of the test set, calculate the relative difference between the real-time error and the benchmark value through an error distribution model, adjust the width of the confidence interval, and output an interval prediction result set.

[0072] The above are only the preferred embodiments of the present invention, and do not limit the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as the technical content of the present invention is not departed from, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A wind speed prediction method based on entropy clustering, characterized in that, It includes the following steps: S1: Perform multi-scale decomposition on the training set wind speed sequence through the permutation entropy algorithm to generate a subsequence set, calculate the probability distribution of the multi-subsequences using a sliding window, set a dynamic classification threshold based on the normalized entropy value, and divide the subsequence set into an entropy classification result set; S2: Input the original wind speed sequence of the training set and the entropy classification result set into the GRU network for training the gated state parameters, generate a component mapping model, and input the original wind speed sequence of the test set into the component mapping model to output the entropy component set of the test set; S3: Input the high-entropy components in the entropy component set of the test set into a bidirectional gated recurrent unit to generate a high-entropy prediction result, input the medium-entropy components into a long short-term memory model to generate a medium-entropy prediction result, input the low-entropy components into a support vector regression model to generate a low-entropy prediction result, allocate initial weights based on the entropy value ratio, use the chaotic mapping algorithm to iteratively optimize the weight parameters, and perform weighted superposition on the three types of prediction results to generate a wind speed point prediction result.

2. The wind speed prediction method based on entropy clustering according to claim 1, characterized in that The entropy classification result set specifically includes high-entropy subsequence categories, medium-entropy subsequence categories, and low-entropy subsequence categories. The entropy component set of the test set includes high-entropy components, medium-entropy components, and low-entropy components. The wind speed point prediction result specifically refers to the superposition value of the high-entropy prediction result, medium-entropy prediction result, low-entropy prediction result, and the optimized weight parameters.

3. The wind speed prediction method based on entropy clustering according to claim 2, wherein, The normalized entropy value is calculated through the linear mapping formula wherein represents the normalized value of the permutation entropy of the current window, represents the permutation entropy value of the i-th sliding window, represents the maximum value of the permutation entropy values in all sliding windows, represents the minimum value of the permutation entropy values in all sliding windows, and the dynamic classification threshold is the coordinate value corresponding to the extreme point of the second derivative of the entropy value distribution density curve after mapping. Among them, this extreme point represents the position where the density change is the most significant and is used as the classification boundary for class division; The dynamic classification threshold is the coordinate value corresponding to the extreme point of the second derivative of the entropy value distribution density curve after mapping; The GRU network includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is equal to the time step of the wind speed sequence. The hidden layer has 2 layers and each layer contains 128 neurons. The number of neurons in the output layer is 32; The chaotic mapping algorithm uses the Logistic mapping function to generate a random sequence, where is the chaotic sequence value at the nth iteration, is the chaotic sequence value at the (n + 1)th iteration. The fitness function adopts the mean square error form, that is , where represents the mean square error of the prediction result, is the number of samples, is the jth true wind speed value, is the jth predicted wind speed value. The gradient descent method is used to update the weight perturbation amplitude. When optimizing the weights, the initial weights are allocated according to the entropy ratio. Finally, through multiple iterations of optimization, the convergent optimal weight parameters are obtained.

4. The wind speed prediction method based on entropy clustering according to claim 3, wherein, The specific steps for obtaining the entropy classification result set are as follows: S101: Obtain the training set wind speed sequence, call the permutation entropy algorithm to set the embedding dimension parameter and the delay parameter, construct a multi-dimensional phase space according to the embedding dimension parameter, perform time-delay slicing on the sequence through the delay parameter, reconstruct the phase space trajectories of multiple time scales, and decompose the original sequence based on the similarity difference between the trajectories to generate multi-scale decomposition subsequences; S102: Based on the multi-scale decomposition subsequences, set the covering length parameter and the moving step parameter of the sliding window, successively intercept the local data segments of the subsequences according to the step parameter, count the frequency of the occurrence of different wind speed values in each window, divide the frequency by the window length parameter, calculate the probability distribution of the corresponding wind speed values in multiple windows, and generate a probability distribution matrix; The window length parameter is uniformly in seconds, and the moving step parameter is 1 second; S103: Extract the permutation entropy value of each window from the probability distribution matrix, locate the maximum and minimum values of the entropy value sequence, map the entropy value to the 0-1 interval, analyze the distribution density curve of the mapped entropy value, calculate the second derivative of the curvature with respect to the entropy value change, locate the entropy value coordinate corresponding to the derivative extreme point, and set the coordinate value as the classification threshold to divide the high-entropy and low-entropy categories, generating a dynamic classification threshold result; The second derivative is calculated by the numerical difference method, and the extreme point determination criterion is that the derivative sign changes from positive to negative or from negative to positive.

5. The wind speed prediction method based on entropy clustering according to claim 4, wherein The specific steps for obtaining the entropy component set of the test set are as follows: S201: Obtain the original wind speed sequence of the training set and the entropy classification result set. Construct the input layer of the GRU network based on the time step characteristics of the wind speed sequence. At the same time, normalize and initialize the weights of the input layer according to the multi-class probability distribution in the entropy classification result set. Use the gradient descent method to calculate the weight errors of the forget gate and the input gate step by step in time, and update the parameters of the gated unit through chain derivation iteration to generate a set of gated state parameters; The above-mentioned normalization initialization adopts the Xavier method, and the learning rate of the gradient descent method is set to ; S202: Call the set of gated state parameters, divide the wind speed sequence of the training set into subsequence segments. At the same time, establish a mapping index between the subsequence and the hidden layer state based on the entropy classification label. Calculate the error gradient between the hidden layer state and the label through backpropagation, and update the output gate weight and the parameters of the fully connected layer to generate the component mapping model parameters; S203: Divide the wind speed sequence of the test set into subsequence segments based on the component mapping model parameters, input it into the GRU network to calculate the hidden layer state vector, extract the neuron activation values corresponding to the entropy classification label, and splice the activation value sequences step by step in time to generate the test set entropy component set.

6. The wind speed prediction method based on entropy clustering according to claim 5, characterized in that The specific steps for obtaining the wind speed point prediction result are as follows: S301: Call the test set entropy component set. Input the high-entropy components into the bidirectional gated recurrent unit based on the entropy classification label, expand the forward and backward hidden state transmissions in time steps, calculate the output values of the multi-time step gated units. Input the medium-entropy components into the long short-term memory model, screen the time-dependent features according to the forget gate and the input gate states. Input the low-entropy components into the support vector regression model, map the feature space using the radial basis kernel function and solve the regression hyperplane to generate a set of component prediction results; The bandwidth parameter of the radial basis kernel function , and the sequential minimal optimization algorithm is used to solve the hyperplane; S302: Based on the initial entropy value ratio of the high-entropy, medium-entropy, and low-entropy components in the set of component prediction results, calculate the proportion of the high-entropy component prediction result in the total entropy value, and use the proportion as the initial weight distribution coefficient. At the same time, call the Logistic mapping function in the chaos mapping algorithm to generate a random sequence, use the mean square error of the prediction result as the fitness function, and iteratively update the perturbation amplitude of the weight parameters to generate optimized weight parameters; The update step size of the perturbation amplitude is ; S303: Call the set of component prediction results and the optimized weight parameters, perform a scalar product of the high-entropy prediction result and the high-entropy weight coefficient in the optimized weights. Synchronously calculate the products of the medium-entropy and low-entropy prediction results and the corresponding weights, and align and superimpose the three types of weighted results step by step in time to generate the wind speed point prediction result.

7. The wind speed prediction method based on entropy clustering according to claim 6, wherein, The method further includes: S4: Establish a dynamic error benchmark set based on the error standard deviations of multiple entropy classification intervals in the training set. Input the wind speed point prediction result and the real-time error of the test set into the error distribution model, adjust the width of the confidence interval according to the relative relationship between the error and the benchmark value, and output a set of interval prediction results.

8. The wind speed prediction method based on entropy clustering according to claim 7, characterized in that, The set of interval prediction results includes dynamic error benchmarks, real-time error distribution data, confidence interval width adjustment parameters, and probability interval ranges; The dynamic error benchmark set is a sequence of error standard deviations arranged in ascending order of entropy classification labels, and the confidence interval width adjustment formula is , where is the width of the current confidence interval,[[]] is the proportionality coefficient,[[]] is the wind speed prediction error value at the t-th moment,[[]] is the error benchmark value of the k-th entropy classification interval,[[]] and The difference between them reflects the prediction uncertainty. Based on this difference, the confidence interval width is dynamically adjusted, and the probability interval range is jointly determined by the error distribution and the width parameter.

9. The wind speed prediction method based on entropy clustering according to claim 8, characterized in that The specific steps for obtaining the set of interval prediction results are as follows: S401: Obtain the prediction error sequences corresponding to multiple entropy classification intervals of the training set, divide the error data by intervals, calculate the standard deviation of the error values within each entropy classification interval, sort the standard deviations in ascending order according to the entropy classification labels, and generate a dynamic error benchmark set; S402: Call the dynamic error benchmark set, align the wind speed point prediction results with the real-time error of the test set by time steps. At the same time, based on the normal distribution setting of the error distribution model, calculate the absolute value of the difference between the error value at each time step and the corresponding entropy classification benchmark value, and use the maximum likelihood estimation method to update the mean and variance parameters of the distribution model to generate error distribution parameters; The initial mean of the normal distribution is , and the variance is ; S403: Based on the error distribution parameters, linearly expand or shrink the width of the confidence interval according to the proportion of the current time step error value exceeding the benchmark value. Use the mean of the distribution model as the center of the interval, and calculate the upper and lower bounds in combination with the adjusted width to generate an interval prediction result set.

10. A wind speed prediction system based on entropy clustering, characterized in that, The system is used to implement the wind speed prediction method based on entropy clustering according to any one of claims 1-9. The system includes: A decomposition clustering module, which is used to decompose the wind speed sequence of the training set through the permutation entropy algorithm to generate a subsequence set, divide the subsequences by a sliding window, calculate the probability distribution of the wind speed values within the window, generate a dynamic classification threshold based on the normalized entropy value function, divide the subsequences into a high-entropy component set, a medium-entropy component set, and a low-entropy component set, integrate them into an entropy classification result set, and transfer the entropy classification result set to the component mapping module; The medium-entropy component set is a set of subsequences whose normalized entropy values are within the middle 30% interval of the dynamic classification threshold; A component mapping module, which is used to obtain the original wind speed sequence of the training set and the entropy classification result set, train through a GRU network to generate a component mapping model, input the wind speed sequence of the test set into the model to output the entropy component set of the test set, and transfer the entropy component set of the test set to the weight prediction module; A weight prediction module, which is used to call the entropy component set of the test set, input the high-entropy component set into a bidirectional gated recurrent unit to generate a high-entropy prediction value, input the medium-entropy component set into a long short-term memory model to generate a medium-entropy prediction value, input the low-entropy component set into a support vector regression model to generate a low-entropy prediction value, allocate initial weights based on the entropy value ratio, optimize the weight parameters using a chaotic mapping algorithm, and superimpose to generate a wind speed point prediction result, and transfer the wind speed point prediction result to the interval correction module; An interval correction module, which is used to establish a dynamic error benchmark set according to the prediction error standard deviation of the entropy classification interval of the training set, call the wind speed point prediction result and the real-time error of the test set, calculate the relative difference between the real-time error and the benchmark value through the error distribution model, adjust the width of the confidence interval, and output an interval prediction result set.

Images