Medium-and-long-term wind speed prediction method based on hybrid expert decomposition and adaptive moving average

Abstract

The invention discloses a medium-and-long-term wind speed prediction method and system mixed with expert decomposition and adaptive moving average, and belongs to the technical field of wind speed prediction, and the method comprises the steps: carrying out the preprocessing of an original wind speed sequence, obtaining a smooth wind speed sequence, and inputting the smooth wind speed sequence into an MOEA model; extracting a period term and a trend term of the wind speed sequence for the first time in the encoder through a decomposition module; a first branch of the decoder performs decomposition of a multi-head self-correlation mechanism and a hybrid expert decomposition module for multiple times, a trend term decomposed each time and a periodic term decomposed each time continue to be transmitted in the first branch, and the trend term decomposed each time is transmitted to a second branch; and superposing the first trend term and the trend term decomposed each time in a decoder, and adding the superposed trend term and the finally predicted periodic term to obtain a wind speed prediction result. According to the method, the influence of the distribution offset in the wind speed sequence on the wind speed prediction result is relieved, the feature extraction capability of the decomposition technology on the wind speed sequence is reserved, and the problem of too high time cost in practical application is also avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of wind speed prediction, and in particular to a medium- and long-term wind speed prediction method based on hybrid expert decomposition and adaptive sliding average. Background Art

[0002] Wind speed forecasting is a core component of wind power forecasting. Depending on the input variables, wind speed forecasting can be categorized as univariate or multivariate. Multivariate wind speed forecasting models typically rely on time series data, such as wind direction and temperature. Multivariate wind speed forecasting models require input of future forecast values ​​for these factors, but these forecast values ​​themselves are subject to errors, which can lead to error accumulation and further impact the accuracy of the final wind speed forecast. Univariate wind speed forecasting, on the other hand, directly uses historical wind speed data to predict future wind speeds and is therefore immune to error accumulation. However, existing univariate wind speed forecasting methods have limitations in capturing the long-term dependencies of wind speed series, resulting in low prediction accuracy. To improve the accuracy of univariate wind speed forecasting, many studies have employed decomposition techniques as data preprocessing, combining the decomposed wind speed subsequences with a prediction model to form a hybrid model. However, such hybrid models require splicing the newly generated wind speed series onto the end of the historical wind speed series to form a longer series for decomposition before entering the subsequent prediction process. This approach incurs a significant time cost. Therefore, studying a univariate wind speed prediction model with the ability to capture long-term dependencies and solving the time cost problem brought by decomposition technology are of great significance for improving the accuracy of medium- and long-term wind speed forecasts.

[0003] Among traditional wind speed forecasting methods, physical methods typically utilize numerical weather forecasts (NWP) and geographic terrain data for wind speed prediction. Although these methods do not rely on historical data, they struggle to cope with complex weather patterns. Statistical methods such as moving averages, Markov models, and autoregressive models primarily rely on historical data for wind speed prediction, but are less effective for nonlinear and non-stationary sequences. In recent years, with the development of deep learning, models such as recurrent neural networks (RNNs, LSTMs, GRUs) have achieved promising results in wind speed forecasting. Temporal convolutional networks (TCNs), in particular, have been shown to outperform recurrent neural networks in capturing long-term dependencies in sequences, demonstrating superior performance in a variety of areas, including traffic flow, short-term passenger demand, power converter defects, and wind speed forecasting.

[0004] Since its introduction in 2017, the Transformer model has rapidly developed in the fields of natural language processing and computer vision, and has subsequently been applied to time series analysis, particularly wind speed forecasting. While traditional Transformer models have achieved some success in wind speed forecasting, they still face challenges due to their high computational load, large memory requirements, and low efficiency when processing long sequences.

[0005] Furthermore, methods that use the estimated probability distribution function (PDF) of wind speed as prior knowledge or employ decomposition algorithms to divide wind speed series into subsequences of varying frequencies for prediction still face the problem of prohibitively high time costs in practical applications. When predicting a newly generated wind speed series, it often needs to be spliced ​​onto the tail of a historical wind speed series, resulting in an extremely long series for subsequent decomposition and prediction. This approach is clearly unacceptable. Therefore, capturing long-term dependencies and improving prediction accuracy have become key challenges in medium- and long-term wind speed forecasting. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by this invention is that, in response to the high time cost associated with traditional decomposition techniques as preprocessing, this patent proposes and utilizes, for the first time, an improved MOEA model based on an Autoformer for medium- and long-term univariate wind speed forecasting. By incorporating decomposition technology as a submodule of the Autoformer model, this not only addresses the high time cost associated with traditional decomposition techniques, but also retains the advantages of decomposition technology in wind speed series analysis.

[0008] To further mitigate the impact of wind speed series distribution shift on forecast results, this patent improves the decomposition module of Autoformer into a Mixed Expert Decomposition Module (MOEDecomp). This module extracts complex trend terms from the wind speed series using an adaptive window sliding average technique, mitigating the impact of distribution shift on wind speed forecast results.

[0009] To solve the above technical problems, the present invention provides the following technical solution: a medium- and long-term wind speed prediction method based on hybrid expert decomposition and adaptive sliding average, which includes the following steps:

[0010] The original wind speed sequence is preprocessed by filling missing values ​​with mean and wavelet soft threshold to obtain a smooth wind speed sequence;

[0011] The wind speed sequence is input into the MOEA model respectively. The MOEA model includes an encoder and a decoder. The encoder and decoder are composed of a multi-head autocorrelation mechanism module, a mixture of experts decomposition module and a feedforward neural network module cross-connected.

[0012] The processed wind speed sequence is input into the hybrid expert decomposition module of the MOEA model, and the wind speed sequence is decomposed by the module in the encoder to extract the periodic term and trend term;

[0013] The decoder consists of an upper branch and a lower branch. The upper branch of the decoder performs multiple decompositions using the multi-head autocorrelation mechanism and the mixed expert decomposition module. The trend term and the periodic term after each decomposition are continuously transferred to the upper branch, and the trend term after each decomposition is transferred to the lower branch to ensure that the trend term and the periodic term in the intermediate process are uncorrelated.

[0014] In the decoder, the first trend term is superimposed with the trend term after each decomposition and added to the last predicted period term to obtain the final wind speed prediction result.

[0015] As a preferred solution of the medium- and long-term wind speed prediction method of hybrid expert decomposition and adaptive sliding average described in the present invention, wherein: the wavelet soft threshold includes,

[0016] The original wind speed sequence is decomposed using wavelet transform, and the continuous wavelet transform is replaced by discrete wavelet transform. Assuming that the pure signal of the wind speed sequence is f(t), and the sample signal of the wind speed sequence containing noise is x(t), the basic structure is x(t) = f(t) + σn(t), where t is the time point and n(t) is the signal that follows the normal distribution N(0, σ 2 ) of white Gaussian noise, σ represents the noise intensity;

[0017] The wavelet function of discrete wavelet transform is performed using binary wavelet, let a=2 m ,b=n2 m ,m,n∈Z, where m represents the scale and n represents the parameter related to time translation;

[0018] Putting a and b into the waveform sequence obtained by scaling and time-shifting the mother wavelet, we get the waveform sequence equation of the binary wavelet:

[0019]

[0020] Among them, ψ m,n (t) is the wavelet sequence, ψ(·) represents the mother wavelet function;

[0021] By incorporating a and b into the continuous wavelet transform definition of the mother wavelet ψ, the corresponding DWT is obtained,

[0022]

[0023] Among them, W x represents the wavelet transform coefficients, represents the conjugate of ψ(·);

[0024] The multi-level DWT decomposition is performed by Mallat algorithm, and the corresponding scaling coefficients and wavelet coefficients are obtained after decomposition;

[0025] The scaling coefficients contain low-frequency information, and the wavelet coefficients carry high-frequency information.

[0026] As a preferred solution of the medium- and long-term wind speed prediction method of hybrid expert decomposition and adaptive sliding average described in the present invention, the wavelet soft threshold also includes calculating the threshold λ for distinguishing noise signals from valid signals. According to different threshold functions, WTD is divided into wavelet hard threshold denoising and wavelet soft threshold denoising:

[0027]

[0028] Among them, W x is the wavelet transform coefficient, λ represents the threshold, sign(·) represents the sign function, and |W x | is the evaluated wavelet transform coefficient, WTD is wavelet threshold denoising, WSTD is wavelet soft threshold denoising, and WHTD is wavelet hard threshold denoising;

[0029] The noise wavelet transform coefficients are set to zero and the wind speed time series is reconstructed.

[0030] As a preferred solution of the medium- and long-term wind speed forecasting method of hybrid expert decomposition and adaptive sliding average described in the present invention, the hybrid expert decomposition module includes improving the one-dimensional sliding average decomposition algorithm with a fixed window in the Autoformer model into a MOEDecomp algorithm based on a multi-scale sliding average filter;

[0031] Various trend components are extracted from the wind speed sequence and combined into comprehensive trend terms through different weights. The comprehensive trend terms are removed from the original wind speed sequence to obtain the periodic terms:

[0032] X trend =SoftMax(L(x)*(F(x)))

[0033] X seasonal =xX trend

[0034] Where x represents the input wind speed sequence, L(x) represents the linear transformation function Linear(x), F(·) represents a one-dimensional sliding average filter with different window sizes, and SoftMax(L(x)) is the weight matrix of the integrated trend component. The final comprehensive trend term X is obtained through the weight matrix. trend ;X seasonal represents the periodic term extracted from the original wind speed series.

[0035] As a preferred solution of the medium- and long-term wind speed forecasting method of hybrid expert decomposition and adaptive sliding average described in the present invention, wherein: the multi-head autocorrelation mechanism module includes, the MOEA model includes, through the autocorrelation mechanism, finding periodic dependence and time delay aggregation;

[0036] The periodic dependency is found to estimate different periods of the time series. For each period, the time series is divided into multiple subsequences.

[0037] Multiply the aligned subsequences by the corresponding probability correlation coefficients for aggregation;

[0038] Input three query matrices Q, key matrices K, and value matrices of the same dimensions. Regarding the description of the dimensions, L represents the sequence length, C represents the channel dimension, and S represents the length of the aligned subsequences.

[0039] The similarity of subsequences is determined by the autocorrelation coefficient between subsequences. The larger the autocorrelation coefficient, the greater the similarity. According to the random process theory, the discrete time series {χ t The autocorrelation coefficient R tt (τ) is calculated as follows:

[0040]

[0041] Among them, χ t and χ t-τ Represents the sequence { t} and its τ-delayed sequence {χ t-τ}, based on the Wiener-Khinchin theory, perform Fourier transform on matrices K and Q, and concatenate the transformed matrices into S xx (f);

[0042]

[0043] S χχ (f) Inverse Fourier transform produces multiple autocorrelation coefficients:

[0044]

[0045] Among them, τ∈{1,···,L}, F represents the fast Fourier transform and the inverse operation F- 1 , * represents the conjugate operation, represents the complex conjugate of α, S χχ (f) is the frequency domain representation.

[0046] As a preferred solution of the medium- and long-term wind speed forecasting method of hybrid expert decomposition and adaptive sliding average described in the present invention, the time delay aggregation includes aggregating the corresponding subsequence information according to the autocorrelation coefficient calculated in the search period dependency, that is, selecting the top k time delays τ1,···,τ with the largest correlation coefficient k Implement sequence-level connections;

[0047] Use the SoftMax function to perform probabilistic processing on the autocorrelation coefficients calculated by finding periodic dependence to obtain the probability of each coefficient

[0048] According to the period corresponding to the first k autocorrelation coefficients, the wind speed sequence in the V matrix is ​​rolled and translated, that is, the sequence is aligned according to the possible period, which is the ROLL operation;

[0049] Will Multiply with the corresponding subsequence and aggregate all subsequences to form a sequence.

[0050] As a preferred solution of the medium- and long-term wind speed forecasting method of the hybrid expert decomposition and adaptive sliding average described in the present invention, wherein: the multi-head autocorrelation mechanism module also includes an input dimension of d model The query matrix, key matrix and value matrix are divided into d model / h Dimension Q i ,K i ,V i ,i∈{1,2,3,···,h};

[0051] Q i , K i 、V i Through different linear transformations, it is mapped to the corresponding subspace and Q is calculated. i and K i The autocorrelation coefficient of the first k highest correlation coefficients is selected to select the time delay τ j ,j∈{1,2,3,···,k};

[0052] According to the time delay τ j ,j∈{1,2,3,···,k} for V i Aggregation is performed by merging the output results of h heads and performing a linear transformation:

[0053] MultiHead(Q,K,V)=W output *Concat(head1,head2,...,head h )

[0054] Among them, head i=Auto-Correlation(Q i ,K i ,V i ), W output Represents trainable weights, Concat(·) represents the concatenation operation, and Auto-Correlation(·) represents the autocorrelation process.

[0055] Another object of the present invention is to provide a medium- and long-term wind speed prediction system based on hybrid expert decomposition and adaptive sliding average.

[0056] To solve the above technical problems, the present invention provides the following technical solutions: a medium- and long-term wind speed prediction system based on hybrid expert decomposition and adaptive sliding average, comprising: a preprocessing module and a MOEA model;

[0057] The preprocessing module performs mean missing value filling and wavelet soft threshold preprocessing operations on the original wind speed sequence to obtain a smooth wind speed sequence;

[0058] The MOEA model, MOEA encoder and MOEA decoder, MOEA encoder and MOEA decoder respectively include a multi-head autocorrelation mechanism module, a mixed expert decomposition module, and a feedforward neural network module;

[0059] The wind speed sequence is input into the MOEA model, which includes an encoder and a decoder. The processed wind speed sequence is input into the hybrid expert decomposition module of the MOEA model. In the encoder, the wind speed sequence is decomposed through the decomposition module to extract the periodic term and trend term for the first time, and the first periodic term and the first trend term are obtained.

[0060] The first branch of the decoder performs multiple decompositions using the multi-head autocorrelation mechanism module and the mixed expert decomposition module. The trend term and the periodic term after each decomposition are continuously transferred to the first branch, and the trend term after each decomposition is transferred to the second branch to ensure that the trend term and the periodic term in the intermediate process are uncorrelated.

[0061] In the decoder, the first trend term is superimposed with the trend term after each decomposition and added to the last predicted period term to obtain the final wind speed prediction result.

[0062] The present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the medium- and long-term wind speed prediction method based on hybrid expert decomposition and adaptive sliding average are implemented.

[0063] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the medium- and long-term wind speed prediction method based on hybrid expert decomposition and adaptive sliding average are implemented.

[0064] The beneficial effects of the present invention are as follows: the present invention is used for medium- and long-term single-variable wind speed forecasting. By improving the original fixed-window sliding average decomposition module into a hybrid expert decomposition module, the adaptive window sliding average decomposition technology is used to alleviate the influence of distribution offset in the wind speed sequence on the wind speed forecast results.

[0065] At the same time, the present invention uses the hybrid expert decomposition module as an internal sub-model of the prediction model instead of a preprocessing technology, and nests it with the autocorrelation mechanism module to form a hierarchical prediction framework. This not only retains the decomposition technology's ability to extract features of wind speed sequences, but also avoids the problem of excessively high time costs in practical applications.

[0066] Good performance is demonstrated on three public real-world datasets, indicating that the proposed method can effectively handle the task of medium- and long-term univariate wind speed forecasting and has certain generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0068] Figure 1 The overall architecture of the MOEA model of the medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average is provided in one embodiment of the present invention.

[0069] Figure 2 This is the time aggregation process of a medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average provided by an embodiment of the present invention.

[0070] Figure 3 A multi-head autocorrelation mechanism of a medium- to long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average is provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0071] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0072] Example 1, reference Figures 1 to 3 , which is the first embodiment of the present invention, provides a medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average, including:

[0073] like Figure 1 As shown, S1: the original wind speed sequence is preprocessed by filling missing values ​​with mean and performing wavelet soft thresholding to obtain a smoothed wind speed sequence.

[0074] Wind power data often contains noise due to sensor failures, data transmission errors, and other factors, which severely impacts the accuracy of wind speed prediction models. To effectively mitigate this noise, wavelet soft thresholding is used to reduce noise, preserving useful information and preventing loss. The mother wavelet can be scaled and time-shifted using a scale factor a and a time-shift factor b, generating a series of daughter wavelets to extract target features at varying resolutions.

[0075] S101: The original wind speed series is decomposed using a wavelet transform, replacing the continuous wavelet transform with a discrete wavelet transform. The CWT involves continuously varying a and b. However, in practical applications, the CWT is computationally complex and difficult to implement. Therefore, the DWT is an alternative. Compared to the CWT, the DWT discretizes the scale and transformation factors, reducing the computational burden with minimal signal loss.

[0076] In order to simulate the denoising of the wind speed sequence, it is assumed that the pure signal of the wind speed sequence is f(t), and the sample signal of the wind speed sequence containing noise is x(t). The basic structure is x(t) = f(t) + σn(t), where t is the time point and n(t) is the signal that follows the normal distribution N(0, σ 2 ) of white Gaussian noise, σ represents the noise intensity;

[0077] The wavelet function of discrete wavelet transform is performed using binary wavelet, let a=2 m ,b=n2 m ,m,n∈Z, where m represents the scale and n represents the parameter related to time translation;

[0078] Putting a and b into the waveform sequence obtained by scaling and time-shifting the mother wavelet, we get the waveform sequence equation of the binary wavelet:

[0079]

[0080] Among them, ψ m,n (t) is the wavelet sequence, ψ(·) represents the mother wavelet function;

[0081] In order to decompose the sample signal using discrete wavelet transform (DWT), the corresponding DWT is obtained by incorporating a and b into the continuous wavelet transform definition of the mother wavelet ψ.

[0082]

[0083] Among them, W x Represents the wavelet transform coefficient, "<·>" is a mathematical symbol in the wavelet sequence definition formula, indicating that this formula involves two terms within the symbol; represents the conjugate of ψ(·).

[0084] Mallat algorithm is often used in discrete wavelet transform, and scaling functions and wavelet equations are often used as low-pass and high-pass filters.

[0085] The multi-level DWT decomposition is performed using the Mallat algorithm, and the corresponding scaling coefficients and wavelet coefficients are obtained after decomposition.

[0086] The scaling coefficients contain low-frequency information, and the wavelet coefficients carry high-frequency information:

[0087]

[0088] Among them, c m,n represents the scale coefficient on the decomposed m scale, d m,n is the wavelet coefficient on the m scale after decomposition, ψ(·) represents the value when x(t) is projected onto V m The mother wavelet function in space is When x(t) is projected onto W m The mother wavelet function in space time.

[0089] S102: Calculate the threshold λ for distinguishing noise signals from valid signals. Presetting the threshold is essential in wavelet threshold denoising (WTD), which will determine the effect of the denoising process.

[0090] According to the different threshold functions, WTD is divided into wavelet hard threshold denoising and wavelet soft threshold denoising:

[0091]

[0092]

[0093] Among them, W x is the wavelet transform coefficient, λ represents the threshold, sign(·) represents the sign function, and |Wx | is the evaluated wavelet transform coefficient, WTD is wavelet threshold denoising, WSTD is wavelet soft threshold denoising, and WHTD is wavelet hard threshold denoising;

[0094] WHTD suffers from an unavoidable problem: the discontinuity of the hard threshold function. This discontinuous function can trigger oscillations during the wavelet reconstruction process, ultimately reducing the final denoising effect. In contrast, the estimated wavelet coefficients obtained from the soft threshold function of WSTD maintain good continuity. In other words, the soft threshold function overcomes the shortcomings of the hard threshold and is more suitable for the data preprocessing of this patent.

[0095] S103: Set the noise wavelet transform coefficients to zero and reconstruct the wind speed time series.

[0096] S2: The wind speed sequence is input into the MOEA model respectively. The MOEA model includes an encoder and a decoder. The encoder and decoder are composed of a multi-head autocorrelation mechanism module, a mixture of experts decomposition module and a feedforward neural network module cross-connected.

[0097] S3: The processed wind speed sequence is input into the hybrid expert decomposition module of the MOEA model, and the wind speed sequence is decomposed through the module in the encoder to extract the periodic term and trend term.

[0098] Specifically, the hybrid expert decomposition module improves the one-dimensional sliding average decomposition algorithm with a fixed window in the Autoformer model into a MOEDecomp algorithm based on a multi-scale sliding average filter;

[0099] Various trend components are extracted from the wind speed sequence and combined into a comprehensive trend term (also called a seasonal term) through different weights. The comprehensive trend term is removed from the original wind speed sequence to obtain the periodic term:

[0100] X trend =SoftMax(L(x)*(F(x)))

[0101] X seasonal =xX trend

[0102] Where x represents the input wind speed sequence, L(x) represents the linear transformation function Linear(x), F(·) represents a one-dimensional sliding average filter with different window sizes, and SoftMax(L(x)) is the weight matrix of the integrated trend component. The final comprehensive trend term X is obtained through the weight matrix. trend ;X seasonal represents the periodic term extracted from the original wind speed series.

[0103] like Figure 1 The content shown,

[0104] S4: The decoder consists of an upper branch and a lower branch. The upper branch of the decoder performs multiple decompositions using the multi-head autocorrelation mechanism and the mixed expert decomposition module. The trend term of each decomposition and the periodic term after each decomposition continue to be transmitted to the upper branch, and the trend term after each decomposition is transmitted to the lower branch to ensure that the trend term and the periodic term in the intermediate process are uncorrelated.

[0105] Furthermore, for time series such as wind speed, similar phases in different periods exhibit similar characteristics. Therefore, MOEA designs an autocorrelation mechanism, which is called Auto-Correction.

[0106] The MOEA model includes,finding periodic dependence and time-delayed aggregation through the,autocorrelation mechanism;

[0107] The periodic dependency is found to estimate different periods of the time series. For each period, the time series is divided into multiple subsequences.

[0108] Multiply the aligned subsequences by the corresponding probability correlation coefficients for aggregation;

[0109] Input three query matrices Q, key matrices K, and value matrices of the same dimensions. Regarding the description of the dimensions, L represents the sequence length, C represents the channel dimension, and S represents the length of the aligned subsequences.

[0110] The similarity of subsequences is determined by the autocorrelation coefficient between subsequences. The larger the autocorrelation coefficient, the greater the similarity. According to the random process theory, the discrete time series {χ t The autocorrelation coefficient R tt (τ) is calculated as follows:

[0111]

[0112] Among them, χ t and χ t-τ Represents the sequence { t} and its τ-delayed sequence {χ t-τ}, based on the Wiener-Khinchin theory, perform Fourier transform on matrices K and Q, and concatenate the transformed matrices into S xx (f);

[0113]

[0114] S χχ (f) Inverse Fourier transform produces multiple autocorrelation coefficients:

[0115]

[0116] Where τ∈{1,···,L}, F represents the fast Fourier transform and the inverse operation F-1, * represents the conjugate operation, represents the complex conjugate of α, S χχ (f) is the frequency domain representation;

[0117] These autocorrelation coefficients reflect the multiple possible periods of the sequences in the Q and K matrices. Since they are the same as the two K matrices, the correlation between them is the autocorrelation of the sequence in the Q matrix.

[0118] Furthermore, based on the obtained multiple autocorrelation coefficients, the corresponding subsequence information is aggregated according to the autocorrelation coefficients calculated in the search for periodic dependence, that is, the autocorrelation coefficients are arranged in descending order, and the top k time delays τ1,···,τ with the largest correlation coefficients are selected. k Implement sequence-level connections;

[0119] Use the SoftMax function to perform probabilistic processing on the autocorrelation coefficients calculated by finding periodic dependence to obtain the probability of each coefficient

[0120] According to the period corresponding to the first k autocorrelation coefficients, the wind speed sequence in the V matrix is ​​rolled and translated, that is, the sequence is aligned according to the possible period, which is the ROLL operation;

[0121] Will Multiply with the corresponding subsequence and aggregate all subsequences to form a sequence.

[0122] It should also be noted that the multi-head mechanism improves the model's feature extraction capabilities by capturing data representations in different representation spaces, enabling each head to focus on the data's characteristics from a specific perspective. The multi-head autocorrelation mechanism (MultiHead) enhances the model's robustness to noise and anomalies and accelerates training by simultaneously learning different data representations. The flexibility of this mechanism enables it to adapt to complex real-world application scenarios.

[0123] Specifically, the input dimension is d model The query matrix, key matrix and value matrix are divided into d model / h Dimension Q i ,K i ,V i ,i∈{1,2,3,···,h};

[0124] Q i , K i 、V i Through different linear transformations, it is mapped to the corresponding subspace and Q is calculated. i and K i The autocorrelation coefficient of the first k highest correlation coefficients is selected to select the time delay τj ,j∈{1,2,3,···,k};

[0125] According to the time delay τ j ,j∈{1,2,3,···,k} for V i Aggregation is performed by merging the output results of h heads and performing a linear transformation:

[0126] MultiHead(Q,K,V)=W output *Concat(head1,head2,…,head h )

[0127] Among them, head i =Auto-Correlation(Q i ,K i ,V i ), W output Represents trainable weights, Concat(·) represents the concatenation operation, and Auto-Correlation(·) represents the autocorrelation process.

[0128] S4: In the decoder, the trend term is superimposed with the trend term after each decomposition, and added to the last predicted period term to obtain the final wind speed prediction result.

[0129] Example 2 is the second embodiment of the present invention, which provides a medium- and long-term wind speed prediction method based on hybrid expert decomposition and adaptive sliding average. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0130] This experiment uses two indicators to evaluate the proposed model: Mean Square Error (MSE) and Mean Absolute Error (MAE).

[0131] MSE and MAE are used to measure the difference between the predicted value and the true value. The smaller the value, the higher the prediction accuracy and the closer the predicted value is to the true value. The expressions of MSE and MAE are as follows:

[0132]

[0133]

[0134] Among them, y i(pred) ,i∈{1,2,3,……,N} and y i(true) ,i∈{1,2,3,……,N} represents the true value and predicted value of wind speed respectively, and N represents the number of data points corresponding to a specific prediction step.

[0135] This paper uses three datasets for model validation: SPO, Sujawal and SDWPF.

[0136] The SPO, Sujawal, and SDWPF datasets were obtained from the National Oceanic and Atmospheric Administration (NOAA) Global Monitoring Laboratory (GML), the Energy Sector Management Assistance Program (ESMAP), and the 2022 Baidu KDD Cup Competition, respectively.

[0137] In this experiment, the last month of each dataset was used as the testing set, the second-to-last month’s data was used as the validation set, and the rest were used as the training set.

[0138] The main data preprocessing process includes missing value filling, normalization, etc. For missing values, this study uses the mean filling algorithm (MFA) for supplementary processing, and then normalizes the wind speed series using the Z score method.

[0139] After preprocessing and dividing the three data sets, the mainstream simple neural network prediction models, feedforward neural network (FFNN), residual feedforward neural network (R-FFNN), interval probability distribution learning model (IPDL) and the model MOEA proposed in this paper were used for learning, and finally the model performance was verified on the test set.

[0140] It should be noted that the input of this experiment is a 96-hour wind speed sequence. Four different experiments were conducted to predict the wind speed values ​​for the next 24 hours, 48 ​​hours, 72 hours, 96 hours, and 120 hours, respectively, to provide a more comprehensive evaluation of the model. The verification results are shown in Table 1:

[0141] Table 1 Prediction results of different models on three datasets

[0142]

[0143] Example 3 is the third embodiment of the present invention, which differs from the first two embodiments in that:

[0144] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0145] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0146] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0147] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0148] Example 4 is the fourth embodiment of the present invention, which provides a medium- and long-term wind speed forecasting system based on hybrid expert decomposition and adaptive sliding average, including: a preprocessing module and a MOEA model;

[0149] The preprocessing module performs mean missing value filling and wavelet soft threshold preprocessing operations on the original wind speed sequence to obtain a smooth wind speed sequence;

[0150] The MOEA model, MOEA encoder and MOEA decoder, MOEA encoder and MOEA decoder respectively include a multi-head autocorrelation mechanism module, a mixed expert decomposition module, and a feedforward neural network module;

[0151] The processed wind speed sequence is input into the hybrid expert decomposition module of the MOEA model, and the wind speed sequence is decomposed by the module in the encoder to extract the periodic term and trend term;

[0152] The decoder consists of an upper branch and a lower branch. The upper branch of the decoder performs multiple decompositions using the multi-head autocorrelation mechanism and the mixed expert decomposition module. The trend term and the periodic term after each decomposition are continuously transferred to the upper branch, and the trend term after each decomposition is transferred to the lower branch to ensure that the trend term and the periodic term in the intermediate process are uncorrelated.

[0153] In the decoder, the trend term is superimposed with the trend term after each decomposition and added to the last predicted period term to obtain the final wind speed prediction result.

[0154] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average is characterized by: include, The original wind speed sequence is preprocessed by filling missing values ​​with mean and wavelet soft threshold to obtain a smooth wind speed sequence; The wind speed sequence is input into the MOEA model respectively. The MOEA model includes an encoder and a decoder. The encoder and decoder are composed of a multi-head autocorrelation mechanism module, a mixture of experts decomposition module and a feedforward neural network module cross-connected. The processed wind speed sequence is input into the hybrid expert decomposition module of the MOEA model, and the wind speed sequence is decomposed by the module in the encoder to extract the periodic term and trend term; The decoder consists of an upper branch and a lower branch. The upper branch of the decoder performs multiple decompositions using the multi-head autocorrelation mechanism and the mixed expert decomposition module. The trend term and the periodic term after each decomposition are continuously transferred to the upper branch, and the trend term after each decomposition is transferred to the lower branch to ensure that the trend term and the periodic term in the intermediate process are uncorrelated. In the decoder, the trend term is superimposed with the trend term after each decomposition and added to the last predicted period term to obtain the final wind speed prediction result.

2. The medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average according to claim 1 is characterized in that: The wavelet soft threshold includes: The original wind speed sequence is decomposed using wavelet transform, and the continuous wavelet transform is replaced by discrete wavelet transform. Assuming that the pure signal of the wind speed sequence is f(t), and the sample signal of the wind speed sequence containing noise is x(t), the basic structure is x(t) = f(t) + σn(t), where t is the time point and n(t) is the signal that follows the normal distribution N(0, σ 2 ) of white Gaussian noise, σ represents the noise intensity; The wavelet function of discrete wavelet transform is performed using binary wavelet, let a=2 m ,b=n2 m ,m,n∈Z, where m represents the scale and n represents the parameter related to time translation; Putting a and b into the waveform sequence obtained by scaling and time-shifting the mother wavelet, we get the waveform sequence equation of the binary wavelet: Among them, ψ m,n (t) is the wavelet sequence, ψ(·) represents the mother wavelet function; By incorporating a and b into the continuous wavelet transform definition of the mother wavelet ψ, the corresponding DWT is obtained, Among them, W x represents the wavelet transform coefficients, represents the conjugate of ψ(·); The multi-level DWT decomposition is performed by Mallat algorithm, and the corresponding scaling coefficients and wavelet coefficients are obtained after decomposition; The scaling coefficients contain low-frequency information, and the wavelet coefficients carry high-frequency information.

3. The medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average according to claim 2 is characterized in that: The wavelet soft threshold also includes calculating the threshold λ for distinguishing noise signals from valid signals. According to the different threshold functions, WTD is divided into wavelet hard threshold denoising and wavelet soft threshold denoising: Among them, W x is the wavelet transform coefficient, λ represents the threshold, sign(·) represents the sign function, and |W x | is the evaluated wavelet transform coefficient, WTD is wavelet threshold denoising, WSTD is wavelet soft threshold denoising, and WHTD is wavelet hard threshold denoising; The noise wavelet transform coefficients are set to zero and the wind speed time series is reconstructed.

4. The medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average according to claim 3 is characterized by: The hybrid expert decomposition module includes improving the one-dimensional sliding average decomposition algorithm with a fixed window in the Autoformer model into a MOEDecomp algorithm based on a multi-scale sliding average filter; Various trend components are extracted from the wind speed sequence and combined into comprehensive trend terms through different weights. The comprehensive trend terms are removed from the original wind speed sequence to obtain the periodic terms: X trend =SoftMax(L(x)*(F(x))) X seasonal =x-X trend Where x represents the input wind speed sequence, L(x) represents the linear transformation function Linear(x), F(·) represents a one-dimensional sliding average filter with different window sizes, and SoftMax(L(x)) is the weight matrix of the integrated trend component. The final comprehensive trend term X is obtained through the weight matrix. trend ;X seasonal represents the periodic term extracted from the original wind speed series.

5. The medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average according to claim 4 is characterized in that: The multi-head autocorrelation mechanism module includes, the MOEA model includes, finding periodic dependence and time delay aggregation through the autocorrelation mechanism; The periodic dependency is found to estimate different periods of the time series. For each period, the time series is divided into multiple subsequences. Multiply the aligned subsequences by the corresponding probability correlation coefficients for aggregation; Input three query matrices Q, key matrices K, and value matrices of the same dimensions. Regarding the description of the dimensions, L represents the sequence length, c represents the channel dimension, and s represents the length of the aligned subsequences. The similarity of subsequences is determined by the autocorrelation coefficient between subsequences. The larger the autocorrelation coefficient, the greater the similarity. According to the random process theory, the discrete time series {χ t The autocorrelation coefficient R tt (τ) is calculated as follows: Among them, χ t and χ t-τ Represents the sequence { t } and its τ-delayed sequence {χ t-τ }, based on the Wiener-Khinchin theory, perform Fourier transform on matrices K and Q, and concatenate the transformed matrices into S xx (f); S χχ (f) Inverse Fourier transform produces multiple autocorrelation coefficients: Among them, τ∈{1,···,L}, F represents the fast Fourier transform and the inverse operation F- 1 , * represents the conjugate operation, represents the complex conjugate of α, S χχ (f) is the frequency domain representation.

6. The medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average according to claim 4 is characterized in that: The time delay aggregation includes aggregating the corresponding subsequence information according to the autocorrelation coefficient calculated in the search cycle dependency, that is, selecting the first k time delays τ1,···,τ with the largest correlation coefficients. k Implement sequence-level connections; Use the SoftMax function to perform probabilistic processing on the autocorrelation coefficients calculated by finding periodic dependence to obtain the probability of each coefficient According to the period corresponding to the first k autocorrelation coefficients, the wind speed sequence in the V matrix is ​​rolled and translated, that is, the sequence is aligned according to the possible period, that is, the ROLL operation; Multiply with the corresponding subsequence and aggregate all subsequences to form a sequence.

7. The medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average according to claim 4 is characterized in that: The multi-head autocorrelation mechanism module also includes an input dimension of d model The query matrix, key matrix and value matrix are divided into d model / h Dimension Q i ,K i ,V i ,i∈{1,2,3,···,h}; Q i , K i 、V i Through different linear transformations, it is mapped to the corresponding subspace and Q is calculated. i and K i The autocorrelation coefficient of the first k highest correlation coefficients is selected to select the time delay τ j ,j∈{1,2,3,···,k}; According to the time delay τ j ,j∈{1,2,3,···,k} for V i Aggregation is performed by merging the output results of h heads and performing a linear transformation: MultiHead(Q,K,V)=W output *Concat(head1,head2,...,head h ) Among them, head i =Auto-Correlation(Q i ,K i ,V i ), W output Represents trainable weights, Concat(·) represents the concatenation operation, and Auto-Correlation(·) represents the autocorrelation process.

8. A medium- and long-term wind speed forecasting system based on hybrid expert decomposition and adaptive sliding average, which uses the medium- and long-term wind speed forecasting method based on hybrid expert decomposition and adaptive sliding average according to any one of claims 1 to 7, characterized in that: include: Preprocessing module and MOEA model; The preprocessing module performs mean missing value filling and wavelet soft threshold preprocessing operations on the original wind speed sequence to obtain a smooth wind speed sequence; The MOEA model, MOEA encoder and MOEA decoder, MOEA encoder and MOEA decoder respectively include a multi-head autocorrelation mechanism module, a mixed expert decomposition module, and a feedforward neural network module; The processed wind speed sequence is input into the hybrid expert decomposition module of the MOEA model, and the wind speed sequence is decomposed by the module in the encoder to extract the periodic term and trend term; The decoder consists of an upper branch and a lower branch. The upper branch of the decoder performs multiple decompositions using the multi-head autocorrelation mechanism and the mixed expert decomposition module. The trend term and the periodic term after each decomposition are continuously transferred to the upper branch, and the trend term after each decomposition is transferred to the lower branch to ensure that the trend term and the periodic term in the intermediate process are uncorrelated. In the decoder, the trend term is superimposed with the trend term after each decomposition and added to the last predicted period term to obtain the final wind speed prediction result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the medium- and long-term wind speed prediction method based on hybrid expert decomposition and adaptive sliding average are implemented as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the medium- and long-term wind speed prediction method based on hybrid expert decomposition and adaptive sliding average are implemented as described in any one of claims 1 to 7.

Images