Offshore wind power point and interval prediction method and device based on ICEEMDAN-VMD-LSTM-KDE

Abstract

The invention discloses an offshore wind power point and interval prediction method and device based on ICEEMDAN-VMD-LSTM-KDE, and the method comprises the following steps: S1, carrying out the deep analysis of offshore wind power data through employing a two-layer decomposition algorithm, so as to extract key data features; s2, predicting each component obtained by decomposition by using an LSTM algorithm to obtain a respective prediction result, and obtaining an ultimate point prediction result through calculation; and S3, based on a point prediction result, measuring probability density distribution of prediction errors by using a KDE technology, and determining a prediction interval of the offshore wind power according to the probability density distribution. Compared with the prior art, the method has the advantages that the accurate offshore wind power point prediction value can be provided, the distribution range of the point prediction value can be provided, and the like.

Description

Technical Field

[0001] The present invention relates to the field of offshore wind power prediction methods, and in particular to an offshore wind power point and interval prediction method and device based on ICEEMDAN-VMD-LSTM-KDE. Background Art

[0002] In order to achieve the goal of low-carbon development, my country is vigorously developing wind power. Among them, offshore wind power has been highly valued and vigorously developed by many countries due to its advantages such as long power generation time and proximity to traditional power load centers. By the end of 2023, my country's total installed capacity of offshore wind power will be about 37.29GW, ranking first in the world. The rapid development of offshore wind farms has brought new opportunities and challenges to the power system. Accurate and reliable offshore wind power forecasting is essential for optimizing grid dispatching and improving the flexibility and economy of the power system. However, offshore wind power is affected by complex atmospheric boundary layer processes, marine environment and intermittent wind resources, showing strong randomness and uncertainty, which brings great challenges to accurate prediction.

[0003] In recent years, domestic and foreign experts and scholars have conducted extensive and in-depth research on offshore wind power forecasting. They mainly use physical models, machine learning models and other methods to make point forecasts of offshore wind power based on historical data, and rarely consider the impact and value of interval forecasts on actual decision-making. Given the increasing demand for offshore wind power forecast accuracy for actual decision-making purposes, providing accurate point forecasts and corresponding interval estimates has greater practical significance.

[0004] However, there are still two new problems in the point and interval prediction of offshore wind power:

[0005] 1) How to effectively conduct in-depth analysis of offshore wind power data and extract key data features to improve the accuracy of point prediction of offshore wind power;

[0006] 2) How to establish a simple and feasible interval prediction technology based on accurate point prediction results to achieve interval prediction of offshore wind power. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an offshore wind power point and interval prediction method and device based on ICEEMDAN-VMD-LSTM-KDE, so as to effectively perform in-depth analysis on offshore wind power data and extract key data features to improve the point prediction accuracy of offshore wind power. At the same time, a simple and feasible interval prediction technology is established in combination with the KDE algorithm to realize offshore wind power interval prediction.

[0008] A technical solution to achieve the above purpose is: a method for offshore wind power point and interval prediction based on ICEEMDAN-VMD-LSTM-KDE, comprising the following steps:

[0009] S1, based on the deep analysis of offshore wind power data using a two-layer decomposition algorithm to extract key data features;

[0010] S2, using LSTM algorithm to predict each component obtained by decomposition, obtain their respective prediction results, and obtain the ultimate point prediction result through calculation;

[0011] S3, based on the point prediction results, uses KDE technology to measure the probability density distribution of the prediction error, and determines the prediction interval of offshore wind power accordingly.

[0012] Furthermore, in S1, a two-layer decomposition algorithm is used to perform in-depth analysis on offshore wind power data to extract key data features, which specifically includes the following steps:

[0013] S1.1, ICEEMDAN is used to decompose the original offshore wind speed time series into a finite number of intrinsic mode functions, thereby capturing the dynamic changes inherent in the time series;

[0014] S1.2, for the difficult-to-predict high-frequency intrinsic mode function IMF1, the VMD algorithm is used to further decompose it to obtain sub-components.

[0015] Furthermore, the specific calculation steps of S1.1 are as follows:

[0016] (1) Add I groups of noise w to the time series signal x to be decomposed (i) , construct a sequence x of noisy mixed signals (i) (i=1,2,...,k), the formula is as follows:

[0017] x (i) =x+β o E1(w (i) )

[0018] Where: β o is the weight coefficient of the noise; E1(·) represents the first modal component after EMD decomposition of the noise;

[0019] (4) Calculate the noisy mixed signal x (i) The local mean of , and the first set of residuals r1 and the first modal component IMF1 are calculated using local mean decomposition. The formula is as follows:

[0020]

[0021] IMF1=x-r1

[0022] Where: N(·) is the local mean of the noisy mixed signal;

[0023] (5) Add noise E2(w (i) ), use the same method as (2) to obtain the second group of residuals r2 and modal components IMF2;

[0024]

[0025] IMF2=r1-r2

[0026] (4) Similarly, we can get the residual r of the kth group k and modal component IMF k ;

[0027]

[0028] (5) Repeat the above steps until all the IMFs of the modal components in the original sequence are obtained.

[0029] Furthermore, the specific calculation steps of S1.2 are as follows:

[0030] The core of VMD is to construct and optimize the following variational problem:

[0031]

[0032] Where: {u k} is the mode function obtained by decomposition; {ω k} is the corresponding center frequency. for u k Fourier transform of , K is the preset number of modes;

[0033] After constructing the above variational problem, the alternating direction multiplier method is used to solve the above variational problem. By introducing auxiliary variables and Lagrange multipliers, the original problem is transformed into a sub-problem that is easier to solve. At the same time, in each iteration, the modal {u k} and the center frequency {ω k} until convergence.

[0034] Furthermore, in S2, the LSTM algorithm is used to predict each component obtained by decomposition, obtain respective prediction results, and obtain the ultimate point prediction result by calculation, which specifically includes the following steps:

[0035] S2.1, using LSTM to predict the intrinsic mode functions and sub-components obtained from the two decompositions to obtain sub-prediction results;

[0036] S2.2, all sub-prediction results are calculated to obtain the final point prediction result.

[0037] Furthermore, the specific calculation steps of S2.1 are:

[0038] (2) The forget gate removes the old memory unit c by using the sigmoid activation function t-1 After removing some information, the output f of the forget gate t The formula is as follows:

[0039] f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f )

[0040] Where: σ is the activation function of sigmoid; Wx f , W hf , W cf is the weight coefficient matrix of the forget gate; h t-1 is the output value at time t-1; x t is the input value at time t; b f is the network layer bias of the forget gate;

[0041] (2) The input gate uses the sigmoid activation function to control the information at time t, which is calculated as follows:

[0042] i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i )

[0043] Where: W hi , W xi , W ci is the weight coefficient matrix of the input gate; b i is the network layer bias of the input gate;

[0044] (3) Because the forget gate removes some data information at time t-1, and the input gate adds new information at time t, the cell state is updated, and the calculation formula is as follows:

[0045] c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1+b c )

[0046] Where: tanh is the hyperbolic tangent activation function; W xc , W hc is the weight coefficient matrix of the input gate; b c is the network layer bias;

[0047] (4) The output gate also uses the sigmoid activation function to control the output result. The calculation formula is as follows:

[0048] o t =σ(W xo x t +W ho h t-1 +W co c t +b o )

[0049] Where: W ho , W xo , W co is the weight coefficient matrix of the output gate; b o is the network layer bias of the output gate;

[0050] (5) Output result h t By memory unit c t The result after processing by the tanh layer is determined together with the output of the output gate, and the calculation formula is as follows:

[0051] h t =o t tanh(c t ).

[0053] Furthermore, the specific calculation method of S2.2 is to linearly sum all sub-prediction results, and the result obtained is the final point prediction result.

[0054] Furthermore, in S3, the KDE technique is used to measure the probability density distribution of the prediction error, and the specific steps of determining the prediction interval of the offshore wind power are as follows:

[0055] S3.1, obtain the prediction error based on the point prediction result and the actual value;

[0056] S3.2, KDE is used to measure the probability density distribution of the prediction error, and based on this, the prediction interval of offshore wind power at a given confidence level is determined.

[0057] Furthermore, the specific calculation steps of S3.2 are as follows:

[0058] The KDE algorithm places a kernel around each data point and then stacks all these kernels to form a whole density estimate. Mathematically, the kernel density estimate can be expressed as:

[0059]

[0060] Where: is the estimated density at point x; n is the number of samples; x i is the data sample point; h is the bandwidth parameter, which controls the width of the kernel and affects the smoothness of the estimate; K is the kernel function, which is used to control how to get from each data point x i Contribute to the estimated density; when the confidence level is 1-α, if the power value of the offshore wind turbine meets:

[0061] P(G low <G<G up )=1-α

[0062] Then the interval [G low ,G up ] is the prediction interval of power, G low , G up represent the lower and upper limits respectively.

[0063] Further, it is based on the above-mentioned offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE, including an electronic device and a computer-readable storage medium;

[0064] The electronic device comprises a processor, a memory and a communication bus, wherein the memory stores a computer-readable program executable by the processor;

[0065] The communication bus realizes the connection and communication between the processor and the memory;

[0066] The processor is used to execute the computer readable program;

[0067] The computer-readable storage medium stores one or more programs that can be executed by one or more processors.

[0068] Compared with the prior art, the present invention has the following advantages:

[0069] The present invention proposes an innovative hybrid prediction model that combines ICEEMDAN, VMD, LSTM and KDE to perform point and interval predictions of offshore wind power. The model first uses a two-layer decomposition technique to process wind power data, which can effectively extract and separate different components in wind power, such as periodic components and random fluctuations. Then, an LSTM network is used to learn and predict the processed data. LSTM can effectively process and memorize information in long time series, which is particularly important for data such as wind power that has a high degree of time correlation. Finally, in order to evaluate the uncertainty of the prediction and provide more comprehensive prediction information, KDE is introduced to generate a probability distribution of the predicted power. This can not only provide a single prediction value (point prediction), but also provide a distribution range of the prediction value (interval prediction). BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The present invention is a flow chart of the method.

[0071] Figure 2 Decomposition results for ICEEMDAN.

[0072] Figure 3 VMD decomposition result.

[0073] Figure 4 Point prediction results.

[0074] Figure 5 is the PDF of the absolute error of the fitted point predictions for each distribution method.

[0075] Figure 6 is the CDF curve.

[0076] Figure 7 It is the interval prediction result under each distribution. DETAILED DESCRIPTION

[0077] In order to better understand the technical solution of the present invention, the following is a detailed description through specific embodiments:

[0078] The present invention provides an offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE. The specific inventive concept can be divided into two parts: data feature extraction and point and interval prediction using a two-layer decomposition algorithm (ICEEMDAN-VMD).

[0079] First, it is the feature extraction stage. In this stage, ICEEMDAN is used to decompose the original offshore wind speed time series into a limited number of intrinsic mode functions (IMFs) to capture the inherent dynamic changes in the time series. For the difficult-to-predict high-frequency intrinsic mode function IMF1, the VMD algorithm is used to further decompose it to obtain a series of sub-components. Secondly, it is the point prediction and interval prediction stage. In this stage, LSTM is used to predict the intrinsic mode functions and sub-components obtained by the two decompositions to obtain sub-prediction results, and the sub-prediction results are summed to obtain the final point prediction results. The KDE algorithm is used to analyze the point prediction error and generate the offshore wind power prediction interval under the corresponding confidence value.

[0080] The specific steps of the present invention are described as follows:

[0081] S1: Use a two-layer decomposition algorithm (ICEEMDAN-VMD) to extract data features

[0082] ICEEMDAN is used to decompose the original offshore wind speed time series into a limited number of intrinsic mode functions IMF, thereby capturing the inherent dynamic changes in the time series. For the difficult-to-predict high-frequency intrinsic mode function IMF1, the VMD algorithm is used to further decompose it into a series of sub-components. The principles of ICEEMDAN and VMD algorithms are as follows.

[0083] S1.1: ICEEMDAN is used to decompose the original offshore wind speed time series into a finite number of intrinsic mode functions, thereby capturing the dynamic variations inherent in the time series.

[0084] Both ICEEMDAN and CEEMDAN are methods developed based on EMD. The difference between ICEEMDAN and CEEMDAN is that ICEEMDAN does not directly add Gaussian white noise in the decomposition process, but selects the Kth IMF component after the white noise is decomposed by EMD to determine each IMF component after the original sequence is decomposed, thereby realizing the frequency domain division of the original signal and obtaining subsequences containing different high and low frequencies. The specific calculation steps are as follows:

[0085] 1) Add I groups of noise w to the time series signal x to be decomposed (i) , construct a sequence x of noisy mixed signals (i) (i=1,2,...,k), the formula is as follows:

[0086] x (i) =x+β o E1(w (i) )(1)

[0087] Where: βo is the weight coefficient of the noise; E1(·) represents the first modal component after EMD decomposition of the noise.

[0088] 2) Calculate the noisy mixed signal x (i) The local mean of , and the first set of residuals r1 and the first modal component IMF1 are calculated using local mean decomposition. The formula is as follows:

[0089]

[0090] IMF1=x-r1(3)

[0091] Where: N(·) is the local mean of the noisy mixed signal.

[0092] 3) Add noise E2(w (i) ), use the same method as 2) to obtain the second group of residuals r2 and modal components IMF2.

[0093]

[0094] IMF2=r1-r2(5)

[0095] 4) Similarly, the residual r of the kth group can be obtained k and modal component IMF k .

[0096]

[0097] 5) Repeat the above steps until all the modal component IMFs in the original sequence are obtained.

[0098] S1.2, for the difficult-to-predict high-frequency intrinsic mode function IMF1, the VMD algorithm is used to further decompose it to obtain sub-components.

[0099] The basic idea of ​​VMD is to decompose the signal by solving a variational problem. The algorithm tries to find a set of modes so that the bandwidth of each mode is as small as possible, and the original signal can be expressed as a linear combination of these modes. The core of VMD is to construct and optimize the following variational problem:

[0100]

[0101] Where: {u k} is the mode function obtained by decomposition; {ω k} is the corresponding center frequency. for u k Fourier transform of , K is the preset number of modes.

[0102] After constructing the above variational problem, the alternating direction multiplier method is used to solve the above variational problem. This method transforms the original problem into a sub-problem that is easier to solve by introducing auxiliary variables and Lagrange multipliers. At the same time, in each iteration, the modal {u k} and the center frequency {ω k} until convergence.

[0103] S2: Point and interval prediction

[0104] In this stage, LSTM is used to predict the intrinsic mode functions and subcomponents obtained from the two decompositions to obtain sub-prediction results, and the sub-prediction results are summed to obtain the final point prediction results. The KDE algorithm is used to analyze the point prediction error and generate the offshore wind power prediction interval under the corresponding confidence value. The principles of the LSTM and KDE algorithms are shown below.

[0105] S2.1: Use LSTM to predict the intrinsic mode functions and sub-components obtained from the two decompositions to obtain sub-prediction results.

[0106] The long short-term memory (LSTM) neural network is an improvement of the recurrent neural network (RNN). The main difference is that LSTM adds a structure to the algorithm to determine whether the data is valuable. This structure is called a "cell", which can solve the problem of RNN gradient disappearance and explosion. LSTM is suitable for processing problems related to time series, so it can predict wind power series. The principle of the LSTM algorithm is as follows.

[0107] 1) The forget gate removes the old memory unit c by using the sigmoid activation function t-1 After removing some information, the output f of the forget gate t The formula is as follows:

[0108] f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f )(8)

[0109] Where: σ is the activation function of sigmoid; Wx f , W hf , W cf is the weight coefficient matrix of the forget gate; h t-1 is the output value at time t-1; x t is the input value at time t; b f is the network layer bias of the forget gate.

[0110] 2) The input gate controls the information at time t by using the sigmoid activation function, which is calculated as follows:

[0111] i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i )(9)

[0112] Where: W hi , W xi , W ci is the weight coefficient matrix of the input gate; b i is the network layer bias of the input gate.

[0113] 3) Because the forget gate removes some data information at time t-1, the input gate adds new information at time t, and the cell state is updated. The calculation formula is as follows:

[0114] c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c )(10)

[0115] Where: tanh is the hyperbolic tangent activation function; W xc , W hc is the weight coefficient matrix of the input gate; b c is the network layer bias.

[0116] 4) The output gate also uses the sigmoid activation function to control the output result. The calculation formula is as follows:

[0117] o t =σ(W xo x t +W ho h t-1 +W co c t +b o )(11)

[0118] Where: W ho , W xo , W co is the weight coefficient matrix of the output gate; b o is the network layer bias of the output gate.

[0119] 5) Output result ht By memory unit c t The result after processing by the tanh layer is determined together with the output of the output gate, and the calculation formula is as follows:

[0120] h t =o t tanh(c t )(12)

[0121] S2.2: Linearly sum all sub-prediction results to obtain the final point prediction result.

[0122] S3: Based on the point prediction results, the KDE technique is used to measure the probability density distribution of the prediction error, and the prediction interval of offshore wind power is determined accordingly. The specific steps are as follows:

[0123] S3.1, obtain the prediction error based on the point prediction result and the actual value;

[0124] S3.2, KDE is used to measure the probability density distribution of the prediction error, and based on this, the prediction interval of offshore wind power at a given confidence level is determined. The principle of the KDE algorithm is shown below.

[0125] KDE is a nonparametric method for estimating the probability density function (PDF) of a random variable. KDE can be viewed as a smoothing technique that estimates the probability density function of the entire dataset by smoothing independent and identically distributed data samples. The basic idea of ​​kernel density estimation is to place a kernel (or window) around each data point and then stack all these kernels to form an overall density estimate. Mathematically, kernel density estimation can be expressed as:

[0126]

[0127] Where: is the estimated density at point x; n is the number of samples; x i is the data sample point; h is the bandwidth parameter, which controls the width of the kernel and affects the smoothness of the estimate; K is the kernel function, which is used to control how to get from each data point x i Contribute to the estimated density. Common kernel functions include Gaussian kernel, triangular kernel, and rectangular kernel. When the confidence level is 1-α, if the power value of the offshore wind turbine meets the following conditions:

[0128] P(G low <G<G up )=1-α(14)

[0129] Then the interval [G low ,G up ] is the prediction interval of power, G low , G uprepresent the lower and upper limits respectively. Specific embodiment:

[0131] The data used in this example is collected by the SCADA system from a wind turbine in an offshore wind farm in my country, with a sampling interval of 10 minutes. After cleaning the collected initial data, the discontinuous data on the time scale caused by sensor failure is interpolated by polynomial fitting, and some data with abnormal values ​​are replaced by linear interpolation, resulting in a total of 4448 data points, of which the first 3704 data points are used as the training set and the last 744 data points are used as the test set.

[0132] 1. Evaluation Metrics

[0133] Three commonly used point prediction evaluation indicators are selected: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R Squared, R2); two commonly used interval prediction evaluation indicators are selected: interval coverage (PI normalized averaged width, PINAW) and Prediction Interval Coverage Probability (PICP).

[0134] (1) Point prediction indicators

[0135] In order to verify the accuracy of the wind power point prediction method proposed in this paper, the following indicators are selected for evaluation. 2 The closer the value is to 1, the better the prediction effect is; the smaller the value of RMSE and MAE, the better the prediction effect is.

[0136]

[0137] Where: G i is the actual value of the wind power of the unit; G' i is the predicted value of wind power of the unit; m is the number of data samples; is the average wind power value of the unit.

[0138] (2) Interval prediction index

[0139] In order to verify the accuracy of the interval prediction model of the method proposed in this paper, the following indicators are selected for evaluation.

[0140]

[0141] Where: If the predicted value is within the prediction interval, then c i =1, otherwise c i=0. R is the range of variation of the predicted target value, which is used to normalize the average bandwidth. The larger the PICP, the more reliable the prediction interval. In actual prediction, the PICP should be greater than or close to the predetermined confidence level; otherwise, the prediction interval is considered invalid. In addition, while maintaining the required PICP, a smaller PINAW indicates a narrower prediction interval and a higher interval prediction accuracy.

[0142] Since PICP and PINAW only provide one-sided evaluation criteria and cannot fully reflect the comprehensive performance of interval prediction results, the CWC evaluation index is introduced. CWC solves the limitations of existing indicators by considering the coverage probability and the width of the prediction interval. The calculation formula of CWC is as follows:

[0143] CWC=PINAW[1+γ PICP e -η(PICP-σ) ,η>0](20)

[0144]

[0145] Where: η is 1, σ is the given confidence level. The smaller the CWC value, the better the interval prediction effect.

[0146] 2. ICEEMDAN decomposition results

[0147] The ICEEMDAN decomposition algorithm is used to decompose the original power data to obtain 11 IMFs. Figure 2 It can be seen that the ICEEMDAN decomposition algorithm can effectively decompose the original power data, and the complexity of IMF1-IMF11 gradually decreases.

[0148] 3. VMD decomposition results

[0149] Depend on Figure 2 The decomposition results show that although the ICEEMDAN decomposition algorithm can effectively decompose the original sequence, the decomposed IMF1 still has a high complexity. The secondary decomposition can further reduce the complexity of the sequence and improve the prediction accuracy. Therefore, this paper uses the VMD algorithm to further decompose IMF1 and decompose it into 4 subcomponents. The results after decomposition are as follows: Figure 3 shown.

[0150] 4. Analysis of point prediction results

[0151] LSTM is used to predict each IMF sequence and each sub-component obtained by secondary decomposition. In order to reflect the effectiveness of the prediction model proposed in this paper, the ICEEMDAN-VMD-LSTM prediction model proposed in this paper is compared with the LSTM and ICEEMDAN-LSTM prediction models. The error analysis of the prediction results is shown in Table 1. Figure 4 As shown in Table 1, the data obtained can show that the prediction accuracy evaluation indicators RMSE, MAE, R 2 It can be seen that R using ICEEMDAN-VMD-LSTM 2 Compared with ICEEMDAN-LSTM, it is improved by 0.5579%, and RMSE and MAE are reduced by 60.24% and 63.68% respectively. It can be seen that the secondary decomposition algorithm (ICEEMDAN-VMD) proposed by the present invention can further improve the prediction accuracy compared with the primary decomposition algorithm.

[0152] Table 1 Error analysis of point prediction results

[0153]

[0154] 5. Analysis of interval prediction results

[0155] Based on the point prediction results obtained by ICEEMDAN-VMD-LSTM, KDE is used to measure the probability density distribution of the prediction error, and the prediction interval of offshore wind power is determined accordingly. The bandwidth of KDE is set to 30, and the kernel function is selected as the KDE rectangular kernel function (KDE(box)). In order to prove the effectiveness of KDE, it is compared and analyzed with the normal distribution estimation (NDE), t-distribution estimation (TDE) and logical distribution estimation (LDE). Figure 5 is the PDF of the absolute error of the fitting point prediction under various distribution methods. Secondly, the PDF is integrated to obtain the error density integral function (cumulat ive density function, CDF), such as Figure 6 The interval prediction results under each distribution are shown as follows Figure 7Table 2 shows the evaluation index results of interval prediction using KDE (box), NDE, LDE and TDE methods at 85% confidence level. It can be seen that the interval prediction using the ICEEMDAN-VMD-LSTM-KDE (box) method has a lower CWC while meeting the PICP index.

[0156] Table 2 Error analysis of interval prediction results

[0157]

[0158] Those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention, and are not intended to limit the present invention. As long as they are within the spirit of the present invention, any changes or modifications to the above embodiments will fall within the scope of the claims of the present invention.

Claims

1. A method for offshore wind power point and interval prediction based on ICEEMDAN-VMD-LSTM-KDE, characterized in that: The steps include: S1, based on the deep analysis of offshore wind power data using a two-layer decomposition algorithm to extract key data features; S2, using LSTM algorithm to predict each component obtained by decomposition, obtain their respective prediction results, and obtain the ultimate point prediction result through calculation; S3, based on the point prediction results, uses KDE technology to measure the probability density distribution of the prediction error, and determines the prediction interval of offshore wind power accordingly.

2. The offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE according to claim 1 is characterized in that: In S1, a two-layer decomposition algorithm is used to conduct in-depth analysis of offshore wind power data to extract key data features, which specifically includes the following steps: S1.1, ICEEMDAN is used to decompose the original offshore wind speed time series into a finite number of intrinsic mode functions, thereby capturing the dynamic changes inherent in the time series; S1.2, for the difficult-to-predict high-frequency intrinsic mode function IMF1, the VMD algorithm is used to further decompose it to obtain sub-components.

3. The offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE according to claim 2 is characterized in that: The specific calculation steps of S1.1 are as follows: (1) Add I groups of noise w to the time series signal x to be decomposed (i) , construct a sequence x of noisy mixed signals (i) (i=1,2,...,k), the formula is as follows: x (i) =x+β o E1(in (i) ) Where: β o is the weight coefficient of the noise; E1(·) represents the first modal component after EMD decomposition of the noise; (2) Calculate the noisy mixed signal x (i) The local mean of , and the first set of residuals r1 and the first modal component IMF1 are calculated using local mean decomposition. The formula is as follows: IMF1=x-r1 Where: N(·) is the local mean of the noisy mixed signal; (3) Add noise E2(w (i) ), use the same method as (2) to obtain the second group of residuals r2 and modal components IMF2; IMF2=r1-r2 (4) Similarly, we can get the residual r of the kth group k and modal component IMF k ; (5) Repeat the above steps until all the IMFs of the modal components in the original sequence are obtained.

4. The offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE according to claim 2 is characterized in that: The specific calculation steps of S1.2 are as follows: The core of VMD is to construct and optimize the following variational problem: Where: {u k } is the mode function obtained by decomposition; {ω k } is the corresponding center frequency. for u k Fourier transform of , K is the preset number of modes; After constructing the above variational problem, the alternating direction multiplier method is used to solve the above variational problem. By introducing auxiliary variables and Lagrange multipliers, the original problem is transformed into a sub-problem that is easier to solve. At the same time, in each iteration, the modal {u k } and the center frequency {ω k } until convergence.

5. The offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE according to claim 1 is characterized in that: In S2, the LSTM algorithm is used to predict each component obtained by decomposition, obtain respective prediction results, and obtain the ultimate point prediction result by calculation, which specifically includes the following steps: S2.1, using LSTM to predict the intrinsic mode functions and sub-components obtained from the two decompositions to obtain sub-prediction results; S2.2, all sub-prediction results are calculated to obtain the final point prediction result.

6. The offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE according to claim 5 is characterized in that: The specific calculation steps of S2.1 are: (1) The forget gate removes the old memory unit c by using the sigmoid activation function t-1 After removing some information, the output f of the forget gate t The formula is as follows: f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f ) Where: σ is the activation function of sigmoid; Wx f , W hf , W cf is the weight coefficient matrix of the forget gate; h t-1 is the output value at time t-1; x t is the input value at time t; b f is the network layer bias of the forget gate; (2) The input gate uses the sigmoid activation function to control the information at time t, which is calculated as follows: i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i ) Where: W hi , W xi , W ci is the weight coefficient matrix of the input gate; b i is the network layer bias of the input gate; (3) Because the forget gate removes some data information at time t-1, and the input gate adds new information at time t, the cell state is updated, and the calculation formula is as follows: c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c ) Where: tanh is the hyperbolic tangent activation function; W xc , W hc is the weight coefficient matrix of the input gate; b c is the network layer bias; (4) The output gate also uses the sigmoid activation function to control the output result. The calculation formula is as follows: o t =σ(W xo x t +W ho h t-1 +W co c t +b o ) Where: W ho , W xo , W co is the weight coefficient matrix of the output gate; b o is the network layer bias of the output gate; (5) Output result h t By memory unit c t The result after processing by the tanh layer is determined together with the output of the output gate, and the calculation formula is as follows: h t =o t fishy t )。 7. The offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE according to claim 5 is characterized in that: The specific calculation method of S2.2 is to linearly sum all sub-prediction results, and the result obtained is the final point prediction result.

8. The offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE according to claim 1 is characterized in that: In S3, the KDE technique is used to measure the probability density distribution of the prediction error, and the specific steps for determining the prediction interval of offshore wind power are as follows: S3.1, obtain the prediction error based on the point prediction result and the actual value; S3.2, KDE is used to measure the probability density distribution of the prediction error, and based on this, the prediction interval of offshore wind power at a given confidence level is determined.

9. The offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE according to claim 8, characterized in that: The specific calculation steps of S3.2 are as follows: The KDE algorithm places a kernel around each data point and then stacks all these kernels to form a whole density estimate. Mathematically, the kernel density estimate can be expressed as: Where: is the estimated density at point x; n is the number of samples; x i is the data sample point; h is the bandwidth parameter, which controls the width of the kernel and affects the smoothness of the estimate; K is the kernel function, which is used to control how to get from each data point x i Contribute to the estimated density; when the confidence level is 1-α, if the power value of the offshore wind turbine meets: P(G low <G<G up )=1-a Then the interval [G low ,G up ] is the prediction interval of power, G low , G up represent the lower and upper limits respectively.

10. An offshore wind power point and interval prediction device based on ICEEMDAN-VMD-LSTM-KDE, characterized by: It is based on an offshore wind power point and interval prediction method based on ICEEMDAN-VMD-LSTM-KDE as described in any one of claims 1 to 9, comprising an electronic device and a computer-readable storage medium; The electronic device comprises a processor, a memory and a communication bus, wherein the memory stores a computer-readable program executable by the processor; The communication bus realizes the connection and communication between the processor and the memory; The processor is used to execute the computer readable program; The computer-readable storage medium stores one or more programs that can be executed by one or more processors.

Images