Offshore wind speed synthesis method and wind power prediction method

Abstract

The invention relates to an offshore wind speed synthesis method and a wind power prediction method. The method is suitable for the technical field of offshore wind power. The technical problem to be solved by the invention is to provide an offshore wind speed synthesis method and a wind power prediction method. According to the technical scheme, the offshore wind speed synthesis method comprises the steps of generating wind speed data under a second time scale on the basis of an average wind speed under a first time scale in combination with a Weibull distribution model and wind speed time characteristics; and generating wind speed data under a third time scale based on the wind speed data under the second time scale in combination with a Vor-karman power spectral density function and an autoregressive moving average model.

Description

Technical Field

[0001] This invention relates to a method for synthesizing offshore wind speed and a method for predicting wind power. It is applicable to the field of offshore wind power technology. Background Technology

[0002] Wind energy, with its clean and pollution-free advantages, is increasingly valued by countries worldwide, driven by policies aimed at energy conservation, emission reduction, and air pollution control. In recent years, countries globally have been vigorously developing wind power. Offshore wind power is a crucial direction for global wind power development. From the perspective of offshore wind power construction and planning, due to the richer marine and wind energy resources found in deep-sea areas (greater than 100 km offshore and deeper than 50 m), offshore wind resources are mainly concentrated in small areas far from load centers. The development speed of wind power exceeds local electricity demand, making local consumption difficult and leading to widespread curtailment of wind power in many wind farms, severely impacting the development of the wind power industry.

[0003] To address the issue of wind curtailment and promote large-scale wind power development, the most effective approach is to connect wind power to the grid and transmit it to power load centers, thus realizing the development strategy of "electricity coming from afar." However, due to the randomness and volatility of wind power, which can impact the voltage and frequency of the main power grid, it is necessary to limit the volatility of wind power when connecting it to the grid on a large scale.

[0004] my country's wind power development started late, resulting in insufficient historical wind resource data and a lack of wind power forecasting and grid planning coordination. This has led to unprecedented challenges for my country's wind power grid integration compared to other countries. The incomplete historical wind speed data significantly complicates early-stage site selection and capacity planning for wind farms. Furthermore, wind farm surveys require substantial costs for data collection and other related work.

[0005] Applying hybrid energy storage systems to wind power regulation requires careful consideration not only of real-time control strategies but also crucial determination of capacity. This necessitates preliminary research into actual wind speeds and power output to obtain the required wind power output over the desired timescale, providing a reference for wind farm and energy storage capacity planning.

[0006] Currently, there are many wind resource planning methods and wind speed distribution models, but most of them only focus on the long-term distribution characteristics of wind speed and are mainly applied to wind farm and power system dispatch planning. These methods are not suitable for short-term wind speed synthesis at the control time level. This makes it difficult to accurately predict wind power output and provides accurate basic data for wind farm and energy storage capacity planning. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method for synthesizing offshore wind speed and a method for predicting wind power, in view of the above-mentioned problems.

[0008] The technical solution adopted in this invention is: a method for synthesizing sea wind speed, comprising: Based on the average wind speed at the first time scale, combined with the Weibull distribution model and the time characteristics of wind speed, wind speed data at the second time scale is generated. Based on wind speed data at the second time scale, wind speed data at the third time scale is generated by combining the Vor-karman power spectral density function and the autoregressive moving average model.

[0009] The wind speed time characteristics include: simulating the wind speed time series using an autoregressive model, and simulating the daily shape curve of the wind speed using the COS function.

[0010] The generation of wind speed data at a second time scale, based on the average wind speed at the first time scale and combined with the Weibull distribution model and the time characteristics of wind speed, includes: A wind speed time series z with autocorrelation on the second time scale was generated based on an autoregressive model. Based on the average wind speed and daily shape curve at the first time scale, a wind speed time series d that satisfies the daily shape of wind speed is generated using the second time scale as the sampling time. After the wind speed time series z and d are transformed by probability into normal distribution series with mean 0 and variance 1, they are added together to obtain the wind speed time series k. By combining the Weibull distribution model, an inverse Weibull transform is performed on the wind speed time series k to obtain wind speed data at the second time scale.

[0011] The simulation of wind speed time series using an autoregressive model includes: ; in, It is white noise. For the first Wind speed at each sampling time, For the first Wind speed at each sampling time, The autocorrelation coefficient is the unit-time autocorrelation coefficient, which represents the wind speed autocorrelation coefficient with the sampling time as the time interval. It describes the degree of dependence of the average wind speed at a certain sampling point on the average wind speed over the previous time period.

[0012] The method of using the COS function to simulate the diurnal shape curve of wind speed includes: ; in The average wind speed is measured on the first timescale of the month. The moment of maximum wind speed, For the intensity of the sun's shape, Sampling time, This indicates the wind speed at any given time of day.

[0013] The shape parameters of the Weibull distribution model The scale parameters of the Weibull distribution model are determined based on the annual average wind speed. Determined based on annual average wind speed.

[0014] The wind speed data based on the second time scale, combined with the Vor-karman power spectral density function and the autoregressive moving average model, generates wind speed data at the third time scale, including: By combining the Vor-karman power spectral density function and the autoregressive moving average model, turbulent flow at the third time scale is generated. Based on the wind speed data at the second time scale and the turbulence at the third time scale, wind speed data at the third time scale is generated.

[0015] The method of combining the Vor-karman power spectral density function and the autoregressive moving average model to generate turbulent flow at the third time scale includes: By fitting the autoregressive moving average model with the Vor-karman power spectral density function, a wind speed turbulence autoregressive moving average model with power spectral density distribution characteristics is obtained. Based on the wind speed turbulence autoregressive moving average model, the turbulence flow rate at the third time scale is obtained.

[0016] A wind power prediction method, comprising: Based on the aforementioned marine wind speed synthesis method, a historical wind speed time series is generated, which in turn generates a historical power time series. Based on historical power time series and numerical weather prediction data, input variables are constructed and then fed into the wind power prediction model to obtain the predicted wind power value.

[0017] The input variables, constructed based on historical power time series and numerical weather prediction data, include: Phase space variables of historical power time series are extracted using phase space reconstruction theory; Candidate variables are constructed based on phase space variables of historical power time series and numerical weather prediction data; Input variables are selected from candidate variables based on prediction step size and minimum redundancy maximum correlation criteria.

[0018] The wind power prediction model is constructed based on an adaptive neurofuzzy inference model.

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the marine wind speed synthesis method.

[0020] A marine wind speed synthesis device has a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the steps of the marine wind speed synthesis method.

[0021] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the wind power prediction method.

[0022] A wind power prediction device has a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the steps of the wind power prediction method.

[0023] The beneficial effects of this invention are as follows: This invention utilizes the Weibull distribution model of wind speed and the time characteristics of wind speed to synthesize wind speed data, effectively simulating the distribution characteristics and temporal variation patterns of actual wind speed, making the synthesized wind speed data every 10 minutes closer to reality. This is crucial for subsequent wind power prediction based on reliable wind speed data, improving the accuracy of the prediction.

[0024] This invention considers the turbulent characteristics of wind speed and combines the Vor-karman power spectral density function to reconstruct the second-level wind speed within 10 minutes, further refining the wind speed data to obtain second-level wind speed data for one year. This provides more detailed and accurate input data for wind power prediction, and can more accurately reflect the changes in wind speed over a short period of time, which is beneficial to improving prediction accuracy.

[0025] This invention utilizes phase space reconstruction theory to select phase variables representing the wind speed evolution law from synthetic data, ensuring that the selected variables can accurately reflect the key information of wind speed changes and provide effective input for wind power prediction.

[0026] This invention combines phase variables with numerical weather prediction information as candidate input data, integrating multiple factors that influence wind power prediction, broadening the data dimensions, and enabling the model to learn and predict from a more comprehensive perspective, thus improving prediction reliability. Utilizing the minimum redundancy, maximum relevance criterion to filter candidate input data removes redundant information, retains variables most relevant to wind power prediction, improves model efficiency, avoids overfitting due to excessive irrelevant or redundant information, and further enhances prediction accuracy.

[0027] This invention employs an adaptive neurofuzzy inference model to construct a wind power prediction model. This model combines the learning capabilities of neural networks with the expressive power of fuzzy logic, enabling effective modeling of complex nonlinear relationships. By inputting the selected optimal input variables into this model, its advantages can be fully utilized to accurately predict wind power, making it suitable for predicting nonlinear systems like wind power that are influenced by various complex factors. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the power spectral density curve of actual second-level wind speed according to an embodiment of the present invention.

[0029] Figure 2 This is a schematic diagram of the Vor-karman power spectral density according to an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram of the wind speed per second per year synthesis process according to an embodiment of the present invention.

[0031] Figure 4 This is a schematic diagram of the wind speed curve over a day according to an embodiment of the present invention.

[0032] Figure 5 This is a schematic diagram illustrating the determination of phase space reconstruction parameters according to an embodiment of the present invention.

[0033] Figure 6 This is a schematic diagram of the CumminSc curve for feature selection at different time windows in an embodiment of the present invention.

[0034] Figure 7 This is a schematic diagram comparing the prediction results of different prediction step sizes for a test month according to an embodiment of the present invention.

[0035] Figure 8 This is a schematic diagram illustrating the error variation trend of different prediction step sizes in an embodiment of the present invention. Detailed Implementation

[0036] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0037] In the description of this invention, "multiple" means two or more. The use of "first" and "second" is for distinguishing technical features only and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or the order of the indicated technical features. Furthermore, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art.

[0038] Example 1: This example is a method for synthesizing sea wind speed, specifically including the following steps: S100. Based on the average wind speed at the first time scale, combined with the Weibull distribution model and the time characteristics of wind speed, wind speed data at the second time scale is generated, where the first time scale is a monthly scale and the second time scale is a 10-minute time scale.

[0039] This embodiment is based on the Weibull distribution model of wind speed and the temporal characteristics of wind speed. Based on the monthly average wind speed of a year, it synthesizes wind speed data for every 10 minutes of the year that have Weibull distribution characteristics and temporal characteristics of wind speed.

[0040] In this embodiment, the wind speed time characteristics include: simulating the wind speed time series using an autoregressive model, and simulating the daily shape curve of the wind speed using the COS function.

[0041] Since the Weibull distribution of wind speed only represents the probability distribution of wind speed over a long period of time and is a statistical probability result, it is impossible to reconstruct the temporal characteristics of wind speed using only this distribution characteristic. Therefore, the hourly autocorrelation coefficient and diurnal shape intensity of wind speed are introduced. An autoregressive model is used to simulate the wind speed time series, and the COS function is used to simulate the diurnal shape curve of wind speed in order to simulate the wind speed time series.

[0042] wind speed autocorrelation coefficient per unit time Although wind speed has a strong randomness, studies have shown that there is a certain correlation between wind speeds at different times. An autoregressive model was used to simulate the wind speed time series, and the autoregressive model is shown in equation (1): (1) in It is white noise. For the first Wind speed at each sampling time, For the first Wind speed at each sampling time. The autocorrelation coefficient is the unit-time autocorrelation coefficient, which represents the wind speed autocorrelation coefficient with the sampling time as the time interval. It describes the degree of dependence of the average wind speed at a certain sampling point on the average wind speed over the previous time period.

[0043] One-hour autocorrelation coefficient of wind speed If the sampling time is k minutes, then Thus, the autocorrelation coefficient per unit time is obtained. The value of .

[0044] Daily wind speed intensity: Although wind speed is highly random and uncertain, the daily wind speed variation pattern does have certain characteristics: the daily wind speed curve is represented by two parameters: the time of maximum wind speed within a day. and the daily shape intensity of the wind speed curve . This refers to the moment when the highest wind speed occurs within a day, recorded over a 24-hour period. Wind speed and diurnal shape intensity This refers to the correlation between wind speed data at a specific moment and its position within the broader context of the day; that is, the degree of similarity between the times of maximum wind speed each day and the shape of the daily wind speed distribution. .

[0045] use If a function is used to simulate the diurnal shape of wind speed, then the wind speed curve for a day is represented as follows: (2) in The average monthly wind speed The moment of maximum wind speed, For the intensity of the sun's shape, Sampling time, in hours. This indicates the wind speed at any given time of day.

[0046] The two-parameter Weibull distribution is a model that reflects the probability distribution of wind speed data. It generally describes the wind speed distribution characteristics on a scale of 10 minutes or more. Its probability distribution density function is shown in equation (3): (3) in, For wind speed, For Weibull shape parameters, The shape parameters of the Weibull distribution are known. and scale parameters Then, the time series of wind speed is fitted. With average annual wind speed It has the following relationship: (4) Given the average annual wind speed, the Weibull shape parameters can be obtained according to equation (4). Furthermore, after analyzing the first derivative properties of wind resource evaluation indicators—average wind speed, average wind energy density, and maximum available wind energy time—with respect to the Weibull parameter, the approximate relationship between the scale parameters and the annual average wind speed is obtained as follows: (5) Choose appropriate scale parameters based on the characteristics of different regions.

[0047] Synthesis of 10-minute average wind speeds: Since the Weibull distribution describes the wind speed distribution characteristics over a long time scale (greater than or equal to 10 minutes), if the monthly average wind speed is known, wind speed data for every 10 minutes of the year can be generated using the diurnal shape curve of wind speed and its correlation with unit time. Then, a Weibull transformation is performed on this data to obtain the annual wind speeds conforming to the Weibull distribution with a 10-minute time scale, which can be used for wind speed research. The specific process is as follows: S110, Based on the autocorrelation coefficient per unit time Equation (1) generates an autocorrelation wind speed time series in 10-minute intervals. ,in, ; S120. Based on the monthly average wind speed and formula (2), the sampling time is 10 minutes. Generate wind speed time series that satisfy the daily shape of wind speed. ; S130, Wind speed time series and After being transformed by probability into normal distribution sequences with a mean of 0 and a variance of 1, the wind speed time series is obtained by summing them, thus exhibiting both temporal correlation and diurnal shape properties. ; S140. Select appropriate Weibull shape parameters according to the characteristics of different regions. and scale parameters ,right Performing an inverse Weibull transform yields 10-minute average wind speed data that satisfy the Weibull two-parameter distribution. ; S200: Based on the wind speed data at the second time scale, combined with the Vor-karman power spectral density function and the autoregressive moving average model, wind speed data at the third time scale is generated, where the third time scale is a 1-second time scale. S210, combining the Vor-karman power spectral density function and the autoregressive moving average model, generates the turbulent flow rate at the third time scale.

[0048] In this embodiment, an autoregressive moving average model is fitted with the Vor-karman power spectral density function to obtain an autoregressive moving average model of wind speed turbulence with power spectral density distribution characteristics; based on the autoregressive moving average model of wind speed turbulence, the turbulence flow rate on the second time scale is obtained.

[0049] S220. Based on the wind speed data at the second time scale and the turbulence at the third time scale, wind speed data at the third time scale is generated.

[0050] In step S100 of this embodiment, the Weibull distribution describes the characteristics of medium- and long-term wind speed distribution. In actual wind speed simulation, the time scale of its synthesized wind speed is too large and not suitable for real-time adjustment. Therefore, it is necessary to study the synthesis of short-term components of wind speed.

[0051] Turbulence refers to short-term (generally less than 10 minutes) fluctuations in wind speed. Its main causes are friction between airflow and the ground surface, as well as thermal effects resulting from differences in air density and temperature changes, leading to vertical movement of air masses. Turbulence is largely dependent on environmental roughness, ground stability, and obstacles. It typically describes three-dimensional random phenomena, and its actual mathematical models are very complex.

[0052] In this embodiment, a model combining medium- and long-term average wind speeds with turbulence components is used to determine short-term wind speeds during wind speed time series synthesis. In this example, wind speed data with a 1-second timescale represents the 10-minute average wind speed. With turbulence The synthesis is shown in equation (6): (6) Among them, turbulence intensity It is an important parameter describing the characteristics of atmospheric turbulent motion, and its value is generally 1. .

[0053] The distribution characteristics of wind speed turbulence components are described by power spectral density, treating wind speed fluctuations as the result of the superposition of components of different frequencies. In wind speed simulation, the Vor-karman power spectral density function expression is shown in (7): (7) in For frequency, For turbulence intensity, The turbulence length scale is (m). The average wind speed over 10 minutes is used in this embodiment. .

[0054] This embodiment uses an autoregressive moving average (ARMA) model to synthesize wind speed turbulence time series. To ensure the ARMA model exhibits turbulence characteristics, it is primarily obtained by fitting the ARMA model to the Vor-karman power spectrum. The autoregressive moving average model used here for wind speed turbulence fitting is expressed with a sampling time of 1 second as follows: (8) The autoregressive order is 2. , These are the autoregressive coefficients, with a moving average order of 1. The moving average coefficient, It is zero-mean white noise with variance . Given the parameters of the Vor-karman power spectral function, the key is to use the power spectral density to solve for the autoregressive coefficients and moving average coefficients of the ARMA model. By performing a z-transform on (8), the power spectral density of the autoregressive moving average model can be obtained as follows: (9) In this example, it is known that the autoregressive coefficient and autocorrelation coefficient of ARMA have the relationship shown by the Yule-Walker equation, as shown in (10): (10) in Let the order be the autoregressive order. Let be the order of the moving average. The autoregression order in this application is 2, and the moving average order is 1. Expanding (10) yields: (11) The autocorrelation coefficient between wind speed turbulence sequences is defined as: (12) Based on the Fourier transform relationship between equations (9) and (12), we obtain: (13) The autoregressive coefficients of the ARMA model can be obtained by simultaneously solving equations (9), (11), and (13). , .

[0055] In this embodiment, let Then the sequence autocorrelation coefficient Use sequence autocorrelation coefficient Represented as: (14) in Additionally, based on the autocorrelation coefficient... Definition Know (15) in .make By combining equations (14) and (15), the moving average coefficients of the ARMA model can be obtained. and zero-mean white noise variance .

[0056] After obtaining the above parameters, the turbulent flow rate per second over 10 minutes is calculated using the ARMA model. Then, based on the average wind speed data per 10 minutes within a year obtained from equation (6) and step S100, the wind speed data per second for a year can be synthesized.

[0057] The power spectral density curve of the wind speed was obtained by verifying the second-level wind speed over a 10-minute period, as shown in the figure. Figure 1 As shown. Figure 2 The image shows the Vor-karman power spectral density curve. (Comparison) Figure 1 and 2 It can be seen that the power spectral density curve of the actual wind speed at the second level has a certain degree of fit with the Vor-karman power spectral density curve.

[0058] Wind speed synthesis simulation: This embodiment uses MATLAB for simulation. The input data is the monthly average wind speed of a certain region, as shown in Table 1. At the same time, according to equations (4) and (5), the wind speed characteristics of a certain region are used to select the parameters in the wind speed synthesis process, as shown in Table 2.

[0059] Table 1 Average wind speed for a certain month

[0060] Table 2. Wind speed composite parameters

[0061] Based on Tables 1 and 2, the synthesis of the wind speed sequence was simulated using MATLAB software. The specific operation procedure is as follows: Figure 3 As shown.

[0062] First, input the monthly average wind speed. Then, generate the average wind speed for every 10 minutes of the month based on the Weibull distribution of wind speed and the time characteristics of wind speed. .

[0063] use Based on turbulence characteristics, the average wind speed for each second of the month is generated using the ARMA model. This process is repeated until the 12th month, at which point the wind speed data for each second of the year will be output.

[0064] Selecting wind speed data for one day yields the wind speed curve for that day, as shown below. Figure 4 As shown, the wind speed distribution in the medium and long term shows a relatively obvious trend, while the short-term wind speed fluctuates to some extent.

[0065] In practical applications, the Weibull parameters and turbulence scale in the wind speed synthesis process of this embodiment are also taken from the existing local historical wind speed data to restore the local data as much as possible.

[0066] This embodiment utilizes the Weibull distribution model of wind speed and its temporal characteristics to synthesize wind speed data. This effectively simulates the distribution characteristics and temporal variation patterns of actual wind speed, making the synthesized wind speed data every 10 minutes closer to reality. This is crucial for subsequent wind power prediction based on reliable wind speed data, improving prediction accuracy. Considering the turbulent characteristics of wind speed and combining it with the Vor-Karman power spectrum to reconstruct the second-level wind speed within 10 minutes, the wind speed data is further refined, resulting in second-level wind speed data for one year. This provides more detailed and accurate input data for wind power prediction, more accurately reflecting wind speed changes over a short period of time, thus improving prediction accuracy.

[0067] Example 2: This example is a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the marine wind speed synthesis method described in Example 1.

[0068] Example 3: This example is a marine wind speed synthesis device, which has a memory and a processor. The memory stores a computer program, which is executed by the processor to implement the steps of the marine wind speed synthesis method described in Example 1.

[0069] Example 4: This example is a wind power prediction method, specifically including: A. Based on the marine wind speed synthesis method described in Example 1, a historical wind speed time series is generated, and then a historical power time series is generated.

[0070] After synthesizing second-level wind speeds using the method described in Example 1, the data is input to a simulated wind turbine or based on a wind turbine model to obtain second-level wind power, providing data for real-time control of the wind farm.

[0071] B. Based on historical power time series and numerical weather forecast data, construct input variables and input the input variables into the wind power prediction model to obtain the predicted wind power value.

[0072] B1. Extract phase space variables from historical power time series using phase space reconstruction theory; B2. Construct candidate variables based on phase space variables of historical power time series and numerical weather prediction data; B3. Based on the prediction step size and the minimum redundancy maximum correlation criterion, input variables are selected from the candidate variables and input into the wind power prediction model to obtain the wind power prediction value.

[0073] Because weather systems exhibit chaotic characteristics, their dynamic evolution is sensitive to initial values. As the forecast time step increases, the correlation between historical power time series and future power generation rapidly decays, potentially even worsening forecast performance. Therefore, it is necessary to combine time series data with numerical weather prediction data to provide as comprehensive and effective input variables as possible.

[0074] Numerical weather prediction (NMR) forecasts weather changes by solving thermodynamic and fluid dynamic equations based on system meteorological data. However, it can only provide rough, short-term weather feature forecasts and is not perfectly suited to specific local conditions. NMR typically uses data to provide supplementary information, such as short-term forecasts of meteorological elements like wind speed, wind direction, irradiance, temperature, humidity, and air pressure.

[0075] For prediction tasks with different time steps, these related variables, including historical power time series and NWP data, may have different impacts on the prediction target. Therefore, a two-stage input feature selection method is used in the proposed prediction method. First, through phase space reconstruction technology, this embodiment selects phase variables representing the system evolution law from the historical power time series with chaotic properties, and then combines the phase variables with NWP element information as candidate inputs; then, the optimal input variables are selected based on the minimum redundancy maximum relevance (mRMR) criterion.

[0076] The Lyapunov exponent is used to indicate that the time series of distributed renewable energy generation has chaotic characteristics. Therefore, the nonlinear dynamic characteristics of the power time series can be extracted and recovered using phase space reconstruction theory, and the finite-dimensional phase space of the time evolution of the sampling system can be reconstructed using the delay factor technique.

[0077] In the phase space reconstruction process, since the real dataset is finite and noisy, determining a suitable delay factor τ and a good embedding dimension m is not only very important, but also often quite difficult.

[0078] For estimating the delay factor τ and the embedding dimension m, a simplified approach is to assume that they are unrelated and choose them independently. When determining the delay factor τ first, methods such as autocorrelation function, multiple autocorrelation, and mutual information are used. Then, the GP algorithm or pseudo nearest neighbor algorithm is used to further determine the embedding dimension m.

[0079] However, generally speaking, the delay factor and the embedding dimension are interdependent, and the delay time window should be estimated comprehensively. This allows for the selection of m and τ. The C_C method (Correlation Integral-based Method) is used to estimate the delay time window. The C_C method is used to determine the optimal phase space reconstruction parameters for historical time series, which reduces the computational complexity and improves the accuracy of subsequent prediction models.

[0080] Phase space reconstruction is an effective method for analyzing dynamic patterns in chaotic time series data. (Time series) By reconstructing it into a multidimensional phase space vector To represent the dynamic characteristics of the system, the reconstructed phase space vector is given by equation (16): (16) in, , , It is the embedding dimension. It is a delay factor. In this application, the C_C method is constructed using two correlation integrals to reconstruct a given time series. To simplify the candidate input format.

[0081] The correlation integral of an embedded time series is defined by the following formula: (17) in, , Represents the infinity norm, and t represents the time index. It is the Heaviside function, that is The correlation integral is a cumulative distribution function, representing the probability that the distance between any two points in the phase space is less than the search radius r. To investigate the nonlinear correlation of time series and eliminate stray interference related to time correlation, it is necessary to process the given time series... Divide into disjoint subsequences. Disjoint time series are represented by the formula: (18) in, Indicates the length of the subsequence. , This indicates that the integer part of the numerical value is retained.

[0082] Furthermore, construct statistical functions. As a dimensionless measure of correlation in nonlinear time series. For any time index t in the aforementioned disjoint time subsequences, the statistical function... The definition is shown in the formula: (19) when When this happens, we can obtain the following formula: (20)

[0083] If the time series data conforms to independent and identically distributed (i.e., for fixed values ​​m and t) and At that time, statistical functions The result will be zero. However, due to the finite nature of real-world datasets and the correlation of sequence components, the statistical function... It is usually non-zero. For all search radii r, maximizing the difference in the statistical function is given by the formula: (twenty one) N, m, and r are estimated using Brock–Dechert–Scheinkman (BDS) statistics, and their values ​​are: , , ,in The standard deviation of a time series is represented by the following formula: (twenty two) (twenty three) (twenty four) Changing the time index t, we have: when The function crosses zero for the first time or When the function first reaches its minimum value, this is considered to be the optimal delay factor, denoted as . .when When the function reaches its global minimum, the optimal embedding time window is considered to be at this point. And then according to the formula The optimal embedding dimension *m* was calculated. Thus, the phase space reconstruction parameter delay factor was determined using the C_C method. And the embedding dimension m.

[0084] After determining the phase space reconstruction parameters based on the characteristics of the time series, the historical power time series is extracted as a phase space variable, denoted as... ; in Let t be the power generation at time t.

[0085] To utilize numerical weather prediction information, taking wind power generation as an example, five meteorological factors that may affect the output of wind turbines are selected to construct NWP candidate input variables, denoted as... ; in , , , , These represent the predicted times in the future. The values ​​of wind speed, wind direction, temperature, humidity, and air pressure predicted by NWP.

[0086] Therefore, the total candidate variables will include phase space variables and numerical weather prediction variables, denoted as... It can be seen that the dimension of the candidate variable is , where m is the embedding dimension of the phase space reconstruction.

[0087] For candidate input variables composed of phase space variables and numerical weather prediction variables, this embodiment uses the minimum redundancy maximum correlation (mRMR) criterion, with mutual information (MI) as the metric, for screening. Mutual information measures the linear and nonlinear correlations between variables. The basic idea of ​​the feature selection algorithm based on mutual information is to maximize the joint mutual information between the selected variable subset S and the target output Y. This allows for the search of a compact and information-rich input space within the original variable set. The minimum redundancy maximum correlation technique aims to find candidate variables that have maximum dependence on the prediction target while having minimum redundancy among variables within a relatively subset.

[0088] The concept of mutual information comes from information entropy, which applies to any two random variables. and The joint information entropy between them is defined as follows: (25) In the formula: Let X be the joint probability density function of random variables X and Y. The conditional entropy of Y given variable X is defined as: (26) Conditional entropy Mutual information (MI) represents the uncertainty of random variable Y given random variable X, and is less than or equal to the joint entropy of the two variables. MI is the amount of information shared between two variables, defined as follows: (27) Mutual information (MI) can be understood as the amount of information provided by variable X when reducing the uncertainty of variable Y, and it possesses symmetry. MI is zero if and only if the two random variables are independent of each other.

[0089] The Minimum Redundancy Maximum Correlation (mRMR) criterion considers both maximum correlation and minimum redundancy to identify a subset of input information variables. Since simple combinations of individual input information variables may not achieve good predictive performance, both the information provided by a single input variable and the redundancy between variables should be taken into account. Under the mRMR criterion, a single input variable ( The information content score is calculated using the following formula: (28) In the formula: V is the set of all candidate attribute variables; S is the set of input variables that have been selected; The number of variables in the set. The first term on the right-hand side of the equation refers to the target t and the candidate input variables ( Mutual information is used between them. To measure the candidate variable ( The first item is the strength of the relevance of information in the prediction process; the second item is to select those variables that have the least similarity or redundancy among the variables, so that the selected set is more representative or provides information about the whole set.

[0090] In implementation, the input variables are selected using a stepwise search incremental forward method according to equation (28), resulting in a larger information content score. This indicates a higher likelihood of it being used as an input variable. In the first step, the maximum relevance component (MI) of all candidate attribute variables is calculated, i.e., the mutual information with the prediction target. The variable with the largest MI is determined as the first input variable, as shown in the equation: (29) The selection (ranking) of the remaining variables is based on a comparison of the information content scores obtained from formula (28). In step m ( In ), the remaining variables are ( ),in For an ordered set of selected (or sorted) variables in (m-1) Therefore, the variables chosen in m are as follows: (30)

[0091] As each input variable is selected, the algorithm dynamically moves forward one step, thus the process incrementally evaluates all candidate variables until step [step number missing]. All attribute variables (V) were evaluated and ranked according to the mRMR criterion. The information content score for each variable after ranking can be calculated according to equation (28), as follows: (31)

[0092] After quantitatively determining and ranking the information contribution rates of all candidate variables, the number of final input variables is determined based on the cumulative information contribution scores of the variables. The cumulative information contribution rate is defined as follows: ,when When the maximum value is reached, the corresponding ordered set of variables is formed. The final input variable represents the feature attribute with minimum redundancy and maximum relevance, which will be used for regression pattern mining in subsequent machine learning algorithms.

[0093] After obtaining the input variable V corresponding to the prediction target Y, the sample dataset required for training the learning model is constructed using historical data. This embodiment further divides each sample into patterns based on the feature space constructed by the selected variables, that is, dividing the obtained historical training dataset into multiple subsets, where samples in the same subset have similar weather patterns. This greatly reduces the complexity of the training process and uses a lightweight learning model, which is more conducive to improving the model's generalization performance.

[0094] In this embodiment, the process of partitioning the sample dataset is implemented using the k-means clustering algorithm, and the specific method is as follows:

[0095] First, the sample input space is normalized using Z-score standardization (also known as zero-mean normalization). ,in , These represent the mean and standard deviation for each feature dimension, respectively. N is the number of instances in the sample dataset, and D is the dimension of the sample input space, which is the number of input variables after filtering.

[0096] Secondly, the k-means clustering algorithm is used to normalize the training set. The subsets are divided into k distinct pattern categories using unsupervised learning. Let the cluster center be randomly initialized as Then the k-means algorithm will minimize the Euclidean distance of the cluster. Iteratively solve for the cluster centers until... It no longer changes, achieving convergence of the clustering algorithm. Expressions such as: (32) After determining the cluster centers, the samples in the training set are grouped into subsets that are closest to the cluster centers. For subset k, , The number of samples included is denoted as Each sample subset Both can be used to train a lightweight learning model. When performing prediction tasks, the input vector in the feature space will also perform an online prediction and inference process based on the learning model corresponding to its nearest cluster center.

[0097] In this embodiment, the wind power prediction model is built based on the Adaptive Neuro-Fuzzy Interference System (ANFIS). The ANFIS model is a data-driven modeling method where the fuzzy membership function and fuzzy rules are obtained through learning from a large amount of historical data, rather than being based on human experience or intuition. It combines the nonlinearity and adaptability of neural networks with the advantages of fuzzy inference systems in handling complex systems, making it suitable for predicting multivariate nonlinear systems. The ANFIS model is based on the Takagi-Sugeno inference method, creating a nonlinear mapping from the input space to the output space using fuzzy IF-THEN rules. The algorithm parameters include prerequisite parameters and conclusion parameters.

[0098] The mathematical description of each layer of the ANFIS structure is as follows: Layer 1 (Fuzzification): The input variables are fuzzified, and the membership values ​​of the input quantities in each fuzzy set are determined by the membership function (MF). , , In the formula: , The segmented fuzzy set; This refers to the membership degree in the corresponding case; For a Gaussian curve, the membership function is as follows: As shown, its parameter set The prerequisite parameters for constructing ANFIS.

[0099] Layer 2 (Rule-based Reasoning): The neurons in this layer introduce the fuzzy input membership degrees from the upper layer and calculate the activation weights under each reasoning rule, i.e. .

[0100] Layer 3 (Normalization): Each neuron in this layer receives input from all neurons in the previous layer and normalizes the activation weights under each rule, i.e. .

[0101] Layer 4 (Defuzzification): This layer uses the rule interpretation function to inverse fuzzify the activation weights under each rule, and calculates the given rule. The weighted subsequent value, The parameter set of the linear interpretation function used. The conclusion parameters that constitute ANFIS.

[0102] Layer 5 (Output): This layer sums the outputs of all deblurred neurons, i.e. The final output of ANFIS is obtained. .

[0103] Machine learning problems ultimately boil down to optimization problems. In supervised learning on training data, the preconditions that determine the shape of the membership function and the postconditions of the rule interpretation function are optimized. Particle Swarm Optimization (PSO) is used to heuristically search for the network training parameters, minimizing the root mean square error of the normalized mean squared error of the sample data. The objective function is given by the formula: (33) In the formula: For input The sample vector is the output value of the ANFIS network; These are the actual measured values ​​for the corresponding samples; N represents the power-on capacity; N represents the number of sample data.

[0104] Each particle in PSO identifies and updates its local optimum (Pbest), and also collectively searches for the global optimum (Gbest) in the swarm. The position and velocity functions in PSO can be expressed as: (34) In the formula, and For random variables, and is a positive constant, and w is the inertia factor; and These are the position vector and velocity vector of the i-th particle, respectively; and These represent the individual optimal solution and the social (global) optimal solution for the i-th particle, respectively.

[0105] The following example demonstrates the effectiveness of the proposed hybrid prediction method combining feature selection and classification training. To verify the reliability of the examples as broadly as possible, this embodiment uses real data collected from actual wind farms for performance evaluation. The test month is June 2017, specifically the summer wind farm output prediction (capacity 45MW). The historical data provided by the wind farm database includes NWP data for power generation (direct measurement), with a time series resolution of 15 minutes, i.e., 96 observations per day. In selecting the training dataset, the wind farm used data from the three months prior to the test month, i.e., March to May 2017 (approximately 8640 observations).

[0106] Using wind power time-series data from the training set, the phase space reconstruction parameters are determined through the C_C method. Since wind power output in different regions has different spatiotemporal characteristics, this method adaptively determines the optimal parameters for each wind farm based on data information. Figure 5 It was observed that when t equals 26 There is a first local minimum, therefore the optimal delay factor is... Set it to 26. For example... Figure 5 (b) It is concluded that, The optimal embedding window corresponding to the global minimum point Therefore, the embedding dimension m is calculated to be 3.

[0107] Phase space reconstruction transforms historical time series data into representative m-dimensional phase space variables. Let the current time be t, then the obtainable phase space input variables are: ; in It is the wind power output value measured at time t.

[0108] For ease of description, these m-dimensional input variables are denoted as the feature set. The meteorological variables considered in NWP data include wind speed, trigonometric wind direction, temperature, humidity, and atmospheric pressure, denoted as the feature set. Based on the phase space reconstruction parameters obtained by the C_C method, the total candidate input variable set for a certain wind farm can be obtained as follows:

[0109] Wind farm: {Hin1, Hin2, Hin3, WindV, WindD, Temp, Hum, AirP}

[0110] The aforementioned candidate variable set will be ranked using the minimum redundancy maximum relevance (mRMR) criterion to compare the strength of the beneficial information contributed by each input variable to the prediction target. This will be achieved by observing the cumulative information contribution rate. When the curve stops increasing or grows very slowly, the feature ranked highest is selected as the optimal input variable. Furthermore, for different time window prediction tasks, the optimal input variable is adaptively selected based on the different prediction step sizes, thereby improving the accuracy of wind power prediction in different time windows.

[0111] Based on the proposed minimum redundancy maximum relevance feature filtering criterion Figure 6The curve changes of an electric field under different time window prediction tasks (e.g., 1 hour, 2 hours, 3 hours, and 4 hours in advance from (a) to (d)) are shown. The input variables for different prediction step sizes are selected accordingly based on the changing trend of the CumInSc curve. When the cumulative information contribution rate curve reaches its maximum value, it is considered that adding additional variables after the extreme point will not add more useful information. Therefore, the candidate variables before the peak point of the CumInSc curve constitute the optimal or near-optimal input variable space of the prediction model. The detailed ranking of each candidate variable under different step size prediction conditions is shown in Table 3, where the selected input variables are highlighted by shading. It can be seen that the ranking of candidate variables is similar for tasks with similar prediction steps. For prediction tasks under different time windows, this method can rank the effective information intensity of the candidate variables' contribution to the prediction target, thereby achieving maximum relevance and minimum redundancy input space compression and variable selection, effectively reducing input dimensionality and interference information.

[0112] Table 3. Ranking and selection of candidate variables for different time windows.

[0113] For prediction tasks within different time windows, after selecting the corresponding sample input space, the ANFIS learning model, which combines the Particle Swarm Optimization (PSO) algorithm, was trained and fitted to different clustering pattern subsets. The main structural parameters of the PSO-ANFIS model are shown in Table 4. The performance of the trained prediction model was evaluated on the test set. Figure 7 The results of continuous forecasting using the proposed forecasting method at different forecasting steps of 150 hours in the test month are presented.

[0114] Table 4 Main parameter settings for the wind power prediction model

[0115] To evaluate the effectiveness and accuracy of prediction methods, the normalized root mean square error (nRMSE) and normalized mean absolute error (nMAE) are used as performance metrics, and are given by the following formulas. Generally speaking, the smaller these error measures are, the better the prediction performance of the corresponding model method.

[0116] (35) (36) in, Let be the predicted power at time i in the time series. Let N be the actual measured power at time i, and N be the length of the statistical time series. This represents the operating capacity of the wind turbines in the wind farm at the corresponding time.

[0117] To evaluate the proposed minimum redundancy maximum correlation wind power prediction input variable selection method (mRMR-IVS model), a detailed comparative study was conducted with other feature selection methods on tasks with different prediction step sizes. The benchmark feature selection methods were the phase space reconstruction (PSR-IVS) model and the principal component analysis (PCA-IVS) model. For the PSR-IVS prediction model, the input variables included NWP variables and phase space reconstruction variables from historical time series. For the PCA-IVS prediction model, the original variables were composed of NWP variables and historical time series data, and then principal component analysis (PCA) was used to map the original input space to the principal component space. The PCA method automatically reduced the original attribute variables to appropriate principal component variables, and the independent principal components well preserved the key features of the original variables. After selecting or extracting the input variables, all prediction models and their training methods based on the input variable selection method were as described in the previous section, thus ensuring consistency among the comparison methods.

[0118] Table 5 presents the multi-step prediction results (including nRMSE and nMAE) of different prediction models applied to wind farms. As shown in Table 5, compared with the two comparative models, the proposed PSR-IVS prediction method has the smallest error across all 16 prediction step lengths. Among the two comparative methods, the phase space reconstruction-based model (PSR-IVS) performs better in short-term predictions than the principal component analysis-based model (PCA-IVS), but its error increases in longer prediction step lengths. The PSR-IVS prediction method is more targeted in its refined selection of input variables, resulting in better performance than the comparative methods in overall multi-step predictions.

[0119] Figure 8 The graphs are plotted on the nRMSE and nMAE values ​​for two wind farms based on different prediction step lengths, providing a more intuitive comparison of the relationship between prediction error and prediction time window. The curve trend shows that the error level increases with the increase of the prediction step length, which is consistent with objective laws. The results show that the error curves in the two benchmark models are not strictly monotonically increasing, especially for tasks with intermediate prediction step lengths, because the input variable selection of the two models cannot adapt to every prediction task. In the proposed model, the error trend is stable and monotonically increasing, indicating that the mRMR-IVS-based model adaptively selects the optimal or near-optimal input variables for different prediction steps, striving for the best possible prediction accuracy. Test results show that the proposed hybrid intelligent model effectively selects the most valuable input variables in different meteorological models, thus exhibiting better adaptability and generalization.

[0120] Table 5 Comparison of multi-time-window prediction performance under different feature selection methods

[0121] To further verify the effectiveness of continuous-time multi-step prediction in the ultra-short-term wind power forecasting task (i.e., prediction results of continuous predictions for the next 0-4 hours in 15-minute increments), 100 consecutive 4-hour time series were randomly selected during the experimental month's prediction process, and statistical performance indicators were calculated. The nRMSE and nMAE indices for each overall time series were calculated. Then, based on the frequency of occurrence of these 100 results, the frequency distribution of different error levels was obtained. The frequency distribution of the average error and typical error levels of a certain wind farm is shown in Table 6.

[0122] For both comparative models, the proposed mRMR-IVS-based prediction method exhibits the smallest error level in the ultra-short-term continuous prediction of wind power. The results in Table 6 verify that the proposed intelligent model has sufficiently high accuracy in the distribution statistics under different error levels. This clearly demonstrates that the proposed method has better prediction performance than the baseline method.

[0123] Table 6 shows the average error and probability distribution of error levels for each method.

[0124] This embodiment selects phase variables representing the evolution of wind speed from synthetic data, ensuring that the selected variables accurately reflect key information about wind speed changes and provide effective input for wind power prediction. Combining phase variables with numerical weather prediction elements as candidate input data integrates multiple factors affecting wind power prediction, broadening the data dimension and enabling the model to learn and predict from a more comprehensive perspective, thus improving prediction reliability. Using the minimum redundancy, maximum correlation criterion to filter candidate input data removes redundant information, retains variables most relevant to wind power prediction, improves model efficiency, avoids overfitting due to excessive irrelevant or redundant information, and further enhances prediction accuracy.

[0125] This embodiment employs an adaptive neuro-fuzzy inference model, which combines the learning ability of neural networks with the linguistic expression capabilities of fuzzy logic, enabling effective modeling of complex nonlinear relationships. By inputting the selected optimal input variables into this model, its advantages can be fully utilized to accurately predict wind power, making it suitable for predicting nonlinear systems like wind power that are influenced by a variety of complex factors.

[0126] Example 5: This example is a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the wind power prediction method described in Example 4.

[0127] Example 6: This example is a wind power prediction device, which has a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the wind power prediction method described in Example 4.

[0128] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0129] In the foregoing description of this specification, references to terms such as "one embodiment," "another embodiment," or "some embodiments" indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0130] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0131] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for synthesizing sea wind speed, characterized in that, include: Based on the average wind speed at the first time scale, combined with the Weibull distribution model and the time characteristics of wind speed, wind speed data at the second time scale is generated. Based on wind speed data at the second time scale, wind speed data at the third time scale is generated by combining the Vor-karman power spectral density function and the autoregressive moving average model.

2. The method for synthesizing marine wind speed according to claim 1, characterized in that, The wind speed time characteristics include: simulating the wind speed time series using an autoregressive model, and simulating the daily shape curve of the wind speed using the COS function.

3. The method for synthesizing marine wind speed according to claim 2, characterized in that, The generation of wind speed data at a second time scale, based on the average wind speed at the first time scale and combined with the Weibull distribution model and the time characteristics of wind speed, includes: A wind speed time series z with autocorrelation on the second time scale was generated based on an autoregressive model. Based on the average wind speed and daily shape curve at the first time scale, a wind speed time series d that satisfies the daily shape of wind speed is generated using the second time scale as the sampling time. After the wind speed time series z and d are transformed by probability into normal distribution series with mean 0 and variance 1, they are added together to obtain the wind speed time series k. By combining the Weibull distribution model, an inverse Weibull transform is performed on the wind speed time series k to obtain wind speed data at the second time scale.

4. The method for synthesizing marine wind speed according to claim 2 or 3, characterized in that, The simulation of wind speed time series using an autoregressive model includes: ； in, It is white noise. For the first Wind speed at each sampling time, For the first Wind speed at each sampling time, The autocorrelation coefficient is the unit-time autocorrelation coefficient, which represents the wind speed autocorrelation coefficient with the sampling time as the time interval. It describes the degree of dependence of the average wind speed at a certain sampling point on the average wind speed over the previous time period.

5. The method for synthesizing marine wind speed according to claim 2 or 3, characterized in that, The method of using the COS function to simulate the diurnal shape curve of wind speed includes: ； in The average wind speed is measured on the first timescale of the month. The moment of maximum wind speed, For the intensity of the sun's shape, Sampling time, This indicates the wind speed at any given time of day.

6. The method for synthesizing marine wind speed according to claim 1, characterized in that, The shape parameters of the Weibull distribution model The scale parameters of the Weibull distribution model are determined based on the annual average wind speed. Determined based on annual average wind speed.

7. The method for synthesizing marine wind speed according to claim 1, characterized in that, The wind speed data based on the second time scale, combined with the Vor-karman power spectral density function and the autoregressive moving average model, generates wind speed data at the third time scale, including: By combining the Vor-karman power spectral density function and the autoregressive moving average model, turbulent flow at the third time scale is generated. Based on the wind speed data at the second time scale and the turbulence at the third time scale, wind speed data at the third time scale is generated.

8. The method for synthesizing marine wind speed according to claim 7, characterized in that, The method of combining the Vor-karman power spectral density function and the autoregressive moving average model to generate turbulent flow at the third time scale includes: By fitting the autoregressive moving average model with the Vor-karman power spectral density function, a wind speed turbulence autoregressive moving average model with power spectral density distribution characteristics is obtained. Based on the wind speed turbulence autoregressive moving average model, the turbulence flow rate at the third time scale is obtained.

9. A method for predicting wind power output, characterized in that, include: Based on the marine wind speed synthesis method according to any one of claims 1 to 8, a historical wind speed time series is generated, and then a historical power time series is generated; Based on historical power time series and numerical weather prediction data, input variables are constructed and then fed into the wind power prediction model to obtain the predicted wind power value.

10. The wind power prediction method according to claim 9, characterized in that, The input variables, constructed based on historical power time series and numerical weather prediction data, include: Phase space variables of historical power time series are extracted using phase space reconstruction theory; Candidate variables are constructed based on phase space variables of historical power time series and numerical weather prediction data; Input variables are selected from candidate variables based on prediction step size and minimum redundancy maximum correlation criteria.

11. The wind power prediction method according to claim 9, characterized in that, The wind power prediction model is constructed based on an adaptive neurofuzzy inference model.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the marine wind speed synthesis method according to any one of claims 1 to 8.

13. A marine wind speed synthesis device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the marine wind speed synthesis method according to any one of claims 1 to 8.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the wind power prediction method according to any one of claims 9 to 11.

15. A wind power prediction device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the wind power prediction method according to any one of claims 9 to 11.

Images