A wind power probability prediction method

Abstract

The present invention provides a method for probabilistic prediction of wind power, which includes calculating the Moran's index value of the region where the wind farm cluster is located, extracting spatial features, and forming a final input matrix; decomposing the wind power time series data to obtain preliminary components, and further decomposing the preliminary components to obtain several components; calculating entropy values using the refined composite multi-scale weighted permutation entropy, and comprehensively measuring and analyzing using the mean method to obtain several comprehensive components; obtaining the final deterministic prediction result and verifying the performance according to the FAAM-Stacking model; and obtaining the wind power interval prediction result according to the error of the final deterministic prediction result and using the distributed kernel density estimation method, and verifying the wind power interval prediction result. The present invention effectively avoids the problems of poor decomposition and the influence of different bandwidths of each component on the estimation error, and can effectively improve the prediction accuracy of wind power.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and particularly relates to a method for probabilistic prediction of wind power. Background Art

[0002] The world is practicing the concept of green development, aiming at the direction of clean and low-carbon, and vigorously developing renewable energy represented by wind power. However, the intermittency and volatility of wind power pose higher requirements for the flexibility of the power grid. Due to the continuous growth of global electricity demand and the background of high wind power penetration, accurate prediction of wind farm wind power is an effective method to quantify the uncertainty of wind energy and ensure the stable scheduling of wind energy resources.

[0003] Wind power prediction can be divided into two categories: deterministic prediction and uncertainty prediction according to the form of the final prediction result. Among them, the deterministic prediction method can be further divided into single-model and combined-model prediction. For example, scholars such as Tuerxun W used a single LSTM model to predict wind power. However, due to the strong volatility of wind energy, it is difficult for a single model to effectively learn the relationship between features, and the combined model overcomes the weaknesses of the single model and has become the main method for deterministic prediction. Scholars such as Hanifi used a CNN-LSTM model to predict wind power, which can effectively process long-term sequence data and further improve the prediction accuracy. However, the CNN-LSTM model has an over-reliance on gradient information, which may lead to problems such as gradient disappearance or explosion. The Stacking model can integrate multiple base models, learn from each other's strengths and make up for each other's weaknesses, and solve the problem of gradient explosion. Qu et al. used the D-Stacking model to predict the power of a wind farm cluster according to the different amplitudes of the decomposed subsequences, and achieved higher prediction accuracy than combined models such as CNN-LSTM, but did not consider the influence of different amplitudes on the prediction result. Scholars such as Kong used the method of attention mechanism to focus on the main features, thereby improving the accuracy of the model. However, there are few reports on the prediction method combining the attention mechanism and Multi-Stacking (FAMM-Stacking) in the deterministic prediction of wind farm cluster power, which is worthy of in-depth study. However, deterministic prediction is difficult to characterize the uncertainty of wind power, and probabilistic prediction is a powerful tool to quantify the prediction result deviation caused by wind power uncertainty. While giving the deterministic prediction result, it depicts the probability interval of wind power fluctuation, which is more conducive to reasonably arranging reserve capacity, thermal power unit startup plans, etc.

[0004] There are many methods for wind power probability prediction. According to whether it depends on a specific type of probability distribution form, wind power probability prediction can be divided into parametric prediction and non-parametric prediction. Parametric probability prediction requires assuming that the wind power or prediction error distribution follows a Gaussian distribution, Weibull distribution, etc. For example, scholars such as Shirzadi N used the Weibull distribution to simulate its probability distribution. However, wind energy has strong volatility and randomness, and the prediction error will change with the change of prediction conditions. The parametric distribution assumption is difficult to characterize the true probability distribution of the prediction object. Non-parametric probability prediction is a probability prediction method without giving any distribution shape assumptions, effectively avoiding the problems of parametric distribution probability prediction. Common non-parametric probability prediction methods include quantile regression, kernel density estimation, etc. Some scholars used an adaptive quantile regression algorithm to predict the quantiles of wind power output, but for multiple quantiles, separate modeling and calculation are required, resulting in low calculation efficiency. Scholars such as Yang proposed a non-parametric kernel density estimation (KDE) method, which improved the calculation efficiency and realized the probability distribution modeling of wind power. However, the selection of the bandwidth parameter has a great impact on the estimation result. Some researchers optimized KDE using particle swarm optimization (PSO) and grid search (GS) respectively, and achieved better prediction results than the unoptimized KDE method. However, the optimized KDEs all perform probability prediction on the overall deterministic prediction result after superposition, ignoring the influence of different bandwidths of each component on the estimation error. In order to minimize the estimation error as much as possible, distributed KDE is proposed. By combining different components with an attention mechanism, the error distribution of the deterministic prediction result is fitted to obtain the final probability prediction.

[0005] To further improve the accuracy of wind power prediction, scholars generally consider it from two aspects: time series decomposition and feature extraction. In terms of time series decomposition, there are mainly various decomposition methods such as EMD, ICEEMDAN, and VMD. Nahid used EMD decomposition and combined it with the CNN-LSTM model to obtain the wind power prediction results. However, there is a mode mixing phenomenon in the signals decomposed by EMD. Scholars such as Emeksiz used ICEEMDAN decomposition, which can suppress the mode mixing phenomenon to a certain extent, with relatively small reconstruction error and better prediction effect. Hu et al. studied the use of VMD to decompose the historical wind power time series, which can well suppress the mode mixing phenomenon of the signals and improve the accuracy of wind power prediction. However, the high-frequency strong non-stationary components generated by VMD decomposition will still lead to large prediction errors. At the same time, VMD penalizes the internal jumps based on the L2 smoothing stage transition and performs poorly in dealing with burst signals. Scholars such as Sibtain combined the VMD-ICEEMDAN secondary decomposition technique with the LSTM model for prediction, which solved the problem of large prediction errors caused by the high-frequency non-stationary components generated by VMD decomposition, but still did not solve the problem of poor decomposition of burst signals. Parri and Teeparthi used successive variational mode decomposition (SVMD), which constructs a variational model to describe different modes of the signal and extracts the modes by optimizing the algorithm, with higher flexibility and better adaptation to more complex burst signals. Therefore, a rarely reported SVMD-ICEEMDAN secondary decomposition method is proposed to effectively solve the problems of poor prediction effect of high-frequency signals and poor decomposition of burst signals in complex time series. However, after SVMD-ICEEMDAN secondary decomposition, there are a large number of subsequences, increasing the time complexity of the model. Some researchers use permutation entropy and sample entropy to measure the complexity of subsequences and aggregate the subsequences. However, single-scale information entropy is difficult to reflect the complex changes of each subsequence after wind power decomposition. Therefore, refined composite multi-scale weighted permutation entropy (RCMWPE) is designed to measure the subsequences, so as to aggregate the subsequences more reasonably. Based on this, a method of combining SVMD-ICEEMDAN decomposition with RCMWPE (SIR method) is proposed to process wind power data.

[0006] In terms of feature extraction, some researchers have analyzed the spatio-temporal correlation of wind farm clusters based on directed graph networks and traditional correlation analysis. However, traditional correlations, such as Pearson correlation, are not friendly to the processing of extreme value correlations. Some scholars use the Copula structure to better simulate and show the extreme value data correlations existing in the original data, analyze the spatio-temporal characteristics of wind farms, and achieve better performance than traditional correlation analysis. However, graph networks mainly analyze the spatial relationships at close range, lack the spatial analysis between regions, and there is a problem of missing spatio-temporal associations that are important for prediction. The Moran index is a tool for effectively analyzing the spatial relationships between regions. For example, scholars such as Qu Zhijian used the Moran index to analyze the relationships between temperature, air pressure, etc. and irradiance in Xinjiang region of China to predict solar irradiance.

[0007] In summary, in the existing technologies, the prediction effect of the first decomposition high-frequency subsequence signal is poor, and the decomposition of the burst signals in complex time series is not good. The different bandwidths of each component will affect the estimation error. Secondly, it is impossible to accurately measure the complexity of each component, which is not conducive to improving the accuracy of wind power prediction. Summary of the Invention

[0008] Based on this, the purpose of the present invention is to provide a wind power probability prediction method to solve the deficiencies in the above-mentioned existing technologies.

[0009] The present invention provides a wind power probability prediction method, and the method includes:

[0010] Collect the geographical data, weather characteristic data, and historical characteristic data of wind farms in the area where the wind farm cluster is located;

[0011] Based on the geographical data, the weather characteristic data, and the historical characteristic data of wind farms, calculate the Moran index value of the area where the wind farm cluster is located, extract the spatial correlation characteristics between the areas where the wind farm cluster is located, and further extract the spatial characteristics by combining the graph network model of Copula. According to the Moran index value and the spatial characteristics, form the final input matrix;

[0012] Use successive variational mode decomposition for the wind power time series data of the historical characteristic data of wind farms to obtain preliminary components, and use adaptive noise complete ensemble empirical mode decomposition for the preliminary components to obtain several components;

[0013] Use refined composite multi-scale weighted permutation entropy to calculate the entropy values of several components, and use the mean method to comprehensively measure and analyze the complexity of several components to obtain several comprehensive components;

[0014] Input the final input matrix and several comprehensive components into the FAAM-Stacking model to obtain the final deterministic prediction result and verify the performance;

[0015] According to the error of the final deterministic prediction result and using the distributed kernel density estimation method, the wind power interval prediction result is obtained, and the wind power interval prediction result is verified.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: through the final input matrix composed of the Moran index value and the spatial features, the geographical, weather, and historical features of the wind farm cluster area can be effectively analyzed and extracted. Through successive variational mode decomposition and adaptive noise complete ensemble empirical mode decomposition for secondary decomposition, and aggregation through fine composite multi-scale weighted permutation entropy, the problems of poor decomposition and the influence of different bandwidths of each component on the estimation error are effectively avoided. By using the FAAM-Stacking model to predict the final deterministic prediction result and obtaining the wind power interval prediction result through the distributed kernel density estimation method, the prediction accuracy of the wind power can be effectively improved.

[0017] Further, the steps of calculating the Moran index value of the area where the wind farm cluster is located based on the geographical data, the weather feature data, and the historical feature data of the wind farm, extracting the spatial correlation features between the areas where the wind farm cluster is located, and further extracting the spatial features by combining the graph network model of Copula, and forming the final input matrix according to the Moran index value and the spatial features include:

[0018] Calculating the Moran index value of the area where the wind farm cluster is located based on the geographical data, the weather feature data, and the historical feature data of the wind farm to screen out the preliminary spatio-temporal features, where the calculation expression of the Moran index value is:

[0019] ;

[0020] ;

[0021] In the formula, represents the Moran index value, represents the total number, represents the element i and the element j the spatial weight between them, represents the deviation of the i th value from the average value, represents the deviation of the j th value from its average value, represents the aggregation of all spatial weights;

[0022] Build a graph network model of Copula, use the location relationship of the regions where the adjacent wind farm clusters are located to form a graph network, screen out the nodes that make up the graph network through Copula theory, and extract relevant spatio-temporal features;

[0023] Combine the preliminary spatio-temporal features and the relevant spatio-temporal features to obtain a new feature matrix, and reduce the dimension of the feature matrix through mutual information to obtain the final input matrix.

[0024] Further, the step of using successive variational mode decomposition for the electric power time series data of the historical feature data of the wind farm and using complete ensemble empirical mode decomposition with adaptive noise for the preliminary components to obtain several components includes:

[0025] Decompose the wind power time series signal in the wind power time series into L-order modes and a residual signal;

[0026] Apply a minimum constraint to the L-order modes and use a first preset filter to minimize the residual signal;

[0027] Use a second preset filter to establish a final constraint condition, and based on the final constraint condition, constrain the L-order modes to obtain preliminary components;

[0028] Add Gaussian white noise to the high-frequency subsequence signal of the successive variational mode decomposition to obtain a first residual value;

[0029] Define several modes generated by the complete ensemble empirical mode decomposition with adaptive noise, and calculate the local mean of the first residual value after adding Gaussian white noise by EMD to obtain a second residual value;

[0030] Repeat the calculation for all the modes and make the remaining residuals satisfy monotonicity to obtain several components.

[0031] Further, the expression for decomposing the wind power time series signal into L-order modes and a residual signal is:

[0032] ;

[0033] In the formula, represents the wind power time series signal, represents the L-order modes, represents the residual signal.

[0034] Further, the step of using refined composite multi-scale weighted permutation entropy to calculate the entropy values of several components and using the mean method to comprehensively measure and analyze the complexity of several components to obtain several comprehensive components includes:

[0035] Average the electric power time series data to construct several coarse-grained time series;

[0036] Perform phase space reconstruction on several of the coarse-grained time series to obtain the reconstructed time series;

[0037] Arrange the reconstructed time series in ascending order to obtain the arranged time series;

[0038] Calculate the weight values of several subsequences in the arranged time series;

[0039] Calculate the weighted probability values of the arranged time series in each arrangement mode based on the weight values of several of the subsequences, and calculate the average frequency of the arranged time series appearing with the weighted probability values in each arrangement mode;

[0040] Calculate the fine composite multi-scale weighted permutation entropy of the arranged time series to obtain multiple entropy values of several components;

[0041] Obtain the mean entropy of several of the components based on the mean method and the multiple entropy values.

[0042] Further, the expression for constructing several coarse-grained time series is:

[0043] ;

[0044] In the formula, represents the coarse-grained time series, represents the scale factor, represents the th scale factor, represents a positive integer, is an integer in represents [1, , represents the discrete time series.

[0045] Further, the steps of inputting the final input matrix and several of the comprehensive components into the FAAM-Stacking model to obtain the final deterministic prediction result and verify the performance include:

[0046] Input the comprehensive components into the multi-model fusion Multi-Stacking model according to different amplitudes to obtain several predicted values;

[0047] Input several of the predicted values into the feedback attention mechanism to obtain the final deterministic prediction result;

[0048] The final determined prediction result is verified by using the standard root mean square error, the mean absolute percentage error, and the coefficient of determination.

[0049] Furthermore, the expression for verifying the final determined prediction result is:

[0050] ;

[0051] ;

[0052] ;

[0053] In the formula, represents the standard root mean square error, represents the mean absolute percentage error, represents the coefficient of determination, represents the difference between the measured maximum value and the minimum value, represents the total number, represents the measured value, represents the predicted value of the wind power, represents the average value of the measured values, represents a positive integer.

[0054] Furthermore, the steps for verifying the wind power interval prediction result include:

[0055] The wind power interval prediction result is verified by using the prediction interval coverage rate and the average width of the prediction interval.

[0056] Furthermore, the expression for verification is:

[0057] ;

[0058] ;

[0059] In the formula, represents the prediction interval coverage rate, represents the average width of the prediction interval, represents the total number, represents a boolean quantity, represents the number of samples, represents the change range of the predicted target value, 、 respectively represent the upper boundary of the th prediction interval and the lower boundary of the th prediction interval. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a flowchart of the wind power probability prediction method in an embodiment of the present invention;

[0061] Figure 2 Schematic diagram of the FAAM-Stacking prediction model in the embodiments of the present invention;

[0062] Figure 3 Comparison chart of interval predictions of different models at a 90% confidence level in the embodiments of the present invention.

[0063] The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. Specific Embodiments

[0064] For ease of understanding the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided so that the disclosure of the present invention is thorough and complete.

[0065] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there may also be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0066] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0067] Please refer to Figure 1 , which shows the wind power probability prediction method in the embodiments of the present invention. The method includes steps S1 to S6:

[0068] S1, collect the geographical data, weather characteristic data and historical characteristic data of the wind farm cluster in the region;

[0069] It should be noted that geographical data of the area where the equipotential field is located and historical characteristic data of the wind farm are collected, including detailed longitude and latitude information of the regional boundary, weather data, and historical wind power data. In this embodiment, data of an electric field cluster in a certain area of China is selected for analysis, including longitude and latitude information of the boundaries of each county and district in the area, detailed historical weather information of 80 wind farms, and alternative characteristics of each wind farm including wind speed, wind direction, wind speed and wind direction at the hub height, temperature, air pressure, air humidity, and historical power generation of the wind turbine.

[0070] It is worth noting that using ArcGIS to collect geographical data of each county and district in a certain area can clearly distinguish the boundaries of each county and district;

[0071] Collect data of the wind farm cluster, the number of wind farms , weather and historical power data of each wind farm , where represents the dimensionality of the characteristic data of the wind farm, represents the temperature characteristic and the wind speed characteristic.

[0072] S2. Calculate the Moran's I value of the area where the wind farm cluster is located based on the geographical data, the weather characteristic data, and the historical characteristic data of the wind farm, extract the spatial correlation characteristics between the areas where the wind farm cluster is located, and further extract the spatial characteristics by combining the graph network model of Copula. A final input matrix is composed according to the Moran's I value and the spatial characteristics;

[0073] Specifically, the step S2 includes steps S21 to S23:

[0074] S21. Calculate the Moran's I value of the area where the wind farm cluster is located based on the geographical data, the weather characteristic data, and the historical characteristic data of the wind farm to screen out preliminary spatio-temporal characteristics. Among them, the calculation expression of the Moran's I value is:

[0075] ;

[0076] ;

[0077] In the formula, represents the Moran's I value, represents the total number, represents the element i and the element j the spatial weight between them, represents the i th value minus the average value, represents the j th value minus its average value, Represents the aggregation of all spatial weights;

[0078] It can be understood that the Moran's Index values between districts and counties within a region are calculated to preliminarily screen out preliminary spatio-temporal characteristics.

[0079] S22. Establish a Copula graph network model. Use the positional relationships of the regions where adjacent wind farm clusters are located to form a graph network. Through Copula theory, screen out the nodes that make up the graph network and extract relevant spatio-temporal characteristics. Among them, the mathematical expression of Copula theory is as follows:

[0080] Suppose are n random variables, and their marginal distributions are respectively Their joint distribution is , then there exists a function that "connects" the marginal distribution and the joint distribution

[0081]

[0082] According to the inverse transformation of the CDF of the marginal distribution, that is , then the expression form of the Copula function can be obtained:

[0083] .

[0084] It can be understood that a Copula graph network model is established. Using the positional relationships between adjacent wind farm nodes, a possible graph network is formed. Then, using Copula theory, suitable nodes are screened out to form a graph network and relevant spatio-temporal characteristics are extracted.

[0085] S23. Combine the preliminary spatio-temporal characteristics and the relevant spatio-temporal characteristics to obtain a new feature matrix, and perform mutual information dimensionality reduction on the feature matrix to obtain the final input matrix.

[0086] S3. Use successive variational mode decomposition for the wind power time series data of the historical characteristics of the wind farm to obtain preliminary components, and use complete ensemble empirical mode decomposition with adaptive noise for the preliminary components to obtain several components;

[0087] Specifically, the step S3 includes steps S31 to S36:

[0088] S31. Decompose the wind power time series signal in the wind power time series into L-order modes and a residual signal;

[0089] It can be understood that the wind power time series signal is decomposed into two parts, namely the L-order mode and the residual signal. Among them, the expression for decomposing the wind power time series signal into the L-order mode and the residual signal is:

[0090] ;

[0091] In the formula, represents the wind power time series signal, represents the L-order mode, represents the residual signal;

[0092] Among them, the residual signal includes the of the unprocessed part and the sum of the modes obtained previously . If is to be satisfied, then the L-order mode needs to be constrained.

[0093] S32. Minimize the constraint on the L-order mode and minimize the residual signal using the first preset filter;

[0094] It can be understood that the L-order mode realizes the minimization constraint, and the expression of the constraint condition is:

[0095] ;

[0096] In the formula, represents the operation of convolution, represents the center frequency of the L-order mode, represents the code number one of the constraint expression, represents the partial derivative with respect to time t, represents the Dirac function, represents the imaginary number, represents the base of the natural logarithm, represents time;

[0097] On the basis that the L-order mode has effective components, the residual signal should be minimized as much as possible. In order to ensure the stable implementation of this constraint, a suitable filter should be selected, and the frequency response expression is:

[0098] ;

[0099] The constraint that should be established, the expression is:

[0100] ;

[0101] In the formula, represents the frequency response, represents the frequency, represents the balance parameter, Code number two representing the constraint expression Represents a filter

[0102] S33, establish the final constraint condition using the second preset filter, and constrain the L-th order mode based on the final constraint condition to obtain the preliminary component

[0103] It can be understood that after the above two constraint conditions, the L-th order mode and the L-1 mode cannot be distinguished. Using the idea of constraint, select an appropriate filter, and the frequency response expression is the constraint idea, select an appropriate filter, and the frequency response expression is

[0104] ;

[0105] In the formula represents the impulse response of the filter represents an integer

[0106] Thus, a constraint can be established, and the expression is

[0107]

[0108] In the formula Code number three representing the constraint expression represents the i th impulse response

[0109] Thus, the final constraint is to ensure that the signal can be completely reconstructed, and the expression is

[0110]

[0111] In the formula represents the L-th order mode represents the unprocessed part of the signal

[0112] Therefore, the problem of extracting modal components can be considered as a problem of minimizing three constraints, and the expression is

[0113] ;

[0114] In the formula is to balance , , ;

[0115] After successive variational mode decomposition, the original wind power data is decomposed into 6 subsequences

[0116] S34, add Gaussian white noise to the high-frequency subsequence signal of successive variational mode decomposition to obtain the first residual value

[0117] It should be noted that the high-frequency components of the successive variational mode decomposition still have problems of instability and poor prediction effect. By using the complete ensemble empirical mode decomposition with adaptive noise algorithm through the local mean method, the problems of mode distortion and instability can be effectively solved. Specifically:

[0118] Add Gaussian white noise to the high-frequency subsequence signal of the successive variational mode decomposition. The expression is:

[0119] ;

[0120] In the formula, represents the signal after adding Gaussian white noise, represents the high-frequency subsequence signal of the successive variational mode decomposition, represents the standard deviation of the added Gaussian white noise, represents Gaussian white noise with both mean and variance equal to zero, is a function for calculating IMF1. Among them, the expression of the first residual value is:

[0121]

[0122] In the formula, represents the first residual value, represents the high-frequency sub-signal after one decomposition the total number of data, represents the IMF calculation function.

[0123] S35, define several modes generated by the complete ensemble empirical mode decomposition with adaptive noise, and calculate the local mean of the first residual value after adding Gaussian white noise by EMD to obtain the second residual value;

[0124] It should be noted that define K as the k-th mode generated by the complete ensemble empirical mode decomposition with adaptive noise. When k = 1, the first mode is obtained. The expression is:

[0125]

[0126] In the formula, represents the first mode;

[0127] When k = 2, calculate the local mean of after adding Gaussian white noise by EMD to obtain the second residual. The expression is:

[0128]

[0129] In the formula, represents the second residual value, represents the standard deviation of the added Gaussian white noise, Indicates the calculation of the second IMF function;

[0130] When Calculate the k-th residual and the k-th mode, and the expression is:

[0131]

[0132]

[0133] In the formula, Indicates the k-th residual, Indicates the (k - 1)-th residual, Indicates the standard deviation of the Gaussian white noise added on the basis of C_(k - 1), Indicates the calculation of the k-th IMF function, Indicates the k-th IMF value.

[0134] S36. Recursively calculate all the modes and make the remaining residuals satisfy monotonicity to obtain several components;

[0135] It can be understood that all IMFs are obtained by recursive calculation until the last remaining residual Satisfies monotonicity and cannot be decomposed any further. Finally, the original signal is decomposed into two parts, and the expression is:

[0136]

[0137] In the formula, Indicates the original signal, Indicates the i -th IMF value. Finally, the complete ensemble empirical mode decomposition with adaptive noise decomposes the highest frequency component into 14 new components again. Adding the remaining components of the successive variational mode decomposition, a total of 19 subsequences are obtained.

[0138] S4. Calculate the entropy values of several of the said components using the refined composite multi-scale weighted permutation entropy, and comprehensively measure and analyze the complexity of several of the said components using the mean method to obtain several composite components;

[0139] Specifically, the step S4 includes steps S41 to S47:

[0140] S41. Average the electric power time series data to construct several coarse-grained time series;

[0141] It should be noted that decomposing into 19 components for the second time increases the time complexity. Therefore, considering using the information entropy to measure the complexity of subsequences and aggregating subsequences to reduce the time complexity. Specifically, first, the electric power time series data within the window is averaged to construct several coarse-grained time series. The expression for constructing several coarse-grained time series is:

[0142] ;

[0143] In the formula, represents the coarse-grained time series, represents the scale factor, represents the th scale factor, represents a positive integer, is an integer in represents [1, , represents the discrete time series.

[0144] S42. Perform phase space reconstruction on several of the coarse-grained time series to obtain the reconstructed time series;

[0145] It can be understood that the expression for performing phase space reconstruction on the coarse-grained time series is:

[0146] ;

[0147] In the formula, represents the coarse-grained time series, represents the reconstructed time series, , ⋯, respectively represent the coarse-grained time series under , the coarse-grained time series under , represents the delay factor, represents the embedding dimension.

[0148] S43. Arrange the reconstructed time series in ascending order to obtain the arranged time series;

[0149] It can be understood that the expression for the arranged time series is:

[0150] ;

[0151] In the formula, , , , respectively represent the sorted coarse-grained time series under ​ The coarsened time series after sorting, at The coarsened time series after sorting, where represents the number of embeddings.

[0152] S44, calculate the weight values of several subsequences in the permuted time series;

[0153] Among them, the calculation expression for the weight value of each subsequence is:

[0154] ;

[0155] In the formula, represents coarsened time series, represents the mean of the time series, where The calculation expression of is:

[0156] .

[0157] S45, calculate the weighted probability value of the permuted time series in each permutation mode based on the weight values of several said subsequences, and calculate the average frequency of the occurrence of the weighted probability value of the permuted time series in each permutation mode;

[0158] It can be understood that the characteristic information of any subsequence is represented by the weight value and the permutation mode . For this time series, there are k permutation modes, then the expression of the weighted probability value of each permutation mode is:

[0159] ;

[0160] Among them, , , , respectively represent the first reconstructed component, the second reconstructed component, the th reconstructed component, represents the factorial of, represents the average frequency of the occurrence of the coarsened time series under the action of the permutation mode .

[0161] S46, calculate the refined composite multi-scale weighted permutation entropy of the permuted time series to obtain multiple entropy values of several components;

[0162] ​It can be understood that the expression for calculating the refined composite multi-scale weighted permutation entropy of the arranged time series is as follows:

[0163]

[0164] In the formula, represents the refined composite multi-scale weighted permutation entropy, represents the arranged time series, represents the average frequency of

[0165] S47. Based on the mean method and multiple entropy values, the mean entropy of several components is obtained;

[0166] It can be understood that the expression for obtaining the mean entropy of several components is as follows:

[0167]

[0168] AVR represents the mean entropy, represents the mean of the refined composite multi-scale weighted permutation entropy, represents the refined composite multi-scale weighted permutation entropy. In this embodiment, according to the similar mean entropy, the components are reconstructed, and finally 4 wind power subsequences with different amplitudes are obtained .

[0169] S5. Input the final input matrix and several comprehensive components into the FAAM-Stacking model to obtain the final deterministic prediction result and verify the performance;

[0170] Specifically, step S5 includes steps S51 to S53:

[0171] S51. Input the comprehensive components into the multi-model fusion Multi-Stacking model according to different amplitudes to obtain several prediction values;

[0172] It can be understood that the aggregated wind power subsequence is analyzed for amplitude, and according to different amplitudes, it is respectively input into the multi-model fusion Multi-Stacking model, as shown in Figure 2 to obtain the prediction values ;

[0173] S52. Input several prediction values into the feedback attention mechanism to obtain the final deterministic prediction result;

[0174] It can be understood that the prediction values are input into the feedback attention mechanism to obtain the final deterministic prediction result, and the expression is:

[0175] ;

[0176] In the formula, represents the predicted result, represents the first weight, represents the second weight, represents the third weight, represents the first amplitude, represents the second amplitude, represents the third amplitude;

[0177] It should be noted that the Multi-Stacking model with the back-feed attention mechanism is the FAAM-Stacking model.

[0178] S53. Verify the finally determined prediction result by using the standard root mean square error, mean absolute percentage error, and coefficient of determination;

[0179] It can be understood that the expression for verifying the finally determined prediction result is:

[0180] ;

[0181] ;

[0182] ;

[0183] In the formula, represents the standard root mean square error, represents the mean absolute percentage error, represents the coefficient of determination, represents the difference between the measured maximum and minimum values, represents the total number, represents the measured value, represents the predicted value of wind power, represents the average value of the measured values, represents an integer.

[0184] S6. According to the error of the finally determined prediction result and using the distributed kernel density estimation method, obtain the wind power interval prediction result and verify the wind power interval prediction result;

[0185] It can be understood that in specific implementation, first clarify the calculation formula of kernel density estimation. The calculation formula of kernel density estimation is:

[0186]

[0187] In the formula, represents the estimated probability density function, represents the kernel function, denotes taking the absolute value; denotes the indicator function, denotes a random event, n denotes the number of error samples, denotes the window width.

[0188] It should be noted that the distributed kernel density estimation method is based on kernel density estimation. For the prediction results of each component, different attentions are given to the probability density function, so as to reduce errors and achieve the purpose of improving wind power prediction. The expression of the distributed kernel density estimation method is:

[0189]

[0190]

[0191] In the formula, denotes the window width, denotes the th prediction error sample, denotes the weight coefficient of; denotes the kernel density expression with a bandwidth of ; denotes the expression of the distributed kernel density, and Q denotes the total number of bandwidths. denotes the sth window width, denotes an integer;

[0192] The expression of the said verification is:

[0193] ;

[0194] ;

[0195] In the formula, denotes the prediction interval coverage rate, denotes the average width of the prediction interval, denotes the total number, denotes a boolean quantity, denotes the number of samples, denotes the change range of the prediction target value, 、 respectively denote the upper boundary of the th prediction interval and the lower boundary of the th prediction interval, is the confidence level, The smaller it is, the better the sharpness of the prediction interval.

[0196] It should be noted that the specific evaluation result value is obtained from the prediction error through the above evaluation index formula, indicating the effectiveness of the present invention. The present invention is compared with four benchmark models, namely LSTM, CNN-LSTM, and CNN-BiGRU. A case study of a wind farm cluster in a certain autonomous region of a certain area in China is used for analysis. The comparison results are as Figure 3 shown. At a 90% confidence level, the prediction interval of the proposed method is the smallest, indicating that the effect of the present invention is the best.

[0197] In summary, the wind power probability prediction method in the above embodiments of the present invention can effectively analyze and extract the geographical, weather, and historical characteristics of the area where the wind farm cluster is located through the final input matrix composed of the Moran index value and spatial features. Through successive variational mode decomposition and adaptive noise complete ensemble empirical mode decomposition for secondary decomposition, and aggregation through fine composite multi-scale weighted permutation entropy, the problems of poor decomposition and the influence of different bandwidths of each component on the estimation error are effectively avoided. The final deterministic prediction result is predicted through the Multi-Stacking model with a feedback attention mechanism, and the wind power interval prediction result is obtained through the distributed kernel density estimation method, thereby effectively improving the prediction accuracy of wind power.

[0198] In the description of this specification, the description of reference terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0199] The above-described embodiments only represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention should be subject to the appended claims.

Claims

1. A wind power probability prediction method, characterized in that: The method comprises: Collect geographical data, weather characteristics data and historical characteristics data of the wind farm cluster area; Calculate the Moran index value of the area where the wind farm cluster is located based on the geographic data, the weather characteristic data and the historical characteristic data of the wind farm, extract the spatial correlation characteristics between the areas where the wind farm cluster is located, and further extract the spatial characteristics in combination with the Copula graph network model, and form a final input matrix according to the Moran index value and the spatial characteristics; The wind power time series data of the wind farm historical characteristic data is decomposed by successive variational mode to obtain preliminary components, and the preliminary components are decomposed by adaptive noise complete set empirical mode to obtain several components. The step specifically includes: Decomposing the wind power time series signal in the wind power time series into L-order modes and residual signals; Performing a minimization constraint on the L-order mode, and minimizing the residual signal using a first preset filter; A final constraint condition is established using a second preset filter, and the L-order mode is constrained based on the final constraint condition to obtain a preliminary component, wherein the final constraint is to ensure that the signal is completely reconstructed; Adding Gaussian white noise to the high-frequency subsequence signal of the successive variational mode decomposition to obtain a first residual value; Defining a plurality of modes generated by the adaptive noise complete set empirical mode decomposition, and calculating the local mean of the first residual value by EMD plus Gaussian white noise to obtain a second residual value; Repeat the calculation of all the modes and make the remaining residuals satisfy monotonicity to obtain several components; The entropy values ​​of the components are calculated by using a fine composite multi-scale weighted permutation entropy, and the complexity of the components is comprehensively measured and analyzed by using a mean value method to obtain a plurality of comprehensive components; Inputting the final input matrix and the plurality of the integrated components into a MultiStacking model with a feed-back attention mechanism, i.e., a FAAM-Stacking model, to obtain a final deterministic prediction result and verify the performance; A distributed kernel density estimation method is used according to the error of the final deterministic prediction result to obtain a wind power interval prediction result, and the wind power interval prediction result is verified.

2. The wind power probability prediction method according to claim 1, characterized in that: The step of calculating the Moran's index value of the area where the wind farm cluster is located based on the geographic data, the weather characteristic data and the historical characteristic data of the wind farm, extracting the spatial correlation features between the areas where the wind farm cluster is located, and further extracting the spatial features in combination with the Copula graph network model, and forming the final input matrix according to the Moran's index value and the spatial features includes: The Moran index value of the area where the wind farm cluster is located is calculated based on the geographic data, the weather characteristic data and the historical characteristic data of the wind farm to screen out preliminary spatiotemporal characteristics, wherein the calculation expression of the Moran index value is: ； ； In the formula, represents the Moran's index value, Indicates the total number, Representation elements i and elements j The spatial weight between Indicates i The deviation of a value from the mean, Indicates j The deviation of a value from its mean, represents the aggregation of all spatial weights; Establishing a Copula graph network model, using the positional relationship of the adjacent wind farm clusters to form a graph network, screening out the nodes constituting the graph network through Copula theory and extracting relevant spatiotemporal features; The preliminary spatiotemporal features and the related spatiotemporal features are combined to obtain a new feature matrix, and the feature matrix is ​​subjected to mutual information dimensionality reduction to obtain a final input matrix.

3. The wind power probability prediction method according to claim 1, characterized in that: The wind power time series signal is decomposed into L-order modes and residual signals as follows: ； In the formula, represents the wind power time series signal, represents the L-order mode, represents the residual signal.

4. The wind power probability prediction method according to claim 1, characterized in that: The step of using fine composite multi-scale weighted permutation entropy to calculate the entropy values ​​of the components and using the mean method to comprehensively measure and analyze the complexity of the components to obtain the multiple comprehensive components includes: Averaging the wind power time series data to construct a plurality of coarse-grained time series; Reconstructing the phase space of the plurality of coarse-grained time series to obtain a reconstructed time series; Arranging the reconstructed time series in ascending order to obtain an arranged time series; Calculating weight values ​​of a plurality of subsequences in the arranged time series; Calculating a weighted probability value of the arranged time series in each arrangement mode based on the weight values ​​of the plurality of subsequences, and calculating an average frequency of occurrence of the weighted probability value in the arranged time series in each arrangement mode; Calculating the refined composite multi-scale weighted permutation entropy of the permuted time series to obtain multiple entropy values ​​of several components; Based on the mean value method and the plurality of entropy values, mean entropies of the plurality of components are obtained, and the components are reconstructed according to similar mean entropies.

5. The wind power probability prediction method according to claim 4, characterized in that: The expression for constructing several coarse-grained time series is: ； In the formula, represents the coarse-grained time series, represents the scale factor, Indicates The scale factor, represents a positive integer, for Integers in It means [1, ], Represents a discrete time series.

6. The wind power probability prediction method according to claim 1, characterized in that: The step of inputting the final input matrix and the plurality of the comprehensive components into the FAAM-Stacking model to obtain the final deterministic prediction result and verify the performance comprises: Inputting the comprehensive components and the final input matrix into a multi-model fusion Multi-Stacking model according to different amplitudes to obtain a number of prediction values; Inputting the plurality of predicted values ​​into a feed-back attention mechanism to obtain a final deterministic prediction result; The final prediction results are verified using standard root mean square error, mean absolute percentage error and coefficient of determination; 7. The wind power probability prediction method according to claim 6, characterized in that: The expression for verifying the final prediction result is: ； ； ； In the formula, represents the standard root mean square error, represents the mean absolute percentage error, represents the coefficient of determination, It represents the difference between the maximum and minimum values ​​measured. Indicates the total number, Indicates the measured value, represents the predicted value of wind power, represents the average value of the measured values, Represents a positive integer.

8. The wind power probability prediction method according to claim 1, characterized in that: The step of verifying the wind power interval prediction result comprises: The prediction interval coverage and the prediction interval average bandwidth are used to verify the wind power interval prediction results.

9. The wind power probability prediction method according to claim 8, characterized in that: The expression of the verification is: ； ； In the formula, represents the prediction interval coverage, represents the average bandwidth of the prediction interval, Indicates the total number, Represents a Boolean quantity, represents the number of samples, Indicates the range of change of the predicted target value. , Respectively represent The upper boundary of the prediction interval, The lower boundary of the prediction interval.

Images