Short temporary wind speed extrapolation method based on decomposition algorithm and sample entropy

Abstract

The invention discloses a short temporary wind speed extrapolation method based on a decomposition algorithm and sample entropy, and the method comprises the steps: obtaining minute-level wind speed data of a construction point at different construction heights; according to the method, multi-layer minute-level wind speed historical observation data of a vertical gradient observation tower or minute-level wind speed historical observation data of an automatic meteorological station close to a construction point location are utilized, noise reduction is carried out on original wind speed data through an ICEEDMAN decomposition algorithm, high-frequency noise components are removed, and a wind speed sequence after noise reduction is generated; a wavelet packet decomposition algorithm is applied to the de-noised wind speed sequence for multiple times of decomposition, the complexity of each wind speed component is judged through sample entropy calculation to determine the number of decomposition layers, and a suitable wind speed component and a model are selected for modeling, so that a short temporary wind speed extrapolation model construction module is constructed; and generating a minute-level wind speed short-critical extrapolation result of the construction height in the future one hour. The method can effectively improve the extrapolation precision of the short temporary wind speed.

Description

Technical Field

[0001] The invention belongs to the field of electrical digital data processing, and in particular relates to a short-term wind speed extrapolation method based on a decomposition algorithm and sample entropy. Background Art

[0002] Modern numerical prediction models are the most important means of identifying, monitoring and warning of meteorological disasters. Although the resolution of the model has been greatly improved, its prediction ability (temporal and spatial precision and accuracy) under sudden weather processes is still insufficient and has large errors, and it is unable to accurately depict the temporal and spatial variation characteristics of wind; while users in the energy or engineering fields have very high requirements for the accuracy of wind speed.

[0003] The main purpose of time series analysis is to predict the future based on existing historical data. Commonly used models include moving average model (MA), autoregressive model (AR), autoregressive moving average model (ARMA), seasonal autoregressive moving average model (SARIMA) and random walk model (RW). With the development of artificial intelligence, machine learning (DT / SVM / LSTM) and deep learning methods are also gradually being used in time series extrapolation. The introduction of integrated algorithms has developed machine learning models with higher accuracy, providing new ideas for time series extrapolation. Traditional statistical methods do not adequately exploit the nonlinear characteristics of wind. When the wind is too volatile, the wind speed signal becomes unstable, resulting in low wind speed prediction accuracy, a sharp drop in prediction ability, and a short predictable time. In the past decade, with the rapid development of artificial intelligence technology and its widespread application in other fields, artificial intelligence technology can better explore the nonlinear change patterns of discontinuous variables compared to general statistical methods. Therefore, artificial intelligence technology has been applied to real-time prediction services in the wind power field (average wind speed and wind power prediction for every 15 minutes in the next 4 hours). Although the above method is good for predicting the average wind speed for 15 minutes, it does not involve higher-frequency wind speed predictions (such as 1 minute or 2 minutes), and still cannot solve specific problems in engineering operations services. Summary of the invention

[0004] The purpose of the present invention is to provide a short-term wind speed extrapolation method based on a decomposition algorithm and sample entropy. The method utilizes pure historical minute-level wind speed observation data from a vertical gradient tower or an automatic weather station, and through cluster analysis, signal decomposition and coupling with an artificial intelligence algorithm, as well as a wind profile power exponent algorithm based on wind direction and wind speed classification, can provide intelligent short-term wind speed predictions at different construction heights, with a prediction time of 1 hour.

[0005] The purpose of the present invention can be achieved through the following technical solutions: A short-term wind speed extrapolation method based on a decomposition algorithm and sample entropy, the method obtains minute-level wind speed data of a construction site at different construction heights, utilizes the multi-layer minute-level wind speed historical observation data of a vertical gradient observation tower or the minute-level wind speed historical observation data of an automatic weather station near the construction site, performs noise reduction on the original wind speed data through an ICEEDMAN decomposition algorithm, removes high-frequency noise components, generates a noise-reduced wind speed sequence, applies a wavelet packet decomposition algorithm to the noise-reduced wind speed sequence for N times, and the number of decompositions N is determined by the sample entropy of the high-frequency components obtained by the decomposition; determines the complexity of each wind speed component through sample entropy calculation, selects a suitable wind speed component and a model for modeling and prediction, constructs a short-term wind speed extrapolation model construction module, and generates a short-term wind speed extrapolation result of the future 1 hour at the construction height at the minute-level wind speed.

[0006] Specifically, a complex deep learning model is used to model the wind speed components with high sample entropy values, and an integrated machine learning algorithm is used to model the wind speed components with low sample entropy values. Finally, the extrapolated results of each component are superimposed to generate the short-term extrapolation results of minute-level wind speed.

[0007] Specifically, the short-term wind speed extrapolation model construction module includes a wind speed denoising submodule, a wind speed signal decomposition submodule, a wind speed component complexity evaluation submodule, a model construction submodule, a wind speed synthesis module, and an effect evaluation module.

[0008] Specifically, the wind speed synthesis module superimposes the prediction results of each component extrapolation model to complete the reconstruction of wind speed extrapolation and realize high-precision minute-level short-term wind speed extrapolation.

[0009] Specifically, when the construction site has real-time wind speed observation data at the construction height H0, and the time interval is consistent with the observation interval of the gradient tower, the data is used as input, and the short-term wind speed extrapolation model is applied for extrapolation to output the minute-level wind speed extrapolation result at the H0 height for the next hour.

[0010] Specifically, the applicable scope of the short-term wind speed extrapolation model is the area with a radius R centered on the gradient tower, and the elevation standard deviation in the area is not higher than 15 meters, and the wind speed correlation between any two points in the area is not less than 0.55.

[0011] Specifically, if the construction site does not meet the conditions within the applicable range of the short-term wind speed extrapolation model, the short-term wind speed extrapolation model is established using the historical minute-level wind speed observation data at an altitude of 10 meters from the nearest automatic weather station within 10 kilometers of the construction site.

[0012] Specifically, the short-term extrapolated value of the minute-level wind speed at a height of 10 meters at the meteorological station is calculated through the wind profile power exponential model to diagnose the future wind speed value at the construction height.

[0013] Specifically, the K-means clustering algorithm is used to cluster the wind speed samples at different height layers of the gradient tower to form stratified wind speed samples, and the short-term wind speed extrapolation model of the corresponding height interval layer is established.

[0014] The present invention also discloses a short-term wind speed extrapolation method based on a decomposition algorithm and sample entropy, the method specifically comprising: [1] Determine whether there is real-time wind speed observation data at the construction site at the construction height H0, and whether the time interval is consistent with the observation interval of the gradient tower; [2] When the construction site meets condition [1], determine whether the location of the construction site is within the applicable range of the short-term wind speed extrapolation model established by the vertical gradient tower layering; [3] When the construction site meets condition [2], the real-time observation at the construction site H0 height is used as the data input of the vertical short-term wind speed extrapolation model corresponding to the H0 height, and the model output data is the short-term wind speed extrapolation value at the H0 height in the next 1 hour; [4] When the construction site does not meet conditions [1] or [2], find the nearest meteorological station within 10 km of the construction site, use the historical wind speed observation data of the meteorological station to establish a short-term wind speed extrapolation model at a height of 10 m, and output the short-term extrapolated value of the future wind speed at a height of 10 m, that is, the short-term extrapolated value of the future wind speed at a height of 10 m at the construction site; [5] Under the condition of 4, the short-term extrapolated value of future wind speed at a height of 10 m is used as the input of the wind profile power exponential model based on the vertical gradient tower to diagnose the wind speed value of the corresponding time period at the required construction height, that is, the short-term extrapolated value of future wind speed at the height of the construction point; [6] When the construction point does not meet condition [2], determine whether it meets condition [4].

[0015] The beneficial effect of the above method is: the unstable wind speed signal is converted into a relatively stable wind speed signal through signal decomposition, and then extrapolated through the machine learning algorithm model, thereby improving the accuracy of short-term wind speed prediction. This method provides a fast and generalized method for short-term wind speed prediction in wind farms or engineering construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below in conjunction with the accompanying drawings.

[0017] Figure 1 It is a schematic diagram of the short-term wind speed extrapolation system based on the decomposition algorithm and sample entropy of the present invention; Figure 2 It is a schematic diagram of the workflow of the short-term wind speed extrapolation model construction module of the present invention; Figure 3 It is a flow chart of the short-term wind speed extrapolation method based on the decomposition algorithm and sample entropy of the present invention; Figure 4 Schematic diagram for determining the optimal number of layers by the elbow rule; Figure 5 A schematic diagram of using wavelet packets to decompose the original sequence and calculate the sample entropy of each component; Figure 6 This is a topographical diagram showing point distribution. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary operation and maintenance personnel in this field without creative work are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, the short-term wind speed extrapolation system based on decomposition algorithm and sample entropy of the present invention includes a quality control module, a classification wind profile power index module, a clustering and stratification module, a short-term wind speed extrapolation model construction module and an effect evaluation module.

[0020] 1. Quality control module The minute-level observation data are initially screened by controlling the reasonable range of the actual observed wind speed [0m / s, 75m / s) and the reasonable range of the wind direction observation value [0°, 360°). The data with poor observation quality and abnormal observation instrument periods are eliminated according to the quality control code of the monitoring instrument; the rigid value data that remains unchanged for 1 hour are eliminated, and the data quality of wind speeds above level 12 (V≥32.7m / s) is reconfirmed, and compared with the maximum wind speed of the meteorological stations upstream and downstream of the observation point to comprehensively determine whether it is abnormal sample data. For example, if the wind speed at the research point exceeds 3.5 times the average wind speed of the meteorological stations within a radius of 5km, it is considered an outlier and eliminated.

[0021] 2. Classification wind profile power index module The wind profile power index module is classified based on the combination of wind speed and wind direction. In this module, the quality control processing of wind speed in the vertical direction must meet the following four points at the same time, which is recorded as effective wind speed: the hourly average wind speed difference between two height layers with an interval of no more than 20m above 50m is less than 2.0m / s; the hourly average wind speed difference between two height layers with an interval of no more than 20m between 10m~50m is less than 3.0m / s; the hourly average wind speed difference between two height layers with an interval of no more than 20m above 30m is less than 2.0m / s; the wind direction difference of effective wind speed (3m / s~25m / s) between two height layers with an interval of no more than 20m above 30m is less than 22.5°.

[0022] The wind speed of each layer must be above 0.5m / s and the observations of each layer at the same time must be valid wind speeds in order to participate in the calculation of the wind profile power index. Taking the wind speed at a height of 10m as the benchmark, the wind speed frequency in each interval is counted with 1m / s as a wind speed interval. When counting the wind speed frequency in each wind speed interval, the intermediate value is used to represent it, such as using 1m / s wind speed to represent the wind speed in the interval [0.5m / s, 1m / s); through the wind speed frequency curve, the wind speed interval merging threshold is found, and the low-frequency wind speed segment with a frequency less than 5% is merged with the previous wind speed interval, so that the wind speed frequency value in the merged wind speed interval is equivalent to the wind speed frequency value of the previous interval, so that each wind speed division interval has statistical significance; in each wind speed interval, the wind direction is combined with the 8 directions for classification, and the wind profile power index α under each category is calculated by the least squares method (that is, the residual sum of squares of the wind speed at each point of the measured wind profile and the fitted wind profile is minimized). Finally, the wind speed at a certain height layer is calculated according to the power exponential formula based on the measured wind speed at the fitting reference point (10m height): , where α is the wind profile power exponent, ν1 is the wind speed at height Z1, in m / s, and ν2 is the wind speed at height Z2, in m / s.

[0023] When there is no wind direction reference in real-time observation during actual application, the average wind profile power index in each wind speed range can be used. The average wind profile power index α ave The weighted average of the wind profile power exponents in each wind direction within the wind speed range is obtained: , where N is the number of wind directions involved in the statistical calculation, f i is the wind direction frequency counted in the i-th wind direction, α i is the wind profile power index calculated in the i-th wind direction, when f i When it is less than 1%, it does not participate in the weight w i calculate.

[0024] 3. Clustering and Hierarchical Module The K-means clustering algorithm is used to cluster the observation sequences at different heights of the gradient tower in the same time period, and the wind speed samples at similar height layers are classified into one category, which are used as subsequent modeling samples to participate in the short-term model modeling. The established model can be used as a short-term extrapolation model for wind speed at similar height layers. For example, the clustering algorithm classifies the wind speed at heights of 10m, 20m, 40m, and 50m into one category, and the observation data at these heights are used to train the short-term extrapolation model for wind speed. The trained model can be used as a short-term extrapolation model for wind speed at heights of 10-50m.

[0025] Among them, when using the clustering algorithm, the K value of the clustering algorithm is determined by the elbow rule. The elbow rule is characterized by the cost function, which is characterized by the cost function of different K values. As the K value increases, the average distortion will continue to decrease and the number of samples contained in each class will decrease, so the sample will be closer to its center of gravity. However, as the value continues to increase, the improvement effect of the average distortion will continue to decrease. Therefore, it is more reasonable to find the K corresponding to the position where the distortion decreases the most in the process of increasing the K value. The cost function is the sum of the distortion degrees of each class and the sum of the squares of the distances between the positions of its internal members. The optimal solution is to minimize the cost function.

[0026] 4. Short-term wind speed extrapolation model construction module The flow chart of the short-term wind speed extrapolation model construction module is as follows: Figure 2 The short-term wind speed extrapolation model construction module includes: wind speed denoising submodule, wind speed signal decomposition submodule, wind speed component complexity evaluation submodule, model construction submodule, wind speed synthesis module, and effect evaluation module.

[0027] (1) Wind speed noise reduction submodule At each wind speed observation height, the ICEEDMAN decomposition algorithm is used to decompose the original wind speed sequence to obtain each modal component; the correlation coefficient between each modal component and the original wind speed sequence is calculated, and the correlation coefficient method is used for screening to remove the high-frequency components containing white noise with the lowest correlation coefficient, and then the remaining components are superimposed to regenerate the denoised wind speed sequence.

[0028] (2) Wind speed signal decomposition submodule The wavelet packet decomposition algorithm (WPD) is used to decompose the denoised wind speed sequence to form multiple component features, which is equivalent to feature enhancement. The number of WPD decompositions N is determined by the sample entropy of the high-frequency components obtained by decomposition. When the maximum value of the sample entropy (SE) of each high-frequency component after N (N=1, 2, 3...) decompositions is less than 0.6, the decomposition can be stopped.

[0029] (3) Wind speed component complexity evaluation submodule The sample entropy (SE) of the wind speed component is calculated to determine the complexity of the wind speed component sequence. The higher the sample entropy, the greater the complexity of the sequence. For each component sequence after WPD decomposition, in order to reduce the risk of error superposition caused by too many components in the prediction, appropriate components can be selected for modeling according to the sample entropy value of each component. When the difference between the sample entropy SE_SA of the low-frequency component SA obtained after the Nth decomposition and the sample entropy SE_S of the sequence S before its decomposition (N-1 times) is less than 0.1, the sequence S before decomposition is selected for modeling, instead of using the two new components SA and SD obtained after S decomposition, so as to reduce the number of components.

[0030] For a time series consisting of N data {x(n)}=x(1),x(2),x(N), the sample entropy SE is calculated as follows: A vector sequence of dimension m is formed according to the sequence number. ,in . m consecutive x values ​​starting from the i-th point.

[0031] Define the distance d[Xm(i),Xm(j)] between vectors Xm(i) and Xm(j) as the absolute value of the maximum difference between the corresponding elements of the two, that is: d [X m ( i ) ,X m ( j )] = max 0≤k≤m-1 (|x(i+k)-x(j+k)|) .

[0032] For a given Xm(i), count the number of j (1≤j≤N–m) whose distance between Xm(i) and Xm(j) is less than or equal to r, and record it as B i For 1≤i≤N–m, define for: .

[0033] Calculate Ratio: Calculate the ratio , which represents the probability of continuity when the dimension increases from m to m+1.

[0034] Calculate the final sample entropy value: , that is, taking the negative logarithm of the above ratio and taking the limit (when N→∞), we get the value of sample entropy. Alternatively, the limit value can be replaced by the average value of a series of data.

[0035] (4) Model construction submodule Among the wind speed components selected for model building, the appropriate model building is selected according to the size of the sample entropy value: the wind speed components with sample entropy greater than 0.2 are modeled and predicted using complex deep learning models, including but not limited to CNN-LSTM or Transformer-BiLSTM, and the wind speed components with sample entropy less than or equal to 0.2 are modeled and predicted using simple machine learning models, such as integrated machine learning algorithms, such as boost series Xgboost, Catboost, LightGBM, etc. Because the complex model has a large number of parameters, it is suitable for predicting complex component features; a simple machine learning model is used to predict low-frequency simple component features, because the low-frequency component features are relatively simple, and if a complex model is still used, overfitting is likely to occur, which is not effective.

[0036] By defining a window of appropriate size and using the window to move along the time series, the original data can be divided into multiple sub-windows, which are used as input samples for model training. Among them, the optimal sliding window can be selected by choosing different historical sample lengths N (N = 80, 100, 120, 140, 160, 180, 200, 220, 240...). By calculating the average relative error of wind speed extrapolation, the N value corresponding to the lowest average relative error is the optimal sliding window for model building.

[0037] (5)Wind speed synthesis module Add the wind speed component extrapolation results of the final extrapolation models of each component to complete the reconstruction of the predicted component to achieve high-precision minute-level short-term wind speed prediction.

[0038] (6)Effect evaluation module The effect evaluation module uses real-time observation data independent of the samples involved in model building as model input and tests the extrapolation effect. After comparing the requirements of business indicators, industry standards, and the extrapolation effects of other numerical extrapolation products using indicators such as absolute error, root mean square error, correlation coefficient, and high wind hit rate, it is comprehensively considered whether to adjust the model.

[0039] According to the correlation coefficient, root mean square error, and relative error, four (relatively poor, very poor, relatively good, and very good) comprehensive evaluation indicators are defined to evaluate and classify the extrapolation effects of test samples. First, calculate the correlation coefficient (Corr), relative error (RE), and root mean square error (RMSE) between extrapolation and observation for a single test sample. Second, calculate the average correlation coefficient (Corr_), average relative error (RE_), and average root mean square error (RMSE_) for all test samples. The conditions satisfied by each evaluation indicator are as follows: Very poor: All three conditions of RE > RE_, RMSE > RMSE_, and Corr < Corr_ are satisfied simultaneously Relatively poor: Two of the three conditions of RE > RE_, RMSE > RMSE_, and Corr < Corr_ are satisfied Relatively good: One of the three conditions of RE > RE_, RMSE > RMSE_, and Corr < Corr_ is satisfied Very good: None of the three conditions of RE > RE_, RMSE > RMSE_, and Corr < Corr_ are satisfied When in all time periods of the test samples, the extrapolation effects of 60% of the samples meet the relatively good or very good level, and the average root mean square error and average absolute error of the relatively good or very good samples are both below 1.5 m / s, the average relative error is below 0.35, and the average correlation coefficient is above 0.5, the test samples pass the evaluation and the model is not adjusted.

[0040] Another embodiment of the present invention discloses a short-term wind speed extrapolation method based on a decomposition algorithm and sample entropy. The method adopts a short-term wind speed extrapolation system based on a decomposition algorithm and sample entropy, and takes the wind speed at a construction height H0 of a certain construction point as an example. Figure 3 As shown: [1] Determine whether there is real-time wind speed observation data at the construction site at height H0, and whether the time interval is consistent with the observation interval of the gradient tower.

[0041] [2] When the construction site meets condition [1], determine whether the location of the construction site is within the applicable range of the short-term wind speed extrapolation model established by the vertical gradient tower layering.

[0042] Scope of application of the model: With the location of the vertical gradient observation tower as the center and a square area with a radius of R kilometers (R=1, 2, 3, 4, 5...), the elevation standard deviation within the area is calculated. When the elevation standard deviation within the square area with a radius of R kilometers is not higher than 15m, and the wind speed correlation between any two points with wind speed observation records in the area is not less than 0.55, it can be considered that the wind field in the area is similar, and any point in the area can use the short-term wind speed extrapolation model established based on the vertical gradient tower layering.

[0043] [3] When the construction site meets condition [2], the real-time observation at the H0 height of the construction site is used as the data input of the vertical short-term wind speed extrapolation model corresponding to the H0 height, and the model output data is the short-term wind speed extrapolation value at the H0 height in the next 1 hour.

[0044] [4] When the construction site does not meet conditions [1] or [2], find the nearest meteorological station within 10 km of the construction site, and use the historical wind speed observation data of the meteorological station to establish a short-term wind speed extrapolation model at a height of 10 m (the modeling method is consistent with the gradient tower multi-layer observation modeling); the input data of the model is the wind speed observation data of the meteorological station, and the output data is the short-term extrapolated value of the future wind speed at a height of 10 m. It can be considered that this extrapolated value can represent the short-term extrapolated value of the future wind speed at a height of 10 m where the construction site is located.

[0045] [5] When condition 4 is met, the short-term extrapolated value of future wind speed at a height of 10 m is used as the input of the wind profile power exponential model based on the vertical gradient tower to diagnose the wind speed value of the corresponding time period at the required construction height. It can be considered that the extrapolated diagnostic value can represent the short-term extrapolated value of future wind speed at the height of the construction point.

[0046] [6] When the construction point does not meet condition [2], determine whether it meets condition [4].

[0047] Specific application and result analysis The vertical wind observation data of a vertical gradient observation tower in Shenzhen in the past year were selected, of which the observation record heights totaled 12 floors (10m, 20m, 40m, 50m, 80m, 100m, 150m, 160m, 200m, 250m, 300m, 320m), and the observation record period was from May 2023 to May 2024. The wind speed data from May 2023 to March 2024 were used for short-term wind speed extrapolation modeling and calculation of wind profile power index based on wind direction and wind speed classification, and the data from April to May 2024 were used for short-term wind speed extrapolation model verification and testing.

[0048] The K-means clustering algorithm is used to cluster the 12-story wind speeds with observed records, and the elbow rule is used to confirm the K value, such as Figure 4 As shown in the figure, when K is 2, it is more reasonable. Therefore, the K-means algorithm is used to divide the 12-story wind speed into two categories. The wind speed sequence in the height range of 10-100m is one category, recorded as C1, and the wind speed sequence in the height range of 100-320m is another category, recorded as C2. Short-term wind speed extrapolation models can be established for C1 and C2 respectively, recorded as M1 and M2. Among them, M1 can make short-term extrapolation of minute-level wind speed at a height of 10-100m for service points within the applicable range of the model, and M2 can make short-term extrapolation of minute-level wind speed at a height of 100-320m for service points within the applicable range of the model.

[0049] Taking the wind speed at a height of 10m of a vertical gradient observation tower in Shenzhen as a benchmark, the wind speed is divided into 4 wind speed segments and the wind direction is divided into 8 directions according to the wind speed frequency statistical rules to calculate the wind profile power index. At the same time, the average wind profile power index in each wind speed interval is calculated based on the wind direction frequency statistical results.

[0050] The model construction process is introduced by taking the M2 model as an example. The wind speed sequences at heights of 150m, 160m, 200m, 250m, 300m, and 320m from May 2023 to March 2024 are used as training samples of the M2 model. The original wind speed sequence is decomposed using the wavelet packet decomposition algorithm, and the sample entropy of each layer of decomposed components is calculated, as shown in Figure 2. Figure 5 As shown in the figure, when the sample entropy of each component after decomposition is less than 0.6, the decomposition can be stopped. Therefore, the wavelet packet decomposition algorithm is finally used to perform a three-layer decomposition of the original wind speed series.

[0051] Selection of wind speed components for modeling: After three-layer WPD decomposition, by comparing the difference between the sample entropy of the low-frequency component after each round of decomposition and the sample entropy of the sequence before decomposition to see if it is greater than 0.1, a total of 6 wind speed components, aaa, ad, daa, dad, dda and ddd, are finally selected for modeling (while satisfying that the sample entropy of the selected components is within 0.6). Among them, the component aaa (sample entropy 0.19) with a sample entropy less than or equal to 0.2 is predicted using the Xgboost algorithm, while the components (ad, daa, dad, dda and ddd) with sample entropy greater than 0.2 are predicted using the complex artificial intelligence model CNN-LSTM.

[0052] Selecting an appropriate sliding window size can help the model better capture the characteristics of the time series. Select different window sizes of the sequence to participate in modeling, calculate the relative error of wind speed extrapolation, and the N value corresponding to the lowest relative error is the best sliding window for model modeling. The average relative error decreases with the increase of the window size. When the window size is 200 or more, the error decreases slowly. In order to extract longer dependencies in the sequence, select the historical sequence window size of 200 to participate in model building.

[0053] The denoised wind speed is used to predict the original wind speed after signal decomposition, which is equivalent to adding (enriching) the input features. The average root mean square error and relative error are lower than the prediction results of the original data without decomposition. The error results can be further reduced by using different models for combined prediction of the decomposed signals. The idea of ​​the combined prediction model is to use a complex model to predict the complex component characteristics of the data. Because the complex model has a large number of parameters (CNN-LSTM), it is suitable for predicting high-frequency complex component characteristics, but the low-frequency component characteristics are relatively simple. If a complex model is still used, it will be easy to overfit, but the effect is not good. Therefore, for the low-frequency simple component characteristics, we use a simple model (Xgboost) to predict, and then reconstruct the predicted components to achieve high-precision prediction.

[0054] Taking the M2 model as an example, the wind speed data from May 2023 to March 2024 (heights include: 150m, 160m, 200m, 250m, 300m, 320m) are used for modeling, and the wind speed at 150m height from April to May 2024 is used for model verification and testing. The above test and evaluation scenario is that there is real-time observation data of wind speed at a height of 150m (target height). The wind speed sequence at this height can be used as the input of the M2 model based on gradient tower modeling, and the short-term extrapolation of wind speed at a height of 150m can be directly completed. The experimental results show that the extrapolation effect of 64% of the test samples is at a relatively good level or above, with a root mean square error between 0.38-0.57m / s, an absolute error between 0.28-0.43m / s, a relative error between 0.08-0.15, and a correlation coefficient between 0.53-0.76.

[0055] like Figure 6 As shown, the circle is the vertical gradient tower Q, the triangle is the meteorological station P1, P2, P3, and the rectangular area (R = 2, 4, 6, 8km...). According to the defined model application range, the vertical gradient tower location Q is the center, and the square area with a radius of R kilometers (R = 1, 2, 3, 4, 5...) is used to calculate the elevation standard deviation in the area. When the elevation standard deviation in the square area with a radius of R kilometers is not higher than 15m, and the wind speed correlation between any two points with wind speed observation records in the area is not less than 0.55, it can be considered that the wind field in the area is similar, and any point in the area can use the short-term wind speed extrapolation model established based on the vertical gradient tower layering. In this experiment, according to the elevation standard deviation within the regional range, the short-term wind speed extrapolation model established based on the vertical gradient tower layering is applicable within its 6km range. When there are real-time wind speed observation data at other service points, they can be input into the model for future wind speed prediction.

[0056] When there is no real-time observation at the height of the target point where the short-term wind speed extrapolation is required, it is necessary to find a meteorological station near the point and use the historical wind speed observation data of the meteorological station (at a height of 10m) to model Mq. The modeling method is the same as above, and then the real-time wind speed of the meteorological station is used as the input of the model Mq. The output of the model is diagnosed to the target height through the wind profile power exponent module, so as to complete the short-term wind speed extrapolation at the target point height. To verify the extrapolation effect of the above process, the point where the vertical gradient tower is located is used as the target point, and the target height is 150m. The extrapolation result of the model Mq built based on the meteorological station P1 is tested to the accuracy of the wind speed prediction at a height of 150m through the power exponential formula of wind speed variation with height. The observation test of the meteorological station P1 is used at a height of 10m, and the observation test of the vertical gradient tower 150m is used at a height of 150m. The horizontal distance between P1 and the vertical gradient tower is 400m.

[0057] The test results show that the short-term wind speed at a height of 10m can be diagnosed to a height of 150m through the wind profile power index. According to the extrapolation results of the wind speed at a height of 150m, 53% of the test samples have an extrapolation effect of relatively good or above, with a root mean square error between 0.72-1m / s, an absolute error between 0.6-0.86m / s, a relative error between 0.19-0.26, and a correlation coefficient between 0.28-0.56. The overall extrapolation effect is good.

[0058] Taking the wind speed of the meteorological station as input, the short-term wind speed at a height of 10m is extrapolated through the model Mq built by the meteorological station P1. The test results show that 63% of the samples have a relatively good or above level of extrapolation effect, with a root mean square error between 0.3-0.48m / s, an absolute error between 0.21-0.35m / s, a relative error between 0.1-0.15, and a correlation coefficient between 0.54-0.77. The overall extrapolation effect is good. From the single extrapolation result, it can be seen that the short-term wind speed extrapolation diagnosed at a height of 150m is similar to the actual observation trend, and the wind speed value is equivalent, which can better extrapolate the future wind speed changes at high levels.

[0059] In the several embodiments provided by the present invention, it should be understood that the disclosed methods can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative.

[0060] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the basic features of the present invention.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary operation and maintenance personnel in the field should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A short-term wind speed extrapolation method based on decomposition algorithm and sample entropy, characterized in that: This method obtains minute-level wind speed data at different construction heights of the construction site, uses the multi-layer minute-level wind speed historical observation data of the vertical gradient observation tower or the minute-level wind speed historical observation data of the automatic weather station near the construction site, and uses the ICEEDMAN decomposition algorithm to denoise the original wind speed data, remove high-frequency noise components, and generate a denoised wind speed sequence. The denoised wind speed sequence is decomposed multiple times using the wavelet packet decomposition algorithm, and the complexity of each wind speed component is calculated by sample entropy. The appropriate wind speed component and model are selected for modeling, and a short-term wind speed extrapolation model construction module is constructed to generate the short-term extrapolation result of the minute-level wind speed at the construction height for the next hour.

2. The method according to claim 1, characterized in that A complex deep learning model is used to model the wind speed components with high sample entropy values, and an integrated machine learning algorithm is used to model the wind speed components with low sample entropy values. Finally, the extrapolated results of each component are superimposed to generate the short-term extrapolation results of minute-level wind speed.

3. The method according to claim 1, characterized in that The short-term wind speed extrapolation model construction module includes a wind speed denoising submodule, a wind speed signal decomposition submodule, a wind speed component complexity evaluation submodule, a model construction submodule, a wind speed synthesis module, and an effect evaluation module.

4. The method according to claim 3, characterized in that The wind speed synthesis module superimposes the prediction results of each component extrapolation model.

5. The method according to claim 1, characterized in that When the construction site has real-time wind speed observation data at the construction height H0, and the time interval is consistent with the observation interval of the gradient tower, the data is used as input and the short-term wind speed extrapolation model is applied for extrapolation to output the minute-level wind speed extrapolation value at the H0 height for the next hour.

6. The method according to claim 1, characterized in that The applicable scope of the short-term wind speed extrapolation model is the area with a radius R centered on the gradient tower, and the elevation standard deviation in the area is not higher than 15 meters, and the wind speed correlation between any two points in the area is not less than 0.

55.

7. The method according to claim 6, characterized in that If the construction site does not meet the conditions within the applicable range of the short-term wind speed extrapolation model, the short-term wind speed extrapolation model shall be established using the historical minute-level wind speed observation data at an altitude of 10 meters from the nearest automatic weather station within 10 kilometers of the construction site.

8. The method according to claim 1, characterized in that The short-term extrapolated value of minute-level wind speed at a height of 10 meters of the meteorological station is calculated through the wind profile power exponential model to diagnose the future wind speed value at the construction height.

9. The method according to claim 1, characterized in that: The K-means clustering algorithm is used to cluster the wind speed samples at different height layers of the gradient tower to form stratified wind speed samples, and the short-term wind speed extrapolation model of the corresponding height interval layer is established.

10. A short-term wind speed extrapolation method based on a decomposition algorithm and sample entropy, the method specifically comprising: [1] Determine whether there is real-time wind speed observation data at the construction site at height H0, and whether the time interval is consistent with the observation interval of the gradient tower; [2] When the construction site meets condition [1], determine whether the location of the construction site is within the applicable range of the short-term wind speed extrapolation model established by the vertical gradient tower layering; [3] When the construction site meets condition [2], the real-time observation at the construction site H0 height is used as the data input of the vertical short-term wind speed extrapolation model corresponding to the H0 height, and the model output data is the short-term wind speed extrapolation value at the H0 height in the next 1 hour; [4] When the construction site does not meet conditions [1] or [2], find the nearest meteorological station within 10 km of the construction site, use the historical wind speed observation data of the meteorological station to establish a short-term wind speed extrapolation model at a height of 10 m, and output the short-term extrapolated value of the future wind speed at a height of 10 m, that is, the short-term extrapolated value of the future wind speed at a height of 10 m at the construction site; [5] Under the condition of 4, the short-term extrapolated value of future wind speed at a height of 10 m is used as the input of the wind profile power exponential model based on the vertical gradient tower to diagnose the wind speed value of the corresponding time period at the required construction height, that is, the short-term extrapolated value of future wind speed at the height of the construction point; [6] When the construction point does not meet condition [2], determine whether it meets condition [4].

Images