Wind power prediction method based on multi-dimensional space-time wind speed feature fusion

Abstract

The invention belongs to the technical field of wind power prediction, and particularly relates to a wind power prediction method based on multi-dimensional space-time wind speed feature fusion. Through multi-dimensional space-time wind speed feature fusion, the space-time change rule of multiple parameters such as wind speed, wind direction, temperature, air pressure and the like is comprehensively considered, so that the capturing capability of the prediction model on the wind power change trend is improved, the feature fusion mode not only overcomes the limitation of single-dimensional data prediction, but also enhances the generalization capability of the model, and the prediction efficiency is improved. The prediction result is closer to the actual value, and the prediction precision is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power prediction, and particularly relates to a wind power prediction method based on the fusion of multi-dimensional spatio-temporal wind speed characteristics. Background Art

[0002] With the increasing global awareness of environmental protection and the profound transformation of the energy structure, wind energy, as a clean and renewable natural energy, is gradually becoming an important part of the global energy supply. Its rich reserves and wide distribution characteristics make the development and utilization of wind energy have great potential and broad prospects. However, the randomness and intermittency inherent in wind energy pose significant challenges to the operation and scheduling of wind farms and the stable operation of power grids.

[0003] To ensure the safe and reliable operation of the power system and improve energy utilization efficiency, it is particularly important to accurately predict wind power. Currently, various prediction methods have been developed in the field of wind power prediction. These methods have improved the prediction accuracy to a certain extent, but there are still many deficiencies.

[0004] Most of the existing wind power prediction methods are based on a single mathematical model or simple feature extraction techniques. When dealing with complex and variable wind energy data, these methods often fall short, and there are obvious limitations in prediction accuracy and adaptability. Specifically, these methods often ignore the influence of wind speed characteristics near the wind farm during the prediction process and do not fully utilize the spatio-temporal correlation of wind speed data. At the same time, there is also a lack of in-depth analysis and extraction of the interval change rules in historical wind speed data, which further limits the improvement of prediction accuracy.

[0005] To overcome the deficiencies of the existing methods and improve the accuracy of wind power prediction, the present invention proposes a wind power prediction method based on the fusion of multi-dimensional spatio-temporal wind speed characteristics. Summary of the Invention

[0006] To overcome the problems in the prior art, the present invention proposes a wind power prediction method based on the fusion of multi-dimensional spatio-temporal wind speed characteristics.

[0007] The technical solution of the present invention to solve the above technical problems is as follows:

[0008] The present invention provides a wind power prediction method based on the fusion of multi-dimensional spatio-temporal wind speed characteristics, including the following steps:

[0009] Step 100: Obtain data and preprocess the obtained data; wherein, the obtained data includes historical meteorological data of the location of the wind farm, historical wind power output data corresponding to the historical meteorological data, and historical wind speed data of the anemometer towers of other surrounding wind farms introduced.

[0010] Step 200: Extract multi-dimensional refined spatio-temporal wind speed features based on the preprocessed data. The multi-dimensional refined spatio-temporal wind speed features include wind speed quarterly features, wind speed monthly features, daily average wind speed maximum and minimum features, and monthly average wind speed maximum and minimum features;

[0011] Step 300: Select a prediction model according to the prediction period. If the prediction period is less than the preset period threshold, a short-term prediction model is used for wind power prediction; otherwise, a medium- and long-term prediction model is used for wind power prediction. The dung beetle optimization algorithm is used to optimize the parameter combinations of the short-term prediction model and the medium- and long-term prediction model respectively with the prediction accuracy as the goal, and the optimal parameter combinations optimized by the dung beetle optimization algorithm are obtained;

[0012] Step 400: Based on the optimal parameter combinations, use the short-term prediction model and the medium- and long-term prediction model to perform wind power prediction respectively, and correct the residuals of the prediction results respectively, so as to obtain the input feature combinations for training the final short-term prediction model and medium- and long-term prediction model;

[0013] Step 500: Obtain the input feature combinations for the period to be predicted, and predict the wind power using the trained medium- and long-term prediction model or short-term prediction model according to the prediction period and the optimal parameter combinations.

[0014] Further, the meteorological data includes wind speed, wind direction, temperature, humidity, and air pressure.

[0015] Further, in step 100, the obtained data is preprocessed, and the preprocessing includes missing value filling, data deduplication, and stationarity processing of the time series.

[0016] Further, in step 300, the medium- and long-term prediction model uses the FITS model.

[0017] Further, in step 300, when the FITS model is used for medium- and long-term wind power prediction training, in the optimization process of the dung beetle optimization algorithm, the parameter combination of the FITS model is optimized with the accuracy of each round of model prediction results as the objective function. The parameter combination includes learning rate, training batch, regularization parameter, number of iterations, and optimizer type. The highest accuracy prediction result and its corresponding parameter combination are recorded in each round, and the optimal parameter combination is retained.

[0018] Further, in step 300, the short-term prediction model uses the Cat Boost model.

[0019] Further, in step 300, when the Cat Boost model is trained for short-term wind power prediction, during the optimization process of the dung beetle optimization algorithm, the parameter combination of the Cat Boost model is optimized with the accuracy of the prediction result of each round as the objective function. The parameter combination includes the number of trees, the maximum depth of the trees, the learning rate, the regularization coefficient, and the minimum number of leaf nodes. The prediction result with the highest accuracy and its corresponding parameter combination are recorded in each round, and the optimal parameter combination is retained.

[0020] Further, in step 400, based on the optimal parameter combination, the medium and long-term prediction model is used to perform wind power prediction respectively, and the prediction result is corrected for residuals, including:

[0021] Based on the optimal parameter combination, the FITS model performs medium and long-term wind power prediction to generate the corresponding medium and long-term wind power prediction result;

[0022] Calculate the residual between the medium and long-term wind power prediction result and the real wind power data;

[0023] Using the grey relational analysis method, calculate the correlation degree between the input features and the residuals, and select the feature combination with a high correlation degree with the residuals as the new input feature combination;

[0024] Using the new input feature combination, re-use the FITS model for prediction, and superimpose the prediction result with the residual result calculated previously.

[0025] Further, in step 400, based on the optimal parameter combination, the short-term prediction model is used to perform wind power prediction respectively, and the prediction result is corrected for residuals, including:

[0026] Based on the optimal parameter combination, the Cat Boost model performs short-term wind power prediction to generate the corresponding short-term wind power prediction result;

[0027] Calculate the residual between the short-term wind power prediction result and the real wind power data;

[0028] Using the grey relational analysis method, calculate the correlation degree between the input features and the residuals, and select the feature combination with a high correlation degree with the residuals as the new input feature combination;

[0029] Using the new input feature combination, re-use the Cat Boost model for prediction, and superimpose the prediction result with the residual result calculated previously.

[0030] Compared with the prior art, the present invention has the following technical effects:

[0031] (1) The present invention solves the problems of insufficient prediction accuracy and poor stability existing in the existing wind power prediction technologies through multi-dimensional spatio-temporal wind speed feature fusion, adaptive multi-model fusion technology, etc.; it not only improves the prediction accuracy and stability, but also enhances the prediction efficiency, providing more accurate and reliable technical support for the scheduling and decision-making in the wind power generation industry;

[0032] (2) Through multi-dimensional spatio-temporal wind speed feature fusion, the present invention comprehensively considers the spatio-temporal variation laws of multiple parameters such as wind speed, wind direction, temperature, and air pressure, thereby improving the ability of the prediction model to capture the changing trend of wind power. This feature fusion method not only overcomes the limitations of single-dimensional data prediction, but also enhances the generalization ability of the model, making the prediction results closer to the actual values and significantly improving the prediction accuracy;

[0033] (3) By combining the advantages of multiple prediction models and dynamically adjusting and selecting models according to real-time data and prediction scenarios, the present invention realizes the complementary advantages between different models, improving the stability and reliability of the prediction results;

[0034] (4) The present invention uses the dung beetle optimization algorithm, which can automatically iterate the most reasonable dynamic parameter combinations of different time series lengths according to model training, avoiding the simplification of parameters under different time series length predictions and enabling the model to play the greatest role in the task;

[0035] (5) The method of residual correction by means of grey relational analysis in the present invention can dynamically evaluate the correlation strength between each input feature and the target variable, optimize the feature combination accordingly, eliminate redundant features and supplement key information; in addition, through residual correction, this method can also effectively eliminate the deviation between the prediction result and the actual value, improving the accuracy and reliability of the prediction. The flexibility and universality of this method enable it to show significant advantages in various prediction scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0037] Figure 1 is the flow diagram of the present invention;

[0038] Figure 2 is the comparison between the four-hour wind power prediction curve and the real wind power curve of the present invention;

[0039] Figure 3Comparison between the predicted wind power curve and the actual wind power curve of the present invention for one day

[0040] Figure 4 Comparison between the predicted wind power curve and the actual wind power curve of the present invention for four days Specific embodiments

[0041] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following describes in detail the specific embodiments, structures, features and effects of the technical solutions proposed according to the present invention in combination with the accompanying drawings and preferred embodiments. The specific features, structures or characteristics in one or more embodiments can be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs.

[0042] In one embodiment of the present invention, refer to Figure 1 , a wind power prediction method based on multi-dimensional spatio-temporal wind speed feature fusion is provided, including the following steps:

[0043] Step 100: Obtain data and preprocess the obtained data; wherein, the obtained data includes historical meteorological data of the location of the wind farm, historical wind power output data corresponding to the historical meteorological data, and historical wind speed data of the anemometer towers of other surrounding wind farms introduced.

[0044] Step 200: Based on the preprocessed data, extract multi-dimensional refined spatio-temporal wind speed features, and the multi-dimensional refined spatio-temporal wind speed features include wind speed quarterly features, wind speed monthly features, daily average wind speed maximum and minimum features, and monthly average wind speed maximum and minimum features.

[0045] Step 300: Select a prediction model according to the prediction period. If the prediction period is less than the preset period threshold, a short-term prediction model is used for wind power prediction; otherwise, a medium- and long-term prediction model is used for wind power prediction; the dung beetle optimization algorithm is used to optimize the parameter combinations of the short-term prediction model and the medium- and long-term prediction model respectively with the prediction accuracy as the goal, and the optimal parameter combinations optimized by the dung beetle optimization algorithm are obtained.

[0046] Step 400: Based on the optimal parameter combinations, use the short-term prediction model and the medium- and long-term prediction model to perform wind power prediction, and respectively correct the residuals of the prediction results to obtain the input feature combinations for training the final short-term prediction model and medium- and long-term prediction model.

[0047] Step 500: Obtain the input feature combination of the period to be predicted, and predict the wind power using the trained medium- and long-term prediction model or short-term prediction model according to the prediction period and the optimal parameter combination.

[0048] The following elaborates on each of the above steps in detail:

[0049] Step 100: Obtain data and preprocess the obtained data; among them, the obtained data includes historical meteorological data of the location of the wind farm, historical wind power output data corresponding to the historical meteorological data, and historical wind speed data of the anemometer towers of other surrounding wind farms introduced.

[0050] As an example, this step may include the following steps:

[0051] Step 110: Obtain meteorological station meteorological data of the prediction day, wind power at each time point of each day during holidays and the previous N days, history of the location of this wind farm, historical wind power output data corresponding to the historical meteorological data, historical wind speed data of the surrounding wind farms, and historical meteorological data of the surrounding wind farms to form an initial data set.

[0052] Obtain meteorological station meteorological data of the prediction day: Obtain the meteorological data of the prediction day, including key meteorological elements such as wind speed, wind direction, temperature, humidity, and air pressure.

[0053] Wind power at each time point of each day during holidays and the previous N days: Collect the wind power output data at each time point of each day within the past N days of the wind farm. Pay special attention to the data during holidays because factors such as wind power demand and grid dispatching during holidays may be different from those on weekdays and have an impact on wind power prediction.

[0054] Historical meteorological data of the location of this wind farm: Obtain the real meteorological data of the location of the wind farm over a period of time in the past, including wind speed, wind direction, temperature, humidity, etc., for analyzing the relationship between historical weather meteorological data and wind power output.

[0055] Real wind speed data of the surrounding wind farms in the past 4 years: Obtain the wind speed data of the past 4 years from the anemometer towers of other surrounding wind farms. Introduce the historical wind speed data of the anemometer towers of other surrounding wind farms to supplement the wind speed data of the location of the wind farm, which helps the model better capture the spatial distribution characteristics of the wind speed and improve the prediction accuracy.

[0056] Step 120: Preprocess the obtained data, and the data preprocessing includes missing value filling, data deduplication, and stationarity processing of the time series.

[0057] Missing value filling: Fill in the data missing due to reasons such as acquisition equipment, and use interpolation method and Backfill fitting to generate the missing data, including but not limited to the filling of numerical data, the filling of time type data, etc. Among them, Backfill fitting is a data preprocessing method mainly used to handle missing values in the data set.

[0058] Data deduplication: Identify and delete duplicate rows or records in the data caused by reasons such as data collection devices.

[0059] Stationarity processing of time series: The wind speed time series data fluctuates greatly and often has non-stationarity. Convert the non-stationary sequence into a stationary sequence to learn the fluctuation period and peak value in the stationary state of historical data.

[0060] The difference method is a commonly used method for stationary transformation of non-stationary time series. The difference operation can eliminate the trend or seasonal components in the time series, thereby making the series stationary.

[0061] The specific steps of the difference method are as follows: First-order difference: Calculate the difference between the observed values of two adjacent time points, that is, X_diff(t) = X(t) - X(t - 1). This method is used to remove the linear trend.

[0062] Step 200: Based on the preprocessed data, extract multi-dimensional refined spatio-temporal wind speed features, and the multi-dimensional refined spatio-temporal wind speed features include wind speed quarterly features, wind speed monthly features, daily average wind speed maximum and minimum features, and monthly average wind speed maximum and minimum features.

[0063] Feature engineering is an indispensable part of machine learning and occupies a very important position in the field of machine learning. The main role of feature engineering is to extract features from the original data to improve the training effect and performance.

[0064] The features processed by the feature engineering in this embodiment include:

[0065] (1) Add the wind speed features of the previous prediction period to enhance the meteorological influence of adjacent two days;

[0066] (2) Calculate the average wind speed of each day and find the maximum and minimum values of the average wind speed among all days;

[0067] Statistical law of historical wind speed range: The daily average wind speed varies between [7.19, 8.47] m / s, and the daily variation of the average wind speed shows a single-peak type, with the peak value approximately appearing around [20, 22] o'clock and the trough value appearing around [11, 13] o'clock;

[0068] (3) Calculate the average wind speed of each month by grouping by month, and find the maximum and minimum values of the average wind speed for each group of data; The monthly average wind speed at the hub height varies in the range of [5.85, 9.03] m / s, and the wind speed is relatively large from November to April of the following year, and relatively small from May to October;

[0069] (4)Collect the wind speed data of the anemometers with similar geographical locations in the past 4 years. For the data of each anemometer, calculate the wind speed characteristics annually, such as the annual average wind speed, the maximum and minimum wind speeds, etc. Consider the wind speed of each year as an independent wind speed characteristic.

[0070] After feature analysis and feature engineering, it is necessary to normalize the data to eliminate the dimensional differences between different features, thereby making the training of the algorithm more convenient. The normalization methods include but are not limited to maximum-minimum normalization, mean normalization, etc.

[0071] The maximum-minimum normalization formula is as follows:

[0072]

[0073] In the formula, a i is the normalized value, a is the value to be normalized, a min is the minimum value of the feature column where a is located in the data, and a max is the maximum value of the feature column where a is located in the data.

[0074] During the normalization process, the maximum and minimum values need to be recorded, and the inverse normalization operation needs to be performed on the subsequent prediction results.

[0075] Step 300: Select a prediction model according to the prediction period. If the prediction period is less than the preset period threshold, use a short-term prediction model for wind power prediction; otherwise, use a medium-long-term prediction model for wind power prediction; use the dung beetle optimization algorithm to optimize the parameter combinations of the short-term prediction model and the medium-long-term prediction model respectively with the prediction accuracy as the goal, and obtain the optimal parameter combinations optimized by the dung beetle optimization algorithm.

[0076] The short-term prediction model uses the Cat Boost model, and the medium-long-term prediction model uses the FITS model for model training; and the dung beetle optimization algorithm (DBO) is introduced. Use the dung beetle optimization algorithm to optimize the parameter combinations of the Cat Boost model and the FITS model respectively with the prediction accuracy as the goal, and finally use the optimal parameter combinations optimized by the dung beetle optimization algorithm to train the Cat Boost and FITS models respectively.

[0077] The network structure of the Cat Boost model is mainly composed of a series of boosting trees, which are continuously optimized during the iteration process to capture the complex relationships in the data. Cat Boost can automatically process categorical features and avoid cumbersome feature engineering. Through strategies such as ordered boosting and gradient steps, it further optimizes the model training process. Each tree node contains splitting conditions for dividing data subsets to minimize the prediction error. This structure enables Cat Boost to efficiently and accurately handle various prediction tasks.

[0078] The FITS model only uses about 10k parameters, which makes it very suitable for running on edge devices and provides possibilities for various application scenarios. By interpolating the frequency representation of the provided segments to generate an extended time series segment, the FITS model is efficient and accurate in processing time series data, especially in occasions that require medium- and long-term prediction and anomaly detection.

[0079] Using the dung beetle optimization algorithm can find the optimal parameter combination of the Cat Boost model and the FITS model by simulating the unique behavior of dung beetles. This process not only imitates the process of dung beetles rolling dung balls but also incorporates the characteristics of time series data and model complexity.

[0080] When training the FITS model for long-term wind power prediction, the dung beetle optimization algorithm optimizes the parameter combination of the FITS model with the accuracy of the model prediction result in each round as the objective function. The parameter combination includes the learning rate, training batch, regularization parameter, number of iterations, and optimizer type. Record the prediction result with the highest accuracy and the corresponding parameter combination in each round, and finally retain the optimal solution, that is, the optimal parameter combination.

[0081] Among them, for long-term prediction tasks, due to the long prediction time window and relatively high data stability, using the optimal FITS parameter combination of human experience as the initial array for optimization can improve the search efficiency, increase the population size N, reduce the search rolling radius coefficient k to improve the fineness of each search, reduce the step size b to reduce the dung beetle position update amplitude, and increase the maximum number of iterations (T_max) to reduce the rolling speed, so as to have more time to evaluate and adjust the parameters and conduct a more refined search after finding a better solution. It should be noted that in this embodiment, long-term refers to four days and medium-term refers to one day.

[0082] For medium-term prediction tasks, the prediction time window is shorter than that of long-term, and the data stability is a little worse than that of long-term. Here, in order to further improve the optimization efficiency, directly use the relatively stable long-term prediction FITS parameter combination obtained as the initial value, reduce the population N, increase the rolling radius coefficient k, increase the search step size b, and reduce the maximum number of iterations (T_max) to improve the search randomness and efficiency of the optimal combination on the basis of stability.

[0083] When training the Cat Boost model for short-term wind power prediction, the parameter combination of the Cat Boost model is optimized with the accuracy of the prediction result of each round as the objective function. The parameter combination includes the number of trees, the maximum depth of the tree, the learning rate, the regularization coefficient, and the minimum number of leaf nodes. The result with the highest accuracy and the corresponding parameter combination are recorded in each round, and finally the optimal solution, that is, the optimal parameter combination, is retained. Among them, the short-term prediction can be 4 hours.

[0084] Since the prediction time window is the shortest, the stability of the data is poor due to the influence of seasons and sudden weather, and the randomness of the prediction result accuracy is relatively large. The parameter combination of the initial Cat Boost model is defined by human experience for short-term prediction. Since short-term prediction is relatively frequent, to improve the optimization efficiency, in the optimization process of the dung beetle optimization algorithm, it is necessary to further reduce the population size N compared with medium- and long-term prediction, increase the search rolling radius coefficient k, increase the step size b, increase the randomness of the dung beetle position update, and reduce the maximum number of iterations (T_max) to improve the rolling speed, so that the model can cope with sudden extreme data changes and quickly find a more suitable parameter combination for processing this kind of data.

[0085] In this embodiment, the dung beetle optimization algorithm is introduced to achieve precise optimization of the parameters of the Cat Boost model and the FITS model. It not only simulates the unique behavior of dung beetles rolling dung balls, but also fully considers the characteristics of time series data and model complexity, providing an effective solution for medium- and long-term and short-term wind power prediction. In long-term prediction, using the stability of the data, through fine parameter adjustment and increasing the number of iterations, the optimization effect of the FITS model parameters is ensured; in medium-term prediction, on the basis of long-term prediction, by adjusting the search strategy and parameter combination, the optimization efficiency is improved; while in short-term prediction, more attention is paid to the flexibility and adaptability of the model. By reducing the number of iterations and increasing the search randomness, the Cat Boost model can quickly respond to extreme weather changes and find a more suitable parameter combination for processing short-term unstable data. This adaptive multi-model fusion wind power prediction method not only improves the prediction accuracy, but also provides strong support for the reliable operation and efficient management of wind power generation.

[0086] By constructing a prediction model and training and validating the model using historical data, high-precision prediction of future wind power is achieved. The prediction results can provide strong support for the operation and dispatching of wind farms and the stable operation of power grids. Transfer learning can effectively reduce the training time, enabling the training of a model with better performance even when using a small sample size and reducing the occurrence of overfitting.

[0087] The accuracy formula is as follows:

[0088]

[0089] In the above formula, VALUE capacity is the capacity of the fan unit, VALUE pred is the predicted value, VALUE true is the true value, E i is the relative error predicted at each time point. N is 16 for short-term, 96 for medium-term, and 4 * 96 for long-term (calculated based on one point every 15 minutes); accuracy is the precision.

[0090] Step 400: Based on the optimal parameter combination, use the short-term prediction model and the medium- and long-term prediction model to perform wind power prediction, and respectively correct the residuals of the prediction results to obtain the input feature combination for training the final short-term prediction model and medium- and long-term prediction model.

[0091] As an example, this step may include the following steps:

[0092] Step 410: Use the optimal parameter combination optimized by the dung beetle optimization algorithm and the Cat Boost model to perform short-term wind power prediction to generate the corresponding short-term prediction result; also use the optimal parameter combination optimized by the dung beetle optimization algorithm and the FITS model to perform medium- and long-term wind power prediction to generate the corresponding medium- and long-term wind power prediction results.

[0093] Step 420: Calculate the residuals between the short-term prediction result, the medium- and long-term wind power prediction result and the true wind power data respectively.

[0094] The residual reflects the deviation between the model prediction and the actual situation and is the key information for optimizing the prediction accuracy.

[0095] Step 430: Use the grey relational analysis method to calculate the correlation degree between the input features and the residuals; select the feature combination with a high correlation degree with the residuals as the new input feature combination.

[0096] Grey relational analysis is an effective tool for evaluating the closeness of the relationship between different variables, especially suitable for dealing with situations where information is incomplete or the data volume is limited. Based on grey relational analysis, select the feature combination with a high correlation degree with the residuals as the new input feature combination. These feature combinations are considered to have greater potential for improving the prediction results.

[0097] Step 440: Use the new feature combination to re-use the Cat Boost model and the FITS model for prediction, and superimpose the prediction results with the residual results calculated previously to improve the accuracy of the prediction results and enhance the robustness of the model.

[0098] This method of residual correction aims to obtain a wind power prediction result closer to the actual situation by correcting the bias in the model prediction. The whole process not only combines the advantages of the Cat Boost model and the FITS model, but also further improves the accuracy and reliability of the prediction through residual correction and grey relational analysis.

[0099] Step 500: Obtain the input feature combination for the time period to be predicted, and use the trained medium- and long-term prediction model or short-term prediction model to predict the wind power according to the prediction period and the optimal parameter combination.

[0100] Experimental results:

[0101] As can be seen from Table 1, this table shows the average accuracy of wind power prediction of different models in different prediction periods at a meteorological station during a specific test date (from October 26, 2023 to November 4, 2023). Table 1 lists eight models: LGBM, LinearRegression, ridge, Crossformer, Cat Boost, FITS, DBO + Cat Boost + residual correction, and DBO + FITS + residual correction, and compares their performances in three different prediction periods of 4 hours, 1 day, and 4 days.

[0102] Table 1 Performances of different models in different prediction periods of different time cycles

[0103]

[0104] In the 4-hour prediction period, the DBO + Cat Boost + residual correction model performed the best, with an average accuracy of 0.8710. Followed by the LGBM model, with an average accuracy of 0.853. The FITS model and the DBO + FITS + residual correction model also performed relatively well, reaching average accuracies of 0.8432 and 0.8496 respectively. In contrast, the Linear Regression and Crossformer models performed slightly worse in this prediction period.

[0105] In the 1-day prediction period, the DBO + FITS + residual correction model performed the most outstandingly, with an average accuracy of 0.8751, significantly higher than other models. The performances of the Cat Boost and FITS models followed closely, with average accuracies of 0.8296 and 0.8612 respectively. The LinearRegression, ridge, and Crossformer models performed relatively poorly in this prediction period.

[0106] In the 4-day prediction period, the DBO+FITS+residual correction model continued to maintain the leading position, with an average accuracy of 0.7998. The DBO+Cat Boost+residual correction model also showed good prediction ability, with an average accuracy of 0.7840. In contrast, the average accuracies of other models were lower than those of these two models.

[0107] Generally speaking, the DBO+FITS+residual correction model performed the best in the 1-day and 4-day prediction periods, while the DBO+Cat Boost+residual correction model performed the best in the 4-hour prediction period. The performance differences of different models in different prediction periods may reflect their prediction abilities and adaptabilities at different time scales.

[0108] According to Figure 2 、 Figure 3 and Figure 4 the comparison information of the wind power prediction curve (pre) and the true wind power curve (true) shown, the vertical axis represents the wind power, and the horizontal axis is the time axis with a 15-minute interval for each point. Referring to Figures 2 - 4 , a comprehensive analysis of the model performance in different prediction periods can be carried out. Using the DBO+Cat Boost+residual correction model in the 4-hour prediction period, the prediction model can better grasp the change trend of wind power, but there are still deviations. Using the DBO+FITS+residual correction model in the 1-day prediction period, the prediction accuracy has been improved, especially when the wind power is stable. Using the DBO+FITS+residual correction model in the 4-day prediction period, due to the large time span and the increase of uncertain factors, there is a large gap between the prediction and the actual situation, especially at the wind power peak.

[0109] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A wind power prediction method based on multi-dimensional spatiotemporal wind speed feature fusion, characterized in that: The following steps are involved: Step 100: Acquire data and pre-process the acquired data; wherein the acquired data includes historical meteorological data of the location of the wind farm, historical wind power output data corresponding to the historical meteorological data, and historical wind speed data of wind towers of other surrounding wind farms; Step 200: extracting multi-dimensional refined spatiotemporal wind speed features based on the preprocessed data, wherein the multi-dimensional refined spatiotemporal wind speed features include wind speed quarterly features, wind speed monthly features, daily average wind speed maximum and minimum features, and monthly average wind speed maximum and minimum features; Step 300: Select a prediction model according to the prediction period. If the prediction period is less than a preset period threshold, use the short-term prediction model to predict wind power; otherwise, use the medium- and long-term prediction model to predict wind power; use the dung beetle optimization algorithm to optimize the training parameters of the short-term prediction model and the medium- and long-term prediction model with prediction accuracy as the goal, and obtain the optimal parameter combination after optimization by the dung beetle optimization algorithm; Step 400: Based on the optimal parameter combination, the short-term prediction model and the medium- and long-term prediction model are used to respectively predict the wind power, and residual correction is performed on the prediction results respectively, so as to obtain the input feature combination for training the final short-term prediction model and the medium- and long-term prediction model; Step 500: Obtain an input feature combination for the time period to be predicted, and predict wind power using a trained medium- and long-term prediction model or a short-term prediction model according to the prediction period and the optimal parameter combination.

2. The method for wind power prediction based on multi-dimensional spatiotemporal wind speed feature fusion according to claim 1 is characterized in that: The meteorological data include wind speed, wind direction, temperature, humidity and air pressure.

3. The method for wind power prediction based on multi-dimensional spatiotemporal wind speed feature fusion according to claim 1 is characterized in that: In step 100, the acquired data is preprocessed, and the preprocessing includes missing value filling, data deduplication and time series stationarity processing.

4. The wind power prediction method based on multi-dimensional spatiotemporal wind speed feature fusion according to claim 1 is characterized in that: In step 300, the medium- and long-term prediction model adopts the FITS model.

5. The method for wind power prediction based on multi-dimensional spatiotemporal wind speed feature fusion according to claim 4 is characterized in that: In step 300, when the FITS model is conducting medium- and long-term wind power prediction training, during the optimization process of the dung beetle optimization algorithm, the parameter combination of the FITS model is optimized with the accuracy of the model prediction results of each round as the objective function. The parameter combination includes learning rate, training batch, regularization parameter, number of iterations, and optimizer type. The highest accuracy prediction result and its corresponding parameter combination are recorded in each round, and the optimal parameter combination is retained.

6. The method for wind power prediction based on multi-dimensional spatiotemporal wind speed feature fusion according to claim 1 is characterized in that: In step 300, the short-term prediction model adopts the Cat Boost model.

7. The method for wind power prediction based on multi-dimensional spatiotemporal wind speed feature fusion according to claim 6 is characterized in that: In step 300, when the Cat Boost model is used for short-term wind power prediction training, in the optimization process of the dung beetle optimization algorithm, the parameter combination of the Cat Boost model is optimized with the accuracy of the prediction results of each round of the model as the objective function, and the parameter combination includes the number of trees, the maximum depth of the tree, the learning rate, the regularization coefficient and the minimum number of leaf nodes. The prediction result with the highest accuracy and its corresponding parameter combination are recorded in each round, and the optimal parameter combination is retained.

8. The method for wind power prediction based on multi-dimensional spatiotemporal wind speed feature fusion according to claim 5 is characterized in that: In step 400, based on the optimal parameter combination, the medium- and long-term prediction models are used to respectively perform wind power prediction, and residual correction is performed on the prediction results, including: Based on the optimal parameter combination, the FITS model performs medium- and long-term wind power forecasts and generates corresponding medium- and long-term wind power forecast results; Calculate the residual between the medium- and long-term wind power forecast results and the actual wind power data; The grey correlation analysis method is used to calculate the correlation between the input features and the residuals, and the feature combination with the highest correlation with the residuals is selected as the new input feature combination; Using the new input feature combination, the FITS model is reused for prediction, and the prediction results are superimposed with the residual results calculated previously.

9. The method for wind power prediction based on multi-dimensional spatiotemporal wind speed feature fusion according to claim 8, characterized in that: In step 400, based on the optimal parameter combination, the short-term prediction model is used to respectively perform wind power prediction, and residual correction is performed on the prediction results, including: Based on the optimal parameter combination, the Cat Boost model performs short-term wind power forecasting and generates corresponding short-term wind power forecasting results; Calculate the residual between the short-term wind power prediction result and the actual wind power data; The grey correlation analysis method is used to calculate the correlation between the input features and the residuals, and the feature combination with the highest correlation with the residuals is selected as the new input feature combination; Using the new input feature combination, the Cat Boost model is reused for prediction, and the prediction results are superimposed with the residual results calculated previously.

Images