Wind power prediction method and system based on meteorological characteristics and deep learning model

Abstract

The invention relates to a wind power prediction method and system based on meteorological characteristics and a deep learning model, belongs to the technical field of power systems, and solves the problem of high calculation cost caused by complexity of an existing model structure and parameter selection. The method comprises the following steps: acquiring meteorological data through a multi-source data fusion technology and a real-time data acquisition technology; preprocessing the meteorological data; constructing time features, original features, statistical features under a time window and combined features to generate a training set and a test set; constructing an XGBoost model and an LSTM neural network model, training by using the training set and the test set to generate a first wind power prediction model and a second wind power prediction model, and optimizing model parameters by using a Bayesian optimization model; and inputting the real-time feature data set into the first and second wind power prediction models to respectively generate a first prediction value and a second prediction value, and obtaining a combined wind power prediction value by using an error reciprocal method. Through automatic feature selection and parameter adjustment, calculation cost and dependence on professional knowledge are reduced.

Description

Technical Field

[0001] The present application relates to the technical field of power systems, and in particular to a method and system for predicting wind power based on meteorological characteristics and deep learning models. Background Art

[0002] In the current new energy field, wind power prediction technology plays a vital role in improving the intelligence level of power grid dispatching, optimizing power resource allocation, and ensuring the stability and reliability of the power grid. Although existing technologies have achieved certain results in wind power prediction, there are still several problems that need to be solved.

[0003] Existing technical solutions mainly include physical methods based on numerical weather forecasting, statistical methods that rely on historical data, spatial correlation methods that use geographic location information, and prediction models that apply advanced algorithms such as deep learning. Although these methods have their own advantages, they also have some different limitations. For example, physical methods are highly dependent on the quality and acquisition cost of meteorological data; statistical methods are limited by the need for a large amount of historical data and may ignore the complexity of meteorological factors; the applicability of spatial correlation methods is limited by the similarity and distribution characteristics of geographic locations; although deep learning methods can handle complex nonlinear relationships, the complexity of their model structure and parameter selection leads to high computational costs and dependence on professional knowledge. Summary of the invention

[0004] In view of the above analysis, the embodiments of the present application aim to provide a wind power prediction method and system based on meteorological characteristics and deep learning models, so as to solve the problem that the complexity of the existing model structure and parameter selection leads to high computing cost and dependence on professional knowledge.

[0005] On the one hand, an embodiment of the present application provides a wind power prediction method based on meteorological characteristics and a deep learning model, characterized in that it includes: obtaining meteorological data related to wind power through multi-source data fusion technology and real-time data acquisition technology, wherein the meteorological data includes real-time meteorological data and historical meteorological data; preprocessing the meteorological data, wherein the data preprocessing includes outlier processing and normalization processing; constructing a historical feature data set or a real-time feature data set based on the preprocessed historical meteorological data or real-time meteorological data, wherein the historical feature data set or the real-time feature data set both include time features, original features, statistical features under a time window and Combined features, wherein the historical feature data set and the corresponding wind power are divided into a training set and a test set; an XGBoost model and an LSTM neural network model are constructed and the training set and the test set are used to train the XGBoost model and the LSTM neural network model to generate a first wind power prediction model and a second wind power prediction model, respectively, and the model parameters are optimized using a Bayesian optimization model; the real-time feature data set is input into the first wind power prediction model and the second wind power prediction model, respectively, to generate a first prediction value and a second prediction value, respectively, and then the inverse error method is used to obtain the combined wind power prediction value.

[0006] The beneficial effects of the above technical solution are as follows: This application proposes a wind power prediction method based on meteorological characteristics and deep learning models. The purpose of improving the deep learning model is to improve the accuracy and efficiency of wind power prediction by deeply exploring meteorological characteristics and optimizing model parameters. This method can not only automatically screen meteorological characteristics that have a greater impact on wind power, but also reduce the prediction cost by reducing the demand for computing resources. In addition, this application reduces the dependence on professional knowledge through automated feature selection and parameter adjustment processes, making the model easier to deploy and apply.

[0007] Based on the further improvement of the above method, obtaining meteorological data through multi-source data fusion technology and real-time data acquisition technology further includes: obtaining multi-source historical meteorological data from different meteorological stations, satellite images and historical wind power databases; fusing the multi-source historical meteorological data through the following weighted average method formula to generate fused multi-source meteorological data:

[0008]

[0009] in, is the fused multi-source meteorological data, w i is the weight of the ith data source, x iis the data of the i-th data source; and using the Internet of Things device to collect key meteorological data in real time as real-time meteorological data, wherein the key meteorological data includes wind speed, wind direction, temperature, air pressure and humidity.

[0010] The beneficial effects of the above technical solution are as follows: Multi-source data fusion technology and real-time data acquisition technology are combined to ensure that the model can predict wind power based on comprehensive and up-to-date data. Multi-source data fusion provides a comprehensive historical and environmental background by integrating data from multiple sources such as meteorological stations, satellite images and historical wind power data, while real-time data acquisition technology obtains key meteorological parameters in real time through IoT devices to capture the latest meteorological changes. This combination not only improves the comprehensiveness and timeliness of the data, but also enhances the adaptability and prediction accuracy of the model to different meteorological conditions, so that the wind power prediction model can consider both historical trends and immediate changes, thereby more accurately predicting wind power output.

[0011] Based on further improvements of the above method, using IoT devices to collect key meteorological data in real time further includes: deploying anemometers, wind vanes, temperature sensors, barometers and humidity sensors in wind farms; sampling the anemometers, wind vanes, temperature sensors, barometers and humidity sensors to respectively collect wind speed, wind direction, temperature, air pressure and humidity in real time, wherein the wind speed is used as the primary factor in determining wind power output, and its changes directly affect the energy efficiency of wind turbines; the wind direction is used to determine whether the wind wheel can effectively capture wind energy; the temperature and air pressure are used for their changes to act on air density, thereby affecting the wind speed and the power coefficient of the wind turbine.

[0012] Based on the further improvement of the above method, the preprocessing of the meteorological data further includes: when the key meteorological data collected in real time is missing multiple continuous data, directly deleting; when the key meteorological data collected in real time is missing non-multiple continuous data, calculating the i-th missing data by the following average value formula:

[0013] x i =(x i-1 +x i+1 ) / 2;

[0014] Among them, x i is the i-th missing data; x i-1 and x i+1 is the data before and after the missing value. If x i+1 If it is a missing value, then take x i+2 When the key meteorological data after outlier processing has different units, the key meteorological data after outlier processing is normalized by the following formula:

[0015] xa =(xx min ) / (x max -x min );

[0016] Among them, x and x a are the key meteorological data before and after normalization; x max and x min are the maximum and minimum values ​​in the key meteorological data sets, respectively.

[0017] Based on the further improvement of the above method, constructing a historical feature data set or a real-time feature data set based on the preprocessed historical meteorological data or real-time meteorological data further includes: the original features are wind speed, wind direction, temperature, humidity and air pressure features directly obtained from the preprocessed meteorological data:

[0018] Y=[Y1,Y2,…,Y n ] T ;

[0019] Among them, Y i =[y i1 ,y i2 , …, y iM ](i=1,2,…,n) is the feature of the sample at the i-th moment, y ik (k=1,2,…,M) is Y i The kth feature in , n is the number of samples, M is the number of meteorological features; the time feature is the time feature of the month, day, week, hour and minute of the year extracted from the time; the combination feature includes linear combination feature and nonlinear combination feature, which are respectively expressed by the following formulas: additive feature, subtractive feature, multiplicative feature and divisive feature:

[0020] f a =y ij +y ik ;

[0021] f is =y ij -y ik ;

[0022] f im =y ij ×y ik ;

[0023] f id =y ij ÷y ik ;

[0024] The statistical features are used to represent the data fluctuations within the time window:

[0025]

[0026] in, They are time windows [it w ,i+t w ] the mean, standard deviation and maximum value of the kth feature in t w is a time window parameter, indicating the number of moments in the time window; and is divided into the training set and the test set based on the time feature, the original feature, the statistical feature and the combined feature.

[0027] Based on the further improvement of the above method, optimizing the model parameters using the Bayesian optimization model further includes: selecting r feature combinations [p1, p2, ..., p r ] and calculate the corresponding loss function [L(p1),L(p2),…,L(p r )], where L(p s )(s=1,2,…,r) is the loss function corresponding to the sth feature combination, wherein the sample data set includes a training set and a test set; the sample parameters and the corresponding loss function are combined to obtain D r :

[0028] D r ={(p s ,L(p s )),s=1,2,…,r};

[0029] Where D r represents the data set collected after the rth iteration in the Bayesian optimization process; p s represents the sth parameter combination; L(p s ) indicates that in the parameter combination p s Below, the loss function value of the model;

[0030] It is estimated that the loss function follows a Gaussian distribution, [L(p1), L(p2), ..., L(p r )] is:

[0031]

[0032] in, represents a multivariate normal distribution, where 0 is the mean vector and K is the covariance matrix; K is an r×r matrix, where r is the number of parameter combinations evaluated; each element k(p i ,p j ) represents the parameter combination p i and p j The covariance of the loss function between ; the point with the maximum value is selected as the sampling point through the following function E(·) formula:

[0033]

[0034] Among them, P(·) is the probability density function, P(y|p s ) is p s The probability of y appearing when it appears, P(p s |y) is when y appears s The probability of occurrence, P(p s ) is p s The probability of occurrence; γ is the distribution proportion; l(p s ) is when y≤y * The distribution function when g(p s ) is when y>y * The probability density function at the time; according to the sampling point, the optimal parameter combination p is obtained new :

[0035] p new = argmaxE(p s );

[0036] Among them, argmaxE(p s ) represents finding the parameter that maximizes the mathematical expectation function; updating L(p based on the optimal parameter combination s )’s Gaussian distribution:

[0037]

[0038] k′=[k(p new ,p1),k(p new ,p2),…,k(p new ,p r )] T ;

[0039] According to the following formula, L(p new ) distribution P(L(p new |D r ,p new )):

[0040] P(L(p new |D r ,p new ))~N(μ,σ 2 );

[0041] μ=(k′) T K -1 [L(p1),L(p2),···,L(p r )] R ;

[0042] σ 2 =k(p new,p new )-(k′) T K -1 k′;

[0043] Where L(pi) represents the loss function value under the i-th parameter combination pi; pi represents the i-th parameter combination; p new represents a new parameter combination; K represents the covariance matrix of the loss function value of the known parameter combination; k′ represents the covariance vector between the new parameter combination and the known parameter combination; k(p new ,p new ) represents the covariance of the new parameter combination itself; P(L(p new |D r ,p new )) indicates that in a given data set D r and the new parameter combination p new Under the condition of , the probability distribution of the loss function value of the new sampling point; μ represents the predicted mean of the loss function value of the new sampling point; σ 2 Represents the predicted variance of the loss function value of the new sampling point; the sampling point selection step and the update L(p) are performed in a cyclic manner s ) until the maximum number of iterations is reached and the current optimal output point is calculated.

[0044] Based on the further improvement of the above method, constructing an XGBoost model and an LSTM neural network model and using the training set and the test set to train the XGBoost model and the LSTM neural network model to generate a first wind power prediction model and a second wind power prediction model respectively further includes: expressing the t-th round model and the corresponding objective function of the XGBoost model by the following formula:

[0045]

[0046] in, Represents the predicted value of the lth sample after the tth iteration; represents the predicted value of the lth sample after the t-1th iteration; f t (x i ) represents the model function of the i-th sample in the t-th iteration; represents the loss function of the tth iteration; g i represents the gradient boost weight of the i-th sample; h i represents the diagonal elements of the Hessian matrix of the i-th sample; Ω(f t ) represents the model function f in the tth iteration t The regularization term of ;

[0047] The LSTM neural network model is represented by the following formula:

[0048] f t =σ(W f [h t-1 ,x t ]+b f );

[0049] i t =σ(W i [h i-1 ,x t ]+b i );

[0050]

[0051] O t =σ(W o [h t-1 ,x t ]+b o );

[0052] h t =o t tanhC t ;

[0053] Among them, f t 、i t , O t are the forget gate output, input gate output, and output gate output at the current time t respectively; W f , b f are the weight and bias of the forget gate respectively; W i , b i are the weight and bias of the input gate respectively; W c , b c are the weight and bias of the updated value respectively; W o , b o are the weight and bias of the output gate respectively; σ(·) represents the Sigmoid activation function; x t , C t 、h t are the input, cell state and output value of the LSTM neural network unit at the current time t; C t-1 、h t-1 are the cell state at the previous moment and the output value of the LSTM neural network unit; [h t-1 ,x t ] means merging matrices with equal number of rows into columns; is the candidate cell state; “*” represents dot product; the XGBoost model and the LSTM neural network model are trained using the training set and the XGBoost model and the LSTM neural network model are tested using the test set to generate the first wind power prediction model and the second wind power prediction model respectively.

[0054] Based on the further improvement of the above method, the real-time feature data set is input into the first wind power prediction model and the second wind power prediction model to generate a first prediction value and a second prediction value respectively, and then the combined wind power prediction value is obtained by using the inverse error method. The method further includes: inputting the real-time feature data set into the first wind power prediction model and the second wind power prediction model to generate a first prediction value f respectively. 1t and the second predicted value f 2t ; The first prediction error ε1 and the second prediction error ε2 are calculated by the following formula:

[0055]

[0056] Where n is the number of data points, y i is the ith actual value, is the i-th predicted value; the first weight ω1 and the second weight ω2 are calculated by the following formula:

[0057]

[0058] The combined wind power forecast value is calculated by the following formula:

[0059] f t =ω1f 1t +ω2f 2t ,t=1,2,…,n。

[0060] On the other hand, an embodiment of the present application provides a wind power prediction system based on meteorological characteristics and a deep learning model, including: a data acquisition module, used to acquire meteorological data related to wind power through multi-source data fusion technology and real-time data acquisition technology, wherein the meteorological data includes real-time meteorological data and historical meteorological data; a preprocessing module, used to preprocess the meteorological data, wherein the data preprocessing includes outlier processing and normalization processing; a data set construction module, used to construct a historical feature data set or a real-time feature data set based on the preprocessed historical meteorological data or real-time meteorological data, wherein the historical feature data set or the real-time feature data set both include time features, original features, statistical features under a time window, and combined features. Characteristic, wherein the historical feature data set and the corresponding wind power are divided into a training set and a test set; a model construction training module, used to construct an XGBoost model and an LSTM neural network model and use the training set and the test set to train the XGBoost model and the LSTM neural network model to generate a first wind power prediction model and a second wind power prediction model respectively, wherein the model parameters are optimized using a Bayesian optimization model; a prediction module, used to input the real-time feature data set into the first wind power prediction model and the second wind power prediction model respectively to generate a first prediction value and a second prediction value respectively; a prediction value combination module, used to obtain a combined wind power prediction value using an inverse error method.

[0061] Based on the further improvement of the above system, the data acquisition module further includes: a data acquisition submodule, which is used to acquire multi-source historical meteorological data from different meteorological stations, satellite images and historical wind power databases; a data fusion submodule, which is used to perform data fusion on the multi-source historical meteorological data by the following weighted average method to generate fused multi-source meteorological data:

[0062]

[0063] in, is the fused multi-source meteorological data, w i is the weight of the ith data source, x i is the data of the i-th data source; and a data acquisition submodule, which is used to use the Internet of Things device to collect key meteorological data in real time as real-time meteorological data, wherein the key meteorological data includes wind speed, wind direction, temperature, air pressure and humidity.

[0064] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0065] 1. This application aims to improve the accuracy and efficiency of wind power forecasting by deeply exploring meteorological features and optimizing model parameters. This method can not only automatically screen meteorological features that have a greater impact on wind power, but also reduce the forecasting cost by reducing the demand for computing resources. In addition, this application reduces the reliance on professional knowledge through automated feature selection and parameter adjustment processes, making the model easier to deploy and apply.

[0066] 2. By deeply exploring meteorological characteristics and applying improved deep learning models combined with Bayesian optimization parameter adjustment technology, we can address the volatility and randomness problems in wind power forecasting, thereby enhancing the dispatching capacity and operational stability of the power system.

[0067] 3. This application innovatively combines the XGBoost and LSTM models, using XGBoost's ability to handle nonlinear relationships and classification problems, and LSTM's advantages in capturing long-term dependencies in time series. This model combination significantly improves the accuracy of forecasting changes in wind power, while maintaining the model's adaptability and flexibility to changes in various meteorological conditions. In addition, a probabilistic model of the objective function is established through Bayesian optimization technology to efficiently search for the optimal solution in the parameter space. Finally, through the model prediction weighted combination strategy of the inverse error method, the weights are dynamically adjusted according to the prediction error of each model, and the combined prediction results are optimized, making the final prediction results more stable and accurate, and improving the model's response speed and prediction accuracy to short-term fluctuations in wind power.

[0068] In the present application, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present application will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are only for the purpose of illustrating specific embodiments and are not to be considered as limiting the present application. The same reference symbols denote the same components throughout the accompanying drawings.

[0070] Figure 1 This is a flow chart of a method for predicting wind power based on meteorological characteristics and deep learning models according to an embodiment of the present application;

[0071] Figure 2 This is a short-term wind power prediction flow chart of an embodiment of the present application;

[0072] Figure 3 A structural diagram of an LSTM neural network according to an embodiment of the present application; and

[0073] Figure 4 This is a block diagram of a wind power prediction system based on meteorological characteristics and deep learning models according to an embodiment of the present application. DETAILED DESCRIPTION

[0074] The preferred embodiments of the present application are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of the present application and are used together with the embodiments of the present application to illustrate the principles of the present application, and are not used to limit the scope of the present application.

[0075] refer to Figure 1 A specific embodiment of the present application discloses a method for predicting wind power based on meteorological characteristics and a deep learning model, comprising: in step S101, obtaining meteorological data related to wind power through multi-source data fusion technology and real-time data acquisition technology, wherein the meteorological data includes real-time meteorological data and historical meteorological data, and both the real-time meteorological data and the historical meteorological data include wind speed, wind direction, temperature, air pressure and humidity; in step S102, preprocessing the meteorological data, wherein the data preprocessing includes outlier processing and normalization processing; in step S103, constructing a historical feature data set or a real-time feature data set based on the preprocessed historical meteorological data or real-time meteorological data, and both the historical feature data set or the real-time feature data set include Time features, original features, statistical features under time windows and combined features, wherein the historical feature data set or the real-time feature data set is divided into a training set and a test set; in step S104, an XGBoost model and an LSTM neural network model are constructed and the training set and the test set are used to train the XGBoost model and the LSTM neural network model to generate a first wind power prediction model and a second wind power prediction model, respectively, wherein the model parameters are optimized using a Bayesian optimization model; in step S105, the real-time feature data set is respectively input into the first wind power prediction model and the second wind power prediction model to generate a first prediction value and a second prediction value, respectively, and then the combined wind power prediction value is obtained using the inverse error method.

[0076] Compared with the prior art, in the wind power prediction method based on meteorological characteristics and deep learning models provided in this embodiment, the improved deep learning model aims to improve the accuracy and efficiency of wind power prediction by deeply mining meteorological characteristics and optimizing model parameters. This method can not only automatically screen meteorological characteristics that have a greater impact on wind power, but also reduce the prediction cost by reducing the demand for computing resources. In addition, this application reduces the dependence on professional knowledge through automated feature selection and parameter adjustment processes, making the model easier to deploy and apply.

[0077] The embodiment of the present application deeply mines meteorological features and applies an improved deep learning model combined with Bayesian optimization parameter adjustment technology to address the volatility and randomness problems in wind power forecasting, thereby enhancing the dispatching capability and operational stability of the power system. First, by mining the time window of meteorological features [it w , i+t w ] and the corresponding combined features plus the time feature; then the Bayesian optimization parameter is used to select the optimal time window parameter t w The number of features n of the feature combination um1 、n um2 , find the optimal time window and reduce feature redundancy; finally, a trained wind power prediction model is obtained.

[0078] In the following, reference Figures 1 to 3 , each step in the wind power prediction method based on meteorological characteristics and deep learning model according to the embodiment of the present application is described in detail.

[0079] refer to Figure 1 and Figure 2 In step S101, meteorological data related to wind power is obtained through multi-source data fusion technology and real-time data acquisition technology, wherein the meteorological data includes real-time meteorological data and historical meteorological data, and both the real-time meteorological data and the historical meteorological data include wind speed, wind direction, temperature, air pressure and humidity.

[0080] Specifically, obtaining meteorological data through multi-source data fusion technology and real-time data collection technology further includes: obtaining multi-source historical meteorological data from different meteorological stations, satellite images and historical wind power databases; fusing the multi-source meteorological data through the following weighted average method formula (1) to generate fused multi-source historical meteorological data. Using IoT devices to collect key meteorological data in real time further includes: deploying anemometers, wind vanes, temperature sensors, barometers and humidity sensors in wind farms; sampling anemometers, wind vanes, temperature sensors, barometers and humidity sensors to collect wind speed, wind direction, temperature, air pressure and humidity in real time, respectively, wherein wind speed is used as the primary factor determining wind power output, and its change directly affects the energy efficiency of wind turbines; wind direction is used to determine whether the wind wheel can effectively capture wind energy; temperature and air pressure are used to act on air density due to their changes, thereby affecting wind speed and the power coefficient of wind turbines.

[0081] Data collection. Multi-source data fusion and real-time data acquisition technology are introduced in the data collection step. In terms of multi-source data fusion, by integrating data from different meteorological stations, satellite images, historical wind power data and other sources, more comprehensive meteorological characteristics can be provided, thereby enhancing the model's adaptability to different meteorological conditions and prediction accuracy. For these multi-source data, the weighted average method is used to obtain the fused data, and the formula is as follows:

[0082]

[0083] Where: is the fused data, w i is the weight of the ith data source, x i is the data of the i-th data source.

[0084] In addition, in terms of real-time data collection, IoT devices are used to collect key meteorological parameters in real time. When building an accurate wind power prediction model, the five basic meteorological factors of wind speed, wind direction, temperature, air pressure and humidity are mainly considered. Therefore, anemometers, wind vanes, temperature sensors, barometers and humidity sensors are deployed in wind farms. This ensures the timeliness of data and improves the response speed and accuracy of the model to rapidly changing meteorological conditions. Wind speed is the primary factor that determines wind power output, and its changes directly affect the energy efficiency of wind turbines. Wind direction determines whether the wind rotor can effectively capture wind energy. Changes in temperature and air pressure act together on air density, which in turn affects wind speed and the power coefficient of wind turbines. Although humidity has a smaller impact, it cannot be ignored under extreme conditions. The actual wind power data and meteorological data of the wind farm for one and a half years are selected as training samples. The sample data includes time, wind speed, wind direction, temperature, pressure, humidity and actual power. The sampling interval of wind farm data is 15 minutes. The sample data is then divided into training and test sets in a ratio of 9:1, that is, a total of 47,328 sets of data in the first 493 days are used as training sets to train the model, and a total of 5,280 sets of data in the last 55 days are used as test sets to evaluate the model.

[0085] In step S102, the meteorological data is preprocessed, wherein the data preprocessing includes outlier processing and normalization processing. The meteorological data preprocessing further includes: when the key meteorological data collected in real time lacks multiple continuous data, it is directly deleted; when the key meteorological data collected in real time lacks more than multiple continuous data, the i-th missing data is calculated by the following average value formula (2); when the key meteorological data after outlier processing has different units, the key meteorological data after outlier processing is normalized by the following formula (3).

[0086] Data preprocessing: In the data preprocessing stage, in order to ensure the quality of the data and the effectiveness of model training, we adopted a series of automated methods to handle outliers and perform data normalization.

[0087] b1. Data outliers. In the process of data collection, due to the influence of various factors such as natural or human factors, some individual data may not conform to the actual situation or be missing. In order to restore the objective authenticity of the data and thus bring better analysis results, it is necessary to identify and process the abnormal data. Traditional outlier detection methods often require manual participation. By introducing automated statistical methods, we reduce manual intervention and improve the efficiency and accuracy of data processing. In order to improve the continuity of the collected data, the abnormal and missing data in the collected data are processed. This article regards outliers as missing values. Since wind power data changes continuously, the correlation between the previous and next two data is strong. If multiple consecutive data are not missing, the average value method is used to fill in the gaps; if multiple consecutive data are missing, the direct deletion method is adopted. The calculation formula is as follows:

[0088] x i =(x i-1 +x i+1 ) / twenty two)

[0089] Where: x i is the i-th missing data; x i-1 and x i+1 is the data before and after the missing value. If x i+1 If it is a missing value, then take x i+2 .

[0090] b2. Data normalization. When the data input into the model have different units, the numerical differences are often large. At the same time, the analysis is more difficult due to the different units. In order to minimize the differences between different data, all input data are normalized. The calculation formula is as follows:

[0091] x a =(xx min ) / (x max -x min ) (3)

[0092] Where: x and x a are the data before and after normalization; x max and x min are the maximum and minimum values ​​in the data set, respectively.

[0093] In step S103, a historical feature data set or a real-time feature data set is constructed based on the preprocessed historical meteorological data or real-time meteorological data, wherein the historical feature data set or the real-time feature data set both include time features, original features, statistical features under a time window, and combined features, wherein the historical feature data set and the corresponding wind power are divided into a training set and a test set.

[0094] Feature engineering. Features are the core factor that determines the accuracy of the prediction model. Therefore, it is necessary to deeply mine the meteorological features related to wind power and construct more effective features. n ] T Indicates that Y i =[y i1 ,y i2 , …, y iM ](i=1,2,…,n) is the feature of the sample at the i-th moment, y ik (k=1, 2, ..., M) is the kth feature in Yi, n is the number of samples, and M is the number of meteorological features. The following features are constructed.

[0095] 1) Original features: Refers to the basic information obtained directly from meteorological data, such as wind speed, wind direction, temperature, humidity and air pressure.

[0096] That is, Y = [Y1, Y2, ..., Y n ] T .

[0097] 2) Time features: Extract five time features from the time: the month of the year, the day of the month, the day of the week, the hour of the day, and the minute of the hour.

[0098] 3) Combination features. Including linear combination features and nonlinear combination features, used to mine linear and nonlinear relationships between features. Specifically include:

[0099] Additive features: f a =y ij +y ik

[0100] Subtractive characteristics: f is =y ij -y ik

[0101] Multiplicative characteristics: f im =y ij ×y ik

[0102] Except for sexual characteristics: f id =y ij ÷y ik

[0103] 4) Statistical features. Statistical features are used to represent the data fluctuations within a time window. This feature can be expressed as:

[0104]

[0105] Where: They are [it w , i+t w ]The mean, standard deviation and maximum value of the kth feature in the time window; t w is the time window parameter, which indicates the number of moments in the time window; the historical feature data set and the corresponding wind power are divided into a training set and a test set.

[0106] Compared with the prior art, the present application introduces finer-grained time features (such as days of the week and minutes) to enable the model to more accurately identify and predict seasonal and cyclical changes in wind power. This multi-dimensional time feature significantly improves the prediction model's ability to identify changes in wind power, thereby improving the prediction accuracy. In addition, the present invention reveals the complex relationship between different meteorological factors by constructing additive, subtractive, multiplicative and divisive combination features. These combined features enhance the model's understanding of the interaction between meteorological factors and improve the model's prediction accuracy for the complex nonlinear relationship between wind power and meteorological conditions. Finally, the present invention provides a quantitative description of data fluctuations by calculating statistical features such as the mean, standard deviation and maximum value within the time window. These statistical features enable the model to more accurately capture short-term fluctuations in wind power and improve the model's responsiveness to real-time changes and prediction accuracy.

[0107] In step S104, an XGBoost model and an LSTM neural network model are constructed and trained using a training set and a test set to generate a first wind power prediction model and a second wind power prediction model, respectively, wherein the model parameters are optimized using a Bayesian optimization model.

[0108] Build the model. In the model building stage, this application uses two advanced machine learning models, XGBoost and LSTM, and uses Bayesian optimization technology to tune the model parameters to improve prediction accuracy and generalization ability. The specific process is as follows:

[0109] First, after the original data is processed, feature variables are derived through feature engineering, and then the pre-processed data is divided into a test set and a training set. The data of the training set is first put into the XGBoost model and the LSTM neural network model for training. The XGBoost model integrates multiple decision trees and uses the gradient boosting method to optimize the model performance and improve the modeling ability of nonlinear relationships. Compared with a single decision tree, the integration method significantly improves the model's prediction accuracy and ability to capture complex data relationships. The principle and formula of the XGBoost model are as follows: For a data set containing n samples and m-dimensional features, a model containing K decision trees can be used express,

[0110]

[0111] F represents the function space composed of all tree models.

[0112] F={f (x) =w q(x)}(q:R m →T,w∈R T ) ↓ (8)

[0113] q represents the mapping relationship from x to leaf nodes, and w represents the weight on the leaf. The model objective function is defined as (9) and (10)

[0114]

[0115] T represents the number of leaf nodes, the penalty constant of the revenue function for splitting leaf nodes is γ, and the penalty function coefficient of the L2 regular term of the objective function is λ.

[0116] The model prediction for round t is:

[0117]

[0118] At this time, the objective function can be transformed into:

[0119]

[0120] After removing the constant term, we get the objective function:

[0121]

[0122] in, Represents the predicted value of the lth sample after the tth iteration; represents the predicted value of the lth sample after the t-1th iteration; f t (x i ) represents the model function of the i-th sample in the t-th iteration; represents the loss function of the tth iteration; g i represents the gradient boost weight of the i-th sample; h i represents the diagonal elements of the Hessian matrix of the i-th sample; Ω(f t ) represents the model function f in the tth iteration t The regularization term of ;

[0123] refer to Figure 3 , the LSTM network structure is designed with a special gating mechanism, which effectively solves the problem of gradient vanishing or exploding when traditional RNN processes long sequence data. It can also capture long-term dependencies in time series data and improve the model's ability to predict the trend of wind power changes. In addition, LSTM controls the flow of information through parameterized forget gates, input gates, and output gates, enhancing the model's adaptability to time dynamic characteristics, allowing LSTM to flexibly process time series of different lengths and improve the model's adaptability and prediction accuracy.

[0124] The LSTM neural network can be represented by time expansion as a chain structure, and there are 4 neural network layers in its repeating module. The corresponding parameters are:

[0125] f t =σ(W f [h t-1 ,x t ]+b f ) (14)

[0126] i t =σ(W i [h t-1 ,x t ]+b i ) (15)

[0127]

[0128] O t =σ(W o [h t-1 ,x t ]+b o ) (18)

[0129] h t =O t tanhC t (19)

[0130] Where: f t 、i t , O t are the forget gate output, input gate output, and output gate output at the current time t respectively; W f , b fare the weight and bias of the forget gate respectively; W i , b i are the weight and bias of the input gate respectively; W c , b c are the weight and bias of the updated value respectively; W o , b o are the weight and bias of the output gate respectively; σ(·) represents the Sigmoid activation function; x t , C t 、h t are the input, cell state and output value of the LSTM neural network unit at the current time t; C t-1 、h t-1 are the cell state at the previous moment and the output value of the LSTM neural network unit; [h t-1 , x t ] means merging matrices with equal number of rows into columns; is the candidate cell state; “*” indicates dot product.

[0131] The XGBoost model and the LSTM neural network model are trained using the training set and tested using the test set to generate a first wind power prediction model and a second wind power prediction model, respectively.

[0132] In order to improve the prediction accuracy and generalization ability of the model, reduce training time, avoid overfitting, and ensure that the model can efficiently process large-scale data sets, it is also necessary to optimize the model parameters through Bayesian optimization to achieve the best effect. The Bayesian optimization process is often used to select the optimal parameters for machine learning. In order to reduce the amount of model calculations and find the optimal feature combination, this paper optimizes the time window parameter t w , the number of features n to extract statistical features um1 and the number of features n to be combined um2 Perform Bayesian optimization to constrain feature dimensions and model parameter tuning, and construct sample parameter combinations [p1, p2, ..., p q ], where p s (s=1, 2, ..., q) is the sth feature combination, and q is the number of feature combinations. The steps of adjusting parameters using Bayesian optimization are as follows.

[0133] 1) Gaussian process estimation. First, select r feature combinations [p1, p2, ..., p r ], and calculate the corresponding loss functions [L(p1), L(p2), …, L(p r )], where L(p s)(s=1,2,…,r) is the loss function corresponding to the sth feature combination, where the sample data set includes a training set and a test set; then the sample parameters and the loss function are combined to obtain D r :

[0134] D r ={(p s ,L(p s )),s=1,2,…,r} (20)

[0135] Where D r represents the data set collected after the rth iteration in the Bayesian optimization process; p s represents the sth parameter combination; L(p s ) indicates that in the parameter combination p s Below, the loss function value of the model;

[0136] The estimated loss function follows a Gaussian distribution, then [L(p1), L(p2), …, L(p r )] is:

[0137]

[0138] represents a multivariate normal distribution, where 0 is the mean vector and K is the covariance matrix; K is an r×r matrix, where r is the number of parameter combinations evaluated; each element k(p i ,p j ) represents the parameter combination p i and p j The covariance of the loss functions between them.

[0139] 2) Select sampling points. The new sampling point is determined by comparing the loss function expectation of the new sampling point with the current sampling point. This mathematical expectation is called the expected improvement (EI) function. This paper uses the tree structure Parzen estimation algorithm to establish the EI function. Assume that p + is the current optimal parameter combination, which is L(p + ) Large value y * As the threshold, the mathematical expectation function E(·) can be obtained by equations (23) and (24).

[0140]

[0141] Where: P(·) is the probability density function, P(y|p s ) is p s The probability of y appearing when it appears, P(p s |y) is when y appears sThe probability of occurrence, P(p s ) is p s The probability of occurrence; γ is the distribution proportion; For y≤y * The distribution function when g(p s ) is y>y * The probability density function of .

[0142] The point where the function E(·) takes the maximum value is taken as the sampling point to obtain the new optimal parameter combination p new ,

[0143] p new = argmaxE(p s )

[0144] Where: argmaxE(p s ) means finding the parameters that maximize the mathematical expectation function.

[0145] 3) Update L(p) s ). After adding new samples, the new Gaussian distribution is:

[0146]

[0147] k′=[k(p new ,p1),k(p new ,p2),…,k(p new ,p r )] T (26)

[0148] According to the following formula, L(p new )’s new distribution P(L(p new |D r ,p new )).

[0149] P(L(p new |D r ,p new ))~N(μ,σ 2 ) (27)

[0150] μ=(k′) T K -1 [L(p1),L(p2),···,L(p r )] T (28)

[0151] σ 2 =k(p new ,p new )-(k′) T K -1 k′ (29)

[0152] Where: L(pi) represents the loss function value under the i-th parameter combination pi; pi represents the i-th parameter combination; p new represents a new parameter combination; K represents the covariance matrix of the loss function value of the known parameter combination; k′ represents the covariance vector between the new parameter combination and the known parameter combination; k(p new ,p new ) represents the covariance of the new parameter combination itself; P(L(p new |D r ,p new )) indicates that in a given data set D r and the new parameter combination p new Under the condition of , the probability distribution of the loss function value of the new sampling point; μ represents the predicted mean of the loss function value of the new sampling point; σ 2 Represents the predicted variance of the loss function value of the new sampling point;

[0153] 4) Repeat steps 2) and 3) until the maximum number of iterations is reached, and calculate the current optimal output point.

[0154] Bayesian optimization parameter adjustment is to intelligently select new parameter points for evaluation by constructing a probability model of the loss function.

[0155] For example, suppose you use an LSTM neural network model for wind power forecasting and want to optimize two parameters of the model: the learning rate and the number of hidden layer units.

[0156] 1. Define the objective function: Our objective function is the mean squared error (MSE) of the LSTM model on the validation set.

[0157] 2. Initialization: Randomly select several parameter points (e.g., learning rate = 0.01, number of hidden layer units = 50 and learning rate = 0.001, number of hidden layer units = 100) for evaluation.

[0158] 3. Construct a Gaussian process model: Use the data from these initial points to construct a Gaussian process model of the target function.

[0159] 4. Select sampling points: Use the Gaussian process model and the expected improvement (EI) criterion to select new parameter points. For example, the model may suggest that the next evaluation point is learning rate = 0.005 and number of hidden layer units = 75.

[0160] 5. Evaluate new points: Train the LSTM model on the new parameter points and evaluate its performance on the validation set.

[0161] 6. Update the Gaussian process model: Update the Gaussian process model according to the new evaluation results.

[0162] 7. Iteration: Repeat steps 4-6 until the preset number of iterations is reached or the performance improvement is no longer significant.

[0163] 8. Determine the optimal parameters: Among all the evaluated parameter points, select the parameter combination that minimizes the MSE on the validation set as the optimal parameter.

[0164] In step S105, the real-time characteristic data set is input into the first wind power prediction model and the second wind power prediction model respectively to generate a first prediction value and a second prediction value respectively, and then the combined wind power prediction value is obtained by using the inverse error method.

[0165] Finally, the two models are combined for prediction. In wind power prediction, the idea of ​​the combined model is to improve the accuracy of the prediction by combining the advantages of different models. As two different models, LSTM and XGBoost each have their own advantages in processing time series data and complex nonlinear relationships. LSTM is a recurrent neural network that is good at capturing long-term dependencies in time series data, while XGBoost is an ensemble learning method based on gradient boosting, which performs well in dealing with nonlinear and classification problems. When combining the prediction results of the two models, we can use the weighted average method. Specifically, we can assign weights according to the performance of each model on the validation set. We use the inverse error method to perform a weighted combination of the two predicted time series data, which is a strategy for determining weights, which calculates weights based on the prediction error of each model.

[0166] The core idea of ​​the inverse error method is that the smaller the prediction error of a model, the better its prediction performance is, so it should have a greater weight in the combined prediction. Formula (30) represents the combined prediction result f t are the two model prediction values ​​f 1t and f 2t where ω1 and ω2 are the corresponding weight coefficients. The calculation formulas (31) and (32) of the weight coefficients show that the weight is inversely proportional to the prediction error of the model. Specifically, if the error ε1 of the LSTM model is smaller than the error ε2 of the XGBoost model, then the weight ω1 of the LSTM model in the combined prediction will be larger, and vice versa.

[0167] The specific steps are as follows:

[0168] 1. For each time point t, get the predicted value f of the LSTM and XGBoost models 1t and f 2t .

[0169] 2. Calculate the prediction errors ε1 and ε2 of each model. The error calculation formula is as shown in formula (33), where n is the number of data points, y iis the ith actual value, is the i-th predicted value.

[0170] 3. Calculate the weight coefficients ω1 and ω2 based on the error value. Formulas are as follows (31) and (32)

[0171] 4. Apply formula (30) to calculate the combined prediction value f t .

[0172] 5. Repeat steps 1-4 until the predicted values ​​for all time points have been combined.

[0173] f t =ω1f 1t +ω2f 2t ,t=1,2,…,n (30)

[0174]

[0175] This application innovatively combines the XGBoost and LSTM models, using XGBoost's ability to handle nonlinear relationships and classification problems, as well as LSTM's advantages in capturing long-term dependencies in time series. This model combination significantly improves the accuracy of forecasting changes in wind power, while maintaining the model's adaptability and flexibility to changes in various meteorological conditions. And through Bayesian optimization technology, a probabilistic model of the objective function is established to efficiently search for the optimal solution in the parameter space. Finally, through the model prediction weighted combination strategy of the inverse error method, the weights are dynamically adjusted according to the prediction error of each model, and the combined prediction results are optimized, making the final prediction results more stable and accurate, and improving the model's response speed and prediction accuracy to short-term fluctuations in wind power.

[0176] Model prediction and validation. Use the test set data to predict the trained model and use the mean absolute percentage error (y MAPE ) and root mean square error (y RMSE ) are used to evaluate the performance of the prediction model. The smaller the calculated values ​​of the above two indicators are, the more accurate the load prediction results are.

[0177]

[0178] n is the total number of prediction results; y real (i) y fore (i) is the actual value and predicted value of wind power at the i-th moment.

[0179] refer to Figure 4Another specific embodiment of the present application discloses a wind power prediction system based on meteorological characteristics and deep learning models, including: a data acquisition module 401, used to acquire meteorological data related to wind power through multi-source data fusion technology and real-time data acquisition technology, wherein the meteorological data includes real-time meteorological data and historical meteorological data, and both the real-time meteorological data and the historical meteorological data include wind speed, wind direction, temperature, air pressure and humidity; a preprocessing module 402, used to preprocess the meteorological data, wherein the data preprocessing includes outlier processing and normalization processing; a data set construction module 403, used to construct a historical feature data set or a real-time feature data set based on the preprocessed historical meteorological data or real-time meteorological data, wherein the historical feature data set or the real-time feature data set includes time Features, original features, statistical features under time windows and combined features, wherein the historical feature data set and the corresponding wind power are divided into a training set and a test set; a model construction training module 404, used to construct an XGBoost model and an LSTM neural network model and use the training set and the test set to train the XGBoost model and the LSTM neural network model to generate a first wind power prediction model and a second wind power prediction model respectively, wherein the Bayesian optimization model is used to optimize the model parameters; a prediction module 405, used to input the real-time feature data set into the first wind power prediction model and the second wind power prediction model respectively to generate a first prediction value and a second prediction value respectively; a prediction value combination module 406, used to obtain a combined wind power prediction value using the inverse error method.

[0180] The data acquisition module 401 further includes: a data acquisition submodule, which is used to acquire multi-source historical meteorological data from different meteorological stations, satellite images and historical wind power databases; and a data fusion submodule, which is used to fuse the multi-source historical meteorological data by the following weighted average method to generate fused multi-source meteorological data:

[0181]

[0182] in, is the fused multi-source meteorological data, w i is the weight of the ith data source, x i is the data of the i-th data source; and a data acquisition submodule, which is used to use the Internet of Things device to collect key meteorological data in real time as real-time meteorological data, wherein the key meteorological data includes wind speed, wind direction, temperature, air pressure and humidity.

[0183] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0184] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed in the present application should be covered within the protection scope of the present application.

Claims

1. A wind power prediction method based on meteorological characteristics and deep learning model, characterized in that: include: Acquire meteorological data related to wind power through multi-source data fusion technology and real-time data acquisition technology, wherein the meteorological data includes real-time meteorological data and historical meteorological data; Preprocessing the meteorological data, wherein the data preprocessing includes outlier processing and normalization processing; Constructing a historical feature data set or a real-time feature data set based on the preprocessed historical meteorological data or the real-time meteorological data, wherein the historical feature data set or the real-time feature data set both include time features, original features, statistical features under a time window, and combined features, wherein the historical feature data set and the corresponding wind power are divided into a training set and a test set; Constructing an XGBoost model and an LSTM neural network model and training the XGBoost model and the LSTM neural network model using the training set and the test set to generate a first wind power prediction model and a second wind power prediction model, respectively, wherein the model parameters are optimized using a Bayesian optimization model; The real-time characteristic data set is respectively input into the first wind power prediction model and the second wind power prediction model to generate a first prediction value and a second prediction value respectively, and then a combined wind power prediction value is obtained by using an inverse error method.

2. The wind power prediction method based on meteorological characteristics and deep learning model according to claim 1 is characterized in that: Acquiring meteorological data through multi-source data fusion technology and real-time data acquisition technology further includes: Acquire multi-source historical meteorological data from different weather stations, satellite images and historical wind power databases; The multi-source historical meteorological data are fused by the following weighted average formula to generate fused multi-source meteorological data: in, is the fused multi-source meteorological data, w i is the weight of the ith data source, x i is the data of the ith data source; and The key meteorological data are collected in real time using Internet of Things devices as real-time meteorological data, wherein the key meteorological data include wind speed, wind direction, temperature, air pressure and humidity.

3. The wind power prediction method based on meteorological characteristics and deep learning model according to claim 2 is characterized in that: Using IoT devices to collect key meteorological data in real time further includes: Deploy anemometers, wind vanes, temperature sensors, barometers, and humidity sensors at wind farms; The anemometer, the wind vane, the temperature sensor, the barometer and the humidity sensor respectively collect wind speed, wind direction, temperature, air pressure and humidity in real time, wherein: The wind speed is used as the primary factor in determining wind power output, and its change directly affects the energy efficiency of wind turbines; The wind direction is used to determine whether the wind rotor can effectively capture wind energy; The temperature and air pressure, through their changes, act on the air density, thereby affecting the wind speed and the power coefficient of the wind turbine.

4. The wind power prediction method based on meteorological characteristics and deep learning model according to claim 2 is characterized in that: Preprocessing the meteorological data further includes: When the key meteorological data collected in real time is missing multiple consecutive data, it will be deleted directly; When the key meteorological data collected in real time is missing non-consecutive data, the i-th missing data is calculated by the following average formula: x i =(x i-1 +x i+1 ) / 2; Among them, x i is the i-th missing data; x i-1 and x i+1 is the data before and after the missing value. If x i+1 If it is a missing value, then take x i+2 ; When the key meteorological data after outlier processing has different units, the key meteorological data after outlier processing is normalized by the following formula: x a =(x-x min ) / (x max -x min ); Among them, x and x a are the key meteorological data before and after normalization; x max and x min are the maximum and minimum values ​​in the key meteorological data sets, respectively.

5. The method for predicting wind power based on meteorological characteristics and deep learning model according to claim 4 is characterized in that: Constructing a historical feature data set or a real-time feature data set based on the preprocessed historical meteorological data or real-time meteorological data further includes: The original features are wind speed, wind direction, temperature, humidity and air pressure features directly obtained from the pre-processed meteorological data: Y=[Y1,Y2,…,Y n ] T ; Among them, Y i =[y i1 ,y i2 ,…,yi M ](i=1,2,…,n) is the feature of the sample at the i-th moment, y ik (k=1,2,…,M) is Y i The kth feature in , n is the number of samples, and M is the number of meteorological features; The time feature is to extract the time feature of the month, day, week, hour and minute of the current year from the time; The combination features include linear combination features and nonlinear combination features, and the additive features, subtractive features, multiplicative features and divisive features are respectively expressed by the following formulas: f ia =y ij +y ik ; f is =y ij -y ik ; f im =y ij ×y ik ; f id =y ij ÷y ik ; The statistical features are used to represent the data fluctuations within the time window: in, They are time windows [it w ,i+t w ] the mean, standard deviation and maximum value of the kth feature in t w is the time window parameter, indicating the number of moments in the time window.

6. The method for predicting wind power based on meteorological characteristics and deep learning model according to claim 5, characterized in that: Optimizing model parameters using the Bayesian optimization model further includes: Select r feature combinations [p1, p2, ..., p r ] and calculate the corresponding loss function [L(p1),L(p2),…,L(p r )], where L(p s )(s=1,2,…,r) is the loss function corresponding to the sth feature combination, wherein the sample data set includes a training set and a test set; Combine the sample parameters and the corresponding loss function to get D r : D r ={(p s ,L(p s )),s=1,2,…,r}; Where D r represents the data set collected after the rth iteration in the Bayesian optimization process; p s represents the sth parameter combination; L(p s ) indicates that in the parameter combination p s Below, the loss function value of the model; It is estimated that the loss function follows a Gaussian distribution, [L(p1), L(p2), ..., L(p r )] is: in, represents a multivariate normal distribution, where 0 is the mean vector and K is the covariance matrix; K is an r×r matrix, where r is the number of parameter combinations evaluated; each element k(p i ,p j ) represents the parameter combination p i and p j The covariance of the loss functions between them; The point with the maximum value is selected as the sampling point by the following function E(·) formula: Among them, P(·) is the probability density function, P(y|p s ) is p s The probability of y appearing when it appears, P(p s |y) is when y appears p s The probability of occurrence, P(p s ) is p s The probability of occurrence; γ is the distribution proportion; l(p s ) is when y≤y * The distribution function when g(p s ) is when y>y * The probability density function of ; The optimal parameter combination p is obtained according to the sampling points new : p new =argmaxE(p s ); Among them, argmaxE(p s ) represents finding the parameter that maximizes the mathematical expectation function; Based on the optimal parameter combination, L(p s )’s Gaussian distribution: k′=[k(p new ,p1),k(p new ,p2),…,k(p new ,p r )] T ; According to the following formula, L(p new ) distribution P(K(p new |D r ,p new )): (L(p new |D r ,p new ))~N(μ,σ 2 ); μ=(k′) T TO -1 [L(p1),L(p2),…,L(p r )] T ; σ 2 =k(p new ,p new )-(k′) T K -1 k′; Where L(pi) represents the loss function value under the i-th parameter combination pi; pi represents the i-th parameter combination; p new represents a new parameter combination; K represents the covariance matrix of the loss function value of the known parameter combination; k′ represents the covariance vector between the new parameter combination and the known parameter combination; k(p new ,p new ) represents the covariance of the new parameter combination itself; P(L(p new |D r ,p new )) indicates that in a given data set D r and the new parameter combination p new Under the condition of , the probability distribution of the loss function value of the new sampling point; μ represents the predicted mean of the loss function value of the new sampling point; σ 2 Represents the predicted variance of the loss function value of the new sampling point; The steps of selecting sampling points and updating L(p) are performed in a cyclic manner. s ) until the maximum number of iterations is reached and the current optimal output point is calculated.

7. The method for predicting wind power based on meteorological characteristics and deep learning model according to claim 6, characterized in that: Constructing an XGBoost model and an LSTM neural network model and training the XGBoost model and the LSTM neural network model using the training set and the test set to generate a first wind power prediction model and a second wind power prediction model, respectively, further includes: The t-th round model of the XGBoost model and the corresponding objective function are expressed by the following formula: in, Represents the predicted value of the lth sample after the tth iteration; represents the predicted value of the lth sample after the t-1th iteration; f t (x i ) represents the model function of the i-th sample in the t-th iteration; represents the loss function of the tth iteration; g i represents the gradient boost weight of the i-th sample; h i represents the diagonal elements of the Hessian matrix of the i-th sample; Ω(f t ) represents the model function f in the tth iteration t The regularization term of ; The LSTM neural network model is represented by the following formula: f t =σ(W f [h t-1 ,x t ]+b f ); i t =σ(W i [h t-1 ,x t ]+b i ); The t =σ(W o [h t-1 ,x t ]+b o ); h t =Oh t underC t ; Among them, f t 、i t , O t are the forget gate output, input gate output, and output gate output at the current time t respectively; W f 、b f are the weight and bias of the forget gate respectively; Wi and bi are the weight and bias of the input gate respectively; W c 、b c are the weight and bias of the updated value respectively; W o 、b o are the weight and bias of the output gate respectively; σ(·) represents the Sigmoid activation function; x t , C t 、h t are the input, cell state and output value of the LSTM neural network unit at the current time t; C t-1 、h t-1 are the cell state at the previous moment and the output value of the LSTM neural network unit; [h t-1 ,x t ] means merging matrices with equal number of rows into columns; is the candidate cell state; "*" indicates dot product; The XGBoost model and the LSTM neural network model are trained using the training set, and the XGBoost model and the LSTM neural network model are tested using the test set to generate the first wind power prediction model and the second wind power prediction model, respectively.

8. The method for wind power prediction based on meteorological characteristics and deep learning model according to claim 7, characterized in that: Inputting the meteorological data to be predicted into the first wind power prediction model and the second wind power prediction model to generate a first prediction value and a second prediction value respectively, and then obtaining a combined wind power prediction value by using an inverse error method further includes: Inputting the real-time feature data set into the first wind power prediction model and the second wind power prediction model respectively to generate a first prediction value f1t and a second prediction value f2t respectively; The first prediction error ε1 and the second prediction error ε2 are calculated by the following formula: Where n is the number of data points, yi is the ith actual value, is the i-th predicted value; The first weight ω1 and the second weight ω2 are calculated by the following formula: The combined wind power forecast value is calculated by the following formula: f t =ω1f 1t +ω2f 2t ,t=1,2,…,n。 9. A wind power prediction system based on meteorological characteristics and deep learning model, characterized in that: include: A data acquisition module, used to acquire meteorological data related to wind power through multi-source data fusion technology and real-time data acquisition technology, wherein the meteorological data includes real-time meteorological data and historical meteorological data; A preprocessing module, used for preprocessing the meteorological data, wherein the data preprocessing includes outlier processing and normalization processing; A data set construction module, used to construct a historical feature data set or a real-time feature data set based on the preprocessed historical meteorological data or real-time meteorological data, wherein the historical feature data set or the real-time feature data set both include time features, original features, statistical features under a time window, and combined features, wherein the historical feature data set and the corresponding wind power are divided into a training set and a test set; A model building training module, used for building an XGBoost model and an LSTM neural network model and training the XGBoost model and the LSTM neural network model using the training set and the test set to generate a first wind power prediction model and a second wind power prediction model respectively, wherein the model parameters are optimized using a Bayesian optimization model; A prediction module, configured to input the real-time feature data set into the first wind power prediction model and the second wind power prediction model to generate a first prediction value and a second prediction value respectively; The prediction value combination module is used to obtain the combined wind power prediction value by using the inverse error method.

10. The wind power prediction system based on meteorological characteristics and deep learning model according to claim 9, characterized in that: The data acquisition module further comprises: The data acquisition submodule is used to obtain multi-source historical meteorological data from different meteorological stations, satellite images and historical wind power databases; The data fusion submodule is used to perform data fusion on the multi-source historical meteorological data by the following weighted average method to generate fused multi-source meteorological data: in, is the fused multi-source meteorological data, w i is the weight of the ith data source, x i is the data of the ith data source; and The data collection submodule is used to use the Internet of Things devices to collect key meteorological data in real time as real-time meteorological data, wherein the key meteorological data includes wind speed, wind direction, temperature, air pressure and humidity.

Images