Wind power prediction error estimation method based on meteorological prediction error

Abstract

The invention provides a wind power prediction error estimation method based on a meteorological prediction error, and the method comprises the steps: firstly analyzing the correlation between data, and determining main influence factors of the wind power prediction error according to the magnitude of the correlation; a multi-dimensional random variable kernel density estimation method is adopted to construct a weather forecast-forecast error condition probability distribution model; constructing an input feature data set based on the meteorological prediction error data set and additional feature correlation analysis, constructing a multi-neural network mixed error estimation model, and training by using the input feature data set and the wind power prediction error data set; and finally, randomly generating a large-sample weather forecast error sample according to the input weather forecast based on a Monte Carlo simulation method, constructing an input feature data set, inputting the input feature data set into the mixed error estimation model to obtain an error estimation sample data set, and finally fitting probability distribution of errors by using a kernel density estimation method. And error probability distribution estimation is completed. According to the invention, accurate estimation of wind power prediction error estimation is realized.

Description

Technical Field

[0001] The invention belongs to the field of wind power energy and relates to a method for estimating wind power prediction errors based on meteorological prediction errors. Background Art

[0002] As the most mature and large-scale power generation method among new energy sources, wind power has an increasing penetration rate in the power system. However, in the actual production process, new energy sources also have the disadvantage of being easily affected by natural meteorological factors. The uncertainty such as randomness and volatility caused by the influence of the natural environment will also bring new major challenges to the safe and stable operation of the power grid. Therefore, it is necessary to predict the future processing of wind power generation to eliminate this uncertainty as much as possible. However, errors are inevitable in the prediction of wind power. At present, general short-term wind power prediction methods usually use numerical weather prediction data (NWP) as input to make deterministic predictions on the expected value of future wind power output. However, due to the influence of factors such as the prediction error of NWP data itself and the deviation between the wind power prediction model and the ideal regression model, there are still relatively obvious errors in the short-term prediction of wind power, and such errors are usually manifested in the form of probability distribution with uncertain distribution type and uncertain parameters.

[0003] At present, the existing and relatively mature wind power error estimation methods mainly focus on directly fitting the distribution characteristics of the error, using methods such as kernel density estimation (KDE) and Gaussian Mixture Module (GMM) to fit the probability density characteristic curve of the prediction error. On this basis, there are also many methods that consider the segmented fitting method based on the predicted power size classification. However, such statistical methods do not effectively consider the impact of meteorological forecast errors on wind power prediction errors, as well as the error distribution characteristics of meteorological forecast errors themselves, and cannot effectively estimate the error of the prediction system that uses meteorology for wind power prediction. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for estimating wind power prediction errors using meteorological prediction errors. In view of the problem that the error in the numerical weather prediction data (NWP) used in the prediction system in the short-term wind power prediction will affect the short-term wind power prediction error, resulting in insufficient error estimation accuracy based solely on the wind power prediction error, the present invention analyzes the distribution characteristics of the meteorological forecast error and its impact on the wind power prediction error of the existing prediction system, further establishes a meteorological forecast data-meteorological forecast error mapping model and a meteorological forecast error-wind power prediction value-wind power prediction error mapping model, and realizes accurate estimation of the wind power prediction error.

[0005] The technical solution adopted by the present invention to solve the technical problem is:

[0006] The present invention first analyzes the correlation between individual data, mainly uses the maximum mutual information correlation coefficient to perform correlation analysis on meteorological error and wind power prediction, meteorological forecast data and meteorological error, determines the main influencing factors of wind power prediction error according to the size of the correlation, and uses the kernel density estimation method of multidimensional random variables to construct a meteorological forecast-forecast error conditional probability distribution model; further constructs an input feature data set based on the correlation analysis of the meteorological prediction error data set and additional features, constructs a multi-neural network hybrid error estimation model and uses the input feature data set and wind power prediction error data set for training; finally, based on the Monte Carlo simulation method, a large sample meteorological forecast error sample is randomly generated according to the input meteorological forecast and an input feature data set is constructed, which is input into the hybrid error estimation model to obtain an error estimation sample data set, and finally uses the kernel density estimation method to fit the probability distribution of the error to complete the probability distribution estimation of the error.

[0007] Here are the steps:

[0008] Step 1: Collect historical weather forecast data, historical measured weather data, historical wind power output forecast data, historical wind power output test data and corresponding timestamp data of the target wind farm station; clean the collected data, calculate the weather forecast error series and new energy error series, and perform maximum value normalization on the data;

[0009] Step 2: Calculate the correlation coefficient between wind power forecast error and each element of the original weather forecast sequence, screen important weather forecast features, and construct a sample feature matrix; use the kernel density estimation method of multidimensional random variables to construct a weather forecast-forecast error conditional probability distribution model;

[0010] Step 3: Based on the correlation analysis between the meteorological forecast error dataset and the additional features, an input feature dataset is constructed, and a multi-neural network hybrid error estimation model is constructed and trained using the input feature dataset and the wind power forecast error dataset;

[0011] Step 4: Based on the Monte Carlo simulation method, a large sample of meteorological forecast error samples is randomly generated according to the input meteorological forecast and an input feature data set is constructed. The data set is input into the hybrid error estimation model to obtain an error estimation sample data set. Finally, the kernel density estimation method is used to fit the probability distribution of the error to complete the probability distribution estimation of the error.

[0012] Furthermore, step 2 first calculates the maximum mutual information coefficient MIC between each meteorological forecast feature and the wind speed prediction error, and obtains the amount of information contributed by the change of the error of different meteorological features to the change of the wind power prediction error, so as to measure the correlation:

[0013]

[0014]

[0015] In the formula, p(ε i ,ε P ) is the i-th meteorological forecast feature ε i (t) and wind power forecast error series ε P The joint distribution probability of (t), a, b are the number of intervals divided by the algorithm for the two variables respectively, and the value of B is 0.6 power of the amount of data;

[0016] After obtaining the MIC coefficient, the meteorological feature errors are arranged in descending order according to MIC, and the meteorological features with the top 3 MIC coefficients and the remaining MIC coefficients greater than 0.5 are taken as important meteorological features to construct the meteorological forecast sample feature matrix Q;

[0017] The same method is used to calculate the correlation coefficient between meteorological forecast error and wind power forecast error time, screen the data features and construct the corresponding meteorological forecast error sample matrix E;

[0018] The joint probability distribution model of weather forecast sample characteristics is established using a multivariate kernel density estimation method based on adaptive covariance matrix transformation.

[0019] Furthermore, for the new weather forecast feature sample q new , the probability distribution of its error is obtained by the following method:

[0020] 2.1. First, calculate its difference with the historical sample q i The Mahalanobis distance between:

[0021]

[0022] Where ∑ is the covariance matrix between the selected meteorological features:

[0023] ∑ ik = Cov(Q i , Q k )

[0024] In the formula, Q i and Q k are the i-th feature sample and the k-th feature sample of the weather forecast sample feature matrix Q;

[0025] 2.2. Set the initial neighborhood radius r0 and the maximum neighborhood radius r max and the minimum number of samples M, filter out new The historical samples around the area are used as the target meteorological feature sample matrix Q new And the target weather forecast error sample matrix E new The minimum sample size M is determined based on the total sample size, taking 2% to 5% of the total data sample size.

[0026] Q new = {q i |d m (q new ,q i )≤r i}

[0027] E new ={e i |d m (q new ,q i )≤r i}

[0028] If the number of samples m in the target matrix is ​​less than M, the neighborhood radius is increased by the following formula, and the samples are re-screened until the number of samples m in the target matrix is ​​greater than or equal to M or the neighborhood radius increases to the maximum value.

[0029]

[0030] If there are no samples in the target matrix, find new Repeat the above process for the historical sample with the smallest Mahalanobis distance;

[0031] 2.3. After obtaining the target meteorological error matrix E new After that, the covariance matrix ∑ between the meteorological error features in the target matrix is ​​recalculated new The covariance matrix is ​​combined with the multi-random variable kernel density estimation method to establish a joint probability distribution model of weather forecast sample characteristics.

[0032]

[0033]

[0034] Where h is the adaptive bandwidth, d m (E new ) is the target meteorological error matrix E new The Mahalanobis distance between each sample vector and the origin, σ(x) represents the standard deviation of the sequence x, and d is the target meteorological error matrix E new The number of meteorological forecast error characteristics included.

[0035] Furthermore, step 3 specifically includes the following steps:

[0036] 3.1. First, the correlation between non-meteorological error features and wind power forecasting errors is analyzed, and the features with correlation > 0.4 are selected as the common input features of the model, which together with the meteorological forecast error sample matrix E constitute the training data set D0;

[0037] 3.2. Construct a multi-neural network hybrid error estimation model to estimate the wind power prediction error and specify the maximum number of neural network models N max Construct the first neural network, denoted as Model1, using a multi-layer perceptron MLP with a feature reorganization attention layer as the base learner to fit the error estimation mapping.

[0038] The input data of the model x = [x1, x2, x3, .., x d ]∈R d×1 , d is the dimension of the input data x,

[0039] First, after reorganizing the feature attention layer:

[0040] q=W q x+b q

[0041] k=W k x+b k

[0042]

[0043] v=W v x+b v

[0044] h attn =Norm(ev)

[0045] Among them, W q , W k and W v All are linear transformation layers with dimension d model ×d,d modelis the model dimension, which is one of the model's hyperparameters

[0046] Then, through the multi-layer perceptron layer, the number of layers of the perceptron is n mlp is the hyperparameter of the model;

[0047] h mlp,i =ReLU(h mlp,i-1 W mlp,i +b mlp,i )

[0048]

[0049] Among them, h mlp,0 =h attn , W mlp，i-1 is a linear transformation matrix with dimension d mlp，i ×d mlp,i-1 , d mlp，i Take 4d model ,h o It is the hidden vector output by the MLP model;

[0050] Finally, through an output layer Get the power error estimate ε of the model output o :

[0051] ε o =h o W o

[0052] 3.3. Dataset D i The model is first trained on the training set and then tested on the validation machine after each training. The model training is considered complete when the loss function MAE of the model on the validation set does not decrease more than 3 times. The model is then used to test D i Perform error estimation and obtain the error estimation data set

[0053] 3.4. Then filter out the data samples with large error estimation deviation and set the acceptable maximum error estimation deviation Δε m , filter D i The training data samples in form a new training data set D i+1 :

[0054]

[0055] The newly generated training dataset D i+1 Total number of samples m i+1 Greater than 5% of the total data N, update the Model at the same time i The dataset D i :

[0056]

[0057] Afterwards, for D i+1 Building a new neural network model i+1 , complete the Model through the training method in step 3 i+1 Initial training of Dataset to Model i Perform the same secondary training as step 3;

[0058] 3.5. Repeat 3.3 and 3.4 until the number of models reaches the maximum number or the error estimation deviations of all data samples are within the maximum deviation Δε m After all the error estimation neural network models are built and trained, the routing network G(x) of the hybrid error estimation model is constructed; the error estimation model Model that the input data sample should enter is determined according to its characteristics. i ,

[0059] The routing network is a multi-layer MLP model with a softmax function as the output layer:

[0060] h G =MLP G (x)

[0061] G(x)=α=soFtmax(h G )

[0062] The largest component α in the weight vector α max =max(α) is recorded as the error estimation model corresponding to the data sample. The training strategy of step 3.3 is adopted. The routing network is fully trained with the full data set D0 and the error estimation model parameters are frozen.

[0063] Finally, the trained routing network and each error estimation model are combined to form a hybrid error estimation model.

[0064] Furthermore, step 4 is based on the Monte Carlo method for a given weather forecast sample q new Next, based on the step 2 Randomly generate meteorological error sample data set E′ new , further combined with other features related to step 3.1 to form the error estimation input data set D′ new ; Input the error estimate into the data set D′ new Input into the mixed error estimation model constructed in step 3 to obtain the error estimation sample data set ε′ new ,

[0065] Get the error estimation sample data set ε′ new Then, the conditional probability distribution model of the error is established by the kernel density estimation method.

[0066]

[0067] ε i ∈ε′ new

[0068]

[0069] The weather forecast data q is obtained from this new Probability distribution of wind power forecast under certain conditions.

[0070] Furthermore, the weather forecast error sequence ε in step 1 is Q (t) and the new energy error series ε P (t) is calculated as:

[0071]

[0072]

[0073] Among them, X Q , are the measured data of meteorological data containing Q meteorological characteristics and the corresponding measured data of meteorological forecast data; P and The wind power measured data and the prediction system predict the power prediction data.

[0074] Furthermore, the non-meteorological error characteristics in step 3 include wind power forecast value, average value of recent wind power measured power, degree of fluctuation of recent wind power measured power, recent wind power forecast average value and recent wind power forecast standard deviation.

[0075] The advantages and positive effects of the present invention are:

[0076] 1. The wind farm adopts the meteorological error estimation method proposed in the present invention to obtain the joint conditional probability distribution of each meteorological error based on historical data, and then calculates the dynamic covariance matrix within the Mahalanobis distance neighborhood according to different input meteorological conditions during the calculation process, so that the constructed probability distribution is closer to the probability distribution of meteorological errors under similar meteorological conditions in real situations.

[0077] 2. After the wind farm adopts the hybrid error estimation model proposed in the present invention, it can make a deterministic estimate of the wind power prediction error of the original prediction system based on the meteorological prediction error and other characteristics that affect the wind power prediction error. The strategy of adopting the structure of a multi-neural network hybrid model + routing network and data set segmentation iteration and model retraining can effectively reduce the difficulty and time of clustering, and realize adaptive clustering and error estimation in the retraining process.

[0078] 3. The Monte Carlo simulation method proposed in the present invention adopted by the wind farm generates and samples data according to the input meteorological forecast data by combining the above two models, constructs a sample data set and then obtains the error estimation data set, and finally obtains the probability distribution of the error through the kernel density estimation method, thereby realizing the estimation of the error distribution under given meteorological forecast conditions, and providing a new error distribution reference basis for the error estimation and probabilistic prediction of the wind farm power prediction in addition to the previous error probability distribution obtained according to the power. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is a flow chart of the method of step 2 of the present invention;

[0080] Figure 2 is the new weather forecast feature sample q new Flowchart of the probability distribution method of error:

[0081] Figure 3 This is a flow chart of the method of step 3 of the present invention;

[0082] Figure 4 Schematic diagram of the reorganized feature attention layer in step 3;

[0083] Figure 5 This is a schematic diagram of the multi-layer perceptron in step 3;

[0084] Figure 6 This is a flow chart of the method of step 4 of the present invention; DETAILED DESCRIPTION

[0085] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.

[0086] A method for estimating wind power prediction error based on meteorological prediction error, the specific implementation steps are as follows:

[0087] Step 1: Collect historical weather forecast data, historical measured weather data, historical wind power output forecast data, historical wind power output test data, and corresponding timestamp data (in the format of year-month-day-hour-minute). The time resolution of historical data is 15 minutes; ensure that the collected weather data contains at least 70m (or 100m) above the ground layer height wind speed, wind direction, temperature, humidity, and air pressure (including forecast and measurement); historical new energy output forecast and measured data at least include the historical output active power of the power station; historical wind power output forecast data must be guaranteed to be the power forecast system currently used in the system

[0088] Secondly, the collected data are cleaned, including the identification and processing of missing and abnormal data. For data with a time interval of less than 2 hours (8 points at a resolution of 15 minutes), linear interpolation is used to supplement them to eliminate possible abnormal data or abnormal time points.

[0089] Calculate the weather forecast error series ε Q (t) and the new energy error series ε P (t):

[0090]

[0091]

[0092] Among them, X Q , are the measured data of meteorological data containing Q meteorological characteristics and the corresponding measured data of meteorological forecast data; P and The wind power measured data and the prediction system predict the power prediction data.

[0093] Normalize the data to ensure that all data are on a similar scale:

[0094]

[0095] In the formula, x′ is the processed data, and max(|x|) is the maximum value after the absolute value of the x sequence is taken.

[0096] Step 2: This step calculates the correlation coefficient between the wind power prediction error and the elements of the original meteorological forecast sequence, screens important meteorological forecast features, and constructs a sample feature matrix; further, the multidimensional kernel density estimation method is used to establish the conditional probability distribution of the meteorological forecast error.

[0097] The maximum mutual information coefficient (MIC) between each meteorological forecast feature and the wind speed prediction error is calculated to obtain the amount of information contributed by the change in the error of different meteorological features to the change in the wind power prediction error, so as to measure the correlation:

[0098]

[0099]

[0100] In the formula, p(ε i , ε P ) is the i-th meteorological forecast feature ε i (t) and wind power forecast error series ε P (t), in order to simplify the calculation, the two variables are usually divided into intervals, and the amount of data falling into a certain grid and the total number of wallpapers are taken as the approximate joint probability; a, b are the number of intervals divided by the algorithm for the two variables respectively, and the value of B is generally 0.6 power of the amount of data.

[0101] After obtaining the MIC coefficient, the meteorological feature errors are arranged in descending order according to MIC, and the meteorological features with the top 3 MIC coefficients and the remaining MIC coefficients greater than 0.5 are taken as important meteorological features to construct the meteorological forecast sample feature matrix Q.

[0102] Similarly, the same method is used to calculate the correlation coefficient between the meteorological forecast error and the wind power forecast error time, screen the data features and construct the corresponding meteorological forecast error sample matrix E.

[0103] Considering that the correlation between different meteorological factors will change in different time periods, this needs to be fully considered when calculating the distance between different meteorological feature vectors.

[0104] The joint probability distribution model of weather forecast sample characteristics is established using a multivariate kernel density estimation method based on adaptive covariance matrix transformation.

[0105] For a new weather forecast feature sample q new , the probability distribution of its error can be obtained by the following method:

[0106] 2.1. First, calculate its difference with the historical sample q i The Mahalanobis distance between:

[0107]

[0108] Where ∑ is the covariance matrix between the selected meteorological features:

[0109] ∑ik = Cov(Q i , Q k )

[0110] In the formula, Q i and Q k are the i-th feature sample and the k-th feature sample of the weather forecast sample feature matrix Q.

[0111] 2.2. Set the initial neighborhood radius r0 and the maximum neighborhood radius r max and the minimum number of samples M, filter out new The historical samples around the area are used as the target meteorological feature sample matrix Q new And the target weather forecast error sample matrix E new The minimum sample size M can be determined based on the total sample size, and can be 2% to 5% of the total data sample size.

[0112] Q new = {q i |d m (q new ,q i )≤r i}

[0113] E new ={e i |d m (q new ,q i )≤r i}

[0114] If the number of samples m in the target matrix is ​​less than M, the neighborhood radius is increased by the following formula, and the samples are rescreened until the number of samples m in the target matrix is ​​greater than or equal to M or the neighborhood radius increases to the maximum value.

[0115]

[0116] If there are no samples in the target matrix, find new Repeat the above process for the historical sample with the smallest Mahalanobis distance.

[0117] 2.3. After obtaining the target meteorological error matrix E new After that, the covariance matrix ∑ between the meteorological error features in the target matrix is ​​recalculated new The covariance matrix is ​​combined with the multi-random variable kernel density estimation method to establish a joint probability distribution model of weather forecast sample characteristics.

[0118]

[0119]

[0120] Where h is the adaptive bandwidth, d m (E new ) is the target meteorological error matrix E new The Mahalanobis distance between each sample vector and the origin, σ(x) represents the standard deviation of the sequence x, and d is the target meteorological error matrix E new The number of meteorological forecast error characteristics included.

[0121] Step 3: This step uses the historical meteorological forecast error data set and the historical wind power forecast error data set to directly construct a meteorological forecast error-wind power forecast error hybrid neural network mapping model;

[0122] 3.1. First, the correlation between other data features except meteorological errors and wind power forecast errors is analyzed, including the predicted power fluctuation degree of the original system, the recent average error of wind power forecasts, the recent fluctuation degree of wind power measured power, etc. The features with higher correlation (>0.4) are selected as the common input features of the model, and together with the meteorological forecast error sample matrix E, they constitute the training data set D0.

[0123] Table 1 lists the non-meteorological error characteristics considered by the present invention.

[0124]

[0125] Among them, N a For the recent look-back window distance, it can be 1-3 days according to actual needs, that is, 96 to 288 points.

[0126] 3.2. Construct a multi-neural network hybrid error estimation model to estimate the wind power prediction error and specify the maximum number of neural network models N max The first neural network is constructed, denoted as Model 1. The present invention uses a multi-layer perceptron (MLP) with a feature reorganization attention layer as a base learner to fit the error estimation mapping.

[0127] The input data of the model x = [x1, x2, x3, .., x d ]∈R d×1 , d is the dimension of the input data x.

[0128] First, after reorganizing the feature attention layer:

[0129] q=W q x

[0130] k=W k x

[0131]

[0132] v=Wv x+b v

[0133] h attn =Norm(ev)

[0134] Among them, W q , W k and W v All are linear transformation layers with dimension d model ×d,d model is the model dimension, which is one of the model's hyperparameters

[0135] Then, through the multi-layer perceptron layer, the number of layers of the perceptron is n mlp is the hyperparameter of the model;

[0136] h mlp,i =ReLU(h mlp,i-1 W mlp，i +b mlp,i )

[0137]

[0138] in, W mlp,i-1 is a linear transformation matrix with dimension d mlp，i ×d mlp,i-1 , d mlp,i Generally, 4d can be used uniformly model , can also be determined according to actual conditions; h o It is the hidden vector output by the MLP model;

[0139] Finally, through an output layer The power error estimate ε of the model output can be obtained o :

[0140] ε o =h o W o

[0141] 3.3. Dataset D i The model is shuffled and divided into a training set and a validation set with a ratio of 8:2 (which can be adjusted). The model is first trained on the training set, and then tested on the validation machine after each training of the training set. The model training is considered complete when the MAE of the model loss function on the validation set does not decrease more than 3 times. i Perform error estimation and obtain the error estimation data set

[0142] 3.4. Then filter out the data samples with large error estimation deviation and set the acceptable maximum error estimation deviation Δε m , filter Di The training data samples in form a new training data set D i+1 :

[0143]

[0144] The newly generated training dataset D i+1 Total number of samples m i+1 It should be greater than 5% of the total data N to ensure that the training data is sufficient; at the same time, update the Model i The dataset D i :

[0145]

[0146] Afterwards, for D i+1 Building a new neural network model i+1 , complete the Model through the training method in step 3 i+1 Initial training of Dataset to Model i Perform the same secondary training as step 3.

[0147] 3.5. Repeat 3.3 and 3.4 until the number of models reaches the maximum number or the error estimation deviations of all data samples are within the maximum deviation Δε m After all the error estimation neural network models are built and trained, the routing network G(x) of the hybrid error estimation model is constructed; its main function is to determine the error estimation model Model that it should enter according to the characteristics of the input data sample i .

[0148] The routing network used in this invention is a multi-layer MLP model with a softmax function as the output layer:

[0149] h G =MLP G (x)

[0150] G(x) = α = softmax(h G )

[0151] The largest component α in the weight vector α max =max(α) is recorded as the error estimation model corresponding to the data sample. A training strategy similar to step 3.3 is adopted, and the routing network is fully trained using the full data set D0 and the error estimation model parameters are frozen.

[0152] Finally, the trained routing network and each error estimation model are combined to form a hybrid error estimation model.

[0153] Step 4: This step is based on the Monte Carlo method for a given weather forecast sample q new Next, based on the step 2 Randomly generate meteorological error sample data set E′ new , further combined with other features related to step 3.1 to form the error estimation input data set D′ new ; Input the error estimate into the data set D′ new Input into the mixed error estimation model constructed in step 3 to obtain the error estimation sample data set ε′ new .

[0154] Get the error estimation sample data set ε′ new After that, the conditional probability distribution model of the error can be established by the kernel density estimation method.

[0155]

[0156] ε i ∈ε′ new

[0157]

[0158] Thus, the weather forecast data q is obtained new Probability distribution of wind power forecast under certain conditions.

[0159] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept, which all belong to the protection scope of the present invention.

Claims

1. A method for estimating wind power prediction error based on meteorological prediction error, characterized in that: Here are the steps: Step 1: Collect historical weather forecast data, historical measured weather data, historical wind power output forecast data, historical wind power output test data and corresponding timestamp data of the target wind farm station; clean the collected data, calculate the weather forecast error series and new energy error series, and perform maximum value normalization on the data; Step 2: Calculate the correlation coefficient between wind power prediction error and each element of the original meteorological forecast sequence, screen important meteorological forecast features, and construct a sample feature matrix; The conditional probability distribution model of weather forecast-forecast error is constructed by using the kernel density estimation method of multidimensional random variables. Step 3: Based on the correlation analysis between the meteorological forecast error dataset and the additional features, an input feature dataset is constructed, and a multi-neural network hybrid error estimation model is constructed and trained using the input feature dataset and the wind power forecast error dataset; Step 4: Based on the Monte Carlo simulation method, a large sample of meteorological forecast error samples is randomly generated according to the input meteorological forecast and an input feature data set is constructed. The data set is input into the hybrid error estimation model to obtain an error estimation sample data set. Finally, the kernel density estimation method is used to fit the probability distribution of the error to complete the probability distribution estimation of the error.

2. The method for estimating wind power prediction error based on meteorological prediction error according to claim 1, characterized in that: Step 2: First, calculate the maximum mutual information coefficient MIC between each meteorological forecast feature and the wind speed prediction error, and obtain the information contribution of the change of the error of different meteorological features to the change of wind power prediction error, so as to measure the correlation: In the formula, p(ε i ,ε P ) is the i-th meteorological forecast feature ε i (t) and wind power forecast error series ε P The joint distribution probability of (t), a, b are the number of intervals divided by the algorithm for the two variables respectively, and the value of B is 0.6 power of the amount of data; After obtaining the MIC coefficient, the meteorological feature errors are arranged in descending order according to MIC, and the meteorological features with the top 3 MIC coefficients and the remaining MIC coefficients greater than 0.5 are taken as important meteorological features to construct the meteorological forecast sample feature matrix Q; The same method is used to calculate the correlation coefficient between meteorological forecast error and wind power forecast error time, screen the data features and construct the corresponding meteorological forecast error sample matrix E; The joint probability distribution model of weather forecast sample characteristics is established using a multivariate kernel density estimation method based on adaptive covariance matrix transformation.

3. The method for estimating wind power prediction error based on meteorological prediction error according to claim 2, characterized in that: For the new weather forecast feature sample q new , the probability distribution of its error is obtained by the following method: 2.

1. First, calculate its difference with the historical sample q i The Mahalanobis distance between: Where ∑ is the covariance matrix between the selected meteorological features: ∑ ik =Cov(Q i ,Q k ) In the formula, Q i and Q k are the i-th feature sample and the k-th feature sample of the weather forecast sample feature matrix Q; 2.

2. Set the initial neighborhood radius r0 and the maximum neighborhood radius r max and the minimum number of samples M, filter out new The historical samples around the area are used as the target meteorological feature sample matrix Q new And the target weather forecast error sample matrix E new The minimum sample size M is determined based on the total sample size, taking 2% to 5% of the total data sample size. Q new ={q i |d m (q new ,q i )≤r i } E new ={e i |d m (q new ,q i )≤r i } If the number of samples m in the target matrix is ​​less than M, the neighborhood radius is increased by the following formula, and the samples are re-screened until the number of samples m in the target matrix is ​​greater than or equal to M or the neighborhood radius increases to the maximum value. If there are no samples in the target matrix, find new Repeat the above process for the historical sample with the smallest Mahalanobis distance; 2.

3. After obtaining the target meteorological error matrix E new After that, the covariance matrix ∑ between the meteorological error features in the target matrix is ​​recalculated new The covariance matrix is ​​combined with the multi-random variable kernel density estimation method to establish a joint probability distribution model of weather forecast sample characteristics. Where h is the adaptive bandwidth, d m (E new ) is the target meteorological error matrix E new The Mahalanobis distance between each sample vector and the origin, σ(x) represents the standard deviation of the sequence x, and d is the target meteorological error matrix E new The number of meteorological forecast error characteristics included.

4. The method for estimating wind power prediction error based on meteorological prediction error according to claim 3, characterized in that: Step 3 specifically includes the following steps: 3.

1. First, the correlation between non-meteorological error features and wind power forecasting errors is analyzed, and the features with correlation > 0.4 are selected as the common input features of the model, which together with the meteorological forecast error sample matrix E constitute the training data set D0; 3.

2. Construct a multi-neural network hybrid error estimation model to estimate the wind power prediction error and specify the maximum number of neural network models N max Construct the first neural network, denoted as Model1, using a multi-layer perceptron MLP with a feature reorganization attention layer as the base learner to fit the error estimation mapping. The input data of the model x = [x1, x2, x3, .., x d ]∈R d×1 , d is the dimension of the input data x, First, after reorganizing the feature attention layer: q=W q x+b q k=W k x+b k v=W v x+b v h attn =Norm(ev) Among them, W q , W k and W v All are linear transformation layers with dimension d model ×d,d model is the model dimension, which is one of the model's hyperparameters Then, through the multi-layer perceptron layer, the number of layers of the perceptron is n mlp is the hyperparameter of the model; h mlp,i =ReLU(h mlp,i-1 W mlp，i +b mlp,i ) Among them, h mlp,0 =h attn , W mlp，i-1 is a linear transformation matrix with dimension d mlp，i ×d mlp,i-1 , d mlp,i Take 4d model ,h o It is the hidden vector output by the MLP model; Finally, through an output layer Get the power error estimate ε of the model output o : e o =h o W o 3.

3. Dataset D i The model is first trained on the training set and then tested on the validation machine after each training. The model training is considered complete when the loss function MAE of the model on the validation set does not decrease more than 3 times. The model is then used to test D i Perform error estimation and obtain the error estimation data set 3.

4. Then filter out the data samples with large error estimation deviation and set the acceptable maximum error estimation deviation Δε m , filter D i The training data samples in form a new training data set D i+1 : The newly generated training dataset D i+1 Total number of samples m i+1 Greater than 5% of the total data N, update the Model at the same time i The dataset D i : Afterwards, for D i+1 Building a new neural network model i+1 , complete the Model through the training method in step 3 i+1 Initial training of Dataset to Model i Perform the same secondary training as step 3; 3.

5. Repeat 3.3 and 3.4 until the number of models reaches the maximum number or the error estimation deviations of all data samples are within the maximum deviation Δε m After all the error estimation neural network models are built and trained, the routing network G(x) of the hybrid error estimation model is constructed; the error estimation model Model that the input data sample should enter is determined according to its characteristics. i , The routing network is a multi-layer MLP model with a softmax function as the output layer: h G =MLP G (x) G(x)=α=softmax(h G ) The largest component α in the weight vector α max =max(α) is recorded as the error estimation model corresponding to the data sample. The training strategy of step 3.3 is adopted. The routing network is fully trained with the full data set D0 and the error estimation model parameters are frozen. Finally, the trained routing network and each error estimation model are combined to form a hybrid error estimation model.

5. The method for estimating wind power prediction error based on meteorological prediction error according to claim 4, characterized in that: Step 4: Based on the Monte Carlo method, the given weather forecast sample q new Next, based on step 2 Randomly generate meteorological error sample data set E′ new , further combined with other features related to step 3.1 to form the error estimation input data set D′ new ; Input the error estimate into the data set D′ new Input into the mixed error estimation model constructed in step 3 to obtain the error estimation sample data set ε′ new , Get the error estimation sample data set ε′ new Then, the conditional probability distribution model of the error is established by the kernel density estimation method. e i ∈ε′ new The weather forecast data q is obtained from this new Probability distribution of wind power forecast under certain conditions.

6. The method for estimating wind power prediction error based on meteorological prediction error according to claim 5, characterized in that: The weather forecast error sequence ε in step 1 is Q (t) and the new energy error series ε P (t) is calculated as: Among them, X Q , are the measured data of meteorological data containing Q meteorological characteristics and the corresponding measured data of meteorological forecast data; P and The wind power measured data and the prediction system predict the power prediction data.

7. The method for estimating wind power prediction error based on meteorological prediction error according to claim 6, characterized in that: The non-meteorological error characteristics in step 3 include the wind power forecast value, the average value of the recent wind power measured power, the fluctuation degree of the recent wind power measured power, the recent wind power forecast average value and the recent wind power forecast standard deviation.

Images