Regional wind power prediction method, device and medium

Abstract

The present invention belongs to the technical field of wind power prediction, discloses a regional wind power prediction method, device and medium, proposes a coefficient index for determining the strength of time correlation and volatility, performs characteristic analysis on wind farms, classifies wind farms into 4 situations, and designs different prediction methods. For wind farms with strong time correlation and strong volatility, empirical mode decomposition denoising and principal component analysis are used to reduce the dimension of data, and prediction is performed in combination with the CNN‑LSTM model. For wind farms with strong waves at times and weak waves, strong waves at times and weak waves at times, BiLSTM, random forest and liquid neural network are used respectively, and the prediction accuracy is optimized through outlier processing, data filling and feature selection. Finally, bagging ensemble learning, adaboost ensemble learning and LightGBM ensemble learning are selected for prediction, and the model effect is further screened and improved by comparing their advantages and disadvantages, and the optimal ensemble method is selected to predict the total predicted power of the entire region.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and more specifically, to a method, device and medium for regional wind power prediction. Background Art

[0002] As an important renewable energy source, wind power plays a key role in improving energy security and environmental protection. Electricity is a secondary energy source and is crucial to energy conservation, emission reduction and social development. Therefore, efficiently converting wind and solar energy into electricity and stably supplying it to users is the core goal of the energy revolution. However, the volatility and intermittency of wind power lead to high uncertainty. The impact of this uncertainty on the reliability, energy quality, economy and social welfare of the power system has increased significantly with the increase in wind power penetration. The stability of the power system reflects its ability to cope with disturbances and avoid collapse, while the abundance reflects the system's ability to meet the power and electricity needs of users. As the third largest power source after thermal power and hydropower, the introduction of wind power has significantly increased the randomness and uncertainty of the system. This poses a challenge to the safe and stable operation of the power system. Therefore, accurate wind power prediction technology has become the key to ensuring the stable operation of the power system. When using meteorological data to predict wind power, the different environmental factors of each wind farm and the randomness and uncertainty of meteorological data further increase the complexity of the prediction model. Domestic prediction of wind power systems focuses on modeling and prediction of a single wind farm, and the model of each wind farm is different. There are multiple wind farms in a region, and multiple prediction models greatly increase the difficulty and complexity of wind power management. A single prediction model is difficult to adapt to each wind farm, resulting in the need to conduct modeling training for each wind farm. Summary of the invention

[0003] The object of the present invention is to provide a regional wind power prediction method, device and medium to improve the convenience, accuracy and stability of the prediction.

[0004] In order to achieve the above object, the present invention adopts the following technical solution:

[0005] According to a first aspect of the present invention, a method for regional wind power prediction is provided, the method comprising:

[0006] Obtain humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, photovoltaic power generation sequence, and wind power sequence of the wind farm in the target area;

[0007] Based on the geographical location information, the wind farms in the target area are clustered to obtain multiple target sub-wind farms. The characteristics of the multiple target sub-wind farms are divided using the time correlation index and the volatility index; the time correlation index is the autocorrelation analysis coefficient and the periodicity coefficient, and the volatility index is the variation coefficient and the variance coefficient;

[0008] Assign time correlation and volatility labels to multiple target sub-wind farms respectively, and divide the target area wind farms into multiple target sub-clusters according to the strength of time correlation and volatility. Each target sub-cluster uses a different prediction model to predict wind power.

[0009] The prediction models of each target sub-cluster are integrated using different learning methods, multiple integrated models are established, and model evaluation is performed. The integrated model with the best performance is selected as the final prediction model, and regional wind power prediction is achieved based on the final prediction model.

[0010] According to a second aspect of the present invention, a regional wind power prediction device is provided, the device comprising:

[0011] A data acquisition module is configured to acquire a humidity sequence, an air pressure sequence, a temperature sequence, a sine sequence of wind directions at different heights, a cosine sequence of wind directions at different heights, a wind speed sequence at different heights, an illumination sequence, a photovoltaic power generation sequence, and a wind power sequence of a wind farm in a target area;

[0012] The characteristic classification module is configured to cluster the wind farms in the target area based on the geographical location information to obtain multiple target sub-wind farms, and to perform characteristic classification on the multiple target sub-wind farms using the time correlation index and the volatility index; wherein the time correlation index is the autocorrelation analysis coefficient and the periodicity coefficient, and the volatility index is the variation coefficient and the variance coefficient;

[0013] The first prediction module is configured to assign time correlation and volatility labels to multiple target sub-wind farms respectively, and divide the target area wind farms into multiple target sub-clusters according to the strength of the time correlation and volatility, and each target sub-cluster uses a different prediction model to predict wind power;

[0014] The second prediction module is configured to integrate the prediction models of each target sub-cluster using different learning methods, establish multiple integrated models, and perform model evaluation, select the integrated model with the best performance as the final prediction model, and realize regional wind power prediction based on the final prediction model.

[0015] According to a third aspect of the present invention, there is provided a readable storage medium storing one or more programs, wherein the one or more programs can be executed by one or more processors to implement the method as described above.

[0016] The present invention has at least the following beneficial effects:

[0017] The present invention divides the characteristics of the target sub-wind farms through time correlation indicators and volatility indicators, and constructs different target sub-clusters based on the characteristic division results. Different target sub-clusters use different prediction models for prediction, and finally integrates multiple prediction models through a screened learning model to obtain a final prediction model. The regional wind power prediction implemented based on the final prediction model can improve the convenience and accuracy of the prediction and the stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flow chart of a regional wind power prediction method according to an embodiment of the present invention is shown.

[0019] Figure 2 An autocorrelation coefficient analysis diagram according to an embodiment of the present invention is shown.

[0020] Figure 3 A power prediction line graph according to an embodiment of the present invention is shown.

[0021] Figure 4 A prediction error histogram according to an embodiment of the present invention is shown.

[0022] Figure 5 A structural diagram of a regional wind power prediction device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and specific embodiments, but are not intended to limit the present invention. For the various steps described herein, if there is no necessity for a causal relationship between each other, the order in which they are described as examples herein should not be regarded as a limitation, and those skilled in the art should know that they can be adjusted in order, as long as the logic between them is not destroyed, resulting in the inability to implement the entire process.

[0024] The embodiment of the present invention provides a method for predicting regional wind power. Figure 1 The flowchart of the method is shown in FIG. The method comprises the following steps:

[0025] S1. Obtain humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, photovoltaic power generation sequence, and wind power sequence of the wind farm in the target area.

[0026] It should be noted that in step S1, the humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, photovoltaic power generation sequence, and wind power sequence of the wind farm in the target area obtained are the historical humidity data, air pressure data, temperature data, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed data at different heights, illumination data, photovoltaic power generation data, and wind power data of the wind farm in the target area. The purpose is to provide a data basis for constructing the final prediction model in subsequent steps.

[0027] S2. Based on the geographic location information, the wind farms in the target area are clustered to obtain multiple target sub-wind farms, and the characteristics of the multiple target sub-wind farms are divided using time correlation indicators and volatility indicators; the time correlation indicators are the autocorrelation analysis coefficient and the periodicity coefficient, and the volatility indicators are the variation coefficient and the variance coefficient.

[0028] In some embodiments, the specific implementation process of step S2 is as follows:

[0029] The characteristics of multiple target sub-wind farms are distinguished, and labels of time correlation and volatility are assigned respectively. Through fuzzy C-means clustering, according to the strength of their time correlation and volatility, autocorrelation analysis (ACF) and periodic coefficient are used to perform different methods of data preprocessing and wind power prediction on the four sub-clusters. Taking time correlation and volatility as indicators, the characteristics of wind farms in the region are divided.

[0030] The autocorrelation analysis coefficient of the target sub-wind farm is calculated by the following formula:

[0031]

[0032] In the formula, n represents the number of power sequences, represents the mean of the power series, k represents the lag order, r k Represents the autocorrelation analysis coefficient.

[0033] Then r k After finding the absolute value, sum it up and then find the average to get the time correlation index.

[0034]

[0035] When the index is less than or equal to 0.01, it is judged that the time correlation between the data points is weak, and when the index is greater than 0.01, it is judged that the time correlation between the data points is strong.

[0036] Based on the wind power sequence x of the target sub-wind farm in the target area wind farm 1 -x n, divided by month, construct the characteristic matrix X according to the monthly wind power sequence, perform periodic analysis on the characteristic matrix X, and obtain the periodic coefficient:

[0037]

[0038] In the formula, m is the number of samples, d is i is the difference in rank value of each power data point, r s is the periodic coefficient

[0039] When s When it is greater than 0.8, the data is judged to have strong periodicity, and when it is less than or equal to 0.8, the data is judged to have weak periodicity. Only when the ACF coefficient and the periodicity coefficient are both judged to be strong can its time correlation be judged to be strong, otherwise it is weak.

[0040] The coefficient of variation is calculated using the following formula:

[0041]

[0042] In the formula, CV represents the coefficient of variation, n represents the number of power sequences, represents the mean of the power series, x i Represents the i-th sample value.

[0043] When the CV index is less than or equal to 100%, it is judged as weak volatility, and when the CV index is greater than 100%, it is judged as strong volatility.

[0044] Among them, the variance coefficient is a volatility indicator of the variance obtained by normalizing the data of the entire data period.

[0045] For wind power series x 1 -x n Normalize and get a normalized sequence:

[0046]

[0047] In the formula, t represents the normalized data, t i represents the i-th sample value, t min Represents the minimum value in the wind power sequence, t max Indicates the maximum value in the wind power sequence;

[0048] Calculate the variance of the normalized sequence and get the variance coefficient σ 2 :

[0049]

[0050] In the formula, Represents the average value of the normalized series.

[0051] When the index is less than or equal to 0.06, it is judged that the volatility between data points is weak, and when the index is greater than 0.06, it is judged that the volatility between data points is strong; the volatility can only be judged as strong when the coefficient of variation and the coefficient of variance are both judged as strong, otherwise it is weak.

[0052] S3. Time correlation and volatility labels are respectively assigned to multiple target sub-wind farms. According to the strength of the time correlation and volatility, the wind farms in the target area are divided into multiple target sub-clusters. Each target sub-cluster uses a different prediction model to predict wind power.

[0053] In some embodiments, the specific implementation process of step S3 is as follows:

[0054] S31. Based on the characteristic division in step S2, characteristic labels are assigned to the target sub-wind farm, and the target sub-wind farm is divided into four target sub-clusters. The characteristic labels of the four target sub-clusters are respectively strong time correlation and strong volatility; strong time correlation and weak volatility; weak time correlation and strong volatility; weak time correlation and weak volatility.

[0055] S32. Predict the four target subclusters in different ways respectively.

[0056] S321. For target subclusters with strong temporal correlation and volatility, the prediction method is as follows:

[0057] First, obtain the wind power sequence x of k target sub-wind farms in the target sub-cluster 1 -x k , in n trials, by adding the white noise time series to the power series to construct a new time series Y n (t), based on the empirical mode decomposition (EMD) algorithm, the time series Y contaminated by noise n (t) is decomposed into a series of intrinsic mode functions and the sum of residuals, and N iterations are performed to obtain the final intrinsic mode function, which is the power sequence after denoising decomposition:

[0058]

[0059]

[0060] In the formula, m is the number of iterations of the sifting algorithm, n is the number of integrations, and Y n (t) is the power sequence after adding noise, is the final decomposed power sequence, are the eigenmode functions decomposed at time t.

[0061] Then, principal component analysis (PCA) is used to reduce the dimension of the meteorological data:

[0062]

[0063] In the formula, x n represents the nth wind power, a nn represents the coefficient of the principal component, x′ n Represents the principal component.

[0064] Compress the data to minimize the error, improve the prediction speed, select CNN-LSTM for prediction, extract its spatial features through CNN, and extract its time series features through LSTM. The process of LSTM extracting its time series features is as follows:

[0065] Input gate: i t =σ(W i [x t ,h t -1]+b i )

[0066] Forget gate: f t =σ(W f [x t ,h t -1]+b f )

[0067] Output gate: o t =σ(W o [x t ,h t -1]+b o )

[0068] Candidate cell state: c t ′=tanh(W c [x t ,h t -1]+b c )

[0069] Cell state: c t =f t *c t-1 +i t *c t '

[0070] Hidden state: h t =o t *tanh(c t )

[0071] In the formula, i tIndicates the activation value of the input gate, which indicates the state of the input gate at the current time t. It determines how much of the current input information will be stored in the cell state. t represents the input vector at the current time t, represents the data of the input sequence at time t, W i Represents the weight matrix of the input gate, acting on the input x t and the weight matrix of the previous hidden state h(t-1) used to calculate the input gate activation value i t ,h t represents the hidden state vector at the current time t, represents the output of LSTM at time t, and is also used for calculation at the next time. i Represents the bias vector of the input gate, which is used to calculate the input gate activation value i t , f t represents the activation value of the forget gate, which indicates the state of the forget gate at the current time t. It determines how much information in the cell state c(t-1) at the previous time will be retained. σ represents the activation function, and W f Represents the weight matrix of the forget gate, acting on x t and the weight matrix of the previous hidden state h(t-1), used to calculate the forget gate activation value f t , b f Represents the bias vector of the forget gate, which is used to calculate the activation value f of the forget gate t , O t Represents the activation value of the output gate, indicating the state of the output gate at the current time t. It determines how much information in the current cell state ct will be output, W o Represents: The weight matrix of the output gate, the weight matrix acting on the input and the previous hidden state h(t-1), used to calculate the activation value O of the output gate t , b o Represents the bias vector of the output gate, which is used to calculate the output gate activation value, W c represents the weight matrix used to generate candidate cell states, b c represents the bias vector used to generate candidate cell states, c t Represents the cell state at the current time t, including the memory content of the LSTM unit, c t-1 Represents the cell state at the previous time t-1.

[0072] S322: For a target sub-cluster with strong time correlation and weak volatility, first obtain the wind power sequence x of k target sub-wind farms in the target sub-cluster. 1 -x k, Use the interquartile range method to perform outlier tests on the power sequence, and return the outlier data and missing data values. Select the average number of terms N < K, and fill the data with moving averages:

[0073]

[0074] Where k is the sequence sample number, and N is the average number of terms. is used to calculate the next data value to be filled. is the data value at the current moment. is the first data value within the selected time observation interval.

[0075] Obtain the humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, light sequence, and photovoltaic power generation sequence of the target sub-wind farm, construct them into a feature matrix, and construct the filled power sequence into matrix X1. Add the feature matrix and matrix X1 to obtain a new feature matrix X2:

[0076]

[0077] Where x ak represents the humidity sequence of the kth target sub-wind farm, and x bk represents the air pressure sequence of the kth target sub-wind farm, and x ck represents the temperature sequence of the kth target sub-wind farm, and x dk represents the wind direction sine sequence at different heights of the kth target sub-wind farm, and x ik represents the filled power sequence of the kth target sub-wind farm;

[0078] Sort the feature sequences in X2 respectively and calculate the difference in their rank values, and substitute them into the Spearman correlation analysis formula to obtain the importance degree of each feature:

[0079]

[0080] Where d i is the difference in rank values for each power data point, and ρ is the importance of the sequence in the feature matrix X 2 ;

[0081] After performing importance analysis, select the sequences with importance greater than or equal to 0.8 as input data, perform reverse processing on the input sequence data, and select the LSTM model for prediction.

[0082] S323. For the target sub-cluster with weak time correlation and strong volatility, the prediction method is as follows:

[0083] First, obtain the wind power sequences x of k target sub-wind farms in the target sub-cluster1 -x k , in the n -th experiment, by adding a white - noise time series to the power series, a new time series Y n (t) is constructed. Based on the empirical mode decomposition (EMD) algorithm, the time series Y n (t) contaminated by noise is decomposed into the sum of a series of intrinsic mode functions and a residue. After N -th iteration, the final intrinsic mode functions are obtained, which are the power series after noise - reduction decomposition:

[0084]

[0085]

[0086] where m is the number of iterations of the sifting algorithm, n is the number of integrations, Y n (t) is the power series after adding noise, is the power series after final decomposition, are the respective intrinsic mode functions decomposed at time t;

[0087] The decomposed power series is subjected to abnormal - sequence rejection, and the remaining sequences are re - combined to obtain a new power series. Then, principal component analysis (PCA) is used to reduce the dimension of meteorological data such as humidity sequence, air - pressure sequence, temperature sequence, sine sequences of wind directions at different heights, cosine sequences of wind directions at different heights, wind - speed sequences at different heights, illumination sequence, photovoltaic - power generation sequence, etc. of the target sub - wind farm to improve the prediction speed. The meteorological data after dimension reduction and the power series after decomposition and recombination are used as the input data set to obtain feature data. Among n power data, m data are randomly selected to construct a sample subset, and among i features, k features are randomly selected to construct a feature subset, where k << i. A decision tree is constructed using the selected samples and features. At each node, an optimal feature is selected for splitting to maximize its information gain. Through continuous iteration, the prediction stops when each node contains only one sample. By constructing multiple decision trees, the solution with the best prediction effect is selected as the output, which is the final prediction result.

[0088] S324. For a wind farm with weak time - correlation and weak volatility, first obtain the wind - power sequences x 1 -x k of k sub - wind farms in the target area. The inter - quartile range method is used to test for outliers in the power sequence, and the outlier data and missing - data values are returned. Select the average number of terms N < K, and fill the data by moving average:

[0089]

[0090] where k is the number of sequence samples and N is the average number of terms.

[0091] According to the humidity sequence, pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, and photovoltaic power generation sequence of the target sub-wind farm, the first characteristic matrix is ​​constructed, and the filled power sequence is constructed as the second characteristic matrix. The first characteristic matrix and the second characteristic matrix are added to obtain the characteristic matrix X 2 :

[0092]

[0093] In the formula, x ak represents the humidity sequence of the kth target sub-wind farm, x bk represents the air pressure sequence of the kth target sub-wind farm, x ck represents the temperature sequence of the kth target sub-wind farm, x dk represents the sinusoidal sequence of wind directions at different heights of the kth target sub-wind farm, x ik represents the power sequence after filling of the kth target sub-wind farm;

[0094] Take X2 as the input data set and use the XGBoost algorithm for prediction. For a certain feature x, the Shapley value is calculated for all possible feature combinations and weighted summed to obtain the importance of the relevant features:

[0095]

[0096] In the formula, X represents the feature subset used in the model, x represents the eigenvalue vector of the power sequence, p is the number of features, val(X) represents the output value under the feature combination X, and val(X∪{x j}) represents the marginal contribution of the j-th eigenvalue vector to X after it is added to the feature subset X. Indicates the importance of the relevant features, x p represents the pth eigenvalue vector, x j represents the j-th eigenvalue vector;

[0097] After the importance analysis, the sequences with importance greater than or equal to 0.8 were selected as input data, and the liquid neural network was selected for prediction.

[0098] S4. Integrate the prediction models of each target sub-cluster using different learning methods, establish multiple integrated models, and perform model evaluation to select the integrated model with the best performance as the final prediction model, and realize regional wind power prediction based on the final prediction model.

[0099] The specific steps of step 4 are as follows

[0100] The above four models are integrated using three ensemble learning methods. The three ensemble learning methods are divided into two categories, one is bagging and the other is boosting. Bagging ensemble learning is applicable to the case where the prediction effects of the four models are relatively good. It is integrated through sampling learning to curb its overfitting, while boosting ensemble learning is applicable to the case where the prediction effects of the four models are relatively poor. It is gradually enhanced to curb its underfitting. If the environmental characteristics of the wind farms in the two regions are different, the results of time correlation and volatility analysis are also different, and the prediction effects of the four fixed prediction models are also different, so it is necessary to compare different ensemble methods to further reduce the error. By selecting bagging ensemble learning, adaboost ensemble learning, and LightGBM ensemble learning for prediction, by comparing their advantages and disadvantages, the optimal ensemble method is selected as the final prediction model, and then the wind farms in the region are divided into n target subclusters by geographical features using the mean clustering method, and the real-time meteorological data of the n target subclusters are predicted using the above ensemble methods, and the prediction results are superimposed to obtain the total predicted power of the region.

[0101] The following two examples are combined to further illustrate the feasibility and advancement of the present invention.

[0102] Example 1:

[0103] The characteristics of the regional wind farms are divided by performing hierarchical clustering analysis on the humidity data, air pressure data, temperature data, wind direction sine at different heights, wind direction cosine at different heights, wind speed data at different heights, light data, photovoltaic power generation data, and wind power data of the regional wind farms. Next, the wind farms are classified and integrated for prediction. The specific steps are as follows:

[0104] Step 1: Collect humidity data, air pressure data, temperature data, wind direction sine data at different heights, wind direction cosine data at different heights, wind speed data at different heights, light data, photovoltaic power generation data, and wind power data of a wind farm in a certain area of ​​Xinjiang. Data preprocessing includes cleaning and standardization to ensure the accuracy and consistency of the data.

[0105] Step 2: Characterize the wind farms in the region, with indicators including time correlation and volatility. For example, after analyzing the indicators of n wind farms in a certain area of ​​Xinjiang, they are divided into four types. The indicators are mainly used to distinguish the strength of the time correlation and volatility of wind power data. After the division, the prediction models of the wind farms are automatically integrated and screened to select the optimal prediction model.

[0106] The evaluation indicators of time correlation are autocorrelation analysis (ACF) and periodicity coefficient.

[0107] The periodicity coefficient is the absolute average of the Spearman correlation coefficients between each month. If it is greater than 0.8, it is considered to be strong periodicity, and if it is less than or equal to 0.8, it is considered to be weak periodicity. Only when both the ACF coefficient and the periodicity coefficient are strong can the time correlation be considered to be strong, otherwise it is considered to be weak. Figure 2 This is the image after power series autocorrelation coefficient analysis.

[0108] The evaluation indicators of volatility are the coefficient of variation and the coefficient of variance.

[0109] If the coefficient of variation is equal to 100%, the volatility is judged to be weak; if it is greater than 100%, the volatility is judged to be strong.

[0110] The coefficient of variance is the variance volatility index obtained after normalization of the entire data period. If it is less than or equal to 0.06, the volatility is judged to be weak; if it is greater than 0.06, the volatility is judged to be strong. The coefficient of variation and the coefficient of variance must be strong at the same time to judge the volatility to be strong, otherwise it is weak.

[0111] Step 3: Divide the regional wind farms into four subclusters based on time correlation and volatility:

[0112] 1. Strong time correlation and volatility.

[0113] 2. Strong time correlation and weak volatility.

[0114] 3. Weak time correlation and strong volatility.

[0115] 4. Weak time correlation and volatility.

[0116] Step 4: Use appropriate prediction models and data processing methods for wind farms of different properties.

[0117] Strong time correlation and volatility: Use empirical mode decomposition (EEMD) to reduce the noise of wind power data, remove abnormal sequences and reorganize them. Use principal component analysis (PCA) to reduce the dimension of data and compress data to improve prediction speed. Select CNN-LSTM for prediction, extract spatial features through CNN, and extract time series features through LSTM.

[0118] Strong time correlation and weak volatility: Use the interquartile range method to test outliers, and the moving average method to fill in outliers and missing values. Use the Spearman correlation coefficient to analyze the importance of meteorological data, and select data with importance greater than or equal to 0.8 as input. Select the BiLSTM model for prediction to improve prediction accuracy.

[0119] Table 1 shows the results of Spearman correlation analysis of some meteorological data.

[0120] Table 1 Results of Spearman correlation analysis of some data

[0121]

[0122] Weak time correlation and strong volatility: Use the EEMD method to de-noise the wind power data, remove abnormal sequences and then reorganize. Use PCA to reduce the data dimension and compress the data to improve the prediction speed. Select random forest for prediction, analyze the relationship between input features and target variables, identify features that have a significant impact on volatility, reduce the overfitting risk of a single decision tree, and improve the stability and accuracy of the prediction.

[0123] Weak time correlation and volatility: Use the interquartile range method to test outliers, and the moving average method to fill in outliers and missing values. Use the SHAP interpretability model to interpret the XGBoost algorithm to analyze the importance of meteorological data, and select data with importance greater than or equal to 0.8 as input. Select liquid neural network for prediction, which has the advantages of fast prediction speed, small size, free network structure, and high prediction accuracy.

[0124] Step 5: Use three ensemble learning methods to integrate the above four models. The three ensemble learning methods belong to bagging and boosting.

[0125] Bagging ensemble learning is suitable for when the prediction effects of the four models are good, and it can curb overfitting by learning ensemble through sampling.

[0126] Boosting ensemble learning is suitable for curbing underfitting through gradual enhancement learning when the prediction effect of the four models is poor.

[0127] Compare the advantages and disadvantages of bagging ensemble learning, adaboost ensemble learning, and LightGBM ensemble learning, and select the optimal ensemble method as the final prediction model.

[0128] Step 6: Use the mean clustering method to divide the wind farms in the region into n subclusters according to geographical characteristics.

[0129] The above ensemble method is used to predict these n subclusters respectively, and the prediction results are superimposed to obtain the total prediction power of the area.

[0130] Example 2:

[0131] The wind power prediction experiment with integrated prediction and the wind power prediction experiment without integrated prediction were selected to obtain the wind power prediction value image and the LSTM model relative error image, where the wind power prediction value takes the 5000th to 5100th data, such as Figure 3 to Figure 4 shown.

[0132] The embodiment of the present invention also provides a regional wind power prediction device, such as Figure 5 As shown, the device comprises:

[0133] The data acquisition module 501 is configured to acquire a humidity sequence, an air pressure sequence, a temperature sequence, a sine sequence of wind directions at different heights, a cosine sequence of wind directions at different heights, a wind speed sequence at different heights, an illumination sequence, a photovoltaic power generation sequence, and a wind power sequence of a wind farm in a target area;

[0134] The characteristic classification module 502 is configured to cluster the wind farms in the target area based on the geographical location information to obtain a plurality of target sub-wind farms, and to perform characteristic classification on the plurality of target sub-wind farms using a time correlation index and a volatility index; wherein the time correlation index is an autocorrelation analysis coefficient and a periodicity coefficient, and the volatility index is a variation coefficient and a variance coefficient;

[0135] The first prediction module 503 is configured to assign time correlation and volatility labels to multiple target sub-wind farms respectively, and divide the target area wind farms into multiple target sub-clusters according to the strength of the time correlation and volatility, and each target sub-cluster uses a different prediction model to perform wind power prediction;

[0136] The second prediction module 504 is configured to integrate the prediction models of each target sub-cluster using different learning methods, establish multiple integrated models, and perform model evaluation, select the integrated model with the best performance as the final prediction model, and realize regional wind power prediction based on the final prediction model.

[0137] It should be noted that the various device structures described in this embodiment belong to the same technical concept as the previously described method, and achieve the same technical effect through the same principle, which will not be repeated here.

[0138] An embodiment of the present invention further provides a readable storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement the method described in any of the above embodiments.

[0139] In addition, although exemplary embodiments have been described herein, the scope includes any and all embodiments based on the present invention with equivalent elements, modifications, omissions, combinations (e.g., various embodiments intersecting schemes), adaptations or changes. The elements in the claims will be interpreted broadly based on the language adopted in the claims, and are not limited to the examples described in this specification or during the implementation of this application, and the examples will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered as examples only, and the true scope and spirit are indicated by the following claims and the full scope of their equivalents.

[0140] The above description is intended to be illustrative rather than restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. For example, those of ordinary skill in the art can use other embodiments when reading the above description. In addition, in the above-mentioned specific embodiments, various features can be grouped together to simplify the present invention. This should not be interpreted as a feature of an invention that is not claimed for protection being necessary for any claim. On the contrary, the subject matter of the present invention may be less than all the features of the embodiments of a specific invention. Thus, the following claims are incorporated into the specific embodiments as examples or embodiments, wherein each claim is independently used as a separate embodiment, and it is considered that these embodiments can be combined with each other in various combinations or arrangements. The scope of the present invention should be determined with reference to the full scope of the equivalent forms of the attached claims and these claims.

Claims

1. A regional wind power prediction method, characterized in that: The method comprises: Obtain humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, photovoltaic power generation sequence, and wind power sequence of the wind farm in the target area; Based on the geographical location information, the wind farms in the target area are clustered to obtain multiple target sub-wind farms. The characteristics of the multiple target sub-wind farms are divided using the time correlation index and the volatility index; the time correlation index is the autocorrelation analysis coefficient and the periodicity coefficient, and the volatility index is the variation coefficient and the variance coefficient; Assign time correlation and volatility labels to multiple target sub-wind farms respectively, and divide the target area wind farms into multiple target sub-clusters according to the strength of time correlation and volatility. Each target sub-cluster uses a different prediction model to predict wind power. The prediction models of each target sub-cluster are integrated using different learning methods to establish multiple integrated models, and the models are evaluated to select the integrated model with the best performance as the final prediction model, and regional wind power prediction is realized based on the final prediction model; The time correlation index and volatility index are calculated as follows: Get the wind power sequence x1-x of the target sub-wind farm in the target area wind farm n ; The autocorrelation analysis coefficient of the target sub-wind farm is calculated by the following formula: In the formula, n represents the number of power sequences, represents the mean of the power series, k represents the lag order, r k represents the autocorrelation analysis coefficient; Based on the wind power sequence x1-x of the target sub-wind farm in the target area wind farm n , divided by month, construct the characteristic matrix X according to the monthly wind power sequence, perform periodic analysis on the characteristic matrix X, and obtain the periodic coefficient: In the formula, m is the number of samples, d is i is the difference in rank value of each power data point, r s is the periodic coefficient; The coefficient of variation is calculated using the following formula: In the formula, CV represents the coefficient of variation, n represents the number of power sequences, represents the mean of the power series, x i represents the i-th sample value; For wind power sequence x1-x n Normalize and get a normalized sequence: In the formula, t represents the normalized data, t i represents the i-th sample value, t min Represents the minimum value in the wind power sequence, t max Indicates the maximum value in the wind power sequence; Calculate the variance of the normalized sequence and get the variance coefficient σ 2 : In the formula, represents the average value of the normalized series; Multiple target sub-wind farms are respectively assigned labels of time correlation and volatility. According to the strength of time correlation and volatility, the wind farms in the target area are divided into multiple target sub-clusters, including: Assigning label characteristics to multiple target sub-wind farms according to the strength of the time correlation and volatility of the power sequence, wherein the label characteristics are divided into strong time correlation and strong volatility, strong time correlation and weak volatility, weak time correlation and strong volatility, and weak time correlation and weak volatility; Cluster multiple target sub-wind farms with label characteristics, and classify those with the same label characteristics into one category, thus obtaining four target sub-clusters; Each target sub-cluster uses different prediction models for wind power forecasting, including: For the target sub-clusters whose label characteristics are strong in time correlation and volatility, the prediction method is as follows: Get the wind power sequence x1-x of k target sub-wind farms in the target area k , a power series is obtained by adding a white noise time series to construct a new time series Yn(t). Based on the empirical mode decomposition, the new time series Yn(t) is decomposed into a series of intrinsic mode functions and the sum of residuals. After multiple iterations, the power series after denoising is obtained. The power sequence after denoising and decomposition is subjected to abnormal sequence elimination, and the remaining sequences are reorganized to obtain the decomposed and reorganized power sequence, and the humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, and photovoltaic power generation sequence of the target sub-wind farm are reduced in dimension by using principal component analysis, and the meteorological data after dimension reduction and the decomposed and reorganized power sequence are used as input data sets, and a first prediction model is selected for prediction, wherein the first prediction model includes a convolutional neural network and an LSTM network, wherein the convolutional neural network is used to extract the spatial features of the sequence, and the LSTM network is used to extract the temporal features of the sequence, and the spatial features and temporal features of the sequence are used to predict the power sequence; Each target sub-cluster uses different prediction models for wind power prediction, including: For the target sub-clusters whose label characteristics are strong in time correlation and volatility, the prediction method is as follows: Obtain the wind power sequences x1 - x of k target sub - wind farms in the target sub - cluster k , use the interquartile range method to conduct outlier tests on the power sequences, and return the outlier data and missing data values. Select the average number of terms N < K, and perform moving average filling on the power sequences: In the formula, k is the sequence sample number, N is the average number of items, To calculate the next data value to be filled, is the data value at the current moment, It is the first data value in the selected time observation interval; According to the humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, and photovoltaic power generation sequence of the target sub-wind farm, the first characteristic matrix is ​​constructed, and the filled power sequence is constructed as the second characteristic matrix. The first characteristic matrix and the second characteristic matrix are added to obtain the characteristic matrix X2: In the formula, x ak represents the humidity sequence of the kth target sub-wind farm, x bk represents the air pressure sequence of the kth target sub-wind farm, x ck represents the temperature sequence of the kth target sub-wind farm, x dk represents the sinusoidal sequence of wind directions at different heights of the kth target sub-wind farm, x ik represents the power sequence after filling of the kth target sub-wind farm; Sort the sequences in the feature matrix X2 and calculate the difference in rank values. The importance of each sequence is calculated using the following formula: Where, d i is the difference in rank value of each power data point, ρ is the importance of the sequence in the feature matrix X2; Select the sequence with importance greater than or equal to 0.8 as the input data, perform reverse processing on the input sequence data, select the LSTM model for prediction, and when all time steps are calculated, superimpose the result vector of the forward sequence with the result vector of the reverse sequence to obtain the final power prediction result; Each target sub-cluster uses different prediction models for wind power prediction, including: For the target sub-clusters whose label characteristics are weak time correlation and strong volatility, the prediction method is as follows: Get the wind power sequence x1-x of k target sub-wind farms in the target sub-cluster k , by adding the white noise time series to the power series to construct a new time series Yn(t), based on the empirical mode decomposition, the noise-contaminated time series Yn(t) is decomposed into a series of intrinsic mode functions and the sum of residuals, and multiple iterations are performed to obtain the power series after denoising decomposition: The power sequence after denoising and decomposition is subjected to abnormal sequence elimination, and the sequences that are not eliminated are reorganized to obtain the decomposed and reorganized power sequence. The humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, and photovoltaic power generation sequence of the target sub-wind farm are reduced in dimension by using principal component analysis. The reduced-dimensional meteorological data and the decomposed and reorganized power sequence are used as the input data sets of the CNN-LSTM model to obtain feature data. Multiple data are randomly selected from the decomposed and reorganized power sequence to construct a sample subset. Several features are randomly selected from multiple feature data to construct a feature subset. A decision tree is constructed based on the sample subset and the feature subset. At each node, an optimal feature is selected for splitting. Through continuous iteration, the prediction is stopped when each node contains only one sample. In this way, multiple decision trees are constructed, and the solution with the best prediction effect is selected as the final prediction result. Each target sub-cluster uses different prediction models for wind power prediction, including: For the target sub-clusters whose label characteristics are weak time correlation and weak volatility, the prediction method is as follows: Obtain the wind power sequences x1 - x of k target sub - wind farms in the target sub - cluster k , use the interquartile range method to conduct outlier tests on the power sequences, and return the outlier data and missing data values. Select the average number of terms N < K, and perform moving average filling on the data: In the formula, k is the number of sequence samples, and N is the number of average items; According to the humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, and photovoltaic power generation sequence of the target sub-wind farm, the first characteristic matrix is ​​constructed, and the filled power sequence is constructed as the second characteristic matrix. The first characteristic matrix and the second characteristic matrix are added to obtain the characteristic matrix X2: In the formula, x ak represents the humidity sequence of the kth target sub-wind farm, x bk represents the air pressure sequence of the kth target sub-wind farm, x ck represents the temperature sequence of the kth target sub-wind farm, x dk represents the sinusoidal sequence of wind directions at different heights of the kth target sub-wind farm, x ik represents the power sequence after filling of the kth target sub-wind farm; Taking the feature matrix X2 as the input data set, the XGBoost algorithm is used for prediction. The Shapley value of all possible sequence combinations in the feature matrix X2 is calculated and weighted summed to obtain the importance of the relevant features: In the formula, X represents the feature subset used in the model, x represents the eigenvalue vector of the power sequence, p is the number of features, val(X) represents the output value under the feature combination X, and val(X∪{x j }) represents the marginal contribution of the j-th eigenvalue vector to X after it is added to the feature subset X. Indicates the importance of the relevant features, x p represents the pth eigenvalue vector, x j represents the j-th eigenvalue vector; Sequences with importance greater than or equal to 0.8 are selected as input data and predicted using the liquid neural network.

2. The method according to claim 1, characterized in that Using time correlation index and volatility index, multiple target sub-wind farms are divided into characteristics, including: When the autocorrelation analysis coefficient of the target sub-wind farm is greater than 0.01, it is judged that the autocorrelation of the target sub-wind farm is strong; when the autocorrelation analysis coefficient of the target sub-wind farm is less than or equal to 0.01, it is judged that the autocorrelation of the target sub-wind farm is weak; When the periodicity coefficient of the target sub-wind farm is greater than 0.8, it is judged that the periodicity of the target sub-wind farm is strong; when the periodicity coefficient of the target sub-wind farm is less than or equal to 0.8, it is judged that the periodicity of the target sub-wind farm is weak; When the autocorrelation analysis coefficient and the periodicity coefficient of the target sub-wind farm are both judged to be strong, the time correlation of the target sub-wind farm is judged to be strong; when the autocorrelation analysis coefficient and the periodicity coefficient of the target sub-wind farm are not both judged to be strong, the time correlation of the target sub-wind farm is judged to be weak; When the coefficient of variation of the target sub-wind farm is greater than 100%, it is judged that the target sub-wind farm has strong diversity; when the coefficient of variation of the target sub-wind farm is less than or equal to 100%, it is judged that the target sub-wind farm has weak diversity; When the variance coefficient of the target sub-wind farm is greater than 0.06, it is judged that the discreteness of the target sub-wind farm is strong; when the variance coefficient of the target sub-wind farm is less than or equal to 0.06, it is judged that the discreteness of the target sub-wind farm is weak; When the diversity and discreteness of the target sub-wind farm are both judged to be strong, the volatility of the target sub-wind farm is judged to be strong; when the diversity and discreteness of the target sub-wind farm are not both judged to be strong, the volatility of the target sub-wind farm is judged to be weak.

3. A regional wind power prediction device, characterized in that: The device comprises: A data acquisition module is configured to acquire a humidity sequence, an air pressure sequence, a temperature sequence, a sine sequence of wind directions at different heights, a cosine sequence of wind directions at different heights, a wind speed sequence at different heights, an illumination sequence, a photovoltaic power generation sequence, and a wind power sequence of a wind farm in a target area; The characteristic classification module is configured to cluster the wind farms in the target area based on the geographical location information to obtain multiple target sub-wind farms, and to perform characteristic classification on the multiple target sub-wind farms using the time correlation index and the volatility index; wherein the time correlation index is the autocorrelation analysis coefficient and the periodicity coefficient, and the volatility index is the variation coefficient and the variance coefficient; The first prediction module is configured to assign time correlation and volatility labels to multiple target sub-wind farms respectively, and divide the target area wind farms into multiple target sub-clusters according to the strength of the time correlation and volatility, and each target sub-cluster uses a different prediction model to predict wind power; The second prediction module is configured to integrate the prediction models of each target sub-cluster using different learning methods, establish multiple integrated models, and perform model evaluation, select the integrated model with the best performance as the final prediction model, and realize regional wind power prediction based on the final prediction model; Multiple target sub-wind farms are respectively assigned labels of time correlation and volatility. According to the strength of time correlation and volatility, the wind farms in the target area are divided into multiple target sub-clusters, including: Assigning label characteristics to multiple target sub-wind farms according to the strength of the time correlation and volatility of the power sequence, wherein the label characteristics are divided into strong time correlation and strong volatility, strong time correlation and weak volatility, weak time correlation and strong volatility, and weak time correlation and weak volatility; Cluster multiple target sub-wind farms with label characteristics, and classify those with the same label characteristics into one category, thus obtaining four target sub-clusters; Each target sub-cluster uses different prediction models for wind power prediction, including: For the target sub-clusters whose label characteristics are strong in time correlation and volatility, the prediction method is as follows: Get the wind power sequence x1-x of k target sub-wind farms in the target area k , a power series is obtained by adding a white noise time series to construct a new time series Yn(t). Based on the empirical mode decomposition, the new time series Yn(t) is decomposed into a series of intrinsic mode functions and the sum of residuals. After multiple iterations, the power series after denoising is obtained. The power sequence after denoising and decomposition is subjected to abnormal sequence elimination, and the remaining sequences are reorganized to obtain the decomposed and reorganized power sequence, and the humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, and photovoltaic power generation sequence of the target sub-wind farm are reduced in dimension by using principal component analysis, and the meteorological data after dimension reduction and the decomposed and reorganized power sequence are used as input data sets, and a first prediction model is selected for prediction, wherein the first prediction model includes a convolutional neural network and an LSTM network, wherein the convolutional neural network is used to extract the spatial features of the sequence, and the LSTM network is used to extract the temporal features of the sequence, and the spatial features and temporal features of the sequence are used to predict the power sequence; Each target sub-cluster uses different prediction models for wind power prediction, including: For the target sub-clusters whose label characteristics are strong in time correlation and volatility, the prediction method is as follows: Obtain the wind power sequences x1 - x of k target sub - wind farms in the target sub - cluster k , use the interquartile range method to conduct outlier tests on the power sequences, and return the outlier data and missing data values. Select the average number of terms N < K, and perform moving average filling on the power sequences: In the formula, k is the sequence sample number, N is the average number of items, To calculate the next data value to be filled, is the data value at the current moment, It is the first data value in the selected time observation interval; According to the humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, and photovoltaic power generation sequence of the target sub-wind farm, the first characteristic matrix is ​​constructed, and the filled power sequence is constructed as the second characteristic matrix. The first characteristic matrix and the second characteristic matrix are added to obtain the characteristic matrix X2: In the formula, x ak represents the humidity sequence of the kth target sub-wind farm, x bk represents the air pressure sequence of the kth target sub-wind farm, x ck represents the temperature sequence of the kth target sub-wind farm, x dk represents the sinusoidal sequence of wind directions at different heights of the kth target sub-wind farm, x ik represents the power sequence after filling of the kth target sub-wind farm; Sort the sequences in the feature matrix X2 and calculate the difference in rank values. The importance of each sequence is calculated using the following formula: Where, d i is the difference in rank value of each power data point, ρ is the importance of the sequence in the feature matrix X2; Select the sequence with importance greater than or equal to 0.8 as the input data, perform reverse processing on the input sequence data, select the LSTM model for prediction, and when all time steps are calculated, superimpose the result vector of the forward sequence with the result vector of the reverse sequence to obtain the final power prediction result; Each target sub-cluster uses different prediction models for wind power prediction, including: For the target sub-clusters whose label characteristics are weak time correlation and strong volatility, the prediction method is as follows: Get the wind power sequence x1-x of k target sub-wind farms in the target sub-cluster k , by adding the white noise time series to the power series to construct a new time series Yn(t), based on the empirical mode decomposition, the noise-contaminated time series Yn(t) is decomposed into a series of intrinsic mode functions and the sum of residuals, and multiple iterations are performed to obtain the power series after denoising decomposition: The power sequence after denoising and decomposition is subjected to abnormal sequence elimination, and the sequences that are not eliminated are reorganized to obtain the decomposed and reorganized power sequence. The humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, and photovoltaic power generation sequence of the target sub-wind farm are reduced in dimension by using principal component analysis. The reduced-dimensional meteorological data and the decomposed and reorganized power sequence are used as the input data sets of the CNN-LSTM model to obtain feature data. Multiple data are randomly selected from the decomposed and reorganized power sequence to construct a sample subset. Several features are randomly selected from multiple feature data to construct a feature subset. A decision tree is constructed based on the sample subset and the feature subset. At each node, an optimal feature is selected for splitting. Through continuous iteration, the prediction is stopped when each node contains only one sample. In this way, multiple decision trees are constructed, and the solution with the best prediction effect is selected as the final prediction result. Each target sub-cluster uses different prediction models for wind power prediction, including: For the target sub-clusters whose label characteristics are weak time correlation and weak volatility, the prediction method is as follows: Obtain the wind power sequences x1 - x of k target sub - wind farms in the target sub - cluster k , use the interquartile range method to conduct outlier tests on the power sequences, and return the outlier data and missing data values. Select the average number of terms N < K, and perform moving average filling on the data: In the formula, k is the number of sequence samples, and N is the number of average items; According to the humidity sequence, air pressure sequence, temperature sequence, wind direction sine sequence at different heights, wind direction cosine sequence at different heights, wind speed sequence at different heights, illumination sequence, and photovoltaic power generation sequence of the target sub-wind farm, the first characteristic matrix is ​​constructed, and the filled power sequence is constructed as the second characteristic matrix. The first characteristic matrix and the second characteristic matrix are added to obtain the characteristic matrix X2: In the formula, x ak represents the humidity sequence of the kth target sub-wind farm, x bk represents the air pressure sequence of the kth target sub-wind farm, x ck represents the temperature sequence of the kth target sub-wind farm, x dk represents the sinusoidal sequence of wind directions at different heights of the kth target sub-wind farm, x ik represents the power sequence after filling of the kth target sub-wind farm; Taking the feature matrix X2 as the input data set, the XGBoost algorithm is used for prediction. The Shapley value of all possible sequence combinations in the feature matrix X2 is calculated and weighted summed to obtain the importance of the relevant features: In the formula, X represents the feature subset used in the model, x represents the eigenvalue vector of the power sequence, p is the number of features, val(X) represents the output value under the feature combination X, and val(X∪{x j }) represents the marginal contribution of the j-th eigenvalue vector to X after it is added to the feature subset X. Indicates the importance of the relevant features, x p represents the pth eigenvalue vector, x j represents the j-th eigenvalue vector; Sequences with importance greater than or equal to 0.8 are selected as input data and predicted using the liquid neural network. 4 . A non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, executes the method according to claim 1 .

Images