A wind power short-term power prediction correction method considering unit classification

Abstract

A correction method of wind power short-term power prediction considering unit classification, the average wind speed of the wind farm is taken as the optimization target, and the ideal power curve is substituted to obtain the predicted power P1; the wind turbine units are classified into several representative groups, and several groups of predicted wind speeds are generated to obtain the predicted power P2 of the wind farm; the initial predicted power is corrected through the error back propagation algorithm; the correction method can organically combine multiple NWP and internal information of the unit, effectively avoid extreme false positives, improve the accuracy of the assessment day, and reduce the short-term power prediction assessment power.

Description

Technical Field

[0001] The present application relates to the field of wind power generation, and in particular to a correction method for short-term wind power prediction considering unit classification. Background Art

[0002] The stability of wind power generation is significantly affected by factors such as wind direction and speed. This is why wind power generation suffers from drawbacks such as volatility and intermittency. With the rapid increase in the number of wind farms and wind turbine capacity, a growing number of researchers are focusing on predicting wind farm output power, developing advanced reactive power control technologies for large wind farms, and developing more accurate models to simulate the dynamic performance of wind farms under various operating conditions. Consequently, more accurate wind power prediction methods are essential to ensuring that wind farms can become irreplaceable renewable energy power plants.

[0003] Traditional wind power forecasting methods can be categorized into physical models, statistical models, and a combination of the two. With the improvement of meteorological numerical model performance, the current mainstream approach is a combined approach. This involves using machine learning methods to perform error correction and post-processing on numerical weather prediction (NWP) wind speed forecasts to obtain more accurate wind speeds. Wind power forecasts are then made based on these optimized wind speeds. By incorporating multi-location NWP information, the problem of single-location NWP failing to accurately describe wind farm meteorological information can be effectively avoided, improving forecast accuracy. Researchers simultaneously considered multi-location NWP and atypical meteorological characteristics, performing feature selection based on the maximum relevance-minimum redundancy principle and principal component analysis to reduce overall redundancy and effectively shorten the training time of the forecast model. Furthermore, by considering the temporal correlation of wind turbine data, cluster analysis was used to generate data for different weather types. Correction and post-processing of the NWP data for each weather type further enhanced forecasting capabilities. Researchers also considered power trends to achieve even better performance, which can, to a certain extent, address the problem of large power fluctuations, also known as the ramping problem.

[0004] However, these studies only focus on a single wind speed within a wind farm, lacking sufficient consideration of internal variations within large wind farms and failing to utilize historical information about each turbine within the unit. Unit classification is an important tool for studying the power generation patterns of large wind farms. Related research has also demonstrated that power output models based on wind farm unit classification can effectively represent actual power output and reduce the complexity and time of the power modeling process. In recent years, research on short-term power forecasting using unit classification has shown that different wind turbines within a wind farm vary significantly due to the influence of terrain, wakes, and turbulence, and that power forecasting based on unit classification is superior to power forecasting based on a single wind speed. Unfortunately, these studies simply model power forecasts for all subclasses of the unit and use a superposition method to sum the predicted power. This is affected by the prediction accuracy of each subclass, and the cumulative model error results in low accuracy. Summary of the Invention

[0005] To address these shortcomings, the present invention proposes a power prediction and correction model that considers multiple NWPs and turbine classification. This model performs wind speed correction post-processing based on multiple NWP information and a single wind speed, and generates an initial power forecast using the ideal power curve. Hierarchical clustering is then used to classify wind turbines within the wind farm into representative groups, generating several sets of predicted wind speeds. The initial power forecast is then corrected using an error backpropagation algorithm.

[0006] A correction method for wind power short-term power forecast considering unit classification is characterized in that the method comprises the following steps:

[0007] Step 1: Feature selection of multiple NWPs; taking the average wind speed of the wind farm as the optimization target, the optimized forecast wind speed is obtained by the XGBoost model, and substituted into the ideal power curve to obtain the predicted power P1;

[0008] Step 2: Based on hierarchical clustering, classify the wind turbines in the wind farm into several representative groups and generate several groups of predicted wind speeds; divide the S wind turbines into G groups, and use the average wind speed of each group as the optimization target to obtain the optimized forecast wind speed W of each group. Substitute it into the ideal power curve to obtain the predicted power of each group, and accumulate it to obtain the wind farm predicted power P2;

[0009] When classifying the units, the agglomerative method in the hierarchical clustering method is used. A bottom-up strategy is adopted. The similarity measurement vector matrix of all wind turbines in the data unit is input, and wind turbines with similar distances are gradually merged into one category until all wind turbines are clustered into one large category.

[0010] Step 3: Correct the initial prediction power through the error back propagation algorithm. First, initialize the structure and weights of the BP network. Then, forward calculate the input signal and output signal of each layer of neurons in the BP network based on the input sample. Calculate the reverse error based on the expected output and correct the weight. If the error is less than the given value or the number of iterations exceeds the set value, the learning is terminated and the initial prediction power is finally corrected.

[0011] Advantages of the technical solution of the present invention:

[0012] Through case analysis, the proposed method can organically combine multiple NWPs and internal unit information, effectively avoid extreme false alarms, improve the accuracy of assessment days, and reduce the power consumption of short-term power forecast assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Wind power prediction block diagram DETAILED DESCRIPTION

[0014] The present invention will be further described below in conjunction with the accompanying drawings. It should be understood that the contents described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0015] Explanation of relevant technical terms:

[0016] 1. Numerical Weather Prediction (NWP)

[0017] 2. Error Back Propagation Training (BP)

[0018] 3. Extreme Gradient Boosting (XGBoost)

[0019] 4. Extreme Learning Machine (ELM)

[0020] 5. Fuzzy C Means (FCM)

[0021] like Figure 1 shown.

[0022] Step 1: Feature selection for multiple NWPs. Taking the average wind speed of the wind farm as the optimization target, the XGBoost model is used to obtain the optimized forecast wind speed. Substituting this into the ideal power curve, the predicted power P1 is obtained.

[0023] To cover a wider area and diverse terrain, NWP meteorological data should be collected from four to nine typical locations around the wind farm. Each location should include information on five parameters: meridional wind, zonal wind, temperature, humidity, and pressure, at four altitude levels. This means a maximum of 180 dimensions of data per moment. NWP forecasts are provided for the next 72 hours, updated daily, with a temporal resolution of 15 minutes.

[0024] The preliminary power forecast model uses a combination of physical and statistical models to predict wind speed. To account for NWP information at multiple locations, the Extreme Gradient Boosting (XGBoost) model, suitable for processing high-dimensional samples, was used as the statistical model. The NWP wind speed forecast, corrected by the XGBoost model, was substituted into the ideal power curve to obtain the initial power forecast.

[0025] NWP is used as an input feature, and the predicted wind speed is obtained after the XGBoost model. The accuracy of the predicted wind speed is the key to the accuracy of the power prediction. XGBoost is a modified version of the Gradient Boosting (GBoost) algorithm. Boosting is a serial integration method. The t-th weak learner optimizes the residual part of the previous t-1 learners. Learners are gradually introduced to maximize the reduction of the objective function, thereby obtaining a strong learner. Because multiple NWP data have multi-dimensional (180) features, XGBoost samples the features when constructing the learner and adds regularization terms to prevent overfitting. At the same time, the training speed is fast, which is suitable for post-processing of NWP wind speed correction. The process of training the XGBoost model is as follows:

[0026] (1) For all samples, initialize a CART tree and predict the output

[0027]

[0028] Where: c is the predicted output set.

[0029] (2) Calculate the negative gradient r for all samples ti for

[0030]

[0031] (3) Using the negative gradient, fit a CART regression tree to obtain the t-th regression tree, whose corresponding leaf node area is R tj , j = 1, 2, ..., J. Where J is the number of leaf nodes of regression tree t.

[0032] (4) For leaf region j = 1, 2, ..., J, calculate the best fitting value c tj for

[0033]

[0034] (5) After T rounds of updates, the strong learner is finally predicted to output

[0035]

[0036] Step 2: In order to introduce the internal difference information of the wind farm, the wind turbines of unit S are classified into G groups based on hierarchical clustering, and the average wind speed of the subclasses of group G is generated. The predicted wind speed of group G is generated by correction using the XGBoost model.

[0037] First, based on hierarchical clustering, the wind turbines within the wind farm are classified into several representative groups to generate several sets of predicted wind speeds. Then, the S wind turbines are divided into G groups. The average wind speed of each group is used as the optimization target to obtain the optimized forecast wind speed W for each group. Substituting this into the ideal power curve yields the predicted power for each group, and summing them up to obtain the wind farm predicted power P2.

[0038] When classifying turbine units, hierarchical clustering does not require a priori specification of the number of subclasses. Compared to methods such as the k-means center method (kmeans) and the fuzzy C-means center method (FCM), constraints can be added to adjust the number of subclasses. Based on this, the agglomerative approach within hierarchical clustering employs a bottom-up strategy, inputting a matrix of similarity metrics for all turbines within the unit. It gradually merges turbines with similar distances into a single class, until all turbines are clustered into a single large class.

[0039] The specific process of hierarchical clustering is as follows:

[0040] 1) Let N be the number of samples, M be the number of wind turbines, select M1 wind turbines connected to the grid in the same time period, and construct an N×M1 matrix;

[0041] 2) Based on the Euclidean distance, Pearson correlation coefficient, and entropy correlation coefficient, the similarity measurement matrix between the M1 wind turbines is calculated.

[0042] 3) Use the squared deviation method to calculate the distance d between different classes. Each time a class is reduced, the squared deviation S X To increase, choose S X The two categories with the smallest increase are merged, as shown in formula (1).

[0043]

[0044] Where: K is the total number of clusters in the data set; is the mean of the distances.

[0045] 4) If the number of classes is not equal to 1, repeat step 3), otherwise go to step 5).

[0046] 5) Ensure that the number of fans in a class does not exceed half of the total number of fans, and take the maximum number of subclasses.

[0047] 6) Calculate the clustering evaluation standard index and determine the method for calculating the similarity measure.

[0048] Step 3: Correct the initial prediction power using the back propagation (BP) algorithm. First, initialize the structure and weights of the BP neural network. Then, forward-calculate the input and output signals of each neuron layer of the BP neural network based on the input sample. Calculate the reverse error based on the expected output and correct the weights. If the error is less than a given value or the number of iterations exceeds a set value, learning ends and the initial prediction power is finally corrected.

[0049] BP neural networks can approximate any nonlinear mapping with arbitrary precision; they can learn and adapt to unknown information; and their distributed information storage and processing structure offers a degree of fault tolerance. Therefore, the resulting system possesses robustness and is suitable for handling complex problems. Through case analysis, the proposed method effectively combines multiple NWPs with internal unit information, effectively avoiding extreme false alarms, improving the accuracy of assessment days, and reducing the amount of power required for short-term power forecast assessments.

[0050] The input data includes wind speed W(t), wind direction D(t), and wind power P(t) at time t, as well as wind speed W(t+1) and wind direction D(t+1) at time t+1. The output is wind power P(t+1) at time t+1.

[0051] The output formula of the hidden layer of the BP neural network is shown in formula (6):

[0052]

[0053] Where: f is the transfer function; w ij is the weight of the i-th neuron in the input layer and the j-th neuron in the hidden layer; x i is the input data; a j is the threshold of the jth neuron in the hidden layer.

[0054] The output layer prediction formula is shown in formula (7):

[0055]

[0056] Where: h j is the output of the jth neuron in the hidden layer; w jp is the weight of the jth neuron in the hidden layer and the pth neuron in the output layer; b p is the threshold of the p-th neuron in the output layer.

[0057] Finally, it should be noted that the foregoing description is merely an explanation of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail, those skilled in the art will be able to modify the aforementioned technical solutions or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A correction method for wind power short-term power forecast considering unit classification, characterized in that: The method comprises the following steps: Step 1: Feature selection of multiple NWPs; taking the average wind speed of the wind farm as the optimization target, the optimized forecast wind speed is obtained by the XGBoost model, and substituted into the ideal power curve to obtain the predicted power P1; Step 2: Based on hierarchical clustering, classify the wind turbines in the wind farm into several representative groups and generate several groups of predicted wind speeds; divide the S wind turbines into G groups, and use the average wind speed of each group as the optimization target to obtain the optimized forecast wind speed W of each group. Substitute it into the ideal power curve to obtain the predicted power of each group, and accumulate it to obtain the wind farm predicted power P2; When classifying the units, the agglomerative method in the hierarchical clustering method is used. A bottom-up strategy is adopted. The similarity measurement vector matrix of all wind turbines in the data unit is input, and wind turbines with similar distances are gradually merged into one category until all wind turbines are clustered into one large category. Step 3: Correct the initial prediction power through the error back propagation algorithm. First, initialize the structure and weights of the BP network. Then, forward calculate the input signal and output signal of each layer of neurons in the BP network based on the input sample. Calculate the reverse error based on the expected output and correct the weight. If the error is less than the given value or the number of iterations exceeds the set value, the learning is terminated and the initial prediction power is finally corrected.

2. The correction method for wind power short-term power forecast considering unit classification according to claim 1 is characterized in that: In step one: To cover a larger area of ​​the wind farm and various terrains, NWP meteorological data for 4-9 typical locations around the wind farm should be selected. Each location contains information on meridional wind, zonal wind, temperature, humidity, and air pressure parameters for four altitude layers. NWP is 72-hour forecast data, updated once a day, and the time resolution is interpolated to 15 minutes.

3. The correction method for wind power short-term power forecast considering unit classification according to claim 1 is characterized in that: In step one: The process of training the XGBoost model is as follows: (1) For all samples, initialize a CART tree and predict the output Where: c is the predicted output set; (2) Calculate the negative gradient r for all samples ti for (3) Using the negative gradient, fit a CART regression tree to obtain the t-th regression tree, whose corresponding leaf node area is R tj , j = 1, 2, ..., J; Where J is the number of leaf nodes of regression tree t; (4) For leaf region j = 1, 2, ..., J, calculate the best fitting value c tj for (5) After T rounds of updates, the strong learner is finally predicted to output 4. The correction method for wind power short-term power forecast considering unit classification according to claim 3 is characterized in that: In step 2: The specific process of hierarchical clustering is as follows: 1) Let N be the number of samples, M be the number of wind turbines, select M1 wind turbines connected to the grid in the same time period, and construct an N×M1 matrix; 2) Based on the Euclidean distance, Pearson correlation coefficient, and entropy correlation coefficient, calculate the similarity measurement matrix between the M1 wind turbines; 3) Use the squared deviation method to calculate the distance d between different classes. Each time a class is reduced, the squared deviation S X To increase, choose S X The two categories with the smallest increase are merged, as shown in formula (1), Where: K is the total number of clusters in the data set; is the mean of the distances; 4) If the number of classes is not equal to 1, repeat step 3), otherwise go to step 5); 5) Ensure that the number of fans in a class does not exceed half of the total number of fans, and take the maximum number of subclasses; 6) Calculate the clustering evaluation standard index and determine the method for calculating the similarity measure.

5. The correction method for wind power short-term power forecast considering unit classification according to claim 1 is characterized in that: In step three: The input data includes wind speed W(t), wind direction D(t), wind power P(t) at time t, as well as wind speed W(t+1) and wind direction D(t+1) at time t+1. The output is wind power P(t+1) at time t+1. The output formula of the hidden layer of the BP neural network is shown in formula (6): Where: f is the transfer function; w ij is the weight of the i-th neuron in the input layer and the j-th neuron in the hidden layer; x i is the input data; a j is the threshold of the jth neuron in the hidden layer; The output layer prediction formula is shown in formula (7): Where: h j is the output of the jth neuron in the hidden layer; w jp is the weight of the jth neuron in the hidden layer and the pth neuron in the output layer; b p is the threshold of the p-th neuron in the output layer.

Images