Wind power forecasting method based on XGBoost extended financial factors

Abstract

A wind power prediction method based on XGBoost and extended financial factors includes a time window-based feature generation method to obtain potential correlations in wind power generation data; a wind power K-line-based feature generation method to generate financial factors in the wind power generation data; and an XGBoost algorithm based on gradient boosting regression trees to predict ultra-short-term wind power generation data. This invention, based on existing wind power generation data, generates data containing potential power correlations within a short period through time windows, combines the time windows to generate wind power-specific K-lines, and generates unique wind power KDJ and MACD indicators from these K-lines. Finally, the dataset containing meteorological data, time window data, and financial factors is fed into the XGBoost model to predict future wind power generation, achieving extremely high prediction accuracy and extremely short prediction time.

Description

Technical Field

[0001] The present invention belongs to the field of wind power technology, as well as the field of time series data analysis and mining technology, and also belongs to the field of big data analysis and application technology, and particularly relates to a wind power prediction method based on XGBoost extended financial factors. Background Art

[0002] Wind farms conduct power market bidding and wind farm operation and maintenance by conducting wind power forecasting. Based on the forecast results of the wind power forecasting system, the grid dispatching department makes reasonable power generation plan arrangements, reduces the system's spinning reserve capacity, and improves the economic efficiency of grid operation. At the same time, by predicting the fluctuation of wind power in advance, the operation mode and response measures are reasonably arranged to improve the safety and reliability of the grid. However, due to the single data characteristics and the long running time of the existing model, the forecasting model cannot be deployed in the actual environment of the wind farm. Summary of the invention

[0003] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a wind power prediction method based on XGBoost extended financial factors, which achieves the purpose of fast and accurate prediction of ultra-short-term wind power by constructing a time window for existing data, calculating financial factors, and using the XGBoost algorithm based on the gradient boosting regression tree.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] A wind power prediction method based on XGBoost extended financial factors includes the following steps:

[0006] Step 1, forming an original time series with original data of the wind farm, wherein the original data includes temperature and pressure data monitored by the meteorological department and wind speed, wind direction, and real-time power monitored by the wind farm;

[0007] Step 2, selecting the power at the current time point, the power c minutes before the current time point, and the difference between the power at the current time point and the power c minutes before the current time point in the original time series to form a time window;

[0008] Step 3, select the original time series, the power at the current time point is the starting power of 1 hour, the power 60 minutes after the current time point is the ending power of 1 hour, the maximum power within 60 minutes is the maximum power within 1 hour, and the minimum power within 60 minutes is the minimum power within 1 hour; all the data in the original time series are drawn according to this step. Wind power K line;

[0009] Step 4, based on the wind power K-line, calculate the financial factors MACD and KDJ that are unique to wind power;

[0010] Step 5, combining the original data, the time window generated in step 2, and the financial factors MACD and KDJ calculated in step 4 into a new time series, and inputting them into the XGBoost algorithm based on the gradient boosting regression tree to predict future wind power generation.

[0011] Compared with the prior art, the present invention aims at the problem of too few features in the process of predicting wind power data, and adopts the method of building a time window and calculating financial factors to achieve the purpose of increasing data features, and finally passes the new time series into the prediction model to predict the ultra-short-term wind power. The present invention can solve the problem of too few features in the original wind power data to a great extent, improve the accuracy of the prediction model and reduce the prediction time, and provide more accurate prediction data for the power grid dispatching department and wind farms. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0013] Figure 2 It is a wind power generation power K-line diagram drawn by the present invention.

[0014] Figure 3 It is a schematic diagram of the XGBoost algorithm process in an embodiment of the present invention.

[0015] Figure 4 It is an experimental result diagram and a feature importance score diagram of Example 1 of the present invention; wherein, (a) is a comparison between the true value and the predicted value of the first group of experiments, (b) is a comparison between the true value and the predicted value of the second group of experiments, (c) is a comparison between the true value and the predicted value of the third group of experiments, (d) is a comparison between the true value and the predicted value of the fourth group of experiments, (e) is a comparison between the true value and the predicted value of the fifth group of experiments, (f) is a comparison between the true value and the predicted value of the sixth group of experiments, (g) is a comparison between the true value and the predicted value of the seventh group of experiments, (h) is the feature importance score of the first group of experiments, (i) is the feature importance score of the second group of experiments, (j) is the feature importance score of the third group of experiments, (k) is the feature importance score of the fourth group of experiments, (l) is the feature importance score of the fifth group of experiments, (m) is the feature importance score of the sixth group of experiments, and (n) is the feature importance score of the seventh group of experiments.

[0016] Figure 5 It is an experimental result diagram and a feature importance score diagram of Example 2 of the present invention; wherein, (a) is a comparison between the true value and the predicted value of the first group of experiments, (b) is a comparison between the true value and the predicted value of the second group of experiments, (c) is the feature importance score of the first group of experiments, and (d) is the feature importance score of the second group of experiments.

[0017] Figure 6 It is a work flow chart of the present invention. DETAILED DESCRIPTION

[0018] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0019] like Figure 1 As shown, the present invention is a wind power prediction method, comprising the following steps:

[0020] Step 1, the original time series is formed with the original data of the wind farm. In the present invention, the original data should at least include the temperature and pressure data monitored by the meteorological department and the wind speed, wind direction, and real-time power monitored by the wind farm. For example, when only the above data are used without other data, the purpose of the present invention can still be achieved, that is, the original data can also only consist of the temperature and pressure data monitored by the meteorological department and the wind speed, wind direction, and real-time power monitored by the wind farm.

[0021] The original time series format of the present invention is a two-dimensional array of time and monitoring data, an example of which is shown in Table 1:

[0022] Table 1, schematic diagram of two-dimensional array

[0023]

[0024] Step 2: Select the power at the current time point, the power c minutes before the current time point, and the difference between the power at the current time point and the power c minutes before the current time point in the original time series to form a time window. The potential correlation in the wind power data can be obtained through the feature generation method based on the time window.

[0025] For example, one method of constructing a time window in this step is as follows:

[0026] For time series data A, let the data value at time i+1 be t i+1 is the data value to be predicted, and the time series data value t at time i is taken i , i-1 time series data value t i-1 , the data difference between time i and time i-1 A time window is constructed to obtain new time series data B. The conversion algorithm is as follows:

[0027]

[0028] In the formula, n is the total number of moments divided in the time series data A in the present invention. In the present invention, for convenience of calculation and moderate amount of calculation, c=15 is selected, that is, the power after 15 minutes is predicted.

[0029] Step 3: Select the original time series, the power at the current time point is the start power of 1 hour, the power 60 minutes after the current time point is the end power of 1 hour, the maximum power within 60 minutes is the maximum power within 1 hour, and the minimum power within 60 minutes is the minimum power within 1 hour; all the data in the original time series are drawn according to this step. The feature generation method based on the K-line is used to generate financial factors in the wind power data in the next step.

[0030] For example, a calculation method of the wind power generation power K line in this step is as follows:

[0031] The first step is to integrate the sampling data of every 5 adjacent sampling points into one piece of data according to the frequency of sampling once every c minutes. Taking the sampling interval of 15 minutes (that is, c is 15) and there are 96 sampling points every day as an example, the integration method is shown in Table 2 and Table 3:

[0032] Table 2, sampling data (unit: MW)

[0033] 00:00 00:15 00:30 00:45 01:00 01:15 01:30 01:45 02:00 02:15 5.6582 7.2793 8.5115 8.5169 7.1129 6.1613 7.6255 8.4773 10.9395 11.7586

[0034] Table 3, integrated data

[0035] 1 5.6582 7.2793 8.5115 8.5169 7.1129 2 7.2793 8.5115 8.5169 7.1129 6.1613 3 8.5115 8.5169 7.1129 6.1613 7.6255 4 8.5169 7.1129 6.1613 7.6255 8.4773 5 7.1129 6.1613 7.6255 8.4773 10.9395 6 6.1613 7.6255 8.4773 10.9395 11.7586

[0036] In the second step, the maximum value among the five adjacent sampling points is regarded as the highest price in the financial market; the minimum value among the five sampling points is regarded as the lowest price in the financial market; the power output value of the first sampling point is regarded as the opening price in the financial market; the power output value of the last sampling point is regarded as the closing price in the financial market; when the closing price is higher than the opening price, the opening price is below and the closing price is above, and the rectangular column between the two is drawn in red; when the closing price is lower than the opening price, the opening price is above and the closing price is below, and the rectangular column between the two is drawn in green.

[0037] In the embodiment, the drawing results are as follows Figure 2 As shown, the vertical axis unit is megawatt, represented by MW, and the horizontal axis unit is hour, represented by h.

[0038] Step 4, based on the wind power K-line, calculate the financial factors MACD and KDJ that are unique to wind power.

[0039] In the step, the KDJ method for wind power generation is proposed, which integrates the concept of moving average and forms a relatively accurate basis for power increase and decrease. The KDJFP (KDJ For Power) indicator is mainly used to study the relationship between the highest power output value, the lowest power output value and the sampling end value during the design process, and also integrates some advantages of momentum concept, strength index and moving average. Therefore, the trend of wind power generation can be judged relatively quickly and intuitively.

[0040] KDJ is based on ultra-short-term wind power forecast, and its cycle is selected as 3 hours. Assume that the power at the end of 3 hours is C 3 ; The maximum power within 3 hours is H 3 ; The minimum power within 3 hours is L 3 ; The immature random value is RSV; K is a quick confirmation indicator, and the K value at the xth hour is K x ; D is the slow trunk indicator, and the D value at the xth hour is D x ; K value and D value float between 0 and 100. J is the trend direction indicator, and the J value of the xth hour is J x ; The RSV value at the xth hour is RSV x ; K x-1 K value for the previous 3 hours. If it does not exist, use 50 instead. D x-1 D value for the previous 3 hours. If it does not exist, use 50 instead. The calculation method is as follows:

[0041]

[0042] In this step, the MACD method for wind power generation is also proposed, which represents the possible development and change trend of the current wind power generation by the discreteness and aggregation of the fast and slow moving averages. The MACDFP (MACD For Power) indicator mainly studies the relationship between the fast moving average and the slow moving average during the design process, and calculates the deviation between the fast moving average and the slow moving average. Through the combined analysis of MACDFP and KDJFP, the accuracy of power prediction can be further improved.

[0043] MACD is based on ultra-short-term wind power forecast. Its fast moving average period is selected as 3 hours, and its slow moving average period is selected as 5 hours. Suppose the power at the end of x hours is C x ; The fast moving average of x hours is The x-hour slow moving average is The difference value for x hours is DIF x ; The smoothed moving average of x hours is DEA x ; Represents the EMA of the previous 3 hours (the fast moving average period is 3 hours, here the previous 3 hours refers to the previous period) fast If the value does not exist, 0 is used instead; Represents the EMA of the previous 5 hours slow If the value does not exist, use 0 instead; the bar value for x hour is The calculation method is as follows:

[0044]

[0045] In step 5, the original data, the time window generated in step 2, and the financial factors MACD and KDJ calculated in step 4 are combined into a new time series, and input into the XGBoost algorithm based on the gradient boosting regression tree to predict future wind power generation.

[0046] In this step 5, the new time series consisting of meteorological data, time windows, and financial factors is input into the XGBoost algorithm based on the gradient boosting regression tree to predict the wind power generation in the next c minutes. The XGBoost algorithm consists of multiple CART trees, and the integrated learning of multiple CART trees is realized through the gradient boosting regression method. In the training stage, each decision tree learns the residual between the target value and the sum of the predicted values ​​of all previous trees. Each newly added tree is trained on the basis of the trained tree. At the same time, when selecting the split point, the optimal split point will be selected according to the optimal feature. This process is similar to a balanced binary tree. Finally, multiple decision trees make a decision together, and the results of all trees are added up as the final prediction result. The brief process is as follows:

[0047] In the first step, XGBoost reads the new time series into the model and selects representative features to build a CART tree (the way of selecting features is similar to that of a balanced binary tree, that is, after classification by features, the number of left subtrees is the same as that of right subtrees);

[0048] The second step is to build all CART trees according to the algorithm design (such as the number of trees, the depth of the tree, etc.);

[0049] The third step is to test the test set. The XGBoost algorithm will obtain the prediction results in each CART tree based on the test data, and sum up all the prediction results to get the final prediction result.

[0050] The algorithm process is shown as follows Figure 3 As shown:

[0051] Assume that the number of trees in the XGBoost design tree is 2 and the depth is 3; after the test data (wind speed = 3.8, J = 12, D = 39, dif = 55) is passed into the model, the prediction result is: 0.612 + 0.566 = 1.178 (MW)

[0052] Based on the original wind power data, the present invention generates data containing potential power correlation in a short time through a time window, generates a K-line unique to wind power in combination with the time window, and generates a unique wind power KDJ and MACD through the wind power K-line. Finally, the data set containing meteorological data, time window data, and financial factors is passed into the XGBoost model to predict future wind power. Experiments show that this method has extremely high prediction accuracy and extremely short prediction time, which is of great significance for deployment in wind farms. The following is the specific experimental process of the present invention.

[0053] In a specific embodiment of the present invention, the hardware is a computer, and the configuration includes the following hardware environment: CPU: 2 Intel Xeon 6130 processors (2.1GHz / 16c) / 2666MHz / 10.4GT; GPU: 6 16G_TESLA-P100_4096B_P_CAC; memory: 16 32G ECC Registered DDR42666; software environment: operating system: UBAntu 16.04; machine learning framework: XGBoost; language and development environment: Python 3.6, AnAcondA 3.

[0054] The proposed method was evaluated in a wind farm dataset in Inner Mongolia, China, and the Tennet dataset (wind power dataset of German Tennet Power Company). In order to prove that the method has a high accuracy and the proposed new features have a high importance, multiple groups of comparative experiments and feature importance evaluation experiments were conducted. The basic information of the dataset is shown in Table 4.

[0055] Table 4 Basic information of the dataset

[0056] Dataset Number of features Number of samples Missing rate Tennet Dataset 15 200000 0％ A wind farm dataset in Inner Mongolia, China 15 12500 <1％

[0057] Embodiment 1

[0058] A dataset of a wind farm in Inner Mongolia, China, where the original data consists of two parts: NWP data (historical meteorological data) and power data of the wind farm. NWP data was collected by the wind farm at a frequency of 15 minutes from January 1 to May 22, 2019, and corrected according to historical weather forecasts. There are 13,493 entries in the NWP data. Each entry includes 9 fields: date and time occupy 2 fields, and meteorological characteristics (such as wind speed, wind direction, air density, and atmospheric pressure) occupy the other 7 fields. From January 1 to May 21, 2019, the wind farm's supervisory control and data acquisition (SCADA) collected actual power data at a frequency of 15 minutes. There are 13,493 entries in the data.

[0059] Preliminary observations indicate that the numerical weather prediction data is of high quality without any missing or outliers. This is because the data is collected from more than one source and cross-checked. The most important column in the power data provides the actual output power. In the power data, empty elements were found at 68 time points, zero elements were found at 562 time points, and abnormal elements were found at 343 time points (abnormal here means that the meteorological conditions are normal, but the output power is much lower than the normal output power). This may be caused by a SCADA failure or a transmission failure, and deletion measures are taken.

[0060] Through the above analysis, the weather forecast data was spliced ​​with the power data. After splicing, the time and date fields were converted to indexes and not used for machine learning. Other features were subjected to principal component analysis. Features that had little effect on the output power (such as altitude) were deleted, leaving 6 features (wind speed, wind direction, temperature, humidity, atmospheric pressure, and air density) out of 12,520 valid and continuous entries. Finally, 12,520 valid entries were obtained and converted into the initial dataset.

[0061] According to the time window construction method and financial factor calculation method of the present invention, the data set is reconstructed, with a total of 12,500 data, each of which has 15 features, namely K-FP, D-FP, and J-FP (FP (for power) represents a financial derivative indicator unique to wind power generation, which is used to distinguish it from K, D, and J in the financial market, that is, K in the formula x , D x , J x ), DIF-FP, DEA-FP, bar-FP (also known as DIF in the formula) calculated in Algorithm 2 x 、DEA x , ), the power 15 minutes ago, the current power, the change in the time window, the wind speed, wind direction, humidity, temperature, air pressure, and air density in the original time series. The forecast target is the output power 15 minutes after the current time point, and its unit is MW.

[0062] In order to verify the positive impact of time window and financial factors on prediction accuracy, multiple experiments were conducted. In all experiments, the training set accounted for 80% of the data set and the test set accounted for 20% of the data set. Among them, meteorological conditions refer to wind speed, wind direction, humidity, temperature, air pressure, and air density; time window refers to power 15 minutes ago, current power, and change; financial factors refer to K-FP, D-FP, J-FP, DIF-FP, DEA-FP, and bar-FP.

[0063] The input data features of the first group of experiments are meteorological conditions, with a total of 6 features; the input data features of the second group of experiments are time windows, with a total of 3 features; the input data features of the third group of experiments are financial factors, with a total of 6 features; the input data features of the fourth group of experiments are meteorological conditions and time windows, with a total of 9 features; the input data features of the fifth group of experiments are meteorological conditions and financial factors, with a total of 12 features; the input data features of the sixth group of experiments are meteorological conditions, time windows, and financial factors, with a total of 15 features; the input data features of the seventh group of experiments are time windows and financial factors, with a total of 9 features.

[0064] The experimental results are as follows Figure 4 As shown in Table 5, the F score in the feature importance score graph refers to the number of times the feature appears in the gradient boosting regression tree, that is, the number of times the feature is used as a split node in all boosting trees.

[0065] Table 5 Prediction results on a wind farm dataset in Inner Mongolia, China

[0066] Group MAE MSE RMSE AR <![CDATA[R 2 score]]> Run time 1 2.394 20.563 4.535 87.62％ 0.836 0.968s 2 1.285 9.072 3.012 93.23％ 0.926 2.173s 3 1.086 5.904 2.430 94.36％ 0.954 2.235s 4 0.959 5.716 2.390 94.89％ 0.954 2.193s 5 1.070 5.475 2.339 94.40％ 0.956 2.259s 6 0.932 5.154 2.270 95.24％ 0.959 3.325s 7 0.969 5.183 2.276 94.99％ 0.960 2.985s

[0067] In Table 5, taking Experiment 1 as the benchmark experiment, and comparing Experiment 2 with Experiment 3, the addition of the time window and financial factors reduced the MSE by 55.88% and 71.29% respectively, while the accuracy increased by 5.61% and 6.74% respectively. Figure 4 As can be seen from (a), (b), and (c), the predicted curve of Experiment 1 is poorly consistent with the actual curve, while the predicted curves of Experiments 2 and 3 are significantly more consistent with the actual curve. Therefore, the addition of financial factors and time windows significantly improves the accuracy of wind power generation prediction.

[0068] In Table 5, by comparing Experiments 4 and 5, it can be found that the prediction accuracy of the hybrid model of time window and meteorological data is slightly higher than that of the hybrid model of financial factors and meteorological data. Figure 4In (d) and (e), we can see that the predicted curves of the two models are in good agreement with the actual curves. Figure 4 From (k) and (l), we can see that the importance of the time window and financial factors in the new model is much higher than that of meteorological data.

[0069] In Table 5, Experiment 6 is a mixed model of financial factors, time windows, and meteorological data. All its evaluation indicators are the highest among the seven experiments. Figure 4 In (f), it can be seen that the model prediction curve and the actual curve are in good agreement, reflecting that the model constructed by the present invention has a high prediction accuracy. Therefore, the model of Experiment 6 is the best model. Figure 4 As can be seen from (m), the importance scores of most financial factors in this model are at the top, reflecting the importance of financial factors to this prediction model.

[0070] Combined with Table 5 Figure 4 (f), (g), (m), (n), compared with Experiments 6 and 7, it can be found that meteorological data has no significant impact on the prediction model. Meteorological data can be regarded as unnecessary auxiliary information and deleted to reduce system overhead and model training time. Combined with the evaluation indicators, the model of Experiment 7 is a cost-effective model.

[0071] Taking the first group of experiments as the benchmark experiment, the optimal model proposed by the present invention improves the accuracy by about 8.69%, reduces MAE by about 61%, MSE by about 74.9%, RMSE by about 49.9%, and R 2 The score increased by about 14.7%. It can be seen that the prediction model based on the XGBoost algorithm combined with financial factors and time windows has extremely high training accuracy and extremely short training time. It is a machine learning model that can rival or even exceed deep learning.

[0072] Embodiment 2

[0073] TenneT is a leading European electricity transmission system operator (TSO), with its main business in the Netherlands and Germany. The present invention obtains the short-term wind power data of Tennet in Germany from 2015 to 2020 through ENTSO-E, including land wind power data set and ocean wind power data set. The present invention constructs models and predicts future power based on these two data sets.

[0074] Since the land wind power generation data set and the ocean wind power generation data set have the same size and dimension, the following processing scheme of the present invention is applicable to both data sets.

[0075] ENTSO-E disclosed the wind power generation situation from 00:00 on January 1, 2015 to 23:45 on September 30, 2020. The original data contained a total of 201,600 data, excluding meteorological data and other factors, and only contained power values ​​every 15 minutes. Because the power data is the sum of the output power of all the company's wind power equipment in Germany, there are no null and outlier problems, and no special processing is required. According to the time window construction method and financial factor calculation method of the present invention, the data set was reconstructed. Finally, the data set contains 200,000 data, each of which contains 9 features: K-FP, D-FP, J-FP, DIF-FP, DEA-FP, BAr-FP, data from the previous 15 minutes, current data, and change. The prediction target is the output power 15 minutes after the current time point, and its unit is 10 2 MW.

[0076] The present invention conducts two groups of experiments on the Tennet dataset, namely, land wind power prediction experiment and ocean wind power prediction experiment. The training set accounts for 80% of the dataset, and the test set accounts for 20% of the dataset. The input data of the first group of experiments is land wind power data, including all features; the input data of the second group of experiments is ocean wind power data, including all features.

[0077] Table 6 Prediction results on the German power company Tennet dataset

[0078] Group MAE MSE RMSE AR <![CDATA[R 2 score]]> Run time 1 0.455 1.182 1.087 98.87％ 0.994 14.877s 2 0.227 0.569 0.754 98.62％ 0.992 15.365s

[0079] Through Table 6 and Figure 5 , the positive impact of time window and financial factors on prediction is further confirmed. Because the power data in the data set is the wind power generation data of Tennet in Germany, its power value is more stable, and the possibility of being affected by wind turbine failure or local environment is greatly reduced. Therefore, the model constructed by the present invention has a higher accuracy on this data set.

[0080] Finally, the workflow diagram of the present invention is as follows Figure 6 shown.

[0081] Although the present invention has been described through preferred embodiments, the present invention is not limited to the embodiments described herein but includes various changes and modifications that may be made without departing from the scope of the present invention.

Claims

1. A wind power prediction method based on XGBoost extended financial factor, characterized in that: The steps include: Step 1, forming an original time series with original data of the wind farm, wherein the original data includes temperature and pressure data monitored by the meteorological department and wind speed, wind direction, and real-time power monitored by the wind farm; Step 2, selecting the power at the current time point, the power c minutes before the current time point, and the difference between the power at the current time point and the power c minutes before the current time point in the original time series to form a time window; Step 3, select the original time series, the power at the current time point is the starting power of 1 hour, the power 60 minutes after the current time point is the ending power of 1 hour, the maximum power within 60 minutes is the maximum power within 1 hour, and the minimum power within 60 minutes is the minimum power within 1 hour; all the data in the original time series are drawn according to this step. Wind power K line; Step 4, based on the wind power K-line, calculate the financial factors MACD and KDJ that are unique to wind power; Step 5, combining the original data, the time window generated in step 2, and the financial factors MACD and KDJ calculated in step 4 into a new time series, and inputting them into the XGBoost algorithm based on the gradient boosting regression tree to predict future wind power generation.

2. The wind power prediction method based on XGBoost extended financial factor according to claim 1 is characterized in that: In the step 1, the original time series format is a two-dimensional array of time and monitoring data.

3. The wind power prediction method based on XGBoost extended financial factor according to claim 1 is characterized in that: In step 2, the method for constructing the time window is as follows: For time series data A, let the data value at time i+1 be t i+1 is the data value to be predicted, and the time series data value t at time i is taken i , i-1 time series data value t i -1, the data difference between time i and time i-1 Construct a time window to obtain new time series data B. The conversion process is as follows: ↓ Where n is the total number of moments divided in the time series data A.

4. The wind power prediction method based on XGBoost extended financial factor according to claim 1 is characterized in that: In step 3, the method for drawing the wind power generation power K line is as follows: The first step is to integrate the sampling data of every 5 adjacent sampling points into one piece of data at a frequency of once every c minutes; In the second step, the maximum value among the five adjacent sampling points is regarded as the highest price in the financial market; the minimum value among the five sampling points is regarded as the lowest price in the financial market; the power output value of the first sampling point is regarded as the opening price in the financial market; the power output value of the last sampling point is regarded as the closing price in the financial market; when the closing price is higher than the opening price, the opening price is below and the closing price is above, and the rectangular column between the two is drawn in red; when the closing price is lower than the opening price, the opening price is above and the closing price is below, and the rectangular column between the two is drawn in green.

5. The wind power prediction method based on XGBoost extended financial factor according to claim 1 is characterized in that: In step 4, KDJ is based on ultra-short-term wind power forecast, and its cycle is selected as 3 hours. The power at the end of 3 hours is C3; the maximum power within 3 hours is H3; the minimum power within 3 hours is L3; the immature random value is RSV; K is a quick confirmation indicator, and the K value of the xth hour is K x ; D is the slow trunk indicator, and the D value at the xth hour is D x ; J is the trend direction indicator, and the J value in the xth hour is J x ; The RSV value at the xth hour is RSV x ; K x-1 K value for the previous 3 hours. If it does not exist, use 50 instead. D x-1 D value for the previous 3 hours. If it does not exist, use 50 instead. The calculation method is as follows: J x =3K x -2D x MACD is based on ultra-short-term wind power forecast. Its fast moving average period is selected as 3 hours, and its slow moving average period is selected as 5 hours. Suppose the power at the end of x hours is C x ; The fast moving average of x hours is The x-hour slow moving average is The difference value for x hours is DIF x ; The smoothed moving average of x hours is DEA x ; Represents the EMA of the previous 3 hours fast If the value does not exist, 0 is used instead; Represents the EMA of the previous 5 hours sl o w If the value does not exist, use 0 instead; the bar value for x hour is The calculation method is as follows:

6. The wind power prediction method based on XGBoost extended financial factor according to claim 4 is characterized in that: The K and D values ​​fluctuate between 0 and 100.

7. The wind power prediction method based on XGBoost extended financial factor according to claim 1 is characterized in that: The XGBoost algorithm is composed of multiple CART trees, and the ensemble learning of multiple CART trees is achieved through the gradient boosting regression method. In the training stage, each decision tree learns the residual between the target value and the sum of the prediction values ​​of all previous trees. Each newly added tree is trained on the basis of the trained tree. At the same time, when selecting the split point, the optimal split point will be selected according to the optimal features. Finally, multiple decision trees make a decision together, and the results of all trees are accumulated as the final prediction result.

Images