A long-term wind power cluster power forecasting method considering the evolution of significant weather information

Abstract

The present invention belongs to the technical field of wind power generation prediction. Taking the poor availability of existing long-forecast numerical weather forecasts as the starting point, a long-forecast wind power cluster power prediction method taking into account the evolution law of significant wind processes and power constraints is proposed. The significant wind process identification method considering autocorrelation analysis predicts future power trends; explores an 8-15-day power forecasting method based on variational mode decomposition and multivariate linear regression model; establishes a power-power-trend relationship, and uses a historical similar trend process matching method to achieve power reconstruction to complete the prediction. The method proposed in the present invention is a prediction method that takes into account power and power trends and is suitable for long-forecast periods. This model is simple to calculate, has high prediction performance, clear physical meaning, effective prediction results, and strong practicality.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation prediction, and more specifically, to a long-term wind power cluster power prediction method taking into account the evolution law of significant weather information. Background Art

[0002] Wind power is one of the rapidly developing new energy sources nowadays, and its power generation accounts for an increasingly higher proportion in today's energy structure. At the same time, as new energy gradually becomes the main power source of the new power system, large-scale wind power grid connection brings huge challenges to the safe operation of the power system. The current wind power forecast has a maximum forecast period of 7 days, and research on power forecasts for 8-15 days is still blank. Breaking through the long-forecast power forecast is the key support to ensure the safe and stable operation of the power system, and it is also an important reference for the long-term operation planning of the electrical system.

[0003] The long-term forecast of wind power refers to the forecast for the next 8-15 days from the forecast time, with a time resolution of 15 minutes. The significance of the long-term forecast of wind power is to provide a reference for large-scale maintenance plans of wind farms and long-term energy planning of thermal power plants.

[0004] Existing wind power forecasting methods generally establish a mapping relationship between historical input data and future power output, and can directly predict future power values ​​based on historical data. Such a forecasting framework performs well in ultra-short-term and short-term forecasts, but in the long-term forecast period, due to the long forecast time scale, this mapping relationship cannot be effectively established. Summary of the invention

[0005] In order to solve the above-mentioned problems existing in the prior art, the present invention provides a long-term wind power cluster power prediction method taking into account the evolution law of significant weather information, and provides a long-term wind power cluster power prediction scheme taking into account the evolution law of significant weather information, which has clear physical meaning, is scientific, reasonable and effective, has reliable and stable predictions, and can utilize long-time scale trend information.

[0006] The present invention provides a method for predicting wind power cluster power in a long-term forecast period taking into account the evolution law of significant weather information, the method comprising:

[0007] Autocorrelation analysis and weather evolution law characterization are performed on the power series and wind speed series respectively to obtain the power trend series and wind speed trend series;

[0008] Based on the power trend sequence and the wind speed trend sequence, the power trend and the amount of electricity are predicted to obtain a predicted power trend and a predicted amount of electricity;

[0009] Based on the matching of historical similar trend processes under power constraints, the power prediction value is obtained.

[0010] Preferably, autocorrelation analysis and weather evolution law characterization are performed on the power sequence and the wind speed sequence respectively to obtain a power trend sequence and a wind speed trend sequence, including:

[0011] The autocorrelation coefficient of the power series or wind speed series is calculated by the following formula:

[0012]

[0013] In the formula, ρ ik is the autocorrelation coefficient of the power series or wind speed series of the lag k time points on the i-th day, i is the number of days, k is the number of lag time points, N is the number of sequence time points, t is the sequence time point, x it is the wind power value or wind speed value at time t on the i-th day, is the mean of the power series or wind speed series on the i-th day, x it-k is the wind power value or wind speed value at k time points lagging behind time t on the i-th day;

[0014] Based on the calculated autocorrelation coefficient of the power sequence or wind speed sequence, the daily average autocorrelation coefficient of the power sequence or wind speed sequence is calculated by the following formula;

[0015]

[0016] In the formula, is the daily average autocorrelation coefficient of the power series or wind speed series, and n is the total number of days;

[0017] Determine the window length for dividing the power sequence or the wind speed sequence according to the daily average autocorrelation coefficient of the power sequence or the wind speed sequence;

[0018] Based on the window length of the power sequence or the wind speed sequence, the weather evolution law of the power sequence and the wind speed sequence is characterized to obtain a power trend sequence and a wind speed trend sequence.

[0019] Preferably, based on the window length of the power sequence or the wind speed sequence, the power sequence and the wind speed sequence are characterized by the weather evolution law to obtain the power trend sequence and the wind speed trend sequence, including:

[0020] Perform linear fitting on the power series and the wind speed series respectively to obtain the first line segment and the second line segment;

[0021] Taking the slopes of the first line segment and the second line segment as the first elements of the first weather process quantitative index and the second weather process quantitative index respectively;

[0022] Determining a first time window and a second time window based on the window lengths of the power sequence and the wind speed sequence;

[0023] The last value in each first time window minus the first value is used as the second element of the first weather process quantitative index, and the last value in each second time window minus the first value is used as the second element of the second weather process quantitative index;

[0024] The variation trend of the power sequence is determined according to the second element and the first element of the quantitative index of the first weather process, and the variation trend of the wind speed sequence is determined according to the second element and the first element of the quantitative index of the second weather process;

[0025] The changing trends of the power series and wind speed series are quantified and represented respectively to form a power trend series and a wind speed trend series.

[0026] Preferably, based on the power trend sequence and the wind speed trend sequence, the power trend and the amount of electricity are predicted to obtain the predicted power trend and the predicted amount of electricity, including:

[0027] Based on the multivariate linear regression model, the wind speed trend sequence is taken as input and the power trend sequence is taken as output, a mapping relationship is established to obtain the power trend sequence at the future moment;

[0028] The line segments obtained based on the power sequence are integrated, with days as the time window and one sampling point as the time step, to obtain the power sequence, and the power sequence is decomposed using variational mode decomposition to obtain several intrinsic mode components and one residual mode component, and the several intrinsic mode components and one residual mode component are predicted respectively through the power prediction model, and the prediction results are added to obtain the predicted power.

[0029] Preferably, the historical similar trend process matching under the power constraint includes at least one of trend matching, first power matching, second power matching, first combination matching and second combination matching.

[0030] Preferably, the trend matching includes: globally matching the predicted power trend with the power trend sequence, selecting the power corresponding to the best historical similar trend segment and splicing them to form a power prediction value;

[0031] Preferably, the first power matching includes: matching the predicted power with the first historical power set, and selecting the power corresponding to the most similar power as the power prediction value; the first historical power set is obtained by integrating the line segments obtained based on the power sequence, taking the day as the time window, and calculating using the first sliding step size;

[0032] The second power matching includes: matching the predicted power with the second historical power set, and selecting the power corresponding to the most similar power as the power prediction value; the second historical power set is obtained by integrating the line segments obtained based on the power sequence, taking days as the time window, and calculating using a second sliding step size.

[0033] Preferably, the first combination matching includes:

[0034] With the predicted power as the first constraint and the predicted power trend as the second constraint, the distance between each predicted power and the historical power set is calculated based on the following formula (3):

[0035]

[0036] Where D q Indicates the distance between the predicted power and the historical power set, j is the sequence length, n is the total length of the sequence, data pre,j is the predicted value, data his,j is the historical value;

[0037] Take the power values ​​corresponding to multiple historical power values ​​with the smallest distance as the quasi-power prediction value;

[0038] The quasi-power prediction value is converted into a trend feature format according to the trend division and quantification method, and the distance sum of each element with the predicted trend feature is calculated. Then, the trend feature distances of the same day are added together to form the daily trend feature distance. The quasi-prediction power value with the smallest daily trend feature distance is taken as the final prediction power.

[0039] Preferably, the second combination matching includes:

[0040] With the predicted power trend as the first constraint and the predicted electricity as the second constraint, the predicted power trend is globally searched and matched against all historical sub-sequence power trend sets. Based on the set error, multiple historical similar values ​​are selected in each sub-sequence trend, and all historical similar values ​​are spliced ​​into multiple complete day power series trends in time series. The electricity of each complete day power is calculated, and the day power with the smallest difference from the predicted electricity is taken as the predicted power.

[0041] Preferably, after obtaining the power prediction value by matching the historical similar trend process under the power constraint, the method further includes:

[0042] Based on the measured total power data of the wind power cluster and the average NWP wind speed data of the corresponding wind power cluster predicted 8-15 days in advance, the overall prediction results of the wind power cluster power for 8-15 days are obtained;

[0043] The evaluation index is used to perform error analysis on the overall prediction results of wind power cluster power for 8-15 days; wherein, the evaluation index includes selecting the root mean square error and the mean absolute error, and the calculation formulas are:

[0044]

[0045] In the formula, p k and p k,preThey represent the actual value and predicted value of power at time k respectively, Cap is the total installed capacity of the wind farm, N is the number of samples in the prediction section, RMSE is the root mean square error, and MAE is the mean absolute error.

[0046] The present invention has at least the following beneficial effects:

[0047] The long-term wind power cluster power prediction method proposed in the present invention, which takes into account the evolution law of significant weather information, fully considers the availability of long-term forecast NWP, designs a special trend extraction and utilization method for the situation where its numerical value is inaccurate but the trend is available, and provides an available long-term wind power cluster power prediction method, while improving the effectiveness and reliability of its prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flow chart of a method for predicting wind power cluster power in the long term taking into account the evolution law of significant weather information according to an embodiment of the present invention is shown.

[0049] Figure 2 A schematic diagram of the power average autocorrelation coefficient according to an embodiment of the present invention is shown.

[0050] Figure 3 The schematic diagram of the quantitative representation of the weather process according to the embodiment of the present invention is shown. In the figure, the horizontal axis represents the time step, the vertical axis represents the power value, a represents the slope after the linear fitting of the power value of each time window, and b represents the intercept after the linear fitting. The change trend of the weather process can be expressed as a tuple (q 1 ,c 1 ) indicates that q 1 The first element of the quantitative index of the weather process, c 1 The second element of the quantitative index of the weather process, y end Represents the last element of the original data of the weather process, y start Represents the first element of the raw data of the weather process.

[0051] Figure 4 The flowchart of the decomposition and reconstruction power prediction method according to an embodiment of the present invention is shown. In the figure, IMF 1 、IMF 2 、IMF 3 ...IMF n represents the different modal components of the original power sequence after VMD decomposition, (x 1 ,x 2 ,x 3 ,...,x t ) represents the electric quantity characteristic between the first moment and the tth moment, (x 2 ,x 3 ,x4 ,...,x t+1 ) represents the electric quantity characteristic between the 2nd moment and the t+1th moment, (x 3 ,x 4 ,x 5 ,...,x t+2 ) represents the electric quantity characteristic between the 3rd moment and the t+2th moment, (x n-t+1 ,x n-t+2 ,x n-t+3 ,...,x n ) represents the power characteristics from the n-t+1th moment to the nth moment, Y t+1 represents the predicted power value at time t+1, Y t+2 represents the predicted power value at time t+2, Y t+3 represents the predicted power value at time t+3, Y n+1 Indicates the predicted power value at the n+1th moment.

[0052] Figure 5 A schematic diagram of the relationship between power consumption and trend according to an embodiment of the present invention is shown. In the figure, T 1 Indicates the first trend segment with a time window width of 8, T 12 Indicates the 12th trend segment, T 13 Indicates the 13th trend segment, T 24 Indicates the 24th trend segment, Q 1 Indicates the first element in the power set, with a time span of 96 sampling points, Q 2 Represents the second element in the electric quantity set, Q 3 Represents the third element in the electric quantity set, Q 13 Represents the 13th element in the power collection.

[0053] Figure 6 A schematic diagram of power prediction results on the 8th day according to different methods according to an embodiment of the present invention is shown.

[0054] Figure 7 A schematic diagram of power prediction results on the 15th day according to different methods of an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0055] 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.

[0056] Based on the previous research status, the embodiment of the present invention proposes a long-term wind power cluster power prediction method taking into account the evolution law of significant weather information. Figure 1 , respectively, are the flow charts of this method. Figure 1 As shown, the long-term wind power cluster power prediction method taking into account the evolution law of significant weather information includes the following steps S1 to S3, which are described in detail as follows.

[0057] S1. Perform autocorrelation analysis and weather evolution law characterization on the power series and wind speed series respectively to obtain the power trend series and wind speed trend series.

[0058] In this embodiment, the purpose of performing autocorrelation analysis and weather evolution law characterization on the power sequence and wind speed sequence is to divide a reasonable time window; then based on the divided time windows, trend features of the wind speed sequence and the power sequence are extracted to obtain subsequences, and after quantitative characterization of the subsequences, a new ordered set is formed according to the time series, that is, a power trend sequence and a wind speed trend sequence are obtained.

[0059] It should be noted that the power sequence is specifically a wind power sequence, and the wind speed sequence is specifically a NWP wind speed sequence; NWP stands for numerical weather forecast.

[0060] In some embodiments, the wind power sequence and the NWP wind speed sequence are time series, and their fluctuations have a certain continuity. All wind power sequences usually have a certain autocorrelation. Therefore, this embodiment performs autocorrelation analysis on the wind power sequence and the NWP wind speed sequence, thereby giving a reasonable window length to divide the power sequence and the numerical weather forecast (Numerical Weather Prediction, NWP) wind speed, so as to form a database for subsequent historical matching. The autocorrelation coefficient analysis expression is as shown in formula (1):

[0061]

[0062] In the formula, ρ ikis the autocorrelation coefficient of the power series or wind speed series of the i-th day lagged by k time points, i is the number of days, k is the number of lagged time points, N is the number of sequence time points, t is the sequence time point, x it is the wind power value or wind speed value at time t on the i-th day, x i is the mean of the power series or wind speed series on the i-th day, x it-k is the wind power value or wind speed value at k time points lagging behind time t on the i-th day.

[0063] If the autocorrelation coefficient is greater than 0.4, it is considered to have a certain correlation. The daily average autocorrelation coefficient is analyzed to obtain the reasonable window length for sequence division. The calculation expression of the daily average autocorrelation coefficient is shown in formula (2):

[0064]

[0065] In the formula, is the daily average autocorrelation coefficient of the power series or wind speed series, and n is the total number of days.

[0066] In this embodiment, the window length for dividing the power sequence or the wind speed sequence is determined according to the daily average autocorrelation coefficient of the power sequence or the wind speed sequence; and based on the window length of the power sequence or the wind speed sequence, the weather evolution law of the power sequence and the wind speed sequence is characterized to obtain the power trend sequence and the wind speed trend sequence.

[0067] In some embodiments, the change trend of the subsequence is represented by two elements, so as to facilitate the prediction of the future change trend of the sequence. The change trend of the subsequence can be represented by a tuple, and the specific quantitative characterization steps are as follows: linear fitting is performed on the power and NWP wind speed respectively, and the slope of the line segment is taken as the first element of the quantitative index of the weather process; the last value of each time window minus the first value is the second element of the weather process. The trend features of the NWP wind speed and power sequences are extracted, and the subsequences are quantitatively characterized to form a new ordered set according to the time sequence.

[0068] Specifically, based on the window length of the power sequence or wind speed sequence, the weather evolution law of the power sequence and the wind speed sequence is characterized to obtain the power trend sequence and the wind speed trend sequence, including the following steps:

[0069] S11. Perform linear fitting on the power sequence and the wind speed sequence respectively to obtain a first line segment and a second line segment.

[0070] S12. Take the slopes of the first line segment and the second line segment as the first elements of the first weather process quantitative index and the second weather process quantitative index respectively.

[0071] S13. Determine a first time window and a second time window based on the window lengths of the power sequence and the wind speed sequence.

[0072] S14. The last value in each first time window minus the first value is used as the second element of the first weather process quantitative index, and the last value in each second time window minus the first value is used as the second element of the second weather process quantitative index.

[0073] S15. Determine the change trend of the power sequence according to the second element and the first element of the first weather process quantitative index, and determine the change trend of the wind speed sequence according to the second element and the first element of the second weather process quantitative index.

[0074] S16. Quantify and characterize the changing trends of the power sequence and the wind speed sequence respectively to form a power trend sequence and a wind speed trend sequence.

[0075] S2. Based on the power trend sequence and the wind speed trend sequence, the power trend and the amount of electricity are predicted to obtain a predicted power trend and a predicted amount of electricity.

[0076] In this embodiment, after step S1, the power sequence and the NWP wind speed sequence are converted into corresponding trend sequences of the two. The trend sequence only indicates the trend direction of the original data, and does not characterize the numerical value. A simple multivariate linear regression model is used, with the NWP wind speed trend sequence as input and the power trend sequence as output, to establish a mapping relationship and obtain the power trend sequence at the future moment. The electric quantity is calculated by an integral calculation method to integrate the line segments drawn in the power sequence. The day is used as the time window and a sampling point is used as the time step to calculate the electric quantity sequence. The variational mode decomposition (VMD) is used to decompose the electric quantity sequence to obtain several intrinsic mode components and one residual mode component. After each mode component is predicted by the electric quantity prediction model, the prediction results are added to obtain the final electric quantity prediction result.

[0077] Specifically, step S2 includes the following steps:

[0078] S21. Based on the multivariate linear regression model, with the wind speed trend sequence as input and the power trend sequence as output, a mapping relationship is established to obtain the power trend sequence at the future moment.

[0079] In this embodiment, the multivariate linear regression model can be obtained by training with a first training data set, wherein the first training data set includes a historical wind speed trend sequence and an actual power trend sequence, and is obtained by training with a supervised training method.

[0080] S22. Integrate the line segments obtained based on the power sequence, take days as the time window and one sampling point as the time step, and obtain the power sequence. Use variational mode decomposition to decompose the power sequence to obtain several intrinsic mode components and one residual mode component. After predicting the several intrinsic mode components and one residual mode component respectively through the power prediction model, add the prediction results to obtain the predicted power.

[0081] In this embodiment, the power prediction model can be a multivariate linear regression model, which can be obtained by training a blank multivariate linear regression model based on the second training data set. For example, the second training data set can include the input features and actual output of the power prediction model, and a supervised training method is used to obtain the corresponding weight matrix, and the blank multivariate linear regression model is configured using the weight matrix to obtain the power prediction model, where the input features can be obtained by decomposing the historical power sequence through variational mode decomposition, and the actual output is the actual collected power value.

[0082] S3. Based on the historical similar trend process matching under the power constraint, the power prediction value is obtained.

[0083] In this embodiment, the method of matching historical similar trend processes under power constraints can be one or more. Specifically, it can be at least one of trend matching, first power matching, second power matching, first combination matching and second combination matching. Each matching method has different prediction effects.

[0084] Among them, the trend matching includes: globally matching the predicted power trend with the power trend sequence, selecting the power corresponding to the best historical similar trend segment and splicing it to form a power prediction value;

[0085] The first power matching includes: matching the predicted power with the first historical power set, and selecting the power corresponding to the most similar power as the power prediction value; the first historical power set is obtained by integrating the line segments obtained based on the power sequence, taking the day as the time window, and calculating using the first sliding step size;

[0086] The second power matching includes: matching the predicted power with the second historical power set, and selecting the power corresponding to the most similar power as the power prediction value; the second historical power set is obtained by integrating the line segments obtained based on the power sequence, taking days as the time window, and using a second sliding step size for calculation.

[0087] The first combination matching includes: taking the predicted power as the first constraint and the predicted power trend as the second constraint, and calculating the distance between each predicted power and the historical power set based on the following formula (3):

[0088]

[0089] Where D q Indicates the distance between the predicted power and the historical power set, j is the sequence length, n is the total length of the sequence, data pre,j is the predicted value, data his,j is the historical value;

[0090] The power values ​​corresponding to multiple historical electricity quantities with the smallest distance are taken as quasi-power prediction values; the quasi-power prediction values ​​are converted into trend feature formats according to trend division and quantification methods, and the distance sum of each element is calculated with the predicted trend feature, and then the trend feature distances of the same day are added together to form the daily trend feature distance, and the quasi-prediction power value with the smallest daily trend feature distance is taken as the final predicted power.

[0091] The second combination matching includes: taking the predicted power trend as the first constraint and the predicted electricity as the second constraint, globally searching and matching the predicted power trend with all historical sub-sequence power trend sets, selecting multiple historical similar values ​​in each sub-sequence trend based on the set error, splicing all historical similar values ​​into multiple complete day power sequence trends in time series, and calculating the electricity of each complete day power, taking the day power with the smallest difference from the predicted electricity as the predicted power.

[0092] In some embodiments, after obtaining the power prediction value based on the historical similar trend process matching under the power constraint, the method further includes:

[0093] Based on the measured total power data of the wind power cluster and the average NWP wind speed data of the corresponding wind power cluster predicted 8-15 days in advance, the overall prediction results of the wind power cluster power for 8-15 days are obtained;

[0094] The evaluation indexes are used to conduct error analysis on the overall prediction results of wind power cluster power for 8-15 days; the evaluation indexes include the root mean square error and mean absolute error, and the calculation formulas are:

[0095]

[0096] In the formula, p k and p k,pre They represent the actual value and predicted value of power at time k respectively, Cap is the total installed capacity of the wind farm, N is the number of samples in the prediction section, RMSE is the root mean square error, and MAE is the mean absolute error.

[0097] The following embodiments of the present invention will describe the method of the present invention in detail with reference to specific data.

[0098] In order to maximize the accuracy of wind power prediction, this embodiment mainly focuses on the long-term wind power cluster power prediction, and the proposed method requires a large amount of historical data to support it. This embodiment uses a cluster wind farm for actual verification, with a total installed capacity of 1790MW, a data length from January 1, 2021 to December 31, 2022, a power and NWP sampling interval of 15 minutes, a total of 96 sampling points per day, and a training set, validation set, and test set ratio of 10:1:1. Based on the above data, the method includes the following steps:

[0099] Step 1: The power is divided by day, and the lagged autocorrelation calculation is performed on the power sequence of each day from 0 to 50 time points. 51 autocorrelation coefficients are obtained every day. The autocorrelation coefficients of the same lagged time points are averaged to obtain the daily average autocorrelation coefficients of different lagged time points. The autocorrelation coefficients greater than 0.4 are considered to have a certain correlation. The daily average autocorrelation coefficients are analyzed to obtain the reasonable window length for sequence division. The daily average autocorrelation coefficient of power is as follows: Figure 2 As shown, in order to process the power and NWP wind speed in units of days, the window length of the sequence division is set to 8, which satisfies the autocorrelation analysis results and facilitates the subsequent combined analysis with the power.

[0100] Step 2: Use two elements to represent the changing trend of the subsequence, so as to facilitate the prediction of future sequence changing trends. The changing trend of the subsequence can be represented by a tuple. The specific quantitative representation method is: use 8 as the time window length and 8 as the step size for sliding processing, perform linear fitting on power and NWP wind speed respectively, and take the slope of the line segment as the first element of the quantitative index of the weather process; the last value of each time window minus the first value is the second element of the weather process. Trend features are extracted from the NWP wind speed and power sequences, and after the subsequence is quantitatively represented, a new ordered set is formed according to the time series. The quantitative representation of the subsequence is expressed as follows: Figure 3 shown.

[0101] Step 3: After steps 1 and 2, the autocorrelation analysis of the historical actual cluster power is performed to obtain a trend segment division window of 8 time points. Then the NWP wind speed and cluster power are divided into several trend segments, and each trend segment is quantitatively characterized to form a corresponding set of trend feature elements. Finally, a mapping relationship is established with the NWP wind speed trend feature as input and the power trend feature as output to predict the future power trend feature.

[0102] The electricity quantity is calculated by the integral method, and the line segments drawn in the power sequence are integrated. The day is used as the time window and one sampling point is used as the time step to calculate the electricity quantity sequence. Two different electricity quantity sets are proposed, and the only difference between electricity quantity sets 1 and 2 is the sliding step size. The time window width of both is 96, the sliding step size of electricity quantity set 1 is 96, and the sliding step size of electricity quantity set 2 is 8. Electricity quantity set 1 is used for electricity quantity forecasting, and electricity quantity set 2 is only used for historical matching. The electricity quantity sequence is decomposed into several intrinsic mode components and one residual mode component using variational mode decomposition. After each mode component is predicted by a simple multivariate linear regression model, the prediction results are added together to obtain the final electricity quantity forecast result. The process of the decomposition and reconstruction electricity prediction method is as follows: Figure 4 shown.

[0103] Step 4: The method proposed in the present invention involves the mutual conversion of electric quantity, power and trend, so the correlation between electric quantity, power and trend is given as follows: Figure 5 This leads to the matching of historical similar trend processes under power constraints, and then the power prediction results are obtained. In the power reconstruction process, this paper also gives a variety of search and matching methods. Different power reconstruction methods have different effects. The specific descriptions of each method are as follows:

[0104] 1) Trend matching: The predicted power trend is globally matched with the historical power trend set, and the power corresponding to the best historical similar trend segment is selected to form the predicted value;

[0105] 2) Power matching 1: Match the predicted power with the historical power set 1, and select the power corresponding to the most similar power as the power prediction value;

[0106] 3) Power matching 2: The set selected for matching is historical power collection 2;

[0107] 4) Combination matching 1: Double constraint matching, with predicted power as the first constraint and predicted trend as the second constraint. First, the distance between each predicted power and the historical power set is calculated according to the given formula (3), and the power values ​​corresponding to the multiple historical power with the smallest distance are taken as the quasi-power prediction value. Secondly, the quasi-power prediction value is converted into a trend feature format according to the trend division and quantification method, and the distance sum of each element is calculated with the predicted trend feature, and then the trend feature distances of the same day are added to form the daily trend feature distance. Finally, the quasi-predicted power value with the smallest daily trend feature distance is the final predicted power.

[0108] 5) Combination matching 2: Double constraint matching, with the predicted trend as the first constraint and the predicted power as the second constraint. From the perspective of data value distribution and global trend, the predicted subsequence power trend is globally searched and matched against the entire historical subsequence power trend set. Multiple historical similar values ​​with small errors are selected for each subsequence trend, and then the multiple historical values ​​selected from all subsequence trends are spliced ​​into multiple complete day power sequence trends in time sequence. The power of each complete day is calculated, and the power of the day with the smallest difference from the predicted power is taken as the predicted value.

[0109] Step 5: Analyze the measured data and NWP wind speed data of a cluster. The total installed capacity of the power plant is 1790MW. The data sampling interval is 15 minutes. According to steps 1 to 4, the 8-15-day prediction results of the entire wind power are obtained under different methods. The results are shown in Table 1.

[0110] Table 1 Power prediction results of different methods

[0111]

[0112]

[0113] It can be seen that the modeling prediction of power mapping using NWP alone is no longer reliable in the power forecast of 8-15 days. The power accuracy of power matching 2 is higher than that of power matching 1 in most forecast scales of 8-15 days, and the average forecast error of 8-15 days is 0.24% lower, indicating that the accuracy of power matching is affected by the size and quality of the power data set. Using a reasonable method to expand the historical power data set is conducive to improving the power prediction accuracy of the power matching method. Combination matching 1 is the best method in most forecast scales of 8-15 days, and the average forecast error of 8-15 days is 1.09% higher than the second place, indicating that the most accurate forecast power is obtained by trend selection under the constraint of power value, and the selection of power or trend alone is not the best. The average forecast error of combination matching 2 for 8-15 days is 9.52% higher than that of combination matching 1, indicating that the method with priority of power value constraint is better than the method with priority of trend constraint. Although the power forecast error law of 8-15 days generally increases with the lengthening of the forecast scale, it no longer has a strict error growth.

[0114] Visualize the power prediction results, such as Figure 6 and Figure 7. It can be seen that no matter which power reconstruction method is used in the present invention, the prediction curve on the 8th day is obviously better than the prediction curve on the 15th day, and there is no significant reverse peak in the prediction curve on the 15th day. The power restored by matching only with the trend has a good effect on the trend prediction of some actual values, but the numerical value cannot be close to the actual value at all, which has a huge defect. However, due to the time series characteristics of the power itself and the constraints of the quality of the historical data set, whether it is the power prediction curve of the 8th day or the 15th day, reverse peaks will still appear at some sample points. From the overall trend point of view, the combined matching method 1 under the dual constraint matching has the best effect of fitting the actual power curve. The present invention proposes a dual constraint historical matching method with power constraint priority, which can effectively reduce the impact of historical power trend caused by the use of power constraint alone, and can also overcome the "step" phenomenon caused by the priority use of trend constraint. On the prediction scale of 8-15 days, the average RMSE and MAE are reduced by 4.29% and 3.3% compared with the traditional prediction method, which improves the reliability of the prediction.

[0115] Under the double-constraint historical matching method (combination matching 1) with power constraint priority, the prediction effect of using different models is explored.

[0116] Table 2 Comparison of prediction errors of different models for combination matching 1

[0117]

[0118]

[0119] The prediction error results of each model of combination matching 1 are shown in Table 2. Among them, MLR has the lowest prediction error, and the average RMSE and MAE for 8-15 days are 1.36% and 1.01% lower than the second place Xgboost. Deeper neural network models such as CNN and LSTM have many hyperparameters, diverse and complex structural layers, and are sensitive to parameter adjustments, which makes this type of neural network very unstable for prediction and also depends on the amount of training data. However, there are not many training data sets for electricity, and the regularity of each mode after decomposition is strong and simple. Using a complex model may have a counterproductive effect, making this type of model have the largest prediction error in the dual-constraint historical matching method with electricity constraint priority. Relatively simple models such as ELM and Xgboost are less sensitive to parameter adjustment, have low training data requirements, and have better overall prediction effects than the former.

[0120] In summary, the long-term wind power cluster power prediction method proposed in the present invention, which takes into account the evolution law of significant weather information, fully considers the availability of long-term forecast NWP, designs a special trend extraction and utilization method for the situation where its numerical value is inaccurate but the trend is available, and provides an available long-term forecast method for wind power cluster power, while improving the effectiveness and reliability of its prediction.

[0121] The specific embodiments of the present invention have been described in detail, but are not limited to these embodiments. Any obvious changes made by those skilled in the art based on the teachings of the present invention are within the scope of protection of the present invention.

Claims

1. A method for predicting wind power cluster power in the long term taking into account the evolution law of significant weather information, characterized in that: The method comprises: Autocorrelation analysis and weather evolution law characterization are performed on the power series and wind speed series respectively to obtain the power trend series and wind speed trend series; Based on the power trend sequence and the wind speed trend sequence, the power trend and the amount of electricity are predicted to obtain a predicted power trend and a predicted amount of electricity; Based on the matching of historical similar trend processes under power constraints, the power forecast value is obtained; Autocorrelation analysis and weather evolution law characterization are performed on the power series and wind speed series respectively to obtain the power trend series and wind speed trend series, including: The autocorrelation coefficient of the power series or wind speed series is calculated by the following formula: In the formula, ρ ik is the autocorrelation coefficient of the power series or wind speed series of the lag k time points on the i-th day, i is the number of days, k is the number of lag time points, N is the number of sequence time points, t is the sequence time point, x it is the wind power value or wind speed value at time t on the i-th day, is the mean of the power series or wind speed series on the i-th day, x it-k is the wind power value or wind speed value at k time points lagging behind time t on the i-th day; Based on the calculated autocorrelation coefficient of the power sequence or wind speed sequence, the daily average autocorrelation coefficient of the power sequence or wind speed sequence is calculated by the following formula; In the formula, is the daily average autocorrelation coefficient of the power series or wind speed series, and n is the total number of days; Determine the window length for dividing the power sequence or the wind speed sequence according to the daily average autocorrelation coefficient of the power sequence or the wind speed sequence; Based on the window length of the power sequence or the wind speed sequence, characterize the weather evolution law of the power sequence and the wind speed sequence to obtain a power trend sequence and a wind speed trend sequence; Based on the window length of the power sequence or the wind speed sequence, the weather evolution law of the power sequence and the wind speed sequence is characterized to obtain a power trend sequence and a wind speed trend sequence, including: Perform linear fitting on the power series and the wind speed series respectively to obtain the first line segment and the second line segment; Taking the slopes of the first line segment and the second line segment as the first elements of the first weather process quantitative index and the second weather process quantitative index respectively; Determining a first time window and a second time window based on the window lengths of the power sequence and the wind speed sequence; The last value in each first time window minus the first value is used as the second element of the first weather process quantitative index, and the last value in each second time window minus the first value is used as the second element of the second weather process quantitative index; The variation trend of the power sequence is determined according to the second element and the first element of the quantitative index of the first weather process, and the variation trend of the wind speed sequence is determined according to the second element and the first element of the quantitative index of the second weather process; The changing trends of the power series and wind speed series are quantified and represented respectively to form a power trend series and a wind speed trend series.

2. The method according to claim 1, characterized in that Based on the power trend sequence and the wind speed trend sequence, the power trend and the amount of electricity are predicted to obtain the predicted power trend and the predicted amount of electricity, including: Based on the multivariate linear regression model, the wind speed trend sequence is taken as input and the power trend sequence is taken as output, a mapping relationship is established to obtain the power trend sequence at the future moment; The line segments obtained based on the power sequence are integrated, with days as the time window and one sampling point as the time step, to obtain the power sequence, and the power sequence is decomposed using variational mode decomposition to obtain several intrinsic mode components and one residual mode component, and the several intrinsic mode components and one residual mode component are predicted respectively through the power prediction model, and the prediction results are added to obtain the predicted power.

3. The method according to claim 1, characterized in that The historical similar trend process matching under the power constraint includes at least one of trend matching, first power matching, second power matching, first combination matching and second combination matching.

4. The method according to claim 3, characterized in that The trend matching includes: globally matching the predicted power trend with the power trend sequence, selecting the power corresponding to the best historical similar trend segment and splicing them to form a power prediction value.

5. The method according to claim 3, characterized in that: The first power matching includes: matching the predicted power with the first historical power set, and selecting the power corresponding to the most similar power as the power prediction value; the first historical power set is obtained by integrating the line segments obtained based on the power sequence, taking the day as the time window, and calculating using the first sliding step size; The second power matching includes: matching the predicted power with the second historical power set, and selecting the power corresponding to the most similar power as the power prediction value; the second historical power set is obtained by integrating the line segments obtained based on the power sequence, taking days as the time window, and calculating using a second sliding step size.

6. The method according to claim 3, characterized in that The first combination matching includes: With the predicted power as the first constraint and the predicted power trend as the second constraint, the distance between each predicted power and the historical power set is calculated based on the following formula (3): Where D q Indicates the distance between the predicted power and the historical power set, j is the sequence length, n is the total length of the sequence, data pre,j is the predicted value, data his,j is the historical value; Take the power values ​​corresponding to multiple historical power values ​​with the smallest distance as the quasi-power prediction value; The quasi-power prediction value is converted into a trend feature format according to the trend division and quantification method, and the distance sum of each element with the predicted trend feature is calculated. Then, the trend feature distances of the same day are added together to form the daily trend feature distance. The quasi-prediction power value with the smallest daily trend feature distance is taken as the final prediction power.

7. The method according to claim 3, characterized in that The second combination matching includes: With the predicted power trend as the first constraint and the predicted electricity as the second constraint, the predicted power trend is globally searched and matched against all historical sub-sequence power trend sets. Based on the set error, multiple historical similar values ​​are selected in each sub-sequence trend, and all historical similar values ​​are spliced ​​into multiple complete day power sequence trends in time series. The electricity of each complete day power is calculated, and the day power with the smallest difference from the predicted electricity is taken as the predicted power.

8. The method according to claim 1, characterized in that After obtaining the power prediction value by matching the historical similar trend process under the power constraint, the method further includes: Based on the measured total power data of the wind power cluster and the average NWP wind speed data of the corresponding wind power cluster predicted 8-15 days in advance, the overall prediction results of the wind power cluster power for 8-15 days are obtained; The evaluation index is used to perform error analysis on the overall prediction results of wind power cluster power for 8-15 days; wherein, the evaluation index includes selecting the root mean square error and the mean absolute error, and the calculation formulas are: In the formula, p k and p k,pre They represent the actual value and predicted value of power at time k respectively, Cap is the total installed capacity of the wind farm, N is the number of samples in the prediction section, RMSE is the root mean square error, and MAE is the mean absolute error.

Images