A short-term wind power segmented prediction method based on the recognition of turning periods

Abstract

The present invention relates to an ultra-short-term wind power segmented prediction method based on the identification of turning periods, which uses the moving average method to extract the time series trend; adopts the Gaussian window method to smooth the exponential moving average (EMA), and calculates the time series change rate α at each moment; adaptively adjusts the time window width based on the window adjustment strategy of local time series characteristics; based on the inflection point detection strategy of the double fixed-time sliding window, introduces α as one of the criteria, extracts and divides the turning weather mutation periods; uses the improved GRU algorithm for point prediction of the turning section time series, and combines the improved Attention mechanism of the CRS algorithm; uses probability prediction for the flat section time series, establishes a time series pattern-power prediction error probability density distribution model by the empirical distribution estimation method, and conducts wind power probability prediction based on the variable bandwidth kernel density estimation method; combines the point prediction and the probability prediction time series segmented prediction to obtain the final prediction result. Compared with the prior art, the present invention has the advantages of improving the model operation efficiency, etc.

Description

Technical Field

[0001] The present invention relates to the technical field of ultra-short-term wind power prediction for high-concentration wind farms, and particularly to an ultra-short-term wind power segmented prediction method based on the identification of turning periods. Background Art

[0002] In recent years, with the large-scale and high-concentration development of wind power, the proportion of wind power connected to the power grid has been increasing day by day. However, with the frequent occurrence of extreme weather, the large wind speed fluctuations caused by it may lead to potential disasters, especially in large-capacity wind farms. In order to formulate effective prevention and control strategies in a timely manner, it is particularly important to predict the wind power under extreme weather in advance. The impact of extreme weather on wind farms is directly reflected in the significant changes in wind speed within a short period of time. The extreme weather power period shows the characteristics of large-amplitude and intense fluctuations in power on the time series scale. Identifying and extracting the extreme weather power period has become the primary task.

[0003] Under extreme meteorological conditions represented by turning weather, the wind power fluctuates violently within a short period of time. The existing ultra-short-term wind power single-value prediction methods lack a prediction model for extreme weather, resulting in poor prediction accuracy and stability; secondly, for the flat power period, the traditional single-value prediction method has higher prediction accuracy. For the extreme weather power period, the probability prediction method has better prediction performance because it can quantify the prediction error. The existing methods lack a prediction method optimization strategy based on the extreme weather period and cannot combine the advantages of single-value prediction and probability prediction throughout the whole period. Therefore, for the prediction under the traditional "simple meteorological mode", it is necessary to further find an ultra-short-term wind power prediction method suitable for the "complex meteorological mode". In particular, the characteristics of turning weather should be considered and an adaptive power mutation identification mechanism should be constructed to further improve the generalization ability of the model.

[0004] The current research on wind power mutation under turning weather is still limited to the description and prediction of wind power ramp events. The existing technologies mainly adopt the following methods: iteratively optimizing the parameters of the probability generation model through the genetic algorithm of the multi-objective fitness function to obtain a large number of prediction scenarios, and evaluating the prediction method through the probability characteristics of the ramp events mined within the scenario capture band. However, this model still needs to improve its robustness under extreme weather; based on the event detection framework, using data-driven algorithms to improve the prediction accuracy. However, in this model, the power ramp events still use traditional evaluation indicators and lack the ability to flexibly handle interference factors such as pseudo inflection points; continuously adjusting through the ensemble learning method to generate probability predictions to quantify the uncertain factors of the prediction. However, this method lacks a precise detection and identification method for wind power ramp events and does not highlight the improvement of the model's accuracy under extreme weather. To sum up, although preliminary research results have been obtained in the detection, identification and advanced prediction of wind power mutation periods, there is still room for improvement in the refined identification of mutation periods and other aspects. Summary of the Invention

[0005] The object of the present invention is to provide a short-term wind power segmented prediction method based on the recognition of turning periods to overcome the defects existing in the above-mentioned prior art.

[0006] The object of the present invention can be achieved by the following technical solutions:

[0007] A short-term wind power segmented prediction method based on the recognition of turning periods includes the following steps:

[0008] 1) Extract the time series trend, obtain the EMA curve representing the short-term development trend of the original wind power data of the wind farm, perform smoothing processing using the Gaussian window method, and then obtain the change rate α at each moment.

[0009] 2) Based on the EMA curve obtained in step 1), formulate a window adjustment strategy using the local feature differences of the EMA curve, set a detection threshold ε. If the distribution difference fluctuation between the window and the previous window is less than the detection threshold ε, then expand the window width to accelerate the detection speed; otherwise, shrink the window width to improve the detection accuracy.

[0010] 3) Based on the window adjustment strategy of the local feature differences formulated in step 2), use the α obtained in step 1) as one of the criteria, and mark the positions where the minimum values of the means appear within two windows as inflection points.

[0011] 4) Improve the traditional power mutation period criterion, merge adjacent mutation periods with the same trend, and completely extract the turning weather mutation periods.

[0012] 5) Based on the extraction result of the adaptive turning period, divide the time series into a turning segment and a flat segment.

[0013] 6) For the flat segment, perform point prediction. Use GRU as the original algorithm for point prediction, introduce an improved Attention mechanism combined with the CRS algorithm, assign different weights to the transition feature vectors of the neural network model, and then transmit the attention weights to the GRU layer to output the training results of the GRU neural network. Read the training loss curve and error curve, observe the vertical distance between the training set and validation set loss curves during the convergence process, and combine the absolute error conditions of the training set and validation set to visually evaluate the convergence performance of the network prediction results.

[0014] 7) For the turning segment, perform probability prediction using the time series pattern - adaptive bandwidth kernel density estimation method for probability prediction.

[0015] 8) Combine steps 6) and 7) to complete the short-term wind power prediction based on the turning period, and obtain the predicted power.

[0016] Further, in step 1), the moving average method is used to extract the time series trend.

[0017] Further, in step 2), the specific steps of formulating the window adjustment strategy by using the local feature difference of the EMA curve include:

[0018] 21) Segment the original power time series into several segments, and perform turning point detection on each segment;

[0019] 22) In the turning point detection, define diff i as the measure of the distribution difference fluctuation between the i-th window and the previous window, denoted as where Vs i is the mean fluctuation of the data distribution of the i-th window of the data to be detected, and Ds i is the difference fluctuation;

[0020] 23) Set a threshold ε. If the value of diff i is less than or equal to the threshold ε, then expand the sliding window width W to increase the detection speed; if the value of diff i is greater than ε, then shrink the sliding window width W to improve the detection accuracy.

[0021] Further, in step 3), a double fixed-time sliding window is used for inflection point detection. The specific content is as follows:

[0022] First, introduce the change rate α obtained at each moment in step 1) as one of the criteria, and set the inflection point to satisfy the condition α = 0; based on the EMA curve, establish two closely connected sliding windows, update the data in the two windows frame by frame, and mark the power value at the junction where the mean difference in the two windows reaches the minimum as the inflection point. Repeat the above steps to obtain the corresponding time series trend inflection point set T ip .

[0023] Further, in step 4), the expression of the improved mutation period criterion is:

[0024]

[0025] In the formula, is the power value of the j-th point in the inflection point set T ip ; is the power value of the (j + 1)-th point in the inflection point set T ip ; is the time passing through the j-th point in the inflection point set T ip ; is the time passing through the (j + 1)-th point in the inflection point set T ip ; λ is the mutation amplitude threshold of the turning period; β is the mutation rate threshold of the turning period; merge adjacent mutation periods with the same trend to completely extract the turning period.

[0026] Further, in step 6), the mathematical expression of the GRU neural network is as follows:

[0027] z t =σ(W z ·[h t-1 ,x t )

[0028] r t =σ(W r ·[h t-1 ,x t )

[0029]

[0030]

[0031] In the formula: z t is the update gate, r t is the reset gate, X t is the current input, is the summary of the input and the past hidden layer state, h t is the output of the hidden layer, W z , W r are trainable parameter matrices.

[0032] The specific steps of introducing an improved Attention mechanism combined with the CRS algorithm to assign different weights to the transition feature vectors of the neural network model are as follows:

[0033] 61) Provide the weight W of the attention layer, and the specific calculation steps of the weight are as follows:

[0034] 611) Calculate the similarity M(Q, K) for the given task query vector Q and the attention variable K;

[0035] 612) Perform the Softmax operation on the obtained similarity for normalization to obtain the normalized similarity η i :

[0036]

[0037] 613) For the weights calculated above, perform a weighted sum of all the obtained weights to obtain the Attention vector;

[0038] 62) Convert the provided weight W of the attention layer into a binary code W B , and the subset W i is the attention weight. Transmit the subset to the GRU neural network, and generate a corresponding loss value in the GRU neural network according to the prediction error in the network;

[0039] 63) According to WB Select the optimal attention weight subset \(W\) according to the loss situation i B and And perform repeated loops on its subset combinations;

[0040] 64) Reconstruct a new attention weight

[0041] Furthermore, in step 7), the specific steps of probability prediction using the time series pattern - adaptive bandwidth kernel density estimation method include:

[0042] 71) Divide the time series pattern based on the power period characteristics, and divide it into five categories: sharp rise, sharp fall, slow rise, slow fall, and oscillation;

[0043] 72) Use the empirical distribution estimation method to establish a time series pattern - wind power prediction error probability density distribution model for each category under different weather type conditions.

[0044] In step 71), taking \(\alpha\) as the division basis, divide the weather types, calculate the power prediction error probability density distribution under the corresponding time series pattern characteristics, and visually reflect the distribution through a box plot; obtain the optimal window width using the progressive integral mean square error method, substitute it into the estimation function, and respectively fit the probability density distribution curves under each time series characteristic to provide a basis for presenting the interval prediction results.

[0045] The ultra - short - term wind power segmented prediction method based on turning period recognition provided by the present invention has at least the following beneficial effects compared with the prior art:

[0046] 1) The method of the present invention proposes a power time series trend discrimination method based on moving average iteration for the situation of variable wind power scenarios under turning weather conditions, fully considering the historical wind power time series characteristics, which is beneficial to improving the accuracy of trend description in the power mutation period under turning weather;

[0047] 2) Aiming at the problem of insufficient extraction of the power time series turning period, the present invention proposes a sliding window width adjustment strategy based on local time series feature differences to improve the integrity of turning period extraction and significantly improve the algorithm effectiveness;

[0048] 3) Aiming at the differences in power time series characteristics under meteorological patterns, a segmented power prediction method of point prediction - probability interval prediction based on time series feature matching is proposed. The point prediction uses an improved GRU algorithm, reducing the training amount, and the probability interval prediction uses the variable bandwidth kernel density estimation method, improving the prediction performance. Description of the Drawings

[0049] Figure 1Schematic flowchart of the ultra-short-term wind power segmented prediction method based on the identification of turning periods in the embodiment;

[0050] Figure 2 Schematic diagram of the window adjustment strategy in the embodiment;

[0051] Figure 3 Schematic flowchart of the improved Attention mechanism combined with CRS in the embodiment;

[0052] Figure 4 Power prediction error distribution under each time series mode in the embodiment. Detailed implementation manners

[0053] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0054] Embodiment

[0055] The present invention relates to an ultra-short-term wind power segmented prediction method based on the identification of turning periods. The method extracts the time series trend based on the moving average method, realizes the division of turning periods through an adaptive window adjustment method, adopts different prediction methods considering different characteristic periods, formulates a segmented prediction strategy of point prediction - probability prediction, and makes accurate predictions on the wind power of the whole period including turning weather.

[0056] The main principle of the ultra-short-term wind power segmented prediction model based on turning weather established by the present invention is as follows:

[0057] Regarding the extraction of turning weather periods, the power periods of extreme weather show the characteristics of large and intense fluctuations in power on the time series scale. To improve the prediction accuracy of wind power under turning weather, identifying and extracting the power periods of extreme weather is the primary task, and it is necessary to achieve accurate and complete extraction of the mutation periods as much as possible. Since the meteorological patterns vary widely in a short time under turning weather, the traditional extraction of time series features of the original power time series is prone to cause misjudgment of trends. Therefore, the power time series trend discrimination method of moving average iteration is selected to improve the accuracy of period trend description. In order to fully extract the turning periods of the power time series, the traditional climbing extraction algorithm represented by the fixed-time sliding window is difficult to accurately and completely extract the turning power periods. The present invention adopts a sliding time window width adjustment strategy based on local time series feature differences to significantly improve the effectiveness of the algorithm.

[0058] In terms of prediction methods, the current single-value prediction method lacks a prediction model for extreme weather, resulting in poor prediction accuracy and stability. For the flat power period, the traditional single-value prediction method has higher prediction accuracy. For the extreme weather power period, the probability prediction method has better prediction performance because it can quantify the prediction error. Therefore, the present invention proposes to adopt a segmented prediction strategy to accurately predict the wind power of the whole period including transitional weather, that is, point prediction is used for the flat period time series, and probability prediction is used for the transitional period time series.

[0059] LSTM is an RNN model. Each unit in the RNN not only processes the input data at the current time point but also processes the output of the previous unit, and finally outputs a single prediction. The basic RNN model only processes the output of the previous unit, so the output of units far away gradually disappears because it is processed multiple times in the middle. To address the above problems, the present invention proposes to use the GRU algorithm as the original algorithm for point prediction. GRU is a simplified variant of the LSTM network and belongs to the gated recurrent neural network. The update gate in GRU is formed by merging the forget gate and the input gate in the LSTM network. The model architecture is simpler, reducing the computational amount and training time while ensuring the prediction accuracy of the model. On the basis of the existing model, an Attention mechanism combined with the CRS algorithm is introduced to guide the weight distribution between time series.

[0060] Traditional wind power interval prediction often only considers the distribution of prediction errors at different power levels, ignoring the impact of power mutations caused by weather type transitions on prediction errors. To address the above problems, the present invention divides the time series pattern based on the characteristics of the power period, which is divided into five categories: sharp rise, sharp fall, slow rise, slow fall, and oscillation. Combining the distribution characteristics of wind power prediction errors in different weather type periods, the prediction performance of the model is improved, and an empirical distribution estimation method is used to establish a time series pattern-wind power prediction error probability density distribution model for each category under different weather type conditions.

[0061] Based on the above principles and design ideas, as Figure 1 shown, the ultra-short-term wind power segmented prediction method based on transitional period recognition of the present invention specifically includes the following steps:

[0062] Step 1: Extract the time series trend. First, obtain the EMA curve representing the short-term development trend of the original data (the original wind power data of the wind farm); second, perform smoothing processing using the Gaussian window method to obtain the change rate α at each moment. The specific operations are as follows:

[0063] 11) Obtain the EMA curve

[0064] The EMA curve uses a statistical processing method to perform weighted averaging on the original data, and then the connected curve is used to observe the change trend of the future trend of the data.

[0065] Obtain the N-day smoothed moving average Y of t N , Y N-1 is the smoothed moving average of N-1 days, that is, the EMA curve is obtained as follows:

[0066]

[0067] 12) Obtain α

[0068] Before the rising or falling trend of the time series changes, the slope of the time series can intuitively reflect the change trend of the time series. The index of the change trend of the power time series can be obtained by tracking the slope of the time series. For the above-obtained EMA curve, first smooth the curve using the Gaussian window method, and then calculate the initial power mutation sensitivity factor α from the current moment change rate. The calculation formula is as follows:

[0069]

[0070] Among them, Y smooth,N (t) is the EMA value before smoothing at time t. Y smooth,N-1 (t - Δt) is the EMA value after smoothing after the rising or falling trend of the time series changes by Δt.

[0071] 13) Trend extraction

[0072] When the original power is above the short-term moving average, α > 0. When the original power is below the short-term moving average, α < 0. Then, according to the aggregation and separation of the short-term moving average and the original power, and further combining the time series characteristics of the moving average itself, the high and low points of the prediction object can be intuitively judged from the image.

[0073] Step 2: Propose an adaptive time window turning period division method to quickly identify the turning power period. First, for the EMA curve obtained in the previous step, use the local feature differences of the EMA curve to calculate the distribution differences between adjacent windows, and then set a detection threshold. If the distribution difference fluctuation between a window and its previous window is less than this value, the window width is enlarged to speed up the detection speed. If it is greater than this value, the window width is reduced to improve the detection accuracy.

[0074] Most power mutation detection algorithms judge whether there is a power time series mutation by the difference between the distribution of the local data to be detected and the standard data distribution. The window adjustment strategy based on time series characteristics can speed up the extraction of turning periods.

[0075] The specific steps of the window adjustment method based on local feature differences are as follows:

[0076] A) Introduce the original sliding window model. First, divide the original power time series into several segments, and then detect the turning points of each segment;

[0077] B) In the turning point detection, calculate the mean fluctuation Vs of the data distribution of the i-th window data of the data to be detected i and the difference fluctuation Ds i , and the expressions are as follows:

[0078]

[0079]

[0080] In the formula, U i represents the data within the i-th window, U represents the entire original power time series, var and std represent variance and standard deviation respectively. max(U) and min(U) represent the maximum and minimum values of the entire original power time series respectively.

[0081] C) Calculate the distribution difference diff between the i-th window and the previous window i .

[0082]

[0083] D) As Figure 2 shown, for the distribution difference diff obtained from the above formula i , set the threshold ε = 0.2. If the value of diff i is less than or equal to ε, it belongs to the same data feature, and the sliding window width W is enlarged; if the value of diff i is greater than ε, it is in the turning period, and the sliding window width W is reduced, finally achieving the purpose of adaptive window adjustment.

[0084] The adaptive window adjustment method is used to identify the power mutation period, so as to further extract the inflection point set, greatly increasing the detection speed and improving the detection accuracy.

[0085] Step 3: Detect the inflection points in the trend by using the method of double fixed-time sliding window. First, based on the above-mentioned window adjustment strategy for local feature differences, introduce α as one of the criteria, and mark the positions where the mean values in the two windows appear as minima as inflection points; then improve the traditional power mutation period criterion, merge adjacent mutation periods with the same trend, and completely extract the turning weather mutation period.

[0086] The specific steps for extracting the inflection point set in the trend are as follows:

[0087] A) In the oscillating output period, there will be a situation of misjudging inflection points. The method of the present invention introduces α as one of the criteria, that is, the inflection point must satisfy the condition α = 0;

[0088] B) Based on the EMA curve, two closely connected sliding windows are established, and the data within the two windows are updated frame by frame. In the present invention, the mean value of the sliding window is used as the benchmark for difference comparison, and the inflection point detection score S c The calculation method is as follows:

[0089]

[0090] In the formula: X 1,i is all the power data within the previous window, and X 2,i is all the power data within the subsequent window. When the difference between the mean values within the two windows reaches the minimum, the power value at the junction point is marked as the inflection point. Repeat the above steps to finally obtain the corresponding time series trend inflection point set T ip .

[0091] Then, the traditional power period criterion is improved, and the improved mutation period criterion is as follows:

[0092]

[0093] In the formula, is the power value of the j-th point in the inflection point set T ip ; is the power value of the (j + 1)-th point in the inflection point set T ip ; is the time passing through the j-th point in the inflection point set T ip ; is the time passing through the (j + 1)-th point in the inflection point set T ip ; λ is the mutation amplitude threshold of the transition period; β is the mutation rate threshold of the transition period; not only the change in the amplitude of the mutation period is considered, but also the mutation rate is considered, so as to exclude the existence of pseudo-inflection points.

[0094] Finally, adjacent mutation periods with the same trend are merged to completely extract the power transition period.

[0095] Step Four: Based on the extraction result of the adaptive transition period, the time series is divided into a transition segment and a flat segment. For the flat segment, point prediction based on the GRU algorithm is used, and for the transition segment, probability prediction using the time series pattern - adaptive bandwidth kernel density estimation method is used.

[0096] Step Five: Input the time series features extracted by the GRU network into the improved Attention mechanism combined with the CRS algorithm, assign different weights to the transition feature vectors of the neural network model, and then input the weighted transition feature vectors into the GRU layer according to the time steps to output the training results of the improved GRU neural network. Read the training loss curve and error curve, observe the vertical distance between the training set and validation set loss curves during the convergence process, and combine the absolute error conditions of the training set and validation set to visually evaluate the convergence performance of the network prediction results.

[0097] The following are the convergence situations represented by three common fitting states:

[0098] 1) When the loss curve of the training set hardly decreases, it is in an underfitting state, which is a non-convergent state;

[0099] 2) When the loss curve of the training set continues to decrease and the loss curve of the validation set no longer decreases at a certain moment, it is in an overfitting state, which is a convergent state but not a perfect convergence;

[0100] 3) When there is no obvious gap between the loss curves of the training set and the validation set, it is in a perfect fitting state and a perfect convergence.

[0101] Furthermore, the present invention improves and optimizes the traditional GRU model, and combines the Attention mechanism of the CRS algorithm. The specific content is as follows:

[0102] Since there are numerous feature quantities input into the model, in order to highlight more critical influencing factors and help the model make more accurate judgments, the present invention proposes an improved Attention mechanism to assign different weights to the transition feature vectors of the neural network model. In the traditional Attention mechanism, the information carried by the content input into the network first will be covered by the information input later, and the semantic vector may not be able to fully represent the information of the entire sequence. Therefore, aiming at the above deficiencies, the present invention proposes an improved Attention mechanism combined with the CRS (Competitive random search) algorithm, which makes up for the deficiency of the network's attention to the feature of different relevant factors on the same time scale and improves the network's attention degree to various relevant factors.

[0103] CRS is used to generate the optimal parameter combination in the attention layer. Figure 3 The running process of CRS is introduced in [reference], and CRS consists of four parts: "I, II, III, IV".

[0104] "I" provides the weight W of the attention layer; then it is converted into binary code in "II", and the subset W i is the attention weight, which is transmitted to the GRU neural network, and the corresponding loss value is generated there according to the prediction error in the network. Then, according to the loss situation of W B in "III", the optimal attention weight subset W i B and its subset combination is repeatedly looped. Finally, a new attention weight is reconstructed in "IV"

[0105] The detailed steps of CRS are as follows:

[0106] 1) Randomly generate an attention weight set with length M = n (n is the dimensionality of the model input features).

[0107] 2) Input the subset W i into the attention layer and convert W into a binary code:

[0108] 3) Calculate the prediction error according to the true value y and the predicted value of the GRU model:

[0109] 4) Select the optimal attention weight subset W i B and Each subset consists of binary strings and is evenly divided into n segments. Correspondingly, W i B and are represented by W i B =(F i 1 , F i 2 ,... F i n ) and respectively. F i 1 and are parts of W i B and respectively.

[0110] 5) Randomly select a part of W i B and For example, select the (n - 1)-th segment F i n-1 and However, the number of selected segments is not fixed.

[0111] 6) Obtain the genetic recombination of F i n-1 and F i n-1 and are represented by binary codes of length 6, and they randomly exchange at the corresponding 6 indices to obtain the recombinant segment

[0112] 7) Simulate a gene mutation and reverse the genotype. For example, 0 is reversed to 1. Then replaces in Wi B the corresponding F in i n-1 , forming a new inserted into W B .

[0113] 8) W B is decoded to obtain an updated set of attention weights: W′ = (W′ 1 , W′ 2 ,..., W′ k ,..., W′ M ).

[0114] 9) Repeat steps 2) - 8) until a preset number of times K is reached.

[0115] Furthermore, the mathematical expression of the GRU neural network is:[[]]

[0116] z t = σ(W z · [h t-1 , x t )

[0117] r t = σ(W r · [h t-1 , x t )

[0118]

[0119]

[0120] In the formula: z t is the update gate, r t is the reset gate, X t is the current input, is the summary of the input and the past hidden layer state, h t is the output of the hidden layer, W z , W r are trainable parameter matrices.

[0121] Step 7: For the time - series mutation period, the specific steps of the time - series pattern - adaptive bandwidth kernel density estimation method for probability prediction are as follows:

[0122] 1) Based on the power period characteristics, divide the time - series pattern into five categories: sharp rise, sharp decline, slow rise, slow decline, and oscillation;

[0123] 2) Use the empirical distribution estimation method to establish a time - series pattern - wind power prediction error probability density distribution model for each category under different weather type conditions. The specific content is as follows:

[0124] 21) Suppose there is a set of sample observations \(x\) of size \(m\) for the overall variable \(X\). 1 , \(x\) 2 , … \(x\) p … \(x\) m After rearranging them in ascending order, the order statistic \(x'\) is obtained. 1 ', \(x\) 2 ', … \(x\) p '… \(x\) m ' For any real number \(x\), its empirical distribution expression is:

[0125]

[0126] 22) Calculate the probability distribution of the power prediction error for the corresponding time series pattern features, and obtain the box plot based on the error distributions of multiple meteorological patterns. For example, as shown Figure 4 in the figure, the left and right boundaries of the square are the positions corresponding to the 50% quantiles, the scale line at the center position is the median of the error of this type of pattern, and the external points are outliers (abnormal values). This figure intuitively reflects the distribution of the power prediction error under each time series pattern;

[0127] 23) Obtain the original prediction error probability distribution \(f\) from the empirical distribution error , and estimate the formula for calculating the kernel density of \(f\) e as follows:

[0128]

[0129] In the formula, \(N\) is the sample size of a set of data; \(h\) is the window width, also known as the smoothing parameter; \(K(u)\) is the kernel function, \(u = h\) -1 (\(e - e\) i ) ; \(e\) i is the \(i\)-th sample value in the prediction error data.

[0130] 24) Select the Gaussian kernel function and substitute it into the estimation expression. The specific expression of the Gaussian kernel function is as follows:

[0131]

[0132] Substituting it into the expression gives:

[0133]

[0134] 25) First, introduce the integrated mean square error to judge the difference between the estimated probability density function and the true probability density function \(f\) e \((x)\). The expression is as follows:

[0135]

[0136] Among them MISE is expressed as the sum of the main item AMISE. When h→0 and nh→∞, the AMISE expression is defined as:

[0137]

[0138] In the formula, AMISE is an expression about the window width h. When AMISE takes the minimum value, the optimal h value is reached, that is The optimal window width h can be obtained x The expression is as follows:

[0139]

[0140] 3) Fit the probability density distribution curves under each time series feature respectively to provide a theoretical basis for presenting the interval prediction results.

[0141] Traditional interval prediction only considers historical power levels to classify errors, while considering the physical dynamic process of time series pattern classification can effectively improve the prediction accuracy of interval prediction. The present invention proposes a short-term interval prediction method based on time series pattern classification and the Monte Carlo method, which has the best effect. Under different confidence levels, an interval coverage rate greater than the preset confidence level can be obtained.

[0142] Step eight: Establish a point prediction - probability prediction segmented prediction model to quantitatively analyze the result accuracy of full-time point prediction, probability prediction and segmented prediction point prediction, probability prediction.

[0143] The method of the present invention considers the full-time wind power prediction including turning weather. First, a power time series trend discrimination method based on moving average iteration is proposed, and the Gaussian window method is used to smooth the EMA curve to extract the turning weather trend; secondly, a window adjustment strategy based on local time series features is proposed to adaptively adjust the time window width, and an inflection point detection strategy of a double fixed-time sliding window is adopted. Combined with the criterion, an inflection point set is obtained, and the turning weather mutation period is extracted and divided. The proposed time series segmented prediction algorithm for point prediction and probability interval prediction is used to perform power prediction on the full time period: the improved GRU algorithm is used for point prediction of the turning segment time series, and an improved Attention mechanism combined with the CRS algorithm is proposed to assign different weights to the transition feature vectors of the neural network model; probability prediction is used for the smooth segment time series, and an empirical distribution estimation method is used to establish a time series pattern - power prediction error probability density distribution model, and wind power probability prediction is performed based on the variable bandwidth kernel density estimation method. This model not only conforms to the typical characteristics of mutating weather, but also significantly improves the generalization ability and prediction performance of full-time power prediction.

[0144] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or substitutions within the technical scope disclosed by the present invention, and these modifications or substitutions should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A short-term wind power segmented prediction method based on the identification of turning periods, characterized in that, it includes the following steps: 1) Extract the time series trend, obtain the EMA curve representing the short-term development trend of the original wind power data of the wind farm, perform smoothing processing using the Gaussian window method, and then obtain the change rate α at each moment; 2) Based on the EMA curve obtained in step 1), formulate a window adjustment strategy using the local feature differences of the EMA curve, set the detection threshold ε. If the distribution difference fluctuation between the window and its previous window is less than the detection threshold ε, then widen the window width to accelerate the detection speed. Otherwise, narrow the window width to improve the detection accuracy; 3) Based on the window adjustment strategy of the local feature differences formulated in step 2), use the α obtained in step 1) as one of the criteria, and mark the positions where the minimum values of the means appear within two windows as inflection points; 4) Improve the traditional power mutation period criterion, merge adjacent mutation periods with the same trend, and completely extract the turning weather mutation periods; the expression of the improved mutation period criterion is: Wherein, is the inflection point set T ip the power value of the j-th point in; is the inflection point set T ip the power value of the (j + 1)-th point in; is the inflection point set T ip the time passing through the j-th point in; is the inflection point set T ip the time passing through the (j + 1)-th point in; λ is the mutation amplitude threshold of the turning period; β is the mutation rate threshold for the turning period; merge adjacent mutation periods with the same trend to completely extract the turning period; 5) Based on the adaptive turning period extraction result as the division basis, divide the time series into turning segments and gentle segments; 6) Perform point prediction on the gentle segment. Use GRU as the original algorithm for point prediction, introduce an improved Attention mechanism combined with the CRS algorithm, assign different weights to the transition feature vectors of the neural network model, and then transmit the attention weights to the GRU layer to output the training results of the GRU neural network. Read the training loss curve and error curve, observe the vertical distance between the training set and validation set loss curves during the convergence process, and combine the absolute error conditions of the training set and validation set to intuitively evaluate the convergence performance of the network prediction results; the specific steps of introducing an improved Attention mechanism combined with the CRS algorithm and assigning different weights to the transition feature vectors of the neural network model include: 61) Provide the weight W of the attention layer; 62) Convert the weights W of the provided attention layer into a binary code W B , subset W i is the attention weight, transmit the subset to the GRU neural network, and generate corresponding loss values according to the prediction error in the GRU neural network; Select the optimal attention weight subset \(W\) according to the loss situation of \(W\). B i B and Perform repeated loops on its subset combinations;​ 64) Reconstruct a new attention weight 7) Perform probability prediction on the turning segment using the time series pattern - adaptive bandwidth kernel density estimation method for probability prediction; 8) Combine steps 6) and 7) to complete the short-term wind power prediction based on the turning period, and obtain the predicted power.

2. The short-term wind power segmented prediction method based on the identification of turning periods according to claim 1, characterized in that, in step 1), the time series trend is extracted using the moving average method.

3. The short-term wind power segmented prediction method based on the identification of turning periods according to claim 1, characterized in that, the specific steps of formulating a window adjustment strategy using the local feature differences of the EMA curve in step 2) include: 21) Cut the original power time series into several segments, and perform inflection point detection on each segment; 22) In the turning point detection, define diff i as the measure of the distribution difference fluctuation between the i-th window and the previous window, denoted as where Vs i is the mean fluctuation of the data distribution of the i-th window of the data to be detected, and Ds i is the difference fluctuation; 23) Set a threshold ε. If the value of diff i is less than or equal to the threshold ε, then expand the sliding window width W to increase the detection speed; if the value of diff i is greater than ε, then shrink the sliding window width W to improve the detection accuracy.

4. The short-term wind power segmented prediction method based on the identification of turning periods according to claim 1, characterized in that, in step 3), double fixed-time sliding windows are used for inflection point detection.

5. The short-term wind power segmented prediction method based on the identification of turning periods according to claim 4, characterized in that, The specific content of using a double fixed-time sliding window for inflection point detection is as follows: First, introduce the change rate α obtained at each moment in step 1) as one of the criteria, and set that the inflection point satisfies the condition α = 0; based on the EMA curve, establish two closely connected sliding windows, update the data in the two windows frame by frame, and mark the power value at the junction point where the difference in the means in the two windows reaches the minimum as the inflection point. Repeat the above steps to obtain the corresponding set of inflection points of the time series trend T ip .

6. The ultra-short-term wind power segmented prediction method based on turning period recognition according to claim 1, characterized in that, in step 6), the mathematical expression of the GRU neural network is: z t = σ(W z · [h t-1 , x t ) r t = σ(W r · [h t-1 , x t ) where: z t is the update gate, r t is the reset gate, X t is the current input, is the summary of the input and the past hidden layer state, h t is the hidden layer output, W z and W r are trainable parameter matrices.

7. The ultra-short-term wind power segmented prediction method based on turning period recognition according to claim 1, characterized in that, in step 7), the specific steps of probability prediction using the time series pattern - adaptive bandwidth kernel density estimation method include: 71) Divide the time series pattern based on the power period characteristics, which are divided into five categories: sharp rise, sharp fall, slow rise, slow fall, and oscillation; 72) Use the empirical distribution estimation method to establish a time series pattern - wind power prediction error probability density distribution model for each category under different weather type conditions.

8. The ultra-short-term wind power segmented prediction method based on turning period recognition according to claim 7, characterized in that, in step 71), taking α as the division basis, divide the weather types, calculate the power prediction error probability density distribution under the corresponding time series pattern characteristics, and visually reflect the distribution through a box plot; obtain the optimal window width using the progressive integral mean square error method, substitute it into the estimation function, and respectively fit the probability density distribution curves under each time series characteristic to provide a basis for presenting the interval prediction results.

Images