CN114386324A - Ultra-short-term wind power segmented prediction method based on turning period identification - Google Patents
Ultra-short-term wind power segmented prediction method based on turning period identification Download PDFInfo
- Publication number
- CN114386324A CN114386324A CN202111609901.7A CN202111609901A CN114386324A CN 114386324 A CN114386324 A CN 114386324A CN 202111609901 A CN202111609901 A CN 202111609901A CN 114386324 A CN114386324 A CN 114386324A
- Authority
- CN
- China
- Prior art keywords
- prediction
- time sequence
- turning
- wind power
- window
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 84
- 238000001514 detection method Methods 0.000 claims abstract description 32
- 230000008859 change Effects 0.000 claims abstract description 21
- 230000007246 mechanism Effects 0.000 claims abstract description 15
- 238000009499 grossing Methods 0.000 claims abstract description 9
- 239000000284 extract Substances 0.000 claims abstract description 6
- 238000012935 Averaging Methods 0.000 claims abstract 2
- 230000035772 mutation Effects 0.000 claims description 25
- 238000012549 training Methods 0.000 claims description 17
- 238000000605 extraction Methods 0.000 claims description 13
- 239000013598 vector Substances 0.000 claims description 13
- 238000013528 artificial neural network Methods 0.000 claims description 12
- 230000014509 gene expression Effects 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 8
- 230000000630 rising effect Effects 0.000 claims description 8
- 230000007704 transition Effects 0.000 claims description 8
- 238000012795 verification Methods 0.000 claims description 8
- 238000003062 neural network model Methods 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 6
- 230000010355 oscillation Effects 0.000 claims description 5
- 238000011161 development Methods 0.000 claims description 4
- 150000001875 compounds Chemical class 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- PHTXVQQRWJXYPP-UHFFFAOYSA-N ethyltrifluoromethylaminoindane Chemical compound C1=C(C(F)(F)F)C=C2CC(NCC)CC2=C1 PHTXVQQRWJXYPP-UHFFFAOYSA-N 0.000 claims description 2
- 230000000750 progressive effect Effects 0.000 claims description 2
- 230000009977 dual effect Effects 0.000 claims 2
- 230000011218 segmentation Effects 0.000 abstract 1
- 230000009194 climbing Effects 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000013475 authorization Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 230000007423 decrease Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 206010064571 Gene mutation Diseases 0.000 description 1
- 238000000342 Monte Carlo simulation Methods 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 230000001174 ascending effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000011217 control strategy Methods 0.000 description 1
- 230000010485 coping Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000012850 discrimination method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 238000013139 quantization Methods 0.000 description 1
- 230000006798 recombination Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 230000037221 weight management Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/06—Wind turbines or wind farms
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Geometry (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Algebra (AREA)
- Computational Mathematics (AREA)
- Computer Hardware Design (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Medical Informatics (AREA)
- Pure & Applied Mathematics (AREA)
- Databases & Information Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to an ultra-short-term wind power sectional prediction method based on turning time interval identification, which extracts a time sequence trend by utilizing a mobile averaging method; smoothing an exponential moving average line (EMA) by adopting a Gaussian window method, and calculating the time sequence change rate alpha at each moment; the window adjustment strategy based on the local time sequence characteristics adaptively adjusts the time window width; based on an inflection point detection strategy of a double-timing sliding window, introducing alpha as one of criteria, and extracting and dividing a turning weather sudden change time period; adopting improved GRU algorithm point prediction for a turning section time sequence, and combining an improved Attention mechanism of a CRS algorithm; probability prediction is carried out on the gentle section time sequence, an empirical distribution estimation method is adopted to establish a time sequence mode-power prediction error probability density distribution model, and wind power probability prediction is carried out on the basis of a variable bandwidth kernel density estimation method; and combining the point prediction and the probability prediction to obtain a final prediction result by time sequence segmentation prediction. Compared with the prior art, the method has the advantages of improving the model operation efficiency and the like.
Description
Technical Field
The invention relates to the technical field of ultra-short-term wind power prediction of a high-concentration wind power plant, in particular to an ultra-short-term wind power sectional prediction method based on turning time interval identification.
Background
In recent years, with the development of large-scale and high-concentration wind power, the proportion of wind power accessed to a power grid is increasing day by day. But with the frequent occurrence of extreme weather, the resulting large wind speed fluctuations may lead to a potential disaster, especially in large-capacity wind farms. In order to make an effective prevention and control strategy in time, it is important to predict wind power in extreme weather in advance. The influence of extreme weather on the wind power plant is visually reflected as great change of wind speed in a short time, the extreme weather power period is characterized by great and violent fluctuation of power on a time sequence scale, and the identification and extraction of the extreme weather power period become a primary task.
Under extreme meteorological conditions represented by turning weather, the wind power fluctuates sharply in a short time. The existing ultra-short-term wind power single-value prediction method lacks a prediction model aiming at extreme weather, so that the prediction precision and stability are poor; secondly, for a gentle power period, the traditional single-value prediction method has higher prediction accuracy, for an extreme weather power period, the probability prediction method has better prediction performance due to the quantization of prediction errors, and the existing method lacks a prediction method optimization strategy based on the extreme weather period and cannot have the advantages of single-value prediction and probability prediction in the whole period. Therefore, an ultra-short-term wind power prediction method suitable for the complex meteorological mode needs to be further found for the traditional prediction in the simple meteorological mode. Particularly, the generalization capability of the model is further improved by considering the combination of the characteristics of the turning weather and the construction of an adaptive power mutation identification mechanism.
The current research aiming at the sudden change of the wind power under the turning weather is limited to the description and prediction of the wind power climbing event, and the method mainly adopted in the prior art comprises the following steps: iterative optimization is carried out on probability generation model parameters through a genetic algorithm of a multi-target fitness function to obtain a large number of prediction scenes, and the prediction method is evaluated through the probability characteristics of the slope climbing events excavated in the scene capture zone, but the model still needs to improve the robustness under extreme weather; based on an event detection framework, a data-driven algorithm is adopted to improve the prediction accuracy, but the power ramp event in the model still uses the traditional evaluation index and lacks flexible coping capability on interference factors such as a pseudo-inflection point and the like; through continuous adjustment of an integrated learning method, probability prediction is generated to quantify uncertain factors of prediction, however, the method is lack of a precise detection and identification method for wind power climbing events, and precision improvement of the model under extreme weather is not highlighted. In summary, preliminary research results have been obtained by considering detection and identification and advanced prediction of wind power abrupt change time periods, but a certain promotion space still exists in the aspects of abrupt change time period fine identification and the like.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide an ultra-short-term wind power segment prediction method based on turning period identification.
The purpose of the invention can be realized by the following technical scheme:
an ultra-short-term wind power segmented prediction method based on turning period identification comprises the following steps:
1) extracting a time sequence trend, solving an EMA curve representing the short-term development trend of the original wind power data of the wind power plant, and solving the change rate alpha at each moment after smoothing by adopting a Gaussian window method;
2) based on the EMA curve obtained in the step 1), utilizing the local characteristic difference of the EMA curve to make a window adjustment strategy, setting a detection threshold value epsilon, if the distribution difference fluctuation between a window and the previous window is smaller than the detection threshold value epsilon, expanding the window width to accelerate the detection speed, and otherwise, reducing the window width to improve the detection precision;
3) based on the window adjustment strategy of the local feature difference formulated in the step 2), marking the position of the minimum value of the mean value in the two windows as an inflection point by using alpha obtained in the step 1) as one of the criteria;
4) improving the criterion of the traditional power mutation time interval, combining adjacent same-trend mutation time intervals, and completely extracting the turning weather mutation time interval;
5) dividing the time sequence into a turning section and a gentle section according to the self-adaptive turning time period extraction result as a dividing basis;
6) point prediction is adopted for the gentle section, GRU is adopted as an original algorithm of the point prediction, an improved Attention mechanism combined with a CRS algorithm is introduced, different weights are given to transition characteristic vectors of a neural network model, Attention weights are transmitted to a GRU layer, training results of the GRU neural network are output, training loss curves and error curves are read, longitudinal distances between the training set and the verification set loss curves in a convergence process are observed, and the convergence performance of the network prediction results is visually evaluated in combination with the absolute error conditions of the training set and the verification set;
7) adopting probability prediction for the turning section by adopting a time sequence mode-self-adaptive bandwidth kernel density estimation method;
8) and (5) combining the step 6) and the step 7) to complete ultra-short-term wind power prediction based on the turning time period, and acquiring predicted power.
Further, in the step 1), a time sequence trend is extracted by using a moving average method.
Further, in step 2), the specific step of making a window adjustment strategy by using the local feature difference of the EMA curve includes:
21) the original power time sequence is divided into a plurality of segments, and turning point detection is carried out on each segment;
22) in the inflection Point detection, diff is definediFor measuring the distribution difference fluctuation of the ith window and the previous window, record
Wherein VsiMean fluctuation Ds of data distribution of the ith window of the data to be detectediIs the difference fluctuation;
23) setting a threshold value ε, if diffiIf the value of (d) is less than or equal to the threshold value epsilon, the sliding window width W is enlarged, and the detection speed is increased; if diffiIf the value of (d) is greater than epsilon, the sliding window width W is reduced, and the detection accuracy is improved.
Further, in the step 3), a double-timing sliding window is adopted for inflection point detection. The concrete contents are as follows:
firstly, introducing the change rate alpha of each moment obtained in the step 1) as one of criteria, and setting an inflection point to meet the condition that alpha is 0; based on an EMA curve, establishing two sliding windows which are closely connected, updating data in the two windows frame by frame, marking a power value at a joint point when the mean value difference in the two windows reaches the minimum as an inflection point, repeating the steps, and obtaining a corresponding time sequence trend inflection point set Tip。
Further, in step 4), the expression of the improved mutation period criterion is as follows:
in the formula (I), the compound is shown in the specification,
set T for inflection pointipThe power value of the j point;
set T for inflection pointipThe power value of the j +1 th point;
set T for inflection pointipPassing through the jth point moment;
is a point of inflectionCollection TipThe moment of passing through the j +1 th point; lambda is a break amplitude threshold value in the turning period; beta is a turning period mutation rate threshold; and combining adjacent mutation periods with the same trend to completely extract the turning period.
Further, in step 6), the mathematical expression of the GRU neural network is:
zt＝σ(Wz·[ht-1,xt])
rt＝σ(Wr·[ht-1,xt])
in the formula: z is a radical oftTo refresh the door, rtTo reset the gate, XtIn order to be the current input,
for input and past summary of hidden states, htFor hidden layer output, Wz，WrAnd W is a trainable parameter matrix.
An improved Attention mechanism combined with a CRS algorithm is introduced, and the specific steps of endowing the transition characteristic vectors of the neural network model with different weights comprise:
61) providing the weight W of the attention layer, wherein the specific calculation steps of the weight are as follows:
611) calculating similarity M (Q, K) for a given task query vector Q and an attention variable K;
612) performing Softmax operation on the obtained similarity, and normalizing to obtain normalized similarity etai：
613) Carrying out weighted summation on all obtained weights according to the calculated weights to obtain an Attention vector;
62) converting the provided weights W of the attention layer into binary code WBSubset WiTransmitting the subset to a GRU neural network for attention weighting, and generating a corresponding loss value in the GRU neural network according to a prediction error in the network;
63) according to WBSelecting the optimal attention weight subset W for the loss case ofi BAnd
and repeatedly circulating the subset combination;
64) reconstructing a new attention weight
Further, in step 7), the specific step of probability prediction by using the time-series mode-adaptive bandwidth kernel density estimation method includes:
71) dividing a time sequence mode into five types of violent rising, violent falling, slow rising, slow falling and oscillation based on the power time interval characteristics;
72) and establishing a time sequence mode-wind power prediction error probability density distribution model of each category under different weather type conditions by adopting an empirical distribution estimation method.
Step 71), dividing weather types by taking alpha as a division basis, calculating power prediction error probability density distribution under corresponding time sequence mode characteristics, and visually reflecting the distribution condition through a box line graph; and (4) solving the optimal window width by utilizing a progressive integral mean square error method, substituting the optimal window width into an estimation function, respectively fitting a probability density distribution curve under each time sequence characteristic, and providing a basis for presenting an interval prediction result.
Compared with the prior art, the ultrashort-term wind power segment prediction method based on turning time interval identification at least has the following beneficial effects:
1) aiming at the condition that the wind power scene is changeable under the turning weather condition, the method provides a power time sequence trend judgment method based on moving mean iteration, fully considers the historical wind power time sequence characteristics, and is favorable for improving the accuracy of power mutation period trend description under the turning weather;
2) aiming at the problem of insufficient extraction of the turning time interval of the power time sequence, the invention provides a sliding window width adjusting strategy based on local time sequence characteristic difference so as to improve the extraction integrity of the turning time interval and obviously improve the algorithm effectiveness;
3) aiming at power time sequence characteristic difference in a meteorological mode, a segmented power prediction method based on time sequence characteristic matching point prediction-probability interval prediction is provided, an improved GRU algorithm is adopted for point prediction, training amount is reduced, a variable bandwidth kernel density estimation method is adopted for probability interval prediction, and prediction performance is improved.
Drawings
FIG. 1 is a flow chart of an ultra-short-term wind power segment prediction method based on turning period identification in an embodiment;
FIG. 2 is a diagram illustrating an exemplary window adjustment strategy;
FIG. 3 is a schematic diagram illustrating an improved Attention mechanism for CRS integration in an embodiment;
fig. 4 shows the power prediction error distribution under each timing mode in the embodiment.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments. It is to be understood that the embodiments described are only a few embodiments of the present invention, and not all embodiments. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, shall fall within the scope of protection of the present invention.
Examples
The invention relates to an ultrashort-term wind power sectional prediction method based on turning period identification.
The main principle of the ultra-short-term wind power sectional prediction model based on the turning weather is as follows:
in the aspect of extraction of the turning weather time interval, the extreme weather power time interval is characterized by the fact that the power fluctuates greatly and violently on a time sequence scale. In order to improve the wind power prediction precision in the turning weather, the identification and extraction of the extreme weather power time interval are the primary tasks, and the accurate and complete extraction of the sudden change time interval is realized as far as possible. Because the air image modes in a short time under the turning weather are various, the trend misjudgment is easily caused by the traditional time sequence characteristic extraction of the original power time sequence, and therefore the accuracy of time period trend description is improved by adopting a power time sequence trend judgment method of moving mean iteration. In order to fully extract the power time sequence turning time period, the traditional climbing extraction algorithm represented by a timing sliding window is difficult to accurately and completely extract the turning power time period, and the invention adopts a sliding time window width adjusting strategy with local time sequence characteristic difference to obviously improve the effectiveness of the algorithm.
In the aspect of a prediction method, the conventional single-value prediction method lacks a prediction model aiming at extreme weather, so that the prediction precision and stability are poor. For a gentle power period, the traditional single-value prediction method has higher prediction accuracy, and for an extreme weather power period, the probability prediction method has better prediction performance due to the fact that prediction errors are enabled. Therefore, the invention provides that a segmented prediction strategy is adopted to accurately predict the wind power in the whole time period including the turning weather, namely point prediction is adopted in the time sequence of the gentle section, and probability prediction is adopted in the time sequence of the turning section.
LSTM is a RNN model, where each unit in RNN processes the input data at the current time point, and 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 cell, so that the output of a cell far away is affected by the multiple processing. In order to solve the problems, the invention provides an original algorithm adopting a GRU algorithm as point prediction, wherein the GRU is a simplified variant of an LSTM network and belongs to a gated recurrent neural network. The updating gate in the GRU is formed by combining a forgetting gate and an input gate in the LSTM network, the model architecture is simpler, and the calculation amount and the training time are reduced while the prediction accuracy of the model is ensured. On the basis of the existing model, an Attention mechanism combined with a CRS algorithm is introduced to guide the weight distribution among time sequences.
The traditional wind power interval prediction only considers the distribution condition of prediction errors under different power levels, and the influence of sudden power change caused by weather type conversion on the prediction errors is ignored. Aiming at the problems, the time sequence modes are divided based on the power period characteristics, and are divided into five types of modes such as violent rising, violent falling, slow rising, slow falling and oscillation. And (3) combining the distribution characteristics of the wind power prediction errors in different weather type time periods, improving the prediction performance of the model, and establishing a time sequence mode-wind power prediction error probability density distribution model of each category under different weather type conditions by adopting an empirical distribution estimation method.
Based on the above principle and design concept, as shown in fig. 1, the ultra-short-term wind power segment prediction method based on turning period identification of the present invention specifically includes the following steps:
firstly, extracting a time sequence trend, and firstly, solving an EMA curve representing the short-term development trend of original data (original wind power data of a wind power plant); secondly, smoothing by adopting a Gaussian window method, and solving the change rate at each moment as alpha, wherein the method specifically comprises the following steps:
11) calculating EMA curve
The EMA curve is used for carrying out weighted average on original data in a statistical processing mode, and then the connected curve is used for observing the change trend of the future trend of the data.
Obtaining the N-day smooth moving average value Y of tN，YN-1The EMA curve, which is a smooth moving average for N-1 days, was found as follows:
12) finding alpha
Before the ascending or descending trend of the time sequence is not changed, the time sequence slope can visually reflect the time sequence change trend, and the index of the power time sequence change trend can be searched by tracking the time sequence slope. Aiming at the EMA curve obtained above, firstly, smoothing the curve by adopting a Gaussian window method, and then calculating an initial power mutation sensitivity factor alpha according to the current time change rate, wherein the calculation formula is as follows:
wherein, Ysmooth,NAnd (t) is the EMA value before smoothing processing at the time t. Y issmooth,N-1(t- Δ t) is the EMA value after the time sequence is changed by increasing or decreasing trend and is subjected to smoothing treatment.
13) Trend extraction
When the original power is above the short term average, alpha is greater than 0, and when the original power is below the short term average, alpha is less than 0. And according to the aggregation and separation conditions between the short-term average line and the original power, the time sequence characteristics of the average line are further combined, and the high point and the low point of the prediction object can be visually judged through the image.
And step two, providing a self-adaptive time window turning period dividing method for rapidly identifying the turning power period. Firstly, aiming at the EMA curve obtained in the last step, calculating the distribution difference of adjacent windows by using the local characteristic difference of the EMA curve, then setting a detection threshold value, if the distribution difference fluctuation between a window and the previous window is less than the value, expanding the window width to accelerate the detection speed, and if the distribution difference fluctuation is more than the value, reducing the window width to improve the detection precision.
Most power mutation detection algorithms judge whether power time sequence mutation exists or not through the difference between the local data distribution to be detected and the standard data distribution, and the extraction of the turning time period can be accelerated by carrying out a window adjustment strategy based on the time sequence characteristics.
The window adjusting method based on the local feature difference comprises the following specific steps:
A) introducing an original sliding window model, firstly, segmenting an original power time sequence into a plurality of segments, and then, detecting a turning point of each segment;
B) in the turning point detection, the mean fluctuation Vs of the data distribution of the ith window of the data to be detected is calculatediSum-difference fluctuation DsiThe expressions are respectively as follows:
in the formula of UiRepresents the data in the ith window, U represents the entire raw power timing, and var and std represent the variance and standard deviation, respectively. max (U), min (U) represent the maximum and minimum values of the entire raw power sequence, respectively.
C) Calculating the distribution difference diff between the ith window and the previous windowi。
D) As shown in FIG. 2, the distribution difference diff obtained for the above equationiThe threshold value ∈ is set to 0.2. If diffiIf the value of (A) is less than or equal to epsilon, the data belong to the same data characteristic, and the width W of the sliding window is enlarged; if diffiIf the value of (d) is greater than epsilon, the window is in a turning period, the width W of the sliding window is reduced, and finally the purpose of self-adaptive window adjustment is achieved.
And the power abrupt change time interval is identified by adopting a self-adaptive window adjustment method so as to further extract and obtain an inflection point set, thereby greatly increasing the detection speed and improving the detection precision.
Step three, detecting inflection points in the trend by adopting a double-timing sliding window method, firstly, introducing alpha as one of criteria based on the formulated window adjustment strategy of local feature difference, and marking the position of a minimum value appearing in the mean value in two windows as the inflection point; and then improving the criterion of the traditional power mutation time interval, combining adjacent same-trend mutation time intervals, and completely extracting the turning weather mutation time interval.
The specific steps for extracting the inflection point set in the trend are as follows:
A) in the oscillation output time period, the situation of inflection point misjudgment can occur, the method of the invention introduces alpha as one of the criteria, namely the inflection point must meet the condition that alpha is 0;
B) on the basis of EMA curve, 2 closely-connected sliding windows are established, data in the 2 windows are updated frame by frame, the mean value of the sliding windows is taken as a reference for difference comparison, and inflection point detection score S is obtainedcThe calculation method is as follows:
in the formula: x1,iFor all power data in the previous window, X2,iThe full power data in the latter window. When the mean value difference in the 2 windows reaches the minimum, the power value at the mark combining point is an inflection point, the steps are repeated, and finally the corresponding time sequence trend inflection point set T is obtainedip。
Then, the traditional power period criterion is improved, and the improved mutation period criterion is as follows:
in the formula (I), the compound is shown in the specification,
set T for inflection pointipThe power value of the j point;
set T for inflection pointipThe power value of the j +1 th point;
set T for inflection pointipPassing through the jth point moment;
set T for inflection pointipThe moment of passing through the j +1 th point; lambda is a break amplitude threshold value in the turning period; beta is a turning period mutation rate threshold; not only the change in the amplitude of the mutation period but also the mutation rate is taken into account, thereby excluding the presence of a false inflection point.
And finally, combining adjacent mutation time periods with the same trend, and completely extracting the power turning time period.
And step four, dividing the time sequence into a turning section and a gentle section according to the division basis of the self-adaptive turning time period extraction result, adopting point prediction based on a GRU algorithm for the gentle section, and adopting a time sequence mode-self-adaptive bandwidth kernel density estimation method for the turning section to carry out probability prediction.
Inputting the time sequence characteristics extracted by the GRU network into an improved Attention mechanism combined with a CRS algorithm, endowing different weights to the transition characteristic vectors of the neural network model, then inputting the transition characteristic vectors subjected to weight management into the GRU layer according to time steps, outputting the training result of the improved GRU neural network, reading a training loss curve and an error curve, observing the longitudinal distance between the training set and a verification set loss curve in the convergence process, and visually evaluating the convergence performance of the network prediction result by combining the absolute error conditions of the training set and the verification set.
The convergence represented by the three common fitting states is as follows:
1) when the loss curve of the training set is almost not reduced, the state is an under-fitting state and is a non-convergence state;
2) when the loss curve of the training set continuously decreases, the loss curve of the verification set does not decrease until a certain moment, and the loss curve is in an overfitting state and in a convergence state but in imperfect convergence;
3) and when the loss curves of the training set and the verification set have no obvious interval, the state is a perfect fitting state and is perfect convergence.
Further, the invention improves and optimizes the traditional GRU model, combines with the Attention mechanism of CRS algorithm, and has the following specific contents:
because the number of the features of the input model is large, in order to highlight more key influence factors and help the model to make more accurate judgment, the invention provides an improved Attention mechanism, and transition feature vectors of a neural network model are endowed with different weights. In the conventional Attention mechanism, information carried by content input to a network firstly is covered by information input later, and a semantic vector may not completely represent information of a whole sequence. Therefore, in view of the above disadvantages, the present invention provides an Improved authorization mechanism (Improved authorization mechanism) combined with crs (comprehensive random search) algorithm, which makes up the deficiency of the network concerning different related factors on the same time scale, and improves the degree of Attention of the network concerning various related factors.
The CRS is used to generate the optimal combination of parameters in the attention layer. The operation process of the CRS is described in fig. 3, and the CRS is composed of four parts of "I, II, III, and IV".
"I" provides the weight W of the attention layer; then through the conversion in "II" into binary code, subset WiFor attention weighting, is transmitted to the GRU neural network where a corresponding penalty value is generated based on the prediction error in the network. Then, according to W in "IIIBTo select the optimal attention weight subset Wi BAnd
and iteratively loops through combinations of subsets thereof. Finally, a new attention weight is reconstructed in "IV
CRS detailed procedure as follows:
1) an attention weight set of length M-n (n is the model input feature dimension) is randomly generated,
2) will be a subset WiInputting attention layer, converting W into binary code:
3) according to the truth value y and the predicted value of the GRU model
To calculate the prediction error:
4) selecting an optimal attention weight subset W based on error feedbacki BAnd
each subset is composed of a binary string and is evenly divided into n segments. Accordingly, Wi BAnd
from Wi B＝(Fi 1,Fi 2,...Fi n) And
and (4) showing. Fi 1And
are respectively Wi BAnd
a part of (a).
5) Randomly decimating the portion Wi BAnd
for example, the n-1 th section F of both are selectedi n-1And
however, the number of segments selected is not fixed.
6) Obtaining Fi n-1And
the genetic recombination of (3). Fi n-1And
represented by binary code with length of 6, and randomly exchanging the binary code with the binary code on corresponding 6 indexes to obtain a recombined segment
7) Simulates a gene mutation and reverses
The genotype of (a). For example, 0 is inverted to 1. Then the
Substituted at Wi BOf (5) to (F)i n-1Form a new
Is inserted into WBIn (1).
8)WBDecoded to obtain an updated set of attention weights: w '(W)'1,W′2,...,W′k,...,W′M)。
9) And repeating the steps 2) to 8) until the preset time number K is reached.
Further, the mathematical expression of the GRU neural network is:
zt＝σ(Wz·[ht-1,xt])
rt＝σ(Wr·[ht-1,xt])
in the formula: z is a radical oftUpdate door rtReset gate, XtIn order to be the current input,
summary of input and past hidden layer states, htFor hidden layer output, Wz，WrAnd W is a trainable parameter matrix.
Step seven, aiming at the time-allowed mutation period, the probability prediction method by adopting the time sequence mode-self-adaptive bandwidth kernel density estimation method comprises the following specific steps:
1) dividing a time sequence mode into five modes of violent rising, violent falling, slow rising, slow falling, oscillation and the like based on the power time interval characteristics;
2) an empirical distribution estimation method is adopted to establish a time sequence mode-wind power prediction error probability density distribution model of each category under different weather types, and the specific content is as follows:
21) assume that there is a set of m-capacity sample observations X for the global variable X1,x2,…xp…xmRearranging the order of the1′,x2′,…xp′…xm' for any real number x, the empirical distribution is expressed as:
22) and calculating power prediction error probability distribution according to corresponding time sequence mode characteristics, and solving a box line graph based on the error distribution of various meteorological modes. For example, as shown in fig. 4, the left and right boundaries of the square are respectively the 50% quantile corresponding point locations, the central position scale mark is the median of the error of the pattern, the external points are outliers (outliers), and the graph visually represents the power prediction error distribution under each time sequence pattern;
23) derived from empirical distributionOriginal prediction error probability distribution ferrorEstimate calculation feThe formula for the core density is as follows:
wherein N is the number of samples of a set of data; h is the window width, also called the smoothing parameter; k (u) is a kernel function, u ═ h-1(e-ei)；eiIs the i-th sample value in the prediction error data.
24) Using Gaussian kernel function substitution estimation
In the expression, the specific expression of the gaussian kernel function is as follows:
substituting into the expression:
25) firstly, the integral mean square error is introduced to judge the probability density function obtained by estimation
And the true probability density function fe(x) The difference between the two is as follows:
wherein
MISE is expressed as the sum of the main terms AMISE, and when h → 0, nh → ∞, the AMISE expression is defined as:
where AMISE is an expression for window width h that reaches an optimal value of h when AMISE takes a minimum value, i.e.
The optimal window width h can be obtainedxThe expression is as follows:
3) and respectively fitting the probability density distribution curves under each time sequence characteristic, and providing a theoretical basis for presenting interval prediction results.
The traditional interval prediction only considers the historical power level to classify the errors, and the physical dynamic process considering the time sequence mode classification can improve the prediction accuracy of the interval prediction. The short-term interval prediction method based on time sequence mode classification and the Monte Carlo method has the best effect, and the interval coverage rate larger than the preset confidence level can be obtained under different confidence levels.
And step eight, establishing a point prediction-probability prediction sectional prediction model, and quantitatively analyzing the result precision of full-time point prediction, probability prediction, sectional prediction point prediction and probability prediction.
The method considers the whole-time wind power prediction including the turning weather. Firstly, a power time sequence trend discrimination method based on moving average iteration is provided, an EMA curve is subjected to smoothing processing by adopting a Gaussian window method, and a turning weather trend is extracted; secondly, a window adjusting strategy based on local time sequence characteristics is provided for adaptively adjusting the time window width, an inflection point detection strategy of a double-timing sliding window is adopted, an inflection point set is obtained by combining criteria, and a turning weather sudden change time period is extracted and divided. The proposed time sequence subsection prediction algorithm of point prediction and probability interval prediction carries out power prediction on the whole time period: adopting improved GRU algorithm point prediction for a turning section time sequence, providing an improved Attention mechanism combined with a CRS algorithm, and endowing different weights to a neural network model transition characteristic vector; probability prediction is carried out on the gentle section time sequence, an empirical distribution estimation method is adopted to establish a time sequence mode-power prediction error probability density distribution model, and wind power probability prediction is carried out on the basis of a variable bandwidth kernel density estimation method. The model is fit for typical characteristics of catastrophe weather, and the generalization capability and prediction performance of full-time power prediction are obviously improved.
While the invention has been described with reference to specific embodiments, the invention is not limited thereto, and those skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope of the invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (10)
1. An ultra-short-term wind power segmented prediction method based on turning period identification is characterized by comprising the following steps:
1) extracting a time sequence trend, solving an EMA curve representing the short-term development trend of the original wind power data of the wind power plant, and solving the change rate alpha at each moment after smoothing by adopting a Gaussian window method;
2) based on the EMA curve obtained in the step 1), utilizing the local characteristic difference of the EMA curve to make a window adjustment strategy, setting a detection threshold value epsilon, if the distribution difference fluctuation between a window and the previous window is smaller than the detection threshold value epsilon, expanding the window width to accelerate the detection speed, and otherwise, reducing the window width to improve the detection precision;
3) based on the window adjustment strategy of the local feature difference formulated in the step 2), marking the position of the minimum value of the mean value in the two windows as an inflection point by using alpha obtained in the step 1) as one of the criteria;
4) improving the criterion of the traditional power mutation time interval, combining adjacent same-trend mutation time intervals, and completely extracting the turning weather mutation time interval;
5) dividing the time sequence into a turning section and a gentle section according to the self-adaptive turning time period extraction result as a dividing basis;
6) point prediction is adopted for the gentle section, GRU is adopted as an original algorithm of the point prediction, an improved Attention mechanism combined with a CRS algorithm is introduced, different weights are given to transition characteristic vectors of a neural network model, Attention weights are transmitted to a GRU layer, training results of the GRU neural network are output, training loss curves and error curves are read, longitudinal distances between the training set and the verification set loss curves in a convergence process are observed, and the convergence performance of the network prediction results is visually evaluated in combination with the absolute error conditions of the training set and the verification set;
7) adopting probability prediction for the turning section by adopting a time sequence mode-self-adaptive bandwidth kernel density estimation method;
8) and (5) combining the step 6) and the step 7) to complete ultra-short-term wind power prediction based on the turning time period, and acquiring predicted power.
2. The ultra-short-term wind power segment prediction method based on turning period identification as claimed in claim 1, wherein in step 1), the time sequence trend is extracted by using a moving averaging method.
3. The ultra-short-term wind power segment prediction method based on turning period identification as claimed in claim 1, wherein in the step 2), the specific step of utilizing the EMA curve local feature difference to make the window adjustment strategy comprises:
21) the original power time sequence is divided into a plurality of segments, and turning point detection is carried out on each segment;
22) in the inflection Point detection, diff is definediFor measuring the distribution difference fluctuation of the ith window and the previous window, record
Wherein VsiMean fluctuation Ds of data distribution of the ith window of the data to be detectediIs the difference fluctuation;
23) setting a threshold value ε, if diffiIf the value of (d) is less than or equal to the threshold value epsilon, the sliding window width W is enlarged, and the detection speed is increased; if it isdiffiIf the value of (d) is greater than epsilon, the sliding window width W is reduced, and the detection accuracy is improved.
4. An ultra-short-term wind power segment prediction method based on turning period identification as claimed in claim 1, wherein in step 3), a dual timing sliding window is adopted for inflection point detection.
5. The ultra-short-term wind power subsection prediction method based on turning period identification as claimed in claim 4, wherein the specific contents of the inflection point detection by adopting the dual timing sliding window are as follows:
firstly, introducing the change rate alpha of each moment obtained in the step 1) as one of criteria, and setting an inflection point to meet the condition that alpha is 0; based on an EMA curve, establishing two sliding windows which are closely connected, updating data in the two windows frame by frame, marking a power value at a joint point when the mean value difference in the two windows reaches the minimum as an inflection point, repeating the steps, and obtaining a corresponding time sequence trend inflection point set Tip。
6. An ultra-short-term wind power subsection prediction method based on turning period identification as claimed in claim 5, wherein in step 4), the expression of the improved abrupt change period criterion is as follows:
in the formula (I), the compound is shown in the specification,
set T for inflection pointipThe power value of the j point;
set T for inflection pointipThe power value of the j +1 th point;
set T for inflection pointipPassing through the jth point moment;
set T for inflection pointipThe moment of passing through the j +1 th point; lambda is a break amplitude threshold value in the turning period; beta is a turning period mutation rate threshold; and combining adjacent mutation periods with the same trend to completely extract the turning period.
7. The ultra-short-term wind power segment prediction method based on turning period identification as claimed in claim 1, wherein in step 6), the mathematical expression of the GRU neural network is as follows:
zt＝σ(Wz·[ht-1,xt])
rt＝σ(Wr·[ht-1,xt])
in the formula: z is a radical oftTo refresh the door, rtTo reset the gate, XtIn order to be the current input,
for input and past summary of hidden states, htFor hidden layer output, Wz，WrAnd W is a trainable parameter matrix.
8. The ultra-short term wind power segment prediction method based on turning period identification as claimed in claim 7, wherein the specific step of introducing an improved Attention mechanism combined with CRS algorithm to assign different weights to the transition feature vectors of the neural network model in step 6) comprises:
61) providing the weight W of the attention layer, wherein the specific calculation steps of the weight are as follows:
611) calculating similarity M (Q, K) for a given task query vector Q and an attention variable K;
612) performing Softmax operation on the obtained similarity, and normalizing to obtain normalized similarity etai：
613) Carrying out weighted summation on all obtained weights according to the calculated weights to obtain an Attention vector;
62) converting the provided weights W of the attention layer into binary code WBSubset WiTransmitting the subset to a GRU neural network for attention weighting, and generating a corresponding loss value in the GRU neural network according to a prediction error in the network;
63) according to WBSelecting the optimal attention weight subset W for the loss case ofi BAnd
and repeatedly circulating the subset combination;
64) reconstructing a new attention weight
9. The ultra-short-term wind power segment prediction method based on turning period identification as claimed in claim 1, wherein in step 7), the concrete step of probability prediction by adopting a time sequence mode-adaptive bandwidth kernel density estimation method comprises:
71) dividing a time sequence mode into five types of violent rising, violent falling, slow rising, slow falling and oscillation based on the power time interval characteristics;
72) and establishing a time sequence mode-wind power prediction error probability density distribution model of each category under different weather type conditions by adopting an empirical distribution estimation method.
10. The ultra-short-term wind power sectional prediction method based on turning period identification as claimed in claim 9, wherein in step 71), the weather types are divided based on α as a division basis, the power prediction error probability density distribution under the corresponding time sequence mode characteristics is calculated, and the distribution is visually reflected through a box line graph; and (4) solving the optimal window width by utilizing a progressive integral mean square error method, substituting the optimal window width into an estimation function, respectively fitting a probability density distribution curve under each time sequence characteristic, and providing a basis for presenting an interval prediction result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111609901.7A CN114386324A (en) | 2021-12-27 | 2021-12-27 | Ultra-short-term wind power segmented prediction method based on turning period identification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111609901.7A CN114386324A (en) | 2021-12-27 | 2021-12-27 | Ultra-short-term wind power segmented prediction method based on turning period identification |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114386324A true CN114386324A (en) | 2022-04-22 |
Family
ID=81197461
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111609901.7A Pending CN114386324A (en) | 2021-12-27 | 2021-12-27 | Ultra-short-term wind power segmented prediction method based on turning period identification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114386324A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114898810A (en) * | 2022-05-20 | 2022-08-12 | 厦门大学 | Method for detecting growth condition of microorganism |
| CN115078909A (en) * | 2022-06-16 | 2022-09-20 | 国网江苏省电力有限公司宿迁供电分公司 | Non-ideal pulse leading edge detection method and device based on energy inflection point |
| CN115442271A (en) * | 2022-08-29 | 2022-12-06 | 云南电网有限责任公司迪庆供电局 | Network performance index time sequence data anomaly detection method |
| CN116340796A (en) * | 2023-05-22 | 2023-06-27 | 平安科技（深圳）有限公司 | Time sequence data analysis method, device, equipment and storage medium |
| CN116881743A (en) * | 2023-09-06 | 2023-10-13 | 国网上海市电力公司 | Meteorological turning weather identification method and system for photovoltaic power prediction |
| CN117007979A (en) * | 2023-10-07 | 2023-11-07 | 深圳市众航物联网有限公司 | Power output power failure abnormality early warning method based on data driving |
| CN117272120A (en) * | 2023-11-21 | 2023-12-22 | 国能日新科技股份有限公司 | Multi-factor continuous real-time data prediction method and device based on data classification |
| CN117476136A (en) * | 2023-12-28 | 2024-01-30 | 山东松盛新材料有限公司 | High-purity carboxylate synthesis process parameter optimization method and system |
-
2021
- 2021-12-27 CN CN202111609901.7A patent/CN114386324A/en active Pending
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114898810A (en) * | 2022-05-20 | 2022-08-12 | 厦门大学 | Method for detecting growth condition of microorganism |
| CN115078909A (en) * | 2022-06-16 | 2022-09-20 | 国网江苏省电力有限公司宿迁供电分公司 | Non-ideal pulse leading edge detection method and device based on energy inflection point |
| CN115078909B (en) * | 2022-06-16 | 2024-03-15 | 国网江苏省电力有限公司宿迁供电分公司 | Non-ideal pulse leading edge detection method and device based on energy inflection point |
| CN115442271A (en) * | 2022-08-29 | 2022-12-06 | 云南电网有限责任公司迪庆供电局 | Network performance index time sequence data anomaly detection method |
| CN115442271B (en) * | 2022-08-29 | 2023-09-26 | 云南电网有限责任公司迪庆供电局 | Network performance index time sequence data anomaly detection method |
| CN116340796B (en) * | 2023-05-22 | 2023-12-22 | 平安科技（深圳）有限公司 | Time sequence data analysis method, device, equipment and storage medium |
| CN116340796A (en) * | 2023-05-22 | 2023-06-27 | 平安科技（深圳）有限公司 | Time sequence data analysis method, device, equipment and storage medium |
| CN116881743B (en) * | 2023-09-06 | 2024-01-19 | 国网上海市电力公司 | Meteorological turning weather identification method and system for photovoltaic power prediction |
| CN116881743A (en) * | 2023-09-06 | 2023-10-13 | 国网上海市电力公司 | Meteorological turning weather identification method and system for photovoltaic power prediction |
| CN117007979A (en) * | 2023-10-07 | 2023-11-07 | 深圳市众航物联网有限公司 | Power output power failure abnormality early warning method based on data driving |
| CN117272120A (en) * | 2023-11-21 | 2023-12-22 | 国能日新科技股份有限公司 | Multi-factor continuous real-time data prediction method and device based on data classification |
| CN117476136A (en) * | 2023-12-28 | 2024-01-30 | 山东松盛新材料有限公司 | High-purity carboxylate synthesis process parameter optimization method and system |
| CN117476136B (en) * | 2023-12-28 | 2024-03-15 | 山东松盛新材料有限公司 | High-purity carboxylate synthesis process parameter optimization method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114386324A (en) | Ultra-short-term wind power segmented prediction method based on turning period identification | |
| CN109754113B (en) | Load prediction method based on dynamic time warping and long-and-short time memory | |
| CN108564192B (en) | Short-term photovoltaic power prediction method based on meteorological factor weight similarity day | |
| CN111401755B (en) | Multi-new-energy output scene generation method, device and system based on Markov chain | |
| CN115688579A (en) | Basin multi-point water level prediction early warning method based on generation of countermeasure network | |
| CN113095550A (en) | Air quality prediction method based on variational recursive network and self-attention mechanism | |
| CN112329990A (en) | User power load prediction method based on LSTM-BP neural network | |
| CN114596726B (en) | Parking berth prediction method based on interpretable space-time attention mechanism | |
| CN116187835A (en) | Data-driven-based method and system for estimating theoretical line loss interval of transformer area | |
| CN114819340A (en) | Time sequence prediction method for daily load of natural gas | |
| CN117498319A (en) | Distributed photovoltaic short-term power prediction method based on fusion clustering and VQC-LSTM | |
| Wibawa et al. | Long Short-Term Memory to Predict Unique Visitors of an Electronic Journal | |
| CN116646927A (en) | Wind power prediction method based on segmented filtering and longitudinal and transverse clustering | |
| CN113884807B (en) | Power distribution network fault prediction method based on random forest and multi-layer architecture clustering | |
| CN115564155A (en) | Distributed wind turbine generator power prediction method and related equipment | |
| CN115905855A (en) | Improved meta-learning algorithm MG-copy | |
| CN117856248A (en) | Power load prediction method based on frequency decomposition and cluster correction | |
| CN117688362A (en) | Photovoltaic power interval prediction method and device based on multivariate data feature enhancement | |
| CN117521501A (en) | NSGA-based LSTM prediction energy power generation method and system | |
| CN117408424A (en) | Short-term traffic flow prediction method based on PSO-SVR-LSTM | |
| CN117096860A (en) | Overhead transmission line current-carrying capacity interval prediction method and equipment based on LSSVM model | |
| Chen et al. | Interval prediction of photovoltaic power using improved NARX network and density peak clustering based on kernel mahalanobis distance | |
| CN114219122A (en) | Ultra-short-term prediction method for generating power of small hydropower station based on grid meteorological data | |
| Chen | Brain Tumor Prediction with LSTM Method | |
| Song et al. | A multi-task spatio-temporal fusion network for offshore wind power ramp events forecasting |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |