CN115600725A - Wind power plant ultra-short term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS - Google Patents
Wind power plant ultra-short term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS Download PDFInfo
- Publication number
- CN115600725A CN115600725A CN202211103988.5A CN202211103988A CN115600725A CN 115600725 A CN115600725 A CN 115600725A CN 202211103988 A CN202211103988 A CN 202211103988A CN 115600725 A CN115600725 A CN 115600725A
- Authority
- CN
- China
- Prior art keywords
- data
- beats
- network model
- improved
- wind direction
- 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 52
- 241000282461 Canis lupus Species 0.000 claims abstract description 69
- 238000005457 optimization Methods 0.000 claims abstract description 43
- 230000007246 mechanism Effects 0.000 claims abstract description 6
- 238000001514 detection method Methods 0.000 claims abstract description 5
- 238000012549 training Methods 0.000 claims description 21
- 238000012545 processing Methods 0.000 claims description 16
- 238000010606 normalization Methods 0.000 claims description 14
- 241000160777 Hipparchia semele Species 0.000 claims description 4
- 230000008859 change Effects 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000011160 research Methods 0.000 abstract description 4
- 238000003062 neural network model Methods 0.000 abstract description 2
- 238000000354 decomposition reaction Methods 0.000 description 10
- 230000008569 process Effects 0.000 description 7
- 239000013598 vector Substances 0.000 description 6
- 230000006870 function Effects 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 3
- 238000012804 iterative process Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 230000002829 reductive effect Effects 0.000 description 3
- 230000003247 decreasing effect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000010248 power generation Methods 0.000 description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 241001028048 Nicola Species 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000005299 abrasion Methods 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 190000008236 carboplatin Chemical compound 0.000 description 1
- 229960004562 carboplatin Drugs 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000009191 jumping Effects 0.000 description 1
- 230000000670 limiting effect Effects 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000036961 partial effect Effects 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 239000000758 substrate Substances 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 230000002195 synergetic effect Effects 0.000 description 1
- 238000004642 transportation engineering Methods 0.000 description 1
- 230000017105 transposition Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Economics (AREA)
- General Physics & Mathematics (AREA)
- Human Resources & Organizations (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Strategic Management (AREA)
- Computing Systems (AREA)
- Tourism & Hospitality (AREA)
- Artificial Intelligence (AREA)
- General Business, Economics & Management (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Marketing (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Operations Research (AREA)
- Life Sciences & Earth Sciences (AREA)
- Primary Health Care (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Water Supply & Treatment (AREA)
- Public Health (AREA)
- Molecular Biology (AREA)
- Quality & Reliability (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a CEEMDAN-IGWO-N-BEATS-based wind power plant ultra-short term wind direction prediction method. In order to reduce the volatility of original wind direction data signals, the prediction method of the invention decomposes original wind direction sequence signals into a series of intrinsic mode components and a group of residual components by using a CEEMDAN algorithm; in order to avoid the gray wolf optimization algorithm from falling into a local optimal solution and improve the local search precision, a dynamic adjustment nonlinear convergence factor strategy is adopted, the convergence of the algorithm is improved by utilizing a disturbance factor selection mechanism based on stagnation detection and a boundary-crossing reset strategy based on an influence coefficient, the hyperparameter in an N-BEATS model is optimized, and then each component is predicted; and finally, combining the component prediction results to obtain a final prediction result. The prediction method is superior to the traditional neural network model in prediction precision, generalization and interpretability in the wind direction time sequence prediction research of the wind turbine generator, and can be effectively applied to prediction of the wind direction of the wind turbine generator in an actual scene.
Description
Technical Field
The invention relates to the field of yaw control, in particular to a CEEMDAN-IGWO-N-BEATS-based wind power plant ultra-short-term wind direction prediction method.
Background
As society rapidly develops, human needs for energy have increased, and non-renewable resources have been unable to sustain the increasing demand of human. Wind energy in renewable energy is used as a novel efficient and clean energy, and can solve the shortage of the demand of human on non-renewable resources, so that the vigorous development of wind power generation is an important way for solving the problem. In wind power generation, a yaw system is used as a key component of a wind turbine generator and is an actuating mechanism which can enable the wind turbine generator to finish wind operation quickly and efficiently and reduce power loss of the wind turbine generator. However, the yaw system often causes unit abrasion due to instability of wind direction, and the service life of the unit is affected. Therefore, the accuracy of the prediction result of the ultra-short-term wind direction has important significance on the normal operation of the whole wind turbine generator.
While the deep learning technology is rapidly developed, students at home and abroad carry out deep research in the field of ultra-short-term wind direction prediction. The main research method comprises the following steps: conventional signal decomposition methods such as Empirical Mode Decomposition (EMD), machine learning algorithms such as support vector machine (ARIMA), integrated Moving Average autoregression, deep learning methods such as Long-Short Term Memory network (LSTM), convolutional neural network, and the like. However, due to the characteristics of large fluctuation and poor stability of the ultra-short-term wind direction sequence, the wind direction change rule is difficult to grasp by adopting a single prediction model, so that the advantages of each model are integrated by a combined model for improving the wind direction prediction accuracy, and the aim of improving the prediction accuracy is finally fulfilled. The existing machine learning prediction model is correspondingly improved on the basis of the traditional method, but the prediction accuracy of the problems of large Data scale And complicated abnormal points of a blower Data Acquisition And monitoring Control System (SCADA) still needs to be improved.
Disclosure of Invention
Aiming at the problems of low prediction precision, poor generalization performance, poor interpretability and the like of a single wind direction prediction model of a wind turbine generator set at present, the invention provides a wind power plant ultra-short term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS, namely a wind power plant ultra-short term wind direction prediction method based on a neural network substrate extension analysis optimized by a fully adaptive noise set empirical mode decomposition (CEEMDAN) and an improved Grey wolf optimization algorithm (IGWO). The prediction method is based on an N-BEATS model constructed by a pure deep learning network, and the hyperparameters in the N-BEATS prediction model are optimized by using an Improved Grey Wolf optimization algorithm (IGWO), and the optimized parameters act on modal components decomposed by Complete set empirical mode decomposition (CEEMDAN), and the final prediction result is obtained in a prediction result superposition mode. The model solves the problems of short service life and the like of a unit caused by poor stability, serious noise influence, low prediction precision and the like of an original sequence, and can ensure normal and stable operation of the wind turbine.
The technical scheme for solving the technical problems comprises the following steps: a CEEMDAN-IGWO-N-BEATS-based wind power plant ultra-short term wind direction prediction method is characterized by comprising the following specific implementation steps:
the method comprises the following steps: selecting a wind direction historical data time sequence in a yaw system from fan equipment, and selecting data points according to fixed time intervals to obtain an original wind direction time sequence;
step two: decomposing the original wind direction time sequence by using a CEEMDAN method to obtain K groups of intrinsic modal components and a group of residual components;
step three: respectively normalizing K groups of intrinsic mode components and a group of residual components obtained in the step 2 by K +1 groups of component data according to not less than 70% of data before the time point, and taking the data as the input of a network model; the rest data corresponding to the group component data are used as reference values during training to obtain a training data set; the normalization processing formula is as follows:
in the formula, s i,std Normalized data, s, representing the ith time point of a certain set of component data i Raw data, s, representing the ith time point of the component data min Represents the minimum value, s, of the first 70% or more of the component amount data max Represents the maximum value of the first not less than 70% of the component amount data;
step four: designing improved N-BEATS network model
The basic structure of the improved N-BEATS network model is similar to that of a typical N-BEATS network model, and the improved N-BEATS network model consists of a plurality of layers of stacks, each layer of stacks consists of a plurality of blocks, forward prediction values of the blocks are superposed to obtain forward prediction output of one layer of stacks, and the forward prediction values of each layer of stacks are superposed to obtain a final forward prediction result; the full connection layer number of the FC Stack in the block of the improved N-BEATS network model is set to be m, and the rest part of the setting is the same as that in the typical N-BEATS network model;
step five: training an improved N-BEATS network model using a training data set in combination with an improved grayling optimization algorithm
Taking the number of stacks in the improved N-BEATS network model, the number of blocks in each layer of stacks and the number m of layers of full connection layers of an FC Stack part in each block as hyper-parameters to be substituted into an improved Hui wolf optimization algorithm for optimizing; setting the number of stacks, the number of blocks in each layer of stacks and the range of the number of layers of the fully-connected layer of the FC Stack part in each block as [1,6], [1,8], [128,512], taking the ranges as initial ranges, substituting the initial ranges into an improved wolf optimization algorithm, and initializing a population;
inputting the normalized data of not less than 70% of the time points in a group of component data in the training data set into an improved N-BEATS network model, predicting the data of the rest time points, and calculating the individual fitness by combining the reference values of the rest time points of the corresponding group of component data in the training data set; taking the root mean square error as the individual fitness of the improved grayling optimization algorithm:
where y represents a reference value for the component data,a predictor representing the component data;
the prediction data representing the output of the network model,show thatPredicted value, y, of component data obtained after denormalization min Represents the minimum value, y, in the network model input data before normalization processing max Representing the maximum value before normalization processing in the network model input data;
the calculation formula of the convergence factor a of the improved grayish wolf optimization algorithm is as follows:
in the formula, t max The maximum value of the iteration times is represented as a set value; t represents the current number of iterations;
the improved grey wolf optimization algorithm adopts a disturbance factor selection mechanism based on stagnation detection to simultaneously detect the fitness values of alpha, beta and delta wolfs, if the alpha, beta and delta wolfs are all stagnated, the maximum stagnation period of the population is set, and when the population is stagnated and does not exceed the maximum stagnation time of the population, a disturbance factor with a small step length is added; when the stagnation of the population exceeds the maximum stagnation period of the population, adding a disturbance factor with larger step jump, wherein the formula is as follows:
X new showing the individuals after the position change, X t Representing the current individual position of the wolf, ub is the maximum value upper bound, lb is the minimum value lower bound, and N (0, 1) is a random number which accords with standard normal distribution; mu is an adjustable parameter, where mu 1 The value is smaller, the value range is [0.01],μ 2 The value of (A) is a large value, and the value range is [0.1,0.5 ]](ii) a g represents a stall cycle, and Mg represents a maximum stall cycle;
preserving X added perturbation factor new The current grey wolf individual to-be-updated positionThe medium fitness is better as the next generation individual;
in the formula (24), X t+1 F (X) represents the fitness value of an individual for the next generation of population positions;
the improved grey wolf optimization algorithm adopts a border-crossing resetting strategy based on influence coefficients, and the respective influence coefficients are calculated through the fitness values of alpha, beta and delta wolf, and the formula is as follows:
in the above formula, I α 、I β 、I δ Respectively the respective influence coefficients of alpha, beta and delta wolf, f α 、f β 、f δ Individual fitness of alpha, beta and delta wolf respectively; mu is a variable parameter, set here to a smaller value, ranging from [0.01];
When the iteration times reach the set maximum iteration times, outputting corresponding super-parameter values, namely the number of stacks in the improved N-BEATS network model, the number of blocks in each layer of stacks and the number of layers of full-connection layers of an FC Stack part in each block, and obtaining the improved N-BEATS network model trained by the group of mass data;
step six: processing each group of component data in the training set according to the method in the fifth step to obtain an improved N-BEATS network model trained by each group of component data; respectively carrying out normalization processing on the K +1 component data obtained in the step 2 according to not less than 70% of data after the time point, and respectively inputting the data serving as the input of the network model into the improved N-BEATS network model trained by the corresponding component data to respectively obtain the prediction data output by each network model; performing inverse normalization processing on the prediction data output by each network model to obtain the prediction value of the data at the time point which is not more than 30% of the data of each group of component data; and superposing the predicted values of the data of the time points which are not more than 30% behind each group of component data to obtain the wind direction predicted data of the time points which are not more than 30% behind the original wind direction time sequence.
Compared with the prior art, the invention has the beneficial effects that: in order to reduce the volatility of original wind direction data signals, the prediction method utilizes a CEEMDAN algorithm to decompose original wind direction sequence signals into a series of eigenmode components and a group of residual components; in order to avoid the gray wolf optimization algorithm from being trapped in a local optimal solution and improve the local search precision, a dynamic adjustment nonlinear convergence factor strategy is adopted, the convergence of the algorithm is improved by utilizing a disturbance factor selection mechanism based on stagnation detection and a border-crossing resetting strategy based on an influence coefficient, the hyperparameter in an N-BEATS model is optimized, and then each component is predicted; and finally, combining the component prediction results to obtain a final prediction result. The prediction method is superior to the traditional neural network model in prediction precision, generalization and interpretability in the wind direction time sequence prediction research of the wind turbine generator, and can be effectively applied to prediction of the wind direction of the wind turbine generator in an actual scene.
Drawings
The accompanying drawings, which are included to provide a further understanding of the embodiments of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the principles of the invention.
FIG. 1 is a schematic diagram of the method for predicting the ultra-short term wind direction of the wind farm based on CEEMDAN-IGWO-N-BEATS.
FIG. 2 shows the result of decomposition of the original wind direction time series by the CEEMDAN method according to an embodiment.
FIG. 3 is a block (blocks) structure diagram of a typical N-BEATS network model.
Fig. 4 is a graph of the decreasing process of the convergence factor of the gray wolf optimization algorithm and the improved gray wolf optimization algorithm in the present invention in the iterative process.
FIG. 5 is a schematic step flow chart of an embodiment of a CEEMDAN-IGWO-N-BEATS-based wind farm ultra-short term wind direction prediction method.
FIG. 6 is a comparison graph of a wind direction prediction result of an embodiment of a wind power plant ultra-short term wind direction prediction method (CEEMDAN-IGWO-N-BEATS for short) based on CEEMDAN-IGWO-N-BEATS of the invention and a wind direction prediction result of a traditional N-BEATS method, an IGWO-N-BEATS method and a CEEMDAN-N-BEATS method.
Detailed Description
In order to more clearly illustrate the technical solution of the present invention, the following detailed description of the present invention is provided with reference to the accompanying drawings and examples. The embodiments and descriptions of the present invention are provided only for explaining the present invention and not for limiting the present invention.
The invention provides a CEEMDAN-IGWO-N-BEATS-based ultra-short-term wind direction prediction method for a wind power plant, which comprises the following specific implementation steps of:
the method comprises the following steps: selecting a wind direction historical data time sequence in a yaw system from fan equipment, and selecting data points according to 20s time intervals to obtain an original wind direction time sequence;
step two: decomposing the original wind direction time sequence by using a CEEMDAN method to obtain K groups of intrinsic modal components and a group of residual components; as shown in fig. 2, the time series decomposition of the original wind direction results in nine sets of eigenmode components and one set of residual values.
2.1CEEMDAN decomposition Gaussian white noise w that first follows a standard normal distribution i (t) adding the original signal x (t).
x i (t)=x(t)+w i (t),i=1,2,…,n (1)
Where n is the total amount of signal decomposed by CEEMDAN.
2.2 pairs of x i (t) EMD decomposition of the resulting n modal componentsPerforming ensemble averaging to obtain the first-order component of the signal
In the formula r 1 (t) is the first order residual,is a second order IMF component (denoted by the character IMF2 in the figure, and the other components and so on), ε 1 Expressed as a noise-related parameter, E m (. Cndot.) represents the mth order IMF component after EMD decomposition.
2.3 continue to add noise to the remaining components after CEEMDAN decomposes to the k-th order.
In the formula r k (t) is the calculated K (K =2,3, \8230;, K) order residual,is an IMF component of order k + 1.
2.4 repeat the above steps until the obtained residual function is a monotonic function (less than two extreme values) and the decomposition is not continued.
Where R (t) is the final decomposed residual signal and x (t) is the final original signal.
Step three: respectively normalizing K groups of intrinsic mode components and a group of residual components obtained in the step 2 by K +1 groups of component data according to not less than 70% (80% in the embodiment) of data before the time point, and taking the data as the input of a network model; and taking the rest data corresponding to the group component data as reference values during training to obtain a training data set. The normalization processing formula is as follows:
in the formula s i,std Normalized data, s, representing the ith time point of a certain set of component data i Raw data, s, representing the ith time point of the component data min Represents the minimum value, s, of the first 70% or more of the component amount data max Represents the maximum value of the first not less than 70% of the component amount data.
Step four: designing improved N-BEATS network model
The basic structure of the improved N-BEATS network model is similar to that of a typical N-BEATS network model, and all the improved N-BEATS network model are composed of a plurality of layers of stacks (stacks), each layer of Stack is composed of a plurality of blocks (blocks), forward predicted values of the blocks are superposed to obtain a forward predicted output of one layer of Stack, and the forward predicted values of each layer of Stack are superposed to obtain a final forward predicted result, the difference is that the structure of the blocks (blocks) in the typical N-BEATS network model is shown in FIG. 3, the FC Stack of the first part is a fully-connected layer structure of four layers, the fully-connected layer number of the FC Stack in the blocks of the improved N-BEATS network model is set to m, and the rest parts are set to be the same as that in the typical N-BEATS network model (the structure and principle of the typical N-BEATS network model can be referred to Boris N.Oreshkin, ditriri carboplatin, nicolas, yoshia BEATS network model, BEALS network model: 2019: update, supplementary, 12, 2019, and 9. Said method for predicting.
For the nth block of a k-th layer stack, x is input n Firstly, m full-connection layers of FC Stack are processed in sequence, and the output is h n (ii) a Output h n Then, the two linear projection layers (FC layer, i.e. full connection layer) are respectively processed to generate forward and backward prediction coefficients, which are respectivelyAndthe mathematical formula is as follows:
wherein LINEAR is a LINEAR projection layer, b represents backscast, i.e. backward prediction, and f represents forecast, i.e. forward prediction; h is n Denotes x n Output after FC Stack processing and base vectorOptimizing the accuracy of forward prediction and backward prediction to obtain backward prediction (i.e. backtracking) output values respectivelyAnd forward predicted (i.e. forecast) output values
Base vectorIs a constraint on the function space and is the reason that the N-BEATS model is more interpretable.
One block gets two outputs, a backoff output and a forecast output, the forecast outputs of each block are finally directly added to be used as the output of a stack, the original input of each block is subtracted from the backoff output, and the obtained result is used as the input of the next block. All the stacks sum the outputs of the forecast part and generate the final output, and the output of the backscast enters the next block with stacks for fitting as abstract information of historical data.
The N-BEATS network model is based on forward and backward residual error chaining and a very deep full-link layer as a basic frame, a classical residual error network structure is used for reference, a layered double-residual error network topology structure is provided, the input of a block is subtracted by the backward prediction (i.e. backtracking) output of the block to obtain the residual error output of the block, the residual error output is used as the input of the next block, and for the (N + 1) th block of the kth stack, the input is as follows:
a specific polynomial is introduced for modeling the trend:
in the formula, p is a polynomial degree, and in this embodiment, k represents a kth layer stack, and n is an nth block of the kth layer stack; vector t = [0,1,2, \8230 ], N-2, N-1] T N, wherein N is the length of a prediction window, namely the prediction step length; t denotes transposition.
Step five: training an improved N-BEATS network model using a training data set in combination with an improved grayling optimization algorithm
And taking the number of stacks (stacks) in the improved N-BEATS network model, the number of blocks (blocks) in each layer of stacks and the number m of fully-connected layers of an FC Stack part in each block as hyper-parameters to carry out optimization by adopting an improved Husky optimization algorithm. The number of stacks (stacks), the number of blocks (blocks) in each Stack layer, and the number of layers of the fully-connected layer of the FC Stack part in each block are set to be in the ranges of [1,6], [1,8], [128,512], respectively, and the ranges are taken as initial ranges to be introduced into an improved Grey wolf optimization algorithm to initialize the population.
And inputting the normalized data of not less than 70% of the time points in a group of component data in the training data set into the improved N-BEATS network model, predicting the data of the rest time points, and calculating the individual fitness by combining the reference values of the rest time points of the corresponding group of component data in the training set. Root Mean Square Error (RMSE) is taken as the individual fitness of the improved graywolf optimization algorithm:
where y represents a reference value of the component data,representing the predicted values of the component data.
The prediction data representing the output of the network model,show thatPredicted value, y, of component data obtained after denormalization min Represents the minimum value, y, in the network model input data before normalization processing max Representing the maximum value in the network model input data before normalization processing.
Grey Wolf optimization Algorithm (GWO) is a global stochastic optimization Algorithm that is optimized by learning the predation behavior of Grey wolfs in nature. The wolf pack can be divided into 4 levels, and the levels are alpha, beta, delta and omega from high to low. The hunting process is as follows:
firstly, initializing a wolf population, and simultaneously initializing to obtain a convergence factor, A and C synergistic coefficient vectors:
in the iterative process, the value of a is linearly reduced from 2 to 0; vector r 1 、r 2 Is taken by die [0,1 ]]A random number in between. Calculating the individual fitness of the gray wolf, and reserving the alpha, beta and delta of the 3 gray wolf with the best fitness, wherein the optimization process of the gray wolf is mainly completed by the guidance of the alpha, beta and delta of each generation. Surrounding a prey:
D=|C·X p (t)-X(t)| (17)
X(t+1)=X p (t)-AD (18)
d is the distance between the grey wolf and the prey, X p For prey position, X represents the gray wolf position and t is the current number of iterations. Capturing a prey: based on the useful information of the target position obtained in the equations (17) and (18), the wolf group searches for the target prey under the guidance of α, β and δ wolfs. The mathematical model can be expressed as:
in the formula X α ,X β ,X δ Respectively representing the position vectors of alpha, beta and delta wolfs, D α ,D β ,D δ Respectively represent the current candidate wolf and alpha, beta and delta wolfThe position distance of | A | Liao>1 represents the disperse search of wolf, | A | does not count<1 represents the gray wolf searching intensively in some areas.Showing the updated position of the omega wolf after the wheel is caught.
When the game stops moving, the wolf completes the hunting process by attacking. In order to simulate approaching a prey, the value of a is gradually reduced, and thus the fluctuation range of A is also reduced. In other words, in an iterative process, as the value of a linearly decreases from 2 to 0, its corresponding value of A also changes within the interval [ -a, a ]. When the value of a is within the interval, the next position of the gray wolf can be located anywhere between its current position and the game position. When | a | <1, the wolf pack launches an attack (falls into local optima) to the prey. When | a | >1, the grayish wolf is separated from the game, and it is desired to find a more suitable game (global optimum).
The GWO algorithm has another component C to help find new solutions. As can be seen from the formula (16), C is a random value between [0,2 ]. C represents the random weight of the influence of the position of the wolf on the prey, C >1 represents that the influence weight is large, and conversely, the influence weight is small. This helps the GWO algorithm to behave more randomly and support exploration, while avoiding trapping in local optima during the optimization process. In addition, unlike a, C is non-linearly decreasing, so that it provides a global search in the decision space from the initial iteration to the final iteration. When the algorithm falls into local optimum and is not easy to jump out, the randomness of C plays a very important role in avoiding local optimum, especially in the iteration that finally needs to obtain a global optimum solution (the principles and processes of the Grey wolf optimization algorithm can be referenced by Wang Li Hui, yang Hui, wang Yintang, liu Yong, hu Qing Fang. Dangjiang mouth reservoir runoff forecasting based on GWO-LSTM. [ J ]. Water conservancy and Water transportation engineering report, 2021 (06): 51-59.).
The improved grey wolf optimization algorithm in the invention has the difference compared with the principle and the process of the grey wolf optimization algorithm, namely:
the calculation formula of the convergence factor a of the improved grayish wolf optimization algorithm is as follows:
in the formula, t max The maximum value of the iteration times is represented as a set value; t represents the current number of iterations, and the convergence factor is a nonlinear control convergence factor based on a cosine function.
The improved grey wolf optimization algorithm adopts a disturbance factor selection mechanism based on stagnation detection, alpha, beta and delta wolf fitness values are detected at the same time, if the alpha, the beta and the delta wolf are all stagnated (namely the fitness is unchanged), the guiding effect of the alpha, the beta and the delta wolf on a population is weakened, the maximum stagnation period of the population is set, and when the population is stagnated and does not exceed the maximum stagnation period of the population, a disturbance factor with a small step length is added, searching is carried out in a small range, and the convergence of the later algorithm is prevented from being influenced; when the stagnation of the population exceeds the maximum stagnation period of the population, which indicates that the smaller disturbance factor is not enough to enable the population to jump out of the local optimal solution, the disturbance factor with larger step jump is added, and the formula is as follows:
X new showing the individuals after the position change, X t Representing the current individual position of the wolf, ub is the maximum value upper bound, lb is the minimum value lower bound, and N (0, 1) is a random number which conforms to the standard normal distribution; mu is an adjustable parameter (for adjusting the fluctuation range of the disturbance factor, the disturbance factor is mu 2 (ub-lb). N (0, 1)), in this case μ 1 The value is smaller, the value range is [0.01],μ 2 The value of (A) is a large value, and the value range is [0.1,0.5 ]]. g denotes a stall period and Mg denotes a maximum stall period, where the maximum stall period Mg is set to 3.
Preserving X added perturbation factor new The current grey wolf individual to-be-updated positionThe medium fitness is better as the next generation individual.
In the formula (24), X t+1 For the next generation population position, f (X) represents the fitness value of the individual. The disturbance factors adopt random numbers distributed in a standard normal way to ensure that individuals in the population have the possibility of jumping in different directions and different step lengths, and variable parameters mu are utilized 1 、μ 2 And the jump step length is adjusted, so that the gray wolf optimization algorithm is prevented from being trapped into partial optimization in optimization.
The improved grayish optimization algorithm employs an out-of-range reset strategy based on an influence coefficient. The reset strategy of the original grayish wolf optimization algorithm is to place the boundary-crossing individuals at the boundary, which often causes the boundary-crossing individuals to be still at the boundary after the next update, so that the convergence speed of the algorithm in the early stage is limited, and the local optimization is achieved in the early stage. In order to fully utilize the guiding function of alpha, beta and delta wolf, an out-of-range resetting strategy based on influence coefficients is provided, the respective influence coefficients are calculated through the adaptability values of the alpha, beta and delta wolf, the smaller the adaptability value is, the stronger the individual leadership is, the higher the influence coefficient is, and the formula is as follows:
in the above formula, I α 、I β 、I δ The influence coefficients of alpha, beta and delta wolf respectively, f α 、f β 、f δ Individual fitness of alpha, beta and delta wolf respectively; mu is a variable parameter set to a smaller value here, ranging from 0.01]The reset position is guaranteed to be selected under alpha, beta and delta wolf directions, while a small range of jumps can be made.
And when the iteration times reach the set maximum iteration time, outputting corresponding over-parameter values, namely the number of stacks (stacks) in the improved N-BEATS network model, the number of blocks (blocks) in each layer of stacks and the number of layers of full connection layers of an FC Stack part in each block, so as to obtain the improved N-BEATS network model trained by the group of component data.
Step six: and (4) processing each group of component data in the training set according to the method in the fifth step to obtain an improved N-BEATS network model trained by each group of component data. Respectively normalizing the K +1 component quantity data obtained in the step 2 by not less than 70% (80% in the embodiment) of the data after the time point, and respectively inputting the data serving as the input of the network model into the improved N-BEATS network model which is trained by the corresponding component quantity data to respectively obtain the prediction data output by each network model; performing reverse normalization on the predicted data output by each network model to obtain the predicted value (marked as 100% in the figure) of the data of the time points (30% is the number percentage of the time points, if the time points of the input data of the network model are 70, the time points of the output predicted data are 30, and recursion is carried out, the input time points and the output time points of the network model are 100% in total) of the data of the time points of each group of component dataOr R); and superposing the predicted values of the data of the time points which are not more than 30% behind each group of component data to obtain the wind direction predicted data of the time points which are not more than 30% behind the original wind direction time sequence.
TABLE 1N-BEATS MODEL PARAMETERS TABLE FOR GROUP QUANTITY DATA
The final parameter optimization result of the improved grey wolf optimization algorithm is shown in table 1, and it can be seen from the table that each modal component corresponds to different optimal parameters respectively, and the optimal values are concentrated in the parameter values, which shows that the experimental parameter range is reasonable in setting.
Under the same experimental condition, the wind direction prediction results of the method are compared with the wind direction prediction results of the existing N-BEATS, LSTM, ARIMA, GWO-N-BEATS, IGWO-N-BEATS, PSO-N-BEATS, CEEMDAN-N-BEATS and EMD-N-BEATS methods, and the result shows that the prediction accuracy of the CEEMDAN-IGWO-N-BEATS method designed by the invention is obviously superior to that of the traditional machine learning model, superior to that of a single prediction model and superior to that of the CEEMDAN-N-BEATS and IGWO-N-BEATS combination mode. The evaluation index expression of the predicted result is as follows:
root Mean Square Error (RMSE):
mean Absolute Error (MAE):
mean Absolute Percentage Error (MAPE):
where y represents the true value of the wind direction data,the predicted value of wind direction data is expressed. The predicted results are shown in table 2.
TABLE 2 average value of prediction accuracy for each model
In FIG. 6, by comparing the wind direction prediction effects of the CEEMDAN-IGWO-N-BEATS method with those of the conventional N-BEATS method, IGWO-N-BEATS method, and CEEMDAN-N-BEATS method, it is found that the CEEMDAN-IGWO-N-BEATS model has a higher degree of fitting tendency at some data points with larger fluctuation.
Nothing in this specification is said to apply to the prior art.
Claims (4)
1. A method for predicting the ultra-short-term wind direction of a wind power plant based on CEEMDAN-IGWO-N-BEATS is characterized by comprising the following specific implementation steps of:
the method comprises the following steps: selecting a wind direction historical data time sequence in a yaw system from fan equipment, and selecting data points according to fixed time intervals to obtain an original wind direction time sequence;
step two: decomposing the original wind direction time sequence by using a CEEMDAN method to obtain K groups of intrinsic modal components and a group of residual components;
step three: respectively normalizing K groups of intrinsic mode components and a group of residual components obtained in the step 2 and K +1 groups of component data in front of the time point by not less than 70 percent to be used as the input of a network model; the rest data corresponding to the group component data are used as reference values during training to obtain a training data set; the normalization processing formula is as follows:
in the formula, s i,std Normalized data, s, representing the ith time point of a certain set of component data i Raw data, s, representing the ith time point of the component data min Represents the minimum value, s, of the first not less than 70% of the component amount data max Represents the maximum value of the first not less than 70% of the component amount data;
step four: designing improved N-BEATS network model
The basic structure of the improved N-BEATS network model is similar to that of a typical N-BEATS network model, and the improved N-BEATS network model consists of a plurality of layers of stacks, each layer of stacks consists of a plurality of blocks, forward prediction values of the blocks are superposed to obtain forward prediction output of one layer of stacks, and the forward prediction values of each layer of stacks are superposed to obtain a final forward prediction result; the full connection layer number of the FC Stack in the block of the improved N-BEATS network model is set to be m, and the rest part of the setting is the same as that in the typical N-BEATS network model;
step five: training an improved N-BEATS network model using a training data set in combination with an improved graying optimization algorithm
Taking the number of stacks in the improved N-BEATS network model, the number of blocks in each layer of stacks and the number m of fully-connected layers of an FC Stack part in each block as hyper-parameters to be brought into an improved wolf optimization algorithm for optimization; setting the number of stacks, the number of blocks in each layer of stacks and the range of the number of layers of the fully-connected layer of the FC Stack part in each block as [1,6], [1,8], [128,512], taking the ranges as initial ranges, substituting the initial ranges into an improved wolf optimization algorithm, and initializing a population;
inputting the normalized data of not less than 70% of the time points in a group of component data in the training data set into an improved N-BEATS network model, predicting the data of the rest time points, and calculating the individual fitness by combining the reference values of the rest time points of the corresponding group of component data in the training data set; taking the root mean square error as the individual fitness of the improved grayling optimization algorithm:
where y represents a reference value of the component data,a predictor representing the component data;
the prediction data representing the output of the network model,show thatPredicted value, y, of component data obtained after denormalization min Represents the minimum value, y, of the network model input data before normalization max Representing the maximum value of the network model input data before normalization processing;
the improved grey wolf optimization algorithm has the calculation formula of the convergence factor a as follows:
in the formula, t max Representing the maximum value of the iteration times as a set value; t represents the current number of iterations;
the improved grey wolf optimization algorithm adopts a disturbance factor selection mechanism based on stagnation detection to simultaneously detect the fitness values of alpha, beta and delta wolfs, if the alpha, beta and delta wolfs are all stagnated, the maximum stagnation period of the population is set, and when the population is stagnated and does not exceed the maximum stagnation time of the population, a disturbance factor with a small step length is added; when the stagnation of the population exceeds the maximum stagnation period of the population, adding a disturbance factor with larger step jump, wherein the formula is as follows:
X new showing the individuals after the position change, X t Representing the current individual position of the wolf, ub is the maximum value upper bound, lb is the minimum value lower bound, and N (0, 1) is a random number which conforms to the standard normal distribution; mu is an adjustable parameter, where mu 1 The value is smaller, the value range is [0.01],μ 2 The value of (A) is a large value, and the value range is [0.1,0.5 ]](ii) a g represents a stall cycle, and Mg represents a maximum stall cycle;
preserving X added perturbation factor new The current grey wolf individual to-be-updated positionThe medium fitness is better as the next generation individual;
in the formula (24), X t+1 F (X) represents the fitness value of an individual for the next generation of population positions;
the improved grey wolf optimization algorithm adopts a border-crossing resetting strategy based on influence coefficients, and the respective influence coefficients are calculated through the fitness values of alpha, beta and delta wolf, and the formula is as follows:
in the above formula, I α 、I β 、I δ The influence coefficients of alpha, beta and delta wolf respectively, f α 、f β 、f δ Individual fitness of alpha, beta and delta wolf respectively; mu is a variable parameter, set here to a smaller value, ranging from [0.01];
When the iteration times reach the set maximum iteration times, outputting corresponding super-parameter values, namely the number of stacks in the improved N-BEATS network model, the number of blocks in each layer of stacks and the number of layers of full-connection layers of an FC Stack part in each block, and obtaining the improved N-BEATS network model trained by the group of mass data;
step six: processing each group of component data in the training set according to the method in the fifth step to obtain an improved N-BEATS network model trained by each group of component data; respectively carrying out normalization processing on the K +1 component data obtained in the step 2 according to not less than 70% of data after the time point, and respectively inputting the data serving as the input of the network model into the improved N-BEATS network model trained by the corresponding component data to respectively obtain the prediction data output by each network model; performing inverse normalization processing on the prediction data output by each network model to obtain the prediction value of the data at the time point which is not more than 30% behind each group of component data; and superposing the predicted values of the data of the time points which are not more than 30% behind each group of component data to obtain the wind direction predicted data of the time points which are not more than 30% behind the original wind direction time sequence.
2. The CEEMDAN-IGWO-N-BEATS-based wind farm ultrashort-term wind direction prediction method according to claim 1, characterized in that the fixed time interval in the first step is 20s.
3. The method for predicting the ultra-short-term wind direction of the wind power plant based on the CEEMDAN-IGWO-N-BEATS as claimed in claim 1, wherein in the third step, each group of component data is normalized according to the data of the first 80% of the time point.
4. The CEEMDAN-IGWO-N-BEATS-based wind farm ultra-short term wind direction prediction method according to claim 1, characterized in that in step five, the maximum stall period Mg is set to 3.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211103988.5A CN115600725A (en) | 2022-09-09 | 2022-09-09 | Wind power plant ultra-short term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211103988.5A CN115600725A (en) | 2022-09-09 | 2022-09-09 | Wind power plant ultra-short term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115600725A true CN115600725A (en) | 2023-01-13 |
Family
ID=84842852
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211103988.5A Pending CN115600725A (en) | 2022-09-09 | 2022-09-09 | Wind power plant ultra-short term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115600725A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117471502A (en) * | 2023-10-31 | 2024-01-30 | 北京华云星地通科技有限公司 | Positioning source parameter anomaly detection and correction method, system and electronic equipment |
-
2022
- 2022-09-09 CN CN202211103988.5A patent/CN115600725A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117471502A (en) * | 2023-10-31 | 2024-01-30 | 北京华云星地通科技有限公司 | Positioning source parameter anomaly detection and correction method, system and electronic equipment |
CN117471502B (en) * | 2023-10-31 | 2024-04-02 | 北京华云星地通科技有限公司 | Positioning source parameter anomaly detection and correction method, system and electronic equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wang et al. | Wind power forecasting based on singular spectrum analysis and a new hybrid Laguerre neural network | |
Li et al. | Smart wind speed forecasting approach using various boosting algorithms, big multi-step forecasting strategy | |
Al-qaness et al. | Boosted ANFIS model using augmented marine predator algorithm with mutation operators for wind power forecasting | |
CN110059844B (en) | Energy storage device control method based on ensemble empirical mode decomposition and LSTM | |
CN112733462B (en) | Ultra-short-term wind power plant power prediction method combining meteorological factors | |
CN108365608B (en) | Uncertain optimization scheduling method and system for regional energy Internet | |
CN109615146B (en) | Ultra-short-term wind power prediction method based on deep learning | |
Wang et al. | A hybrid forecasting model based on CNN and informer for short-term wind power | |
CN115511177A (en) | Ultra-short-term wind speed prediction method based on INGO-SWGMN hybrid model | |
CN115275991A (en) | Active power distribution network operation situation prediction method based on IEMD-TA-LSTM model | |
CN115600725A (en) | Wind power plant ultra-short term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS | |
CN114676645B (en) | Non-stationary time sequence prediction method and system | |
CN112329339A (en) | Short-term wind speed prediction method for wind power plant | |
CN115310669A (en) | Ultra-short-term wind power prediction method based on secondary decomposition and IWOA-LSSVM | |
CN112819238A (en) | Short-term wind power prediction method based on chaotic chicken flock optimization algorithm | |
CN115660165A (en) | Modular neural network effluent ammonia nitrogen concentration multi-step prediction method based on double-layer PSO | |
CN116960978A (en) | Offshore wind power prediction method based on wind speed-power combination decomposition reconstruction | |
CN116822743A (en) | Wind power prediction method based on two-stage decomposition reconstruction and error correction | |
Yu et al. | Reinforcement learning-based multi-objective differential evolution for wind farm layout optimization | |
CN114444755A (en) | Wind power short-term power prediction method and device and readable storage medium | |
CN110674460B (en) | E-Seq2Seq technology-based data driving type unit combination intelligent decision method | |
CN112865170A (en) | Scene probability-based load recovery optimization method considering wind power output correlation | |
Tan et al. | New dynamical optimal learning for linear multilayer FNN | |
CN116681154A (en) | Photovoltaic power calculation method based on EMD-AO-DELM | |
Zhao et al. | Cuckoo search algorithm with interactive learning for economic dispatch |
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 |