A wind farm ultra-short-term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS

Abstract

The application discloses a wind direction super-short-term prediction method for a wind farm based on CEEMDAN-IGWO-N-BEATS. In order to reduce the volatility of the original wind direction data signal, the prediction method decomposes the original wind direction sequence signal into a series of intrinsic mode components and a group of residual components by using the CEEMDAN algorithm. In order to avoid the grey 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, a disturbance factor selection mechanism based on stagnation detection and a boundary reset strategy based on an influence coefficient are used to improve the algorithm convergence, and the hyperparameters in the N-BEATS model are optimized, and then each component is predicted. Finally, the final prediction result is obtained by combining the prediction results of each component. The prediction method is superior to the traditional neural network model in terms of prediction accuracy, generalization and interpretability in the research on wind direction time series prediction of wind turbines, and can be effectively applied to the prediction of wind direction of wind turbines in actual scenes.

Description

Technical Field

[0001] This invention relates to the field of yaw control, specifically a method for ultra-short-term wind direction prediction in wind farms based on CEEMDAN-IGWO-N-BEATS. Background Technology

[0002] With rapid societal development, human demand for energy is increasing daily, and non-renewable resources can no longer sustain this growing demand. Wind energy, as a highly efficient and clean new type of renewable energy, can address the shortage of non-renewable resources, making its vigorous development a crucial solution. In wind power generation, the yaw system, a key component of the wind turbine, is the actuator that enables the turbine to quickly and efficiently operate in the wind, reducing power loss. However, the yaw system is often susceptible to wear and tear due to wind instability, affecting the turbine's lifespan. Therefore, the accuracy of ultra-short-term wind direction prediction is of paramount importance for the normal operation of the entire wind turbine.

[0003] While deep learning technology has developed rapidly, scholars both domestically and internationally have conducted in-depth research in the field of ultra-short-term wind direction forecasting. The main research methods include: traditional signal decomposition methods such as Empirical Mode Decomposition (EMD), machine learning algorithms such as Support Vector Machines (Autoregressive Integrated Moving Average Model, ARIMA), and deep learning methods such as Long Short-Term Memory (LSTM) networks and convolutional neural networks. However, due to the large fluctuations and poor stability of ultra-short-term wind direction sequences, it is difficult to grasp the wind direction change patterns using a single prediction model. Therefore, to improve the accuracy of wind direction forecasting, combined models are used to integrate the advantages of various models, ultimately achieving the goal of improving prediction accuracy. Existing machine learning prediction models have all made corresponding improvements to traditional methods, but for the large data scale and complex anomalies of Supervisory Control and Data Acquisition (SCADA) systems, the prediction accuracy still needs to be improved. Summary of the Invention

[0004] To address the problems of low prediction accuracy, poor generalization performance, and poor interpretability of current single-model wind direction prediction for wind turbines, this invention proposes an ultra-short-term wind direction prediction method for wind farms based on CEEMDAN-IGWO-N-BEATS. Specifically, it's an interpretable time series analysis method (N-BEATS) based on neural network basis extension analysis optimized by Completely Adaptive Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and an improved Grey Wolf Optimizer (IGWO) algorithm. This prediction method is based on an N-BEATS model constructed using a pure deep learning network. The hyperparameters in the N-BEATS prediction model are optimized using the improved Grey Wolf Optimizer (IGWO) algorithm, which is applied to each modal component decomposed by Complete Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). The final prediction result is obtained by superimposing the prediction results. This model solves the problems of poor stability of the original sequence, severe noise influence, low prediction accuracy, and short turbine lifespan, ensuring the normal and stable operation of wind turbines.

[0005] The technical solution of this invention to solve the aforementioned technical problem is: a wind farm ultra-short-term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS, characterized in that the specific implementation steps of the prediction method are as follows:

[0006] Step 1: Select historical wind direction data time series from the yaw system of the wind turbine equipment, select data points at fixed time intervals to obtain the original wind direction time series;

[0007] Step 2: Decompose the original wind direction time series using the CEEMDAN method to obtain K sets of intrinsic mode components and one set of residual components;

[0008] Step 3: For the K groups of intrinsic mode components and one group of residual components obtained in Step 2, totaling K+1 components, normalize the data based on the first 70% of each time point, and use this normalized data as input to the network model. The remaining data for each component are used as reference values ​​during training, resulting in the training dataset. The normalization formula is:

[0009]

[0010] In the formula, s i,std s represents the normalized data at the i-th time point of a certain component quantity data. i This represents the raw data of the component quantity data at the i-th time point, s min This represents the minimum value among the first 70% of the component quantity data, s maxThis indicates the maximum value among at least 70% of the component's quantity data;

[0011] Step 4: Design an improved N-BEATS network model

[0012] The basic structure of the improved N-BEATS network model is similar to that of the typical N-BEATS network model, both consisting of multiple stacks. Each stack consists of multiple blocks, and the forward prediction values ​​of multiple blocks are superimposed to obtain the forward prediction output of one stack. The forward prediction values ​​of each stack are superimposed to obtain the final forward prediction result. The number of fully connected layers in the FC Stack of the blocks in the improved N-BEATS network model is set to m, and the settings of the rest are the same as those in the typical N-BEATS network model.

[0013] Step 5: Train the improved N-BEATS network model using the training dataset and the improved gray wolf optimization algorithm.

[0014] The number of stacks, the number of blocks in each stack, and the number of fully connected layers m in the FC Stack of each block in the improved N-BEATS network model are used as hyperparameters in the improved Grey Wolf optimization algorithm for optimization. The ranges of the number of stacks, the number of blocks in each stack, and the number of fully connected layers m in the FC Stack of each block are set to [1,6], [1,8], and [128,512], respectively. These ranges are used as the initial ranges in the improved Grey Wolf optimization algorithm to initialize the population.

[0015] Normalized data from at least 70% of the time points in a set of component data in the training dataset are input into the improved N-BEATS network model to predict the data at the remaining time points. The individual fitness is calculated by combining the reference values ​​of the remaining time points of the corresponding component data in the training set. The root mean square error is used as the individual fitness of the improved Grey Wolf optimization algorithm.

[0016]

[0017] In the formula, y represents the reference value of the component data. This represents the predicted value of the component data;

[0018]

[0019] This represents the predicted data output by the network model. Indicates will The predicted values ​​of the component data obtained after inverse normalization, y min y represents the minimum value in the input data of the network model before normalization. max This represents the maximum value in the input data of the network model before normalization.

[0020] The formula for calculating the convergence factor 'a' of the improved gray wolf optimization algorithm is as follows:

[0021]

[0022] In the formula, t max This represents the maximum number of iterations, which is a set value; t represents the current number of iterations.

[0023] The improved gray wolf optimization algorithm employs a stagnation detection-based perturbation factor selection mechanism. It simultaneously detects the fitness values ​​of α, β, and δ wolves. If all α, β, and δ wolves stagnate, a maximum stagnation period for the population is set. Furthermore, when the population stagnates but does not exceed the maximum stagnation period, a perturbation factor with a small step size is added; when the population stagnates beyond the maximum stagnation period, a perturbation factor with a large step size is added. The formula is as follows:

[0024]

[0025] X new X represents the individual after the position change. t This represents the current position of the individual gray wolf; ub is the upper bound of the maximum value; lb is the lower bound of the minimum value; N(0,1) is a random number that conforms to a standard normal distribution; μ is an adjustable parameter, where μ1 takes a smaller value, ranging from [0.01, 0.05], and μ2 takes a larger value, ranging from [0.1, 0.5]; g represents the stagnation period, and Mg represents the maximum stagnation period.

[0026] Retain X with added perturbation factor new Current location of individual gray wolves awaiting update Those with better fitness will be selected as the next generation of individuals;

[0027]

[0028] In formula (24), X t+1 f(X) represents the fitness value of an individual, which is the position of the next generation population.

[0029] The improved gray wolf optimization algorithm employs an out-of-bounds reset strategy based on influence coefficients. The influence coefficients are calculated using the fitness values ​​of α, β, and δ wolves, as shown in the following formula:

[0030]

[0031]

[0032]

[0033]

[0034] In the above formula, I α I β I δ These are the influence coefficients of α, β, and δ wolves, respectively, f α f β f δ α, β, and δ represent the individual fitness of wolves, respectively; μ is a variable parameter, set to a small value here, ranging from [0.01, 0.05].

[0035] When the number of iterations reaches the set maximum number of iterations, the corresponding hyperparameter values ​​are output, namely the number of stacks in the improved N-BEATS network model, the number of blocks in each stack, and the number of fully connected layers in the FC Stack part of each block, thus obtaining the improved N-BEATS network model trained with the sample data.

[0036] Step Six: Process each component data in the training set according to the method described in Step Five to obtain an improved N-BEATS network model trained on each component data; normalize the K+1 component data obtained in Step 2 according to at least 70% of the data after each time point, and use it as input to the network model, respectively inputting it into the improved N-BEATS network model trained on the corresponding component data to obtain the predicted data output by each network model; perform inverse normalization on the predicted data output by each network model to obtain the predicted values ​​of the data at no more than 30% of the time points after each component data; superimpose the predicted values ​​of the data at no more than 30% of the time points after each component data to obtain the wind direction prediction data at no more than 30% of the time points following the original wind direction time series.

[0037] Compared with existing technologies, the advantages of this invention are as follows: To reduce the volatility of the original wind direction data signal, the prediction method of this invention uses the CEEMDAN algorithm to decompose the original wind direction sequence signal into a series of intrinsic mode components and a set of residual components; to avoid the Grey Wolf optimization algorithm getting trapped in local optima and improve local search accuracy, a strategy of dynamically adjusting the nonlinear convergence factor is adopted, and a disturbance factor selection mechanism based on stagnation detection and an out-of-bounds reset strategy based on influence coefficient are used to improve the convergence of the algorithm, and the hyperparameters in the N-BEATS model are optimized to predict each component; finally, the prediction results of each component are merged to obtain the final prediction result. The prediction method of this invention outperforms traditional neural network models in terms of prediction accuracy, generalization, and interpretability in the study of wind direction time series prediction for wind turbines, and can be effectively applied to the prediction of wind direction for wind turbines in real-world scenarios. Attached Figure Description

[0038] The accompanying drawings, which are included to provide a further understanding of the embodiments of the present invention and form part of this application, are not intended to limit the embodiments of the present invention.

[0039] Figure 1 This is a schematic diagram illustrating the principle of the wind farm ultra-short-term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS of this invention.

[0040] Figure 2 This is the result of decomposing the original wind direction time series using the CEEMDAN method in one embodiment.

[0041] Figure 3 This is a schematic diagram of the structure of blocks in a typical N-BEATS network model.

[0042] Figure 4 The graph shows the decrease in the convergence factor of the Grey Wolf Optimization Algorithm and the improved Grey Wolf Optimization Algorithm in this invention during the iteration process.

[0043] Figure 5 This is a schematic flowchart illustrating the steps of an embodiment of the wind farm ultra-short-term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS of the present invention.

[0044] Figure 6 This is a comparison chart of the wind direction prediction results of one embodiment of the wind farm ultra-short-term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS (hereinafter referred to as CEEMDAN-IGWO-N-BEATS) of the present invention with the wind direction prediction results of the traditional N-BEATS method, IGWO-N-BEATS method and CEEMDAN-N-BEATS method. Detailed Implementation

[0045] To more clearly illustrate the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and examples. The embodiments and descriptions of the present invention are for illustrative purposes only and are not intended to limit the scope of the invention.

[0046] This invention provides a method for ultra-short-term wind direction prediction in wind farms based on CEEMDAN-IGWO-N-BEATS. The specific implementation steps of this prediction method are as follows:

[0047] Step 1: Select historical wind direction data time series from the yaw system of the wind turbine equipment, and select data points at 20-second time intervals to obtain the original wind direction time series;

[0048] Step 2: Decompose the original wind direction time series using the CEEMDAN method to obtain K sets of intrinsic mode components and one set of residual components; for example Figure 2As shown, the original wind direction time series was decomposed to obtain nine sets of intrinsic mode components and one set of residual values.

[0049] 2.1 CEEMDAN Decomposition First, Gaussian white noise w that follows a standard normal distribution is decomposed. i (t) The original signal x(t) is added.

[0050] x i (t)=x(t)+w i (t), i = 1, 2, ..., n (1)

[0051] In the formula, n is the total amount of signal decomposed by CEEMDAN.

[0052] 2.2 for x i (t) Perform EMD decomposition on the resulting n modal components Perform an overall average to obtain the first-order component of the signal.

[0053]

[0054]

[0055]

[0056] In the formula, r1(t) is the first-order residual. The second-order IMF component (represented by the character IMF2 in the attached figure, and so on for other components), ε1 represents the noise-related parameter, E m (·) represents the m-th order IMF component after EMD decomposition.

[0057] 2.3 Continue to add noise to the remaining components after the kth order of CEEMDAN decomposition.

[0058]

[0059]

[0060] In the formula r k (t) represents the calculated k-th order residual (k = 2, 3, ..., K). It is the k+1 order IMF component.

[0061] 2.4 Repeat the above steps until the obtained residual function is a monotonic function (less than two extreme values), and then stop decomposing.

[0062]

[0063]

[0064] In the formula, R(t) is the residual signal after final decomposition, and x(t) is the original signal after final decomposition.

[0065] Step 3: For the K groups of intrinsic mode components and one group of residual components obtained in Step 2, totaling K+1 component data, normalize at least 70% (80% in this embodiment) of the data at each time point, and use this normalized data as input to the network model; the remaining data corresponding to the component data are used as reference values ​​for training, thus obtaining the training dataset. The normalization formula is:

[0066]

[0067] In the formula, s i,std s represents the normalized data at the i-th time point of a certain component quantity data. i This represents the raw data of the component quantity data at the i-th time point, s min This represents the minimum value among the first 70% of the component quantity data, s max This indicates the maximum value among at least 70% of the component's quantity data.

[0068] Step 4: Design an improved N-BEATS network model

[0069] The basic structure of the improved N-BEATS network model is similar to that of the typical N-BEATS network model, both consisting of multiple stacks. Each stack comprises multiple blocks, and the forward predictions of multiple blocks are superimposed to obtain the forward prediction output of one stack. The forward predictions of each stack are then superimposed to obtain the final forward prediction result. The difference lies in the structure of the blocks in the typical N-BEATS network model, as follows: Figure 3 As shown, the first part of its FC Stack is a four-layer fully connected layer structure. The number of fully connected layers in the FC Stack of the block of the improved N-BEATS network model is set to m. The settings of the rest are the same as those in the typical N-BEATS network model (the structure and principle of the typical N-BEATS network model can be found in the following reference: Boris N. Oreshkin, Dmitri Carpov, Nicolas Chapados, Yoshua Bengio. N-BEATS: Neural basis expansion analysis for interpretable time series forecasting. [J]. CoRR, 2019, abs / 1905.10437.).

[0070] For the nth block of a stack at level k, input x nFirst, the data is processed sequentially through m fully connected layers of the FC Stack, and the output is h. n Output h n The data is then processed through two linear projection layers (FC layers, i.e., fully connected layers) to generate forward and backward prediction coefficients, respectively. and The mathematical formula is as follows:

[0071]

[0072] Where LINEAR is a linear projection layer, b represents backcast (predicting backwards), f represents forwardcast (predicting forwards); h n x represents n The output after processing by the FC Stack is then processed by the basis vectors. The accuracy of forward and backward predictions is optimized to obtain the backward prediction (i.e., backtracking) output values. and forward prediction (i.e. forecast) output values

[0073]

[0074] basis vectors The constraint on the function space is the reason why the N-BEATS model is more interpretable.

[0075] Each block produces two outputs: a backcast output and a forward output. The forward outputs of each block are summed to form the output of a stack. The original input of each block is subtracted from the backcast output, and the result is used as the input of the next block. All stacks sum the outputs of the forward portions to generate the final output. The backcast output is used as abstract information of historical data and fed into the next block with a stack for fitting.

[0076] The N-BEATS network model is based on forward and backward residual links and a very deep fully connected layer as its basic framework. Borrowing from classic residual network structures, it proposes a hierarchical dual-residual network topology. The input of a block is subtracted from its backward prediction (i.e., backtracking) output to obtain the residual output of that block, which is then used as the input to the next block. For the (n+1)th block of the k-th stack, its input is:

[0077]

[0078] To simulate trends, a specific polynomial is introduced:

[0079]

[0080] In the formula, p is the polynomial degree, which is set to 4 in this embodiment; k represents the k-th level of the stack; n is the n-th block of the k-th level of the stack; vector t = [0, 1, 2, ..., N-2, N-1] T / N, where N is the prediction window length, i.e. the prediction step size; T represents transpose.

[0081] Step 5: Train the improved N-BEATS network model using the training dataset and the improved gray wolf optimization algorithm.

[0082] The number of stacks, the number of blocks in each stack, and the number of fully connected layers *m* in the FC Stack portion of each block are used as hyperparameters in the improved N-BEATS network model and substituted into the improved Gray Wolf optimization algorithm for optimization. The ranges for the number of stacks, the number of blocks in each stack, and the number of fully connected layers in the FC Stack portion of each block are set to [1,6], [1,8], and [128,512], respectively, and these ranges are used as initial values ​​in the improved Gray Wolf optimization algorithm to initialize the population.

[0083] Normalized data from at least 70% of the time points in a set of component data in the training dataset are input into the improved N-BEATS network model to predict the data at the remaining time points. The individual fitness is then calculated by combining the reference values ​​of the remaining time points of the corresponding component data in the training set. The root mean square error (RMSE) is used as the individual fitness of the improved Grey Wolf optimization algorithm.

[0084]

[0085] In the formula, y represents the reference value of the component data. This represents the predicted value of the component data.

[0086]

[0087] This represents the predicted data output by the network model. Indicates will The predicted values ​​of the component data obtained after inverse normalization, y min y represents the minimum value in the input data of the network model before normalization. max This represents the maximum value in the input data of the network model before normalization.

[0088] The Grey Wolf Algorithm (GWO) is a globally stochastic optimization algorithm that learns from the hunting behavior of grey wolves in nature. The wolf pack can be divided into four levels, from highest to lowest: α, β, δ, and ω. The hunting process is as follows:

[0089] First, initialize the gray wolf population, and simultaneously initialize the convergence factor a and the A and C coordination coefficient vectors:

[0090]

[0091] During the iteration process, the value of 'a' linearly decreases from 2 to 0; the modulo of vectors r1 and r2 is taken as random numbers between [0,1]. The fitness of individual gray wolves is calculated, and the three wolves with the best fitness are retained, along with their α, β, and δ values. The gray wolf optimization process is mainly guided by α, β, and δ in each generation. Encircling the prey:

[0092] D = |C·X p (t)-X(t)| (17)

[0093] X(t+1)=X p (t)-AD (18)

[0094] D is the distance between the gray wolf and its prey, X p Let X represent the prey's location, t represent the gray wolf's location, and t represent the current iteration number. Hunting the prey: Based on the useful information about the target location obtained from equations (17) and (18), the wolf pack searches for the target prey under the guidance of α, β, and δ wolves. This mathematical model can be expressed as:

[0095]

[0096]

[0097]

[0098] In the formula X α X β X δ D represents the position vectors of α, β, and δ wolves, respectively. α D β D δ These represent the current candidate wolf's positional distance from α, β, and δ wolves, respectively. |A|>1 indicates that the gray wolves are searching in a dispersed manner, while |A|<1 indicates that the gray wolves are searching in a concentrated manner within certain areas. This indicates the updated location of ωwolf after this round of capture.

[0099] When the prey stops moving, the gray wolves attack to complete the hunting process. To simulate approaching the prey, the value of 'a' is gradually decreased, thus reducing the fluctuation range of 'A'. In other words, during the iteration process, as the value of 'a' linearly decreases from 2 to 0, the corresponding value of 'A' also varies within the interval [-a, a]. When the value of 'A' is within this interval, the gray wolf's next position can be anywhere between its current position and the prey's position. When |A| < 1, the wolf pack attacks the prey (getting trapped in a local optimum). When |A| > 1, the gray wolves separate from the prey, hoping to find a more suitable prey (global optimum).

[0100] The GWO algorithm also has another component, C, to help discover new solutions. As shown in equation (16), C is a random value between [0, 2]. C represents the random weight of the wolf's location on the prey; C>1 indicates a large influence weight, and vice versa. This helps the GWO algorithm to behave more randomly and support exploration, while avoiding getting trapped in local optima during the optimization process. In addition, unlike A, C decreases non-linearly, thus providing a global search in the decision space from the initial iteration to the final iteration. When the algorithm gets trapped in local optima and is not easy to escape, the randomness of C plays a very important role in avoiding local optima, especially in the final iteration where a global optimal solution needs to be obtained (the principle and process of the gray wolf optimization algorithm can be found in the following reference: Wang Lihui, Yang Huibin, Wang Yintang, Liu Yong, Hu Qingfang. Prediction of inflow runoff of Danjiangkou Reservoir based on GWO-LSTM [J]. Journal of Water Resources and Water Transport Engineering, 2021(06):51-59.).

[0101] The above describes the principle and process of the Grey Wolf optimization algorithm. The improved Grey Wolf optimization algorithm in this invention differs from it in that:

[0102] The formula for calculating the convergence factor 'a' of the improved gray wolf optimization algorithm is as follows:

[0103]

[0104] In the formula, t max The maximum number of iterations is a set value; t represents the current number of iterations. This convergence factor is a nonlinear control convergence factor based on the cosine function.

[0105] The improved gray wolf optimization algorithm employs a stagnation detection-based perturbation factor selection mechanism. It simultaneously detects the fitness values ​​of α, β, and δ wolves. If all α, β, and δ wolves stagnate (i.e., their fitness remains unchanged), their guiding role in the population weakens, and a maximum stagnation period is set for the population. Furthermore, when the population stagnates but does not exceed the maximum stagnation period, a perturbation factor with a small step size is added to search within a small range, avoiding impact on the convergence of later algorithms. When the population stagnates beyond the maximum stagnation period, it indicates that a small perturbation factor is insufficient to help the population escape the local optimum, and a perturbation factor with a larger step size is added, as shown in the following formula:

[0106]

[0107] X new X represents the individual after the position change. t This represents the current position of the individual gray wolf. ub is the upper bound of the maximum value, lb is the lower bound of the minimum value, and N(0,1) is a random number conforming to a standard normal distribution. μ is an adjustable parameter (used to adjust the fluctuation range of the perturbation factor, which is μ2(ub-lb)·N(0,1)). Here, μ1 takes a smaller value, ranging from [0.01, 0.05], and μ2 takes a larger value, ranging from [0.1, 0.5]. g represents the stagnation period, and Mg represents the maximum stagnation period, which is set to 3 here.

[0108] Retain X with added perturbation factor new Current location of individual gray wolves awaiting update Those with better fitness will be selected as the next generation of individuals.

[0109]

[0110] In formula (24), X t+1 Let f(X) represent the position of the next generation population, and let f(X) represent the fitness value of the individual. The perturbation factor uses a random number from a standard normal distribution to ensure that individuals in the population have an equal probability of jumping in different directions and with different step lengths. The step length is adjusted using variable parameters μ1 and μ2 to avoid the Grey Wolf Optimization Algorithm getting trapped in local optima during optimization.

[0111] The improved gray wolf optimization algorithm employs a boundary reset strategy based on influence coefficients. The original gray wolf optimization algorithm's reset strategy places out-of-bounds individuals at the boundary, which often results in these individuals remaining at the boundary after the next update, limiting the algorithm's convergence speed in the early stages and leading to local optima. To fully utilize the guidance of α, β, and δ wolves, a boundary reset strategy based on influence coefficients is proposed. The influence coefficients of α, β, and δ wolves are calculated using their fitness values; the smaller the fitness value, the stronger the individual's leadership and the higher the influence coefficient. The formula is as follows:

[0112]

[0113]

[0114]

[0115]

[0116] In the above formula, I α I β I δ These are the influence coefficients of α, β, and δ wolves, respectively, f α f β f δ α, β, and δ represent the individual fitness of wolves; μ is a variable parameter set to a small value here, ranging from [0.01, 0.05], to ensure that the reset position is selected under the guidance of wolves α, β, and δ, while allowing for small-range jumps.

[0117] When the number of iterations reaches the set maximum number of iterations, the corresponding hyperparameter values ​​are output, namely the number of stacks in the improved N-BEATS network model, the number of blocks in each stack, and the number of fully connected layers in the FCStack part of each block, thus obtaining the improved N-BEATS network model trained with the sample data.

[0118] Step Six: Process each component data in the training set according to the method described in Step Five to obtain an improved N-BEATS network model trained on each component data. Normalize at least 70% (80% in this embodiment) of the K+1 component data obtained in Step 2, and use this normalized data as input to the corresponding improved N-BEATS network model trained on the corresponding component data, obtaining the predicted data output by each network model. Inverse normalize the predicted data output by each network model to obtain the predicted values ​​(marked in the attached figure) of the data at at least 30% of the time points (30% is the percentage of time points; if the input data time points of the network model are 70, then the output predicted data time points are 30, and so on; the total of the input and output time points of the network model is 100%). Alternatively, R); the predicted values ​​of the data at no more than 30% of the time points following each component of the wind direction data are superimposed to obtain wind direction prediction data at no more than 30% of the time points following the original wind direction time series.

[0119] Table 1. N-BEATS Model Parameters for Component Quantities Data

[0120]

[0121] The final parameter optimization results of the improved gray wolf optimization algorithm are shown in Table 1. It can be seen from the table that each modal component corresponds to different optimal parameters, and the optimal values ​​are concentrated within the parameter range, indicating that the experimental parameter range is set reasonably.

[0122] Under the same experimental conditions, the wind direction prediction results of this invention and existing methods such as N-BEATS, LSTM, ARIMA, GWO-N-BEATS, IGWO-N-BEATS, PSO-N-BEATS, CEEMDAN-N-BEATS, and EMD-N-BEATS were compared. The results show that the prediction accuracy of the CEEMDAN-IGWO-N-BEATS method designed in this invention is significantly better than traditional machine learning models, and better than single prediction models. Furthermore, the IGWO-N-BEATS combination is superior to CEEMDAN-N-BEATS. The evaluation index expression for the prediction results is as follows:

[0123] Root Mean Square Error (RMSE):

[0124]

[0125] Mean Absolute Error (MAE):

[0126]

[0127] Mean Absolute Percentage Error (MAPE):

[0128]

[0129] In the formula, y represents the actual value of the wind direction data. This represents the predicted value of the wind direction data. The prediction results are shown in Table 2.

[0130] Table 2. Average prediction accuracy of each model

[0131]

[0132] Figure 6 In this study, by comparing the wind direction prediction performance of the CEEMDAN-IGWO-N-BEATS method with the traditional N-BEATS method, IGWO-N-BEATS method, and CEEMDAN-N-BEATS method, it was found that the CEEMDAN-IGWO-N-BEATS model has a higher degree of trend fit on some data points with greater volatility.

[0133] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A wind farm ultra-short-term wind direction prediction method based on CEEMDAN-IGWO-N-BEATS, characterized in that, The prediction method specifically implements the following steps: Step one: select the yaw system from the fan equipment The historical data time series of the wind direction, select the data points according to the fixed time interval, and obtain the original wind direction time series; Step two: decompose the original wind direction time series using the CEEMDAN method to obtain K sets of intrinsic mode components and a set of residual components; Step three: the K sets of intrinsic mode components and the set of residual components obtained in step two are summed up K+1 group component data, and the data of not less than 70% of the time points are normalized respectively, which are used as the input of the network model; the remaining data of the corresponding component data are used as the reference value during training to obtain the training data set; The normalization processing formula is: wherein, s i,std denotes the normalized data of the i-th time point of a certain component data, s i denotes the original data of the i-th time point of the component data, s min denotes the minimum value in the first not less than 70% of the component data, s max denotes the maximum value in the first not less than 70% of the component data; Step four: design an improved N-BEATS network model The basic structure of the improved N-BEATS network model is similar to that of the typical N-BEATS network model, both of which are composed of multiple layers of stacks, each layer of stack is composed of multiple blocks, the forward prediction values of multiple blocks are superimposed to obtain the forward prediction output of a layer of stack, and the forward prediction values of each layer of stack are superimposed to obtain the final forward prediction result; the number of full connection layers of FC Stack in the block of the improved N-BEATS network model is set to m, and the rest is the same as that in the typical N-BEATS network model; Step five: train the improved N-BEATS network model using the training data set and combining the improved grey wolf optimization algorithm The number of stacks, the number of blocks in each layer of stack, and the number of full connection layers m of the FC Stack part in each block in the improved N-BEATS network model are taken as hyperparameters and brought into the improved grey wolf optimization algorithm for optimization; the range of the number of stacks, the number of blocks in each layer of stack, and the number of full connection layers of the FC Stack part in each block is set to [1, 6], [1, 8], and [128, 512] respectively, which are taken as the initial range and brought into the improved grey wolf optimization algorithm to initialize the population; The normalized data of not less than 70% of the time points in a group of component data in the training data set are input into the improved N-BEATS network model to predict the data of the remaining time points, and the reference values of the remaining time points of the corresponding component data in the training set are combined to calculate the individual fitness; the root mean square error is taken as the individual fitness of the improved grey wolf optimization algorithm: where y represents a reference value of the component data, represents a predicted value of the component data; predicted data representing the output of the network model, representing the output of the network model, predicted value of the component data obtained after the inverse normalization, y min representing the minimum value before the normalization process in the input data of the network model, y max representing the maximum value before the normalization process in the input data of the network model; The calculation formula of the convergence factor a of the improved grey wolf optimization algorithm is: In the formula, t max represents the maximum value of iteration times, which is a set value; t represents the current iteration times; The improved grey wolf optimization algorithm adopts a disturbance factor selection mechanism based on stagnation detection, detects the fitness values of alpha, beta and delta wolves at the same time, and if all alpha, beta and delta wolves are in stagnation, the maximum stagnation period of the population is set, and when the population is in stagnation and has not exceeded the maximum stagnation period of the population, a disturbance factor with a smaller step size is added; when the population is in stagnation and exceeds the maximum stagnation period of the population, a disturbance factor with a larger step size jump is added, and the formula is as follows: X new X represents the individual after the position change. t This represents the current position of the individual gray wolf; ub is the upper bound of the maximum value; lb is the lower bound of the minimum value; N(0,1) is a random number that conforms to a standard normal distribution; μ is an adjustable parameter, where μ1 takes a smaller value, ranging from [0.01, 0.05], and μ2 takes a larger value, ranging from [0.1, 0.5]; g represents the stagnation period, and Mg represents the maximum stagnation period. X with a disturbance factor added new the current gray wolf individual to be updated better fitness in the population as the next generation individual; In Equation (24), X t+1 is the position of the next generation, and f(X) represents the fitness value of the individual. The improved grey wolf optimization algorithm adopts a boundary resetting strategy based on influence coefficient, and the influence coefficients of alpha, beta and delta wolves are calculated based on their fitness values, and the formula is as follows: In the above formula, I α , I β , I δ are the influence coefficients of the alpha, beta and delta wolves respectively, f α , f β , f δ are the individual fitness of the alpha, beta and delta wolves respectively; μ is a variable parameter, which is set to a small value in this case, and the value range is [0.01, 0.05]; When the number of iterations reaches the set maximum number of iterations, the corresponding hyperparameter values are output, i.e., the number of stacks in the improved N-BEATS network model, the number of blocks in each stack layer, and the number of fully connected layers in the FC Stack part of each block, to obtain the improved N-BEATS network model trained by the component data; Step six: each set of component data in the training set is processed according to the method described in step five to obtain the improved N-BEATS network model trained by each set of component data; the last no less than 70% of the data of the K+1 sets of component data obtained in step 2 are normalized as inputs of the network model and input into the improved N-BEATS network model trained by the corresponding set of component data respectively to obtain the predicted data output by each network model respectively; the predicted data output by each network model is de-normalized to obtain the predicted values of the data of the last no more than 30% of the time points of each set of component data; the predicted values of the data of the last no more than 30% of the time points of each set of component data are superimposed to obtain the wind direction prediction data of the time points no more than 30% after the original wind direction time sequence.

2. The CEEMDAN-IGWO-N-BEATS-based wind farm ultra-short-term wind direction prediction method according to claim 1, characterized in that, The fixed time interval in step one is 20s.

3. The CEEMDAN-IGWO-N-BEATS-based wind farm ultra-short-term wind direction prediction method according to claim 1, characterized in that, In step three, the first 80% of the data of each set of component data is normalized respectively.

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 stagnation period Mg is set to 3. In step five, the maximum stagnation period Mg is set to 3.

Images