Wind speed prediction method based on deep learning

Abstract

The invention discloses a wind speed prediction method based on deep learning, and relates to the technical field of wind power generation, and the method comprises the steps: inputting an original wind speed sequence, and carrying out the wind speed decomposition through employing a variational mode decomposition algorithm, and obtaining a group of IMF components with different frequencies; analyzing the obtained IMF components, removing high-frequency components, and combining into a new wind speed sequence; taking the new wind speed sequence as input, and performing wind speed prediction through a KAN-LSTM network; and according to the prediction result and the true value, the prediction precision is evaluated. Therefore, by adopting the wind speed prediction method based on deep learning, harmful high-frequency components containing white noise can be effectively removed, and the wind speed prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and particularly to a wind speed prediction method based on deep learning. Background Art

[0002] With the increase in energy consumption and the emphasis on the ecological environment, wind energy has attracted worldwide attention due to its advantages of being clean, pollution-free, and renewable. With the development of new energy technologies, wind power generation, as a green and renewable energy source, has great development potential.

[0003] Wind speed prediction plays a crucial guiding role in the construction and operation of wind farms. First of all, in the site selection stage, accurate wind speed prediction helps developers evaluate the wind energy resources in different regions, so as to select the best location to maximize the power generation potential. Secondly, the prediction results affect the selection and layout of wind turbines, ensuring the safe and efficient operation of the equipment under various wind speeds. In operation management, wind speed prediction is used to optimize power generation scheduling, ensure the stability of power supply, and support equipment maintenance plans to reduce the risk of failures. In addition, through wind speed data, enterprises can conduct accurate economic evaluations, predict future revenues, and formulate effective risk management strategies. Finally, wind speed prediction also promotes the coordination between wind farms and the power grid, optimizes the matching of power generation and load, thereby enhancing the stability and security of the power grid. In short, wind speed prediction is an important tool for improving the construction efficiency and operation benefits of wind farms.

[0004] At present, the commonly used research methods for wind speed prediction include: physical methods, statistical methods, machine learning methods, and deep learning methods, etc. Among them, physical methods are based on a detailed physical description of the atmosphere, using meteorological data such as air temperature, terrain, and air pressure to predict wind speed. Usually, large computers are required to perform long-term integration of the differential equations of fluid mechanics, and the calculation cost and theoretical requirements are quite high; statistical methods analyze the historical wind speed sequence to explore the essential laws of the wind speed sequence, so as to predict the future wind speed. However, due to the nonlinear and intermittent problems presented by the wind, the accuracy of statistical methods is not high, that is, statistical methods cannot well handle nonlinear problems; while deep learning methods are relatively simple to implement and have relatively lower calculation costs, so they are widely used in non-linear prediction fields such as meteorology and hydrology in practical applications.

[0005] Deep learning methods can adaptively learn and develop models, not completely relying on historical data or physical principles, and are more suitable for dealing with nonlinear problems. For example, recurrent neural networks (RNNs), convolutional neural networks (CNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), and other network algorithms. However, existing deep learning methods all use a single model, which cannot capture complex nonlinear relationships and time variations, limiting their prediction ability and adaptability. Based on this, combined models have emerged to overcome the limitations of individual models, thereby improving the accuracy and stability of predictions. For example, Y. Zhang and Y. Zhao et al. used the MD-PRBF-ARMA-E model to improve the wind speed prediction accuracy. This strategy simultaneously addressed the periodic and aperiodic characteristics of wind speed, but the drawback is that the traditional ARMA model cannot fit complex nonlinear problems well. O. Abedinia and M. Lotfi et al. proposed the IEMD-BaNN-Kmeans hybrid model, which successfully solved the problems of wind speed instability and chaotic prediction. However, the drawback is that after using the IEMD algorithm, the components were not screened, and white noise problems were easily introduced.

[0006] Due to the intermittent and nonlinear effects of the wind, the power generation of wind turbines is unstable, seriously affecting the quality of wind power generation. Therefore, improving the accuracy of wind speed prediction can help wind farms make rapid scheduling decisions, thereby improving the wind energy conversion efficiency, which is of great significance for the development of wind energy. Summary of the Invention

[0007] The purpose of the present invention is to provide a wind speed prediction method based on deep learning, which can remove harmful high-frequency components containing white noise, thereby improving the wind speed prediction accuracy and ensuring the safety and efficiency of high-sea wind turbines and offshore operations.

[0008] To achieve the above object, the present invention provides a wind speed prediction method based on deep learning, including the following steps:

[0009] S1. Input the original wind speed sequence and use the variational mode decomposition algorithm to decompose the wind speed to obtain a set of IMF components with different frequencies;

[0010] S2. Analyze the obtained IMF components, remove the high-frequency components, and combine them into a new wind speed sequence;

[0011] S3. Use the new wind speed sequence as the input and perform wind speed prediction through the KAN-LSTM network;

[0012] S4. Evaluate the prediction accuracy according to the prediction result and the true value.

[0013] Preferably, the variational mode decomposition algorithm is as follows:

[0014] First, decompose the original signal into several modal components. Take the sum of the estimated bandwidths of each mode as the objective function, and the constraint condition is that the sum of all modes is equal to the original signal, and construct a VMD constrained variational model.

[0015] Secondly, introduce the augmented Lagrangian function, and use the quadratic penalty term and the Lagrange multiplier method to transform the VMD constrained variational model into an unconstrained variational problem for solution.

[0016] Preferably, the KAN-LSTM network includes a KAN network and an LSTM neural network;

[0017] Among them, the KAN network replaces the traditional linear weight by defining a learnable univariate function on the edge;

[0018] The LSTM neural network is used to process the long-distance dependence relationship in the sequence data.

[0019] Preferably, evaluating the prediction accuracy includes using the mean absolute error, mean relative error, and root mean square error to evaluate the performance of the model and measure the reliability of the prediction results.

[0020] Therefore, the present invention adopts the above-mentioned wind speed prediction method based on deep learning, and has the following technical effects:

[0021] (1) Divide the original wind speed data into several modal components through the variational mode decomposition algorithm, and then analyze each modal component and remove the high-frequency noise components to improve the quality of the input data and lay a foundation for subsequent prediction.

[0022] (2) Use the processed wind speed data as input, and perform wind speed prediction through the KAN-LSTM network, which can better process the long-distance dependence relationship in the sequence data, thereby improving the accuracy of wind speed prediction.

[0023] Next, through the drawings and embodiments, the technical solutions of the present invention will be further described in detail. Description of the Drawings

[0024] Figure 1 is a flowchart of a wind speed prediction method based on deep learning;

[0025] Figure 2 is a schematic diagram of the KAN network structure in an embodiment of a wind speed prediction method based on deep learning;

[0026] Figure 3 is a schematic diagram of the LSTM neural network structure in an embodiment of a wind speed prediction method based on deep learning;

[0027] Figure 4It is a prediction result graph in an embodiment of a wind speed prediction method based on deep learning. Detailed implementation manners

[0028] The present invention can be more specifically explained through the following embodiments. The purpose of disclosing the present invention is to protect all changes and improvements within the scope of the present invention. The present invention is not limited to the following embodiments.

[0029] As Figure 1 shown, the present invention provides a wind speed prediction method based on deep learning, which predicts the wind speed data at a certain position in the offshore area of Fujian in ERA5, including the following steps:

[0030] S1. Input the original wind speed sequence, and use the variational mode decomposition algorithm to decompose the wind speed to obtain a set of IMF components with different frequencies, specifically as follows:

[0031] (1) Assume that the original signal f is decomposed into K components, ensuring that the decomposition sequence is a modal component with a finite bandwidth having a center frequency, and at the same time, the sum of the estimated bandwidths of each mode is minimized. The constraint condition is that the sum of all modes is equal to the original signal. Then the VMD constrained variational model is:

[0032]

[0033] In the formula, {u k} = {u 1 , u 2 , …, u k} are each modal function, and {w k} = {w 1 , w 2 , …, w k} are the center frequencies of each mode; The partial derivative with respect to time t. In variational mode decomposition (VMD), this symbol is used to represent the dynamic characteristics of the signal changing with time, usually related to the frequency characteristics; δ(t) represents the Dirac function; t represents the time variable, which is the independent variable of the signal f(t).

[0034] (2) Solve the above-mentioned constrained optimization problem, transform the constrained variational problem into an unconstrained variational problem, and introduce the augmented Lagrangian function by utilizing the advantages of the quadratic penalty term and the Lagrange multiplier method, as shown in the following formula:

[0035]

[0036]

[0037] In the formula, α is the penalty parameter, and λ is the center frequency of each mode.

[0038] (3) For all ω ≥ 0, update the functional

[0039]

[0040] where f(ω) is the representation of the original signal f(t) in the frequency domain, i.e., the Fourier transform result of f(t).

[0041] Update the functional ω k :

[0042]

[0043] For all ω ≥ 0, further perform double lifting:

[0044]

[0045] where γ represents noise. When the signal contains strong noise, γ can be set to 0 to achieve a better denoising effect. is the updated frequency-domain representation.

[0046] (4) Do not repeat the update of the functional until the following iterative constraint condition is satisfied:

[0047]

[0048] where ε is the judgment accuracy.

[0049] S2. Since the variation of the offshore wind speed usually has obvious low-frequency characteristics, such as seasonal variation, diurnal variation, and atmospheric circulation, etc., the components in the high-frequency part may reflect short-time scale random perturbations, such as instrument measurement errors, small turbulences, random noises in the environment, etc. Removing these high-frequency components may improve the performance of wind speed prediction. In this embodiment, among the K IMF components obtained by VMD decomposition, N IMF components with the highest frequencies are removed from less to more, and the remaining K - N IMF components are recombined to form a new wind speed sequence. Among them, the size of N is determined according to the final prediction effect, and the N that minimizes the prediction error is selected.

[0050] S3. Use the new wind speed sequence as the input and perform wind speed prediction through the KAN-LSTM network. Among them, the KAN-LSTM network is composed of the KAN network and the LSTM neural network.

[0051] The KAN network structure is constructed based on the Kolmogorov-Arnold representation theorem and is significantly different from the traditional multi-layer perceptron (MLP). This network realizes non-linear mapping by learning univariate functions and defining activation functions on the edges of the network, rather than using fixed activation functions on nodes like MLP. Through the KAN network, it is possible to better handle the non-linear and highly unstable characteristics of sea surface wind speed, thus achieving better prediction results.

[0052] As Figure 2 shown, the KAN network layer consists of univariate functions of a matrix, and each function is parameterized by a spline function, with the expression as follows:

[0053] Given the input The output of KAN is obtained through the combination of multiple layers:

[0054]

[0055] In the formula, Φ l is the function matrix of the l-th layer, expressed as:

[0056]

[0057] In the formula, is a univariate activation function, and x l,i is the input related to the i-th node in the l-th layer.

[0058] The activation value of the KAN network is calculated through the combination of the accumulated sum of the input and non-linear functions, specifically as:

[0059]

[0060] In the formula, x l+1,j is the activation value of the j-th node in the (l + 1)-th layer network, and n l is the number of nodes in the l-th layer. This process is repeated for each layer of the network until the final output layer, and its output result is used as the input of the Long-Short Term Memory (LSTM) network.

[0061] The LSTM neural network is a special type of recurrent neural network (RNN). It controls the flow of information by introducing a gating mechanism and can better handle long-distance dependencies in sequential data. As Figure 3 shown, the core structure of LSTM consists of a series of repeated units. Each unit receives the input at the current time and the hidden state at the previous time as inputs and outputs the hidden state and output at the current time. Among them, each unit is mainly composed of 4 parts, as follows:

[0062] (1) Forget gate, which determines which information to discard from the cell state, and the expression is:

[0063] f t =σ(W f ·[h t-1 ,x t +b f );

[0064] In the formula, σ is the sigmoid function, W f is the weight matrix, b f is the bias term, x t is the input at the current moment, h t-1 is the hidden state at the previous moment. The output f t is a value between 0 and 1, indicating the information forgotten from the cell state at the previous moment.

[0065] (2) Input gate, which determines which new information is stored in the cell state and consists of two parts:

[0066] First, the activation value of the input gate, whose calculation method is similar to that of the forget gate, determines which new information is used for calculation, and the expression is:

[0067] i t =σ(W i ·[h t-1 ,x t +b i ;

[0068] In the formula, W i is the weight matrix of the input gate.

[0069] Second, the candidate cell state, which is generated by the tanh function here, and the expression is:

[0070]

[0071] In the formula, W C is the weight matrix of the candidate cell state.

[0072] (3) Cell state update, which determines how much old information to retain by multiplying the output of the forget gate by the cell state at the previous moment, and then adding the new information determined by the input gate. The expression is:

[0073]

[0074] In the formula, C t 、C t-1 are the cell states at the current moment and the previous moment respectively.

[0075] (4) Output gate, which determines which information to output, and the expression is:

[0076] o t = σ(W o ·[ht -1 ,x t +b o );

[0077] h t = o t *tanh(C t );

[0078] Wherein, o t is the activation value of the output gate, and h t is the hidden state calculated at the current time step.

[0079] S4. As Figure 4 shown, by quantitatively evaluating the model prediction results, the performance of the model can be judged more objectively. In this embodiment, the mean absolute error σ MAE , mean relative error σ MAPE and root mean square error σ RMSE are selected to quantitatively evaluate the prediction results of the model. Among them, the mean absolute error is used to measure the average gap between the prediction results and the actual values, without considering the positive and negative directions of the errors, avoiding the situation where the positive and negative errors cancel each other out. The mean relative error is used to measure the relative gap between the prediction results and the actual values, and can reflect the credibility of the prediction experiment. The root mean square error is used to measure the gap between the prediction results and the actual values, and is more sensitive to larger error values when calculating the errors.

[0080]

[0081]

[0082] Wherein, p(i) is the predicted value, y(i) is the actual value, and N is the number of samples.

[0083] In addition, in this embodiment, the prediction accuracies of the LSTM network, KAN network, KAN-LSTM network and the present implementation method (VMD-KAN-LSTM) are also compared to verify the effectiveness of the solution of this embodiment, as shown in Table 1.

[0084] Table 1 Prediction accuracies of each model

[0085] Model <![CDATA[σ MAE > <![CDATA[σ MAPE > <![CDATA[σ RMSE > LSTM 0.227 2.625％ 0.328 KAN 0.093 2.490％ 0.304 KAN - LSTM 0.112 2.753％ 0.335 VMD - KAN - LSTM 0.065 2.294％ 0.256

[0086] Therefore, the present invention adopts the above-mentioned wind speed prediction method based on deep learning, which can effectively remove the harmful high-frequency components containing white noise and improve the wind speed prediction accuracy.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements do not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A wind speed prediction method based on deep learning, characterized in that: include: S1. Input the original wind speed sequence and use the variational mode decomposition algorithm to decompose the wind speed to obtain a set of IMF components with different frequencies; S2, analyze the obtained IMF components, remove the high-frequency components, and combine them into a new wind speed sequence; S3, taking the new wind speed sequence as input, and performing wind speed prediction through the KAN-LSTM network; S4. Evaluate the prediction accuracy based on the prediction results and the true value.

2. The wind speed prediction method based on deep learning according to claim 1, characterized in that: The variational mode decomposition algorithm is as follows: Firstly, the original signal is decomposed into K modal components, and the sum of the estimated bandwidths of each mode is used as the objective function. The constraint condition is that the sum of all modes is equal to the original signal, and a VMD constrained variational model is constructed. Secondly, the augmented Lagrangian function is introduced, and the quadratic penalty term and Lagrangian multiplier method are used to transform the VMD constrained variational model into an unconstrained variational problem for solution.

3. The wind speed prediction method based on deep learning according to claim 1, characterized in that: The KAN-LSTM network includes the KAN network and the LSTM neural network; Among them, the KAN network replaces the traditional linear weights by defining a learnable univariate function on the edge; LSTM neural networks are used to process long-distance dependencies in sequence data.

4. The wind speed prediction method based on deep learning according to claim 1, characterized in that: The evaluation of prediction accuracy includes using mean absolute error, mean relative error and root mean square error to evaluate the performance of the model and measure the reliability of the prediction results.

Images