A wind speed prediction method

Abstract

The application discloses a wind speed prediction method, comprising the following steps: adaptively decomposing a wind speed time sequence into a plurality of modal components, and dynamically determining the modal length and the modal number in the decomposition process through an adaptive optimization algorithm; for each modal component, calculating its local nonlinear feature respectively, and extracting a nonlinear component through a nonlinear mapping structure containing a learnable base function; for each modal component, calculating its long-term time sequence dependence feature respectively, and capturing a time sequence dependence relationship through an improved gate recurrent unit structure; fusing the local nonlinear feature and the long-term time sequence dependence feature of the same modal component; and performing wind speed prediction output based on the fused feature; the application realizes the improvement of the prediction precision, and provides a systematic methodology and engineering guidance for the prediction of non-stationary time sequences.

Description

Technical Field

[0001] This invention relates to the field of wind speed prediction technology, and more specifically to a wind speed prediction method. Background Technology

[0002] Wind speed forecasting holds crucial strategic importance in the energy industry, particularly in wind energy development, as it directly determines the economic viability and operational efficiency of wind power projects. Accurate wind speed forecasts provide a scientific basis for wind farm site selection, turbine layout optimization, and power generation prediction, helping to reduce investment risks and improve energy output stability. Simultaneously, forecast data supports grid dispatching, balancing the impact of wind energy fluctuations on the power system and preventing supply-demand imbalances. With the accelerated transition to clean energy, wind speed forecasting has become a core support for improving wind energy utilization, ensuring grid security, and promoting the large-scale development of renewable energy.

[0003] Wind speed time series exhibit significant non-stationarity, multi-scale fluctuations, and randomness. Current wind speed forecasting methods primarily rely on physical models and machine learning methods. While physical models offer high accuracy, they are computationally expensive, while machine learning models perform better in forecasting. However, current wind speed forecasting methods still face significant shortcomings: models have poor adaptability to complex terrain or extreme weather conditions; insufficient fusion of multi-source meteorological data leads to inadequate information utilization; deep learning models have weak interpretability, making them difficult to guide practical operation and maintenance; and there is a contradiction between the need for real-time forecasting and the computational efficiency of the models. Therefore, achieving high-precision and robust wind speed forecasting has become an urgent technical challenge to be solved. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a wind speed prediction method that adaptively decomposes wind speed time series into multiple modes, then designs two deep learning models with parallel structures to efficiently capture the local nonlinearity and long-term temporal dependence features of each mode, and fuses the output features of the two models to output the wind speed prediction result, thereby solving the problems existing in the background technology.

[0005] Technical solution: The wind speed prediction method of the present invention includes the following steps:

[0006] (1) The wind speed time series is adaptively decomposed into multiple modal components, and the mode length and number of modes are dynamically determined in the decomposition process by an adaptive optimization algorithm;

[0007] (2) For each modal component, calculate its local nonlinear characteristics and extract the nonlinear components through a nonlinear mapping structure containing learnable basis functions; for each modal component, calculate its long-term temporal dependency characteristics and capture the temporal dependency relationship through an improved gated recurrent unit structure.

[0008] (3) The local nonlinear characteristics of the same modal component are fused with the long-term time-dependent characteristics;

[0009] (4) Wind speed prediction output based on fusion features.

[0010] Furthermore, in step (1), the adaptive decomposition specifically involves: establishing a parameter space with mode length and mode number as optimization variables; using a multi-strategy collaborative optimization algorithm to search for parameters with the goal of minimizing the complexity of each modal component after decomposition; dynamically switching four update rules based on the population state and random decision variables during the optimization process, including local search guided by the current best individual, exploration guided by random individuals, spiral approximation strategy, and Lévy flight random walk mechanism; and outputting the optimal parameter combination when the convergence condition is met and using it for decomposition.

[0011] Furthermore, in step (2), the calculation of local nonlinear features is as follows: performing convolution operations on the modal components to extract linear features; performing nonlinear mapping on the linear features through a set of learnable basis functions, wherein the basis functions can adaptively capture nonlinear relationships.

[0012] Furthermore, in step (2), the calculation of long-term temporal dependency features is as follows: a gated loop unit structure is adopted, including update gate, reset gate and candidate hidden state calculation; the retention and updating of historical information are controlled by the gating mechanism to achieve effective modeling of long-term temporal dependencies.

[0013] Furthermore, in step (3), feature fusion specifically involves concatenating the local nonlinear feature vectors of the same modal component with the long-term time-dependent feature vectors to form a fused feature vector.

[0014] Furthermore, the method also includes: inputting the fusion features of each modal component into a fully connected layer for regression prediction to obtain the final wind speed prediction value.

[0015] An electronic device according to the present invention includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the methods described herein.

[0016] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the methods described herein.

[0017] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention automatically selects the optimal parameters when decomposing wind speed time series. The method of this invention outperforms the comparative methods in four evaluation indicators: root mean square error, mean absolute error, mean absolute percentage error, and coefficient of determination. This invention improves prediction accuracy and provides a systematic methodology and engineering guidance for the prediction of non-stationary time series. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention;

[0019] Figure 2 This is the network diagram for calculating the local nonlinear characteristics of this invention. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0021] like Figure 1 As shown in the figure, an embodiment of the present invention provides a wind speed prediction method, including the following steps:

[0022] Step 1: Decompose the wind speed time series.

[0023] Let the wind speed time series be Decompose it into the following formula: Each mode and residual:

[0024] ;

[0025] ;

[0026] When k=1 When k>1, ;

[0027] in, For the k-th mode, yes Length, It is the unit impulse response function; The residual is the result of the decomposition.

[0028] Step 2: Adaptively determine the parameters for decomposing the wind speed.

[0029] This step adaptively optimizes the parameters in step 1. and .

[0030] First, determine the objective function and parameter space for optimization.

[0031] Secondly, perform parameter optimization as follows:

[0032] (21) Population initialization: Randomly generate N individuals X(i) in the parameter space, each individual is represented as (x i ,y i ). x i Represents Size, y i Represents modenum.

[0033] (22) Calculate the objective function value: For each individual, substitute the values ​​from step 1 to calculate each mode, and then calculate the objective function for each mode.

[0034] (23) Population update: Determine the behavioral modality by using the behavioral decision variable E and rand (a random number in the interval [0,1]).

[0035] ;

[0036] in, This represents the current iteration number. This represents the maximum number of iterations.

[0037] Based on E and rand, update X according to the following rules:

[0038] (a) Rule 1

[0039] When E > 1 and rand < 0.3, update X according to the following formula:

[0040] ;

[0041] in, Let i be the updated position vector of the i-th individual; This is the optimal position for the current population; The scaling factor controls the migration step size, with a value range of F∈[0.5,2], balancing exploration and exploitation; It is a random binary vector; A random number in the interval [0,1]. Let i be the current position vector of the i-th individual; It is a random integer from 1 to N.

[0042] (b) Rule 2

[0043] When E>1, rand When the value is 0.3, X is updated according to the following formula:

[0044] ;

[0045] in, It is a random number in the interval [0,1].

[0046] (c) Rule 3

[0047] When E≤1 and rand<0.5, X is updated according to the following formula:

[0048] ;

[0049] ;

[0050] ;

[0051] in, This represents the Euclidean distance from the current individual to the optimal position. A random number in the interval [0,1]. This is the step size factor.

[0052] (d) Rule 4

[0053] When E≤1, rand When the value is 0.5, X is updated according to the following formula:

[0054] ;

[0055] ;

[0056] in, For direction and step size adjustment; It is a symbolic function; This represents the current iteration number; This represents the maximum number of iterations. This is the function for generating random numbers for Levi's flight.

[0057] (24) Convergence judgment: If the number of iterations reaches Tmax (Tmax is a set value) or the objective function value changes by less than 10 −4 If the condition is met, the optimization will terminate, and the optimal parameter combination (Size) will be output. opt modenum opt Substitute the optimal parameter combination into step 1 to calculate the corresponding mode. .

[0058] Step 3: Calculate the local nonlinear characteristics of each mode in the wind speed time series.

[0059] like Figure 2 As shown, firstly, a convolution operation is performed on each obtained modality to obtain linear features:

[0060] ;

[0061] in, The number of channels in the input feature map. , , The input feature map contains the channel, height, and width indices. , , To output the channel, height, and width indices of the feature map, ∈ RC′×C×Kh×Kw The convolution kernel weight matrix is... , The height and width of the convolution kernel; ∈R C′ It is the bias vector; These are linear eigenvectors.

[0062] Subsequently, the linear features are transformed using a nonlinear activation function:

[0063] ;

[0064] in, The number of basis functions. As basis function weights, These are learnable basis functions.

[0065] Step 4: Calculate the long-term temporal dependency features of each mode in the wind speed time series.

[0066] For each modal time t feature obtained in step 2 ∈R d (d is the input dimension), and the hidden state at time t-1. ∈R m (where m is the dimension of the hidden layers), the calculation process of the recurrent network is as follows:

[0067] (e) Update gate z t

[0068] Controlling the degree of influence of the hidden state in the previous moment on the current moment:

[0069] ;

[0070] in, It is the sigmoid activation function; ∈R m×(m+d) To update the gate weight matrix; ∈R m It is the bias vector; This represents the concatenation operation between the hidden state and the input features.

[0071] (f) Reset the door

[0072] Control the degree to which the hidden state from the previous moment is retained:

[0073] ;

[0074] in, ∈R m×(m+d) To reset the gate weight matrix; ∈R m This is the bias vector.

[0075] (g) Candidate hidden state

[0076] Calculate the candidate hidden state at the current time based on the output of the reset gate:

[0077] ;

[0078] in, It is the hyperbolic tangent activation function; ∈R m×(m+d) The candidate hidden state weight matrix; ∈R m It is the bias vector; This represents element-wise multiplication of a matrix.

[0079] (h) Final hidden state

[0080] By combining the update gate and candidate hidden states, the hidden state at the current moment is generated:

[0081] ;

[0082] in, ∈R m The hidden state at time t contains temporal dependency information of the input sequence.

[0083] Step 5: Integrate local nonlinear features with long-term temporal dependence features.

[0084] The output features F3 and F4 from steps 3 and 4 are concatenated to generate a fused feature vector. ∈R d1+d2 :

[0085] ;

[0086] Here, [;] represents the concatenation operation of feature vectors.

[0087] Step 6: Wind speed prediction output.

[0088] The fused features are input into the fully connected layer to obtain the final wind speed prediction. ∈R 1 :

[0089] ;

[0090] in, ∈R 1×(d1+d2) This is the output layer weight matrix. ∈ R1 This is a bias term.

[0091] To evaluate the performance of the method, this embodiment uses mean squared error, root mean square error, mean absolute error, and coefficient of determination as indicators to assess the method's performance. The smaller the values ​​of mean squared error, root mean square error, and mean absolute error, the better the method's prediction effect, and the closer the predicted value is to the actual value; while the closer the coefficient of determination is to 1, the higher the goodness of fit of the method, and the better the prediction effect.

[0092] This example compares the prediction results of Method 1 of this invention with those of Methods 2 (Temporal Convolutional Network - Bidirectional Long Short-Term Memory - Attention), 3 (Convolutional Neural Network - Bidirectional Long Short-Term Memory), 4 (Bidirectional Gated Recurrent Unit - Attention), 5 (Transformation - Bidirectional Long Short-Term Memory), and 6 (Transformation - Bidirectional Gated Recurrent Unit), conducted on the same test dataset. The methods used in the comparison are the most current mainstream methods. Table 1 shows a comparison of the performance metrics of these prediction models. The method of this invention significantly outperforms the control group in four metrics: mean squared error (0.04357), root mean square error (0.20873), mean absolute error (0.13367), and coefficient of determination (0.96792). For example, compared to Method 6, the method of this invention reduces the mean squared error by 79.2%, the root mean square error by 54.4%, the mean absolute error by 56.4%, and improves the coefficient of determination by 14.4%. The comparison demonstrates that the method of this invention has higher accuracy and stability than the other comparative methods.

[0093] Table 1 Comparison of performance metrics of prediction models

[0094] .

Claims

1. A wind speed prediction method, characterized in that, Includes the following steps: (1) The wind speed time series is adaptively decomposed into multiple modal components, and the modal length and number of modes in the decomposition process are dynamically determined by an adaptive optimization algorithm. The adaptive decomposition is as follows: a parameter space is established with the modal length and number of modes as optimization variables; the optimization objective is to minimize the complexity of each modal component after decomposition, and a multi-strategy collaborative optimization algorithm is used for parameter search; during the optimization process, four update rules are dynamically switched according to the population state and random decision variables. The four update rules include local search guided by the current best individual, exploration guided by random individuals, spiral approximation strategy and Levy flight random walk mechanism; when the convergence condition is met, the optimal parameter combination is output and used for decomposition. (2) For each modal component, calculate its local nonlinear characteristics and extract the nonlinear components through a nonlinear mapping structure containing learnable basis functions; for each modal component, calculate its long-term temporal dependency characteristics and capture the temporal dependency relationship through an improved gated recurrent unit structure. (3) The local nonlinear features of the same modal component are fused with the long-term time-dependent features. The feature fusion is as follows: the local nonlinear feature vector of the same modal component is concatenated with the long-term time-dependent feature vector to form a fused feature vector; the fused features of each modal component are input into the fully connected layer for regression prediction to obtain the final wind speed prediction value. (4) Wind speed prediction output based on fusion features.

2. The wind speed prediction method according to claim 1, characterized in that, In step (2), the calculation of local nonlinear features is as follows: perform convolution operation on the modal components to extract linear features; perform nonlinear mapping on the linear features through a set of learnable basis functions, which can adaptively capture nonlinear relationships.

3. The wind speed prediction method according to claim 1, characterized in that, In step (2), the long-term temporal dependency features are calculated as follows: a gated loop unit structure is adopted, including update gate, reset gate and candidate hidden state calculation; the retention and updating of historical information are controlled by the gating mechanism to achieve effective modeling of long-term temporal dependencies.

4. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method according to any one of claims 1-3.

5. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the method described in any one of claims 1-3.

Images