Wind speed short-term prediction method based on multi-objective optimization and sub-model selection

Abstract

The invention relates to the technical field of wind power generation of a wind power plant, in particular to a wind speed short-term prediction method based on multi-objective optimization and sub-model selection, which comprises the following steps of: eliminating and repairing abnormal data and noise data in a wind speed data sample by considering the existence of the abnormal data and the noise data in wind speed time sequence data. According to the method, phase-space reconstruction, prediction model selection and parameter synchronous optimization are realized by using an improved multi-objective optimization algorithm, so that relatively high prediction precision and prediction stability of the prediction model are realized, and meanwhile, the generalization ability of the prediction model is improved.

Description

Technical Field

[0001] The invention relates to the technical field of wind farm wind power generation, and in particular to a wind speed short-term prediction method based on multi-objective optimization and sub-model selection. Background Art

[0002] Since wind speed time series data from different wind farms and different time periods have different characteristics, a single wind speed prediction model cannot obtain high prediction accuracy for different wind speed samples. In addition, the selection of prediction models in traditional wind speed prediction methods lacks objectivity. For some wind speed time series samples with high linearity and low volatility, the prediction accuracy of statistical prediction models ARIMA and Kalman is relatively high, while for wind speed time series data with high nonlinearity and strong randomness, artificial intelligence methods such as extreme learning machines ELM, SVM and LSSVM and artificial neural networks ANN have better prediction performance. Due to the diversity of wind speed time series, it is difficult for a single prediction model to accurately describe the nonlinear characteristics of different wind speed data.

[0003] Traditional wind speed prediction methods use signal preprocessing technology to decompose different subsequences obtained from wind speed data, and artificially select a single prediction model to predict all subsequences. However, different subsequences have different frequency characteristics, and a single prediction model is often not fully suitable for simulating all subsequence characteristics. Subjectively selecting a single prediction method is often not fully suitable for simulating all subsequence characteristics, nor can it adapt to different scenarios.

[0004] It is difficult for a single prediction model to accurately describe the nonlinear characteristics of different wind speed data, there is a risk of decision-making errors when selecting a single prediction model, and the stability of the wind speed combination prediction model is as important as the wind speed prediction accuracy. A short-term wind speed prediction model based on a multi-objective optimization algorithm and sub-model selection is proposed, which takes into account both model prediction accuracy and prediction stability. In addition, in order to objectively select each sub-model, a short-term wind speed prediction method based on multi-objective optimization and sub-model selection is needed to solve this problem. Summary of the invention

[0005] The purpose of the present invention is to provide a method for short-term wind speed prediction based on multi-objective optimization and sub-model selection to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for short-term wind speed prediction based on multi-objective optimization and sub-model selection, comprising the following steps: S1. Considering the existence of abnormal data and noise data in wind speed time series data, the abnormal data and noise data in wind speed data samples are removed and repaired; S2. Based on the step S1, the wind speed time series data has strong randomness and volatility, the wind speed sample data is normalized, the normalization interval is [0,1], and the sample data set is dynamically randomly divided into a training set and a test set; S3. Based on the training data set and the test data set divided in steps S1 and S2, the wind speed data time series is decomposed into multiple subsequence components with different frequencies and smaller amplitudes based on the variational mode decomposition technique; S4. The optimal number of decomposition layers of the variational modal decomposition technique of step S3 is determined by the center frequency method. The number of decomposed wind speed subsequence components can be determined by obtaining the optimal number of decomposition layers of the variational modal decomposition technique; S5. Based on the decomposition of the training data set and the test data set obtained in step S4, the input variable matrix of the prediction method is constructed using the phase space reconstruction technique; S6. Based on the steps S4 and S5, the input variable matrix is ​​obtained to perform wind speed prediction; a suitable prediction method for each decomposition subsequence is selected from the wind speed method set, and wind speed prediction is performed using the suitable prediction method; S7. Based on the step S6, the wind speed is predicted for different decomposed subsequences, and the improved bat optimization algorithm (chaos enhanced bat algorithm) is used to take the prediction accuracy and the stability of the prediction model as the objective function, and the phase space reconstruction parameters and the internal parameters of the prediction method are synchronously optimized and selected; S8. Accumulating the wind speed decomposition subsequence prediction results based on steps S6 and S7, and adding the prediction value of each time node of the decomposition subsequence to obtain the wind speed prediction value of the time node; S9. Evaluate the wind speed prediction results of step S8, analyze the wind speed prediction results using statistical error analysis indicators, and compare them with the prediction benchmark method and the recently developed typical wind speed prediction method.

[0007] S10. Based on the chaos enhanced bat optimization algorithm, the optimal decomposition method for selecting each decomposition subsequence in the wind speed prediction method set is given, and a wind speed short-term prediction method based on multi-objective optimization and sub-model selection is proposed.

[0008] As a preferred technical solution of the present invention, step S1 includes the following sub-steps: S1.1 Assume that the wind speed time series set {( x i , y i )∈ R 2 | i =0,1, ∙∙, N}, then the difference function of the data set can be expressed asf :[ x 0 , x N ]→ R , then: ( x i , y‘ i )∈ R 2 Defined as fractal interpolation points, x i Point interpolation error is as follows: Fractal dimension value D It is related to the correlation degree of the wind speed time series in the interpolation space, and is used to measure the irregular parameters of the fractal interpolation wind speed time series. The fractal dimension value D and the vertical scale factor d n The relationship is as follows.

[0009] Where: a i =( x i - x i-1 ) / ( x N - x 0 ), d i Represents the vertical scale factor, reflecting the local characteristics of the interpolation curve.

[0010] S1.2 The relative root mean square error (RMSE) and Nash efficiency coefficient (NSE) are used to evaluate the interpolation results of the wind speed time series. NSE is used to indicate the degree of consistency between the fractal interpolation results and the actual wind speed time series, such as: Where: is fractal interpolation, The larger the NSE value is, the higher the degree of agreement between the fractal interpolation and the measured wind speed value is.

[0011] As a preferred technical solution of the present invention, step S2 includes the following sub-steps: S2.1 According to the following formula, the input sample data set Normalization is performed, and the normalization interval is [0,1]: Where: is the normalized data, is the wind speed data sample, is the minimum value in the sample data, is the maximum value in the sample data; S2.2 The sample data set is calculated according to the following formula Perform dynamic random ratio division into training set and test set; The training set is: The test set is: Where: is the cardinality of set P, that is, the number of elements in set P, Generate a random number greater than 0 and less than 1, function Truncate the decimal part to the integer.

[0012] As a preferred technical solution of the present invention, step S3 includes the following sub-steps: S3.1 The augmented Lagrange function without constraints is used to solve the variational mode function, as shown in the formula: S3.2 Determination of the total number of modal functions by the center frequency method K And the parameters are obtained based on the residual minimization method t , such as: In the formula, N is the length of the decomposed signal, u k is a modal function.

[0013] As a preferred technical solution of the present invention, step S5 includes the following sub-steps: S5.1 Assume that the wind speed time series is X = { x i | i =1,2,∙∙∙, N}, it can be reconstructed into a multidimensional phase space, as follows: According to the above formula, the relationship between the phase space reconstruction matrix of the wind speed time series X and its predicted output variable Y is as follows: Each row represents a vector, when determining the phase space reconstruction parameters m and t , you can reconstruct ( N -( m -1) t ) vectors. If the number of wind speed prediction time series is L , then the required initial data is L + ( m −1) t The key issue in using the PSR method to reconstruct the phase space of wind speed time series is how to select the parameters m and t . Use bat intelligent algorithm to optimize and select phase space reconstruction parameters m and t , wind speed is predicted in multiple steps through recursive method. The predicted value is used as the new input phase point to construct the input variable matrix of the prediction model, and the wind speed prediction result is obtained after multiple recursive cycles.

[0014] As a preferred technical solution of the present invention, step S6 includes the following sub-steps: S6.1 Optimize the selection of a prediction method from the prediction method set (ARIMA, ANFIS, BPNN, ELM, LSSVM, LSTM, SVM and WNN) based on the comprehensive evaluation index (CEI) for the subsequence IMF i Make predictions; S6.2 When the CEI of a prediction model for a subsequence prediction value is the smallest, the prediction method is the optimal sub-prediction method.

[0015] Where: MAE i * , MAPE i * , RMSE i * and SDE i * Respectively i The prediction results of the sub-methods MAE , MAPE , RMSE and SDE The standard normalized index values ​​are: In the formula, N The amount of data.

[0016] As a preferred technical solution of the present invention, step S7 includes the following sub-steps: S7.1 In order to improve the population diversity and traversability of the standard bat algorithm, chaotic mapping is also introduced into the bat algorithm. The bat algorithm population is initialized using a logical self-mapping function with good traversal ability. The logical self-mapping function is as follows: In the formula, L i d ∈(-1, 1).

[0017] According to the logical self-mapping function, the individual is transformed to [-1,1] in the bat algorithm, as shown in the formula: Then initialize the population individuals as: In the formula b i d and a i d Represent the maximum and minimum values ​​respectively.

[0018] Pulse frequency and loudness A i All of them are monotonic. In the heuristic bat optimization algorithm, this variable monotonicity is not conducive to generating feasible solution diversity. The chaotic sine mapping function is used to replace the standard bat algorithm emission pulse frequency. r i and loudness A i ; S7.2 In order to prevent the algorithm from falling into the local optimum, the dynamic disturbance coefficient is introduced into the current individual position, and the individual position of the current population is disturbed, so that the individual jumps out of its position and enters the neighborhood space, expanding the space for the bat algorithm to search for feasible solutions and enhancing the flexibility and search ability of the bat algorithm in global optimization. Dynamic disturbance coefficient r and location updates x i ( t+ 1) respectively as follows: In the formula, t andT max are the current iteration number and the maximum iteration number respectively.

[0019] As the number of iterations t Increase and adaptively increase. In the early stage of iteration, the disturbance amplitude of the current bat position is large, which reduces x i ( t )- x * The constraints on the spatial search of individuals in the population towards a wider solution promote the individuals to enter the adjacent area and enhance the global search ability of the bat algorithm; as the number of iterations increases and enters the later stage of iteration, the disturbance amplitude of the current population is reduced, which enhances the bat algorithm's ability to deeply search for local optimal solutions.

[0020] S7.3 In order to fully explore the local search capability of the algorithm, the uniform random perturbation variable λ is changed to tangent random exploration, that is: In the formula, r is(0 , 1) Uniformly distributed random numbers, tan (( r -0.5)× π ) is a tangent random variable. From the relationship between the tangent random variable and the iteration, we can see that it has two main characteristics: (1) Most of the tan (( r -0.5)× π ) values ​​are distributed around 0, which makes the algorithm search near the current optimal solution, enhancing the algorithm's local "exploitation" ability; (2) very few tan (( r -0.5)× π ) value is large, which can enable the algorithm to jump out of the local optimal solution and enter other areas to search for the best solution, making the local optimization strategy of the bat algorithm more flexible instead of blindly searching randomly.

[0021] S7.4 Introduce the average position method into the standard bat algorithm, and use the average position of the individual population to guide the individual bats. In the early stage of the algorithm, the average position of the individual population is large. As the iteration increases, the algorithm gradually converges and the average position gradually decreases. Using the average position of the population to adjust the variable step length can improve the optimization ability of the algorithm in the early stage of iteration, and accelerate the convergence of the algorithm in the later stage of iteration, which can effectively reduce the possibility of the algorithm falling into premature local optimality. The speed update is as follows: In the formula, N and Dare the population size and population dimension respectively.

[0022] S7.5 According to claim 7, based on the chaos-enhanced bat algorithm and the sub-model selection strategy, with prediction accuracy and prediction model stability as the goals, the parameters of the prediction model are optimized and selected to improve the prediction accuracy and stability of the prediction model, as well as the generalization ability of the prediction model.

[0023] As a preferred technical solution of the present invention, step S8 includes the following sub-steps: S8.1 Assume that the wind speed data is X , using the signal processing method OVMD to decompose the wind speed data into X d1 , X d2 , X d3 ∙∙∙ X dN , then the prediction model of each subsequence is as follows: In the formula, for i The predicted value of the subsequence, f Represents the prediction model. The final wind speed prediction result is as follows: As a preferred technical solution of the present invention, step S9 includes the following sub-steps: S9.1 In order to evaluate the deterministic prediction results of the wind speed prediction model, the relative root mean square error (RMSE), mean absolute percentage error (MAPE), mean absolute error (MAE), normalized mean absolute proportional error (MASE), variance (Std) and Pearson correlation coefficient (R) are used to calculate the prediction error, as shown in the following formula: In the formula, N is the number of wind speed samples; and are the measured wind speed value and the predicted wind speed value, respectively. M is the mean wind speed, M =( v 1 ,v 2,v 3 ,…, v N ) / N .

[0024] As a preferred technical solution of the present invention, step S10 includes the following sub-steps: S10.1 The wind speed prediction strategy based on "decomposition-prediction-reconstruction" is a typical collaborative combination method. This method decomposes the wind speed time series into a group of subsequences with different characteristics, uses intelligent algorithms to model each subsequence separately, and makes predictions. According to the sub-model selection strategy, the optimal sub-model of each sub-sequence is selected for time series prediction, and the prediction values ​​of different prediction sub-models are integrated using a collaborative combination method to obtain the final wind speed prediction result. Among them, the predictions of each sub-sequence are independent of each other, but the prediction processes of each sub-sequence are collaborative. Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention is a short-term wind speed prediction method based on multi-objective optimization and sub-model selection. The present invention also uses an improved multi-objective optimization algorithm to realize phase space reconstruction and prediction model selection and simultaneous optimization of its parameters, so as to achieve higher prediction model prediction accuracy and prediction stability, while improving the generalization ability of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solution in the prior art of the present invention, the drawings required for use in the specification will be briefly introduced below.

[0026] Figure 1 is a schematic diagram of a structure in which a dynamic disturbance coefficient varies with an iteration coefficient according to the present invention; Figure 2 is a graph of the relationship between the tangent random variable and iteration according to the present invention; Figure 3 It is a working principle diagram of decomposition-prediction-reconstruction according to the present invention; Figure 4 is a graph of a sinusoidal mapping function and an exponential function according to the present invention; Figure 5 According to the present invention d n Schematic diagram of the relationship between fractal dimension and Figure 6 According to the present invention d n Schematic diagram of the relationship between RMSE; Figure 7 It is a wind speed prediction model diagram solved by the multi-objective optimization bat algorithm according to the present invention. DETAILED DESCRIPTION

[0027] Below, combined with the attached Figure 1-7 And specific implementation methods, further describe the invention: A sub-model selection strategy according to an embodiment of the present invention selects an optimal prediction method from a prediction method set, the method comprising the following steps: S1. Considering the different locations of wind turbines in wind farms and different wind speeds, the historical data of each wind turbine containing wind speed data is obtained as a sample set; affected by factors such as electromagnetic interference, signal acquisition sensors, information transmission channels and the reliability of information receiving devices, the missing wind speed time series data and the existence of abnormal noise data are difficult to avoid, and there is even a period of missing wind speed data, which brings many troubles to wind farm wind power scheduling, unit maintenance planning and operation management; the fractal interpolation algorithm is used to identify abnormal noise and interpolation calculations for wind speed time series data, and the chaos enhanced bat algorithm is used to adaptively optimize the vertical scale factor in the selection algorithm to improve the computational performance of the fractal interpolation algorithm.

[0028] Further, the step S1 includes the following sub-steps: In the fractal interpolation process, the vertical scale factor and the number of iterations need to be adjusted. For example, 288 data sets with 10-minute intervals from 00:00:00 on April 10, 2019 to 24:00:00 on April 11, 2019 at a wind farm in Weihai, Shandong Province, my country are used, and 50 continuous wind speed time series are selected from the data sets as the initial points of interpolation.

[0029] 1) Vertical scale factor Vertical scale factor d i The value directly affects the fractal dimension D . The relationship between the vertical scale factor and the fractal dimension increases with | d n | Increase, fractal dimension D The larger the value is, the more volatile the fractal interpolation fitting curve of the wind speed time series is. The RMSE of the interpolation curve error varies with | d n There is a nonlinear relationship between d n ∈[-0.1,0.126], RMSE d n Increase gradually decrease, d n ∈[0.126, 0.25], RMSE increases with d n Increase gradually, d n = 0.126 is the inflection point of the fractal interpolation curve error. Before the inflection point, RMSE decreases with dn The RMSE decreases gradually as the inflection point increases. d n Increase gradually, inflection point d n = The fractal interpolation dimension corresponding to 0.126 is the optimal interpolation dimension. For different wind speed time series data, there are different vertical scale factors. d i Corresponding to the optimal fractal interpolation dimension, this paper uses the chaos enhanced bat algorithm (CEBA) to optimize the selection of the vertical scale factor d i , using the vertical scale factor stored in the individual positions of the population in CEBA, taking the root mean square of the interpolation error (RMSE) as the fitness function, the global optimal solution is obtained through multiple iterations of the algorithm, that is, minutes [ RMSE ( d 1 ,d 2 ,d 3 ,…,d n ) and | d i |<1.

[0030] 2) Number of iterations Fractal interpolation points m With the number of iterations n The relationship between them is as follows: (2-1) In the formula, L Indicates the number of initial interpolation points.

[0031] The fractal interpolation results after multiple iterations are shown in the table. As can be seen from the table, different numbers of iterations will affect the fractal dimension. D , the number of fractal interpolation points, the interpolation error, and the running time of the algorithm. As the number of iterations increases, the fractal dimension increases when the number of iterations is 2, 3, and 4. D The value and interpolation error RMSE remain basically unchanged, that is, the intensity of the fractal interpolation curve fluctuation and the interpolation error are basically not affected by the change of the number of iterations. However, as the number of iterations increases, the number of interpolation data points increases exponentially, which greatly increases the running time.

[0032] Table Statistical indicators of multiple iterative fractal interpolation The number of fractal interpolation iterations only needs to meet the actual interpolation requirements. According to the above analysis, the number of interpolation iterations is selected as 2.

[0033] S2. Based on the step S1, the wind speed time series data is repaired to have strong randomness and volatility, the wind speed sample data is normalized, the normalization interval is [0,1], and the sample data set is dynamically randomly divided into a training set and a test set; Based on the training data set and the test data set divided in steps S1 and S2, the wind speed data time series is decomposed into a plurality of subsequence components with different frequencies and smaller amplitudes based on the variational mode decomposition technique; Further, the step S2 includes the following sub-steps: Variational mode decomposition technology decomposes each mode function in the process of wind speed data u k and center frequency oh k Iterate and update continuously in the frequency domain, and then use Fourier transform technology to transform the frequency domain { }Transform to time domain{ u k ( t )}, its working process is: 1) Initialize the modal function { }、Center frequency{ }、Lagrange multiplication operator and the number of iterations n ; 2) Order n = n +1, execute loop iteration; 3) Update the mode function and center frequency respectively; 4) Order k = k +1 Repeat step 3) to update the modal function and center frequency until k = K ; 5) Follow Update the Lagrange multiplication operator λ; 6) If Below the set threshold i When VMD ends, we get K natural mode components; otherwise, execute 2).

[0034] Determining the total number of mode functions using the center frequency method K And the parameters are obtained based on the residual minimization method t , such as: In the formula, N is the length of the decomposed signal, u k is a modal function.

[0035] S4. The optimal number of decomposition layers of the variational modal decomposition technique in step S3 is determined by the center frequency method. The number of decomposition layers of the wind speed subsequence components can be determined by obtaining the optimal number of decomposition layers of the variational modal decomposition technique. The center frequency method optimizes the selection of the decomposition layer of the variational modal decomposition as K , then we get K Group wind speed decomposition subsequences; S5. According to the wind speed subsequence obtained in step S4, phase space reconstruction is used to reconstruct the phase space of each decomposed subsequence to obtain an input variable matrix, and the chaos enhanced bat algorithm is used to optimize and select phase space parameters; S6. The wind speed prediction strategy based on "decomposition-prediction-reconstruction" is a typical collaborative combination method. This method decomposes the wind speed time series into a group of subsequences with different characteristics, uses intelligent algorithms to model each subsequence separately, and makes predictions. According to the submodel selection strategy, the optimal submodel of each subsequence is selected for time series prediction, and the prediction values ​​of different prediction submodels are integrated using a collaborative combination method to obtain the final wind speed prediction result. Among them, the predictions of each subsequence are independent of each other, but the prediction processes of each subsequence are collaborative: Wind speed combination prediction method Assume the wind speed data is X , using the signal processing method OVMD to decompose the wind speed data into X d1 , X d2 , X d3 ∙∙∙ X dN , then the prediction model of each subsequence is as follows: In the formula, for i The predicted value of the subsequence, f Represents the prediction model. The final wind speed prediction result is as follows: S7. Based on the step S6, the wind speed is predicted for different decomposed subsequences, and the improved bat optimization algorithm (chaos enhanced bat algorithm) is used to take the prediction accuracy and the stability of the prediction model as the objective function, and the phase space reconstruction parameters and the internal parameters of the prediction method are synchronously optimized and selected; By improving the population diversity and search ability of the standard bat algorithm, the chaotic mapping strategy, dynamic perturbation coefficient, tangent random optimization mechanism and the average position of the population individuals are introduced into the standard bat algorithm, namely the chaos enhanced bat algorithm (CEBA), whose pseudo code is as shown in the algorithm. Wind speed prediction model based on SMOBA solution 8. Based on step S7, the wind speed is predicted for different decomposed subsequences, and the predicted values ​​of different subsequences at each moment are accumulated to obtain the final prediction result: Assume the wind speed data is X , using the signal processing method OVMD to decompose the wind speed data into X d1 , X d2 , X d3 ∙∙∙ X dN , then the prediction results of each subsequence are as follows: In the formula, for i The predicted value of the subsequence, f Represents the subsequence prediction method. The final wind speed prediction result is as follows: S9. Based on the wind speed prediction result of step S8, the statistical indicators RMSE, MAPE, MAE, MASE, and Std are used for evaluation, and the wind speed prediction result is compared with the prediction benchmark method and the recently developed typical wind speed prediction method using statistical error analysis indicators.

[0036] In this embodiment, since wind speed time series data of different wind farms and different time periods have different characteristics, a single wind speed prediction model cannot obtain high prediction accuracy when predicting different wind speed samples. For some wind speed time series samples with high linearity and low volatility, the prediction accuracy of the statistical prediction models ARIMA and Kalman is relatively high, while for wind speed time series data with high nonlinearity and strong randomness, artificial intelligence methods such as extreme learning machine (ELM), SVM and LSSVM and artificial neural network (ANN) have better prediction performance. For example, for some wind farm wind speed data, ARIMA prediction accuracy is higher than ELM prediction accuracy, while the results of testing other wind speed data samples are just the opposite. ELM, SVM and WNN use different characteristic basis functions to capture nonlinear information in wind speed time series data. In some literatures, the prediction accuracy of wavelet neural network (WNN) is higher than that of ELM in some cases, while the prediction accuracy of ELM is higher than that of WNN in some cases. Due to the diversity of wind speed time series, it is difficult for a single prediction model to accurately describe the nonlinear characteristics of different wind speed data. Most of the literature uses signal preprocessing technology to decompose the different subsequences obtained from the wind speed data, and then uses a certain prediction model to predict all subsequences. However, different subsequences have different frequency characteristics, and a single prediction model is often not fully suitable for simulating all subsequence characteristics.

[0037] In view of the fact that it is difficult for a single prediction model to accurately describe the nonlinear characteristics of different wind speed data, the risk of decision-making errors when selecting a single prediction model, and the fact that the stability of the wind speed combination prediction model is as important as the wind speed prediction accuracy, this chapter proposes a short-term wind speed prediction model based on multi-objective optimization and sub-model selection. This chapter mainly focuses on the following aspects: 1) Construct a short-term wind speed prediction model based on the "decomposition-prediction-reconstruction" strategy: First, the wind speed data samples are preprocessed using fractal interpolation algorithm, variational mode decomposition and phase space reconstruction technology to construct the input matrix of the wind speed decomposition subsequence; secondly, different prediction methods are used to predict each wind speed decomposition subsequence; finally, the prediction results of each sub-model are accumulated to obtain the final wind speed prediction value; 2) Parameter solution based on SMOBA multi-objective optimization: Optimizing the selection of sub-models and their parameters with the goal of prediction accuracy and model stability can not only obtain high-precision and stable wind speed predictions at the same time, but also overcome the weaknesses of single-objective optimization algorithms; 3) The sub-model selection strategy is used to optimize the selection of the best sub-model suitable for each sub-sequence from the sub-model set: first, different prediction sub-models, including statistical models and artificial intelligence models, are used to predict each wind speed decomposition sub-series; then, the sub-model selection strategy is used to determine the best prediction model for each wind speed decomposition sub-sequence to ensure the effectiveness of the overall wind speed prediction.

[0038] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for short-term wind speed prediction based on multi-objective optimization and sub-model selection, characterized in that: The following steps are involved: S1. Considering the existence of abnormal data and noise data in wind speed time series data, the abnormal data and noise data in wind speed data samples are removed and repaired; S2. Based on the step S1, the wind speed time series data has strong randomness and volatility, the wind speed sample data is normalized, the normalization interval is [0,1], and the sample data set is dynamically randomly divided into a training set and a test set; S3. Based on the training data set and the test data set divided in steps S1 and S2, the wind speed data time series is decomposed into multiple subsequence components with different frequencies and smaller amplitudes based on the variational mode decomposition technique; S4. The optimal number of decomposition layers of the variational modal decomposition technique of step S3 is determined by the center frequency method. The number of decomposed wind speed subsequence components can be determined by obtaining the optimal number of decomposition layers of the variational modal decomposition technique; S5. Based on the decomposition of the training data set and the test data set obtained in step S4, the input variable matrix of the prediction method is constructed using the phase space reconstruction technique; S6. Based on the input variable matrix obtained in steps S4 and S5, wind speed prediction is performed; a suitable prediction method for each decomposed subsequence is selected from the wind speed method set, and wind speed prediction is performed using the suitable prediction method; S7. Based on the step S6, the wind speed is predicted for different decomposed subsequences, and the improved bat optimization algorithm (chaos enhanced bat algorithm) is used to take the prediction accuracy and the stability of the prediction model as the objective function, and the phase space reconstruction parameters and the internal parameters of the prediction method are synchronously optimized and selected; S8. Accumulating the prediction results of the wind speed decomposition subsequences based on the steps S6 and S7, and adding the prediction values ​​of each time node of the decomposition subsequence to obtain the wind speed prediction value of the time node; S9. Evaluate the wind speed prediction results of step S8, analyze the wind speed prediction results using statistical error analysis indicators, and compare them with the prediction benchmark method and the recently developed typical wind speed prediction method. S10. Based on the chaos enhanced bat optimization algorithm, the optimal decomposition method for selecting each decomposition subsequence in the wind speed prediction method set is given, and a wind speed short-term prediction method based on multi-objective optimization and sub-model selection is proposed.

2. The method for short-term wind speed prediction based on multi-objective optimization and sub-model selection according to claim 1, characterized in that: The step S1 comprises the following sub-steps: S1.1 Assume that the wind speed time series set {( x i , y i )∈ R 2 | i =0,1, ∙∙, N }, then the difference function of the data set can be expressed as f :[ x 0, x N ]→ R , then: ( x i , y‘ i )∈ R 2 Defined as fractal interpolation points, x i Point interpolation error is as follows: Fractal dimension value D It is related to the correlation degree of the wind speed time series in the interpolation space, and is used to measure the irregular parameters of the fractal interpolation wind speed time series. The fractal dimension value D and the vertical scale factor d n The relationship is as follows. Where: a i =( x i - x i-1 ) / ( x N - x 0 ), d i Represents the vertical scale factor, reflecting the local characteristics of the interpolation curve. S1.2 The relative root mean square error (RMSE) and Nash efficiency coefficient (NSE) are used to evaluate the interpolation results of the wind speed time series. NSE is used to indicate the degree of consistency between the fractal interpolation results and the actual wind speed time series, such as: Where: is fractal interpolation, The larger the NSE value is, the higher the degree of agreement between the fractal interpolation and the measured wind speed value is.

3. The method for short-term wind speed prediction based on multi-objective optimization and sub-model selection according to claim 1, characterized in that: The step S2 comprises the following sub-steps: S2.1 According to the following formula, the input sample data set Normalization is performed, and the normalization interval is [0,1]: Where: is the normalized data, is the wind speed data sample, is the minimum value in the sample data, is the maximum value in the sample data; S2.2 The sample data set is calculated according to the following formula Perform dynamic random ratio division into training set and test set; The training set is: The test set is: Where: is the cardinality of set P, that is, the number of elements in set P, Generate a random number greater than 0 and less than 1, function Truncate the decimal part to the integer.

4. The method for short-term wind speed prediction based on multi-objective optimization and sub-model selection according to claim 1, characterized in that: The step S3 comprises the following sub-steps: S3.1 The augmented Lagrange function without constraints is used to solve the variational mode function, as shown in the formula: S3.2 Determination of the total number of modal functions by the center frequency method K And the parameters are obtained based on the residual minimization method τ , such as: In the formula, N is the length of the decomposed signal, u k is a modal function.

5. The method for short-term wind speed prediction based on multi-objective optimization and sub-model selection according to claim 1, characterized in that: The step S5 comprises the following sub-steps: S5.1 Assume that the wind speed time series is X = { x i | i =1,2,∙∙∙, N }, it can be reconstructed into a multidimensional phase space, as follows: According to the above formula, the relationship between the phase space reconstruction matrix of the wind speed time series X and its predicted output variable Y is as follows: Each row represents a vector, when determining the phase space reconstruction parameters m and τ , you can reconstruct ( N -( m -1) τ ) vectors. If the number of wind speed prediction time series is L , then the required initial data is L + ( m −1) τ The key issue in using the PSR method to reconstruct the phase space of wind speed time series is how to select the parameters m and τ . Use bat intelligent algorithm to optimize and select phase space reconstruction parameters m and τ , multi-step wind speed prediction is performed through recursive method. The predicted values ​​are used as new input phase points to construct the input variable matrix of the prediction model. After multiple recursive cycles, the wind speed prediction results are obtained.

6. The method for short-term wind speed prediction based on multi-objective optimization and sub-model selection according to claim 1, characterized in that: The step S6 comprises the following sub-steps: S6.1 Optimize the selection of a prediction method from the prediction method set (ARIMA, ANFIS, BPNN, ELM, LSSVM, LSTM, SVM and WNN) based on the comprehensive evaluation index (CEI) for the subsequence IMF i Make predictions; S6.2 When the CEI of a prediction model for a subsequence prediction value is the smallest, the prediction method is the optimal sub-prediction method. Where: MAE i * , MAPE i * , RMSE i * and SDE i * Respectively i The prediction results of the sub-methods MAE , MAPE , RMSE and SDE The standard normalized index values ​​are: In the formula, N The amount of data.

7. The method for short-term wind speed prediction based on multi-objective optimization and sub-model selection according to claim 1, characterized in that: The step S7 comprises the following sub-steps: S7.1 In order to improve the population diversity and traversability of the standard bat algorithm, chaotic mapping is also introduced into the bat algorithm. The bat algorithm population is initialized using a logical self-mapping function with good traversal ability. The logical self-mapping function is as follows: In the formula, L i d ∈(-1, 1). According to the logical self-mapping function, the individual is transformed to [-1,1] in the bat algorithm, as shown in the formula: Then initialize the population individuals as: In the formula b i d and a i d Represent the maximum and minimum values ​​respectively. Pulse frequency and loudness A i All of them are monotonic. In the heuristic bat optimization algorithm, this variable monotonicity is not conducive to generating feasible solution diversity. The chaotic sine mapping function is used to replace the standard bat algorithm emission pulse frequency. r i and loudness A i , S7.2 In order to prevent the algorithm from falling into the local optimum, the dynamic disturbance coefficient is introduced into the current individual position, and the individual position of the current population is disturbed, so that the individual jumps out of its position and enters the neighborhood space, expanding the space for the bat algorithm to search for feasible solutions and enhancing the flexibility and search ability of the bat algorithm in global optimization. Dynamic disturbance coefficient ρ and location updates x i ( t+ 1) respectively as follows: In the formula, t and T max are the current iteration number and the maximum iteration number respectively. As the number of iterations t Increase and adaptively increase. In the early stage of iteration, the disturbance amplitude of the current bat position is large, which reduces x i ( t )- x * The constraints on the spatial search of individuals in the population towards a wider solution promote the individuals to enter the adjacent area and enhance the global search ability of the bat algorithm; as the number of iterations increases and enters the later stage of iteration, the disturbance amplitude of the current population is reduced, which enhances the bat algorithm's ability to deeply search for local optimal solutions. S7.3 In order to fully explore the local search capability of the algorithm, the uniform random perturbation variable λ is changed to tangent random exploration, that is: In the formula, r is(0 , 1) Uniformly distributed random numbers, tan (( r -0.5)× π ) is a tangent random variable. As shown in Figure 3.3, it has two main characteristics: (1) Most tan (( r -0.5)× π ) values ​​are distributed around 0, which makes the algorithm search near the current optimal solution, enhancing the algorithm's local "exploitation" ability; (2) very few tan (( r -0.5)× π ) value is large, which can enable the algorithm to jump out of the local optimal solution and enter other areas to search for the best solution, making the local optimization strategy of the bat algorithm more flexible instead of blindly searching randomly. S7.4 Introduce the average position method into the standard bat algorithm, and use the average position of the individual population to guide the individual bats. In the early stage of the algorithm, the average position of the individual population is large. As the iteration increases, the algorithm gradually converges and the average position gradually decreases. Using the average position of the population to adjust the variable step length can improve the optimization ability of the algorithm in the early stage of iteration, and accelerate the convergence of the algorithm in the later stage of iteration, which can effectively reduce the possibility of the algorithm falling into premature local optimality. The speed update is as follows: In the formula, N and D are the population size and population dimension respectively. S7.5 According to claim 7, based on the chaos-enhanced bat algorithm and the sub-model selection strategy, with prediction accuracy and prediction model stability as the goals, the parameters of the prediction model are optimized and selected to improve the prediction accuracy and stability of the prediction model, as well as the generalization ability of the prediction model.

8. The method for short-term wind speed prediction based on multi-objective optimization and sub-model selection according to claim 1, characterized in that: The step S8 comprises the following sub-steps: S8.1 Assume that the wind speed data is X , using the signal processing method OVMD to decompose the wind speed data into X d1 , X d2 , X d3 ∙∙∙ X dN , then the prediction model of each subsequence is as follows: In the formula, for i The predicted value of the subsequence, f Represents the prediction model. The final wind speed prediction result is as follows: 。 9. The method for short-term wind speed prediction based on multi-objective optimization and sub-model selection according to claim 1, characterized in that: The step S9 includes the following sub-steps: S9.1 In order to evaluate the deterministic prediction results of the wind speed prediction model, the relative root mean square error (RMSE), mean absolute percentage error (MAPE), mean absolute error (MAE), normalized mean absolute proportional error (MASE), variance (Std) and Pearson correlation coefficient (R) are used to calculate the prediction error, as shown in the following formula: In the formula, N is the number of wind speed samples; and are the measured wind speed value and the predicted wind speed value, respectively. M is the mean wind speed, M =( v 1 ,v 2 ,v 3 ,…, v N ) / N .

10. The method for short-term wind speed prediction based on multi-objective optimization and sub-model selection according to claim 1, characterized in that: The step S10 includes the following sub-steps: S10.1 The wind speed forecasting strategy based on "decomposition-prediction-reconstruction" is a typical collaborative combination method, which decomposes the wind speed time series into a group of subsequences with different characteristics, uses intelligent algorithms to model each subsequence separately, and makes predictions. According to the submodel selection strategy, the optimal submodel of each subsequence is selected for time series forecasting, and the forecast values ​​of different forecast submodels are integrated using the collaborative combination method to obtain the final wind speed forecast result, in which the forecasts of each subsequence are independent of each other, but the forecast processes of each subsequence are collaborative.

Images