Short-term wind power prediction method based on secondary decomposition and hybrid deep neural network

Abstract

The invention discloses a short-term wind power prediction method based on secondary decomposition and a hybrid deep neural network. The short-term wind power prediction method comprises the following steps: firstly, acquiring original wind power data and performing primary modal decomposition; performing K-means clustering to obtain three IMF components including a high frequency IMF component, an intermediate frequency IMF component and a low frequency IMF component; carrying out secondary modal decomposition on the high-frequency IMF component through a VMD algorithm; respectively inputting the stationary subsequences after VMD decomposition and the intermediate frequency and low frequency components of the K-means clustering result into a neural network for combined model prediction; hyper-parameters of the neural network are optimized through an SBOA algorithm, and the network is prevented from falling into a local optimal solution; and superposing the predicted values of all the subsequences to obtain a final prediction result of the short-term wind power. According to the method, the combined prediction model based on quadratic mode decomposition is adopted, and the advantages of various single models can be fully played, so that the overall prediction performance is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction in power systems, and particularly to a short-term wind power prediction method based on quadratic decomposition and hybrid deep neural network. Background Art

[0002] Wind power prediction is crucial for power grid scheduling and load balancing. Especially in areas with a large proportion of wind power, accurate short-term wind power prediction helps grid operators schedule backup power in advance or adjust grid load, thereby reducing the uncertainty of power supply and enhancing the stability and security of the power grid. Traditional single prediction models often struggle to fully capture the complex correlations between various influencing factors, resulting in limited prediction accuracy. Summary of the Invention

[0003] The purpose of the present invention is to provide a short-term wind power prediction method based on quadratic decomposition and hybrid deep neural network, which adopts a combined prediction model based on quadratic modal decomposition and SBOA-TCN-BiGRU-Attention, and can give full play to the advantages of various single models, thereby improving the overall prediction performance.

[0004] In view of the high volatility of wind power signals, the original signal is decomposed into multiple components to reduce the volatility of the original time series and effectively extract the main features in the signal.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A short-term wind power prediction method based on quadratic decomposition and hybrid deep neural network, which includes the following steps:

[0007] Step 1, obtain the original wind power data and perform primary modal decomposition, and cluster it into 3 IMF components of high frequency, medium frequency and low frequency through K-means clustering;

[0008] Furthermore, in Step 1, the CEEMDAN algorithm is used to obtain the original wind power data and perform primary modal decomposition.

[0009] After the primary decomposition, there will still be noise and high-frequency components. In order to further extract useful signal features and reduce the interference of noise on the prediction model, the high-frequency IMF component is subjected to secondary modal decomposition by the VMD algorithm to obtain a stationary subsequence;

[0010] Furthermore, in Step 2, the high-frequency IMF component is subjected to secondary modal decomposition by the VMD algorithm.

[0011] Step 3, construct a neural network model; the neural network includes a Temporal Convolutional Network (TCN), a Bidirectional Gated Recurrent Unit (BiGRU), and an attention mechanism network. Among them, the Temporal Convolutional Network (TCN) performs convolutional operations on the input wind power time series to extract global and long-term dependence features; the Bidirectional Gated Recurrent Unit (BiGRU) mines the forward and backward time dependencies of the sequence based on the global features extracted by the Temporal Convolutional Network (TCN) to capture local dynamic changes; the attention mechanism network sets different attention degrees for the output features of the Temporal Convolutional Network (TCN) and the Bidirectional Gated Recurrent Unit (BiGRU) respectively, so as to focus on key features and key time points, and further improve the prediction accuracy and robustness.

[0012] Step 4, input the stationary subsequence, intermediate frequency, and low-frequency components obtained by the second-order modal decomposition into the neural network model for prediction respectively, and use the SBOA algorithm to optimize the hyperparameters of the neural network to obtain the optimal network hyperparameters.

[0013] Furthermore, the Bidirectional Gated Recurrent Unit (BiGRU) in Step 3 includes a forward GRU unit and a backward GRU unit. The forward GRU unit processes data in the forward direction of the time series, and the backward GRU unit processes data in the reverse direction of the time series; at each time step, the hidden states output by the forward GRU unit and the backward GRU unit are connected and combined to form the final hidden state corresponding to the time step.

[0014] Furthermore, in Step 4, the SBOA algorithm takes the minimum average percentage error between the expected output and the actual output of the neural network as the fitness function to optimize the hyperparameters of the optimization target. The optimization target includes the learning rate, the number of iterations, and the number of neurons in the hidden layer. The SBOA algorithm optimizes the hyperparameters of the neural network to avoid the network falling into a local optimal solution.

[0015] Furthermore, the specific steps of the SBOA algorithm optimization in Step 4 are as follows:

[0016] Step 4-1, obtain the optimization target, and the optimization target includes the learning rate, the number of iterations, and the number of neurons in the hidden layer.

[0017] Step 4-2, initialize the neural network model and the SBOA parameters.

[0018] Step 4-3, generate an initial secretary bird population and input it into the neural network for prediction.

[0019] Step 4-4, calculate the fitness values of all current secretary birds respectively.

[0020] Step 4-5: Update the positions of all secretary birds and determine whether each agent goes out of bounds. If so, execute Step 3-6; otherwise, update the optimal solution and corresponding fitness of the current population and execute Step 3-6.

[0021] Step 4-6: Determine whether the current iteration number t is greater than the maximum iteration number TMAX. If so, output the best network hyperparameter result; otherwise, increment the iteration number by 1 and execute Step 4-4.

[0022] Step 5: Based on the neural network model with the best network hyperparameters, predict the values of each subsequence and superimpose them to obtain the final short-term wind power prediction result.

[0023] The present invention adopts the above technical solutions, and proposes to combine the secondary modal decomposition technology with the TCN-BiGRU-Attention deep learning model optimized by SBOA to form an overall prediction framework. In this combination, the secondary modal decomposition is used to decompose the complex wind power time series into multiple subsequences with clear frequency characteristics, while the deep network structures of TCN, BiGRU and the attention mechanism accurately model the temporal characteristics of each subsequence. The limitations of manual hyperparameter tuning are solved by SBOA to optimize the hyperparameters, and the prediction performance is improved. Description of the Drawings

[0024] The following further describes the present invention in detail with reference to the drawings and specific embodiments;

[0025] Figure 1 It is a schematic diagram of the structure of the bidirectional gated recurrent neural network BiGRU;

[0026] Figure 2 It is a schematic flow diagram of the short-term wind power prediction method based on secondary decomposition and hybrid deep neural network of the present invention;

[0027] Figure 3 It is a schematic diagram of the comparison between the method of the present invention and the prediction results of a single model. Detailed Embodiments

[0028] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.

[0029] As Figures 1 to 3 shown in one of them, the present invention discloses a short-term wind power prediction method based on secondary decomposition and hybrid deep neural network, which includes the following steps:

[0030] Step 1: Obtain the original wind power data and perform primary modal decomposition, and cluster it into 3 IMF components of high frequency, medium frequency and low frequency through K-means.

[0031] Further, in step 1, the CEEMDAN algorithm is used to obtain the original wind power data and perform the first modal decomposition.

[0032] In step 2, there will still be noise and high-frequency components after the first decomposition. In order to further extract useful signal features and reduce the interference of noise on the prediction model, the high-frequency IMF components are subjected to a second modal decomposition by the VMD algorithm to obtain stationary subsequences.

[0033] Further, in step 2, the high-frequency IMF components are subjected to a second modal decomposition by the VMD algorithm.

[0034] In step 3, a neural network model is constructed; the neural network includes a temporal convolutional network (TCN), a bidirectional gated recurrent unit (BiGRU), and an attention mechanism network; among them, the temporal convolutional network (TCN) performs a convolutional operation on the input wind power time series to extract global and long-term dependence features; based on the global features extracted by the temporal convolutional network (TCN), the bidirectional gated recurrent unit (BiGRU) mines the forward and backward time dependence relationships of the sequence, captures local dynamic changes, and enhances the ability to represent non-linear features; the attention mechanism network sets different degrees of attention for the output features of the temporal convolutional network (TCN) and the bidirectional gated recurrent unit (BiGRU) respectively, so as to focus on key features and key time points, and further improve the prediction accuracy and robustness.

[0035] In step 4, the stationary subsequences, intermediate-frequency and low-frequency components obtained by the second modal decomposition are respectively input into the neural network model for prediction, and the SBOA algorithm is used to optimize the hyperparameters of the neural network to obtain the best network hyperparameters.

[0036] Further, in step 3, the bidirectional gated recurrent unit (BiGRU) includes a forward GRU unit and a backward GRU unit. The forward GRU unit processes data in the forward direction of the time series, and the backward GRU unit processes data in the reverse direction of the time series; at each time step, the hidden states output by the forward GRU unit and the backward GRU unit are connected together to form the final hidden state corresponding to the time step.

[0037] Further, in step 4, the SBOA algorithm is used to optimize the hyperparameters of the neural network to avoid the network falling into a local optimal solution.

[0038] Further, in step 4, the SBOA algorithm optimizes the hyperparameters of the optimization target with the minimum average percentage error between the expected output and the actual output of the neural network as the fitness function. The optimization targets include the learning rate, the number of iterations, and the number of neurons in the hidden layer.

[0039] Further, the specific steps of the SBOA algorithm optimization in step 4 are as follows:

[0040] Step 4-1: Obtain the target to be optimized, where the optimization targets include the learning rate, the number of iterations, and the number of neurons in the hidden layer;

[0041] Step 4-2: Initialize the neural network model and the SBOA parameters;

[0042] Step 4-3: Generate the initial secretary bird population and input it into the neural network for prediction;

[0043] Step 4-4: Calculate the fitness values of all current secretary birds respectively;

[0044] Step 4-5: Update the positions of all secretary birds and determine whether each agent goes out of bounds; if so, execute Step 3-6; otherwise, update the optimal solution and the corresponding fitness of the current population and execute Step 3-6;

[0045] Step 4-6: Determine whether the current iteration number t is greater than the maximum iteration number TMAX; if so, output the best network hyperparameter result; otherwise, increment the iteration number by 1 and execute Step 4-4.

[0046] Step 5: Based on the neural network model with the best network hyperparameters, predict the values of each subsequence, and superimpose them to obtain the final short-term wind power prediction result.

[0047] The specific principle of the present invention is described in detail as follows:

[0048] The original wind power is affected by environmental factors such as wind speed, and has characteristics such as volatility and randomness, resulting in low accuracy of direct prediction. The present invention performs secondary modal decomposition on the original wind power data to improve the accuracy and stability of wind power prediction. First, the CEEMDAN algorithm is used to perform primary modal decomposition on the original wind power signal. Then, the K-means clustering algorithm is used to divide each mode after the primary modal decomposition into high-frequency, medium-frequency, and low-frequency signals according to the sample entropy. Finally, the high-frequency signal is subjected to secondary modal decomposition by the VMD algorithm to further reduce the noise and interference in the high-frequency signal, which is more conducive to extracting the features of the signal and further improving the prediction accuracy.

[0049] The SBOA-TCN-BiGRU-Attention model adopted by the neural network model; TCN uses dilated convolutions to insert holes between convolutional kernels, effectively capturing long-term dependencies without increasing the computational complexity, which is very important for modeling long-term wind speed and meteorological data in wind power prediction. At the same time, TCN uses causal convolutions to ensure that the output of each time step only depends on the current and previous time steps, avoiding the use of future information, thus making the prediction more in line with the actual situation. Since TCN can process time series data in parallel and adapt to the large-scale characteristics of wind power data, the training speed is improved. TCN introduces residual connections, allowing the network to directly transmit information across layers, reducing the problem of gradient vanishing and enhancing the training efficiency of the model. In short-term wind power prediction, TCN improves the prediction accuracy and the training speed of the model.

[0050] As Figure 1 shown, the BiGRU model is a recurrent neural network, which consists of two independent GRU units, one processes data forward according to the time series, and the other processes data backward according to the time series. For each time step, the forward GRU and the backward GRU will both output a hidden state. These two hidden states are combined together through a simple connection to form the final hidden state of this time step.

[0051] The core idea of the attention mechanism is to focus the attention on the most important information in the input data and ignore the unimportant information. When processing each input, the model sets different attention degrees for different parts of the input data through the learned weight information.

[0052] For the TCN-BiGRU-Attention combined prediction model, first, use TCN to perform convolutional operations on the input wind power time series to extract global and long-term dependency features, providing a basis for subsequent modeling. Then, based on the global features extracted by TCN, use BiGRU to mine the forward and backward time dependencies of the sequence, capture local dynamic changes, and enhance the representation ability of non-linear features. Subsequently, combining the output features of TCN and BiGRU, focus on key features and key time points through the attention mechanism to further improve the prediction accuracy and robustness.

[0053] As Figure 2As shown, the SBOA algorithm simulates the hunting and escape strategies of the secretary bird. The efficient response ability demonstrated by the secretary bird when facing food and threats inspired the development of this algorithm in optimization problems. The prediction accuracy of the TCN-BiGRU-Attention network is affected by the learning rate, the number of neurons in BiGRU, the key value of the attention mechanism, and the regularization parameter. The present invention uses SBOA to optimize the TCN-BiGRU-Attention network. In order to determine a set of network hyperparameters that can minimize the error of the TCN-BiGRU-Attention network, the minimum average percentage error between the expected output and the actual output of the TCN-BiGRU-Attention network is used as the fitness function in the present invention. The optimization objectives include the learning rate size, the number of iterations, and the number of neurons in the hidden layer. The flow of the SBOA-TCN-BiGRU-Attention network.

[0054] Effect experiment: The data used in the present invention is the measured data of a wind farm in a certain area, and the sampling time is from March 1, 2019 to March 31, 2019. The sampling interval is 15 minutes, and there are a total of 2,976 data points. The data set is divided into a test set and a training set according to 1:30, and 96 sample points on March 31 are used as test samples. This example is implemented based on Matlab.

[0055] In the experiment of the present invention, the root mean square error (RMSE), the mean absolute error (MAE), and the correlation coefficient are used as the evaluation indicators for wind power prediction.

[0056] To verify the effectiveness of the combined model of the present invention, first, the method of the present invention is compared with the currently common single prediction models, such as Figure 3 As shown, the single prediction model can roughly predict the wind power, but the accuracy is not as high as the combined prediction model proposed by the present invention. As shown in Table 1, compared with the RMSE and MAE of the single model, the method of the present invention has at least improved by 4.6401 MW and 3.7221 MW, and at most improved by 5.4514 MW and 4.5824 MW. And the correlation coefficient of the method of the present invention is closest to 1, and the fitting degree is the highest, which reflects the superiority of the combined prediction method based on signal decomposition. This is because decomposing the original wind power signal can reduce the interference of noise, improve the accuracy of the training result, and be closer to the actual value.

[0057] Table 1 Comparison of error indicators between the method of the present invention and the single model

[0058] Model RMSE (MW) MAE (WM) <![CDATA[R 2 (%)]]> BIGRU 7.6444 6.2750 95.9960％ LSSVM 6.8331 5.4329 96.7683％ TCN 7.4639 6.2932 96.1540％ BP 7.1668 5.5613 96.4591％ The method of the present invention 2.193 1.7108 99.6688％

[0059] Secondly, to verify the superiority of the method of the present invention in the combined model, a comparative experiment was conducted between CEEMDAN-VMD-SBOA-TCN-BiGRU-Attention and other five combined models. As can be seen from Figure 3 the prediction results of each combined model shown, the prediction result of the method of the present invention has the highest coincidence degree with the actual value and the best prediction result. Through the comparative experiment, CEEMDAN-VMD-TCN-BiGRU-Attention was compared with the CEEMDAN-VMD-CNN-LSTM and CEEMDAN-VMD-CNN-BiLSTM models. It can be analyzed from Table 2 that since TCN and BiGRU usually have a relatively high model complexity and can better adapt to the complex wind power time series, the generalization ability of the model is improved. In particular, the Attention mechanism enables the model to pay more attention to important features, thereby improving the prediction accuracy.

[0060] Next, to verify the superiority of the secondary mode decomposition relative to the primary mode decomposition, the present invention compared CEEMDAN-VMD-TCN-BiGRU-Attention with the VMD-TCN-BiGRU-Attention and CEEMDAN-VMD-TCN-BiGRU-Attention models. It can be seen from the analysis of Table 2 that the prediction accuracy of the secondary mode decomposition is higher than that of the two primary mode decomposition methods.

[0061] Table 2 Comparison of error indexes of combined models

[0062] Model RMSE (MW) MAE (WM) <![CDATA[R 2 (%)]]> CEEMDAN-VMD-CNN-LSTM 7.5234 6.2357 96.1679％ CEEMDAN-VMD-CNN-BiLSTM 4.9563 4.0098 98.3025％ VMD-TCN-BiGRU-Attention 4.1387 3.3458 98.7022％ CEEMDAN-TCN-BiGRU-Attention 3.9517 3.3503 98.9003％ CEEMDAN-VMD-TCN-BiGRU-Attention 3.7635 2.9009 98.9052％ The method of the present invention 2.193 1.7108 99.6688％

[0063] In addition, by comparing CEEMDAN-VMD-TCN-BiGRU-Attention with the method proposed by the present invention, the RMSE and MAE of the prediction results of the SBOA algorithm are reduced by 41.74% and 41.04% respectively, while it is increased by 0.77%. The results show that after adding the SBOA algorithm, the prediction effect is significantly improved, and the advantages of the TCN-BiGRU-Attention model in the short-term prediction of wind power are fully exerted.

[0064] The present invention combines the secondary mode decomposition technology with the TCN-BiGRU-Attention deep learning model optimized by SBOA to form an overall prediction framework. In this combination, the secondary mode decomposition is used to decompose the complex wind power time series into multiple sub-sequences with clear frequency characteristics, while the deep network structures of TCN, BiGRU and the attention mechanism accurately model the temporal characteristics of each sub-sequence. The SBOA is used to optimize the hyperparameters to solve the limitations of manual parameter tuning and improve the prediction performance.

[0065] Obviously, the described embodiments are some, but not all, of the embodiments of this application. Without conflict, the embodiments in this application and the features in the embodiments may be combined with each other. Generally, the components of the embodiments of this application described and illustrated in the accompanying drawings here may be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts fall within the scope of protection of this application.

Claims

1. A short-term wind power forecasting method based on quadratic decomposition and hybrid deep neural network, characterized by: It includes the following steps: Step 1: Obtain the original wind power data and perform the initial modal decomposition, and cluster it into three IMF components of high frequency, medium frequency and low frequency through K-means; Step 2, perform secondary modal decomposition on the high-frequency IMF components to obtain a stationary subsequence; Step 3, construct a neural network model; the neural network includes a temporal convolutional network TCN, a bidirectional gated recurrent neural network BiGRU and an attention mechanism network; the temporal convolutional network TCN performs a convolution operation on the input wind power time series to extract global and long-term dependency features; the bidirectional gated recurrent neural network BiGRU mines the forward and backward time dependencies of the sequence based on the global features extracted by the temporal convolutional network TCN to capture local dynamic changes; the attention mechanism network sets different attention levels for the output features of the temporal convolutional network TCN and the bidirectional gated recurrent neural network BiGRU, respectively, so as to focus on key features and key time points; Step 4: Decompose the secondary modal into stable subsequences, intermediate frequency components and low frequency components and input them into the neural network model for prediction. Then use the SBOA algorithm to optimize the hyperparameters of the neural network to obtain the best network hyperparameters. Step 5: The neural network model based on the optimal network hyperparameters predicts the predicted values ​​of each subsequence, and superimposes them to obtain the final prediction result of short-term wind power.

2. The short-term wind power forecasting method based on quadratic decomposition and hybrid deep neural network according to claim 1 is characterized by: In step 1, the CEEMDAN algorithm is used to obtain the original wind power data and perform the initial modal decomposition.

3. The short-term wind power forecasting method based on quadratic decomposition and hybrid deep neural network according to claim 1 is characterized by: In step 2, the VMD algorithm is used to perform secondary modal decomposition on the high-frequency IMF components.

4. The short-term wind power forecasting method based on quadratic decomposition and hybrid deep neural network according to claim 1 is characterized by: In step 3, the bidirectional gated recurrent neural network BiGRU includes a forward GRU unit and a backward GRU unit. The forward GRU unit processes data forward according to the time series, and the backward GRU unit processes data in reverse according to the time series. The hidden states output by the forward GRU unit and the backward GRU unit in each time step are connected and combined to form the final hidden state of the corresponding time step.

5. The short-term wind power forecasting method based on quadratic decomposition and hybrid deep neural network according to claim 1 is characterized by: In step 4, the SBOA algorithm is used to optimize the hyperparameters of the optimization target by minimizing the average percentage error between the expected output and the actual output of the neural network as the fitness function. The optimization target includes the learning rate, the number of iterations, and the number of neurons in the hidden layer.

6. The short-term wind power forecasting method based on quadratic decomposition and hybrid deep neural network according to any one of claim 1, characterized in that: The specific steps of SBOA algorithm optimization in step 4 are: Step 4-1, obtain the target to be optimized, the optimization target includes the learning rate, the number of iterations and the number of neurons in the hidden layer; Step 4-2, initializing the neural network model and SBOA parameters; Step 4-3, generate the initial secretary bird population and input it into the neural network for prediction; Step 4-4, calculate the fitness values ​​of all current secretary birds respectively; Step 4-5, update the positions of all secretary birds and determine whether each agent is out of bounds; if so, execute steps 3-6; otherwise, update the optimal solution and corresponding fitness of the current population and execute steps 3-6; Step 4-6, determine whether the current number of iterations t is greater than the maximum number of iterations TMAX; if so, output the optimal network hyperparameter result; otherwise, add 1 to the number of iterations and execute step 4-4.

Images