Short-term wind power prediction method and device based on multi-module integration

Abstract

The invention discloses a short-term wind power prediction method based on multi-module integration. The method comprises the following steps: collecting meteorological data and historical power data of a target wind power plant to form a data set; calculating the correlation between the meteorological data and the historical wind power data, and inputting the features with strong correlation into a short-term wind power prediction model; the model ICEEMDAN layer decomposes the historical power into a plurality of modal components; combining the gas phase data with the gas phase data in the input variable to form a characteristic matrix; the BiTCN layer extracts feature information of the feature matrix; the BiGRU layer captures feature information time sequence dependence, and the MHSA layer carries out weight distribution on time sequence state information; and superposing the prediction value of each modal component output by the dense layer to obtain an overall prediction result. The device comprises a data acquisition unit, a feature selection unit, a sequence decomposition unit and a short-term wind power training and prediction unit. According to the invention, the cooperative work of multiple modules can fully mine the nonlinear relationship between meteorological data and wind power, and the prediction capability of the model under complex meteorological conditions is enhanced.

Description

Technical field:

[0001] The present invention relates to the technical field of wind power prediction, and in particular to a method and device for short-term wind power prediction based on multi-module integration. Background technology:

[0002] With the intensification of global warming and the increasing attention to ecological environment protection and sustainable energy supply, the development and utilization of renewable energy has become a key direction for global scientific research and exploration and the development of the energy industry. As a clean, pollution-free renewable energy, wind power has gradually improved its position in the energy field with its significant green and environmentally friendly characteristics, and has become one of the important ways to replace traditional fossil energy. Large-scale wind power grid connection has put forward more stringent requirements on the stability of the power system. However, meteorological factors such as wind speed, wind direction, temperature, and humidity have a significant impact on wind power, resulting in wind power having the characteristics of randomness, volatility, and intermittency. This makes it difficult to accurately predict the changes in wind power, which brings many challenges to the dispatching and operation of the power grid. Therefore, achieving high-precision short-term wind power forecasting has become a key link in optimizing the dispatching of the power system and ensuring the stable operation of the power grid.

[0003] At present, the commonly used wind power prediction methods are mainly divided into two categories: physical model-based and data-driven model-based. The physical model method relies on wind turbine characteristics and detailed meteorological parameters, and predicts by constructing a physical relationship between meteorological factors and wind power. However, this method has a complex calculation process and is highly dependent on the accuracy of wind turbine parameters. Once there is an error in the wind turbine parameters or the meteorological data is not accurate enough, the accuracy of the prediction results will be greatly affected.

[0004] The method based on data-driven model directly establishes the mapping relationship between meteorological factors and wind power with the help of machine learning or deep learning technology. This type of method is relatively simple to model and has strong adaptability to different scenarios. However, due to the significant non-stationarity of wind power series and the extremely complex and changeable meteorological conditions, traditional data-driven models are difficult to accurately capture the multi-scale characteristics of wind power when processing such data. Especially under extreme weather conditions, the accuracy and robustness of the prediction model are obviously insufficient. In addition, wind power data may also be subject to mutations or strong noise interference. Models that rely solely on mapping relationships have limited adaptability in such complex environments, resulting in poor stability of the prediction results.

[0005] In summary, existing technologies have obvious shortcomings in dealing with problems such as non-stationary data, complex meteorological conditions and power mutations. Summary of the invention:

[0006] In view of the above problems, the present invention provides a method and device for short-term wind power prediction based on multi-module integration, aiming to improve the accuracy and adaptability of short-term wind power prediction.

[0007] A method for short-term wind power prediction based on multi-module integration, the method specifically comprising:

[0008] Step 1: Collect meteorological data and historical power data of the target wind farm to form a data set, wherein the meteorological data is a feature of the historical wind power data; wherein the meteorological data includes wind speed, wind direction, temperature, humidity, air temperature and air pressure at different heights; pre-process the data set: sort out the historical wind power data and remove bad data;

[0009] Step 2: Calculate the correlation between the meteorological data and the historical wind power data, select the features with strong correlation as input variables and input them into the short-term wind power prediction model, wherein the short-term wind power prediction model includes an improved adaptive noise complete empirical mode decomposition algorithm ICEEMDAN layer, a bidirectional temporal convolutional network BiTCN layer, a bidirectional gated recurrent unit BiGRU layer, a multi-head self-attention mechanism MHSA layer and a dense layer;

[0010] Step 3: The ICEEMDAN layer decomposes the historical wind power data into multiple modal components of different complexity; combines the modal components with the gas phase data in the input variables to form a feature matrix; the feature matrix is ​​used as the input of the BiTCN layer to extract feature information;

[0011] Step 4: The feature information is fed to the BiGRU layer to capture the timing dependency. After the MHSA layer weights the time series state information output by the BiGRU layer, the dense layer maps the output of the MHSA layer to the output space through a nonlinear transformation to obtain the predicted value of each historical wind power modal component. The historical wind power modal components are added together to obtain the overall prediction result.

[0012] Preferably, the correlation between the meteorological data and the historical wind power data is calculated using the Pearson correlation coefficient (PCC), wherein the calculation method of the PCC is as follows:

[0013]

[0014] Among them, r XY The range is [-1,1]. The larger its absolute value, the greater the correlation between X and Y.

[0015] Preferably, the BiTCN layer uses extended convolution to expand the receptive field, and introduces a residual block in the BiTCN layer.

[0016] Preferably, the specific method of the ICEEMDAN layer decomposing the historical wind power data into multiple modal components with different complexity is:

[0017] Add white noise E to the historical wind power data X 1 [ω i ], generate auxiliary noise signal X (i) :X (i) =X+β i *E 1 [ω i ], where ω i is the i-th white noise added, β i is the signal-to-noise ratio;

[0018] The modal components of each order are extracted through iterative calculation until the energy of the residual signal is lower than the preset threshold or cannot be further decomposed. Specifically, the hth modal component is:

[0019]

[0020] Among them, R h is the hth order residue, h=1,2,3,…,N, is the hth modal component, M(﹒) is the local average of the generated modal component, and I is the number of noise disturbances.

[0021] Preferably, the method further comprises evaluating the prediction results using mean absolute error (MAE), root mean square error (RMSE), and R square.

[0022]

[0023] Where n is the number of prediction results, y i is the actual value, is the predicted value.

[0024] Preferably, the meteorological data and historical wind power data of the wind farm are obtained through the SCADA system; the method of eliminating bad data by data preprocessing is data cleaning and data interpolation, and the bad data includes missing data, erroneous data, duplicate data and data beyond a reasonable range.

[0025] Preferably, the data set needs to be normalized after preprocessing, and the specific method is as follows:

[0026]

[0027] Among them, Y i is the normalized data, Y is the unnormalized data, and Y min is the minimum value in the input sequence, Y maxis the maximum value in the input sequence.

[0028] Preferably, the dense layer implements nonlinear transformation through an activation function, and generally, the activation function is sigmoid.

[0029] A device for short-term wind power prediction based on multi-module integration, the device comprising:

[0030] Data acquisition unit: used to acquire historical meteorological data of the wind farm and corresponding wind power data to form a data set, and divide the data set into a training set and a test set;

[0031] Feature selection unit: used to calculate the correlation between meteorological data features and historical wind power data using the Pearson correlation coefficient, and select features with stronger correlation as input variables of the short-term wind power prediction model;

[0032] Sequence decomposition unit: used to decompose the historical wind power data with non-smooth characteristics into multiple modal components with different complexity using the ICEEMDAN layer;

[0033] Short-term wind power training and prediction unit: used to use the training set short-term wind power prediction model; and input the test set into the short-term wind power prediction model for prediction, and summarize the prediction results at each moment into the final short-term wind power prediction result.

[0034] The present invention designs a method and device for short-term wind power prediction based on multi-module integration, integrates ICEEMDAN, BiTCN, BiGRU and MHSA multi-modules, and uses ICEEMDAN to decompose the original wind power data with non-smooth characteristics into multiple modal components with different complexity. This decomposition method can reduce the complexity of the data, separate the components of different frequencies and trends in the original data, so that the subsequent model can capture the data characteristics more clearly, thereby effectively processing the non-stationarity of the wind power sequence. The bidirectional time convolution structure (BiTCN) is used to capture the hidden features of the forward and reverse directions, and the long-term dependency of the wind power sequence can be better obtained. BiTCN uses extended convolution to achieve a larger receptive field with fewer layers while maintaining the dimension of the feature map, which helps to capture abnormal situations such as power mutations. At the same time, the introduction of residual blocks avoids the problems of gradient vanishing and slow convergence caused by the increase in the receptive field, ensuring that the model can still efficiently extract features when processing data containing power mutations. In addition, the multi-head self-attention mechanism assigns weights to the time series state information output by the BiGRU layer, which can highlight the impact of relevant information at the moment of power mutation on the prediction results, making the model more sensitive to power mutations and more accurate in prediction. The collaborative work of multiple modules of the present invention can fully explore the complex nonlinear relationship between meteorological data and wind power, and enhance the prediction ability of the model under complex meteorological conditions. At the same time, the present invention calculates the correlation between meteorological data features and historical wind power data through the Pearson correlation coefficient (PCC), and selects features with strong correlation as input variables of the model. This operation can screen out meteorological factors that have a greater impact on wind power, eliminate redundant information, focus the model on key factors, and better adapt to complex meteorological conditions. Description of the drawings:

[0035] Attached Picture 1 Flow chart of the short-term wind power prediction method based on multi-module integration of Example 1 of the present invention. Picture 2 This is a diagram of the sequence decomposition results of ICEEMDAN of Example 1 of the present invention.

[0036] Attached Picture 3 This is a comparison chart of wind power prediction results of different models in Example 1 of the present invention. Specific implementation method:

[0037] In order to make the technical solution of the present invention easier to understand, a method and device for short-term wind power prediction based on multi-module integration disclosed in the present invention are now clearly and completely described in combination with embodiments and drawings.

[0038] like Picture 1 As shown, a method and device for short-term wind power prediction based on multi-module integration, the method specifically includes:

[0039] Step 100: The data source is an open source 200MW wind farm. The SCADA system measures and saves wind speed, wind direction, temperature, humidity, air temperature, air pressure and historical wind power data at 10 / 30 / 50 meters every 15 minutes. The collected data set covers the period from January 1, 2019 to December 31, 2019.

[0040] Step 110: pre-process the data collected in step 100: remove bad data from the collected historical power data. The method of removing bad data is data cleaning and data interpolation. The bad data includes missing data, erroneous data, duplicate data and data beyond a reasonable range. Normalize all data. The specific method is: Among them, Y i is the normalized data, Y is the unnormalized data, and Y min is the minimum value of the input sequence China, Y max is the maximum value in the input sequence.

[0041] Step 120: Calculate the correlation between the meteorological data and the historical wind power data through PCC, select the features with strong correlation as input variables and input them into the short-term wind power prediction model (ICEEMDAN-BiTCN-BiGRU-MSHA), wherein the short-term wind power prediction model includes an improved adaptive noise complete empirical mode decomposition algorithm ICEEMDAN layer, a bidirectional temporal convolutional network BiTCN layer, a bidirectional gated recurrent unit BiGRU layer, a multi-head self-attention mechanism MHSA layer and a dense layer;

[0042] The calculation method of PCC is as follows:

[0043]

[0044] Among them, r XY The range is [-1,1]. The larger its absolute value, the greater the correlation between X and Y.

[0045] Since the traditional TCN only performs forward convolution calculation on the input sequence, it only extracts the forward features of the wind power data and ignores the backward implicit information. Therefore, the present invention adopts a bidirectional temporal convolution structure (BiTCN) to capture the forward and backward hidden features to better obtain the long-term dependency of the wind power sequence.

[0046] The causal nature of causal convolution is characterized by the fact that the input sequence x = (x 1 ,x 2 ,…,x t ) is output as y=(y 1 ,y2 ,…,y t ), and the output y at each moment t Only depends on the input before the current time point. In order to effectively capture the long-term features of the sequence, causal convolution requires more layers or a larger receptive field. Therefore, BiTCN uses dilated convolution to achieve a larger receptive field with fewer layers while maintaining the dimension of the feature map. For input x∈R n One-dimensional sequence and convolution kernel filter: {0,…,k-1}→R, the dilated convolution of element s in the sequence is defined as:

[0047]

[0048] Where k is the size of the convolution kernel, s-di is the past direction, and d is the dilation factor, which determines how many zero vectors are inserted between two adjacent convolution kernels. After each layer of convolution of the input sequence, d grows exponentially. After several convolutions, BiTCN obtains a larger receptive field. However, the increase in the receptive field of BiTCN brings problems such as gradient vanishing and slow convergence. The introduction of residual blocks can avoid these problems while achieving efficient feature extraction of the sequence.

[0049] Since the traditional GRU algorithm can only propagate in one direction along the sequence, it only considers the correlation between the current moment and the past moment. In order to simultaneously consider the impact of past and future moment data on the current data, the BiGRU architecture is constructed. BiGRU is based on the original GRU, and another GRU architecture with the opposite direction is constructed for each group of sequences. Finally, after weighting the hidden layers of the two groups of GRU, the outputs of the two groups of GRU are merged. The expression is as follows:

[0050]

[0051] Where, X t For input, are the outputs of the forward and backward hidden layers at time t; w t and v t Represent the output weights of the forward and backward hidden layers respectively; h t is the output of the two hidden layers after merging at time t; b t is the bias coefficient.

[0052] The attention mechanism is a resource allocation mechanism that processes more important information under limited computing resources. Let Q and (K, V) = [(k 1 ,v 1 ),…,(k n ,v n)] represent the query vector and key-value pair related to the task respectively. In the key-value pair, the key vector K is used to calculate the attention distribution, and the value vector v is used to calculate the aggregate information. By inputting data X, the corresponding Q, K, and V are calculated as: Q = XW Q , K = XW K , V = XW V ;

[0053] Where W Q , W K , W V ∈R D*dv is a trainable parameter. By paying attention to the score function and the input dimension D, Q and K are used to calculate the attention weight distribution α. ​​Then, using α and V, the final comprehensive attention information force Att is calculated:

[0054] In order to simultaneously focus on useful information from different representation subspaces, a multi-head mechanism is applied to the self-attention model to form a multiple-head self-attention (MS) mechanism. Its calculation process is expressed as:

[0055]

[0056] Where Att i =Att(Q i ,K i ,V i ), d v =D / h, where h is the number of heads.

[0057] Step 130: The ICEEMDAN layer decomposes the historical wind power data into multiple modal components of different complexity; combines the modal components with the gas phase data in the input variables to form a feature matrix; the feature matrix is ​​used as the input of the BiTCN layer to extract feature information, and the specific method is as follows:

[0058] Add white noise E to the historical wind power data X 1 [ω i ], generate auxiliary noise signal X (i) :X (i) =X+β 0 *E 1 [ω i ], where ω i is the i-th white noise added, β 0 is the signal-to-noise ratio;

[0059] The modal components of each order are extracted through iterative calculation until the energy of the residual signal is lower than the preset threshold or cannot be further decomposed. Specifically, the hth modal component is:

[0060]

[0061] Among them, R h is the hth order residue, h=1,2,3,…,N, is the hth modal component, M(﹒) is the local average of the generated modal component, and I is the number of noise disturbances.

[0062] Step 140: The feature information is fed to the BiGRU layer to capture the temporal dependency, and the multi-head self-attention mechanism is used to weight the time series state information output by the BiGRU layer.

[0063] Step 150: The dense layer uses an activation function to map the features of the MHSA layer to reallocate weights to the output space through a nonlinear transformation to obtain the predicted values ​​of each historical wind power modal component, and the historical wind power modal components are added together to obtain an overall prediction result. In order to reduce the risk of overfitting, the model uses Dropout regularization, that is, some units are randomly disabled during the training phase.

[0064] In order to verify the effectiveness of the improvement of the present invention, MAE, RMSE and R 2 The three evaluation indicators are used to calculate the prediction errors of 9 models, including the BiTCN model, ICEEMDAN-BiTCN model, ICEEMDAN-BiTCN-A model, BiGRU model, ICEEMDAN-BiGRU model, ICEEMDAN-BiGRU-A model, BiTCN-BiGRU model, ICEEMDAN-BiTCN-BiGRU model and ICEEMDAN-BiTCN-BiGRU-A model. The evaluation results are shown in Table 1.

[0065] Table 1 Comparison of prediction errors of different models

[0066] Kyoto model MAE RMSE <![CDATA[R 2 ]]> BiTCN 4.9068 8.6617 0.9826 ICEEMDAN-BiTCN 4.6995 8.3720 0.9837 ICEEMDAN-BiTCN-A 4.5369 8.0116 0.9851 BiGRU 4.7661 8.3863 0.9837 ICEEMDAN-BiGRU 4.4936 8.0186 0.9851 ICEEMDAN-BiGRU-A 4.3701 7.9015 0.9855 BiTCN-BiGRU 4.4734 8.0808 0.9849 ICEEMDAN-BiTCN-BiGRU 4.2664 7.8371 0.9858 ICEEMDAN-BiTCN-BiGRU-A 3.9142 7.1002 0.9883

[0067] The experimental results show that the model of the present invention shows the best performance in all evaluation indicators. As shown in Table 1, the model combined with the BiTCN-BiGRU network significantly improves the prediction accuracy. Compared with the traditional model (BiTCN or BiGRU used alone), the proposed model has an RMSE of 7.10, which is significantly better than BiTCN (RMSE 8.01) and BiGRU (RMSE 7.90) used alone. The wind power data is decomposed into multiple intrinsic modal components (IMFs) by applying the ICEEMDAN layer, which improves the prediction accuracy of the model. The experimental results show that the MAE is reduced from 4.47 to 4.27, which proves the effectiveness of ICEEMDAN decomposition. After the introduction of the attention mechanism, the prediction accuracy of the model is further improved. The performance of the ICEEMDAN-BiTCN-BiGRU-MHSA model is better than that of BiTCN-BiGRU and ICEEMDAN-BiTCN-BiGRU, verifying the contribution of the attention mechanism to the model accuracy.

[0068] from Picture 2 It can be seen that ICEEMDAN decomposition obtains 15 groups of IMF components and one group of RES components. The sequences are arranged in order from high to low frequency, and the fluctuations of sequences with different frequencies have a certain regularity, which avoids mode aliasing.

[0069] Picture 3 The predicted values ​​and true values ​​of each comparison method are given. The closer the predicted value is to the true value, the higher the prediction accuracy of the method.

[0070] The present invention also provides a device for short-term wind power prediction based on multi-module integration, the device comprising:

[0071] Data acquisition unit: used to acquire historical meteorological data of the wind farm and corresponding wind power data to form a data set, and divide the data set into a training set and a test set;

[0072] Feature selection unit: used to calculate the correlation between meteorological data features and historical wind power data using the Pearson correlation coefficient, and select features with stronger correlation as input variables of the short-term wind power prediction model;

[0073] Sequence decomposition unit: used to decompose the historical wind power data with non-smooth characteristics into multiple modal components with different complexity using the ICEEMDAN layer;

[0074] Short-term wind power training and prediction unit: used to use the training set short-term wind power prediction model; and input the test set into the short-term wind power prediction model for prediction, and summarize the prediction results at each moment into the final short-term wind power prediction result.

[0075] It should be pointed out that for ordinary technicians in this technical field, several improvements, substitutions, modifications and embellishments can be made without departing from the principles and purpose of the present invention. These improvements, substitutions, modifications and embellishments should also be regarded as the scope of protection of the present invention.

Claims

1. A method for short-term wind power forecasting based on multi-module integration, characterized in that: The method specifically comprises: Step 1: Collect meteorological data and historical power data of the target wind farm to form a data set, wherein the meteorological data is a feature of the historical wind power data; wherein the meteorological data includes wind speed, wind direction, temperature, humidity, air temperature and air pressure at different heights; pre-process the data set: sort out the historical wind power data and remove bad data; Step 2: Calculate the correlation between the meteorological data and the historical wind power data, select the features with strong correlation as input variables and input them into the short-term wind power prediction model, wherein the short-term wind power prediction model includes an improved adaptive noise complete empirical mode decomposition algorithm ICEEMDAN layer, a bidirectional temporal convolutional network BiTCN layer, a bidirectional gated recurrent unit BiGRU layer, a multi-head self-attention mechanism MHSA layer and a dense layer; Step 3: The ICEEMDAN layer decomposes the historical wind power data into multiple modal components of different complexity; combines the modal components with the gas phase data in the input variables to form a feature matrix; the feature matrix is ​​used as the input of the BiTCN layer to extract feature information; Step 4: The feature information is fed to the BiGRU layer to capture the timing dependency. After the MHSA layer weights the time series state information output by the BiGRU layer, the dense layer maps the output of the MHSA layer to the output space through a nonlinear transformation to obtain the predicted value of each historical wind power modal component. The historical wind power modal components are added together to obtain the overall prediction result.

2. A method for short-term wind power prediction based on multi-module integration as claimed in claim 1, characterized in that: The correlation between the meteorological data and the historical wind power data is calculated using the Pearson correlation coefficient.

3. The method for short-term wind power prediction based on multi-module integration according to claim 1, characterized in that: The BiTCN layer uses dilated convolution to expand the receptive field, and introduces a residual block in the BiTCN layer.

4. The method for short-term wind power prediction based on multi-module integration according to claim 1, characterized in that: The specific method of the ICEEMDAN layer to decompose the historical wind power data into multiple modal components with different complexity is: Add white noise E1[ω i ], generate auxiliary noise signal X (i) :X (i) =X+β i *E1[ω i ], where ω i is the i-th white noise added, β i is the signal-to-noise ratio; The modal components of each order are extracted through iterative calculation until the energy of the residual signal is lower than the preset threshold or cannot be further decomposed. Specifically, the hth modal component is: Among them, R h is the hth order residue, h=1,2,3,…,N, is the hth modal component, M(﹒) is the local average of the generated modal component, and I is the number of noise disturbances.

5. The method for short-term wind power prediction based on multi-module integration according to claim 1, characterized in that: The method also includes evaluating the prediction results using mean absolute error (MAE), root mean square error (RMSE), and R square.

6. The method for short-term wind power prediction based on multi-module integration according to claim 1, characterized in that: The meteorological data and historical wind power data of the wind farm are obtained through the SCADA system; the method of data preprocessing to eliminate bad data is data cleaning and data interpolation, and the bad data includes missing data, erroneous data, duplicate data and data beyond a reasonable range.

7. A method for short-term wind power prediction based on multi-module integration as claimed in claim 6, characterized in that: After preprocessing the data set, normalization is required. The specific method is as follows: Among them, Y i is the normalized data, Y is the unnormalized data, and Y min is the minimum value of the input sequence China, Y max is the maximum value in the input sequence.

8. The method for short-term wind power prediction based on multi-module integration according to claim 1, characterized in that: The dense layer implements nonlinear transformation through activation function.

9. The method for short-term wind power prediction based on multi-module integration according to claim 1, characterized in that: The short-term wind power prediction model also includes a Dropout layer.

10. A device for short-term wind power prediction based on multi-module integration, characterized in that: The device comprises: Data acquisition unit: used to acquire historical meteorological data of the wind farm and corresponding wind power data to form a data set, and divide the data set into a training set and a test set; Feature selection unit: used to calculate the correlation between meteorological data features and historical wind power data using the Pearson correlation coefficient, and select features with stronger correlation as input variables of the short-term wind power prediction model; Sequence decomposition unit: used to decompose the historical wind power data with non-smooth characteristics into multiple modal components with different complexity using the ICEEMDAN layer; Short-term wind power training and prediction unit: used to use the training set short-term wind power prediction model; and input the test set into the short-term wind power prediction model for prediction, and summarize the prediction results at each moment into the final short-term wind power prediction result.

Images