Ultra-short-term wind power prediction method and device and medium
By introducing coordinate attention mechanism and empirical modal decomposition technology into neural networks, the problem of low accuracy of ultra-short-term wind power power prediction is solved, and higher prediction accuracy and stability are achieved.
Patent Information
- Application Number
- CN202411497049.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-25
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The accuracy of ultra-short-term wind power power prediction in the prior art is low, mainly because neural networks are difficult to effectively understand the spatiotemporal characteristics in various time series data such as wind speed and weather conditions.
By introducing a coordinate attention mechanism into the neural network, and combining empirical modal decomposition, the historical wind power sequence is decomposed into multiple modal components, and input it into the convolutional neural network, long-term memory network and extreme gradient enhancement model, super-supervision Prediction of short-term wind power.
The accuracy and accuracy of ultra-short-term wind power prediction results are improved, and the fluctuation amplitude and complexity are reduced by effectively extracting and utilizing spatial and temporal characteristics, and the analysis ability of prediction results is improved.
Smart Images

Figure CN119026755A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing technology, and in particular to a method, device and medium for predicting ultra-short-term wind power. Background Art
[0002] Wind power generation is one of the main ways to obtain clean energy. With the continuous maturity of related technologies, the total amount of wind power generation is also increasing. In the prior art, neural networks are usually used to predict ultra-short-term wind power, but the accuracy of the predicted ultra-short-term wind power results is low. Therefore, how to improve the accuracy of the prediction results of ultra-short-term wind power has become a technical problem that needs to be solved at present. Summary of the invention
[0003] In view of the above problems, the purpose of the present invention is to provide a method, device and medium for predicting ultra-short-term wind power, so as to improve the accuracy of the prediction results by introducing a coordinate attention mechanism in a neural network. The specific scheme is as follows: In a first aspect, an embodiment of the present application provides a method for predicting ultra-short-term wind power, the method comprising: Obtaining preprocessed wind turbine parameters; the wind turbine parameters include historical wind power sequences and sensor acquisition data; Based on empirical mode decomposition, the preprocessed historical wind power sequence is decomposed into multiple modal components; the multiple modal components have different frequencies; The multiple modal components and the preprocessed sensor data are input into a preset neural network to obtain an ultra-short-term wind power prediction result; the preset neural network introduces a coordinate attention mechanism.
[0004] Optionally, the inputting the multiple modal components and the preprocessed sensor data into a preset neural network to obtain an ultra-short-term wind power prediction result includes: The multiple modal components and the preprocessed sensor collected data are input into a neural network constructed based on a convolutional neural network, the coordinate attention mechanism, a long short-term memory network and an extreme gradient boosting model to obtain the prediction result of the ultra-short-term wind power.
[0005] Optionally, the inputting of the multiple modal components and the preprocessed sensor data into a neural network constructed based on a convolutional neural network, the attention mechanism, a long short-term memory network and an extreme gradient boosting model to obtain the prediction result of the ultra-short-term wind power includes: Inputting the multiple modal components and the preprocessed sensor acquisition data into the convolutional neural network to obtain a feature map; Performing weighted processing on the feature map based on the coordinate attention mechanism to obtain a weighted feature map; Inputting the weighted feature map into the long short-term memory network to obtain key features; The key features are input into the extreme gradient boosting model to obtain the prediction result of the ultra-short-term wind power.
[0006] Optionally, performing weighted processing on the feature map based on the coordinate attention mechanism to obtain a weighted feature map includes: Encoding the height dimension information and the width dimension information in the feature map respectively to obtain an attention map of the height dimension and an attention map of the width dimension; The feature map is recalibrated according to the attention map of the height dimension and the attention map of the width dimension to obtain the weighted feature map.
[0007] Optionally, the decomposing the preprocessed historical wind power sequence into multiple modal components based on empirical mode decomposition includes: Obtaining the local maximum and local minimum of the preprocessed historical wind power sequence; generating an upper envelope and a lower envelope corresponding to the preprocessed historical wind power sequence according to the spline interpolation, the local maximum value and the local minimum value; Calculate the average values of the upper envelope and the lower envelope at each moment to obtain the average value envelope; Calculating the difference between the preprocessed historical wind power sequence and the average value envelope to obtain a first component; If the first component satisfies the intrinsic mode function condition, determining the first-order intrinsic mode function as the first component; Subtracting the first-order intrinsic mode function from the preprocessed historical wind power sequence to obtain a residual; The preprocessed historical wind power sequence is replaced by the residual amount, and a second-order intrinsic mode function is calculated based on the replaced historical wind power sequence until the residual amount is a monotonic function or a constant.
[0008] Optionally, the method further includes: If the first component does not satisfy the inherent mode function condition, the difference between the first component and the average value envelope corresponding to the first component is calculated to obtain the second component; the average value envelope corresponding to the first component is the average value of the upper envelope and the lower envelope corresponding to the first component at each time.
[0009] In a second aspect, an embodiment of the present application further provides a device for predicting ultra-short-term wind power, the device comprising: An acquisition module, used to acquire pre-processed wind turbine parameters; the wind turbine parameters include historical wind power sequences and sensor acquisition data; A decomposition module, used for decomposing the preprocessed historical wind power sequence into multiple modal components based on empirical mode decomposition; the multiple modal components have different frequencies; The prediction module is used to input the multiple modal components and the preprocessed sensor collected data into a preset neural network to obtain an ultra-short-term wind power prediction result; the preset neural network introduces a coordinate attention mechanism.
[0010] Optionally, the prediction module is specifically used to: The multiple modal components and the preprocessed sensor collected data are input into a neural network constructed based on a convolutional neural network, the coordinate attention mechanism, a long short-term memory network and an extreme gradient boosting model to obtain the prediction result of the ultra-short-term wind power.
[0011] Optionally, the prediction module is specifically used to: Inputting the multiple modal components and the preprocessed sensor acquisition data into the convolutional neural network to obtain a feature map; Performing weighted processing on the feature map based on the coordinate attention mechanism to obtain a weighted feature map; Inputting the weighted feature map into the long short-term memory network to obtain key features; The key features are input into the extreme gradient boosting model to obtain the prediction result of the ultra-short-term wind power.
[0012] In a third aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any of the above-mentioned ultra-short-term wind power prediction methods.
[0013] Compared with the prior art, this application has the following beneficial effects: By obtaining the preprocessed wind turbine parameters, based on empirical mode decomposition, the preprocessed historical wind power sequence is decomposed into multiple modal components with different frequencies. Multiple modal components and preprocessed sensor acquisition data are input into a preset neural network to obtain the prediction results of ultra-short-term wind power. Since the preset neural network in the embodiment of the present application introduces a coordinate attention mechanism, the neural network in the embodiment of the present application can effectively extract and utilize spatial features and temporal features to improve the prediction accuracy of the prediction results of ultra-short-term wind power. In addition, since the parameters input into the neural network are the modal components obtained after decomposing the historical wind power sequence, the fluctuation amplitude and complexity are smaller, which is conducive to the analysis of the prediction results, thereby further improving the prediction accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.
[0015] Figure 1 A schematic flow chart of a method for predicting ultra-short-term wind power provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of a neural network provided in an embodiment of the present application; Figure 3 A schematic flow chart of another ultra-short-term wind power prediction method provided in an embodiment of the present application; Figure 4 A schematic diagram of the structure of a unit structure of a long short-term memory network provided in an embodiment of the present application; Figure 5 A schematic diagram of the structure of the coordinate attention mechanism provided in the embodiment of the present application; Figure 6 A schematic diagram of the structure of an ultra-short-term wind power device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0016] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.
[0017] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0018] After analyzing the background technology, the inventor found that the reason why the prediction accuracy of ultra-short-term wind power is low is that the prediction accuracy generated by the neural network used in the prior art is not high when it comes to non-continuous data, and it is difficult for the neural network to better understand the spatiotemporal characteristics contained in various time series data such as wind speed and weather conditions, resulting in the low accuracy of the prior art in generating ultra-short-term wind power prediction results. In response to the above technical problems, the embodiment of the present application provides a method for predicting ultra-short-term wind power, which is as follows: like Figure 1 As shown, the embodiment of the present application provides a method for predicting ultra-short-term wind power, the method comprising: S101: Obtain pre-processed wind turbine parameters; the wind turbine parameters include historical wind power sequences and sensor collection data.
[0019] The historical wind power sequence and the data collected by each sensor are obtained through the Supervisory Control And Data Acquisition (SCADA). The sensor collected data is multi-dimensional meteorological data, including motor speed, blade angle, wind speed and direction, temperature and humidity, wind speed of N range gates of the cabin-type laser radar, etc. The number N can be selected based on actual needs. In the embodiment of this application, N=10.
[0020] The historical wind power sequence and sensor data are used as wind turbine parameters to preprocess the wind turbine parameters. The preprocessing of wind turbine parameters includes: removing abnormal points in wind turbine parameters, filling missing values in wind turbine parameters, and normalizing wind turbine parameters.
[0021] For easier understanding, the following example is given: If there are data corresponding to wind speeds greater than the rated wind speed of the wind turbine, wind turbine shutdown due to fault, wind turbine maintenance, and wind power curtailment in the wind turbine parameters, these data will be removed as abnormal points. If there are missing values in the wind turbine parameters, the missing values will be filled by linear interpolation of adjacent data. This makes the preprocessed wind turbine parameters more accurate than the preprocessed wind turbine parameters, and improves the accuracy of the prediction results generated based on the preprocessed wind turbine parameters.
[0022] S102: Based on empirical mode decomposition, decompose the preprocessed historical wind power sequence into multiple modal components; the multiple modal components have different frequencies.
[0023] Since the original wind power sequence has nonlinear and non-stationary characteristics including spikes and fluctuations, the nonlinearity caused by the spikes and the non-stationarity caused by the fluctuations will affect the accuracy of the subsequent prediction results, resulting in a decrease in the accuracy of the prediction results. Therefore, the embodiment of the present application is based on empirical mode decomposition (EMD) to decompose the original wind power sequence into multiple modal components with different frequencies, each modal component has more stable changes and fewer outliers than the original wind power sequence, thereby reducing the interference of the spikes and fluctuations in the original wind power sequence on the accuracy of the subsequent prediction results, and improving the accuracy of the subsequent prediction results.
[0024] In an optional embodiment, based on empirical mode decomposition, the preprocessed historical wind power sequence is decomposed into multiple modal components, including: Obtain the local maximum and local minimum of the preprocessed historical wind power sequence; According to the spline interpolation, the local maximum and the local minimum, an upper envelope and a lower envelope corresponding to the preprocessed historical wind power sequence are generated; Calculate the average values of the upper envelope and the lower envelope at each moment to obtain the average value envelope; Calculate the difference between the preprocessed historical wind power sequence and the average value envelope to obtain the first component; If the first component satisfies the intrinsic mode function condition, then the first-order intrinsic mode function is determined to be the first component; Subtract the first-order intrinsic mode function from the preprocessed historical wind power sequence to obtain the residual; The preprocessed historical wind power sequence is replaced by the residual, and the second-order intrinsic mode function is calculated based on the replaced historical wind power sequence until the residual is a monotonic function or a constant.
[0025] The preprocessed historical wind power series Divide into multiple regions and determine the maximum value of each region as the local maximum value , the minimum value of each region is a local minimum . Using cubic spline interpolation, local maxima are connected in series to form an upper envelope. Using cubic spline interpolation, local minima are connected in series to form a lower envelope, so as to obtain the upper and lower envelopes corresponding to the historical wind power sequence. Cubic spline interpolation is specifically a method of constructing a smooth curve through a series of shape value points.
[0026] Enter the upper and lower envelopes into the formula The mean value envelope is obtained .
[0027] The preprocessed historical wind power series and the average envelope are input into the formula In the first component . Determine whether the first component satisfies the intrinsic mode function condition. If the first component satisfies the intrinsic mode function condition, determine that the first component is a first-order intrinsic mode function. The intrinsic mode function condition is that the first component is a monotonic function or a constant, so that the obtained intrinsic mode functions are all continuous and stable functions, and the data input to the neural network are all continuous, linear and stable data.
[0028] The preprocessed historical wind power and the first-order natural mode function Enter into formula , and get the remaining amount .
[0029] If the residual is a monotonic function or a constant and the intrinsic modal function cannot be decomposed, the EMD ends. If the residual can still be decomposed into an intrinsic modal function, the obtained residual is used as a new historical wind power sequence, and the above steps are repeated based on the replaced historical wind power sequence until the residual is a monotonic function or a constant. In this way, the preprocessed historical wind power sequence is decomposed into multiple modal components.
[0030] It should be noted that after obtaining multiple modal components, the historical wind power series can be expressed as the sum of n modal components and an average trend component: .in, is the modal component, is the average trend component.
[0031] In an optional embodiment, the above method further includes: If the first component does not meet the intrinsic mode function condition, the difference between the first component and the average value envelope corresponding to the first component is calculated to obtain the second component; the average value envelope corresponding to the first component is the average value of the upper envelope and the lower envelope corresponding to the first component at each time.
[0032] When the first component does not meet the intrinsic mode function condition, the second component is determined based on the difference between the first component and the mean envelope corresponding to the first component, and it is judged whether the second component meets the intrinsic mode function condition. If it does, the second component is determined to be a first-order intrinsic mode function. If not, the next component is determined until the second component meets the intrinsic mode function condition based on the first component. The determined kth component satisfies the intrinsic mode function condition.
[0033] S103: Input multiple modal components and pre-processed sensor data into a preset neural network to obtain an ultra-short-term wind power prediction result; the preset neural network introduces a coordinate attention mechanism.
[0034] Multiple modal components and preprocessed sensor data are taken as input and input into a preset neural network including a coordinate attention mechanism, so that the neural network generates an ultra-short-term wind power prediction result based on the multiple modal components and the preprocessed sensor data.
[0035] This embodiment introduces a coordinate attention mechanism (CA) into a preset neural network. The spatial position information is integrated into the calculation of channel attention through CA, so that when the neural network performs feature extraction, the features obtained include spatial features and temporal features, which facilitates the effective use of multi-dimensional feature information in subsequent feature-based predictions, thereby improving the accuracy and stability of the prediction results output by the neural network based on the input data.
[0036] In an optional embodiment, multiple modal components and preprocessed sensor data are input into a preset neural network to obtain an ultra-short-term wind power prediction result, including: Multiple modal components and preprocessed sensor data are input into a neural network built based on convolutional neural network, coordinate attention mechanism, long short-term memory network and extreme gradient boosting model to obtain the ultra-short-term wind power prediction results.
[0037] Neural networks include convolutional neural networks (CNN), coordinate attention mechanisms, long short-term memory networks (LSTM), and extreme gradient boosting models (XGBoost). The long short-term memory network is specifically a two-layer LSTM.
[0038] like Figure 2As shown, in an optional embodiment, multiple modal components and sensor data are input into a neural network constructed based on a convolutional neural network, an attention mechanism, a long short-term memory network and an extreme gradient boosting model to obtain an ultra-short-term wind power prediction result, including: Input multiple modal components and sensor acquisition data into a convolutional neural network to obtain a feature map; The feature map is weighted based on the coordinate attention mechanism to obtain a weighted feature map; Input the weighted feature map into the long short-term memory network to obtain the key features; The key features are input into the extreme gradient boosting model to obtain the ultra-short-term wind power prediction results.
[0039] like Figure 3 As shown, multiple modal components and sensor data are collected as (N, T1, D input ) is input into CNN, where N is the batch size, T1 is the number of time steps, and D input is the feature dimension. The convolution layer in CNN performs convolution operation on the input data through multiple convolution kernels to extract local features. The convolution operation slides along the time axis and outputs a feature map. Among them, the convolution kernel size is (T1, D con ), D con is the feature dimension of the convolution kernel. A nonlinear activation function is used to perform nonlinear transformation on the feature map to enhance the expressive power. The pooling layer in CNN uses maximum pooling or average pooling operations to reduce the dimension of the feature map and retain important features, thereby improving computational efficiency and preventing overfitting.
[0040] CA is introduced at the connection between CNN and LSTM. After CNN outputs the feature map, the feature map is weighted by CA and the weighted feature map is input into LSTM.
[0041] The unit structure of LSTM is as follows Figure 4 As shown, LSTM introduces a forget gate f t , input gate i t and output gate o t Three gate structures and cell states, through which the discarding and inheritance of information can be controlled. Input gate i t It is used to decide how much new information at the current moment should be stored in the cell state, and the input gate i t It consists of two parts: the Sigmoid layer and the Tanh layer. The Sigmoid layer determines which values need to be updated, and the Tanh layer creates a candidate value vector. The data in the Sigmoid layer and the Tanh layer are multiplied to obtain the final information input to the cell state. tIt is used to determine how many cell states from the previous moment need to be retained to the current moment. Output gate o t Used to decide how much information to output from the cell state to the hidden state at the current moment.
[0042] The process of information transmission and updating in LSTM is shown in the following equation:
[0043] Among them, w f is the weight matrix of the forget gate, w i is the weight matrix of the input gate, w c is the weight matrix of the candidate cell state, w o is the weight matrix of the output gate, σ is the sigmoid function, b f is the bias term of the forget gate, b i is the bias term of the input gate, b c is the bias term of the candidate cell state, b o is the bias term of the output gate, h t-1 is the hidden state at time t-1.
[0044] The key features are input into the extreme gradient boosting model for regression prediction to obtain the ultra-short-term wind power prediction results. XGBoost is a strong classifier built by integrating multiple weak classifiers. The objective function of XGBoost is as follows:
[0045] Among them, X represents the sum of the overall optimization objectives of the XGBoost model, including the error accumulation on all samples and the regularization of the model complexity, which is used to quantify the balance between the performance and complexity of the model on all training samples. Model training is to obtain the ideal model parameters by minimizing X. is the loss function, Represents a tree model, To predict the results, is the true value, Used to measure the training error between the model prediction results and the actual values; The regularization term is used to control the complexity of the model by summing the complexity of the tree. Used to control the number of leaf nodes; T is the number of leaf nodes; The leaf node is used to control the score of the leaf node; is the score of the jth leaf node. Regularization can avoid overfitting of the model. After the model framework is built, it is necessary to use sample data to train the model and optimize the objective function through the previous distribution algorithm. Therefore, the optimization effect of the next step of training depends on the result of the previous step. , indicating that the t-th iteration model is based on the t-1-th iteration model with a sub-model added . The objective function is updated at each iteration as: .
[0046] represents the loss value and regularization term of the model after the current iteration, that is, the comprehensive result of the prediction error and complexity of the model. In order to find the minimum objective function ,exist Performing Taylor second-order expansion at , we get: .
[0047] in is the first-order gradient of the objective function, Remove the constant term that does not affect the optimization result from the second-order gradient of the objective function and replace it with the leaf node score The objective function can be rewritten as follows: .
[0048] The independent variable is , the dependent variable is A quadratic function of one variable. From this, we can calculate the optimal leaf node j for a tree with a fixed structure. And the optimal objective function value at this time is:
[0049] is the gain of node splitting. , Replace the cumulative amount , .
[0050] Error is a key indicator of the quality of wind power prediction results. In order to better evaluate the prediction error of wind power, the embodiment of the present application selects the mean absolute percentage error (MAPE) and the root mean square error (RMSE) for model evaluation: Mean absolute percentage error Y MAPE Specifically:
[0051] Root mean square error Y RMSE Specifically:
[0052] In the formula, is the actual power generation, is the predicted power generation, and n is the total number of prediction points.
[0053] In an optional embodiment, weighted processing is performed on the feature map based on the coordinate attention mechanism to obtain a weighted feature map, including: Encode the height dimension information and width dimension information in the feature map respectively to obtain the attention map of the height dimension and the attention map of the width dimension; According to the attention map of the height dimension and the attention map of the width dimension, the feature map is recalibrated to obtain a weighted feature map.
[0054] The structure of CA is as follows Figure 5 As shown in Figure 1, the size of the feature map when input to CA is C×H×W, where C is the number of channels, H is the height, and W is the width.
[0055] X Avg Pool is: average pooling is performed on each channel along the width to obtain a feature map of the width dimension. The size of the feature map of the width dimension is C×1×W.
[0056] Y Avg Pool is: perform average pooling on each channel along the height to obtain a feature map of the height dimension. The size of the feature map of the height dimension is C×H×1.
[0057] Concat + Conv2d is: concatenate the two pooled feature maps in the channel dimension to form a feature map of C / r×1×(W+H), and fuse and transform the features through a 1×11×1 two-dimensional convolutional layer (Conv2d).
[0058] BatchNorm + Non-linear means: perform batch normalization (BatchNorm) on the feature map after convolution, and increase the expressiveness of the model through non-linear activation functions (such as ReLU).
[0059] Split is: split the batch normalized and activated feature map into two feature maps of C×1×W and C×H×1.
[0060] Conv2d is: the split feature maps are processed separately through another two 1×11×1 two-dimensional convolutional layers.
[0061] Sigmoid is: Apply the Sigmoid activation function to the two convolved feature maps to generate an attention map in the height dimension and an attention map in the width dimension, and use the attention map in the height dimension and the attention map in the width dimension to recalibrate the input feature map in height and width respectively.
[0062] Re-weight is: multiply the attention map activated by Sigmoid by the original C×H×W input feature map to re-weight the original features, strengthen or weaken certain features, and obtain a weighted feature map. Output the weighted feature map C×H×W, so that the feature map input to LSTM is a feature map that has combined the channel information and spatial position information of the original input features.
[0063] By encoding the spatial information of the features output by CNN, the information of the height dimension and width dimension of the feature map are encoded separately to generate attention maps in both the height and width directions. Then, global spatial information is obtained through global average pooling and global maximum pooling operations. Channel attention maps are generated through two independent fully connected layers, and spatial attention maps are generated based on the channel attention maps and combined with spatial position information.
[0064] By obtaining the preprocessed wind turbine parameters, based on empirical mode decomposition, the preprocessed historical wind power sequence is decomposed into multiple modal components with different frequencies. Multiple modal components and preprocessed sensor acquisition data are input into a preset neural network to obtain the prediction results of ultra-short-term wind power. Since the preset neural network in the embodiment of the present application introduces a coordinate attention mechanism, the neural network in the embodiment of the present application can effectively extract and utilize spatial features and temporal features to improve the prediction accuracy of the prediction results of ultra-short-term wind power. In addition, since the parameters input into the neural network are the modal components obtained after decomposing the historical wind power sequence, the fluctuation amplitude and complexity are smaller, which is conducive to the analysis of the prediction results, thereby further improving the prediction accuracy of the prediction results.
[0065] like Figure 6 As shown, an embodiment of the present application provides a device for predicting ultra-short-term wind power, the device comprising: The acquisition module 601 is used to acquire the pre-processed wind turbine parameters; the wind turbine parameters include historical wind power sequences and sensor acquisition data; A decomposition module 602 is used to decompose the historical wind power sequence into multiple modal components based on empirical mode decomposition; the multiple modal components have different frequencies; The prediction module 603 is used to input multiple modal components and sensor collected data into a preset neural network to obtain an ultra-short-term wind power prediction result; the preset neural network introduces a coordinate attention mechanism.
[0066] In an optional embodiment, the prediction module 603 is specifically configured to: Multiple modal components and sensor collected data are input into a neural network built based on convolutional neural network, coordinate attention mechanism, long short-term memory network and extreme gradient boosting model to obtain the ultra-short-term wind power prediction results.
[0067] In an optional embodiment, the prediction module 603 is specifically configured to: Input multiple modal components and sensor acquisition data into a convolutional neural network to obtain a feature map; The feature map is weighted based on the coordinate attention mechanism to obtain a weighted feature map; Input the weighted feature map into the long short-term memory network to obtain the key features; The key features are input into the extreme gradient boosting model to obtain the ultra-short-term wind power prediction results.
[0068] In an optional embodiment, the prediction module 603 is specifically configured to: Encoding the height dimension information and the width dimension information in the feature map respectively to obtain an attention map of the height dimension and an attention map of the width dimension; The feature map is recalibrated according to the attention map of the height dimension and the attention map of the width dimension to obtain the weighted feature map.
[0069] In an optional embodiment, the decomposition module 602 is specifically configured to: Obtaining the local maximum and local minimum of the preprocessed historical wind power sequence; generating an upper envelope and a lower envelope corresponding to the preprocessed historical wind power sequence according to the spline interpolation, the local maximum value and the local minimum value; Calculate the average values of the upper envelope and the lower envelope at each moment to obtain the average value envelope; Calculating the difference between the preprocessed historical wind power sequence and the average value envelope to obtain a first component; If the first component satisfies the intrinsic mode function condition, determining the first-order intrinsic mode function as the first component; Subtracting the first-order intrinsic mode function from the preprocessed historical wind power sequence to obtain a residual; The preprocessed historical wind power sequence is replaced by the residual amount, and a second-order intrinsic mode function is calculated based on the replaced historical wind power sequence until the residual amount is a monotonic function or a constant.
[0070] In an optional embodiment, the decomposition module 602 is further configured to: If the first component does not satisfy the inherent mode function condition, the difference between the first component and the average value envelope corresponding to the first component is calculated to obtain the second component; the average value envelope corresponding to the first component is the average value of the upper envelope and the lower envelope corresponding to the first component at each time.
[0071] By obtaining the preprocessed wind turbine parameters, based on empirical mode decomposition, the preprocessed historical wind power sequence is decomposed into multiple modal components with different frequencies. Multiple modal components and preprocessed sensor acquisition data are input into a preset neural network to obtain the prediction results of ultra-short-term wind power. Since the preset neural network in the embodiment of the present application introduces a coordinate attention mechanism, the neural network in the embodiment of the present application can effectively extract and utilize spatial features and temporal features to improve the prediction accuracy of the prediction results of ultra-short-term wind power. In addition, since the parameters input into the neural network are the modal components obtained after decomposing the historical wind power sequence, the fluctuation amplitude and complexity are smaller, which is conducive to the analysis of the prediction results, thereby further improving the prediction accuracy of the prediction results.
[0072] An embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, any one of the above-mentioned ultra-short-term wind power prediction methods is implemented.
[0073] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.
[0074] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0075] The technical content provided by the present invention is introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for general technical personnel in the field, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.
Claims
1. A method for predicting ultra-short-term wind power. It is characterized in that The method comprises: Obtaining preprocessed wind turbine parameters; the wind turbine parameters include historical wind power sequences and sensor acquisition data; Based on empirical mode decomposition, the preprocessed historical wind power sequence is decomposed into multiple modal components; the multiple modal components have different frequencies; The multiple modal components and the preprocessed sensor collected data are input into a preset neural network to obtain an ultra-short-term wind power prediction result; the preset neural network introduces a coordinate attention mechanism.
2. The method according to claim 1, It is characterized in that The step of inputting the multiple modal components and the pre-processed sensor data into a preset neural network to obtain an ultra-short-term wind power prediction result includes: The multiple modal components and the preprocessed sensor collected data are input into a neural network constructed based on a convolutional neural network, the coordinate attention mechanism, a long short-term memory network and an extreme gradient boosting model to obtain the prediction result of the ultra-short-term wind power.
3. The method according to claim 2, It is characterized in that The multiple modal components and the pre-processed sensor data are input into a neural network constructed based on a convolutional neural network, the coordinate attention mechanism, a long short-term memory network and an extreme gradient boosting model to obtain the prediction result of the ultra-short-term wind power, including: Inputting the multiple modal components and the preprocessed sensor acquisition data into the convolutional neural network to obtain a feature map; Performing weighted processing on the feature map based on the coordinate attention mechanism to obtain a weighted feature map; Inputting the weighted feature map into the long short-term memory network to obtain key features; The key features are input into the extreme gradient boosting model to obtain the prediction result of the ultra-short-term wind power.
4. The method according to claim 3, It is characterized in that The step of performing weighted processing on the feature map based on the coordinate attention mechanism to obtain a weighted feature map includes: Encoding the height dimension information and the width dimension information in the feature map respectively to obtain an attention map of the height dimension and an attention map of the width dimension; The feature map is recalibrated according to the attention map of the height dimension and the attention map of the width dimension to obtain the weighted feature map.
5. The method according to claim 1, It is characterized in that Based on the empirical mode decomposition, the preprocessed historical wind power sequence is decomposed into multiple modal components, including: Obtaining the local maximum and local minimum of the preprocessed historical wind power sequence; Using a three-spline interpolation method, an upper envelope and a lower envelope corresponding to the preprocessed historical wind power sequence are generated according to the local maximum and the local minimum; Calculating the average values of the upper envelope and the lower envelope at each time to obtain an average value envelope; Calculating the difference between the preprocessed historical wind power sequence and the average value envelope to obtain a first component; If the first component satisfies the intrinsic mode function condition, determining the first-order intrinsic mode function as the first component; Subtracting the first-order intrinsic mode function from the preprocessed historical wind power sequence to obtain a residual; The preprocessed historical wind power sequence is replaced by the residual amount, and a second-order intrinsic mode function is calculated based on the replaced historical wind power sequence until the residual amount is a monotonic function or a constant.
6. The method according to claim 5, It is characterized in that The method further comprises: If the first component does not satisfy the inherent mode function condition, the difference between the first component and the average value envelope corresponding to the first component is calculated to obtain the second component; the average value envelope corresponding to the first component is the average value of the upper envelope and the lower envelope corresponding to the first component at each time.
7. A device for predicting ultra-short-term wind power. It is characterized in that The device comprises: An acquisition module, used to acquire pre-processed wind turbine parameters; the wind turbine parameters include historical wind power sequences and sensor acquisition data; A decomposition module, used for decomposing the preprocessed historical wind power sequence into multiple modal components based on empirical mode decomposition; the multiple modal components have different frequencies; The prediction module is used to input the multiple modal components and the preprocessed sensor collected data into a preset neural network to obtain an ultra-short-term wind power prediction result; the preset neural network introduces a coordinate attention mechanism.
8. The device according to claim 7, It is characterized in that The prediction module is specifically used for: The multiple modal components and the preprocessed sensor collected data are input into a neural network constructed based on a convolutional neural network, the coordinate attention mechanism, a long short-term memory network and an extreme gradient boosting model to obtain the prediction result of the ultra-short-term wind power.
9. The device according to claim 8, It is characterized in that The prediction module is specifically used for: Inputting the multiple modal components and the preprocessed sensor acquisition data into the convolutional neural network to obtain a feature map; Performing weighted processing on the feature map based on the coordinate attention mechanism to obtain a weighted feature map; Inputting the weighted feature map into the long short-term memory network to obtain key features; The key features are input into the extreme gradient boosting model to obtain the prediction result of the ultra-short-term wind power.
10. A computer-readable storage medium, It is characterized in that A computer program is stored thereon, wherein when the computer program is executed by a processor, the method for predicting ultra-short-term wind power as claimed in any one of claims 1 to 6 is implemented.
Citation Information
Patent Citations
Short-term wind power prediction method, device and equipment
CN110363360A
Short-long-term prediction method based on online taxi-hailing travel requirements
CN110599767A
Lithium battery capacity estimation method based on improved convolution-long-short term memory neural network
CN111220921A
Offshore wind power prediction method based on RCNN and meteorological time sequence features
CN111815033A
Stack type self-encoding multi-model load prediction method and system based on LSTM
CN112016734A
