End-to-end wind power prediction method and system based on multi-scale feature extraction

Abstract

The invention discloses an end-to-end wind power prediction method and system based on multi-scale feature extraction, and relates to the technical field of smart power grids, and the method comprises the steps: collecting wind power related data from a wind power plant and a weather forecast system; performing data preprocessing on the collected data, performing abnormal value, null value and standardization processing on the data, and dividing the data into a training set and a test set; a Dense-Inception multi-scale feature extraction and wind power prediction model structure is designed; based on the data, optimizing hyper-parameters of the wind power prediction model by adopting a grey wolf optimization algorithm, and carrying out offline secondary training on the model by using the optimized hyper-parameters; and deploying the trained model for online prediction. According to the method, subtle changes in wind power plant data can be effectively captured, the adaptability of the model to a complex environment is enhanced, a split feature reuse strategy is adopted, the utilization rate of feature information is improved, the prediction process is simplified, human intervention is reduced, and the prediction efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of smart grid technology, and in particular to an end-to-end wind power prediction method and system based on multi-scale feature extraction. Background Art

[0002] Renewable energy sources are increasingly being incorporated into distribution networks to avoid greenhouse gas emissions from conventional power generation, meet growing electricity demand, and ensure low-cost electricity supply. With the rapid development of new technologies, the cost of solar and wind power generation has dropped significantly. This shows that using renewable energy to generate electricity is more cost-effective than building new coal-fired power plants. However, integrating renewable energy sources brings uncertainty in power generation and reduces the reliability of the network, especially for large generators such as wind farms. Since these sources are intermittent in nature, there is a mismatch between supply and demand in the real-time operation of the network. By using large energy storage systems or additional generators, this fluctuating power is smoothed over time, resulting in higher maintenance and generation costs. By improving the accuracy of predicting the availability of renewable energy, the size of the energy storage system can be reduced. Improving accuracy is a cost-effective and environmentally friendly solution to address intermittent energy integration.

[0003] In order to improve the prediction accuracy of wind power generation, technologies including physical, statistical and machine learning models have been proposed. The mathematical models of physical and statistical models are inaccurate and cannot be learned and adapted, resulting in high prediction errors. In order to minimize the error, machine learning models including SVM, ANN, LSTM, CNN, Transformer and other models provide more reliable predictions. Deep learning models are able to model the nonlinear relationship between input and output data using flexible network structures, and have higher wind power prediction accuracy than traditional models. However, existing methods have the problem of single feature extraction and no adaptive adjustment of the model's hyperparameters. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is: how to solve the problem that the existing method has a single feature extraction and does not adaptively adjust the hyperparameters of the model.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: an end-to-end wind power prediction method based on multi-scale feature extraction, comprising: collecting wind power related data from wind farms and weather forecast systems; performing data preprocessing on the collected data, performing outlier, null value and standardization processing on the data, and dividing the data into a training set and a test set; designing a Dense-Inception multi-scale feature extraction and wind power prediction model structure; based on the data, optimizing the hyperparameters of the wind power prediction model using the Grey Wolf Optimization Algorithm, and performing offline secondary training on the model using the optimized hyperparameters; and deploying the trained model for online prediction.

[0007] As a preferred solution of the end-to-end wind power prediction method with multi-scale feature extraction described in the present invention, the wind power related data includes power output, wind speed, wind direction, air density, temperature and humidity; the data preprocessing includes removing outliers using the 3σ principle, replacing missing values ​​with the mean before and after, standardizing the data, and dividing the training set and the test set into 7:3.

[0008] As a preferred solution of the end-to-end wind power prediction method of multi-scale feature extraction described in the present invention, wherein: the Dense-Inception multi-scale feature extraction includes dividing the channels of the previous layer into four groups of convolutions on average, each group of convolutions is followed by convolution kernels of different sizes to extract features in parallel, so as to realize feature extraction at multiple scales; a feature reuse strategy is adopted, that is, the previous group of convolutions and the next group of convolutions are added by adding a dimension matching module, so as to realize the reuse of the previous group of convolutions in the next group of convolutions, thereby realizing cross-channel information combination and nonlinear feature transmission, and the dimension matching module is composed of a one-dimensional convolution layer, a maximum pooling layer, a batch normalization and a SELU activation function.

[0009] As a preferred solution of the end-to-end wind power prediction method of multi-scale feature extraction described in the present invention, the wind power prediction model structure includes a convolutional neural network as the backbone, and a certain number of Dense-Inception feature extractions are inserted into the backbone network to form the entire prediction model architecture, which consists of an input layer, a convolutional layer, batch normalization, ReLu activation, 2 Dense-Inception modules, two fully connected layers, Dropout regularization and an output layer. There is a residual connection structure between the two Dense-Inception modules; for the input data set X, the output obtained after the convolutional layer is expressed as:

[0010] X=WX+b

[0011] Among them, S is the weight of the convolution layer, b is the bias term, and X is the output of the convolution layer; the output after the batch normalization layer is expressed as:

[0012]

[0013] Among them, μ is the mean of the small batch data, σ is the standard deviation of the small batch data, γ and β are learnable parameters; the output after ReLU activation is expressed as:

[0014] Z = max(0, Y)

[0015] Among them, Z is the output after ReLU activation; the output obtained by the first Dense-Inception module is expressed as:

[0016] Z1=h(Z)

[0017] Among them, h(·) is the transformation of the first Dense-Inception module on the input Z, Z1 is the output obtained by the first Dense-Inception module; the output obtained by the second Dense-Inception module is expressed as:

[0018] Z2=g(Z+Z1)

[0019] Among them, g(·) is the transformation of the second Dense-Inception module on the input Z+Z1; the output obtained by the first fully connected layer is expressed as:

[0020] O1=W1Z2+b1

[0021] Among them, W1 is the weight of the fully connected layer, b1 is the bias of the fully connected layer; the output obtained by Dropout regularization is expressed as:

[0022] O2=l(O1)

[0023] Where l(·) is the transformation of the input O1 by randomly setting the output of some neurons to zero; the final power prediction output is output through the second fully connected layer and is expressed as:

[0024]

[0025] in, is the model output, W2 is the weight of the second fully connected layer, and b2 is the bias of the fully connected layer.

[0026] As a preferred solution of the end-to-end wind power prediction method based on multi-scale feature extraction described in the present invention, the optimization using the gray wolf optimization algorithm includes a position update strategy that is the average value of the positions of the three best wolves in the population; the optimized hyperparameters include the size of the convolution kernel in the convolution layer and the Dense-Inception multi-scale feature extraction module, the number of neurons in the FC1 and FC2 fully connected layers, and the random inactivation probability of Dropout regularization.

[0027] As a preferred solution of the end-to-end wind power prediction method with multi-scale feature extraction described in the present invention, the gray wolf optimization algorithm includes simulating the gray wolf population hierarchy mechanism and hunting behavior in nature, simulating social hierarchy through four types of gray wolves α, β, δ, ω, simulating the wolf hunting behavior through the process of wolf tracking, encircling, chasing and attacking prey, randomly generating gray wolf populations in the search space, and in order to construct a social hierarchy model of gray wolves, the solution with the best fitness value, the second best solution and the third best solution in the population are regarded as α wolf, β wolf and δ wolf respectively, and the remaining solutions are regarded as ω wolf, which are led by α wolf, β wolf and δ wolf, and ω wolf follows α wolf, β wolf and δ wolf to complete hunting optimization by searching for prey, encircling prey and attacking prey, and finally obtain the optimal solution.

[0028] As a preferred solution of the end-to-end wind power prediction method of multi-scale feature extraction of the present invention, the optimization includes setting an optimization target, using the gray wolf optimization algorithm for optimization, and the optimization target is to minimize the Huber loss function, which is expressed as:

[0029]

[0030] in, is the loss function, y is the actual power value, is the output power value of the prediction model, and θ is the control mechanism coefficient; the offline secondary training includes using the AdamW optimizer to implement the back propagation of the prediction model weights for initial training, performing secondary training on the model based on the best model parameters obtained from the initial training, and achieving secondary training by changing the learning rate update strategy and the number of iterations of the model.

[0031] Another object of the present invention is to provide a system for an end-to-end wind power prediction method based on multi-scale feature extraction, which can solve the problem of end-to-end wind power prediction based on multi-scale feature extraction by constructing an end-to-end wind power prediction system.

[0032] To solve the above technical problems, the present invention provides the following technical solutions: an end-to-end wind power prediction system with multi-scale feature extraction, comprising a data acquisition module, a data preprocessing module, a feature extraction module, an optimization training module and a prediction module; the data acquisition module is used to collect wind power related data from wind farms and weather forecast systems; the data preprocessing module is used to perform data preprocessing on the collected data, perform outlier, null value and standardization processing on the data, and divide the data into a training set and a test set; the feature extraction module is used to design a Dense-Inception multi-scale feature extraction and wind power prediction model structure; the optimization training module is used to optimize the hyperparameters of the wind power prediction model based on the data using the gray wolf optimization algorithm, and use the optimized hyperparameters to perform offline secondary training on the model; the prediction module is used to deploy the trained model for online prediction.

[0033] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the end-to-end wind power prediction method using multi-scale feature extraction as described above are implemented.

[0034] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the end-to-end wind power prediction method using multi-scale feature extraction as described above.

[0035] The beneficial effects of the present invention are as follows: the end-to-end wind power prediction method of multi-scale feature extraction provided by the present invention integrates the Gray Wolf Optimization Algorithm and multi-scale feature extraction technology to achieve end-to-end efficient prediction. By intelligently tuning the hyperparameters of the deep learning model through the Gray Wolf Optimization Algorithm, the optimal hyperparameter combination can be automatically found to improve the prediction accuracy of the model. At the same time, by using the Dense-Inception module for multi-scale feature extraction, this method can effectively capture subtle changes in wind farm data and enhance the adaptability of the model to complex environments. In addition, the present invention also adopts a split feature reuse strategy to further improve the utilization rate of feature information. The end-to-end learning framework simplifies the prediction process, reduces human intervention, and improves prediction efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0037] Figure 1 A flowchart of an end-to-end wind power prediction method based on multi-scale feature extraction provided in the first embodiment of the present invention.

[0038] Figure 2 A schematic diagram of the structure of a Dense-Inception multi-scale feature extraction module of an end-to-end wind power prediction method using multi-scale feature extraction provided in the first embodiment of the present invention.

[0039] Figure 3 A schematic diagram of a wind power prediction model of an end-to-end wind power prediction method based on multi-scale feature extraction provided in a first embodiment of the present invention.

[0040] Figure 4 A schematic diagram of the overall design of an end-to-end wind power prediction method based on multi-scale feature extraction provided in the first embodiment of the present invention.

[0041] Figure 5 A Grey Wolf optimization flow chart of an end-to-end wind power prediction method based on multi-scale feature extraction provided in the first embodiment of the present invention.

[0042] Figure 6 A structural diagram of an end-to-end wind power prediction system with multi-scale feature extraction provided in a second embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0044] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0045] Example 1, reference Figure 1 to Figure 5 , which is the first embodiment of the present invention, and provides an end-to-end wind power prediction method based on multi-scale feature extraction, including: collecting wind power related data from wind farms and weather forecast systems; performing data preprocessing on the collected data, performing outlier, null value and standardization processing on the data, and dividing the data into a training set and a test set; designing a Dense-Inception multi-scale feature extraction and wind power prediction model structure; based on the data, optimizing the hyperparameters of the wind power prediction model using the gray wolf optimization algorithm, and performing offline secondary training on the model using the optimized hyperparameters; and deploying the trained model for online prediction.

[0046] S1. Collect wind power related data from wind farms and weather forecast systems.

[0047] Wind power related data include power output, wind speed, wind direction, air density, temperature and humidity;

[0048] S2. Preprocess the collected data, process outliers, null values ​​and standardize the data, and divide the data into training set and test set.

[0049] Data preprocessing includes removing outliers using the 3σ principle, replacing missing values ​​with the mean before and after, standardizing the data, and dividing the training set and test set into a ratio of 7:3.

[0050] S3. Design the Dense-Inception multi-scale feature extraction and wind power prediction model structure.

[0051] Dense-Inception multi-scale feature extraction involves evenly dividing the channels of the previous layer into 4 groups of convolutions. Each group of convolutions is followed by convolution kernels of different sizes to extract features in parallel, thereby achieving feature extraction at multiple scales.

[0052] A feature reuse strategy is adopted, that is, the previous group of convolutions is added to the next group of convolutions by adding a dimension matching module to achieve the reuse of the previous group of convolutions in the next group of convolutions, thereby realizing cross-channel information combination and nonlinear feature transmission. The dimension matching module consists of a one-dimensional convolution layer, a maximum pooling layer, batch normalization and a SELU activation function.

[0053] like Figure 2As shown in the figure, Dense-Inception divides the channels of the previous layer into four groups of convolutions of 1 / 4, 1 / 4, 1 / 4 and 1 / 4 channels, followed by convolution kernels of different sizes to extract features in parallel, realizing feature extraction at multiple scales. Specifically, the convolution layer and convolution kernel size after each group of convolutions are K1x1 convolution, 1x1 convolution→K2x1 convolution, 1x1 convolution→K3x1 convolution and 1x1 convolution after maximum pooling. In order to reuse feature information, dimension matching channels are constructed between adjacent convolution layer groups, thereby realizing cross-channel information combination and nonlinear feature transmission, further enhancing the information extraction capability. It is worth noting that information combination is achieved by concatenating features. In addition, residual connections are designed to reduce the difficulty of network optimization and avoid problems such as overfitting and abnormal gradients. The dimension matching channel is to match the feature channels between convolution layer groups and the residual connection of the final output. Its structure includes a structure composed of convolution layer, maximum pooling layer, batch normalization and activation function. The activation function chooses SELU. Due to its inherent self-normalization characteristics, the SELU activation function can effectively alleviate the problem of gradient disappearance or explosion during model training, thereby improving the stability and convergence speed of the network. Secondly, the self-normalization mechanism of SELU enables the network output to automatically adjust its scale, enhancing the robustness of the model to input perturbations. Finally, the monotonicity and differentiability of SELU facilitate the use of gradient-based optimization algorithms, simplify the model training process and improve learning efficiency. The expression of the SELU activation function is:

[0054]

[0055] Here, x is the input value, and λ and α are constants, usually set to 1.0507 and 1.67326.

[0056] The wind power prediction model structure is based on the convolutional neural network as the backbone. A certain number of Dense-Inception feature extractions are inserted into the backbone network to form the entire prediction model architecture. It consists of an input layer, a convolutional layer, batch normalization, ReLu activation, two Dense-Inception modules, two fully connected layers, Dropout regularization and an output layer. There is a residual connection structure between the two Dense-Inception modules.

[0057] like Figure 3As shown in the figure, the model consists of an input layer, a convolution layer → batch normalization → activation function, two Dense-Inception multi-scale feature extraction modules, an FC1 fully connected layer, Dropout regularization, and an FC2 fully connected layer. There is a residual connection structure between the two Dense-Inception modules. Dropout regularization is a technique for reducing overfitting and improving the generalization ability of the model when training a neural network. By randomly discarding the activation values ​​of some neurons during the training process, the network is forced to learn more robust and independent feature representations, reducing the risk of complex co-adaptive relationships.

[0058] For the input data set X, the output after the convolution layer is expressed as:

[0059] X=WX+b

[0060] Among them, S is the weight of the convolutional layer, b is the bias term, and X is the output of the convolutional layer.

[0061] The output after the batch normalization layer is expressed as:

[0062]

[0063] Among them, μ is the mean of the mini-batch data, σ is the standard deviation of the mini-batch data, and γ and β are learnable parameters.

[0064] The output after ReLU activation is expressed as:

[0065] Z = max(0, Y)

[0066] Among them, Z is the output after ReLU activation.

[0067] The output obtained by the first Dense-Inception module is expressed as:

[0068] Z1=h(Z)

[0069] Among them, h(·) is the transformation of the input Z by the first Dense-Inception module, and Z1 is the output obtained by the first Dense-Inception module.

[0070] The output obtained by the second Dense-Inception module is expressed as:

[0071] Z2=g(Z+Z1)

[0072] Among them, g(·) is the transformation of the second Dense-Inception module on the input Z+Z1.

[0073] The output obtained by the first fully connected layer is expressed as:

[0074] O1=W1Z2+b1

[0075] Among them, W1 is the weight of the fully connected layer, and b1 is the bias of the fully connected layer.

[0076] The output obtained by Dropout regularization is expressed as:

[0077] O2=l(O1)

[0078] where l(·) is the transformation of the input O1 by randomly setting the outputs of some neurons to zero.

[0079] The final power prediction output is output through the second fully connected layer and is expressed as:

[0080]

[0081] in, is the model output, W2 is the weight of the second fully connected layer, and b2 is the bias of the fully connected layer.

[0082] S4. Based on the data, the hyperparameters of the wind power prediction model are optimized using the Grey Wolf Optimization Algorithm, and the model is trained offline for a second time using the optimized hyperparameters.

[0083] like Figure 4 As shown in the figure, it consists of three submodules: Gray Wolf Optimization Algorithm, Offline Training and Online Prediction. The Gray Wolf Optimization Algorithm is used to optimize the hyperparameters of the prediction model, offline training is used to optimize the weight parameters of the prediction model, and online prediction is used for end-to-end wind power prediction in actual scenarios. The steps for hyperparameter optimization and offline training of the prediction model are as follows:

[0084] S4.1. Set the hyperparameters of the prediction model, including the size of the convolution kernel in the convolution layer and the Dense-Inception multi-scale feature extraction module, the number of neurons in the FC1 and FC2 fully connected layers, and the random inactivation probability of Dropout regularization. The optimized hyperparameters of the designed prediction model are shown in Table 1.

[0085] Table 1 Hyperparameter configuration to be optimized

[0086]

[0087]

[0088] The optimized hyperparameters include the size of the convolution kernel in the convolution layer and the Dense-Inception multi-scale feature extraction module, the number of neurons in the FC1 and FC2 fully connected layers, and the random inactivation probability of Dropout regularization.

[0089] S4.2. Dataset preparation: Collect historical data on wind power, including but not limited to the following parameters: power output, wind speed, wind direction, air density, temperature and humidity.

[0090] S4.3. Preprocess the collected data, handle outliers, null values ​​and standardize the data, and divide them into training set and test set in the ratio of 7:3.

[0091] S4.4. Set the optimization target and use the Grey Wolf Optimization Algorithm for optimization.

[0092] The position update strategy is the average of the positions of the three best wolves in the population.

[0093] The gray wolf optimization algorithm includes simulating the gray wolf population hierarchy mechanism and hunting behavior in nature, simulating the social hierarchy through four types of gray wolves α, β, δ, ω, simulating the wolf's hunting behavior through the process of wolf tracking, encircling, chasing and attacking prey, and randomly generating gray wolf populations in the search space. In order to construct a social hierarchy model of gray wolves, the solution with the best fitness value, the second best solution and the third best solution in the population are regarded as α wolf, β wolf and δ wolf respectively, and the remaining solutions are regarded as ω wolf, which are led by α wolf, β wolf and δ wolf, and ω wolf follows α wolf, β wolf and δ wolf to complete hunting optimization by searching for prey, encircling prey and attacking prey, and finally obtain the optimal solution.

[0094] like Figure 5 As shown in Figure 1, hunting by gray wolves consists of the following three main parts: (1) tracking, chasing, and approaching prey; (2) chasing, surrounding, and harassing prey until it stops moving; and (3) attacking prey. The mathematical model for surrounding prey is as follows:

[0095] D=|C×X p (t)-X(t)|

[0096] X(t+1)=X p (t)-A×D

[0097] Among them, t is the number of instantaneous iterations, X p is the location of the prey, and X is the location of the wolf. The calculation of coefficients A and C is as follows:

[0098]

[0099] Wherein, as the number of iterations decreases, the number of a decreases linearly from 2 to 0. r1 and r2 represent random numbers uniformly selected between [0, 1].

[0100] The gray wolves detect the location of prey under the leadership of wolf α. The gray wolf optimization algorithm assumes that the best solution is wolf α, and the second and third best solutions are wolf β and wolf δ respectively. Therefore, the other wolves in the population will move according to the positions of these three wolves. It can be expressed by the following formula:

[0101]

[0102] Among them, X α , X β and X δ The values ​​of represent the best three wolves in each iteration.

[0103]

[0104] Where the new position of the prey is represented by X p (t+1) is taken as the average of the positions of the three best wolves in the population. The gray wolf completes hunting by attacking the prey. To simulate approaching the prey, A is a random value in the interval [-a, a], where a decreases from 2 to 0 during the iteration. When the |A| value is greater than or equal to 1, the existing search is abandoned to find a better target value. Assuming that the prey is close enough to a value less than 1, the gray wolf is forced to attack the prey. This approach prevents the wolf pack from falling into a local minimum. The search is completed when the gray wolf optimization algorithm reaches the desired number of iterations.

[0105] The optimization goal is to minimize the Huber loss function, which is robust to outliers in the input data and is expressed as:

[0106]

[0107] in, is the loss function, y is the actual power value, is the output power value of the predicted model, θ is the control mechanism coefficient, the control mechanism switches between the mean square error (MSE) and the mean absolute error (MAE), and it is a positive real number.

[0108] S4.5, offline model training for hyperparameter optimization, using the AdamW optimizer to achieve back propagation of the prediction model weights. AdamW has significant improvements over the traditional Adam optimizer in handling weight decay and gradient update bias correction, bringing higher stability and efficiency to the training process. The initial learning rate is set to 0.0005, the batch size is 32, and each model is trained for 30 iterations. When the Huber loss function value is less than 0.05, the iteration is terminated in advance to achieve early stopping and optimal resource configuration.

[0109] S4.6, offline training of the prediction model. The model is trained again based on the best model parameters obtained in S4.5. The model is trained again by changing the learning rate update strategy and the number of iterations. The number of iterations is set to 200, and the learning rate is set to exponential decay with the number of iterations to achieve fine changes in the gradient and find the minimum value. During the second training of the prediction model, the model parameters with the smallest loss function are saved for online verification and prediction.

[0110] Based on the prediction model trained in step S4.6, it can be selectively deployed locally, in the cloud, or on an embedded device, and the output of the prediction value is realized by inputting the monitoring data, thereby realizing end-to-end wind power prediction.

[0111] In summary, the present invention proposes an end-to-end wind power prediction method that integrates the Gray Wolf optimization strategy and the multi-scale feature extraction technology. The method first ensures the accuracy and reliability of the input data through a series of data preprocessing steps. Subsequently, the Gray Wolf optimization algorithm is introduced to intelligently search and optimize the hyperparameters of the deep learning model. The present invention adopts the Dense-Inception multi-scale feature extraction module to perform detailed multi-scale feature extraction on the multi-dimensional data of the wind farm. In addition, through the feature splitting and reuse mechanism, the method enhances the model's ability to integrate information and further improves the accuracy of the prediction. The end-to-end learning framework constructed by the present invention simplifies the entire process from data input to power prediction, eliminates complex feature engineering steps, improves the degree of automation of the prediction process, and has extremely high practical value and broad market application potential.

[0112] Example 2, reference Figure 6 , which is the second embodiment of the present invention, is different from the previous embodiment in that it provides an end-to-end wind power prediction system for multi-scale feature extraction, including: a data acquisition module, a data preprocessing module, a feature extraction module, an optimization training module and a prediction module.

[0113] The data acquisition module is used to collect wind power related data from wind farms and weather forecast systems.

[0114] The data preprocessing module is used to preprocess the collected data, perform outlier, null value and standardization processing on the data, and divide the data into training set and test set.

[0115] The feature extraction module is used to design the Dense-Inception multi-scale feature extraction and wind power prediction model structure.

[0116] The optimization training module is used to optimize the hyperparameters of the wind power prediction model based on data using the Grey Wolf Optimization Algorithm, and use the optimized hyperparameters to perform offline secondary training on the model.

[0117] The prediction module is used to deploy the trained model for online prediction.

[0118] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0119] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0120] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0121] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0122] Embodiment 3 is the third embodiment of the present invention. It is different from the first two embodiments in that it is used to verify the technical effects adopted in the present invention so as to verify the real effects of the method.

[0123] In order to evaluate the effectiveness of the proposed prediction model, mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) and weighted interval accuracy (WIA) were selected as evaluation indicators. The smaller the values ​​of MAE, RMSE and MAPE, the better the prediction effect of the model; the closer the value of WIA is to 1, the higher the consistency between the predicted value and the true value. The calculation formula is as follows:

[0124]

[0125] Among them, y i is the true value, is the predicted value, is the average value of the true value, and m is the number of sample points.

[0126] Using the same input, the support vector machine, artificial neural network, long short-term memory network and convolutional neural network are compared and the results are shown in Table 2. As can be seen from Table 2, the proposed method has the best index, which proves the effectiveness of the proposed method.

[0127] Table 2 Comparison of prediction evaluation indicators of different models

[0128]

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An end-to-end wind power prediction method based on multi-scale feature extraction is characterized by: include, Collect wind power related data from wind farms and weather forecast systems; Preprocess the collected data, remove outliers, remove empty values, and perform standardization on the data, and divide the data into training sets and test sets; Design Dense-Inception multi-scale feature extraction and wind power prediction model structure; Based on the data, the hyperparameters of the wind power prediction model are optimized using the Grey Wolf Optimization Algorithm, and the model is trained offline twice using the optimized hyperparameters. Deploy the trained model for online prediction.

2. The end-to-end wind power prediction method based on multi-scale feature extraction according to claim 1, characterized in that: The wind power related data include power output, wind speed, wind direction, air density, temperature and humidity; The data preprocessing includes removing outliers using the 3σ principle, replacing missing values ​​with the mean before and after, standardizing the data, and dividing the training set and the test set into a ratio of 7:

3.

3. The end-to-end wind power prediction method based on multi-scale feature extraction according to claim 2, characterized in that: The Dense-Inception multi-scale feature extraction includes dividing the channels of the previous layer into four groups of convolutions on average, and each group of convolutions is followed by convolution kernels of different sizes to extract features in parallel, thereby realizing feature extraction at multiple scales; A feature reuse strategy is adopted, that is, the previous group of convolutions is added to the next group of convolutions by adding a dimension matching module to achieve the reuse of the previous group of convolutions in the next group of convolutions, thereby realizing cross-channel information combination and nonlinear feature transmission. The dimension matching module consists of a one-dimensional convolution layer, a maximum pooling layer, batch normalization and a SELU activation function.

4. The end-to-end wind power prediction method based on multi-scale feature extraction according to claim 3, characterized in that: The wind power prediction model structure includes a convolutional neural network as the backbone, a certain number of Dense-Inception feature extractions are inserted into the backbone network to form the entire prediction model architecture, which consists of an input layer, a convolutional layer, batch normalization, ReLu activation, two Dense-Inception modules, two fully connected layers, Dropout regularization and an output layer, and a residual connection structure is provided between the two Dense-Inception modules; For the input data set X, the output after the convolution layer is expressed as: X′=WX+b Among them, W is the weight of the convolutional layer, b is the bias term, and X′ is the output of the convolutional layer; After the batch normalization layer, the output is expressed as, Among them, μ is the mean of the mini-batch data, σ is the standard deviation of the mini-batch data, and γ and β are learnable parameters; After ReLU activation, the output is expressed as, Z=max(0,Y) Among them, Z is the output after ReLU activation; The output obtained by the first Dense-Inception module is expressed as, Z1=h(Z) Where h(·) is the transformation of the input Z by the first Dense-Inception module, and Z1 is the output obtained by the first Dense-Inception module; The output obtained by the second Dense-Inception module is expressed as, Z2=g(Z+Z1) Where g(·) is the transformation of the second Dense-Inception module on the input Z+Z1; The output obtained by the first fully connected layer is expressed as, O1=W1Z2+b1 Among them, W1 is the weight of the fully connected layer, and b1 is the bias of the fully connected layer; The output obtained by Dropout regularization is expressed as, O2=l(O1) where l(·) is the transformation of the input O1 by randomly setting the output of some neurons to zero; The final power prediction output is output through the second fully connected layer and is expressed as, in, is the model output, W2 is the weight of the second fully connected layer, and b2 is the bias of the fully connected layer.

5. The end-to-end wind power prediction method based on multi-scale feature extraction according to claim 4, characterized in that: The optimization using the gray wolf optimization algorithm includes a position update strategy that is an average of the positions of the three best wolves in the population; The optimized hyperparameters include the size of the convolution kernel in the convolution layer and the Dense-Inception multi-scale feature extraction module, the number of neurons in the FC1 and FC2 fully connected layers, and the random inactivation probability of Dropout regularization.

6. The end-to-end wind power prediction method based on multi-scale feature extraction according to claim 5, characterized in that: The gray wolf optimization algorithm includes simulating the gray wolf population hierarchy mechanism and hunting behavior in nature, simulating social hierarchy through four types of gray wolves α, β, δ, ω, simulating the wolf hunting behavior through the process of wolf tracking, encircling, chasing and attacking prey, randomly generating gray wolf populations in the search space, and in order to construct a social hierarchy model of gray wolves, the solution with the best fitness value, the second best solution and the third best solution in the population are regarded as α wolf, β wolf and δ wolf respectively, and the remaining solutions are regarded as ω wolf, which are led by α wolf, β wolf and δ wolf, and ω wolf follows α wolf, β wolf and δ wolf to complete hunting optimization by searching for prey, encircling prey and attacking prey, and finally obtain the optimal solution.

7. The end-to-end wind power prediction method based on multi-scale feature extraction according to claim 6, characterized in that: The optimization includes setting an optimization target and using the gray wolf optimization algorithm for optimization. The optimization target is to minimize the Huber loss function, which is expressed as: in, is the loss function, y is the actual power value, is the output power value of the predicted model, θ is the control mechanism coefficient; The offline secondary training includes using the AdamW optimizer to implement back propagation of the prediction model weights for initial training, performing secondary training on the model based on the best model parameters obtained from the initial training, and implementing secondary training by changing the learning rate update strategy and the number of iterations of the model.

8. A system using the end-to-end wind power prediction method of multi-scale feature extraction according to any one of claims 1 to 7, characterized in that: Including data acquisition module, data preprocessing module, feature extraction module, optimization training module and prediction module; The data acquisition module is used to collect wind power related data from wind farms and weather forecast systems; The data preprocessing module is used to preprocess the collected data, perform outlier, null value and standardization processing on the data, and divide the data into a training set and a test set; The feature extraction module is used to design the Dense-Inception multi-scale feature extraction and wind power prediction model structure; The optimization training module is used to optimize the hyperparameters of the wind power prediction model based on the data using the Grey Wolf Optimization Algorithm, and perform offline secondary training on the model using the optimized hyperparameters; The prediction module is used to deploy the trained model for online prediction.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the end-to-end wind power prediction method based on multi-scale feature extraction according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the end-to-end wind power prediction method based on multi-scale feature extraction according to any one of claims 1 to 7 are implemented.

Images