A method for predicting wind farm power efficiency based on wavelet
Through the wavelet-based wind farm power generation efficiency prediction method, the complexity and interpretability problems of the existing wind power prediction model in multi-scale feature processing are solved, and more efficient wind power prediction and long-term dependency capture are achieved, which improves the training efficiency and prediction accuracy of the model.
Patent Information
- Application Number
- CN202311165632.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-11
- Publication Date
- 2025-08-19
- Estimated Expiration
- 2043-09-11
AI Technical Summary
When dealing with multi-scale features, existing wind power prediction models have problems such as high training complexity, long-term dependency modeling difficulties, complex attention mechanism design, difficult processing of data sparsity and missing values, and poor model interpretability.
The wavelet-based wind farm power generation power efficiency prediction method is used to decompose trend data through moving average technology, and the residual data is decomposed into approximate functions and detail functions by discrete wavelet decomposition. Combined with the Fourier enhanced timing information feature extraction module, the potential data features are mined using a hierarchical double-residual topology structure, and time series correlation is extracted through the U-net structure to optimize the model to improve prediction performance.
It improves the accuracy and interpretability of wind power prediction, reduces the complexity of model training, enhances the ability to capture long-term dependencies, effectively handles data sparsity and missing values, and improves the interpretability and prediction performance of the model.
Smart Images
Figure CN117033987B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wind power generation, and in particular to a method for predicting wind farm power generation efficiency based on wavelet. Background Art
[0002] As one of the most important renewable energy sources, wind power has been highly sought after for its high efficiency, low cost, and environmental benefits, and has experienced rapid growth around the world in recent years. According to the Global Wind Energy Report 2022 released by the Global Wind Energy Council (GWEC), global new wind power installed capacity will reach 93.6 GW in 2021, bringing the cumulative installed wind power capacity to 837 GW, a 12% increase. Furthermore, global wind power tenders reached 88 GW in 2021, a 153% increase over the previous year. The GWEC projects that new installed capacity will reach 557 GW over the next five years, demonstrating that power systems are moving towards a high penetration of renewable energy.
[0003] However, due to meteorological factors such as air pressure, temperature, and the turbine blades themselves, the output power of wind power facilities fluctuates significantly, resulting in poor stability. As the number of wind farms and installed capacity increases, once wind power is integrated into the power grid, these power fluctuations will pose significant challenges to the safe and economic operation of the grid. Therefore, it is crucial to develop reasonable dispatch plans to alleviate the pressure on power system peak and frequency regulation and ensure the safe and economic operation of the power grid. Accurate wind power forecasting facilitates the operation and maintenance of wind farms. Wind power forecasting can be categorized into different types: point power forecasting, interval forecasting, probabilistic forecasting, and scenario forecasting. Point power forecasting based on different forecasting models depends on the wind power forecasting model, primarily including physical methods, statistical methods, and artificial intelligence methods.
[0004] With the rapid development of artificial intelligence technology, it has been applied to wind power forecasting because it can reveal nonlinear relationships in historical data. Convolutional neural networks (CNN) and long short-term memory (LSTM) are two major deep learning models. Mujeeb et al. proposed a wind power forecasting model based on wavelet packet transform and deep convolutional neural network (CNN). Transformer is a deep neural network based on the self-attention mechanism. It was first applied in the field of natural language processing. With the deepening of research, it has been applied to more fields such as computer vision, speech, biology, etc. In recent years, many scholars have introduced attention mechanisms to improve the performance of wind power forecasting models. For example, the attention mechanism is adopted in two LSTM neural networks to adaptively focus on the input features that are more important in the prediction.
[0005] When using the above technology, it was found that the following technical problems exist in the existing technology:
[0006] Most predictions are made only from a single time scale without considering the multi-scale characteristics of the data. Moreover, the converter still needs to be improved in terms of obtaining local information. Existing technologies have some objective shortcomings that need to be recognized and addressed:
[0007] High training complexity: Due to the recursive structure of RNNs, model training can be complex and time-consuming. When processing long sequences, RNNs are prone to vanishing or exploding gradients, which can lead to difficulty in model convergence or unstable prediction results.
[0008] Difficulty in modeling long-term dependencies: Traditional RNN models struggle to effectively capture long-term dependencies. When prediction tasks require considering historical information over a long timeframe, RNNs can suffer from information decay or confusion, leading to degraded prediction performance.
[0009] Complexity of attention mechanism design: While introducing the attention mechanism helps improve model accuracy and interpretability, its design and tuning also present certain challenges. Calculating attention weights requires additional parameters and computational effort, increasing model complexity and training time.
[0010] Data sparsity and missing value handling: In practical applications, wind power forecast data may be sparse or have missing values. Existing technical solutions may face challenges when processing these data features, such as how to fill missing values and how to handle outliers.
[0011] Poor model interpretability and explanation: Although the attention mechanism helps improve the model's predictive performance, its internal decision-making process and prediction results may be difficult to explain. This makes it difficult to understand the basis and reasons for the model's specific prediction results, limiting its interpretability and credibility in practical applications.
[0012] To this end, we design a wavelet-based wind farm power efficiency prediction method to provide another technical solution to the above technical problems. Summary of the Invention
[0013] Based on this, it is necessary to provide a wavelet-based wind farm power efficiency prediction method to address the above technical problems, so as to solve the technical problems raised in the above background technology.
[0014] In order to solve the above technical problems, the present invention adopts the following technical solutions:
[0015] A method for predicting wind farm power efficiency based on wavelet analysis, the steps are as follows:
[0016] S1: Collect wind power data;
[0017] S3: The original time series is converted into trend data by moving average technology, and then the residual data after trend decomposition is decomposed into approximate function and detail function by discrete wavelet decomposition;
[0018] Enhance the temporal information feature extraction module through Fourier;
[0019] The temporal information in the approximate function and the detail function is processed through the temporal information feature extraction module;
[0020] The processed approximate function and detail function are merged through the reconstruction process to restore the standard form of the original data and retain the original time series information;
[0021] S4: Mining latent features in data through a hierarchical dual residual topology structure to improve model performance and trainability.
[0022] As a preferred embodiment of the method for predicting wind farm power generation efficiency based on wavelet provided by the present invention, in step S1, wind power data including environmental characteristics of wind speed, wind direction, temperature and corresponding wind power data are collected;
[0023] The collected data are divided into training set, validation set and test set in a ratio of 7:2:1.
[0024] As a preferred embodiment of the method for predicting wind farm power efficiency based on wavelet provided by the present invention, the trend data obtained by trend decomposition and the detail function and approximate function obtained by wavelet extraction are respectively used as inputs, separated into two subsequences of odd-bit sequence data and even-bit sequence data, and information loss is compensated by interactive learning of the two sequences. The expression is as follows:
[0025]
[0026] in, is the hidden state projected after training through two feedforward neural networks, ⊙ represents the element-wise product method, and φ are training structures built using feedforward neural networks,
[0027] The Fourier enhancement module converts the time domain information into frequency domain information to capture the periodicity, trends and patterns in the data. The expression is as follows:
[0028]
[0029] Among them, X e ' ven ,X o ' ddis the final output of the interactive learning module, where P and U are Fourier enhancement structures based on Fourier transform, P,U = tanh(FE(dropout(LRelu(FE()))), where FE is the Fourier enhancement function.
[0030] As a preferred embodiment of the method for predicting wind farm power efficiency based on wavelet provided by the present invention, in step S3, the correlation between time series data in a larger range is extracted through the U-net structure to capture time series information, and the steps are as follows:
[0031] By downsampling and compressing information, features of a wider range of time series data are extracted, and the output of each layer serves as the input of the next layer;
[0032] At the same time, in the upsampling stage, the downsampling compression information of the same layer is combined and then upsampled and embedded;
[0033] The residual data left after trend decomposition are added to the time series through residual connection to generate a new series with enhanced predictability;
[0034] After the U-net structure, the encoding part is completed to obtain X en , the encoded output X en Input into the fully connected layer for decoding and prediction of X de .
[0035] As a preferred embodiment of the wavelet-based wind farm power efficiency prediction method provided by the present invention, in step S4, the hierarchical double residual topology structure mines the potential features in the data to improve the performance and trainability of the model. The expression is as follows:
[0036]
[0037]
[0038] Among them, f and b are obtained through forward prediction and backward prediction respectively. The forward prediction of f is stacked up, and then the backward prediction of b is subtracted from the current stack input and stacked to the predicted value.
[0039] As a preferred embodiment of the method for predicting wind farm power efficiency based on wavelet provided by the present invention, the method further includes optimizing the model obtained in S4, the steps of:
[0040] Define an appropriate loss function as the optimization objective of the model;
[0041] Use the backpropagation algorithm combined with the Adam optimizer to update and optimize the model parameters to minimize the loss function;
[0042] Perform cross-validation and hyperparameter tuning of the model to select the best model architecture and hyperparameter configuration.
[0043] It can be seen without a doubt that the above-mentioned technical solution of this application can definitely solve the technical problem to be solved by this application.
[0044] At the same time, through the above technical solutions, the present invention has at least the following beneficial effects:
[0045] The present invention provides a wavelet-based wind farm power efficiency prediction method, which uses a wavelet-based Fourier enhanced network model algorithm for wind power prediction and introduces trend decomposition and wavelet transform to adapt to learning and capturing time patterns in a long-term time prediction environment. When the data enters the wavelet transform, detail functions and approximate functions of different scales will be obtained to further extract the time-frequency local characteristics of the sequence, thereby improving the prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0047] Figure 1 Schematic diagram of a feature extraction module based on Fourier enhancement of the present invention;
[0048] Figure 2 This is a schematic diagram of the improved U-Net structure of the present invention;
[0049] Figure 3 It is a flow chart of the solution of the present invention;
[0050] Figure 4 This is a schematic diagram of the network model structure of the present invention. DETAILED DESCRIPTION
[0051] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0052] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.
[0053] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions therein may be combined with each other.
[0054] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.
[0055] Reference Figures 1-4 ,A wind farm power efficiency prediction method based on wavelet.
[0056] 1. Dataset processing:
[0057] Collect historical data related to wind power, including environmental characteristics such as wind speed, wind direction, and temperature, as well as corresponding wind power data. This dataset comes from an actual wind farm dataset in a certain region of China.
[0058] The dataset is divided into training, validation, and test sets in a ratio of 7:2:1. Time series partitioning is typically used to ensure that the training set contains past data, while the validation and test sets contain future data. After partitioning the dataset, the training set can be used for model training and parameter tuning, the validation set for model selection and tuning, and the test set for evaluating model performance and generalization. This allows for a more accurate assessment of the model's predictive effectiveness on unknown data, providing reliable results.
[0059] 2. Model prediction:
[0060] 1) Trend Decomposition and Wavelet Transform: To adapt to learning and capturing temporal patterns in the long-term forecasting environment, we employ a series of processing steps to extract and utilize temporal information. First, we use a moving average technique to detrend the raw data to obtain trend data. However, this processing may result in the loss of some temporal information. Therefore, we further perform residual processing on the trend data. Residual data still contains a wealth of time series-related information. To capture this information, we apply discrete wavelet decomposition (DWT) to decompose the residual data into multiple approximate functions (cA) and detail functions (cD). The approximate functions provide overall trend information of the residual data, while the detail functions capture more subtle temporal variations. To further analyze and utilize the temporal information contained in these approximate and detail functions, we process them using a temporal feature extraction module. This feature extraction process helps extract more representative and useful features, revealing deeper temporal patterns. Finally, we use a reconstruction process to merge the processed approximate and detail functions, restoring the original data to its standard form, thereby preserving the original temporal information.
[0061] 2) Feature extraction module based on Fourier enhancement: To extract more comprehensive temporal information, a multi-layer feedforward neural network with stacked neurons and nonlinear activation functions is established to extract temporal correlation. To compensate for the information loss that may occur due to downsampling, the trend data obtained by decomposing the original data using moving average technology and multiple detail functions and approximate functions obtained through wavelet extraction are used as inputs, respectively, and separated into two sub-sequences: odd-bit sequence data and even-bit sequence data. The information loss is compensated through interactive learning between the two sequences.
[0062]
[0063] is the hidden state projected after training through two feedforward neural networks, where ⊙ represents the element-wise product method, and φ are the training structures we built using feedforward neural networks.
[0064] The time domain information is then converted into frequency domain information through the Fourier enhancement module, which can highlight the frequency domain characteristics of the data, making these characteristics more obvious and identifiable, and can help the model better capture the periodicity, trends and patterns in the data.
[0065]
[0066] As shown in equation (2), X e ' ven ,X o ' dd is the final output of the interactive learning module, where P and U are the Fourier enhancement structures we built based on Fourier transform, P, U = tanh(FE(dropout(LRelu(FE()))), FE is the Fourier enhancement function, and the feature extraction module structure based on Fourier enhancement is as follows Figure 1 shown.
[0067] 3) Improved U-Net structure: In order to effectively extract feature information between a wider range of data, the U-net structure is improved for the analysis of time series data, and a residual connection is added to enable the model to better extract information other than trend data. Using a U-shaped structure for downsampling can expand the receptive field and extract correlations between a larger range of time series data, thereby more effectively capturing time series information. This method can better mine the potential features in the data and provide richer information for subsequent analysis and prediction. The architecture has N layers. It first extracts features of a larger range of time series data by downsampling compressed information, and the output of each layer serves as the input of the next layer; in the upsampling stage, the downsampling compressed information of the same layer is first combined and then upsampled and embedded; then the residual data remaining after the trend decomposition is added to the time series through the residual connection to generate a new series with enhanced predictability. The specific process is as follows. Figure 2 Finally, after the U-net structure, the encoding part is completed to obtain X en , the encoded output X en Input into the fully connected layer for decoding and prediction of X de .
[0068] 4) Double residual:
[0069] The classic deep learning residual network architecture is to pass the result to the next stack, taking the output of the previous stack as input to the next stack and adding the residual to the output of the next stack. These deepening network models can improve trainability, but in the context of this work, simply deepening and performing residual connections will lead to overfitting without well mining some features that cannot be trained. We use a hierarchical double residual topology structure, such as Figure 2 As shown in Figure 2. The proposed architecture has two residual branches, one running on the backward prediction of each layer and the other running on the prediction branch of each layer. This hierarchical dual residual topology can better mine the latent features in the data and improve the performance and trainability of the model. Its operation is described by the following equation:
[0070]
[0071]
[0072] Among them, f and b are obtained through forward prediction and backward prediction respectively. The forward prediction of f is stacked up, and then the backward prediction of b is subtracted from the current stack input and stacked to the predicted value.
[0073] When the training samples are sufficient, we can stack M layers of stack to achieve better prediction accuracy at the cost of a more complex model structure. Specifically, we use the ground truth value to apply intermediate supervision to the output of each stack to facilitate learning intermediate temporal features. The output of the mth intermediate stack, x l , and the input x t-(K-τ)+1:t The process is:
[0074] x l+1 =cat(x l ,x l-1 ) (5)
[0075] Among them, x l It is the output value obtained by reverse prediction and reverse residual of the output of the lth layer, x l-1 is the input of the lth layer.
[0076] The model is able to consider both forward and backward predictions simultaneously, thus better capturing patterns in the data. The backcast part helps the model better understand the data and provides more information for prediction, making the prediction work of downstream blocks easier. This structure also promotes more fluid gradient backpropagation
[0077] Therefore, the original wind power sequence is subjected to trend decomposition and wavelet transform to obtain some feature sequences, which are input into a feature extraction module based on Fourier enhancement to obtain time series features for encoding. In order to further obtain the temporal correlation and correlation of different feature scales, the local features of the sequence are further mined through the U-net structure, and then decoded and predicted through the fully connected layer, and finally the double residual connection is used to improve the prediction performance. The overall framework of the proposed model is as follows Figure 4 shown.
[0078] 3. Model training and optimization:
[0079] 1) Define an appropriate loss function, such as mean squared error (MSE) or mean absolute error (MAE), as the optimization objective of the model.
[0080] 2) Use the backpropagation algorithm combined with the Adam optimizer to update and optimize the model parameters to minimize the loss function.
[0081] 3) Perform cross-validation and hyperparameter tuning of the model to select the optimal model architecture and hyperparameter configuration.
[0082] 4. Model evaluation and prediction:
[0083] 1) Use the test set to evaluate the trained model and calculate the error indicators between the predicted results and the actual wind power, such as the root mean square error (RMSE) and mean absolute error (MAE), to evaluate the prediction performance of the model.
[0084] 2) Use the trained model to predict future wind power and provide short-term wind power prediction results.
[0085] A proposed wavelet-based Fourier-enhanced network model algorithm is used for wind power forecasting. By incorporating trend decomposition and wavelet transforms, it adapts to learning and capturing temporal patterns in long-term forecasting environments. After the data is transformed with wavelets, detail functions and approximate functions at different scales are generated, further extracting local time-frequency features of the sequence, thereby improving forecasting performance. Comparisons of wavelet transforms of different functions show that a combination of the db4 wavelet function and a moving average wavelet function is optimal. The proposed Fourier-enhanced feature extraction module better captures sequence features and is more gradient-friendly without adding additional hyperparameters. The proposed joint model can be combined with different models to adapt to different tasks, demonstrating its versatility.
[0086] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. The preferred embodiments do not describe all details in detail, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.
Claims
1. A method for predicting wind farm power efficiency based on wavelet, characterized in that: Here are the steps: S1: Collect wind power data; S3: The original time series is converted into trend data by moving average technology, and then the residual data after trend decomposition is decomposed into approximate function and detail function by discrete wavelet decomposition; Enhance the temporal information feature extraction module through Fourier; The temporal information in the approximate function and the detail function is processed through the temporal information feature extraction module; The processed approximate function and detail function are merged through the reconstruction process to restore the standard form of the original data and retain the original time series information; S4: Mining latent features in data through hierarchical dual residual topology to improve model performance and trainability; The trend data obtained by trend decomposition and the detail function and approximate function obtained by wavelet extraction are respectively used as inputs and separated into two subsequences: odd-bit sequence data and even-bit sequence data. The information loss is compensated by interactive learning between the two sequences. The expression is as follows: ； in, is the hidden state projected after training through two feedforward neural networks, represents the element-wise product method, and The training structure is built using a feedforward neural network. ; The Fourier enhancement module converts the time domain information into frequency domain information to capture the periodicity, trends and patterns in the data. The expression is as follows: ； in, is the final output of the interactive learning module, where and It is a Fourier enhanced structure based on Fourier transform. , is the Fourier enhancement function; In step S3, the U-net structure is used to extract the correlation between time series data in a larger range to capture the timing information. The steps are as follows: By downsampling and compressing information, features of a wider range of time series data are extracted, and the output of each layer serves as the input of the next layer; At the same time, in the upsampling stage, the downsampling compression information of the same layer is combined and then upsampled and embedded; The residual data left after trend decomposition is added to the time series through residual connection to generate a new series with enhanced predictive power; After the U-net structure, the encoding part is completed. , the encoded output Input into the fully connected layer for decoding prediction .
2. The method for predicting wind farm power efficiency based on wavelet according to claim 1, characterized in that: In the step S1, wind power data is collected, including environmental characteristics of wind speed, wind direction, temperature, and corresponding wind power data; The collected data are divided into training set, validation set and test set in a ratio of 7:2:
1.
3. The method for predicting wind farm power generation efficiency based on wavelet according to claim 1, characterized in that: In the S4 step, the hierarchical double residual topology structure mines the potential features in the data to improve the performance and trainability of the model. The expression is as follows: ； ； in, They are obtained through forward prediction and backward prediction respectively. The forward predictions are stacked up, and then the backward predictions are subtracted from the current stack input. , also stacked to the predicted value.
4. The method for predicting wind farm power efficiency based on wavelet according to claim 1, characterized in that: It also includes optimizing the model obtained in S4, the steps are as follows: Define an appropriate loss function as the optimization objective of the model; Use the backpropagation algorithm combined with the Adam optimizer to update and optimize the model parameters to minimize the loss function; Perform cross-validation and hyperparameter tuning of the model to select the best model architecture and hyperparameter configuration.