Wind power generation prediction method and system
By combining a bidirectional long short-term memory network with a frequency domain attention network and performing ensemble empirical mode decomposition on wind power generation data, the problem of low accuracy and poor robustness of wind power generation prediction models has been solved, achieving higher accuracy and stable prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI INTELLIGENT COMPUTING TECHNOLOGY CO LTD
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-01
AI Technical Summary
Existing wind power generation prediction models have low accuracy and poor robustness, making it difficult to capture the complex nonlinear characteristics and uncertainties in the wind power generation process. In particular, the prediction accuracy is unstable under complex and variable weather conditions.
The ensemble empirical mode decomposition algorithm is used to decompose wind power generation data to obtain multiple intrinsic mode function components and residual terms. An improved bidirectional long short-term memory network and frequency domain attention network are used to construct a wind power generation prediction model. The prediction results of multiple intrinsic mode function components are superimposed and optimized by combining a weighted loss function of log-hyperbolic cosine loss function and Hubel loss function.
It significantly improves the accuracy and stability of wind power generation forecasts, reduces model randomness, and enhances forecast precision and robustness.
Smart Images
Figure CN121960142A_ABST
Abstract
Description
A method and system for predicting wind power generation Technical Field
[0001] This invention relates to the field of wind power generation technology, and in particular to a wind power generation prediction method and system. Background Technology
[0002] Wind power generation, as a clean and renewable energy source, converts natural wind energy into electricity without consuming fossil fuels and producing almost no carbon emissions, making it environmentally friendly and a crucial pathway to achieving sustainable energy development. While wind power can effectively reduce carbon emissions from traditional thermal power generation, its formation is influenced by a variety of complex factors, including atmospheric circulation, topography, and meteorological conditions, resulting in significant randomness and uncertainty in wind power generation. This instability poses a substantial challenge to the stable operation and dispatch of power systems. Therefore, accurately predicting wind power generation can improve the overall energy efficiency management of grid-connected systems.
[0003] Currently, various prediction models have been developed to forecast wind power generation. However, existing wind power generation prediction models generally suffer from low accuracy and poor robustness. On the one hand, these models struggle to fully capture the complex nonlinear characteristics and uncertainties in the wind power generation process, leading to significant deviations between predicted results and actual power generation. On the other hand, when faced with complex and variable weather conditions or abnormal data, the model performance is easily and significantly affected, making it impossible to maintain stable prediction accuracy. Summary of the Invention
[0004] To address the above technical problems, this invention provides a wind power generation prediction method; furthermore, it also provides a wind power generation prediction system.
[0005] The technical problem solved by this invention can be achieved by the following technical solutions:
[0006] A method for predicting wind power generation includes:
[0007] Step S1: Obtain wind power generation data;
[0008] Step S2: Decompose the wind power generation data based on the ensemble empirical mode decomposition algorithm to obtain multiple intrinsic mode function components and residual terms;
[0009] Step S3: Input the multiple intrinsic mode function components into the pre-trained wind power generation prediction model to obtain the prediction results corresponding to the multiple intrinsic mode function components;
[0010] Step S4: The prediction results of the multiple intrinsic mode function components are superimposed to obtain the wind power generation prediction result.
[0011] Preferably, the ensemble empirical mode decomposition algorithm is an improved adaptive noisy complete ensemble empirical mode decomposition algorithm, and step S2 includes:
[0012] Step S21: Add Gaussian white noise to the original wind power generation sequence in the wind power generation data, and perform ensemble empirical mode decomposition and local averaging to obtain the first set of residual terms and the first intrinsic mode function components.
[0013] Step S22: Using the first set of residual terms as a new sequence, Gaussian white noise is added again, and ensemble empirical mode decomposition and local averaging are performed again to obtain the second set of residual terms and the second intrinsic mode function component. The same process is repeated to obtain the Nth set of residual terms and the Nth intrinsic mode function component, where N is a positive integer greater than 1.
[0014] Preferably, the wind power generation prediction model is constructed based on an improved bidirectional long short-term memory network and a frequency domain attention network, and the construction steps of the wind power generation prediction model include:
[0015] Step S301: Obtain the original bidirectional long short-term memory network;
[0016] Step S302: Replace the hyperbolic tangent activation function of the original bidirectional long short-term memory network with a soft sign activation function, and convert the soft maximum output layer into a marginal soft maximum activation function to obtain the improved bidirectional long short-term memory network.
[0017] Step S303: Integrate a frequency domain attention network into the output layer of the improved bidirectional long short-term memory network to obtain the wind power generation prediction model.
[0018] Preferably, each frequency channel of the frequency domain attention network uses a phantom convolutional layer for feature extraction, while feature shuffling is performed in the frequency domain channel dimension.
[0019] Preferably, the loss function used in the wind power generation prediction model is a weighted loss function of the log-hyperbolic cosine loss function and the Hubel loss function.
[0020] On the other hand, a wind power generation prediction system is provided, including the wind power generation prediction method as described above, comprising:
[0021] The data acquisition module is used to acquire wind power generation data;
[0022] The ensemble empirical mode decomposition module, connected to the data acquisition module, is used to decompose the wind power generation data based on the ensemble empirical mode decomposition algorithm to obtain multiple intrinsic mode function components and residual terms;
[0023] The prediction module, connected to the ensemble empirical mode decomposition module, is used to input the multiple intrinsic mode function components into the pre-trained wind power generation prediction model to obtain prediction results corresponding to the multiple intrinsic mode function components;
[0024] The output module, connected to the prediction module, is used to superimpose the prediction results of the multiple intrinsic mode function components to obtain the wind power generation prediction result.
[0025] Preferably, the wind power generation prediction model includes:
[0026] A bidirectional long short-term memory network is used to output predicted components for the multiple intrinsic mode function components;
[0027] A frequency domain attention network, connected to the output layer of the bidirectional long short-term memory network, is used to perform feature fusion on the prediction components output by the bidirectional long short-term memory network to obtain prediction results corresponding to the multiple intrinsic mode function components.
[0028] Preferably, the activation function of the bidirectional long short-term memory network is a soft sign activation function, and the soft maximum output layer of the bidirectional long short-term memory network is converted into a marginal soft maximum activation function.
[0029] Preferably, each frequency channel of the frequency domain attention network uses a phantom convolutional layer for feature extraction, while feature shuffling is performed in the frequency domain channel dimension.
[0030] Preferably, the loss function used in the wind power generation prediction model is a weighted loss function of the log-hyperbolic cosine loss function and the Hubel loss function.
[0031] The advantages or beneficial effects of the technical solution of this invention are as follows:
[0032] This invention decomposes wind power generation data using an ensemble empirical mode decomposition algorithm, reducing the randomness of the original sequence and providing more regular input features for subsequent prediction models. This overcomes the mode aliasing problem of traditional methods and improves the prediction accuracy of the model. Attached Figure Description
[0033] Figure 1 is a flowchart illustrating the wind power generation prediction method in a preferred embodiment of the present invention; Figure 2 is a flowchart illustrating the improved adaptive noise complete set empirical mode decomposition algorithm in a preferred embodiment of the present invention; Figure 3 is a waveform comparison diagram of the original signal, variational mode decomposition, and residual signal of ICEEMDAN decomposition in a preferred embodiment of the present invention; Figure 4 is a flowchart illustrating the construction steps of the wind power generation prediction model in a preferred embodiment of the present invention; Figure 5 is a structural block diagram of the wind power generation prediction system in a preferred embodiment of the present invention; Figure 6 is a structural diagram of the improved bidirectional long short-term memory network in a preferred embodiment of the present invention; Figure 7 is a structural diagram of the frequency domain attention network in a preferred embodiment of the present invention; Figure 8 is a prediction effect diagram of the model of the present invention and the Informer model in a preferred embodiment of the present invention. Detailed Implementation
[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0041] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0042] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.
[0043] In a preferred embodiment of the present invention, based on the uncertainties and randomness existing in the existing wind power generation prediction process, a short-term wind power generation prediction method based on a frequency domain attention mechanism-intelligent optimization model is provided, as shown in Figure 1, including:
[0044] Step S1: Obtain wind power generation data;
[0045] Step S2: Decompose the wind power generation data based on the ensemble empirical mode decomposition algorithm to obtain multiple intrinsic mode function components and residual terms;
[0046] Step S3: Input multiple intrinsic mode function components into the pre-trained wind power generation prediction model to obtain prediction results corresponding to the multiple intrinsic mode function components;
[0047] Step S4: The prediction results of multiple intrinsic mode function components are superimposed to obtain the wind power generation prediction results.
[0048] In this embodiment, the wind power generation data includes wind power generation information within a preset time period. The preset time period can be set to one month, or a shorter time span, such as a week, a day, or even a more precise period, can be selected according to actual needs and specific circumstances; of course, a longer time range can also be selected, such as a quarter, a year, or even longer.
[0049] Wind power generation refers to the total amount of electricity generated by capturing and converting wind energy from flowing wind in nature using wind power generation equipment. This data not only directly reflects the actual power generation capacity of a wind power system within a specific time period, but is also one of the key indicators for measuring the operational effectiveness of wind power projects and evaluating the efficiency of wind energy resource utilization.
[0050] In wind power data, wind power generation is arranged in chronological order to form a data sequence. By analyzing the sequence, we can understand the output variation patterns of wind power over a period of time, such as whether the power generation in a certain season exhibits stable fluctuations, and the peak and trough periods of power generation at different times of the day.
[0051] Wind power generation data can be obtained from historical databases. These databases store a large amount of operational data from wind power systems over a past period. This data records detailed information such as wind power generation, efficiency, and equipment operating status at different times. By mining and analyzing historical databases, we can gain insights into the long-term trends and patterns of wind power generation, such as seasonal fluctuations in power generation and changes in equipment performance over time. This provides a data foundation for long-term power generation planning, equipment maintenance strategies, and overall performance evaluation of wind farms. Alternatively, wind power generation data can be obtained through other methods, such as installing sensors on wind power equipment, including but not limited to wind speed sensors, wind direction sensors, and power sensors to obtain current power generation.
[0052] After acquiring wind power generation data, an ensemble empirical mode decomposition algorithm is applied to decompose the data, obtaining multiple intrinsic mode function components and residual terms. Furthermore, the preferred ensemble empirical mode decomposition algorithm is the improved Complete Ensemble EMD with Adaptive Noise (ICEEMDAN) algorithm. Unlike traditional Empirical Mode Decomposition (EMD) methods, the ICEEMDAN algorithm introduces adaptive noise and a complete ensemble strategy to improve the stability and accuracy of the decomposition.
[0053] Furthermore, in step S2, as shown in Figure 2, the improved adaptive noise complete set empirical mode decomposition algorithm includes:
[0054] Step S21: Add Gaussian white noise to the original wind power generation sequence in the wind power generation data, and perform ensemble empirical mode decomposition and local averaging to obtain the first set of residual terms and the first intrinsic mode function components.
[0055] Step S22: Using the first set of residual terms as the new sequence, Gaussian white noise is added again, and ensemble empirical mode decomposition and local averaging are performed again to obtain the second set of residual terms and the second intrinsic mode function component. The same process is repeated to obtain the Nth set of residual terms and the Nth intrinsic mode function component, where N is a positive integer greater than 1.
[0056] Specifically, considering the non-stationarity and multi-scale characteristics of wind power generation, this embodiment of the invention employs the adaptive ICEEMDAN algorithm to decompose the original power generation sequence. This algorithm adaptively adds Gaussian white noise and integrates it multiple times to decompose the original signal into a series of Intrinsic Mode Function (IMF) components and residual terms with clear physical meaning. Each IMF component represents the fluctuation characteristics at different time scales, while the residual reflects the overall trend of the sequence.
[0057] Figure 3 shows a waveform comparison of the original signal, the residual signal from conventional Variational Mode Decomposition (VMD), and the ICEEMDAN decomposition of this invention. In Figure 3, A1 represents the original signal, and the waveform in the figure is the original wind power generation sequence signal in the wind power generation data; A2 represents the residual signal from Variational Mode Decomposition; A3 represents the residual signal from ICEEMDAN decomposition; and A4 represents a comparison of the VMD residual signal and the ICEEMDAN residual signal.
[0058] As shown in Figure 3, the residual signal amount is small after adaptive ICEEMDAN decomposition, which can effectively extract all effective mode components.
[0059] The embodiments of the present invention, through ICEEMDAN decomposition, can significantly reduce the randomness of the original sequence, providing more regular input features for subsequent prediction models, and effectively overcoming the mode aliasing problem of traditional methods.
[0060] Furthermore, the wind power prediction model is constructed based on an improved Bidirectional Long Short-Term Memory (Bi-LSTM) network and a frequency domain attention network. The Bi-LSTM network is used to output prediction components for multiple intrinsic mode function components, and the frequency domain attention network is used to perform feature fusion on the prediction components output by the Bi-LSTM network to obtain prediction results corresponding to multiple intrinsic mode function components.
[0061] Furthermore, as shown in Figure 4, the steps for constructing a wind power generation prediction model include:
[0062] Step S301: Obtain the original bidirectional long short-term memory network;
[0063] Step S302: Replace the hyperbolic tangent (tanh) activation function of the original bidirectional long short-term memory network with the softsign activation function, and convert the softmax output layer into the marginal softmax activation function to obtain the improved bidirectional long short-term memory network.
[0064] Step S303: Integrate a frequency domain attention network into the output layer of the improved bidirectional long short-term memory network to obtain a wind power generation prediction model.
[0065] Specifically, due to the strong randomness of wind power generation, simple models cannot effectively represent it. To improve the expressive power of the model, this embodiment of the invention uses an improved LSTM model to train multiple effective modal components after ICEEMDAN decomposition.
[0066] The structure of the improved LSTM network is shown in Figure 6, including a forget gate 101, an update gate 102, and an output gate. In Figure 6, C(t) represents the input at the current time step; C(t-1) represents the input at the current time step; a(t-1) represents the hidden state at the previous time step; and a(t) represents the hidden state at the current time step. This represents the model output.
[0067] The forget gate 101 receives the input C(t) of the current time step and the hidden state a(t-1) of the previous time step as input. Through a series of linear transformations and activation operations, the forget gate 101 determines how much information to retain from the cell state of the previous time step. Specifically, the forget gate 101 uses the sigmoid activation function to generate a value between 0 and 1, which is multiplied element-wise by the previous cell state. 0 indicates complete forgetting, and 1 indicates complete retention, allowing the network to effectively discard historical information that is no longer relevant.
[0068] Update gate 102 determines which new information to add to the cell state. Update gate 102 also receives the input C(t) of the current time step and the hidden state a(t-1) of the previous time step as input. Through the sigmoid activation function, the update gate generates an update weight, which controls the degree to which new information updates the cell state. Simultaneously, a candidate cell state is generated through the first activation function 103, namely the Softsign activation function. Compared to the traditional tanh activation function, the Softsign activation function has smoother gradient characteristics, reducing the risk of gradient vanishing or exploding, thus helping the network learn and update information better.
[0069] The cell state is updated by adding the previous cell state processed by forget gate 101 to the candidate cell state processed by update gate 102. Subsequently, the second activation function 104 (also a Softsign activation function) further processes the updated cell state to generate the hidden state a(t) for the current time step. This hidden state will serve as one of the inputs for the next time step and will also be used in the final output calculation.
[0070] Both the first activation function 103 and the second activation function 104 use the Softsign activation function instead of the traditional tanh activation function. Their calculation formula is shown in formula (1) below:
[0071]
[0072] In equation (1): softsign(x) represents the softsign activation function; x represents the input quantity.
[0073] By using the Softsign activation function, the problem of gradient explosion in the model can be effectively avoided.
[0074] The Softsign activation function is used to handle cell state updates and hidden state generation. Its output ranges from -1 to 1 and is continuously differentiable. Compared to the tanh function, the gradient of the Softsign function decreases more slowly when the input is large. This allows the network to better preserve gradient information when processing long sequences of data, improving training stability and convergence speed.
[0075] In the output layer of the network, the third activation function 105 employs the Margin Softmax activation function 105. Margin Softmax is an improvement on the traditional Softmax function, introducing the concept of a margin. By adjusting the margin parameter, Margin Softmax can increase the decision boundary between different classes, making the model more discriminative in classification tasks. Specifically, when calculating class probabilities, Margin Softmax not only considers the similarity between a sample and its class center, but also increases the discriminative power between other classes and the correct class through the margin parameter, thereby improving the model's classification accuracy and robustness.
[0076] By converting the Softmax output layer in the Bi-LSTM model to a Margin Softmax activation function, the accuracy of the model can be improved to some extent.
[0077] Furthermore, as shown in Figure 7, each frequency domain channel of the frequency domain attention network uses phantom convolutional layers for feature extraction, while feature shuffling is performed in the frequency domain channel dimension.
[0078] Specifically, in order to capture the effective features of different modal components, this embodiment of the invention decomposes the existing wind power generation data into multiple effective information components of different frequencies during the early data preprocessing process for training. At the same time, in order to balance the model's prediction performance and inference speed, this embodiment of the invention proposes to use an improved frequency domain attention mechanism to re-integrate and focus on the effective features in the output stage, thereby further improving the model's prediction accuracy.
[0079] The frequency domain attention network consists of multiple frequency domain channels, and each frequency domain channel is operated using the following formula (2):
[0080]
[0081] In equation (2): the superscript k represents the frequency domain channel index, k = 0, 1, ..., n-1, and n represents the total number of frequency domain channels; DDCT represents the two-dimensional discrete Fourier transform function; the superscripts u and v represent the frequency domain coordinate indices, u represents the frequency domain horizontal coordinate, and v represents the frequency domain vertical coordinate; the subscripts h and w represent the spatial dimension indices, h represents the height, and w represents the width; H represents the total height; W represents the total width; X and B represent the two other dimensional features besides the channels, X represents the input feature tensor, and B represents the batch dimension; Freq k This represents the prediction result of the k-th frequency domain channel output.
[0082] To improve the inference speed and prediction accuracy of the model, this invention modifies the frequency domain channel attention mechanism, uses phantom convolution for feature extraction, and adds an additional shuffling mechanism to enhance feature communication across different dimensions, thereby further improving the model's prediction performance.
[0083] Furthermore, the loss function used in the wind power generation prediction model is a weighted loss function of the log-cosh loss function and the Huber loss function.
[0084] Specifically, due to the strong randomness of wind power generation, the accuracy of the prediction model faces certain challenges. To improve the accuracy of the model, this embodiment of the invention adopts a combined loss function, which uses a weighted loss function of Log-Cosh and Huber loss functions as the loss function during model training. Its calculation formula is shown in the following equation (3):
[0085]
[0086] In equation (3): Loss represents the combined loss function; α represents the weight, i.e., the harmonic relationship between the two loss functions; Loss Log-Cosh Represents the Log-Cosh loss function; Loss Huber y represents the Huber loss function; y represents the actual value of wind power generation; y p This represents the predicted value of wind power generation; the subscript i indicates the sample index, i = 1,...,n, where n represents the total number of samples; δ represents the infinitesimal parameter.
[0087] Specifically, addressing the low accuracy and robustness of existing wind power generation prediction models, this invention proposes a short-term wind power generation prediction model based on a frequency domain attention-intelligent optimization model. An adaptive ICEEMDAN mode decomposition method is used to obtain multiple effective frequency components. Subsequently, these decomposed frequency components are input into an improved Bi-LSTM network for training. To balance model performance and inference speed, a low-parameter frequency domain channel attention mechanism is proposed after the Bi-LSTM structure output. Furthermore, to further improve the accuracy of wind power generation prediction results, a combined loss function of Log-Cosh and Hurbur loss functions is proposed for training.
[0088] To comprehensively evaluate the model's performance, multiple indicators are used, including but not limited to Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Squared Error (MSE). The calculation formulas are shown in Equation 4-5.
[0089]
[0090] In equations (4) and (5): y represents the actual value of wind power generation; y p The value represents the predicted value of wind power generation; the subscript i represents the sample index, i = 1,...,m, where m represents the total number of samples participating in the test; MAE represents the mean absolute error; MSE represents the mean squared error; the smaller the MAE and MSE values, the better the model training effect.
[0091] This invention provides a wind power generation prediction system, including the wind power generation prediction method as described above, as shown in Figure 5, comprising:
[0092] Data acquisition module 1 is used to acquire wind power generation data;
[0093] The ensemble empirical mode decomposition module 2, connected to the data acquisition module 1, is used to decompose wind power generation data based on the ensemble empirical mode decomposition algorithm to obtain multiple intrinsic mode function components and residual terms;
[0094] Prediction module 3, connected to ensemble empirical mode decomposition module 2, is used to input multiple intrinsic mode function components into the pre-trained wind power generation prediction model 31 to obtain prediction results corresponding to the multiple intrinsic mode function components;
[0095] Output module 4 is connected to prediction module 3 and is used to superimpose the prediction results of multiple intrinsic mode function components to obtain the wind power generation prediction result.
[0096] Furthermore, as shown in Figure 5, the wind power generation prediction model 31 includes:
[0097] A bidirectional long short-term memory network 311 is used to output predicted components for multiple intrinsic mode function components;
[0098] The frequency domain attention network 312 is connected to the output layer of the bidirectional long short-term memory network 311 and is used to perform feature fusion on the prediction components output by the bidirectional long short-term memory network 311 to obtain prediction results corresponding to multiple intrinsic mode function components.
[0099] Furthermore, as shown in Figure 6, the activation function of the bidirectional long short-term memory network 311 is a soft sign activation function, and the soft maximum output layer of the bidirectional long short-term memory network 311 is converted into a marginal soft maximum activation function.
[0100] Furthermore, as shown in Figure 7, each frequency domain channel of the frequency domain attention network 312 uses a phantom convolutional layer for feature extraction, while simultaneously performing feature shuffling in the frequency domain channel dimension.
[0101] Furthermore, the loss function used in the wind power generation prediction model 31 is a weighted loss function of the log-hyperbolic cosine loss function and the Hubel loss function.
[0102] The input wind power generation data for one consecutive month is decomposed into effective modes with different period frequencies using the adaptive ICEEMDAN algorithm. The final noise data is then input into the LSTM model for training. After training, each mode component is input into the improved frequency domain attention mechanism for feature fusion, further enhancing the model's predictive ability.
[0103] To address the issues of high randomness and low accuracy in wind power generation forecasting, this invention proposes a short-term wind power generation forecasting model based on a frequency domain attention-intelligent optimization model. This model uses an improved LSTM model to predict different modes and then superimposes them. Furthermore, by fusing a frequency domain attention mechanism and the ICEEMDAN mode decomposition algorithm, the prediction accuracy of the model is further improved to a certain extent. At the same time, by using a harmonic loss function as the overall loss function, the prediction performance of the model is better optimized.
[0104] To demonstrate the superiority of the algorithm of this invention, it is compared with existing algorithms, and the results are shown in Table 1.
[0105] Table 1. Performance Comparison of Multiple Models
[0106] Model MAE (%) RMSE (%) BP 10.87 11.21 CNN-LSTM 5.6 8.3 Infomer 35.6 45.9 This invention 0.6 0.8 surface
[0107] Analysis of Table 1 leads to the following conclusions:
[0108] 1) Compared with the BP neural network, the model structure of this invention adopts forward propagation, as well as LSTM model structure, attention mechanism, etc., which makes the representation more complex and has a stronger ability to express features. Therefore, the mean absolute error (MAE) is reduced by 5.37% and the root mean square error (RMSE) is reduced by 6.01%.
[0109] 2) Compared with the CNN-LSTM algorithm, the MAE and RMSE of the model in this invention have decreased. The main reason is that although the introduction of the CNN model can help improve the model's ability to extract effective features to a certain extent, it is still lacking for wind power generation data with strong randomness.
[0110] 3) Although the Informer model has strong feature extraction and model representation capabilities, it focuses excessively on the correlation between long sequences, which leads to the model getting stuck in local optima. Therefore, the model of this invention can effectively predict wind power generation. The comparison results of the two methods are shown in Figure 8. In Figure 8, (a) is the prediction result of the model of this invention, and (b) is the prediction result of the Informer model. The horizontal axis is time, and the vertical axis is wind power generation. The blue line is the actual value of wind power generation, and the red line is the predicted value of wind power generation.
[0111] 4) Compared with the research results of Xiao Bai et al., the model proposed in this invention performs better. The main reason is that this invention directly decomposes the input modal components more thoroughly, thereby obtaining more effective and pure frequency domain quantities for prediction, thus improving the prediction effect of the model.
[0112] To verify that the methods proposed in this invention can effectively improve the accuracy of the model in predicting offshore wind power generation, the main improvements based on the LSTM model are: (A) the improved LSTM algorithm; (B) the ICEEMDAN decomposition mode; (C) the frequency domain attention mechanism; and (D) the combined loss function. Ablation experiments were conducted on these four improvements, and the results are shown in Table 2 below.
[0113] Table 2 Ablation Experiment Results
[0114] ABCDMAERMSE2.83.3√2.23.1√√2.13.1√√√1.31.5√√√√0.60.8 surface
[0115] As shown in Table 2, by using the improved LSTM model in the embodiments of the present invention, the mean absolute error (MAE) decreased by 0.6 and the root mean square error (RMSE) decreased by 0.2, indicating that the improved LSTM can more accurately predict offshore wind power generation.
[0116] After effectively decomposing the modes using the ICEEEMDAN algorithm, the MAE value further decreased by 0.1, indicating that the decomposition mode improved the accuracy of model prediction to some extent.
[0117] By using the frequency domain attention mechanism, experimental results show that the model's MAE and RMSE values were further reduced by 0.8 and 1.6, respectively, indicating that the mechanism enhances the model's attention to local effective features.
[0118] By using a hybrid loss function, the model's performance can be optimized from multiple dimensions, thus further reducing the MAE and RMSE values.
[0119] The above ablation experiments demonstrate that the prediction method proposed in this invention has superior performance and higher accuracy in wind power generation prediction compared to existing advanced algorithms, and can effectively improve the prediction accuracy of the model.
[0120] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.
Claims
1. A method for predicting wind power generation, characterized in that, include: Step S1: Obtain wind power generation data; Step S2: Decompose the wind power generation data based on the ensemble empirical mode decomposition algorithm to obtain multiple intrinsic mode function components and residual terms; Step S3: Input the multiple intrinsic mode function components into the pre-trained wind power generation prediction model to obtain prediction results corresponding to the multiple intrinsic mode function components; Step S4: Superimpose the prediction results of the multiple intrinsic mode function components to obtain the wind power generation prediction result.
2. The wind power generation prediction method according to claim 1, characterized in that, The ensemble empirical mode decomposition algorithm is an improved adaptive noise-complete ensemble empirical mode decomposition algorithm. Step S2 includes: Step S21, adding Gaussian white noise to the original wind power generation sequence in the wind power generation data, and performing ensemble empirical mode decomposition and local averaging to obtain the first set of residual terms and the first intrinsic mode function component; Step S22, using the first set of residual terms as a new sequence, continuing to add Gaussian white noise, and performing ensemble empirical mode decomposition and local averaging again to obtain the second set of residual terms and the second intrinsic mode function component, and so on to obtain the Nth set of residual terms and the Nth intrinsic mode function component, where N is a positive integer greater than 1.
3. The wind power generation prediction method according to claim 1, characterized in that, The wind power generation prediction model is constructed based on an improved bidirectional long short-term memory network and a frequency domain attention network. The construction steps of the wind power generation prediction model include: step S301, obtaining the original bidirectional long short-term memory network; step S302, replacing the hyperbolic tangent activation function of the original bidirectional long short-term memory network with a soft sign activation function, and converting the soft maximum output layer into a marginal soft maximum activation function to obtain the improved bidirectional long short-term memory network; step S303, integrating the frequency domain attention network into the output layer of the improved bidirectional long short-term memory network to obtain the wind power generation prediction model.
4. The wind power generation prediction method according to claim 3, characterized in that, The frequency domain attention network uses phantom convolutional layers to extract features for each frequency domain channel, while also performing feature shuffling along the frequency domain channel dimension.
5. The wind power generation prediction method according to claim 1, characterized in that, The loss function used in the wind power generation prediction model is a weighted loss function of the log-hyperbolic cosine loss function and the Hubel loss function.
6. A wind power generation prediction system, characterized in that, The wind power generation prediction method, as described in any one of claims 1-5, comprises: a data acquisition module for acquiring wind power generation data; an ensemble empirical mode decomposition module connected to the data acquisition module for decomposing the wind power generation data based on an ensemble empirical mode decomposition algorithm to obtain multiple intrinsic mode function components and residual terms; a prediction module connected to the ensemble empirical mode decomposition module for inputting the multiple intrinsic mode function components into a pre-trained wind power generation prediction model to obtain prediction results corresponding to the multiple intrinsic mode function components; and an output module connected to the prediction module for superimposing the prediction results of the multiple intrinsic mode function components to obtain a wind power generation prediction result.
7. The wind power generation prediction system according to claim 6, characterized in that, The wind power generation prediction model includes: a bidirectional long short-term memory network, used to output predicted components for the multiple intrinsic mode function components; and a frequency domain attention network, connected to the output layer of the bidirectional long short-term memory network, used to perform feature fusion on the predicted components output by the bidirectional long short-term memory network to obtain prediction results corresponding to the multiple intrinsic mode function components.
8. The wind power generation prediction system according to claim 7, characterized in that, The activation function of the bidirectional long short-term memory network is a soft sign activation function, and the soft maximum output layer of the bidirectional long short-term memory network is converted into a marginal soft maximum activation function.
9. The wind power generation prediction system according to claim 7, characterized in that, Each frequency domain channel of the frequency domain attention network uses a phantom convolutional layer for feature extraction, while feature shuffling is performed along the frequency domain channel dimension.
10. The wind power generation prediction system according to claim 6, characterized in that, The loss function used in the wind power generation prediction model is a weighted loss function of the log-hyperbolic cosine loss function and the Hubel loss function.