Edge end lightweight wind power plant power prediction system and method based on frequency domain sparse mixing

Abstract

The invention provides an edge end lightweight wind power plant power prediction system and method based on frequency domain sparse mixing, and a data monitoring and storage unit, a database, an edge computing gateway and a power regulation and control and scheduling unit of the system form a closed-loop data processing and power prediction structure. Weight parameters of a neural network are generated and screened by using frequency domain characteristics, model parameters are dynamically screened on the frequency domain dimension based on an extreme value disturbance differentiable soft attention mechanism, and then a time domain weight matrix is reconstructed through sparse frequency domain parameters screened by inverse Fourier transform to replace a dense weight matrix of a traditional full connection layer. According to the scheme, a traditional full connection layer is replaced with a frequency domain parameter learning and time domain weight reconstruction mechanism, and compression from a dense weight matrix to sparse frequency domain parameters is achieved; according to the method, while the global receptive field is kept, the parameter quantity and the calculation overhead are remarkably compressed, and the method is suitable for lightweight wind power prediction in an edge end low-calculation-power environment.

Description

Technical Field

[0001] This invention relates to a wind farm power prediction system and method, and more particularly to a lightweight wind farm power prediction system and method based on frequency domain sparse mixing at the edge, belonging to the field of wind power prediction technology. Background Technology

[0002] Against the backdrop of global energy structure transformation, wind power has become one of the most important renewable energy sources in the power system. In the actual operation of wind power grid connection and electricity market trading, wind farm power forecasting plays a crucial role. For grid dispatching, high-precision power forecasting is the foundation for formulating power generation plans, allocating reserve capacity, and maintaining grid frequency stability, helping to effectively mitigate the impact of wind power fluctuations on grid security. For wind farms themselves, accurate short-term and ultra-short-term forecasts are the core basis for participating in electricity spot market bidding, avoiding penalties under the "two detailed rules" (likely referring to specific regulations or rules), and improving the operational efficiency of power plants. Therefore, deploying a highly reliable and high-precision power forecasting system locally at the wind farm, enabling data "local processing and real-time transmission," has become an indispensable technical link connecting wind farms and the grid dispatching center.

[0003] However, existing wind farm power prediction technologies still have significant limitations in engineering applications. Traditional physical models rely on complex hydrodynamic calculations, are extremely sensitive to local micro-meteorological parameters, and are computationally time-consuming. Early statistical models (e.g., ARIMA), while computationally simple, struggle to effectively capture the nonlinear characteristics of wind speed sequences. In recent years, deep learning-based prediction methods (e.g., LSTM, Transformer) have achieved some breakthroughs in accuracy, but their model structures are increasingly large, often containing millions of parameters, and computational complexity increases quadratically with the length of the input sequence. Meanwhile, wind farms are often located in remote areas, and the computing resources of on-site industrial control computers, edge gateways, and other hardware are limited, making it difficult to support real-time inference for such heavy-duty models. This often leads to prediction tasks relying on cloud execution, resulting in data transmission delays and network security risks. Furthermore, wind power data is affected by seasonal changes, diurnal variations, and extreme weather, exhibiting strong non-stationarity. Its statistical distribution drifts continuously over time, causing a significant decrease in prediction accuracy of general-purpose models after long-term operation at the edge.

[0004] Therefore, it is necessary to propose a lightweight wind farm power prediction system and method based on frequency domain sparse hybridization to adapt to the low computing power environment at the edge and to combat data nonstationarity by capturing global features in the frequency domain. Summary of the Invention

[0005] This invention provides a lightweight edge-end wind farm power prediction system and method based on frequency domain sparse hybridization, to overcome two major bottlenecks faced by existing deep learning prediction models in wind farm power prediction: first, the model computational load is too high, making it difficult to deploy in edge devices of wind farms; second, lightweight models often have insufficient feature extraction capabilities, resulting in prediction accuracy that cannot meet practical requirements. The aim is to significantly reduce the computational resource requirements while ensuring prediction accuracy by optimizing the model structure and inference mechanism, thereby improving the overall computational performance and applicability of the relevant system in edge environments.

[0006] The technical solution of this invention is: a lightweight edge-end wind farm power prediction system based on frequency domain sparse hybrid computing. This system is deployed at the wind farm site and includes wind power generation equipment, a data monitoring and storage unit, a database, an edge computing gateway, and a power regulation and dispatch unit. Its key feature is that the data monitoring and storage unit, the database, the edge computing gateway, and the power regulation and dispatch unit constitute a closed-loop data processing and power prediction structure. The wind power generation equipment is configured within the wind farm, including multiple wind turbine generators and at least one meteorological tower, for real-time acquisition of the wind turbine operating status and wind resource environment within the wind farm. The system collects and uploads multi-source heterogeneous data. The input of the data monitoring and storage unit is communicatively connected to the wind power generation equipment, and its output is connected to the database. This allows for real-time acquisition and uploading of measured data, and the storage of multi-source heterogeneous data into the database. The edge computing gateway is communicatively connected to the database, and integrates a non-stationary data preprocessing module, a Fourier sparse mixture prediction module, and a scheduling data uploading module. The power regulation and scheduling unit is connected to the edge computing gateway to receive prediction results and generate wind farm power scheduling or control commands accordingly, in order to formulate power generation plans or market clearing.

[0007] Furthermore, in the aforementioned lightweight wind farm power prediction system based on frequency domain sparse hybridization at the edge, the input of the non-stationary data preprocessing module is connected to the database and is used to perform data cleaning, outlier removal, and standardization on the input multi-source heterogeneous data to generate a regular feature dataset.

[0008] Furthermore, in the aforementioned lightweight wind farm power prediction system based on frequency domain sparse mixing, the input of the Fourier sparse mixing prediction module is connected to the output of the non-stationary data preprocessing module to receive feature datasets and calculate wind farm power prediction values ​​for a specific future period based on the built-in Fourier sparse mixing prediction model.

[0009] Furthermore, in the aforementioned lightweight wind farm power prediction system based on frequency domain sparse hybridization at the edge, the input end of the scheduling data uploading module is connected to the output end of the Fourier sparse hybrid prediction module, which is used to encapsulate the wind farm power prediction value according to a predetermined communication protocol and data format, and output it to the power regulation and scheduling unit.

[0010] This invention also provides a lightweight edge-end wind farm power prediction method based on frequency domain sparse mixing, comprising the following steps: Real-time acquisition of multi-source heterogeneous data on wind turbine operating status (e.g., active power of each unit) and wind resource environment (e.g., wind speed, wind direction, ambient temperature) from wind power generation equipment; the multi-source heterogeneous data is input into a database for storage via a communication network; when the edge computing gateway sends a request to the database, the latest historical time window data is read from the database via an industrial communication protocol, and the historical time window data is provided to the non-stationary data preprocessing module built into the edge computing gateway. After statistical correction to eliminate data distribution drift, preprocessed data is obtained, and the preprocessed data is then input into the Fourier sparse mixing prediction module built into the edge computing gateway. The Fourier sparse mixing prediction module uses... Based on the parameters of the Fourier sparse mixture prediction model trained offline, key frequency components in the preprocessed data are dynamically identified and filtered through an extreme value perturbation differentiable soft attention mechanism. After the time domain weights are fully reconstructed in the frequency domain, the prediction of ultra-short-term or short-term power of wind farms is realized, and the corresponding wind farm power prediction value is output. After the wind farm power prediction value is denormalized, it is input to the power regulation and dispatch unit through the scheduling data uploading module built into the edge computing gateway. The power regulation and dispatch unit (e.g., power dispatch control center (EMS system) or power trading platform) receives the wind farm power prediction value and generates corresponding wind farm power dispatch or control commands based on the prediction value to perform real-time forward-looking adjustment of the active power output of the wind farm to meet the power control requirements of the power grid dispatching agency.

[0011] Furthermore, in the aforementioned lightweight wind farm power prediction method based on frequency domain sparse mixing at the edge, the Fourier sparse mixing prediction module is configured to specifically perform the following steps: read historical observation data of the wind farm from the database (historical observation data includes, but is not limited to, historical wind farm operation and meteorological observation data); concatenate the historical observation data in chronological order to form a historical observation sequence; perform time-series alignment and feature dimension expansion processing on the historical observation sequence to generate a multi-dimensional time-series input matrix. This multi-dimensional time-series input matrix is ​​composed of one or more input sequences organized and combined; classify the quality of each input sequence according to preset rules; and automatically select the corresponding statistical correction strategy for data sequences of different quality levels to obtain regularized data samples. The Fourier sparse mixing prediction module integrates a sparse linear layer based on Fourier transform. This sparse linear layer based on Fourier transform is used to replace the fully connected layer in traditional deep learning, unlike the fully connected layer which directly stores data in a shape... The dense weight matrix is ​​used in the Fourier transform-based sparse linear layer, which learns a small number of frequency domain feature parameters and combines them with an extreme value perturbation differentiable soft attention mechanism to filter key frequencies. The inverse Fourier transform is used to dynamically restore these sparse frequency domain parameters to time domain weights. Regularized data samples are input into the Fourier transform-based sparse linear layer to calculate the initial feature sequence for subsequent time-channel hybrid modules to perform deep feature extraction. Multiple time-channel hybrid modules are stacked sequentially to perform multi-level feature extraction on the input initial feature sequence. The input of the first time-channel hybrid module is connected to the Fourier transform-based sparse linear layer, and the output of the last time-channel hybrid module is connected to a linear projection layer. The linear projection layer is used to map the length of the feature sequence to the prediction length. Using the statistics saved in the statistical correction stage, the mapped feature sequence is denormalized to output the wind farm power prediction value.

[0012] Furthermore, in the aforementioned lightweight wind farm power prediction method based on frequency domain sparse hybridization, the calculation process of each Fourier sparse linear layer in the Fourier sparse hybrid prediction model is as follows: First, a learnable frequency domain parameter matrix is ​​constructed, and the frequency importance score is calculated to generate extreme value perturbation differentiable soft attention weights. After weighting the frequency domain parameters, the time domain weight matrix is ​​reconstructed using discrete Fourier inverse transform. After using the reconstructed weights to perform a linear transformation on the input data, a linear mapping is output.

[0013] Furthermore, in the aforementioned lightweight wind farm power prediction method based on frequency domain sparse mixing, the time-channel mixing module includes a serial time-mixing sublayer and a channel mixing sublayer.

[0014] Specifically, the time-series mixing sublayer is used to capture the periodic pattern of wind speed changes over time; the channel mixing sublayer is used to capture the coupling relationship between different meteorological variables.

[0015] Compared with existing technologies, the technical solution of this invention introduces a sparse linear layer based on Fourier transform as the core operator in the prediction model. This operator is implemented by improving the underlying computation method of deep learning. It utilizes an extreme value perturbation differentiable soft attention mechanism to dynamically filter model parameters in the frequency domain, retaining key frequency components. Through inverse Fourier transform, the time-domain weight matrix is ​​reconstructed using the filtered sparse frequency domain parameters, which can directly replace the dense weight matrix of traditional fully connected layers, significantly reducing model parameters and computational complexity, thereby improving training efficiency and prediction accuracy. The "frequency domain parameter learning and time domain weight reconstruction" mechanism of this invention can achieve compression from dense weight matrix to sparse frequency domain parameters. This method significantly compresses the number of parameters and computational overhead while maintaining the global receptive field, making it suitable for lightweight wind power prediction in low-computing-power environments at the edge. Attached Figure Description

[0016] Figure 1 This is a system framework diagram of the present invention;

[0017] Figure 2 This is a flowchart of the internal calculation process of the Fourier sparse mixture prediction module of the present invention.

[0018] Figure 3 The graph shows the relative error of the power generation prediction model between Example 1 and Comparative Examples 1, 2, and 3 on the test set sample sequence (i.e., the relative error curve between the predicted value and the actual value).

[0019] Figure 4 The bar chart shows the performance comparison of Example 1 with Comparative Examples 1, 2, and 3 in terms of power generation prediction error and parameter quantity.

[0020] The meanings of the labels in the figure are as follows: 1-Wind power generation equipment, 2-Data monitoring and storage unit, 3-Database, 4-Edge computing gateway, 41-Non-stationary data preprocessing module, 42-Fourier sparse mixed prediction module, 43-Dispatch data transmission module, 5-Power regulation and dispatching unit. Detailed Implementation

[0021] The following will be combined with the appendix Figures 1-4The technical solution of the present invention will be further described in detail below with specific embodiments to facilitate better understanding and implementation by those skilled in the art. It should be noted that the present invention is applicable to power prediction of wind farms with different numbers and models of wind turbines, various spatial boundary conditions, and various layout constraints, possessing wide applicability and good expansion potential. The accompanying drawings are only some feasible embodiments and do not constitute a limitation on the scope of protection and technical applicability of the present invention. For those skilled in the art, based on a full understanding of the core technical concept of the present invention, other forms of wind farm layout schemes can be designed without creative effort according to specific scenarios such as actual wind resource conditions, terrain environment, grid connection requirements, and environmental constraints, achieving technical effects equivalent to or even optimized by the present invention.

[0022] like Figure 1 As shown, this invention provides a lightweight edge-end wind farm power prediction system based on frequency domain sparse hybrid computing. The system is deployed at a wind farm and includes a wind power generation device 1, a data monitoring and storage unit 2, a database 3, an edge computing gateway 4, and a power regulation and scheduling unit 5. The data monitoring and storage unit 2, the database 3, the edge computing gateway 4, and the power regulation and scheduling unit 5 constitute a closed-loop data processing and power prediction structure. The wind power generation device 1 is configured within the wind farm and includes multiple wind turbines and at least one meteorological tower for real-time data collection of wind turbine operating status and wind resource environment within the wind farm. The data monitoring and storage unit 2 has an input terminal that is communicatively connected to the wind power generation equipment 1 and an output terminal that is connected to the database 3. It is used to collect and upload measured data in real time and store multi-source heterogeneous data in the database 3. The edge computing gateway 4 is communicatively connected to the database 3, and the edge computing gateway 4 integrates a non-stationary data preprocessing module 41, a Fourier sparse mixture prediction module 42, and a scheduling data uploading module 43. The power regulation and scheduling unit 5 is connected to the edge computing gateway 4 and is used to receive prediction results and generate wind farm power scheduling or control commands accordingly.

[0023] The prediction method of this system is as follows: The wind farm sensing layer equipment is activated, and the wind power generation equipment 1 (i.e., wind turbines and meteorological towers) collects real-time data on instantaneous wind speed, average wind speed, wind direction, temperature, air pressure, and the active power of each unit within the wind farm. All collected data is aggregated through the wind farm's fiber optic ring network to the data monitoring and storage unit 2 (i.e., the booster station integrated automation subsystem), and stored in the database 3 according to timestamp alignment, forming a historical time-series dataset. The edge computing gateway 4 extracts the most recent period (e.g., the past 24 hours) from the database 3 according to a preset scheduling cycle (e.g., every 15 minutes or every 5 minutes). The observation sequence is used; the non-stationary data preprocessing module 41 and the Fourier sparse mixture prediction module 42 inside the edge computing gateway 4 process and infer the data to generate power prediction curves for future times (e.g., the next 4 hours or 24 hours); then the prediction results are encapsulated into a message format (e.g., E file format) that conforms to the power grid dispatch specifications through the dispatch data uploading module 43, and can be sent to the power control and dispatching unit 5 through the power dispatch data network; after the power control and dispatching unit 5 receives the prediction data, it incorporates it into the power generation plan balance calculation of the whole network and issues the final active power control command to the wind farm.

[0024] like Figure 2 As shown, the Fourier sparse mixture prediction module 42 runs in the edge computing gateway 4 deployed in the wind farm booster station. Its specific data flow and algorithm implementation steps are as follows:

[0025] Step (1) Constructing a multidimensional time-series input matrix: Read the historical observation data of the wind farm for the most recent S time steps from the database, including wind speed, wind direction, temperature, air pressure, and active power. Concatenate these data in chronological order to construct the input tensor X. in :

[0026]

[0027] Where B represents the batch size; S represents the length of the historical observation sequence; and M represents the dimension of the variable features.

[0028] Step (2) Adaptive statistical correction preprocessing of data: In order to eliminate the distribution drift of wind power data caused by seasonality and day-night cycles, each sample instance needs to be cleaned and normalized. The acquired data is thoroughly cleaned, including but not limited to eliminating noise in the data, correcting inconsistent data, and identifying and deleting outliers (specifically including missing value handling, outlier detection and replacement).

[0029] After cleaning the data to ensure its quality, the data is normalized. First, the mean μ of the input sequence in the time dimension is calculated. x and standard deviation σ x :

[0030]

[0031] Next, the input data is standardized to obtain the normalized tensor X. enc :

[0032]

[0033] At the same time, the statistic μ x and σ x Saved to memory for denormalization recovery in subsequent steps.

[0034] Step (3) generates a sparse linear layer based on Fourier transform to replace the fully connected layers in traditional deep learning: this step does not directly store the shape D. out ×D in Instead of using a dense weight matrix, the system dynamically reconstructs time-domain weights by learning sparse parameters in the frequency domain and combining them with an extremum perturbation differentiable soft attention mechanism.

[0035] The calculation process for each Fourier sparse linear layer in the model is as follows:

[0036] Step (31): First, a gating network is used to calculate the importance score of each frequency position based on the input features. When the input is H, the weights of the gating network are... The formula is as follows:

[0037]

[0038] The original scores are transformed and global average pooled to obtain the frequency domain importance score vector. .

[0039] Step (32): The score vector is processed using an extreme value perturbation soft sampling mechanism to generate differentiable sparse attention weights, as follows:

[0040]

[0041] in, Let τ be the noise sampled from the Gumbel(0,1) distribution, and τ be the temperature parameter. The advantage of Gumbel-Softmax over ordinary Softmax is that, while maintaining differentiability, it achieves an effect closer to discrete sampling by introducing random noise. It can be applied to gradient propagation during training and promotes efficient selection of frequency sparsity during inference.

[0042] The mechanism of the temperature parameter is as follows: when τ is large, the attention weight distribution is smooth, and the gradient signal can be effectively propagated to the gating network, which is convenient for training; when τ is small, the attention weights tend to be sparsely distributed, which is closer to the hard selection effect and is beneficial for computational compression in the inference stage. Therefore, during training, the temperature parameter can be set as a learnable parameter or gradually reduced using an annealing strategy.

[0043] Step (33): Initialize the learnable complete spectral parameter matrix By using attention weights to weight the spectral parameters element by element, a weighted frequency domain matrix is ​​constructed:

[0044]

[0045] in, This represents the Hadamard product (i.e., element-wise multiplication). The physical meaning of this operation is that the gated network identifies the most important frequency components for the current input and enhances them by perturbating the soft attention weights through extreme values; at the same time, unimportant frequency components are softly suppressed but still retain a weak gradient path, ensuring that the entire network can continuously optimize end-to-end.

[0046] Step (34): For the weighted frequency domain matrix Performing a two-dimensional discrete inverse Fourier transform, and utilizing the global property of the Fourier transform, ensures that even if only a small number of key frequency components are activated, the reconstructed weight matrix can still cover the entire time domain, thus achieving a global receptive field. The formula for taking the real part as the reconstructed time-domain weight matrix is ​​as follows:

[0047]

[0048] Step (35): Perform a linear transformation on the input H using the reconstruction weights to obtain the output linear mapping, the formula of which is as follows:

[0049]

[0050] Step (4) Decoupling and mixing time-channel features: The input feature H... (l-1) The input is fed into the l-th time-channel mixing module, which contains two sub-layers: serial time mixing and channel mixing. Both sub-layers call the sparse weight generation method based on Fourier transform used in step (3).

[0051] Step (41): In the temporal hybrid sublayer used to capture the periodicity of wind speed changes over time, the input is transposed so that the time dimension S is the last dimension, and then passed sequentially through the Fourier sparse linear layer (dimension: S→D). model ReLU activation function, Fourier sparse linear layer (dimension: D) modelAfter processing by the →S) and Dropout layers, the layers are transposed back to the original dimensions and residual connections are then stacked, as shown in the following formula:

[0052]

[0053] FourierLinear1 has an input dimension of S and an output dimension of D. model The input dimension of FourierLinear2 is D. model The output dimension is S.

[0054] Step (42): In the channel mixing sublayer used to capture the coupling relationship between different meteorological variables, H tmp Passing sequentially through Fourier sparse linear layers (dimension: M→D) model ReLU activation function, Fourier sparse linear layer (dimension: D) model →M) and Dropout layers are processed, and residual connections are superimposed. The formula is as follows:

[0055]

[0056] FourierLinear3 has an input dimension of M and an output dimension of D. model FourierLinear 4 The input dimension is D model The output dimension is M.

[0057] Step (5) Predicting output and inverse normalization: After processing by the L-layer stacked temporal-channel hybrid module, the deep feature representation H is obtained. (L) A linear projection layer maps the sequence length S to the prediction length P, resulting in a normalized prediction value. :

[0058]

[0059] in, These are the projection layer weights.

[0060] Finally, the statistics saved in step (2) are used for inverse normalization to obtain the final wind power prediction result. :

[0061]

[0062] The prediction result This is the final data sent to the power grid dispatch center.

[0063] To illustrate the technical solution of this invention in detail, a specific embodiment 1 is provided below. Embodiment 1 collects historical operational data and meteorological data from a wind farm in Northwest China throughout the year to verify the feasibility and prediction process of the wind farm power prediction system and method. The verification process of Embodiment 1 strictly follows the aforementioned system flow, and the specific steps are as follows:

[0064] Data Acquisition and Non-stationary Preprocessing: Historical observation data of wind farms, covering features such as wind speed, wind direction, temperature, and active power, are acquired. The data is then cleaned to remove obvious outliers and noise. Subsequently, the mean and standard deviation of each sample over time are calculated, and the data is normalized to eliminate dimensional differences and ensure the stability of the input data distribution. Taking into full account the hardware resource limitations of the edge computing gateway, the Fourier sparse mixture prediction model of Example 1 is constructed. The model parameters are set as follows: input sequence length S = 24 (i.e., using 24 historical sample points), prediction length P = 1 (i.e., predicting the power value at the next moment); hidden layer dimension Dmodel = 128; the Fourier sparse mixture prediction model integrates... A sparse linear layer based on Fourier transform dynamically generates sparse weights using an extreme value perturbation differentiable soft attention mechanism, and employs a two-layer stacked time-channel hybrid module to deeply extract multivariate coupled features. The dataset is then divided into training, validation, and test sets in chronological order. The mean squared error (MSE) is used as the loss function, and the Fourier sparse mixture prediction model is trained using the Adam optimizer. During training, the network parameters are iteratively updated using the backpropagation algorithm until the validation set error no longer decreases or reaches a preset number of rounds. The test set data is then input into the trained Fourier sparse mixture prediction model to obtain normalized prediction results, which are then denormalized using the saved statistics (μ, σ) to obtain the final wind power prediction value.

[0065] In the specific implementation, three existing technical solutions with different architectures were selected as comparative examples and deployed together with the embodiments of the present invention in the same low-computing-power hardware environment on the edge side to compare and test the prediction accuracy, response time and resource utilization of each solution.

[0066] Comparative Example 1 (using a fully connected baseline model): The same neural network infrastructure as in Example 1 is used, but instead of using sparse linear layers based on Fourier transform, they are replaced with ordinary fully connected linear layers.

[0067] Comparative Example 2 (using the Transformer model): This uses a deep learning model based on the self-attention mechanism, which represents the mainstream high-precision model used in academia and industry for long sequence prediction, but its computational complexity is relatively high.

[0068] Comparative Example 3 (using LSTM model): Employing a Long Short-Term Memory network, representing a classic recurrent neural network model for time series forecasting.

[0069] like Figure 3 As shown, the prediction deviation curve of the model in Example 1 remains stable, fluctuating closely around the zero axis (y=0). Even at abrupt changes where other models experience significant oscillations near certain sample points, the model of this invention maintains a low deviation amplitude without any obvious spikes. In contrast, the fully connected baseline model in Comparative Example 1 and the LSTM model in Comparative Example 3 exhibit severe jitter at data abrupt changes. This indicates that the model of this invention is not only optimal in overall statistical metrics but also possesses extremely strong robustness in the time series dimension. This model can effectively suppress random noise in wind power data, maintain stable output when wind speed changes drastically, and obtain highly accurate wind farm power prediction values. Therefore, this invention can provide reliable power forecasting information for the accurate formulation and safe dispatch of power grid generation plans, thereby improving the economy and safety of power system operation.

[0070] like Figure 4 As shown, regarding the advantages of lightweight model design, thanks to the newly designed frequency-domain sparse weight generation mechanism, the model of this invention only needs to learn parameters for key frequency components, making its parameter count only about 1% of that of traditional models. Compared with the Transformer model in Comparative Example 2 and the LSTM model in Comparative Example 3, this represents a significant reduction in parameter count, and it also shows a significant decrease compared to the fully connected baseline model in Comparative Example 1. Therefore, the model of this invention greatly reduces the algorithm's consumption of edge hardware storage space and computing resources. Regarding prediction accuracy, although the model of this invention has the fewest parameters, its mean squared error (MSE) and mean absolute error (MAE) are superior to all the comparative models. These results fully demonstrate that by utilizing the global receptive field characteristics of Fourier transform and combining it with a sparse attention mechanism for dynamic selection, this invention maintains extremely high prediction stability while significantly compressing model parameters, achieving an effective balance between "lightweight" and "high accuracy" in edge applications.

[0071] The key technical aspect of this invention is the use of a sparse linear layer based on Fourier transform in the prediction model. This departs from the traditional approach of using frequency as an input feature, improving the underlying computational methods of deep learning rather than simply concatenating features. This invention utilizes frequency domain characteristics to generate and filter the weight parameters of the neural network. Through an extreme value perturbation differentiable soft attention mechanism, it dynamically filters model parameters in the frequency domain, retaining key frequency components. The time-domain weight matrix is ​​reconstructed using the filtered sparse frequency domain parameters via inverse Fourier transform (IFFT), directly replacing the dense weight matrix of traditional fully connected layers. Figure 3 and Figure 4The diagram shows a comparison between the prediction model of the present invention and the traditional models of each comparative example. The prediction model of the present invention is superior to the traditional models of the comparative examples in terms of both model lightweighting and prediction accuracy.

[0072] Thus, by employing the technical solution of this invention, a sparse linear layer based on Fourier transform can compress the parameter complexity of the traditional fully connected layer from O(D²) to the O(K) level (where D is the feature dimension of the model and K represents the number of sparse parameters retained in the frequency domain) while maintaining the global receptive field, making it suitable for lightweight wind power prediction in low-computing-power environments at the edge.

[0073] As can be seen from the above description, compared with the prior art, after adopting the technical solution of this invention, by introducing a sparse linear layer based on Fourier transform as the core operator in the prediction model, this operator is implemented by improving the underlying calculation method of deep learning. It utilizes the extreme value perturbation differentiable soft attention mechanism to dynamically filter the model parameters in the frequency domain dimension, retaining key frequency components, and reconstructing the weight matrix in the time domain using the filtered sparse frequency domain parameters through inverse Fourier transform (IFFT). This significantly compresses the model parameters, adapts to the low computing power environment of edge devices, and can directly replace the dense weight matrix of the traditional fully connected layer. While retaining the fitting ability of the fully connected layer, it significantly reduces the model parameters and computational complexity, thereby improving training efficiency, prediction real-time performance, and accuracy.

[0074] The technical solution, working process, and implementation effects of the present invention have been described in detail above. It should be noted that the described examples are only typical examples of the present invention. In addition, the present invention may have many other specific implementation methods. All technical solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection claimed by the present invention.

Claims

1. A lightweight edge-end wind farm power prediction system based on frequency domain sparse hybridization, the system being deployed at a wind farm site, comprising wind power generation equipment (1), a data monitoring and storage unit (2), a database (3), an edge computing gateway (4), and a power regulation and dispatch unit (5), characterized in that: The data monitoring and storage unit (2), the database (3), the edge computing gateway (4), and the power regulation and scheduling unit (5) constitute a closed-loop data processing and power prediction structure. The wind power generation equipment (1) is configured in the wind farm, including multiple wind turbine units and at least one wind measurement tower, for real-time collection of multi-source heterogeneous data on the wind turbine operation status and wind resource environment in the wind farm. The input end of the data monitoring and storage unit (2) is connected to the wind power generation equipment (1) and its output end is connected to the database (3) for real-time collection and uploading of measured data, and storage of multi-source heterogeneous data in the database (3). The edge computing gateway (4) is connected to the database (3) and integrates a non-stationary data preprocessing module (41), a Fourier sparse mixed prediction module (42), and a scheduling data uploading module (43). The power regulation and scheduling unit (5) is connected to the edge computing gateway (4) for receiving prediction results and generating wind farm power scheduling or control commands accordingly.

2. The lightweight edge-end wind farm power prediction system based on frequency domain sparse hybridization according to claim 1, characterized in that: The input end of the non-stationary data preprocessing module (41) is connected to the database (3) and is used to perform data cleaning, outlier removal and standardization on the input multi-source heterogeneous data to generate a regular feature dataset.

3. The lightweight edge-end wind farm power prediction system based on frequency domain sparse hybridization according to claim 1, characterized in that: The input of the Fourier sparse mixture prediction module (42) is connected to the output of the non-stationary data preprocessing module (41) to receive the feature dataset and calculate the wind farm power prediction value for a specific future period based on the built-in Fourier sparse mixture prediction model.

4. The lightweight edge-end wind farm power prediction system based on frequency domain sparse hybridization according to claim 1, characterized in that: The input end of the scheduling data uploading module (43) is connected to the output end of the Fourier sparse hybrid prediction module (42) to encapsulate the wind farm power prediction value according to a predetermined communication protocol and data format, and output it to the power regulation and scheduling unit (5).

5. A lightweight wind farm power prediction method based on frequency domain sparse mixing at the edge, characterized in that, Includes the following steps: The wind power generation equipment (1) is used to collect multi-source heterogeneous data on the wind turbine operation status and wind resource environment in real time, and the multi-source heterogeneous data is input into the database (3) for storage. When the edge computing gateway (4) sends a request to the database (3), it reads the latest historical time window data in the database (3) and provides the historical time window data to the non-stationary data preprocessing module (41) built into the edge computing gateway (4). After statistical correction, the data distribution drift is eliminated to obtain the preprocessed data. The preprocessed data is then input into the Fourier sparse mixture prediction module (42) built into the edge computing gateway (4). The Fourier sparse mixture prediction module (42) is based on the parameters of the Fourier sparse mixture prediction model completed offline. It dynamically identifies and filters the key frequency components in the preprocessed data through the extreme value perturbation differentiable soft attention mechanism. After the time domain weight is completely reconstructed in the frequency domain, the prediction of the wind farm's ultra-short-term or short-term power is realized, and the corresponding wind farm power prediction value is output. After the wind farm power prediction value is inversely normalized, it is input into the power regulation and dispatch unit (5) through the dispatch data uploading module (43) built into the edge computing gateway (4). The power regulation and dispatch unit (5) receives the wind farm power prediction value and generates corresponding wind farm power dispatch or control instructions based on the prediction value, and performs real-time forward-looking adjustment of the active power output of the wind farm to meet the power control requirements of the power grid dispatching agency.

6. The lightweight wind farm power prediction method based on frequency domain sparse hybridization at the edge as described in claim 5, characterized in that: The Fourier sparse mixture prediction module (42) is configured to specifically perform the following steps: Historical observation data of wind farms are read from the database, and the historical observation data are spliced ​​in chronological order to form a historical observation sequence. The historical observation sequence is then processed for time alignment and feature dimension expansion to generate a multi-dimensional time series input matrix. This multi-dimensional time series input matrix is ​​composed of one or more input sequences. Each input sequence is classified according to preset rules. For data sequences of different quality levels, the corresponding statistical correction strategy is automatically selected to obtain a regularized data sample. The Fourier sparse hybrid prediction module (42) integrates a sparse linear layer based on Fourier transform, which inputs regular data samples into the sparse linear layer based on Fourier transform to calculate the initial feature sequence for subsequent time-channel hybrid module to perform deep feature extraction. Multiple time-channel hybrid modules stacked sequentially are used to perform multi-level feature extraction on the input initial feature sequence. The input of the first time-channel hybrid module is connected to a sparse linear layer based on Fourier transform, and the output of the last time-channel hybrid module is connected to a linear projection layer. The linear projection layer is used to map the length of the feature sequence to the prediction length. Using the statistics saved in the statistical correction stage, the mapped feature sequence is denormalized to output the wind farm power prediction value.

7. The lightweight wind farm power prediction method based on frequency domain sparse hybridization at the edge as described in claim 5, characterized in that: The calculation process for each Fourier sparse linear layer in the Fourier sparse mixture prediction model is as follows: First, a learnable frequency domain parameter matrix is ​​constructed, and the frequency importance score is calculated to generate extreme value perturbation differentiable soft attention weights. After weighting the frequency domain parameters, the time domain weight matrix is ​​reconstructed using discrete inverse Fourier transform. After using the reconstructed weights to perform a linear transformation on the input data, a linear mapping is output.

8. The lightweight wind farm power prediction method based on frequency domain sparse hybridization at the edge as described in claim 5, characterized in that: The timing-channel mixing module includes a serial timing mixing sublayer and a channel mixing sublayer.

9. The lightweight wind farm power prediction method based on frequency domain sparse hybridization at the edge as described in claim 8, characterized in that: The temporal mixing sublayer is used to capture the periodic patterns of wind speed changes over time.

10. The lightweight wind farm power prediction method based on frequency domain sparse hybridization at the edge as described in claim 8, characterized in that: The channel mixing sublayer is used to capture the coupling relationship between different meteorological variables.

Images