Wind power prediction methods, devices, equipment, storage media and program products

Abstract

This application discloses a wind power prediction method, apparatus, equipment, storage medium, and program product, belonging to the field of wind power prediction technology. The method includes: collecting multidimensional meteorological data and historical power data, and preprocessing the multidimensional meteorological data and historical power data to obtain candidate data sequences; constructing a TimesNet time-series extraction module, and obtaining a first wind power prediction result based on the candidate data sequences and the TimesNet time-series extraction module; constructing a STAR channel modeling module, and obtaining a second wind power prediction result based on the candidate data sequences and the STAR channel modeling module; calculating dynamic weights, and adaptively weighting and fusing the first wind power prediction result and the second wind power prediction result according to the dynamic weights to obtain a future wind power prediction result. The embodiments of this application can improve the accuracy of wind power prediction.

Description

Technical Field

[0001] This application relates to the field of wind power prediction technology, and in particular to a wind power prediction method, apparatus, equipment, storage medium and program product. Background Technology

[0002] With the transformation of the global energy structure and the rapid development of renewable energy, wind power has become an important form of clean energy. Accurate wind power forecasting is of great significance for grid dispatch and the stable operation of the power system. High-precision wind power forecasting can not only improve the grid's capacity and reduce system operation risks, but also reduce the need for reserve capacity.

[0003] However, wind power has strong time dependence and multi-scale periodicity. It is affected by various periodic factors such as day-night cycle, seasonal changes, and weather system evolution. Traditional time series forecasting methods, such as LSTM, CNN, and Transformer, are difficult to capture the changing patterns at different time scales at the same time, which leads to insufficient accuracy in wind power forecasting. Summary of the Invention

[0004] The purpose of this application is to provide a wind power prediction method, apparatus, device, storage medium, and program product that can effectively improve the accuracy of wind power prediction.

[0005] To achieve the above objectives, a first aspect of this application provides a wind power prediction method, comprising:

[0006] Collect multidimensional meteorological data and historical power data, and preprocess the multidimensional meteorological data and historical power data to obtain candidate data sequences;

[0007] A TimesNet time series extraction module is constructed, and based on the candidate data sequence and the TimesNet time series extraction module, the first wind power prediction result is obtained;

[0008] A STAR channel modeling module is constructed, and based on the candidate data sequence and the STAR channel modeling module, a second wind power prediction result is obtained;

[0009] Calculate dynamic weights, and adaptively weight and fuse the first wind power prediction result and the second wind power prediction result according to the dynamic weights to obtain the future wind power prediction result.

[0010] Compared with existing technologies, the wind power prediction method provided in this application has the following advantages: by collecting and preprocessing multi-dimensional meteorological data and historical power data, key information affecting wind power is effectively preserved; by combining the TimesNet time series extraction module and the STAR channel modeling module to predict candidate data sequences, the temporal features and complex correlation features in the data can be captured from different dimensions, realizing multi-perspective mining of wind power related laws, and the prediction results of the two models are complementary; and by using dynamic weights for adaptive weighted fusion, the advantages of the two models can be fully utilized, solving the shortcomings of existing methods in capturing temporal features, and ultimately significantly improving the accuracy (precision) and robustness of wind power prediction, thereby providing a more reliable decision-making basis for the efficient scheduling of wind farms and the safe and stable operation of the power grid.

[0011] In some embodiments, the multidimensional meteorological data includes 11 meteorological characteristic elements: wind speed at 10 meters height, wind direction at 10 meters height, wind speed at 30 meters height, wind direction at 30 meters height, wind speed at 50 meters height, wind direction at 50 meters height, wind speed at hub height, wind direction at hub height, air temperature, atmospheric pressure, and relative humidity; the historical power data includes the actual power output of the wind turbine; the time resolution of both the multidimensional meteorological data and the historical power data is 15 minutes; the preprocessing of the multidimensional meteorological data and the historical power data to obtain candidate data sequences includes:

[0012] The multidimensional meteorological data and the historical power data are combined to obtain a multidimensional time series;

[0013] The multidimensional time series is cleaned to obtain a cleaned sequence;

[0014] The cleaned sequence is Z-score normalized to obtain the candidate data sequence.

[0015] In some embodiments, the construction of the TimesNet time-series extraction module, and the obtaining of the first wind power prediction result based on the candidate data sequence and the TimesNet time-series extraction module, includes:

[0016] Stack m TimesBlock blocks sequentially and add residual connections between adjacent TimesBlock blocks to obtain the TimesNet time series extraction module;

[0017] The candidate data sequence is used to predict wind power based on the TimesNet time series extraction module to obtain the first wind power prediction result.

[0018] In some embodiments, constructing the STAR channel modeling module and obtaining a second wind power prediction result based on the candidate data sequence and the STAR channel modeling module includes:

[0019] The STAR channel modeling module is constructed using n STAR layers;

[0020] The candidate data sequence is used to predict wind power based on the STAR channel modeling module to obtain a second wind power prediction result.

[0021] In some embodiments, the calculation of dynamic weights, which involves adaptively weighting and fusing the first wind power prediction result and the second wind power prediction result based on the dynamic weights to obtain the future wind power prediction result, includes:

[0022] The mean, variance, and amplitude of the multidimensional time series are extracted and concatenated with the flattened feature dimensions of the first and second wind power prediction results to form a gating vector.

[0023] The gated vector is sequentially fed into three fully connected layers to output a dynamic weight vector.

[0024] The first wind power prediction result and the second wind power prediction result are weighted and summed using the dynamic weight vector to obtain the future wind power prediction result.

[0025] In some embodiments, the wind power prediction method further includes:

[0026] The mean square error loss function is used to calculate the overall loss value of the future wind power prediction results, and the overall loss value is back-propagated to the TimesNet time series extraction module and the STAR channel modeling module through the backpropagation algorithm;

[0027] Adjust all learnable parameters within the TimesNet time series extraction module and the STAR channel modeling module based on the overall loss value.

[0028] To achieve the above objectives, a second aspect of this application provides a wind power prediction device, the wind power prediction device comprising:

[0029] The acquisition module is used to acquire multidimensional meteorological data and historical power data, and to preprocess the multidimensional meteorological data and historical power data to obtain candidate data sequences;

[0030] The first prediction module is used to construct a TimesNet time series extraction module and obtain a first wind power prediction result based on the candidate data sequence and the TimesNet time series extraction module.

[0031] The second prediction module is used to construct the STAR channel modeling module and obtain the second wind power prediction result based on the candidate data sequence and the STAR channel modeling module.

[0032] The calculation module is used to calculate dynamic weights and adaptively weight and fuse the first wind power prediction result and the second wind power prediction result according to the dynamic weights to obtain the future wind power prediction result.

[0033] To achieve the above objectives, a third aspect of this application provides an electronic device, the electronic device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the method described in the first aspect.

[0034] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium comprising a stored computer program, wherein the computer program, when executed, controls the device containing the computer-readable storage medium to perform the method described in the first aspect.

[0035] To achieve the above objectives, a fifth aspect of the present application provides a computer program product, which includes a computer program or computer instructions, wherein the computer program or computer instructions, when executed by a processor, implement the method described in the first aspect. Attached Figure Description

[0036] Figure 1 This is a flowchart of a wind power prediction method provided in an embodiment of this application;

[0037] Figure 2 yes Figure 1 A flowchart of step S101 in the process;

[0038] Figure 3 yes Figure 1 A flowchart of step S102 in the process;

[0039] Figure 4 yes Figure 1 A flowchart of step S103 in the process;

[0040] Figure 5 This is a dynamic weight change diagram of the TimesNet time series extraction module and the STAR channel modeling module provided in the embodiments of this application;

[0041] Figure 6 This is a comparison chart of the predicted future wind power and the actual wind power provided in the embodiments of this application;

[0042] Figure 7 This is a schematic diagram of a structure for obtaining a predicted value of future wind power, provided in an embodiment of this application.

[0043] Figure 8 This is a schematic diagram of the wind power prediction device provided in an embodiment of this application;

[0044] Figure 9 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0046] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0047] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0048] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0049] With the transformation of the global energy structure and the rapid development of renewable energy, wind power has become an important form of clean energy. Accurate wind power forecasting is of great significance for grid dispatch, stable power system operation, and the economic benefits of wind farms. High-precision wind power forecasting can not only improve the grid's capacity and reduce system operation risks, but also reduce reserve capacity requirements and improve overall economic efficiency. However, wind power forecasting faces many technical challenges that limit further improvements in forecast accuracy.

[0050] Wind power forecasting faces three main technical challenges: First, wind power exhibits strong time dependence and multi-scale periodicity, influenced by various periodic factors such as day-night cycles, seasonal variations, and weather system evolution. Traditional time series forecasting methods, such as LSTM, CNN, and Transformer, struggle to simultaneously capture the changing patterns at different time scales, exhibiting significant deficiencies in handling multi-period coupling and non-stationarity. Second, wind power forecasting involves the synergistic effects of multiple meteorological elements, including wind speed, wind direction, temperature, humidity, and air pressure, with complex nonlinear coupling relationships between variables. Existing forecasting models often employ simple linear combinations or fixed-weight fusion, making it difficult to adaptively learn the importance of different meteorological elements and their dynamic interactions. Third, single modeling strategies have inherent limitations. Time-series modeling methods have limited capabilities in multi-variable feature fusion, channel relationship modeling methods fail to capture long-term dependencies in time series, and fixed modeling architectures are ill-suited to adapt to forecasting needs under different meteorological conditions, resulting in limited model robustness and generalization ability.

[0051] Therefore, how to effectively integrate the advantages of time series feature extraction and multivariate relationship modeling, and adaptively adjust the fusion strategy according to the characteristics of input data to achieve more accurate and robust wind power prediction has become a technical problem that urgently needs to be solved.

[0052] Please see Figure 1 , Figure 1 This is an optional flowchart of the wind power prediction method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S104.

[0053] Step S101: Collect multidimensional meteorological data and historical power data, and preprocess the multidimensional meteorological data and historical power data to obtain candidate data sequences;

[0054] Step S102: Construct a TimesNet time series extraction module, and obtain the first wind power prediction result based on the candidate data sequence and the TimesNet time series extraction module;

[0055] Step S103: Construct the STAR channel modeling module, and obtain the second wind power prediction result based on the candidate data sequence and the STAR channel modeling module;

[0056] Step S104: Calculate the dynamic weights, and perform adaptive weighted fusion of the first wind power prediction result and the second wind power prediction result based on the dynamic weights to obtain the future wind power prediction result.

[0057] Steps S101 to S104 as illustrated in this embodiment effectively preserve key information affecting wind power by collecting and preprocessing multidimensional meteorological data and historical power data. By combining the TimesNet time series extraction module and the STAR channel modeling module to predict candidate data sequences, the temporal features and complex correlation features in the data can be captured from different dimensions, realizing multi-perspective mining of wind power-related laws. The prediction results of the two models are complementary. Furthermore, adaptive weighted fusion through dynamic weights can fully leverage the advantages of the two models, addressing the shortcomings of existing methods in capturing temporal features. Ultimately, this significantly improves the accuracy and robustness of wind power prediction, thereby providing a more reliable decision-making basis for the efficient scheduling of wind farms and the safe and stable operation of the power grid.

[0058] In step S101 of some embodiments, the multidimensional meteorological data includes 11 meteorological characteristic elements: wind speed at 10 meters height, wind direction at 10 meters height, wind speed at 30 meters height, wind direction at 30 meters height, wind speed at 50 meters height, wind direction at 50 meters height, wind speed at hub height, wind direction at hub height, air temperature, atmospheric pressure, and relative humidity. The historical power data includes the actual power output of the wind turbine. The time resolution of both the multidimensional meteorological data and the historical power data is 15 minutes. The candidate data sequence can be obtained by preprocessing the multidimensional meteorological data and the historical power data.

[0059] Please see Figure 2 In some embodiments, step S101 may include, but is not limited to, steps S201 to S203:

[0060] Step S201: Merge the multidimensional meteorological data and historical power data to obtain a multidimensional time series;

[0061] Step S202: Clean the multidimensional time series data to obtain the cleaned sequence. ;

[0062] Step S203, cleaning sequence Z-score standardization is performed to obtain candidate data sequences. .

[0063] In step S201 of some embodiments, the multidimensional time series can be a sequence of data formed by aligning multidimensional meteorological data and historical power data along the time dimension, with each time point containing multiple related feature values. For example, multidimensional meteorological data and historical power data are associated one-to-one according to timestamps to form a unified sequence containing meteorological features and power features.

[0064] In step S202 of some embodiments, data cleaning can be a process of processing the original data to eliminate abnormal or invalid information. The cleaned sequence can be a multidimensional time series obtained after data cleaning, which has removed outliers, missing values, and other problems. Data cleaning includes handling missing values ​​(e.g., using linear interpolation or mean filling), identifying and removing outliers (e.g., using the 3σ criterion or box plot method to detect data deviating from the normal range), and removing duplicate records. For missing, abnormal, or duplicate data existing in the multidimensional time series, appropriate methods are used to correct or remove them, resulting in a continuous, consistent, and cleaned sequence that conforms to actual physical laws, avoiding interference from noisy data in subsequent modeling.

[0065] It should be noted that data cleaning of multidimensional time series results in a cleaned sequence. ,include:

[0066] For short-term missing data, a moving average method is used to calculate the average value based on the sliding window data before and after the missing window to fill the gap; for long-term missing data, samples containing the missing segment are directly removed; for outliers, a box plot identification strategy is used to autonomously decide whether to remove the outlier.

[0067] The calculation formula for box plot recognition is as follows:

[0068] ;

[0069] in, and They are the 25th and 75th percentiles, respectively. The interquartile range; when the observed value exceeds [ , If the value falls within the specified range, it is identified as an outlier. At this point, you can decide whether to remove the outlier.

[0070] In step S203 of some embodiments, Z-score standardization is a data standardization method that transforms data into a distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of differences in the dimensions of different features. Candidate data sequences can be standardized sequences that have undergone Z-score standardization and can be directly used for model training and prediction.

[0071] The Z-score standardization formula is as follows:

[0072] ;

[0073] Among them, candidate data sequences For cleaning sequence The standardized sequence obtained after Z-score standardization For cleaning sequence The mean, For cleaning sequence The standard deviation.

[0074] This application embodiment calculates a cleaning sequence. Mean of each feature dimension and standard deviation Each data point is transformed using the Z-score standardization formula to ensure that the mean of each feature is 0 and the standard deviation is 1, thus eliminating dimensional differences and obtaining a standardized candidate data sequence. This effectively reduces the interference of data noise and format differences on model performance, laying a reliable data foundation for improving the accuracy of wind power prediction.

[0075] In step S102 of some embodiments, the first wind power prediction result can be a prediction result calculated based on the candidate data sequence and the TimesNet time series extraction module.

[0076] Please see Figure 3 In some embodiments, step S102 may include, but is not limited to, steps S301 to S302:

[0077] Step S301: Stack m TimesBlock blocks sequentially and add residual connections between adjacent TimesBlock blocks to obtain the TimesNet time series extraction module;

[0078] Step S302: Based on the TimeNet time series extraction module, wind power prediction is performed on the candidate data sequence to obtain the first wind power prediction result. .

[0079] In step S301 of some embodiments, the TimesBlock is a fundamental component of the TimesNet model, typically containing sub-modules such as temporal decomposition and temporal feature extraction, which can be used to capture local and global temporal patterns in time series. Residual connections are skip connections added between adjacent network modules, adding the output of the previous module to the input of the current module. This can alleviate the gradient vanishing problem in deep network training and enhance feature propagation capabilities. The TimesNet temporal extraction module can be a network structure composed of multiple stacked TimesBlocks connected by residual connections, specifically designed to extract multi-level, multi-scale temporal features from time series.

[0080] In step S302 of some embodiments, the first wind power prediction result can be the wind power prediction value output after the candidate data sequence is processed by the TimesNet time series extraction module.

[0081] In some embodiments, the specific components include:

[0082] Candidate data sequences Splitting by feature channels yields multiple one-dimensional sequences. The feature channel is each variable dimension in a multidimensional time series.

[0083] Based on the Fast Fourier Transform, from multiple one-dimensional sequences The frequency component information and periodic characteristics of wind power data are extracted using the following formula:

[0084] ;

[0085] in, This represents the one-dimensional sequence output by the (m-1)th TimesBlock. One-dimensional sequence Length; For Fast Fourier Transform; Take the magnitude of the FFT result; The amplitude is averaged over the period length to obtain the intensity of each possible period; It is a length of A vector that stores the intensity of all possible periods; from Select the indices of the k periods with the highest intensity; The k most significant components were selected based on their intensity. This indicates that these pre-selected components The corresponding period length.

[0086] Based on the previously calculated period and frequency, the one-dimensional sequence is converted into multiple two-dimensional tensors and padded to ensure consistent output size after subsequent convolution operations; the formula is as follows:

[0087] ;

[0088] in, This indicates a dimensional ascension operation. This indicates padding the end of the sequence with zeros.

[0089] For the transformed two-dimensional tensor, local features within and between periods are extracted using the following formula:

[0090] ;

[0091] in, This represents the operation methods in the Inception Block module, which includes convolutional layers and pooling layers with multiple convolutional kernels.

[0092] The extracted two-dimensional feature tensor is mapped back to multiple one-dimensional tensors, as shown in the following formula:

[0093] ;

[0094] in, This indicates a dimensionality reduction operation. This indicates a length truncation operation to ensure that the length of the output sequence is the same as the length of the input sequence.

[0095] The final result is obtained by weighted summation of the reshaped one-dimensional tensors, with the weights calculated based on the amplitude of each period, as shown in the following formula:

[0096] ;

[0097] in, Represents the normalization function. Indicates weight, This is the multi-period feature fusion result for the current m-th TimesBlock block; Generally, it is expressed as follows:

[0098] ;

[0099] in, The denominator is a natural constant, ensuring that the sum of all weights is 1, i.e. This enables adaptive weighted fusion based on periodic importance.

[0100] After processing by m TimesBlock blocks as described above, the first wind power prediction result is generated through a linear projection layer:

[0101] ;

[0102] in, Indicates the predicted projection layer. This indicates the first wind power prediction result.

[0103] This application introduces a TimesNet time series extraction module and uses Fast Fourier Transform to adaptively identify multiple periodic patterns in wind power data. It combines Inception Block multi-kernel convolution operations to achieve two-dimensional feature extraction, avoiding the gradient vanishing problem of traditional recurrent neural networks (RNN, LSTM) and the limitation of convolutional neural networks (CNN) in capturing long-term time series dependencies. This improves the model's ability to represent complex nonlinear time series features of wind power and the accuracy of long-term prediction.

[0104] In step S103 of some embodiments, the STAR channel modeling module is a module that focuses on modeling the correlation between characteristic channels, used to capture the nonlinear correlation between meteorological factors and power. The second wind power prediction result can be the wind power prediction value output by the STAR channel modeling module.

[0105] A STAR channel modeling module is constructed, which focuses on analyzing cross-channel correlations between different features. It captures the complex nonlinear relationship between meteorological factors and power through attention mechanisms or feature interaction layers, and outputs the second wind power prediction result based on candidate data sequences.

[0106] Please see Figure 4 In some embodiments, step S103 may include, but is not limited to, steps S401 to S402:

[0107] Step S401: Construct a STAR channel modeling module with n STAR layers;

[0108] Step S402: Based on the STAR channel modeling module, wind power prediction is performed on the candidate data sequence to obtain the second wind power prediction result. .

[0109] In step S401 of some embodiments, the STAR layer is the basic unit of the STAR channel modeling module, which typically integrates self-attention, feature transformation, and residual connection mechanisms to capture dependencies and nonlinear interactions between feature channels. The STAR channel modeling module can be a network structure formed by stacking n STAR layers, focusing on mining cross-channel correlations and complex interaction patterns between multidimensional features (such as meteorological factors like wind speed and temperature and power).

[0110] A multi-channel relationship modeling module based on the STAR attention mechanism is constructed to learn the interactions between features: First, the number of STAR layers, n, is determined, and these n STAR layers are stacked sequentially to form a deep network structure. The number n can be adjusted according to the feature dimension and complexity, and no specific limit is imposed here. Within each STAR layer, the association weights of different feature channels are calculated through a self-attention mechanism. The non-linear expressive power is enhanced by a feature transformation layer, and residual connections are added to strengthen feature propagation.

[0111] In step S402 of some embodiments, the second wind power prediction result can be the wind power prediction value output after the candidate data sequence is processed by the STAR channel modeling module.

[0112] During initialization, the candidate data sequence is... Transpose the channel and project it linearly to obtain the channel representation, as shown in the following formula:

[0113] ;

[0114] in, This represents a linear embedding transformation.

[0115] For the j-th STAR layer, global core information is extracted through a two-layer nonlinear transformation, as shown in the following formula:

[0116] ;

[0117] in, and Indicates a fully connected layer. and These represent the output features of the two fully connected layers, Represents the Gaussian error linear unit activation function. This represents the output of the previous STAR layer; FC can generally be represented as:

[0118] ;

[0119] Where W is the learnable weight matrix of the fully connected layer, b is the learnable bias parameter of the fully connected layer, and x is the input feature vector; It can generally be expressed as:

[0120] ;

[0121] Channel attention weights are calculated using the Softmax function, and a global core representation is generated; the formula is as follows:

[0122] ;

[0123] Where L represents the number of feature channels, This represents the channel weight of the j-th layer. The core representation of weighted aggregation. It is a unit vector.

[0124] The local channel information is fused with the weighted aggregated global core representation, as shown in the following formula:

[0125] ;

[0126] in, [;] indicates the channel fusion result; [;] indicates the tensor splicing operation. and This represents the fully connected layer of a converged network. This indicates a modified linear unit activation function; It can generally be expressed as:

[0127] ;

[0128] Channel fusion results Add residual connections to maintain information flow; the formula is as follows:

[0129] ;

[0130] in, This represents the output of the j-th STAR layer. Indicates the first Output of the STAR layer.

[0131] After processing through n STAR layers, the channel relationship features can be mapped to wind power prediction results via a linear prediction layer; the formula is as follows:

[0132] ;

[0133] in, This represents the final output of the nth STAR layer. Indicates the prediction mapping layer, This indicates the second wind power forecast result.

[0134] Inspired by the STAR star topology, this application adopts a centralized information aggregation and redistribution mechanism. By coordinating the information exchange among multidimensional meteorological variables through a global core node, it avoids the shortcomings of the traditional fully connected attention mechanism, which has excessive computational complexity and ignores the coupling relationship of variables by assuming channel independence. This enables efficient modeling of complex nonlinear interactions among multiple meteorological factors such as wind speed, temperature, humidity, and air pressure, and enhances the model's ability to represent multivariate time series data and its generalization performance.

[0135] In some embodiments, step S102 is executed first, followed by step S103; in other embodiments, step S103 is executed first, or both are executed simultaneously, all of which can achieve the technical effects of this application.

[0136] In step S104 of some embodiments, the dynamic weights can be adaptively adjusted weight values ​​based on data characteristics or the prediction scenario, rather than fixed values. Adaptive weighted fusion refers to the process of weighting and summing two prediction results based on dynamic weights to make the fused result more closely resemble the actual scenario. Dynamic weights are calculated, and the first and second prediction results are weighted and fused according to their weights. For example, the weight of the TimesNet result is increased when time-series features are more significant, and the weight of the STAR result is increased when cross-feature correlation is more critical, ultimately obtaining a future wind power prediction result that takes into account both time-series patterns and feature correlations.

[0137] This application adopts a parallel processing architecture of the TimesNet time series extraction module and the STAR channel modeling module, which avoids the error propagation and information bottleneck problems of the front-end modules in the traditional serial architecture. This allows the two modules to extract complementary feature information from the original input independently and simultaneously, overcomes the limitation of the single model structure's limited representation ability, ensures the integrity and independence of time series features and variable relationship features extraction, and improves the overall model's computational efficiency and feature representation comprehensiveness.

[0138] Specifically, the mean, variance, and amplitude of the multidimensional time series are extracted and compared with the first wind power prediction result. Second wind power forecast results Flattened back edge feature dimensions are spliced ​​together to form the input vector of the gating network. The formula is as follows:

[0139]

[0140] in, The mean, variance, and amplitude characteristics of a multidimensional time series can characterize the meteorological environment and power generation status of the current period, providing an important basis for the gating network to judge the reliability of the two prediction modules. Indicates the input of the gating network. This indicates that the prediction result will be flattened; [;] indicates a splicing operation.

[0141] Gating vector The weights are sequentially fed into three fully connected layers, and a dynamic weight vector is output. ,in, It is the first priority corresponding to the first wind power prediction result; It is the second weight corresponding to the second wind power prediction result. ;

[0142] Optionally, the fusion weights of the two prediction modules are calculated using a three-layer gating network, as shown in the following formula:

[0143] ;

[0144] in, , , These represent three fully connected layers. This represents the activation function. Ensure weight normalization. This represents the output features of the first hidden layer. This represents the output features of the first hidden layer. This represents the dynamic weights of the TimesNet time series extraction module. This represents the dynamic weights of the STAR channel modeling module; such as... Figure 5 As shown, Figure 5 This is the dynamic weight change diagram of the TimesNet time series extraction module and the STAR channel modeling module in the embodiment.

[0145] The prediction results from the two modules are adaptively weighted and fused based on the dynamic weight vector to obtain the final prediction result, as shown in the following formula:

[0146] ;

[0147] in, This is the final wind power prediction output of the hybrid model.

[0148] In other words, with , Predicted results for the first wind power output Second wind power forecast results By performing element-wise weighted summation, the future wind power prediction results are obtained. This application constructs an adaptive gated fusion network, avoiding the limitations of traditional fixed-weight fusion methods that cannot adapt to dynamic changes and the shortcomings of simple average fusion that ignore model differences. It dynamically adjusts the fusion strategy based on input data characteristics and the prediction confidence of each module, achieving intelligent collaboration between two heterogeneous mechanisms: time-series modeling and multivariate modeling. Furthermore, the visualization analysis of gating weights provides interpretability of the model's decision-making process, clearly demonstrating the contribution and importance distribution of each sub-module under different meteorological conditions, thus enhancing the model's decision-making transparency and credibility.

[0149] This application's embodiments obtain high-quality candidate data sequences through acquisition and preprocessing. Two prediction results are generated using the TimesNet time-series extraction module (capturing time-series features) and the STAR channel modeling module (capturing feature correlations), respectively. These results are then dynamically weighted and adaptively fused to output the final wind power prediction result. This method combines the advantages of both time-series features and cross-channel correlation features. The dynamic weighted fusion mechanism adaptively leverages the strengths of both models based on data characteristics, effectively compensating for the limitations of a single model and significantly improving the accuracy and robustness of wind power prediction. It is particularly stable under complex weather conditions, providing more reliable decision support for wind farm scheduling and grid planning.

[0150] In some embodiments, this application further includes establishing and training an end-to-end hybrid prediction model; constructing an end-to-end hybrid prediction model architecture with candidate data sequences as a unified input; generating prediction results by parallel processing of a TimesNet-based time series extraction module and a STAR-based channel modeling module; combining the extracted statistical features of the input data with the prediction outputs of the two modules, performing dynamic weight allocation and weighted fusion through an adaptive gated fusion network, and finally outputting a wind power prediction sequence within the future time window; this end-to-end architecture achieves collaborative optimization of all module parameters through a unified loss function, ensuring the organic combination of the three core components of time series modeling, channel relationship learning, and adaptive fusion, thereby improving overall prediction performance.

[0151] This application avoids the problems of inconsistent objectives and local optima among components in traditional step-by-step optimization methods by adopting an end-to-end hybrid learning framework. It realizes global collaborative training and unified optimization of three core components: time-series feature extraction, multivariate relationship modeling, and adaptive fusion. It constructs a complete wind power prediction solution, improves the prediction performance, computational efficiency, and engineering practicality of the overall system, and provides reliable technical support for wind farm power prediction and smart grid scheduling.

[0152] In some embodiments, this application further includes calculating the overall loss value of the future wind power prediction result using the mean square error loss function, and transmitting the overall loss value back to the TimesNet time series extraction module and the STAR channel modeling module through the back propagation algorithm;

[0153] The overall loss value is used to adjust all learnable parameters within the TimesNet time series extraction module and the STAR channel modeling module to make the next predicted wind power value closer to the actual wind power value. The formula for calculating the overall loss value is as follows:

[0154]

[0155] in, This represents the actual wind power output. This indicates the forecast results for future wind power output. For batch size, To predict the length of the time period, This represents the total loss value. Let the mean squared error loss function be . Let be the actual wind power value of the i-th sample at the j-th prediction time point. Let be the predicted wind power value of the i-th sample at the j-th prediction time point.

[0156] This application employs the mean squared error loss function to supervise the learning of the final fusion prediction result. Through the backpropagation algorithm, it simultaneously optimizes all learnable parameters of the three components—TimesNet temporal feature extraction module, STAR multi-channel relationship modeling module, and adaptive fusion network—end-to-end, thereby achieving collaborative training of the overall model.

[0157] In some embodiments, this application further includes using three evaluation metrics to quantitatively analyze the prediction results: mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (CQD). This method can comprehensively evaluate the prediction accuracy, stability, and generalization ability of hybrid prediction models; the formula is as follows:

[0158] ;

[0159] in, This represents the actual wind power output. This represents the predicted future wind power output, where N is the batch size and H is the prediction period length. Let be the actual wind power value of the i-th sample at the t-th prediction time. Let be the predicted wind power value of the i-th sample at the t-th prediction time.

[0160] This application embodiment generates future wind power prediction results based on meteorological data and power data within a historical time window; Figure 6 This is a comparison chart of the predicted future wind power (predicted power value) and the actual wind power (actual power value) based on this embodiment.

[0161] In order to quantitatively evaluate the predictive performance of the method in the embodiments of this application and verify the contribution effect of each core module, an ablation experiment was designed.

[0162] Table 1 Ablation Experiment Results

[0163]

[0164] Table 1 shows the experimental results, indicating that this method exhibits excellent predictive performance across all evaluation metrics, with MAE (0.0883), RMSE (0.111), and [other metrics not specified]. (0.987); Compared with using the TimesNet time series extraction module alone, using the STAR channel modeling module alone, and the simple averaging fusion method, the adaptive fusion strategy of this application has achieved significant improvements in both prediction accuracy and stability, fully verifying the effectiveness and necessity of the parallel architecture design and the adaptive gating fusion mechanism.

[0165] like Figure 7 As shown, Figure 7 This application provides a schematic diagram of a structure for obtaining future wind power prediction values. Figure 7 The detailed process and the meaning of each parameter are explained below from left to right:

[0166] On the left, in the historical data input stage: the system first receives a historical data matrix from time step t-h+l to t, which contains 11 types of meteorological data and wind power data. After data preprocessing such as outlier handling and standardization, this historical data is simultaneously distributed to two parallel prediction modules (TimesNet time series extraction module and STAR channel modeling module) for prediction.

[0167] Intermediate Parallel Processing Stage: The TimesNet temporal feature extraction module (TimesNet temporal extraction module) processes historical data through multiple residually connected TimesBlocks. Within each TimesBlock, FFT is first used for frequency domain analysis to extract k dominant periodic components and their frequency domain weights. Then, the one-dimensional time series data is reshaped into a two-dimensional tensor according to the period length. This tensor is then processed in parallel using multi-scale convolutional kernels from the Inception Block to effectively capture local patterns within the period and long-range dependencies between periods. Finally, the importance of different periodic components is weighted using Softmax normalization, outputting the prediction results from the TimesNet temporal feature extraction module. (Right now The multi-channel relationship modeling module (STAR ​​channel modeling module) below processes data through multiple STAR units connected by residuals. First, an MLP layer aggregates the local sequence information of each channel into a global core representation, encoding the complex dependencies between all channels. Then, the core information is distributed to each channel to achieve efficient cross-channel information transfer. Next, the original input and the core information are concatenated and fused to form an enhanced representation. Subsequent MLP layers then perform nonlinear transformation and dimensionality reconstruction. Finally, the prediction results of the multi-channel relationship modeling are output through residual connections. (Right now ).

[0168] Right-side adaptive fusion stage: Extracting simple statistical features from historical meteorological and power data. (Right now ), combined and The feature representations of the prediction results are concatenated to form the gating decision features. ; Normalized weights are output after passing through an adaptive fusion network. and , representing the credibility of the TimesNet time-series feature extraction module and the multi-channel relationship modeling module, respectively; finally, a weighted fusion formula is used. The system generates wind power forecasts for time steps t+1 to t+p. Historical statistical characteristics serve as environmental information to guide weight allocation decisions. and It participates in adaptive fusion as the actual predicted output.

[0169] The adaptive fusion wind power prediction method proposed in this application, which combines a TimesNet time-series extraction module (also known as a TimesNet time-series feature extraction module) and a STAR channel modeling module (also known as a STAR multi-channel relationship modeling module or multi-channel relationship modeling module), demonstrates multi-dimensional breakthrough value in terms of technical path and theoretical innovation. This method achieves collaborative modeling of complex periodic patterns and coupling relationships between multiple meteorological variables in wind power time-series data by constructing a parallel processing architecture of the TimesNet time-series feature extraction module and the STAR attention-based multi-channel relationship modeling module. This overcomes the technical bottlenecks of limited representation capabilities of traditional single-model structures and information transmission loss in serial architectures. The method introduces an adaptive gated fusion network, dynamically adjusting the weight allocation of each sub-module according to the characteristics of the input data, realizing intelligent collaboration of heterogeneous prediction mechanisms and enhancing the interpretability of the model decision-making process. Through an end-to-end hybrid learning framework, the three core components of time-series feature extraction, multivariate relationship modeling, and adaptive fusion are uniformly optimized, constructing a complete wind power prediction solution. This innovative method effectively solves key technical problems currently faced in the field of wind power prediction, such as long-term reliance on modeling difficulties, improper handling of multivariate coupling relationships, rigid model fusion strategies, and lack of interpretability of prediction results. It provides an important theoretical foundation and technical support for the deep integration of wind farm power prediction and smart grid dispatch.

[0170] The innovative advantages of this method are mainly reflected in the following four dimensions:

[0171] First, the time series feature modeling mechanism breaks through the performance bottleneck of a single model. The TimesNet module adaptively identifies multiple periodic patterns through fast Fourier transform and combines the Inception Block two-dimensional convolution operation to avoid the limitations of gradient vanishing in traditional recurrent neural networks and the long-term reliance on insufficient modeling in convolutional neural networks, thereby improving the ability to analyze complex nonlinear time series features.

[0172] Secondly, the multivariate relationship modeling uses STAR's centralized information aggregation and redistribution mechanism. By coordinating the information exchange between multidimensional meteorological variables through global core nodes, it avoids the shortcomings of the traditional fully connected attention mechanism, which has excessive computational complexity and ignores coupling relationships by assuming channel independence. This achieves efficient modeling of the interaction of multiple meteorological factors.

[0173] Third, the parallel processing architecture eliminates the information bottleneck and error propagation problems in the traditional serial method, enabling the two core modules to independently extract complementary features from the original input, avoiding mutual interference between modules, ensuring the integrity and independence of feature extraction, and improving the overall system's computational efficiency and robustness.

[0174] Fourth, the adaptive gating fusion network achieves dynamic weight allocation and decision transparency. It intelligently adjusts the fusion strategy based on the characteristics of the input data and the prediction confidence of each module, overcoming the problem of insufficient adaptability of the fixed weight method. At the same time, it provides interpretability of the model decision process through weight visualization, enhancing the credibility and practicality of the prediction results.

[0175] Please see Figure 8 This application also provides a wind power prediction device that can implement the above-described wind power prediction method. The device includes:

[0176] The acquisition module 801 is used to acquire multidimensional meteorological data and historical power data, and to preprocess the multidimensional meteorological data and historical power data to obtain candidate data sequences;

[0177] The first prediction module 802 is used to construct a TimesNet time series extraction module and obtain the first wind power prediction result based on the candidate data sequence and the TimesNet time series extraction module.

[0178] The second prediction module 803 is used to construct the STAR channel modeling module and obtain the second wind power prediction result based on the candidate data sequence and the STAR channel modeling module;

[0179] The calculation module 804 is used to calculate dynamic weights and adaptively weight and fuse the first wind power prediction result and the second wind power prediction result according to the dynamic weights to obtain the future wind power prediction result.

[0180] The specific implementation of this wind power prediction device is basically the same as the specific implementation of the wind power prediction method described above, and will not be repeated here.

[0181] Thirdly, embodiments of this application provide an electronic device, see [link to relevant documentation]. Figure 9 The diagram shown is a structural schematic of an electronic device provided in this application.

[0182] like Figure 9 As shown, the device includes:

[0183] Memory 31 is used to store computer programs;

[0184] Processor 32 is used to execute computer programs;

[0185] When the processor 32 executes the computer program, it implements the wind power prediction method as described in any of the above embodiments.

[0186] For example, a computer program may be divided into one or more modules / units, one or more of which are stored in memory 31 and executed by processor 32 to complete this application. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in an electronic device.

[0187] The processor 32 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0188] The memory 31 can be used to store computer programs and / or modules. The processor 32 implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory 31 and calling the data stored in the memory 31. The memory 31 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 31 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0189] It should be noted that the aforementioned electronic devices include, but are not limited to, processors and memory, as will be understood by those skilled in the art. Figure 9 The structural diagram is merely an example of the electronic device described above and does not constitute a limitation on the electronic device. It may include more components than shown in the diagram, or combine certain components, or use different components.

[0190] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed, implements the wind power prediction method of any of the above embodiments.

[0191] It should be understood that the implementation of all or part of the above-described wind power prediction method can also be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-described wind power prediction method. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the relevant jurisdiction. For example, in some relevant jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0192] Fifthly, embodiments of this application also provide a computer program product, which is stored in a storage medium and executed by at least one processor to implement the wind power prediction method of any of the above embodiments.

[0193] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0194] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for predicting wind power output, characterized in that, include: Collect multidimensional meteorological data and historical power data, and preprocess the multidimensional meteorological data and historical power data to obtain candidate data sequences; A TimesNet time series extraction module is constructed, and a first wind power prediction result is obtained based on the candidate data sequence and the TimesNet time series extraction module. The TimesNet time series extraction module is used to adaptively identify multiple periodic patterns in wind power data through fast Fourier transform, and convert one-dimensional time series data into two-dimensional tensors to extract local features within and between periods. A STAR channel modeling module is constructed, and a second wind power prediction result is obtained based on the candidate data sequence and the STAR channel modeling module. The STAR channel modeling module is used to adopt a centralized information aggregation and redistribution mechanism to coordinate the information exchange between multiple variables in the multidimensional meteorological data through a global core node, so as to model the nonlinear coupling relationship between meteorological factors and power. Calculate the dynamic weights, and adaptively weight and fuse the first wind power prediction result and the second wind power prediction result according to the dynamic weights to obtain the future wind power prediction result. The multidimensional meteorological data includes 11 meteorological characteristic elements: wind speed at 10 meters height, wind direction at 10 meters height, wind speed at 30 meters height, wind direction at 30 meters height, wind speed at 50 meters height, wind direction at 50 meters height, wind speed at hub height, wind direction at hub height, air temperature, atmospheric pressure, and relative humidity. The historical power data includes the actual power output of the wind turbine. The time resolution of both the multidimensional meteorological data and the historical power data is 15 minutes. The preprocessing of the multidimensional meteorological data and the historical power data to obtain candidate data sequences includes: merging the multidimensional meteorological data and the historical power data to obtain a multidimensional time series; cleaning the multidimensional time series to obtain a cleaned sequence; and performing Z-score standardization on the cleaned sequence to obtain candidate data sequences. The wind power prediction method further includes: calculating the overall loss value of the future wind power prediction result using the mean square error loss function, and backpropagating the overall loss value to the TimesNet time series extraction module and the STAR channel modeling module through the backpropagation algorithm; and adjusting all learnable parameters inside the TimesNet time series extraction module and the STAR channel modeling module according to the overall loss value.

2. The wind power prediction method as described in claim 1, characterized in that, The construction of the TimesNet time series extraction module, and the obtaining of the first wind power prediction result based on the candidate data sequence and the TimesNet time series extraction module, includes: Stack m TimesBlock blocks sequentially and add residual connections between adjacent TimesBlock blocks to obtain the TimesNet time series extraction module; The candidate data sequence is used to predict wind power based on the TimesNet time series extraction module to obtain the first wind power prediction result.

3. The wind power prediction method as described in claim 1, characterized in that, The construction of the STAR channel modeling module, and the obtaining of the second wind power prediction result based on the candidate data sequence and the STAR channel modeling module, includes: The STAR channel modeling module is constructed using n STAR layers; The candidate data sequence is used to predict wind power based on the STAR channel modeling module to obtain a second wind power prediction result.

4. The wind power prediction method according to any one of claims 1-3, characterized in that, The calculation of dynamic weights involves adaptively weighting and fusing the first wind power prediction result and the second wind power prediction result based on the dynamic weights to obtain the future wind power prediction result, including: The mean, variance, and amplitude of the multidimensional time series are extracted and concatenated with the flattened feature dimensions of the first and second wind power prediction results to form a gating vector. The gated vector is sequentially fed into three fully connected layers to output a dynamic weight vector. The first wind power prediction result and the second wind power prediction result are weighted and summed using the dynamic weight vector to obtain the future wind power prediction result.

5. A wind power prediction device, characterized in that, The wind power prediction device includes: The data acquisition module is used to collect multidimensional meteorological data and historical power data, and preprocess the multidimensional meteorological data and historical power data to obtain candidate data sequences. The multidimensional meteorological data includes 11 meteorological characteristic elements: wind speed at 10 meters height, wind direction at 10 meters height, wind speed at 30 meters height, wind direction at 30 meters height, wind speed at 50 meters height, wind direction at 50 meters height, wind speed at hub height, wind direction at hub height, air temperature, atmospheric pressure, and relative humidity. The historical power data includes the actual power output of the wind turbine. The time resolution of both the multidimensional meteorological data and the historical power data is 15 minutes. The preprocessing of the multidimensional meteorological data and the historical power data to obtain candidate data sequences includes: merging the multidimensional meteorological data and the historical power data to obtain a multidimensional time series; cleaning the multidimensional time series to obtain a cleaned sequence; and performing Z-score normalization on the cleaned sequence to obtain candidate data sequences. The first prediction module is used to construct a TimesNet time series extraction module and obtain a first wind power prediction result based on the candidate data sequence and the TimesNet time series extraction module. The TimesNet time series extraction module is used to adaptively identify multiple periodic patterns in wind power data through fast Fourier transform and convert one-dimensional time series data into two-dimensional tensors to extract local features within and between periods. The second prediction module is used to construct the STAR channel modeling module and obtain the second wind power prediction result based on the candidate data sequence and the STAR channel modeling module. The STAR channel modeling module is used to adopt a centralized information aggregation and redistribution mechanism to coordinate the information exchange between multiple variables in the multidimensional meteorological data through global core nodes, so as to model the nonlinear coupling relationship between meteorological factors and power. The calculation module is used to calculate dynamic weights and adaptively weight and fuse the first wind power prediction result and the second wind power prediction result according to the dynamic weights to obtain the future wind power prediction result. The wind power prediction device further includes: calculating the overall loss value of the future wind power prediction result using the mean square error loss function, transmitting the overall loss value back to the TimesNet time series extraction module and the STAR channel modeling module through the backpropagation algorithm; and adjusting all learnable parameters within the TimesNet time series extraction module and the STAR channel modeling module according to the overall loss value.

6. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the wind power prediction method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the wind power prediction method as described in any one of claims 1 to 4.

8. A computer program product, characterized in that, The computer program product includes a computer program or computer instructions, which, when executed by a processor, implement the wind power prediction method as described in any one of claims 1 to 4.

Images