An ultra-short-term offshore wind power prediction method considering multi-wind farm information

Abstract

The application discloses a kind of considering the ultra-short-term offshore wind power prediction method of multiple wind farm information, including obtaining the historical wind power sequence and meteorological variable sequence of target offshore wind farm and adjacent offshore wind farm;Each sequence is carried out differentiable multivariate variation mode decomposition, and modal component is extracted;Phase attention mechanism is constructed based on modal component, the adjacent wind farm with the highest correlation degree associated with target wind farm is screened, and the instantaneous phase difference of target and adjacent wind farm is calculated by analyzing signal, and dynamic time delay feature is obtained.Target wind farm modal component, adjacent wind farm modal component, dynamic time delay feature, historical power sequence and meteorological variable are carried out feature fusion, as the input of prediction model.The prediction model uses the selective state space model of multilayer Mamba module stacking.The application realizes the effective use and accurate prediction of cross wind farm information, and significantly improves the prediction accuracy under complex weather conditions.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, specifically relating to an ultra-short-term offshore wind power prediction method that considers information from multiple wind farms. Background Technology

[0002] To effectively integrate large-scale wind power and ensure grid security, high-precision wind power forecasting technology has become an indispensable key component. Accurate forecasting not only provides decision support for power system dispatching but also significantly improves the utilization efficiency and economic benefits of wind energy, making it one of the core technologies for realizing the vision of a future smart grid. Therefore, improving the accuracy of wind power forecasting is of great significance for ensuring grid security and enhancing wind power absorption capacity.

[0003] Traditional wind power forecasting methods mostly utilize information from individual wind farms, treating each wind farm as an isolated system and completely ignoring their spatial correlations. This leads to insufficient information utilization and inherent bottlenecks in forecasting performance. However, existing spatiotemporal collaborative forecasting models still face key technical challenges in mining the value of multi-source data. First, signals from different wind farms or meteorological variables often contain different frequency scale characteristics. Existing methods mostly use direct splicing as input, leading to multi-scale feature confusion and difficulty in achieving effective scale alignment. Second, wind propagation paths and intensities change dynamically with meteorological conditions; not all nearby wind farms can provide effective information at all times. Existing models lack dynamic information selection mechanisms, easily introducing irrelevant noise and reducing forecast robustness. Third, wind energy propagation from upstream to downstream involves dynamic physical time delays. Existing technologies mostly simplify this to static delays, failing to accurately capture and quantify this time-varying physical propagation process, severely limiting further improvements in forecast accuracy.

[0004] For example, in the prior art, CN117893362A proposes to predict offshore wind power by fusing multi-source meteorological data and combining attention mechanism with recurrent neural network model. It improves prediction performance by feature screening and weighting of multidimensional meteorological variables and mining time series features using bidirectional gated recurrent unit network. However, these methods, which mainly focus on feature fusion and weight allocation of multi-source meteorological variables, still have the following shortcomings: First, these methods usually directly model the original time series features, lacking the decomposition and alignment of signal components at different time scales, making it difficult to effectively distinguish and utilize the wind power variation patterns at different frequency scales; Second, when utilizing information from neighboring wind farms, spatial relationships are often established through geographical region division or graph structure weighting, lacking a fine-grained screening mechanism based on the dynamic correlation of actual signals, making it difficult to adaptively select the most valuable neighboring wind farm information under different meteorological conditions; In addition, these methods usually mine spatial relationships from the perspective of statistical correlation, failing to explicitly model the dynamic physical time delays that exist in the propagation of wind energy between different wind farms, thus making it difficult to accurately characterize the propagation characteristics of power changes across wind farms; At the same time, sequence modeling methods based on recurrent neural networks are prone to gradient decay or information loss problems when dealing with long-term sequence dependencies, and their ability to model complex long-sequence spatiotemporal correlation features is still limited.

[0005] Therefore, in order to solve the problems of multi-scale alignment difficulties, blind selection of spatial information, and difficulty in accurately quantifying the physical delay of wind energy propagation, it is urgent to develop a prediction method that integrates physical prior knowledge and the advantages of deep learning. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an ultra-short-term offshore wind power prediction method that considers information from multiple wind farms.

[0007] The objective of this invention can be achieved through the following technical solutions: This invention provides an ultra-short-term offshore wind power prediction method that considers information from multiple wind farms, comprising the following steps: The historical wind power sequence of the target offshore wind farm, the historical wind power sequence of neighboring offshore wind farms, and the meteorological variable sequence of the target offshore wind farm are obtained and preprocessed. Adaptive mode decomposition is performed on the preprocessed historical wind power sequence of the target offshore wind farm and the historical wind power sequence of the adjacent offshore wind farm to obtain multiple mode components; The correlation between the target offshore wind farm and each neighboring offshore wind farm is calculated based on the modal components, and the neighboring offshore wind farm with the highest correlation with the target offshore wind farm is selected. Phase features are extracted from the modal components of the target offshore wind farm and the modal components of the selected neighboring offshore wind farms to obtain dynamic time delay features characterizing the propagation relationship between different offshore wind farms. The modal components of the target offshore wind farm, the modal components of neighboring offshore wind farms, dynamic time delay characteristics, historical wind power sequences of the target offshore wind farm, and meteorological variable sequences are fused to obtain the predictive input features; By inputting the predicted input features into the pre-trained prediction model, time-series feature modeling is performed to obtain the ultra-short-term wind power prediction results of the target offshore wind farm.

[0008] Furthermore, the historical wind power sequence of the target offshore wind farm is a sequence of wind power data collected in chronological order during the historical operation of the target offshore wind farm, used to characterize the changes in the power generation of the target offshore wind farm at different times. The historical wind power sequence of the adjacent offshore wind farm is a historical wind power data sequence collected at least one offshore wind farm in the surrounding area of ​​the target offshore wind farm at the same time scale, which is used to reflect the spatial correlation between different offshore wind farms. The meteorological variable sequence of the target offshore wind farm is a sequence of meteorological observation data related to the operating environment of the target offshore wind farm, including wind speed, wind direction, temperature, humidity and air pressure.

[0009] Furthermore, the preprocessing is Min-Max normalization, and the formula is: in, For the normalized data, This is the original data. and These are the minimum and maximum values ​​in the original data, respectively.

[0010] Furthermore, the adaptive mode decomposition of the preprocessed historical wind power sequence of the target offshore wind farm and the historical wind power sequence of neighboring offshore wind farms to obtain multiple mode components specifically includes: A differentiable multivariable variational mode decomposition module is constructed to perform mode decomposition on the preprocessed historical wind power sequences of the target offshore wind farm and the historical wind power sequences of neighboring offshore wind farms. Each input power sequence is represented as the sum of multiple modal components with different center frequencies, as follows: in, Indicates the first Each variable dimension at time... The input power sequence, Indicates the first kThe modal component in the ... Signal components in each variable dimension Indicates the number of modal components; Suppose the input power sequence contains multiple variable dimensions. The modal center frequencies corresponding to each variable dimension are used to construct a center frequency parameter matrix, which is represented as: in, Represents the center frequency parameter matrix. Indicates the first The modal component in the ... Central frequency in each variable dimension Indicates the dimension of the input variable; The bandwidth penalty parameters for each modal component are parameterized and reconstructed. The bandwidth penalty parameters are obtained through exponential mapping and are expressed as follows: in, Indicates the first The modal component in the ... Bandwidth penalty parameters for each variable dimension This represents the corresponding trainable parameters; Through the center frequency parameter matrix and bandwidth penalty parameters The historical wind power sequence of the target offshore wind farm and the historical wind power sequence of neighboring offshore wind farms are iteratively updated and decomposed to obtain multiple modal components corresponding to different time scales.

[0011] Furthermore, the process via the center frequency parameter matrix and bandwidth penalty parameters The historical wind power sequences of the target offshore wind farm and neighboring offshore wind farms are iteratively updated and decomposed to obtain multiple modal components corresponding to different time scales, specifically including: Based on the center frequency parameter matrix and bandwidth penalty parameters Construct a multivariate variational mode decomposition optimization model, with the objective function expressed as: in, Represents the Hilbert transform operator; The imaginary unit; Represents the time variable Differentiation operation; The optimization model is solved under the constraint of input power sequence reconstruction, which is expressed as follows: ; The optimization model is iteratively solved by the alternating direction multiplier method, and the modal components and center frequency parameters are updated until the convergence condition is met, so as to obtain multiple modal components corresponding to the target offshore wind power sequence and the historical wind power sequence of the adjacent offshore wind farm.

[0012] Furthermore, the step of calculating the correlation between the target offshore wind farm and its neighboring offshore wind farms based on modal components, and selecting the neighboring offshore wind farms with the highest correlation to the target offshore wind farm, specifically includes: A phase attention mechanism is constructed for each modal component to calculate the phase attention relationship between the target offshore wind farm and each neighboring offshore wind farm in the 1st modal period. k The association weights under each modality are expressed as follows: in, Indicates the first Target offshore wind farm under each mode With the A nearby offshore wind farm The correlation weight between them; Indicates the target offshore wind farm One modal component; Indicates the number of nearby offshore wind farms One modal component; This represents the attention function used to calculate the correlation between two modal components; Based on the target offshore wind farm and each adjacent offshore wind farm in the first k Association weights under each modality By maximizing the attention weight matrix, the nearest offshore wind farm index with the highest correlation is selected, as shown below: in, This represents the index of the nearest offshore wind farms that have the highest correlation with the target offshore wind farm; This indicates the total number of nearby offshore wind farms.

[0013] Furthermore, the step of extracting phase features from the modal components of the target offshore wind farm and the modal components of the selected neighboring offshore wind farms to obtain dynamic time delay features characterizing the propagation relationship between different offshore wind farms specifically includes: For the target offshore wind farm Modal components Compared with the selected neighboring offshore wind farms Modal components Perform a Hilbert transform to construct an analytic signal, which is represented as: in, Indicates the target offshore wind farm The analytical signal corresponding to each modal component; This indicates the selected neighboring offshore wind farms. The analytical signal corresponding to each modal component; Represents the Hilbert transform operator; It represents the imaginary unit.

[0014] Based on the analytical signal, the target offshore wind farm and neighboring offshore wind farms are calculated at the [number]th [time]. The instantaneous phase in each mode is represented as: in, Indicates the target offshore wind farm Each modal component at time... The instantaneous phase; Indicates the number of nearby offshore wind farms Each modal component at time... The instantaneous phase; A phase angle extraction operator for complex analytic signals, used to calculate the argument of a complex number; The phase difference between the target offshore wind farm and the adjacent offshore wind farm is calculated based on the instantaneous phase and expressed as follows: in, This indicates the distance between the target offshore wind farm and adjacent offshore wind farms on the [missing information]. Phase difference in each mode; The dynamic time delay characteristics between different offshore wind farms are calculated based on the phase difference and the center frequency of the corresponding modal components, and are expressed as follows: in, Indicates the first Dynamic time delay characteristics between the target offshore wind farm and neighboring offshore wind farms under each mode; Indicates the first The center frequency of each modal component.

[0015] Furthermore, the predicted input features are represented as follows: in, To predict input features; Historical wind power sequence of the target offshore wind farm; Indicates the target offshore wind farm One modal component; This indicates the first neighboring offshore wind farm selected through screening. One modal component; Indicates the first Dynamic time delay characteristics between the target offshore wind farm and neighboring offshore wind farms under each mode; Indicates the target wind farm number n A meteorological variable.

[0016] Furthermore, the prediction model is a deep time-series prediction model based on a selective state-space model. This prediction model consists of multiple stacked Mamba modules used to perform long-sequence dependency modeling on the predicted input features, specifically including: The predicted input features Construct as a time series input vector And input a selective state-space model for time series modeling, wherein the discretized state equation of the state-space model is expressed as: in, Indicates time The input vector; Indicates time The hidden state; Represents the state-space model at time 10:00. The output characteristics; This represents the discretized state transition matrix; This represents the discretized input mapping matrix; Indicates the output mapping matrix; Represents the continuous-time state transition matrix; Represents the continuous-time input matrix; This represents the discretization time step parameter; Indicates matrix exponentiation; Represents the identity matrix; In each Mamba module, the input vector The features are divided into main branch features and gated branch features through linear mapping, as follows: in, Indicates the characteristics of the main branch; Indicates the gating branch feature; Indicates the input linear mapping layer; The main branch features are subjected to a one-dimensional convolution operation and then non-linearly mapped using an activation function, as follows: in, Represents convolutional features; This represents a one-dimensional convolution operation; This represents the activation function of the SigmoidLinear Unit; The convolutional features are input into the selective state-space model for temporal modeling to obtain the state-space output features: in, This represents the computational process of the selective state-space model. Output features for the state space; The state space output features and gated branch features are fused element-wise, and the Mamba module output is obtained through linear mapping, as follows: in, Indicates the output characteristics of the Mamba module; This represents element-wise multiplication. Indicates the output linear mapping layer; A deep temporal prediction network is constructed by stacking multiple Mamba modules in a hierarchical manner. The prediction input features are then used for temporal modeling, and the wind power prediction results of the target offshore wind farm are obtained through the output layer.

[0017] Furthermore, the training process of the prediction model includes: The predicted input features are arranged in chronological order to construct a training sample set, which is then paired with the corresponding actual wind power sequence of the target offshore wind farm to form a supervised training data pair. The predicted input features are input into the prediction model for forward propagation calculation to obtain the predicted power result of the target offshore wind farm. A model training loss function is constructed based on the predicted power result and the actual wind power, where the loss function is cross-entropy loss. The gradient of the prediction model parameters is calculated using the backpropagation algorithm based on the loss function, and the prediction model parameters are iteratively updated according to the gradient. Training stops when the loss function meets the preset convergence condition or reaches the maximum number of training rounds, resulting in a trained prediction model. The trained prediction model is then used to predict the ultra-short-term wind power of the target offshore wind farm.

[0018] Compared with the prior art, the present invention has the following advantages: (1) Existing technologies, when processing information from multiple wind farms, mostly directly splice the sequences of different wind farms or meteorological variables into the model, which leads to the problem of multi-scale feature confusion and makes it difficult to achieve effective time scale alignment. This invention uses a differentiable multivariable variational mode decomposition module to perform mode decomposition on the historical power sequences of the target wind farm and neighboring wind farms, and iteratively updates the mode components based on the center frequency parameter matrix and bandwidth penalty parameter to achieve adaptive multi-scale alignment of cross-wind farm signals, thereby effectively distinguishing features of different time scales and improving the sensitivity and prediction accuracy of the prediction model to long-term and short-term changes.

[0019] (2) Existing technologies lack a dynamic screening mechanism when selecting wind farms or features for prediction, which can easily introduce irrelevant information and lead to a decrease in prediction robustness. This invention constructs a phase attention mechanism to calculate the correlation weights of the modal components of the target wind farm and neighboring wind farms, and dynamically selects the neighboring wind farms with the highest correlation to input into the prediction model, thereby achieving adaptive selection of information, reducing interference from irrelevant features, and improving the stability and reliability of prediction results.

[0020] (3) Existing technologies typically simplify the time delay of wind energy propagation to a static delay, which cannot accurately quantify the dynamic physical process of wind energy propagation from upstream wind farms to downstream wind farms, thus limiting the prediction accuracy. This invention performs Hilbert transform on the modal components of the target wind farm and selected neighboring wind farms, calculates the instantaneous phase and obtains the phase difference, and then accurately quantifies the dynamic time delay characteristics of wind energy propagation based on the modal center frequency, thereby achieving explicit modeling of the physical propagation process of wind energy. This improves the prediction model's ability to capture the dynamic dependence between wind farms and enhances the accuracy of ultra-short-term power prediction.

[0021] (4) Existing technologies often employ fixed structures or simple recurrent networks when constructing prediction models, making it difficult to simultaneously handle long sequence dependencies and nonlinear feature interactions. This invention improves the prediction input features by using a deep time-series prediction model composed of multiple stacked Mamba modules. Each Mamba module combines one-dimensional convolution, a selective state-space model, and gated branch features for feature fusion, thereby enhancing the ability to efficiently model long sequence dependencies and perform nonlinear mapping on the prediction input sequence, and further improving the accuracy and adaptability of wind power prediction.

[0022] (5) Existing technologies often only use raw power data or a single type of meteorological variable when constructing the input features of the prediction model, which has limited feature expression capabilities and makes it difficult to fully reflect the comprehensive influence of multiple factors on wind power changes. This invention constructs prediction input features containing multi-source information by fusing the modal components of the target offshore wind farm, the modal components of neighboring offshore wind farms, dynamic time delay features, historical wind power sequences, and meteorological variable sequences. This enables the model to simultaneously utilize power change features, spatial propagation features, and meteorological environment features, thereby enhancing the expressive power of the input data and improving the adaptability of the prediction model to complex operating environments. Attached Figure Description

[0023] Figure 1 This is a flowchart of the ultra-short-term offshore wind power prediction method according to an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] Example 1: This embodiment provides an ultra-short-term offshore wind power prediction method that considers information from multiple wind farms, such as... Figure 1 As shown, it includes the following steps: Step S1: Obtain the historical wind power sequence of the target offshore wind farm, the historical wind power sequence of neighboring offshore wind farms, and the meteorological variable sequence of the target offshore wind farm, and perform preprocessing; In the data preparation phase, the historical wind power sequence of the target offshore wind farm, the historical wind power sequence of neighboring offshore wind farms, and the meteorological variable sequence of the target offshore wind farm are first obtained, and these data are preprocessed.

[0026] The historical wind power sequence of the target offshore wind farm records the power data of the wind farm in chronological order during its past operation. This data can reflect the power variation pattern of the wind farm in different time periods and provide time series characteristics for prediction.

[0027] Historical wind power sequences of nearby offshore wind farms include power data collected at the same time scale from at least one offshore wind farm located around the target wind farm. These data can reveal the spatial correlation between different wind farms and improve the accuracy of predictions by capturing the propagation relationship of wind energy among multiple wind farms.

[0028] The meteorological variable series of the target offshore wind farm includes meteorological observation data related to the wind farm's operating environment, such as wind speed, wind direction, temperature, humidity, and air pressure. These variables have a direct impact on wind power output.

[0029] To improve model training efficiency and prediction stability, all input data are normalized to represent them within a uniform numerical range, avoiding interference from differences in feature units on model parameter learning. Simultaneously, preprocessing reduces the impact of noise and outliers, making input features more stable and thus improving the accuracy of subsequent time-series modeling and wind power prediction.

[0030] Step S2: Perform adaptive mode decomposition on the preprocessed historical wind power sequence of the target offshore wind farm and the historical wind power sequence of neighboring offshore wind farms to obtain multiple mode components, specifically including: Adaptive mode decomposition (EMD) is performed on the preprocessed historical wind power sequences of the target offshore wind farm and neighboring offshore wind farms to extract feature information at different time scales. A differentiable multivariable variational mode decomposition module is constructed to decompose each input power sequence, representing each sequence as the sum of multiple modal components with different center frequencies. That is, the power signal of each input variable at time t can be expressed as: in, Indicates the first Each variable dimension at time... The input power sequence, Indicates the first k The modal component in the th ... Signal components in each variable dimension Indicates the number of modal components; Suppose the input power sequence contains multiple variable dimensions. The modal center frequencies corresponding to each variable dimension are used to construct a center frequency parameter matrix, which is represented as: in, Represents the center frequency parameter matrix. Indicates the first The modal component in the th ... Central frequency in each variable dimension Indicates the dimension of the input variable; The bandwidth penalty parameters for each modal component are parameterized and reconstructed. The bandwidth penalty parameters are obtained through exponential mapping and are expressed as follows: in, Indicates the first The modal component in the th ... Bandwidth penalty parameters for each variable dimension This represents the corresponding trainable parameters; through the trainable parameters The frequency selection range of the modal components is automatically adjusted so that high-frequency and low-frequency signals can be adapted to the power fluctuation characteristics of different time scales, thereby improving the adaptability and accuracy of the decomposition.

[0031] Through the center frequency parameter matrix and bandwidth penalty parameters The historical wind power sequences of the target offshore wind farm and neighboring offshore wind farms are iteratively updated and decomposed to obtain multiple modal components corresponding to different time scales, specifically including: Based on the center frequency parameter matrix and bandwidth penalty parameters Construct a multivariate variational mode decomposition optimization model, with the objective function expressed as: in, Represents the Hilbert transform operator; The imaginary unit; Represents the time variable The optimization objective is to minimize the analytic energy gradient of the modal signal in the frequency domain. Specifically, the analytic signal is constructed through Hilbert transform, and its derivatives are weighted and summed to achieve an accurate characterization of instantaneous frequency and amplitude changes.

[0032] The optimization model is solved under the constraint of input power sequence reconstruction, and the constraint is expressed as follows: ; The optimization model is iteratively solved by the alternating direction multiplier method, and the modal components and center frequency parameters are updated until the convergence condition is met, so as to obtain multiple modal components corresponding to the target offshore wind power sequence and the historical wind power sequence of the adjacent offshore wind farm.

[0033] Step S3: Calculate the correlation between the target offshore wind farm and its neighboring offshore wind farms based on the modal components, and select the neighboring offshore wind farms with the highest correlation to the target offshore wind farm. Specifically, this includes: A phase attention mechanism is constructed for each modal component to calculate the phase attention relationship between the target offshore wind farm and each neighboring offshore wind farm in the 1st modal period. k The association weights under each modality are expressed as follows: in, Indicates the first Target offshore wind farm under each mode With the A nearby offshore wind farm The correlation weight between them; Indicates the target offshore wind farm One modal component; Indicates the number of nearby offshore wind farms One modal component; This represents the attention function used to calculate the correlation between two modal components; The correlation between two modal signals is measured by comparing their amplitude, phase, and instantaneous characteristics.

[0034] Based on the target offshore wind farm and each adjacent offshore wind farm, in the... k Association weights under each modality By maximizing the attention weight matrix, the nearest offshore wind farm index with the highest correlation is selected, as shown below: in, This represents the index of the nearest offshore wind farms that have the highest correlation with the target offshore wind farm; This represents the total number of nearby offshore wind farms. This allows for the dynamic selection of the most relevant wind farm inputs, avoiding the introduction of irrelevant or noisy signals. Through this processing, the model can adaptively capture the spatial correlation between different wind farms and adjust the contribution weights of nearby wind farms for different modes. This provides high-quality input for subsequent dynamic time-delay feature extraction and deep time-series modeling, achieving effective fusion of cross-wind farm information and significantly improving the accuracy of ultra-short-term power prediction, thus enhancing the model's robustness under complex weather conditions.

[0035] Step S4: Extract phase features from the modal components of the target offshore wind farm and the modal components of the selected neighboring offshore wind farms to obtain dynamic time delay features characterizing the propagation relationship between different offshore wind farms. Specifically, this includes: For the target offshore wind farm Modal components Compared with the selected neighboring offshore wind farms Modal components Perform a Hilbert transform to construct an analytic signal, which is represented as: in, Indicates the target offshore wind farm The analytical signal corresponding to each modal component; This indicates the selected neighboring offshore wind farms. The analytical signal corresponding to each modal component; Represents the Hilbert transform operator; It represents the imaginary unit; by analyzing the signal, the instantaneous amplitude and phase information of each modal component can be obtained, thereby mapping the original time-domain signal to the complex plane, which makes it easier to capture the relative phase relationship between wind farms.

[0036] Based on the analytical signal, the target offshore wind farm and the adjacent offshore wind farms are calculated at the [missing information]. The instantaneous phase in each mode is represented as: in, Indicates the target offshore wind farm Each modal component at time... The instantaneous phase; Indicates the number of nearby offshore wind farms Each modal component at time... The instantaneous phase; A phase angle extraction operator for complex analytic signals, used to calculate the argument of a complex number; The phase difference between the target offshore wind farm and neighboring offshore wind farms, calculated based on instantaneous phase, is expressed as follows: in, This indicates the distance between the target offshore wind farm and adjacent offshore wind farms on the [missing information]. Phase difference in each mode; The dynamic time delay characteristics between different offshore wind farms are calculated based on the phase difference and the center frequency of the corresponding modal components, and are expressed as follows: in, Indicates the first Dynamic time delay characteristics between the target offshore wind farm and neighboring offshore wind farms under each mode; Indicates the first The center frequencies of each modal component are used to capture the time-varying path and propagation speed of wind from the upstream wind farm to the target wind farm, dynamically reflecting the spatial transfer process of wind energy. This explicit dynamic time delay modeling avoids the limitations of the static delay assumption in traditional methods, enabling the prediction model to adaptively adjust the time alignment of input features for different meteorological conditions and wind field combinations, thereby significantly improving the accuracy and robustness of ultra-short-term wind power prediction.

[0037] Step S5: Perform feature fusion on the modal components of the target offshore wind farm, the modal components of neighboring offshore wind farms, dynamic time delay features, the historical wind power sequence of the target offshore wind farm, and the meteorological variable sequence to obtain the prediction input features; the prediction input features are expressed as: in, To predict input features; Historical wind power sequence of the target offshore wind farm; Indicates the target offshore wind farm One modal component; This indicates the first neighboring offshore wind farm selected through screening. One modal component; Indicates the first Dynamic time delay characteristics between the target offshore wind farm and neighboring offshore wind farms under each mode; Indicates the target wind farm number n Several meteorological variables were fused. This not only preserved the multi-scale temporal characteristics of each modal component but also explicitly encoded the time delay relationship of wind energy propagation from neighboring wind farms to the target wind farm. Furthermore, by combining meteorological variables, a unified representation of multi-source information was achieved. This multi-dimensional, dynamically aligned input feature enhances the prediction model's ability to capture wind power changes under complex wind field conditions, thereby improving the accuracy and robustness of ultra-short-term power prediction.

[0038] Step S6: Input the predicted input features into the pre-trained prediction model to perform time series feature modeling, and obtain the ultra-short-term wind power prediction results of the target offshore wind farm.

[0039] The prediction model is based on a selective state-space model and consists of multiple stacked Mamba modules. It can effectively capture long-term dependencies in the input feature sequence and dynamically process multimodal and multi-scale information.

[0040] Predict input features Construct as a time series input vector In the prediction model, time series vectors Input a selective state-space model for time series modeling; the discretized state equations are expressed as: in, Indicates time The input vector; Indicates time The hidden state; Represents the state-space model at time 10:00. The output characteristics; This represents the discretized state transition matrix; This represents the discretized input mapping matrix; Indicates the output mapping matrix; Represents the continuous-time state transition matrix; Represents the continuous-time input matrix; This represents the discretization time step parameter; Indicates matrix exponentiation; The identity matrix is ​​represented by the discretization of the continuous-time state matrix A and the input matrix B. This allows for both continuous dynamic characteristics and discrete-time series modeling, enabling accurate characterization of the dynamic changes in wind power.

[0041] In each Mamba module, the input vector The features are divided into main branch features and gated branch features through linear mapping, as follows: in, Indicates the characteristics of the main branch; Indicates the gating branch feature; Indicates the input linear mapping layer; Performing a one-dimensional convolution operation on the main branch features and then applying a non-linear mapping using an activation function is represented as follows: in, Represents convolutional features; This represents a one-dimensional convolution operation; This represents the activation function of the SigmoidLinear Unit; The convolutional features are input into a selective state-space model for time-series modeling to obtain the state-space output features: in, This represents the computational process of the selective state-space model. Output features for the state space; The state-space output features and gated branch features are fused element-wise, and the Mamba module output is obtained through linear mapping, represented as: in, Indicates the output characteristics of Mamba modules; This represents element-wise multiplication. This represents the output linear mapping layer; through a gating mechanism, the model can dynamically control the contribution ratio of each feature branch, effectively suppressing noise information and improving prediction robustness.

[0042] Multiple Mamba modules are stacked hierarchically to form a deep temporal prediction network, performing multi-level, multi-scale temporal modeling on the entire prediction input features. The network output layer maps the features at each time step, ultimately generating ultra-short-term wind power prediction results for the target offshore wind farm. This multi-layer stacking and state-space modeling approach fully utilizes historical power sequences, information from neighboring wind farms, dynamic time delay features, and meteorological variables to achieve high-precision wind power prediction under complex meteorological conditions, improving the efficiency of wind farm scheduling and operation management.

[0043] The training process of the prediction model includes: The predicted input features are arranged in chronological order to construct a training sample set, which is then paired with the corresponding actual wind power sequence of the target offshore wind farm to form a supervised training data pair. The predicted input features are then input into the prediction model for forward propagation calculation to obtain the predicted power result of the target offshore wind farm. A model training loss function is constructed based on the predicted power result and the actual wind power, and the loss function is the mean squared error loss function. The gradient of the prediction model parameters is calculated using the backpropagation algorithm based on the loss function, and the prediction model parameters are iteratively updated according to the gradient. Training stops when the loss function meets the preset convergence condition or reaches the maximum number of training rounds, thus obtaining the trained prediction model. The trained prediction model is then used to predict the ultra-short-term wind power of the target offshore wind farm.

[0044] This invention aims to construct an efficient and accurate ultra-short-term wind power prediction method. By deeply integrating differentiable multivariable variational mode decomposition and phase attention mechanism, it models the dynamic spatiotemporal dependence between wind farms, solves the problems of difficulty in aligning multi-scale signal features and inaccurate quantization of physical propagation delay, and thus achieves rapid and high-precision prediction of wind power.

[0045] The prediction method proposed in this invention can effectively achieve adaptive signal decomposition driven by prediction tasks, accurately quantify the physical time delay characteristics of wind energy propagation across wind farms, and thus optimize the fusion and extraction of multi-source features, improving the prediction model's ability to perceive non-stationary wind power sequences. This method not only significantly improves prediction accuracy and robustness under complex and variable meteorological conditions, but also achieves integrated joint optimization from signal decomposition to prediction output, providing strong support for intelligent scheduling, fluctuation mitigation, and safe and stable operation of high-proportion wind power grid-connected systems.

[0046] Example 2: The test data used in this embodiment comes from five offshore wind farm clusters within 50 kilometers of each other in Jiangsu Province, with installed capacities ranging from 100MW to 400MW. Historical power data for all wind farms were collected with a time resolution of 15 minutes, spanning from 00:00 on November 14, 2022 to 00:00 on November 13, 2023. The data for each wind farm contains 34,945 time steps.

[0047] To facilitate model training and evaluation, the entire dataset was divided into three parts chronologically: the first 70% (24,462 time points) served as the training set for model parameter learning; the middle 10% (3,495 time points) served as the validation set for early stopping detection and optimal model selection during training; and the final 20% (6,988 time points) served as the test set for objective performance evaluation of the final selected model. Mean absolute error and root mean square error were selected as the evaluation metrics for the prediction results.

[0048] The original multi-wind field power data were subjected to Min-Max normalization to eliminate dimensional differences. Then, the prediction model of this invention was constructed: the number of modes K in the DVMMD module was set to a fixed value, and the center frequency matrix was initialized. The hidden layer dimension of the Mamba predictor was set to 64, and the stacking layer was 4. The model was trained end-to-end using the Adam optimizer, with an initial learning rate of 0.0001 and mean squared error as the loss function.

[0049] Three ultra-short-term prediction tasks with different time spans were designed: predictions 15 minutes, 1 hour, and 4 hours in advance. The benchmark comparison models selected were Persistence, LSTM, Transformer, and the prediction model of this invention. Additionally, ablation experiments were added to verify the effectiveness of the PA and DMVMD modules. The comparison of MAE and RMSE indices between the benchmark models and the ablation experiments is shown in Tables 1, 2, and 3. Table 1. Comparison of RMSE indices for prediction results from different models Table 2 Comparison of MAE indices from different model predictions Table 3 Comparison Results of Ablation Experiments Simulation experiments demonstrate that the proposed method, which integrates differentiable modal decomposition and phase attention mechanisms, outperforms existing mainstream deep learning models in terms of power prediction accuracy, trend tracking capability, and robustness under different prediction lead times. Furthermore, this method exhibits significant anti-interference capabilities in long-sequence prediction tasks, effectively addressing the complex spatiotemporal dynamics of offshore wind farms, and providing a high-precision solution for ultra-short-term wind power prediction in practical engineering.

[0050] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for ultra-short-term offshore wind power prediction considering information from multiple wind farms, characterized in that, Includes the following steps: The historical wind power sequence of the target offshore wind farm, the historical wind power sequence of neighboring offshore wind farms, and the meteorological variable sequence of the target offshore wind farm are obtained and preprocessed. Adaptive mode decomposition is performed on the preprocessed historical wind power sequence of the target offshore wind farm and the historical wind power sequence of the adjacent offshore wind farm to obtain multiple mode components; The correlation between the target offshore wind farm and each neighboring offshore wind farm is calculated based on the modal components, and the neighboring offshore wind farm with the highest correlation with the target offshore wind farm is selected. Phase features are extracted from the modal components of the target offshore wind farm and the modal components of the selected neighboring offshore wind farms to obtain dynamic time delay features that characterize the propagation relationship between different offshore wind farms. The modal components of the target offshore wind farm, the modal components of neighboring offshore wind farms, dynamic time delay characteristics, historical wind power sequences of the target offshore wind farm, and meteorological variable sequences are fused to obtain the predictive input features; The predicted input features are input into the pre-trained prediction model to perform time series feature modeling, and the ultra-short-term wind power prediction results of the target offshore wind farm are obtained. The step of extracting phase features from the modal components of the target offshore wind farm and the modal components of the selected neighboring offshore wind farms to obtain dynamic time delay features characterizing the propagation relationship between different offshore wind farms specifically includes: For the target offshore wind farm Modal components Compared with the selected neighboring offshore wind farms Modal components Perform a Hilbert transform to construct an analytic signal, which is represented as: in, Indicates the target offshore wind farm The analytical signal corresponding to each modal component; This indicates the selected neighboring offshore wind farms. The analytical signal corresponding to each modal component; Represents the Hilbert transform operator; Represents the imaginary unit; Based on the analytical signal, the target offshore wind farm and neighboring offshore wind farms are calculated at the [number]th [time]. The instantaneous phase in each mode is represented as: in, Indicates the target offshore wind farm The analytic signal corresponding to each modal component at time 1 The instantaneous phase; Indicates the number of nearby offshore wind farms The analytic signal corresponding to each modal component at time 1 The instantaneous phase; A phase angle extraction operator for complex analytic signals, used to calculate the argument of a complex number; The phase difference between the target offshore wind farm and the adjacent offshore wind farm is calculated based on the instantaneous phase and expressed as follows: in, This indicates the distance between the target offshore wind farm and adjacent offshore wind farms on the [missing information]. Phase difference in each mode; The dynamic time delay characteristics between different offshore wind farms are calculated based on the phase difference and the center frequency of the corresponding modal components, and are expressed as follows: in, Indicates the first Dynamic time delay characteristics between the target offshore wind farm and neighboring offshore wind farms under each mode; Indicates the first The center frequency of each modal component.

2. The ultra-short-term offshore wind power prediction method considering multiple wind farm information as described in claim 1, characterized in that, The historical wind power sequence of the target offshore wind farm is a sequence of wind power data collected in chronological order during the historical operation of the target offshore wind farm, used to characterize the changes in the power generation of the target offshore wind farm at different times. The historical wind power sequence of the adjacent offshore wind farm is a historical wind power data sequence collected at least one offshore wind farm in the surrounding area of ​​the target offshore wind farm at the same time scale, which is used to reflect the spatial correlation between different offshore wind farms. The meteorological variable sequence of the target offshore wind farm is a sequence of meteorological observation data related to the operating environment of the target offshore wind farm, including wind speed, wind direction, temperature, humidity and air pressure.

3. The ultra-short-term offshore wind power prediction method considering multiple wind farm information according to claim 1, characterized in that, The preprocessing is Min-Max normalization, and the formula is: in, For the normalized data, This is the original data. and These are the minimum and maximum values ​​in the original data, respectively.

4. The ultra-short-term offshore wind power prediction method considering multiple wind farm information as described in claim 1, characterized in that, The adaptive mode decomposition of the preprocessed historical wind power sequences of the target offshore wind farm and the historical wind power sequences of neighboring offshore wind farms to obtain multiple mode components specifically includes: A differentiable multivariable variational mode decomposition module is constructed to perform mode decomposition on the preprocessed historical wind power sequences of the target offshore wind farm and the historical wind power sequences of neighboring offshore wind farms. Each input power sequence is represented as the sum of multiple modal components with different center frequencies, as follows: in, Indicates the first Each variable dimension at time... The input power sequence, Indicates the first k The modal component in the th ... Signal components in each variable dimension Indicates the number of modal components; Suppose the input power sequence contains multiple variable dimensions. The modal center frequencies corresponding to each variable dimension are used to construct a center frequency parameter matrix, which is represented as: in, Represents the center frequency parameter matrix. Indicates the first The modal component in the th ... Central frequency in each variable dimension Indicates the dimension of the input variable; The bandwidth penalty parameters for each modal component are parametrically reconstructed. The bandwidth penalty parameters are obtained through exponential mapping and are expressed as follows: in, Indicates the first The modal component in the th ... Bandwidth penalty parameters for each variable dimension This represents the corresponding trainable parameters; Through the center frequency parameter matrix and bandwidth penalty parameters The historical wind power sequence of the target offshore wind farm and the historical wind power sequence of neighboring offshore wind farms are iteratively updated and decomposed to obtain multiple modal components corresponding to different time scales.

5. The ultra-short-term offshore wind power prediction method considering multiple wind farm information according to claim 4, characterized in that, The through the center frequency parameter matrix and bandwidth penalty parameters The historical wind power sequences of the target offshore wind farm and neighboring offshore wind farms are iteratively updated and decomposed to obtain multiple modal components corresponding to different time scales, specifically including: Based on the center frequency parameter matrix and bandwidth penalty parameters Construct a multivariate variational mode decomposition optimization model, with the objective function expressed as: in, Represents the Hilbert transform operator; The imaginary unit; Represents the time variable Differentiation operation; The optimization model is solved under the constraint of input power sequence reconstruction, which is expressed as follows: ; The optimization model is iteratively solved by the alternating direction multiplier method, and the modal components and center frequency parameters are updated until the convergence condition is met, so as to obtain multiple modal components corresponding to the target offshore wind power sequence and the historical wind power sequence of the adjacent offshore wind farm.

6. The ultra-short-term offshore wind power prediction method considering multiple wind farm information according to claim 1, characterized in that, The step of calculating the correlation between the target offshore wind farm and its neighboring offshore wind farms based on modal components, and then selecting the neighboring offshore wind farms with the highest correlation to the target offshore wind farm, specifically includes: A phase attention mechanism is constructed for each modal component to calculate the phase attention relationship between the target offshore wind farm and each neighboring offshore wind farm in the 1st modal period. k The association weights under each modality are expressed as follows: in, Indicates the first Target offshore wind farm under each mode With the A nearby offshore wind farm The correlation weight between them; Indicates the target offshore wind farm One modal component; Indicates the number of nearby offshore wind farms One modal component; This represents the attention function used to calculate the correlation between two modal components; Based on the target offshore wind farm and each adjacent offshore wind farm, in the... k Association weights under each modality By maximizing the attention weight matrix, the nearest offshore wind farm index with the highest correlation is selected, as shown below: in, This represents the index of the nearest offshore wind farms that have the highest correlation with the target offshore wind farm; This indicates the total number of nearby offshore wind farms.

7. The ultra-short-term offshore wind power prediction method considering multiple wind farm information according to claim 1, characterized in that, The predicted input features are represented as follows: in, To predict input features; Historical wind power sequence of the target offshore wind farm; Indicates the target offshore wind farm One modal component; This represents the first neighboring offshore wind farm selected through screening. One modal component; Indicates the first Dynamic time delay characteristics between the target offshore wind farm and neighboring offshore wind farms under each mode; Indicates the target wind farm number n A meteorological variable.

8. The ultra-short-term offshore wind power prediction method considering multiple wind farm information according to claim 1, characterized in that, The prediction model is a deep time-series prediction model based on a selective state-space model. The prediction model consists of multiple stacked Mamba modules, used for long-sequence dependency modeling of the predicted input features, specifically including: The predicted input features Construct as a time series input vector And input a selective state-space model for time series modeling, wherein the discretized state equation of the state-space model is expressed as: in, Indicates time The input vector; Indicates time The hidden state; Represents the state-space model at time 10:

00. The output characteristics; This represents the discretized state transition matrix; This represents the discretized input mapping matrix; Indicates the output mapping matrix; Represents the continuous-time state transition matrix; Represents the continuous-time input matrix; This represents the discretization time step parameter; Indicates matrix exponentiation; Represents the identity matrix; In each Mamba module, the input vector The features are divided into main branch features and gated branch features by linear mapping, as follows: in, Indicates the characteristics of the main branch; Indicates the gating branch feature; Indicates the input linear mapping layer; The main branch features are subjected to a one-dimensional convolution operation and then non-linearly mapped using an activation function, as follows: in, Represents convolutional features; This represents a one-dimensional convolution operation; This represents the activation function of the SigmoidLinear Unit; The convolutional features are input into the selective state-space model for temporal modeling to obtain the state-space output features: in, This represents the computational process of the selective state-space model. Output features for the state space; The state space output features and gated branch features are fused element-wise, and the Mamba module output is obtained through linear mapping, as follows: in, Indicates the output characteristics of the Mamba module; This represents element-wise multiplication. Indicates the output linear mapping layer; A deep temporal prediction network is constructed by stacking multiple Mamba modules in a hierarchical manner. The prediction input features are then used for temporal modeling, and the wind power prediction results of the target offshore wind farm are obtained through the output layer.

9. A method for ultra-short-term offshore wind power prediction considering multiple wind farm information as described in claim 8, characterized in that, The training process of the prediction model includes: A training sample set is constructed by arranging the predicted input features in chronological order, and a supervised training data pair is formed with the corresponding actual wind power sequence of the target offshore wind farm. The predicted input features are input into the prediction model for forward propagation calculation to obtain the predicted power result of the target offshore wind farm. A model training loss function is constructed based on the predicted power result and the actual wind power, and the loss function is the mean squared error loss function. The gradient of the prediction model parameters is calculated through the backpropagation algorithm based on the loss function, and the prediction model parameters are iteratively updated according to the gradient. Training stops when the loss function meets the preset convergence condition or reaches the maximum number of training rounds, and the trained prediction model is obtained. The trained prediction model is then used to predict the ultra-short-term wind power of the target offshore wind farm.

Images