A wind power prediction method based on multi-scale space-time coupling

Abstract

The application discloses a wind power prediction method based on multi-scale space-time coupling, and relates to the technical field of wind energy prediction. The method first collects historical wind speed data of a target turbine and its nearest neighbors, and obtains enhanced features through multi-scale feature extraction and preprocessing; a unique embedding vector is generated for each turbine, which is spliced with the enhanced features to form joint features; the joint features are input into an LSTM encoder to obtain an encoding state, and a spatial context vector is generated through a spatial attention mechanism; enhanced space-time features are obtained through a gating fusion unit, and a multi-layer perceptron is input with a learnable time domain vector of a prediction step to directly regress future 1-12h wind power at one time. The application realizes end-to-end, multi-step and turbine-specific prediction, effectively captures multi-scale space-time coupling rules, reduces prediction error, improves robustness and generalization ability, meets real-time demand of power grid dispatching, and helps large-scale and efficient utilization of wind energy.

Description

Technical Field

[0001] This invention relates to the field of wind energy prediction technology, and more specifically, to a wind power prediction method based on multi-scale spatiotemporal coupling. Background Technology

[0002] In the field of wind energy forecasting, with the advancement of global "dual carbon" goals, wind energy, as a clean and renewable energy source, has become a key pillar of the global power structure transformation. It is projected that by 2030, global installed wind power capacity will exceed 1.2 TW. However, the stochastic and non-stationary nature of wind energy, as well as the time-varying characteristics of the weather-to-power mapping, pose significant challenges to accurate wind energy forecasting. Existing technologies mainly include physical methods, statistical methods, and deep learning methods, but these methods have significant limitations in dealing with multi-scale dynamic coupling (from minutes to hours). The following are the main problems in the development of existing technologies: 1. Physical methods rely on numerical weather prediction (NWP), which has high computational costs and insufficient resolution: Physical methods are based on fluid dynamics equations and provide clear physical interpretations, but they have huge computational overhead, limited spatiotemporal resolution, and cannot capture turbine-level turbulence and wake effects.

[0003] 2. Statistical methods (such as ARIMA) have weak ability to handle nonlinear spatiotemporal coupling: Statistical models are computationally efficient, but they cannot effectively represent strong nonlinear relationships and multivariate spatiotemporal interactions. When faced with the complex dynamics of wind energy data, statistical methods often ignore multi-scale energy cascades, leading to a decrease in prediction accuracy, especially in multi-step prediction scenarios where accumulated lag errors are significant.

[0004] 3. Deep learning models perform poorly in multi-step prediction and lack adaptive capabilities: Although existing deep learning models (such as LSTM and Transformer) perform well in single-step prediction, they are prone to cumulative lag and attention pattern rigidity in multi-step prediction.

[0005] 4. Ignoring turbine-specific behavior and insufficient end-to-end optimization: Existing technologies often employ a two-stage pipeline of "wind speed first, power curve second," which introduces fitting errors and hinders end-to-end optimization. Furthermore, the models fail to capture individual turbine differences (such as geographical location, altitude, local terrain, and historical service records), resulting in unspecific predictions.

[0006] 5. Insufficient real-time performance and robustness: Traditional methods have shortcomings in real-time monitoring and response, especially in the poor generalization ability of cross-wind farm assessment.

[0007] Therefore, there is an urgent need for a wind power prediction method based on multi-scale spatiotemporal coupling. Through innovative module design and architecture integration, this method can achieve end-to-end accurate prediction from historical wind speed data to wind power output, effectively overcoming the limitations of traditional methods in nonlinear processing, multi-step prediction error accumulation, and insufficient generalization ability. This will help wind power play a more stable supporting role in the energy structure transformation and promote the efficient and large-scale utilization of renewable energy under the "dual carbon" target. Summary of the Invention

[0008] This invention provides a wind power prediction method based on multi-scale spatiotemporal coupling. Its purpose is to address the problems existing in the current wind power prediction technology, such as insufficient multi-scale dynamic coupling capture, weak nonlinear spatiotemporal interaction modeling capability, accumulation of multi-step prediction errors, lack of turbine-specific adaptation, and insufficient real-time performance and generalization ability. Through innovative designs such as multi-scale feature extraction and spatiotemporal coupling modeling, it achieves high-precision and robust prediction of wind power.

[0009] Specifically, without relying on high-cost numerical weather forecasting, the model adaptively mines short-, medium-, and long-term temporal patterns of wind energy data, dynamically integrates the spatial correlation effects of neighboring turbines, and accurately encodes the individual differences of each turbine. This effectively reduces the accumulation of errors in multi-step predictions, improves the model's generalization ability and real-time response speed in cross-wind farm scenarios, provides reliable power prediction data for grid dispatch and energy planning, and promotes the large-scale and efficient utilization of wind energy as a clean and renewable energy source, thus contributing to the achievement of global dual-carbon goals.

[0010] To achieve the above objectives, this invention provides a wind power prediction method based on multi-scale spatiotemporal coupling, specifically including the following steps: S1. Collect historical wind speed data of the target turbine and its nearest neighbors, and perform feature extraction and preprocessing on the historical wind speed data to obtain enhanced features; S2. Each turbine learns to generate a unique embedding vector, and the embedding vector is concatenated with the enhanced temporal features to form a joint feature; S3. Input the joint features into the LSTM encoder to obtain the hidden state sequence and the final encoded state; and use the intermediate state of the decoder as the query, dynamically weight the hidden states of neighboring wind turbines through the spatial attention mechanism to generate a spatial context vector. S4. The intermediate state of the decoder and the spatial context vector are adaptively coupled through a gated fusion unit to obtain enhanced spatiotemporal features; S5. The prediction step size is encoded into a learnable time-domain vector, and the learnable time-domain vector and the enhanced spatiotemporal features are input into a multilayer perceptron to directly regress the wind power for the next 1–12 hours in one go, thereby achieving power prediction.

[0011] Preferably, in S1, the specific content of feature extraction and preprocessing includes: S11, using three parallel convolutional kernels to extract short, medium and long-term multi-scale temporal features respectively; S12. The three multi-scale temporal features are spliced ​​together using a gating network to obtain spliced ​​multi-scale features; S13. Calculate the gating weights and dynamically fuse multi-scale temporal features to generate enhanced features.

[0012] Preferably, the expression for the multi-scale temporal features is: ; in, , , These represent short-, medium-, and long-term time characteristics, For a one-dimensional convolutional layer with kernel size n, The activation function for the rectified linear unit is... Historical wind speed data for the target turbine and its nearest neighbor; The splicing multi-scale features The expression is: ; The gating weight The formula for calculation is: ; in, It is the Sigmoid activation function. , These are the weight matrices for the first and second level linear transformations, respectively. , These are the bias terms for the first and second level linear transformations, respectively; The enhanced features The expression is: ; in, This is a dynamic fusion symbol.

[0013] Preferably, the method for generating the embedding vector includes: Assign a unique identifier index i∈{1,2,...,N} to each turbine i, where N is the total number of turbines in the wind farm; and use a learnable embedding matrix. Map the turbine identifier to a low-dimensional dense vector, where... The dimension of the embedded vector; The expression for the embedding vector is: in, Let be the embedding vector of turbine i. This is the unique thermal encoding vector for the i-th turbine; The embedding vector The implicitly encoded individual differences of turbines, such as geographical location, altitude, local terrain, wake effect, and historical operation and maintenance characteristics, are adaptively learned through an end-to-end training process, without the need to explicitly provide physical parameters.

[0014] Preferably, the hidden state sequence in S3 and final encoding status The expression is: ; in, For LSTM encoder modules, This is the set of learnable parameters for the encoder; The attention weight The formula for calculation is: ; in, Let T be the hidden state of the decoder at the current moment, and let T be the transpose operator. and These represent the query weight matrix and the key weight matrix, respectively. Let d be the hidden state vector of the nth neighboring wind turbine, d be the feature dimension, exp(·) represent the result of the natural exponent operation, and j be the index of the decoding step, i.e. the jth prediction time step.

[0015] The spatial context vector The formula for calculation is: ; in, k For the number of neighboring turbines, Let be the value vector of the nth neighbor turbine.

[0016] Preferably, the adaptive coupling process of the gated fusion unit in S4 uses the following formula: ; ; in, The gating coefficient, This is the weight matrix. For bias terms, c For spatial context vectors, To enhance spatiotemporal characteristics.

[0017] Preferably, the power prediction process in S5 is based on a step-size-aware nonlinear mapping mechanism, specifically including: S51. Step-size encoding: The prediction time domain τ∈{1,2,...,12} is mapped to a time-domain encoded vector through a learnable embedding matrix. The expression for the process is: in, (τ) Implicit coding of meteorological evolution patterns, power curve slope changes and error propagation characteristics corresponding to different prediction step sizes; S52, Feature Fusion: Enhancing Spatiotemporal Features With time-domain coding vector (τ) is concatenated to form step-size sensing features. ; S53. Nonlinear Regression: A nonlinear mapping from step-size sensing features to power prediction values ​​is achieved through a multilayer perceptron (MLP). S54. Generation of multi-step prediction sequences: For each time domain τ∈{1,2,...,12} in the future 1–12h, repeat the above steps to calculate the corresponding power prediction values ​​to form a complete prediction sequence.

[0018] Preferably, step size sensing features The expression is: The formula for calculating the predicted power value in S53 is: in, and (L=1,...,L) represents the weight matrix and bias vector of the Lth layer, where L is the number of MLP layers; For target turbine i Power prediction at time t+τ; In S54, the expression for the complete predicted sequence is: .

[0019] Specifically: Multi-scale feature extraction and preprocessing: Historical wind speed data of the target turbine and its nearest neighbors are collected, and enhanced features are obtained by feature extraction and preprocessing of the historical wind speed data. This step implements multi-band filtering through adaptive temporal convolutional blocks (ATC), and uses three parallel convolutional kernels (kernel sizes of 3, 5, and 7) to extract short-term (1-3 hours), medium-term (4-8 hours), and long-term (9-12 hours) multi-scale temporal features respectively. Then, dynamic fusion is performed through a gating network to effectively capture the multi-scale dynamic coupling law of wind speed data and solve the information loss problem caused by single-scale feature extraction. Turbine-specific embedding: A unique embedding vector is generated for each turbine, and this embedding vector is concatenated with the enhanced temporal features to form a joint feature. This step maps turbine identifiers to low-dimensional dense vectors through a learnable embedding matrix, implicitly encoding individual differences in turbine geographical location, altitude, local terrain, wake effect, and historical operation and maintenance characteristics. This enables turbine-specific modeling, overcomes the limitations of traditional methods that rely on fixed power curves, and improves the model's adaptability to different turbines. Spatiotemporal coding and spatial attention mechanism: The joint features are input into the LSTM encoder to obtain the hidden state sequence and the final encoded state; and using the intermediate state of the decoder as the query, the hidden states of neighboring wind turbines are dynamically weighted through a spatial attention mechanism to generate a spatial context vector. This step adopts a spatial attention mechanism similar to beamforming in signal processing, adaptively adjusting the influence weights of neighboring turbines according to the current prediction task, effectively fusing spatial correlation information, capturing the propagation law of wind speed between turbine arrays and wake effects, and improving the model's ability to model spatial dynamic coupling. Gated spatiotemporal coupling: The intermediate state of the decoder and the spatial context vector are adaptively coupled through a gated fusion unit to obtain enhanced spatiotemporal features. This step dynamically balances the contribution ratio of time series information and spatial context information through a gating mechanism, achieving deep coupling in both time and space dimensions. The feature fusion strategy is adaptively adjusted according to different prediction scenarios to enhance the robustness and generalization ability of the model. Step-size-aware power regression: The prediction step size is encoded as a learnable temporal vector, which, together with the enhanced spatiotemporal features, is input into a multilayer perceptron to directly regress wind power for the next 1–12 hours in a single step, achieving end-to-end, multi-step, turbine-specific power prediction. This step dynamically modulates the power mapping function through temporal encoding, enabling the model to adaptively adjust attention weight allocation and feature extraction strategies according to the prediction step size. This effectively solves the problem of error accumulation in multi-step prediction using traditional fixed mapping functions, significantly improving the accuracy and stability of long-term prediction.

[0020] This invention uses time-domain coding (τ) The dynamic modulation power mapping function enables the model to adaptively adjust the attention weight allocation and feature extraction strategy according to the prediction step size, effectively solving the problem of error accumulation in multi-step prediction by the traditional fixed mapping function, and realizing accurate prediction with step size awareness.

[0021] Therefore, the wind power prediction method based on multi-scale spatiotemporal coupling described above, as used in this invention, has the following advantages compared with existing technologies: (1) This invention introduces an adaptive temporal convolution block to achieve multi-band filtering, directly extracting short, medium and long-term time patterns, and dynamically recombining features through gating fusion. This improvement significantly enhances the model's ability to capture the multi-scale dynamics of wind energy data, solves the problems of high computational overhead and insufficient resolution in existing technologies, and reduces the prediction error by 2.9%, especially in 1-hour prediction where the MAE drops to 0.123; (2) This invention dynamically adjusts the influence weights between turbines through a spatial attention module and a gating mechanism, and directly modulates the slope of the power mapping function. This improvement enables the model to adaptively optimize the information path according to different prediction time domains, thereby enhancing the robustness of multi-step prediction. Experiments show that in 12-hour prediction, the RMSE is reduced to 0.244, which is better than the baseline model; (3) This invention learns a low-dimensional representation for each turbine by embedding turbine identity vectors, thus encoding its aerodynamic fingerprint. This improvement enables end-to-end, turbine-specific predictions without explicit power curve fitting, solving the problem of poor generalization ability in existing technologies. In cross-wind farm assessment, the model's MAE is reduced by 1.1%, demonstrating its strong adaptability.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating an embodiment of a wind power prediction method based on multi-scale spatiotemporal coupling according to the present invention. Figure 2 This is a comparison curve of the RMSE performance of different models in an embodiment of the wind power prediction method based on multi-scale spatiotemporal coupling of the present invention. Detailed Implementation

[0024] Figure 1 A flowchart illustrating the implementation of this invention is provided. To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] Example 1 This embodiment comprehensively verifies the performance of the wind power prediction method based on multi-scale spatiotemporal coupled network (MsSTCN) under different prediction time domains and different turbine distribution scenarios using real-world scenario datasets. The core objectives include: 1. Verify the model's prediction accuracy for wind power in short (1-3 hours), medium (4-8 hours), and long (9-12 hours) time domains, and compare its advantages with existing mainstream models; 2. Evaluate the model's generalization ability under different turbine densities and terrain conditions; 3. The contribution of each core module (Adaptive Temporal Convolutional Block (ATC) and Spatial Attention Enhancement Sequence (AES)) was quantified through ablation experiments; 4. Test the real-time computational efficiency of the model and verify its feasibility in meeting the real-time dispatch requirements of the power grid.

[0026] To fully verify the performance of this invention, two real-world wind energy datasets were selected for experiments: the US onshore wind farm dataset (200 turbines, one year of hourly data) and the WIND dataset from the National Renewable Energy Laboratory (NREL) (10×10 turbine array in the Wyoming area). The experiments employed time-segmentation: the US dataset was used for training over the first two months, validation over the following month, and testing over the next nine months; the WIND dataset was used for training over the first eight months and testing over the following four months.

[0027] The model parameters are set as follows: the encoder uses LSTM (hidden layer dimension 128), the decoder uses stride-aware MLP (number of layers 3, activation function ReLU), the ATC convolution kernel size is (3,5,7), the turbine embedding dimension de=32, and the optimizer uses Adam (learning rate 0.001, batch size 64).

[0028] To fully verify the performance advantages of MsSTCN, 11 mainstream wind power prediction models were selected as benchmarks. All benchmark models were trained on the same dataset and preprocessing process, and the parameters were optimized to the optimal state through the validation set to ensure the fairness of the comparison.

[0029] Four commonly used evaluation metrics in the field of wind power prediction are employed to comprehensively measure the model's prediction accuracy, error distribution, and stability: 1. Mean Absolute Error (MAE): Reflects the average deviation between the predicted value and the actual value; the smaller the value, the higher the accuracy. 2. Normalized Absolute Error (NAE): Eliminates the influence of power magnitude, facilitating comparisons across datasets; 3. Root Mean Square Error (RMSE): Sensitive to large errors, reflecting the degree of dispersion of prediction errors; 4. Coefficient of determination : Measures the model’s ability to explain the variation in real data. The value ranges from [0,1]. The closer it is to 1, the better the model fits.

[0030] The calculated data were organized, and Table 1 shows the normalized absolute error (NAE) for wind energy prediction of the US wind farm in the 1-12 hour time domain: Short-term prediction (1-3 hours): MSSTCN, RNN, and LSTM all achieved a 1-hour NAE of 0.123, ranking among the best; however, in the 2-3 hour prediction, MSSTCN (2h: 0.152, 3h: 0.171) significantly outperformed other models, demonstrating the ability of multi-scale temporal feature extraction to capture short-term dynamics.

[0031] Mid-term forecast (4-8 hours): MSSTCN's NAE consistently maintains the lowest proof that the spatial attention mechanism effectively integrates neighbor turbine information, improving the stability of mid-term forecasts.

[0032] Long-term prediction (9-12 hours): MSSTCN has an NAE of 0.244 in 12-hour prediction, which is 0.41% lower than the best baseline model DSTAN (0.245) and 10.62% lower than LSTM (0.273), thus alleviating the problem of long-term prediction error accumulation of traditional deep learning models.

[0033] Average performance: MSSTCN has an average NAE of 0.210, which is an improvement over all benchmark models. It is 1.41% lower than DSTAN and 9.09% lower than LSTM, fully validating the model's comprehensive advantages.

[0034] Table 1 Normalized Absolute Error (NAE) of Wind Energy Forecasting for US Wind Farms

[0035] Table 2 shows a comparison of ablation experiments using the US dataset, regarding the Adaptive Temporal Convolutional Block (ATC): After adding ATC alone, the 1-hour MAE decreased from 0.1259 to 0.1220, a reduction of 3.09%, proving that short, medium and long-term multi-scale feature extraction can effectively capture the dynamic changes in wind speed. The 3-hour MAE decreased by 3.13%, which reflects the role of the gating fusion mechanism in optimizing and recombining multi-scale features and reducing the limitations of single-scale features.

[0036] Regarding Spatial Attention Enhancement Sequences (AES): The advantage is obvious in forecasts longer than 5 hours, with the 8-hour MAE decreasing by 0.95% compared to the baseline and the 12-hour MAE decreasing by 0.16%. The spatial attention module can dynamically adjust the influence weight of neighboring turbines, which effectively mitigates the error accumulation caused by wind speed propagation delay, especially in long-term forecasting.

[0037] Regarding the synergistic effect of modules: The average MAE of the complete model (ATC+AES) decreased to 0.1929, a reduction of 1.83% compared to the baseline, and the test loss decreased by 1.66%. The synergistic effect of multi-scale temporal features and dynamic spatial attention enables accurate modeling in both time and space dimensions, which is the core reason for the improved model performance.

[0038] Table 2 Comparison of ablation experiments (US dataset)

[0039] To more intuitively observe the performance comparison of each model, the RMSE is plotted as a curve, such as... Figure 2 As shown in Table 1, on the US wind farm dataset, this invention (MsSTCN) outperforms the baseline model in 1-12 hour predictions. The 1-hour MAE is 0.123, and the 12-hour MAE is 0.244, representing a 0.4% improvement over the best baseline model (DSTAN). The robustness of multi-step prediction is significant, especially in long-term predictions where error accumulation is reduced.

[0040] Therefore, this embodiment verifies the advantages of the wind power prediction method based on multi-scale spatiotemporal coupled network (MsSTCN) in 1-12 hour prediction by comparing it with 11 benchmark models and conducting ablation experiments.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A wind power prediction method based on multi-scale spatiotemporal coupling, characterized in that, Includes the following steps: S1. Collect historical wind speed data of the target turbine and its nearest neighbors, and perform feature extraction and preprocessing on the historical wind speed data to obtain enhanced features; S2. Each turbine learns to generate a unique embedding vector, and the embedding vector is concatenated with the enhanced temporal features to form a joint feature; S3. Input the joint features into the LSTM encoder to obtain the hidden state sequence and the final encoded state; and use the intermediate state of the decoder as the query, dynamically weight the hidden states of neighboring wind turbines through the spatial attention mechanism to generate a spatial context vector. S4. The intermediate state of the decoder and the spatial context vector are adaptively coupled through a gated fusion unit to obtain enhanced spatiotemporal features; S5. The prediction step size is encoded into a learnable time-domain vector, and the learnable time-domain vector and the enhanced spatiotemporal features are input into a multilayer perceptron to directly regress the wind power for the next 1–12 hours in one go, thereby achieving power prediction.

2. The wind power prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, In S1, the specific content of feature extraction and preprocessing includes: S11, using three parallel convolutional kernels to extract short, medium and long-term multi-scale time features respectively; S12. The three multi-scale temporal features are spliced ​​together using a gating network to obtain spliced ​​multi-scale features; S13. Calculate the gating weights and dynamically fuse multi-scale temporal features to generate enhanced features.

3. The wind power prediction method based on multi-scale spatiotemporal coupling according to claim 2, characterized in that, The expression for the multi-scale temporal feature is: ； in, , , These represent short-, medium-, and long-term time characteristics, For a one-dimensional convolutional layer with kernel size n, The activation function for the rectified linear unit is... Historical wind speed data for the target turbine and its nearest neighbor; The splicing multi-scale features The expression is: ； The gating weight The formula for calculation is: ； in, It is the Sigmoid activation function. , These are the weight matrices for the first and second level linear transformations, respectively. , These are the bias terms for the first and second level linear transformations, respectively; The enhanced features The expression is: ； in, This is a dynamic fusion symbol.

4. The wind power prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, The method for generating the embedding vector includes: Assign a unique identifier index i∈{1,2,...,N} to each turbine i, where N is the total number of turbines in the wind farm; and use a learnable embedding matrix. Map the turbine identifier to a low-dimensional dense vector, where... The dimension of the embedded vector; The expression for the embedding vector is: ； in, Let be the embedding vector of turbine i. This is the unique thermal encoding vector for the i-th turbine; The embedding vector The implicitly encoded individual differences of turbines, such as geographical location, altitude, local terrain, wake effect, and historical operation and maintenance characteristics, are adaptively learned through an end-to-end training process, without the need to explicitly provide physical parameters.

5. The wind power prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, The hidden state sequence described in S3 and final encoding status The expression is: ； in, For LSTM encoder modules, This is the set of learnable parameters for the encoder; The attention weight The formula for calculation is: ； in, Let T be the hidden state of the decoder at the current moment, and let T be the transpose operator. and These represent the query weight matrix and the key weight matrix, respectively. Let d be the hidden state vector of the nth neighboring wind turbine, d be the feature dimension, exp(·) represent the result of the natural exponent operation, and j be the index of the decoding step, i.e. the jth prediction time step. The spatial context vector The formula for calculation is: ； in, k For the number of neighboring turbines, Let be the value vector of the nth neighbor turbine.

6. The wind power prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, The adaptive coupling process of the gated fusion unit described in S4 uses the following formulas: ； ； in, The gating coefficient, This is the weight matrix. For bias terms, c For spatial context vectors, To enhance spatiotemporal characteristics.

7. The wind power prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, The power prediction process described in S5 is based on a step-size-aware nonlinear mapping mechanism, specifically including: S51. Step-size encoding: The prediction time domain τ∈{1,2,...,12} is mapped to a time-domain encoded vector through a learnable embedding matrix. The expression for the process is: in, (τ) Implicit coding of meteorological evolution patterns, power curve slope changes and error propagation characteristics corresponding to different prediction step sizes; S52, Feature Fusion: Enhancing Spatiotemporal Features With time-domain coding vector (τ) is spliced ​​together to form step length sensing features; ； S53. Nonlinear Regression: A nonlinear mapping from step-size sensing features to power prediction values ​​is achieved through a multilayer perceptron (MLP). S54. Generation of multi-step prediction sequences: For each time domain τ∈{1,2,...,12} in the future 1–12h, repeat the above steps to calculate the corresponding power prediction values ​​to form a complete prediction sequence.

8. The wind power prediction method based on multi-scale spatiotemporal coupling according to claim 7, characterized in that, The formula for calculating the predicted power value in S53 is: ； in, and (L=1,...,L) represents the weight matrix and bias vector of the Lth layer, where L is the number of MLP layers; For target turbine i Power prediction at time t+τ; In S54, the expression for the complete predicted sequence is: .

Images