Wind power output prediction method based on large meteorological model and multi-head self-attention mechanism

Abstract

The invention provides a wind power output prediction method based on a large meteorological model and a multi-head self-attention mechanism, and the method comprises the steps: S1, obtaining weather forecast initial field data as the input of the large meteorological model, and generating coarse-resolution weather forecast data based on the large meteorological model; s2, downscaling diffusion is carried out on the coarse-resolution weather forecast data to generate high-resolution weather forecast data; s3, introducing different conditions into the high-resolution weather forecast data, and generating multiple groups of high-resolution ensemble forecast data simulating different forecast scenes; and S4, adopting a multi-head attention mechanism to adaptively aggregate the multiple groups of high-resolution ensemble forecast data, and generating a wind power output prediction result. According to the method, the robustness and generalization ability of the prediction model are enhanced, the adaptability and the calculation efficiency are improved, and a high-quality and high-precision wind power prediction result with higher comprehensiveness is obtained.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy electric power, and in particular relates to a wind power output prediction method based on a large meteorological model and a multi-head self-attention mechanism. Background Art

[0002] As a clean and renewable energy source, wind power plays an important role in the transformation of the global energy structure. However, wind power output has strong randomness and uncertainty, and accurate wind power forecasting is crucial for grid dispatching and stable operation. Traditional wind power output forecasting methods mainly rely on high-precision numerical weather forecasts (NWP). These numerical weather forecast models use complex physical equations to simulate atmospheric movement and predict future weather conditions. However, although numerical weather forecasting methods have achieved high accuracy, due to their high computational cost and often require a long computing cycle, they are difficult to meet the timeliness and high resolution requirements of wind power forecasting.

[0003] In recent years, the application of deep learning in the field of weather forecasting has made significant progress. Meteorological large models based on deep learning, such as deep meteorological models and generative adversarial networks (GANs), can combine a large amount of historical meteorological data and satellite observation data to achieve accurate weather forecasts in an end-to-end manner. Compared with traditional numerical weather forecast models, these meteorological large models based on deep learning not only have significant improvements in computing efficiency, but can also provide more accurate predictions in small-scale areas, which is expected to improve the overall level of wind power forecasting.

[0004] At present, efficient large-scale meteorological model technologies are all trained based on large-scale ERA5 reanalysis meteorological data, and their spatial grid resolution is 0.25°*0.25°. Although this is relatively detailed in global meteorological models, it is still insufficient for local wind farms, especially for local meteorological changes in areas with complex terrain (such as mountainous and coastal areas). The actual meteorological changes in wind farms often require smaller-scale prediction capabilities to provide more detailed wind speed and wind direction change information in order to improve the accuracy of power forecasts. Summary of the invention

[0005] The present invention proposes a wind power output forecasting method based on a large meteorological model and a multi-head self-attention mechanism, which enhances the robustness and generalization ability of the forecasting model, improves the adaptability and computational efficiency, and obtains more comprehensive, high-quality and high-precision wind power forecasting results.

[0006] To achieve the above object, the technical solution of the present invention is achieved as follows: A wind power output prediction method based on a large meteorological model and a multi-head self-attention mechanism, comprising: S1. Obtaining initial weather forecast field data as input to the large weather model, and generating coarse-resolution weather forecast data based on the large weather model; S2, downscaling and diffusing the coarse-resolution weather forecast data to generate high-resolution weather forecast data; S3, introducing different conditions into the high-resolution meteorological forecast data to generate multiple groups of high-resolution ensemble forecast data simulating different forecast scenarios; S4. Adopting a multi-head attention mechanism, adaptively aggregating the multiple groups of high-resolution ensemble forecast data to generate a wind power output forecast result.

[0007] Furthermore, step S2 specifically includes: S201, gradually adding noise to the coarse-resolution weather forecast data, so that the noise-added data is converted into a noise distribution; S202, using the UNet structure as a denoiser to perform reverse denoising on the noisy data; in the reverse denoising process, generating noise-free high-resolution weather forecast data according to the current noise state and conditional input.

[0008] Furthermore, step S3 specifically includes: Different random disturbance sequences are introduced into the high-resolution meteorological forecast data to simulate uncertainty and different forecast paths, generate different ensemble members, and obtain multiple groups of high-resolution ensemble forecast data simulating different forecast scenarios.

[0009] Furthermore, step S4 specifically includes: S401, spatially interpolating the multiple groups of high-resolution ensemble forecast data to wind farm sites, and then performing temporal interpolation to obtain corresponding meteorological forecast data; standardizing meteorological characteristics and wind farm observation data in the meteorological forecast data; S402, the standardized meteorological characteristics and wind farm observation data constitute the input feature vector of the wind power output prediction model, the input feature dimension of the input feature vector is evenly divided into multiple attention heads, and the attention output of each attention head is spliced ​​together to form a new matrix; S403: converting the new matrix into wind power prediction through the fully connected layer of the wind power output prediction model to obtain the final wind power output prediction result.

[0010] On the other hand, the present invention also proposes a wind power output prediction device based on a large meteorological model and a multi-head self-attention mechanism, comprising: Acquisition module: Acquisition of initial weather forecast data as input to the large weather model, and generation of coarse-resolution weather forecast data based on the large weather model; Downscaling module: downscaling and diffusing the coarse-resolution meteorological forecast data to generate high-resolution meteorological forecast data; Ensemble forecast module: introducing different conditions into the high-resolution meteorological forecast data to generate multiple sets of high-resolution ensemble forecast data simulating different forecast scenarios; Multi-head attention prediction module: adopts the multi-head attention mechanism to adaptively aggregate the multiple groups of high-resolution ensemble forecast data to generate wind power output prediction results.

[0011] Furthermore, the downscaling module comprises: Noise adding unit: gradually adding noise to the coarse resolution weather forecast data, so that the noise-added data is converted into noise distribution; Reverse denoising unit: Use the UNet structure as a denoiser to reversely denoise the noisy data; in the reverse denoising process, generate noise-free high-resolution weather forecast data based on the current noise state and conditional input.

[0012] Furthermore, the ensemble forecast module includes: Different random disturbance sequences are introduced into the high-resolution meteorological forecast data to simulate uncertainty and different forecast paths, generate different ensemble members, and obtain multiple groups of high-resolution ensemble forecast data simulating different forecast scenarios.

[0013] Furthermore, the multi-head attention prediction module includes: Interpolation unit: spatially interpolating the multiple sets of high-resolution ensemble forecast data to the wind farm site, and then performing temporal interpolation to obtain corresponding meteorological forecast data; standardizing the meteorological characteristics and wind farm observation data in the meteorological forecast data; Multi-head unit: The standardized meteorological features and wind farm observation data constitute the input feature vector of the wind power output prediction model. The input feature dimension of the input feature vector is evenly divided into multiple attention heads, and the attention output of each attention head is spliced ​​together to form a new matrix; Prediction unit: converts the new matrix into wind power prediction through the fully connected layer of the wind power output prediction model to obtain the final wind power output prediction result.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention realizes high-precision wind power prediction: The present invention is based on a large meteorological model, and uses a downscaling diffusion model to refine coarse-resolution meteorological data into high-resolution data, thereby more accurately reflecting the meteorological characteristics of the area where the wind farm is located. On this basis, a conditional generation model is used to generate a variety of ensemble forecasts to cover the changes in wind power output under different meteorological scenarios, enhancing the model's adaptability to uncertainty and complex meteorological conditions. Finally, the multi-head attention mechanism is used to accurately fuse the information of the ensemble members, making the final wind power prediction results more accurate and stable, thereby significantly improving the applicability and prediction performance of the large meteorological model in the wind power prediction scenario.

[0015] 2. The present invention improves adaptability and computing efficiency: The present invention reduces the reliance on high computing resources in traditional numerical weather forecasting by introducing a large meteorological model and a deep learning downscaling model, which not only significantly improves the speed of large-scale meteorological data processing, but also enhances the applicability of the model. The method has a wide range of adaptability and can be applied to wind farms in different geographical environments such as coastal areas, mountains, and plains, and maintains the stability of the prediction effect under various complex terrain and climatic conditions, providing solid technical support for the comprehensive promotion and reliable operation of wind farms. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the process of Example 1 of the present invention. DETAILED DESCRIPTION

[0017] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0018] In order to make the purpose and features of the present invention more obvious and understandable, the specific implementation methods of the present invention are further described below in conjunction with the accompanying drawings.

[0019] The present invention proposes a wind power output forecasting method based on a large meteorological model and a multi-head attention mechanism. The main design idea is: first, the ECMWF deterministic forecast initial field is inferred using a large meteorological model to obtain coarse-resolution meteorological forecast data. Then, the coarse-resolution meteorological data is further refined based on the downscaling diffusion model of the U-Net architecture. By introducing high-resolution static data and physical constraints, the model improves the temporal and spatial resolution of meteorological data. The encoder-decoder structure of U-Net is particularly suitable for processing downscaling tasks, so that the model can capture local details of meteorological data while retaining large-scale information in the original data. In addition, by introducing the smooth diffusion characteristics of the diffusion model, the downscaling results can be kept smooth and physically reasonable in space. Then, a conditional generation model is used to generate a variety of ensemble forecasts. Through the diverse ensemble members, the changes in wind power output under different meteorological scenarios can be fully reflected, which enhances the robustness and generalization ability of the prediction model. Finally, the multi-head attention mechanism is used to effectively integrate the prediction information of multiple set members. By capturing the similarities and differences between the members, it can better filter and weight the most valuable information to obtain more comprehensive and high-quality wind power prediction results.

[0020] Based on the above design concept, this embodiment Figure 1 As shown, it mainly includes: Step 1: Obtain initial weather forecast field data as input to the large weather model, and generate coarse-resolution weather forecast data based on the large weather model; Download the initial weather forecast field X0 of ECMWF (European Centre for Medium-Range Weather Forecasts) as the meteorological model M LM The input of the meteorological model M is used to obtain the coarse-resolution weather forecast data for the next 24 to 360 hours through inference, with a spatial resolution of 0.25° and a temporal resolution of 6 hours. LM It can be a large meteorological model that has been implemented in the existing technology, such as NVIDIA FourCastNet, Huawei Pangu Meteorological Model, Fengwu Meteorological Model, etc.

[0021] ;in It is the coarse-resolution weather forecast data.

[0022] If you focus on the day-ahead output forecast of a wind farm, you need to infer at least to the next day, that is, a 48-hour forecast.

[0023] Step 2: downscaling and diffusing the coarse-resolution weather forecast data to generate high-resolution weather forecast data; A downscaling diffusion model based on the U-Net framework is used to convert coarse-resolution data into high-resolution data by gradually adding and removing noise, thereby improving the spatial resolution of meteorological forecast data from 0.25° to 0.1° and the temporal resolution from 6 hours to 1 hour.

[0024] Specifically, the downscaling diffusion model first gradually adds noise to the input data to convert it into a noise distribution, and then in the reverse process, through the conditional input and U-Net denoiser, generates denoised high-resolution weather forecast data at each time step. In this way, the downscaling diffusion model uses conditional diffusion and denoising processes to effectively downscale the coarse-resolution data to the target resolution, providing support for high-precision weather forecasts.

[0025] The details of this step include: 2.1 Explanation of input and output data of downscaling diffusion model: Downscaling diffusion model inputs include: Coarse resolution weather forecast data: , with a spatial resolution of 0.25° and a temporal resolution of 3 hours; High-resolution static data: terrain height (H), geographic location (L), etc., with a resolution of 0.1°; Target high-resolution data: ECMWF deterministic forecast data for the same period are obtained as the real target high-resolution data, with a spatial resolution of 0.1° and a temporal resolution of 3 hours.

[0026] Downscaled diffusion model outputs include: High-resolution meteorological data: , with a spatial resolution of 0.1° and a temporal resolution of 1 hour.

[0027] For example, the relevant elements of the downscaling diffusion model are as follows:

[0028] 2.2 Downscaling diffusion model training and reasoning: (1) Forward diffusion process: high-resolution data of the target Gradually add noise to generate a series of noisy data (t represents the number of noise addition steps, and the noise level increases with t) until the data approaches pure Gaussian noise. These noisy data are used to train the model to learn how to recover high-resolution data from data with different noise levels.

[0029] ; in, is the data at the noise step t, is the noise level, which increases gradually with t, and N represents Gaussian distribution. If the final step of noise addition is L, the data of the final step L The mean is close to 0 and the variance is Pure Gaussian noise.

[0030] (2) Reverse diffusion process: training the UNet denoiser , from each noisy sample data Restore the original high-resolution data UNet receives coarse resolution data As a conditional input, the model gradually generates high-resolution data that meets the constraints of the coarse-resolution data during the denoising process.

[0031] The denoising process is the reverse process of the denoising process, which generates 0.1° high-resolution meteorological data by gradually removing the noise. The process is described by the inverse stochastic differential equation (SDE): ; in, is the score function of the diffusion process, estimating the noise level of the data at each step, It is a standard Brownian motion.

[0032] In practice, the scoring function Using UNet Denoiser To approximate. UNet is based on the current data The conditioned input generates denoised 0.1° high-resolution data.

[0033] UNet denoiser: Use the UNet structure as a denoiser to generate noise-free, high-resolution outputs based on the current noise state and conditional input in each reverse denoising step. Use the standard UNet architecture, including encoders, decoders and skip connections.

[0034] UNet Denoiser It is a conditional generative network whose input is noisy meteorological data and coarse-resolution conditional input. Its goal is to generate denoised high-resolution data. The UNet denoiser uses the classifier free bootstrapping technique to enhance the generation effect.

[0035] The UNet denoiser calculates the denoised output using the following formula: ; in: is the guiding strength, is the conditional input, i.e., coarse resolution meteorological data , Represents unconditional input, which is used to control the diversity of generated samples.

[0036] 2.3 Denoising loss function: The loss function during training is based on denoising score matching, and its loss function is: ; in: is the loss weight, which is used to balance the noise level at different time steps, is the output of the UNet denoiser, and E represents the expected value, which is the weighted average of possible values ​​under a certain probability distribution. Indicates that during the training process, the data That is, each iteration randomly selects a training sample and a time step to calculate the loss value of a single sample in the loss function, and then obtains the average value of the overall loss through multiple iterations.

[0037] During the training process, the goal of the downscale diffusion model is to make the generated high-resolution data as close as possible to the real 0.1° high-resolution target data by minimizing the denoising loss function. After the training is completed, the downscale diffusion model can gradually denoise the coarse-resolution data to generate high-resolution output without directly inputting the real high-resolution target data.

[0038] Step 3: introducing different conditions into the high-resolution meteorological forecast data to generate multiple sets of high-resolution ensemble forecast data simulating different forecast scenarios; Generating high-resolution meteorological data based on downscaling diffusion model Finally, uncertainties in initial conditions or physical processes are further introduced to generate diverse forecast paths and obtain multiple sets of high-resolution ensemble forecast data, thereby capturing the uncertainties in meteorological forecasts more comprehensively.

[0039] 3.1 Introduce perturbation to generate set members: When generating different ensemble members, the initial high-resolution meteorological data Introducing different random perturbation sequences Through the conditional diffusion model, each iteration Different perturbations are applied to simulate uncertainties and different forecast paths.

[0040] The specific formula is as follows: ; in, It is Members of the collection, is the perturbation sequence used to generate this member; Indicates the application of disturbance.

[0041] 3.2 Generate N set members: By repeating the above perturbation process, N set members are generated. Each set member introduces different perturbation conditions. , which can simulate different forecast scenarios and form a diverse ensemble forecast. This approach ensures diversity among members, while each member is generated by high-resolution meteorological data, maintaining spatial details and physical consistency.

[0042] Step 4: Adopt a multi-head attention mechanism to adaptively aggregate the multiple groups of high-resolution ensemble forecast data to generate wind power output forecast results.

[0043] 4.1 Data processing: (1) Ensemble forecast interpolation processing: All ensemble forecast members are spatially interpolated to the location of the wind farm site and temporally interpolated to 15 minutes to obtain the corresponding meteorological forecast data.

[0044] (2) All meteorological characteristics of weather forecast data and wind farm observation data are subjected to Z-score standardization: ; Where X is the data to be standardized, For the data after standardization, represents the mean of the data, Represents the standard deviation of the data.

[0045] 4.2 Multi-head attention mechanism: The standardized meteorological characteristics and wind farm observation data constitute the input feature vector of the wind power output prediction model, and a multi-head attention mechanism is introduced.

[0046] (1) Position encoding: The sinusoidal encoding method is used to add position information for each time step. The position encoding formula is as follows: Add the position encoding PE to the normalized input data to obtain the new input data.

[0047] (2) Attention calculation: The input feature dimension d is divided into h attention heads, and the feature dimension of each head is ,therefore . Each head calculates the attention output separately: ; Among them, Attention represents the attention mechanism, Q, K, and V represent query, key, and value, respectively.

[0048] (3) Multi-head attention concatenation: concatenate the outputs of each attention head to form a new matrix.

[0049] ; Among them, MultiHead represents multi-head attention splicing, and the attention output of the j-th attention head is ; ; , , and is a learnable weight matrix used to map the input features into the feature space of each head.

[0050] 4.3 Wind power forecast: The concatenated new matrix is ​​converted into wind power prediction through the fully connected layer of the wind power output prediction model. Through the multi-head attention mechanism, key information channels are dynamically selected, and the new matrix obtained after splicing is predicted by the wind power output prediction model, that is, it is converted into wind power prediction through a fully connected layer to obtain the final wind power output prediction result: ; in, is the output layer mapping function, which converts the features into wind power. For the final predicted wind power output.

[0051] The method described in this embodiment is characterized in that: Downscaling diffusion model: By extending the multi-layer architecture and conditional modeling of U-Net, the resolution of the large meteorological model was improved from 0.25° to 0.1°. In particular, the combination of physical constraints in the inverse diffusion stage improved the precision and physical consistency of the forecast.

[0052] Conditionally generated high-quality ensemble forecasts: Based on different initial disturbances and meteorological conditions, a set of diverse ensemble forecasts are generated through the conditional diffusion generation model, which improves the ability of wind power forecasts to cope with meteorological uncertainties.

[0053] Attention mechanism fuses ensemble forecast information: Processing high-resolution ensemble forecasts through a multi-head attention mechanism can effectively extract and fuse key information, playing a key role in improving the accuracy of the final prediction results.

[0054] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A wind power output prediction method based on a large meteorological model and a multi-head self-attention mechanism, characterized in that: include: S1. Obtaining initial weather forecast field data as input to the large weather model, and generating coarse-resolution weather forecast data based on the large weather model; S2, downscaling and diffusing the coarse-resolution weather forecast data to generate high-resolution weather forecast data; S3, introducing different conditions into the high-resolution meteorological forecast data to generate multiple sets of high-resolution ensemble forecast data simulating different forecast scenarios; S4. Adopting a multi-head attention mechanism, adaptively aggregating the multiple groups of high-resolution ensemble forecast data to generate a wind power output forecast result.

2. The method for predicting wind power output based on a large meteorological model and a multi-head self-attention mechanism according to claim 1 is characterized in that: Step S2 specifically includes: S201, gradually adding noise to the coarse-resolution weather forecast data, so that the noise-added data is converted into a noise distribution; S202, using the UNet structure as a denoiser to perform reverse denoising on the noisy data; in the reverse denoising process, generating noise-free high-resolution weather forecast data according to the current noise state and conditional input.

3. The method for predicting wind power output based on a large meteorological model and a multi-head self-attention mechanism according to claim 1 is characterized in that: Step S3 specifically includes: Different random disturbance sequences are introduced into the high-resolution meteorological forecast data to simulate uncertainty and different forecast paths, generate different ensemble members, and obtain multiple groups of high-resolution ensemble forecast data simulating different forecast scenarios.

4. The method for predicting wind power output based on a large meteorological model and a multi-head self-attention mechanism according to claim 1 is characterized in that: Step S4 specifically includes: S401, spatially interpolating the multiple groups of high-resolution ensemble forecast data to wind farm sites, and then performing temporal interpolation to obtain corresponding meteorological forecast data; standardizing meteorological characteristics and wind farm observation data in the meteorological forecast data; S402, the standardized meteorological characteristics and wind farm observation data constitute the input feature vector of the wind power output prediction model, the input feature dimension of the input feature vector is evenly divided into multiple attention heads, and the attention output of each attention head is spliced ​​together to form a new matrix; S403: converting the new matrix into wind power prediction through the fully connected layer of the wind power output prediction model to obtain the final wind power output prediction result.

5. A wind power output prediction device based on a large meteorological model and a multi-head self-attention mechanism, characterized in that: include: Acquisition module: Acquisition of initial weather forecast data as input to the large weather model, and generation of coarse-resolution weather forecast data based on the large weather model; Downscaling module: downscaling and diffusing the coarse-resolution meteorological forecast data to generate high-resolution meteorological forecast data; Ensemble forecast module: introducing different conditions into the high-resolution meteorological forecast data to generate multiple sets of high-resolution ensemble forecast data simulating different forecast scenarios; Multi-head attention prediction module: adopts the multi-head attention mechanism to adaptively aggregate the multiple groups of high-resolution ensemble forecast data to generate wind power output prediction results.

6. The wind power output prediction device based on a large meteorological model and a multi-head self-attention mechanism according to claim 5 is characterized in that: The downscaling module includes: Noise adding unit: gradually adding noise to the coarse resolution weather forecast data, so that the noise-added data is converted into noise distribution; Reverse denoising unit: Use the UNet structure as a denoiser to reversely denoise the noisy data; during the reverse denoising process, noise-free high-resolution weather forecast data is generated based on the current noise state and conditional input.

7. The wind power output prediction device based on a large meteorological model and a multi-head self-attention mechanism according to claim 5, characterized in that: The ensemble forecast module comprises: Different random disturbance sequences are introduced into the high-resolution meteorological forecast data to simulate uncertainty and different forecast paths, generate different ensemble members, and obtain multiple groups of high-resolution ensemble forecast data simulating different forecast scenarios.

8. The wind power output prediction device based on a large meteorological model and a multi-head self-attention mechanism according to claim 5, characterized in that: The multi-head attention prediction module includes: Interpolation unit: spatially interpolating the multiple sets of high-resolution ensemble forecast data to the wind farm site, and then performing temporal interpolation to obtain corresponding meteorological forecast data; standardizing the meteorological characteristics and wind farm observation data in the meteorological forecast data; Multi-head unit: The standardized meteorological features and wind farm observation data constitute the input feature vector of the wind power output prediction model. The input feature dimension of the input feature vector is evenly divided into multiple attention heads, and the attention output of each attention head is spliced ​​together to form a new matrix; Prediction unit: converts the new matrix into wind power prediction through the fully connected layer of the wind power output prediction model to obtain the final wind power output prediction result.

Images