Weather forecast data resolution improving method for wind power prediction
By combining a collaborative training framework with multiple models, the resolution mismatch problem in wind power forecasting was solved, enabling the conversion from low-resolution meteorological data to high-resolution wind power forecasting data, thereby improving the accuracy of the forecast data and the forecasting capability at fine-grained time scales.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU INST OF ENERGY CONVERSION CHINESE ACAD OF SCI
- Filing Date
- 2026-01-14
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies cannot effectively solve the problem of input-output resolution mismatch in wind power forecasting, especially how to generate high-frequency wind power forecasting data based on low-frequency numerical weather forecast data. Existing interpolation methods suffer from error propagation and target inconsistency.
A collaborative training framework is adopted, including an upsampling model, a site meteorological capture module, a feature fusion model, and a denoising diffusion model. Through end-to-end training, low-resolution meteorological forecast data is transformed into high-resolution wind power prediction data. Noise generation and feature extraction are performed using site measured data and reanalysis data, and model parameters are optimized to achieve high-quality data conversion.
It has achieved a reliable conversion from low-resolution meteorological forecast data to high-resolution wind power prediction data, improving the accuracy of prediction data and the ability to learn the true evolution law and uncertainty distribution on fine-grained time scales.
Smart Images
Figure CN121996925A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological forecast data resolution enhancement technology, and in particular to a method for improving the resolution of meteorological forecast data for wind power prediction. Background Technology
[0002] Short- to medium-term wind power forecasts (24 to 240 hours ahead) are crucial for grid dispatch and trading. To meet the precision requirements of the dispatch system, wind farms need to report power forecast curves with a time resolution of 15 minutes. The core input data for these forecasting tasks relies on the output of global numerical weather prediction models (such as ECMWF and GFS). However, the time resolution of these models is significantly lower than this requirement. For example, ECMWF forecasts have a resolution of 1 hour for the first 90 hours from the start of the forecast period, further decreasing to 3 or 6 hours for subsequent forecast periods.
[0003] Therefore, in the technical process of wind power forecasting, there is an inherent problem of input and output resolution mismatch. Specifically, it is about how to generate high-frequency wind power forecasting data with a time resolution of up to 15 minutes based on low-frequency numerical weather forecast data with a time resolution of only 1 hour or more.
[0004] Currently, the mainstream approach to improving the resolution of weather forecast data still revolves around interpolation techniques. However, these methods have significant shortcomings in practical applications. The following section explains two commonly used interpolation methods:
[0005] The first type of method directly performs mathematical interpolation (such as spline interpolation) on the original low-resolution forecast data without introducing any measured data for correction. While this approach is simple, it cannot correct for the systematic biases inherent in numerical forecasts, thus spreading errors throughout the entire generated sequence. Furthermore, the interpolation methods assume that meteorological elements change uniformly over time, making it difficult to accurately reflect rapid changes in the real atmosphere, such as sharp increases or decreases in wind speed, leading to a lack of forecast details for key weather phenomena.
[0006] The second type of method involves introducing measured data for correction based on interpolation. This typically involves two steps: first, interpolating the low-resolution data, and then building a machine learning model to correct the interpolation results. This two-stage approach has significant drawbacks. First, the interpolation error in the first stage directly affects the input of the model in the second stage, causing error propagation. Second, the goals of the two stages are inconsistent: the first-stage interpolation aims for curve smoothness, while the second-stage correction model aims for a close fit to the measured values. The two are difficult to optimize collaboratively, limiting the improvement of the final result.
[0007] In summary, existing technologies treat "data upscaling" and "data correction" as two separate processes, lacking an end-to-end joint learning mechanism from raw low-resolution data to high-resolution results. Therefore, they cannot completely solve the problem of high-quality input data required for wind power forecasting. Summary of the Invention
[0008] To address the aforementioned issues, this invention proposes a method for improving the resolution of meteorological forecast data for wind power prediction, enabling reliable conversion between low temporal resolution meteorological forecast data and high temporal resolution wind power prediction data.
[0009] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:
[0010] A method for improving the resolution of meteorological forecast data for wind power prediction includes the following steps:
[0011] Acquire meteorological forecast data, station-measured meteorological data, and reanalysis meteorological data, wherein the meteorological variables in the reanalysis meteorological data are consistent with those in the meteorological forecast data;
[0012] A collaborative training framework is constructed, comprising an upsampling model, a station meteorological capture module, a feature fusion model, and a denoising diffusion model. The meteorological forecast data, the station measured meteorological data, and the reanalysis meteorological data are input into the collaborative training framework. Based on the training requirements of the denoising diffusion model, real noise is generated, and predicted noise is obtained by the denoising diffusion model through the calculation of the training input data. At the same time, the station meteorological capture module generates station simulated features for the training stage. The difference between the predicted noise and the real noise, as well as the difference between the station simulated features for the training stage and the station measured meteorological data, are calculated using a loss function. The optimization objective is to minimize the two differences. The parameters of the upsampling model, the station meteorological capture module, the feature fusion model, and the denoising diffusion model are updated through backpropagation to complete the collaborative training.
[0013] The weather forecast data to be processed is input into the trained upsampling model to obtain initial enhanced weather data. After the initial enhanced weather data is processed by the station weather capture module, the station simulation features of the inference stage are obtained. After the initial enhanced weather data and the station simulation features of the inference stage are processed by the feature fusion model, the enhanced weather data of the inference stage is obtained. The initial noise is iteratively denoised based on the trained denoising diffusion model. The target weather forecast data is generated by combining the enhanced weather data of the inference stage.
[0014] In some implementations, the time resolution of the weather forecast data is not less than 1 hour, and the time resolution of the station-measured weather data and the reanalysis weather data is 15 minutes.
[0015] In some implementations, before the collaborative training is performed, the reanalysis meteorological data is adjusted to match the temporal resolution of the station's measured meteorological data by interpolation, using the time resolution of the station's measured meteorological data as a benchmark.
[0016] In some implementations, the loss function is a weighted sum of the mean square error between the predicted noise and the actual noise, and the mean square error between the simulated features of the training stations and the measured meteorological data of the stations. The weighted sum includes weight coefficients, which are freely set according to the actual application scenario.
[0017] In some implementations, the generation process of the real noise and the predicted noise during the collaborative training includes:
[0018] The measured meteorological data from the stations and the reanalysis meteorological data are input into the feature fusion model to obtain the training baseline ground value data.
[0019] Randomly sample time steps and the real noise, and add the real noise to the training benchmark ground truth data based on the diffusion formula to obtain noisy data. The diffusion formula includes a noise scheduling coefficient, which monotonically increases from the initial value to the final value. The scheduling method of the noise scheduling coefficient is linear scheduling or quadratic scheduling.
[0020] The meteorological forecast data is input into the upsampling model to obtain initial enhanced meteorological data for training. The initial enhanced meteorological data is processed by the station meteorological capture module to obtain the station simulation features of the training stage. The initial enhanced meteorological data and the station simulation features of the training stage are processed by the feature fusion model to obtain enhanced meteorological data of the training stage.
[0021] The noise-added data, the time step, and the enhanced meteorological data during the training phase are input into the denoising diffusion model to obtain the predicted noise.
[0022] In some implementations, the iterative denoising process includes:
[0023] Set the iteration step size and generate initial noise from a standard Gaussian distribution;
[0024] For each iteration time step, the current noise data, the current time step, and the enhanced meteorological data of the inference stage are input into the trained denoising diffusion model to obtain the predicted noise of the inference stage;
[0025] Based on the prediction noise calculation in the inference phase, the noise data of the previous iteration time step is calculated until all iteration steps are completed, and the target weather forecast data is obtained.
[0026] In some implementations, the upsampling model is a time-series data processing network model, the station meteorological capture module is a convolutional neural network model, the feature fusion model is a cross-attention network model or a linear superposition model, and the denoising diffusion model is a time-series data prediction model based on a neural network.
[0027] The beneficial effects of this invention are as follows: Unlike traditional solutions, this invention does not directly establish a deterministic mapping relationship from low-resolution meteorological data to high-resolution data. Instead, it adopts a collaborative training framework to simultaneously train a parameterized upsampling model, a station meteorological capture module, a feature fusion model, and a denoising diffusion model. Through an end-to-end training mechanism, the upsampling model learns to generate enhanced meteorological data during the training phase, which serves as intermediate features suitable for subsequent module processing. The station meteorological capture module and the feature fusion model jointly provide precise guiding conditions for the denoising diffusion model, enabling the denoising diffusion model to learn the true evolution law and uncertainty distribution of meteorological forecast data at fine-grained time scales. Ultimately, this achieves the task of transforming low-resolution meteorological forecast data into high-resolution target meteorological forecast data. Attached Figure Description
[0028] Figure 1 This is a flowchart illustrating the method for improving the resolution of meteorological forecast data for wind power prediction disclosed in an embodiment of the present invention.
[0029] Figure 2 This is a schematic diagram of the training process disclosed in an embodiment of the present invention;
[0030] Figure 3 This is a schematic diagram of the reasoning process disclosed in an embodiment of the present invention. Detailed Implementation
[0031] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the content of this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to this invention are shown in the accompanying drawings, not all of them.
[0032] This embodiment proposes a method for improving the resolution of meteorological forecast data for wind power prediction, such as... Figure 1 As shown, it includes the following steps:
[0033] Step 1: Obtain meteorological forecast data, station-measured meteorological data, and reanalysis meteorological data. The meteorological variables in the reanalysis meteorological data are consistent with those in the meteorological forecast data.
[0034] Step 1 of this plan is the data preparation process, which involves only three types of input data: weather forecast data, station-measured weather data, and reanalysis weather data. Specifically:
[0035] Weather forecast data (Low-resolution weather forecast data), its data length is The number of forecast variables is n_features.
[0036] Meteorological data measured at the station (High-resolution real data from the site), data length is The time resolution is 15 minutes. The number of actual meteorological variables measured is n_m, which includes surface wind speed, wind speed at the meteorological tower, air pressure, temperature, humidity, etc.
[0037] This scheme reanalyzes meteorological data using ERA5 reanalysis data. ERA5 reanalysis data is the ECMWF fifth-generation global reanalysis dataset, which integrates global observational data and model simulations to generate consistent historical climate records. Its data length is [data length missing]. The number of forecast variables is n_features. The original ERA5 reanalysis data has a time resolution of 1 hour. Therefore, before co-training, the reanalysis meteorological data is based on station-measured meteorological data. Based on the time resolution, the time resolution of the reanalysis meteorological data is adjusted by interpolation to make it consistent with the time resolution of the station's measured meteorological data. That is, the resolution of the reanalysis meteorological data is upsampled to 15 minutes by interpolation.
[0038] It should be noted that the high resolution and low resolution mentioned in this scheme are relative concepts, referring to meteorological forecast data. Taking ECMWF forecast data as an example, its time resolution is 1 hour for the first 90 hours, gradually decreasing thereafter. Therefore, input data with a resolution lower than 1 hour is considered low-resolution data. However, the target weather forecast data required by this invention needs to have a resolution of 15 minutes, which is used in conjunction with the actual meteorological data measured at the stations. The data also needs to have the same resolution accuracy, so input data with a resolution of 15 minutes is considered high-resolution data.
[0039] In actual implementation, and The value needs to be determined based on the time resolution and the number of days the data covers. For example, if the data length is 1 day, The value is 96. If the resolution is 1 hour, then The value is 24, if If the resolution is 3 hours, then The value is 8.
[0040] Step 2, collaborative training, such as Figure 2 As shown, a collaborative training framework is constructed, comprising an upsampling model, a station meteorological capture module, a feature fusion model, and a denoising diffusion model. Meteorological forecast data, station-measured meteorological data, and reanalysis meteorological data are input into the collaborative training framework. Based on the training requirements of the denoising diffusion model, realistic noise is generated, and predicted noise is obtained through the calculation of the training input data by the denoising diffusion model. Simultaneously, station-simulated features for the training phase are generated through the station meteorological capture module. The difference between predicted noise and realistic noise, as well as the difference between station-simulated features and station-measured meteorological data for the training phase, are calculated using a loss function. Minimizing these two differences is the optimization objective. Backpropagation updates the parameters of the upsampling model, station meteorological capture module, feature fusion model, and denoising diffusion model, completing the collaborative training.
[0041] Step 2 of this scheme is the collaborative training process. First, it is necessary to build a collaborative training framework for the prediction model, which includes four core modules: (1) Upsampling model The parameters are (2) Station meteorological capture module The parameters are (3) Feature fusion model The parameters are (4) Denoising diffusion model The parameters are In collaborative training, gradients are backpropagated from the loss function, while simultaneously updating... , , and Specifically, the collaborative training framework calculates the difference between predicted noise (the noise prediction result output by the denoising diffusion model after data processing) and real noise (artificially randomly generated reference noise), as well as the difference between the station simulated features (the predicted values processed and output by the station meteorological capture module) and the station measured meteorological data (the real values obtained by the station) during the training phase. The optimization objective is to minimize these two differences, enabling the prediction model to gradually master the ability to accurately estimate real noise from noisy data.
[0042] In one example, the generation process of real noise and predicted noise during co-training includes:
[0043] Step 201, Feature Fusion: Integrate the measured meteorological data from the stations. The feature fusion model is input with reanalysis meteorological data to obtain training baseline ground truth data (this data is a baseline reference data after fusing local measured information, used for ground truth comparison in collaborative training). The aforementioned feature fusion model can be a linear superposition model with fixed parameters; for example, the variables in the overlapping part can use high-resolution station measured meteorological data. Other variables are represented by values from reanalysis meteorological data, or by a cross-attention network model that learns how to use station-measured meteorological data. The local details are used to enhance and correct information at corresponding locations in the reanalysis meteorological data, ultimately outputting enhanced, high-resolution training baseline ground truth data. Its dimensions are consistent with those of the reanalysis meteorological data, but it contains richer local information.
[0044] Step 202, Adding noise: Randomly sample time steps and real noise, and add real noise to the training baseline ground truth data based on the diffusion formula to obtain noisy data (training baseline ground truth data superimposed with real noise). The diffusion formula includes a noise scheduling coefficient, which monotonically increases from the initial value to the final value. The scheduling method of the noise scheduling coefficient is linear scheduling or quadratic scheduling.
[0045] Specifically, the noise-adding steps can be found in the following steps:
[0046] First, determine a random time step t, and uniformly sample a batch of time steps t from {1,2,...,T};
[0047] Then, assuming real noise Randomly generate real noise Its dimensions are the same as the training benchmark ground truth data. same;
[0048] Finally, the noise-added data is calculated according to the diffusion formula. (This is a noise sequence) ,in, , . For noise scheduling coefficients, based on the scheduling function from the initial value Monotonically increasing to For example, if a linear scheduling method is used, then the function is obtained. In this scheme, the denoising diffusion model is a time-series data prediction model based on neural networks, such as U-Net, Transformer framework, etc.
[0049] Step 203: Calculate the enhanced meteorological data during the training phase. : Weather forecast data Input the upsampling model to obtain the initial augmented meteorological data used for training. This initial enhanced meteorological data The length is In terms of time dimension, it is consistent with the actual meteorological data measured at the station. Alignment; Initial Enhanced Meteorological Data After processing by the station meteorological capture module (whose goal is to extract key local information), the simulated features of the stations during the training phase are obtained. The training phase site simulation features Dimensions and station-measured meteorological data Consistent; initial enhanced meteorological data Features of site simulation during training phase Enhanced meteorological data for the training phase is obtained after processing by the feature fusion model. .
[0050] It should be noted that the feature fusion model called in step 203 is the same model as the feature fusion model called in step 201. During the call in step 203, the model parameters remain fixed, and only forward computation is performed. In this scheme, the upsampling model is a time-series data processing network model, such as a 1D convolutional neural network model, a recurrent neural network model, or the Transformer framework. Additionally, the station weather capture module uses a convolutional neural network model.
[0051] Step 204, Denoising and Diffusion Prediction: Input the noisy data, time step, and enhanced meteorological data from the training phase into the denoising and diffusion model to obtain the predicted noise.
[0052] The above example yields noisy data. Time step t, and enhanced meteorological data during the training phase. Input together into the denoising diffusion model , obtain predicted noise .
[0053] After forward propagation processing and computation in steps 201-204, the actual noise and corresponding predicted noise are finally obtained, which are used for subsequent training of the prediction model. Specifically, refer to step 205 below to calculate the loss function and perform backpropagation of parameters:
[0054] The loss function is a weighted sum of the mean square error between predicted noise and actual noise, and the mean square error between the simulated features and measured meteorological data from the training phase. The weighted sum includes weighting coefficients, which can be freely set according to the actual application scenario. The loss function can be calculated using the following formula:
[0055] In the formula, Mean square error, The weights can be set according to the real-world scenario. Additionally, the parameters are updated simultaneously using an optimizer (such as Adam) across the various modules of the collaborative training framework. , , and .
[0056] Step 3, high-resolution data generation, such as Figure 3 As shown: The weather forecast data to be processed is input into the trained upsampling model to obtain the initial enhanced weather data. After the initial enhanced weather data is processed by the station weather capture module, the station simulation features of the inference stage are obtained. After the initial enhanced weather data and the station simulation features of the inference stage are processed by the feature fusion model, the enhanced weather data of the inference stage is obtained. The initial noise is iteratively denoised based on the trained denoising diffusion model. The target weather forecast data is generated by combining the enhanced weather data of the inference stage.
[0057] After the co-training in step 2, the denoising and diffusion model has mastered the ability to accurately estimate the added noise from the noisy data. Therefore, the inference process in step 3 can use the trained denoising and diffusion model. From the initial noise Initially, after T iterations, noise is gradually removed to generate high-resolution target weather forecast data, and the features contained in this target weather forecast data are consistent with the features contained in the low-resolution weather forecast data.
[0058] Enhanced high-resolution weather forecast data generation
[0059] (1) Meteorological forecast data Input into the trained upsampling model Initial enhanced meteorological data were obtained. ;
[0060] (2) Initial enhanced meteorological data Input to the station's weather capture module The site simulation features were obtained during the inference phase. ;
[0061] Simulate features of the inference stage site and initial enhanced meteorological data They are input together into the feature fusion model This results in enhanced meteorological data with higher resolution inference stages. .
[0062] (3) Finally, after iterative denoising, the target weather forecast data is obtained.
[0063] In one example, the iterative denoising process described above includes:
[0064] Step 301: Set the iteration step size T and generate initial noise from a standard Gaussian distribution. Initial noise The dimensions are consistent with the reanalysis meteorological data used in step 2 training.
[0065] Step 302, for each iteration time step t, the current noise data Enhance meteorological data at the current time step T and inference phase. Input the trained denoising diffusion model The prediction noise during the inference stage is obtained. ;
[0066] Step 303, based on prediction noise during the inference phase The noise data from the previous iteration time step is calculated, and this process continues until all iterations are completed to obtain the target weather forecast data. Specifically, the predicted noise is calculated based on the noise obtained in step 302. Simulating the backpropagation process and calculating the data before the noise was added, two results are possible: If t>1, ,in, For added random noise, variance , Return to step 301 and enhance meteorological data based on the inference stage. And updated current noise data and time step Calculate noise; If t equals 1, And then the current This is the final output, namely the target weather forecast data.
[0067] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A method for improving the resolution of meteorological forecast data for wind power prediction, characterized in that, Includes the following steps: Acquire meteorological forecast data, station-measured meteorological data, and reanalysis meteorological data, wherein the meteorological variables in the reanalysis meteorological data are consistent with those in the meteorological forecast data; A collaborative training framework is constructed, comprising an upsampling model, a station meteorological capture module, a feature fusion model, and a denoising diffusion model. The meteorological forecast data, the station measured meteorological data, and the reanalysis meteorological data are input into the collaborative training framework. Based on the training requirements of the denoising diffusion model, real noise is generated, and predicted noise is obtained by the denoising diffusion model through the calculation of the training input data. At the same time, the station meteorological capture module generates station simulated features for the training stage. The difference between the predicted noise and the real noise, as well as the difference between the station simulated features for the training stage and the station measured meteorological data, are calculated using a loss function. The optimization objective is to minimize the two differences. The parameters of the upsampling model, the station meteorological capture module, the feature fusion model, and the denoising diffusion model are updated through backpropagation to complete the collaborative training. The weather forecast data to be processed is input into the trained upsampling model to obtain initial enhanced weather data. After the initial enhanced weather data is processed by the station weather capture module, the station simulation features of the inference stage are obtained. After the initial enhanced weather data and the station simulation features of the inference stage are processed by the feature fusion model, the enhanced weather data of the inference stage is obtained. The initial noise is iteratively denoised based on the trained denoising diffusion model. The target weather forecast data is generated by combining the enhanced weather data of the inference stage.
2. The method for improving the resolution of meteorological forecast data for wind power prediction as described in claim 1, characterized in that, The time resolution of the meteorological forecast data is no less than 1 hour, and the time resolution of the station's measured meteorological data and the reanalysis meteorological data is 15 minutes.
3. The method for improving the resolution of meteorological forecast data for wind power prediction as described in claim 1, characterized in that, Before the collaborative training, the reanalysis meteorological data is adjusted to match the temporal resolution of the station's measured meteorological data by interpolation, using the time resolution of the station's measured meteorological data as a benchmark.
4. The method for improving the resolution of meteorological forecast data for wind power prediction as described in claim 1, characterized in that, The loss function is the weighted sum of the mean square error between the predicted noise and the actual noise, and the mean square error between the simulated features of the training site and the measured meteorological data of the site. The weighted sum includes weight coefficients, which can be freely set according to the actual application scenario.
5. The method for improving the resolution of meteorological forecast data for wind power prediction as described in claim 1, characterized in that, During the collaborative training process, the generation process of the real noise and the predicted noise includes: The measured meteorological data from the stations and the reanalysis meteorological data are input into the feature fusion model to obtain the training baseline ground value data. Randomly sample time steps and the real noise, and add the real noise to the training benchmark ground truth data based on the diffusion formula to obtain noisy data. The diffusion formula includes a noise scheduling coefficient, which monotonically increases from the initial value to the final value. The scheduling method of the noise scheduling coefficient is linear scheduling or quadratic scheduling. The meteorological forecast data is input into the upsampling model to obtain initial enhanced meteorological data for training. The initial enhanced meteorological data is processed by the station meteorological capture module to obtain the station simulation features of the training stage. The initial enhanced meteorological data and the station simulation features of the training stage are processed by the feature fusion model to obtain enhanced meteorological data of the training stage. The noise-added data, the time step, and the enhanced meteorological data during the training phase are input into the denoising diffusion model to obtain the predicted noise.
6. The method for improving the resolution of meteorological forecast data for wind power prediction as described in claim 1, characterized in that, The iterative denoising process includes: Set the iteration step size and generate initial noise from a standard Gaussian distribution; For each iteration time step, the current noise data, the current time step, and the enhanced meteorological data of the inference stage are input into the trained denoising diffusion model to obtain the predicted noise of the inference stage; Based on the prediction noise calculation in the inference phase, the noise data of the previous iteration time step is calculated until all iteration steps are completed, and the target weather forecast data is obtained.
7. The method for improving the resolution of meteorological forecast data for wind power prediction as described in claim 1, characterized in that, The upsampling model is a time-series data processing network model, the station meteorological capture module is a convolutional neural network model, the feature fusion model is a cross-attention network model or a linear superposition model, and the denoising diffusion model is a time-series data prediction model based on neural networks.