CN-121996925-A - Weather forecast data resolution improvement method for wind power prediction

CN121996925ACN 121996925 ACN121996925 ACN 121996925ACN-121996925-A

Abstract

The invention discloses a wind power prediction-oriented weather forecast data resolution improving method, which adopts a cooperative training framework, simultaneously trains a parameterized up-sampling model, a site weather capturing module, a feature fusion model and a denoising diffusion model, enables the up-sampling model to learn and generate training-stage reinforced weather data through an end-to-end training mechanism, and serves as intermediate features suitable for processing of subsequent modules, the site weather capturing module and the feature fusion model jointly provide accurate guiding conditions for the denoising diffusion model, and further enables the denoising diffusion model to learn real evolution rules and uncertainty distribution of weather forecast data on a fine-granularity time scale on the basis, and finally achieves the task of converting low-resolution weather forecast data into high-resolution target weather forecast data.

Inventors

Jiang Chunbi
SHU JIE
MA ZETAO
TIAN XIAOYU
WANG TINGTING
BIE KANG
WANG TONGHE
CUI QIONG

Assignees

中国科学院广州能源研究所

Dates

Publication Date: 20260508
Application Date: 20260114

Claims (7)

1. A weather forecast data resolution improving method for wind power prediction is characterized by comprising the following steps: acquiring weather forecast data, site actual measurement weather data and analysis weather data, wherein the analysis weather data is consistent with weather variables of the weather forecast data; Constructing a collaborative training frame comprising an up-sampling model, a site weather capturing module, a characteristic fusion model and a denoising diffusion model, inputting the weather forecast data, the site actual measurement weather data and the analysis weather data into the collaborative training frame, generating real noise based on the training requirement of the denoising diffusion model, calculating training input data through the denoising diffusion model to obtain predicted noise, generating training stage site simulation characteristics through the site weather capturing module, calculating the difference between the predicted noise and the real noise through a loss function, and the difference between the training stage site simulation characteristics and the site actual measurement weather data, taking the minimized two differences as optimization targets, and updating the parameters of the up-sampling model, the site weather capturing module, the characteristic fusion model and the denoising diffusion model in a back propagation mode to finish collaborative training; Inputting the weather forecast data to be processed into the trained up-sampling model to obtain initial enhanced weather data, processing the initial enhanced weather data by the site weather capturing module to obtain inference stage site simulation features, processing the initial enhanced weather data and the inference stage site simulation features by the feature fusion model to obtain inference stage enhanced weather data, carrying out iterative denoising processing on initial noise based on the trained denoising diffusion model, and generating target weather forecast data by combining the inference stage enhanced weather data.
2. The method for improving the resolution of weather forecast data for wind power prediction according to claim 1, wherein the time resolution of the weather forecast data is not less than 1 hour, and the time resolution of the site measured weather data and the re-analyzed weather data is 15 minutes.
3. The method for improving the resolution of weather forecast data for wind power prediction according to claim 1, wherein the time resolution of the weather forecast data is adjusted by interpolation based on the time resolution of the site measured weather data before the collaborative training is performed, so that the time resolution of the weather forecast data is consistent with the time resolution of the site measured weather data.
4. The method for improving the resolution of weather forecast data for wind power prediction according to claim 1, wherein the loss function is a weighted sum of a mean square error of the prediction noise and the real noise and a mean square error of the site simulation feature of the training stage and the site actual measurement weather data, the weighted sum comprises a weight coefficient, and the weight coefficient is freely set according to an actual application scene.
5. The wind power prediction oriented weather forecast data resolution improvement method according to claim 1, wherein in the co-training process, the real noise and the prediction noise generating process includes: Inputting the site actual measurement meteorological data and the analysis meteorological data into the characteristic fusion model to obtain training reference truth value data; Randomly sampling time steps and the real noise, and adding the real noise into the training reference true value data based on a diffusion formula to obtain noise adding data, wherein the diffusion formula contains a noise scheduling coefficient, the noise scheduling coefficient monotonically increases to a termination value from an initial value, and the scheduling mode of the noise scheduling coefficient is linear scheduling or square scheduling; Inputting the weather forecast data into the up-sampling model to obtain initial enhanced weather data for training, processing the initial enhanced weather data by the site weather capturing module to obtain site simulation characteristics of the training stage, and processing the initial enhanced weather data and the site simulation characteristics of the training stage by the characteristic fusion model to obtain enhanced weather data of the training stage; And inputting the noise adding data, the time step and the training stage reinforced meteorological data into the denoising diffusion model to obtain the prediction noise.
6. The wind power prediction oriented weather forecast data resolution enhancement method of claim 1, wherein the iterative denoising process comprises: setting an iteration step length, and generating initial noise from standard Gaussian distribution; for each iteration time step, inputting the current noise data, the current time step and the inference stage enhanced meteorological data into the trained denoising diffusion model to obtain the prediction noise of the inference stage; And calculating noise data of the previous iteration time step based on the prediction noise of the reasoning stage until all iteration steps are completed, and obtaining the target weather forecast data.
7. The wind power prediction oriented weather forecast data resolution improvement method according to claim 1, wherein the up-sampling model is a time sequence data processing network model, the site weather capturing module is a convolution 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 sequence data prediction model based on a neural network.

Description

Weather forecast data resolution improvement method for wind power prediction Technical Field The invention relates to the technical field of weather forecast data resolution enhancement, in particular to a wind power prediction-oriented weather forecast data resolution enhancement method. Background The medium-short term prediction (24 to 240 hours in future) of wind power is a key basis for power grid dispatching and transaction, and in order to meet the fine requirements of a dispatching system, a wind power plant needs to report a power prediction curve with time resolution of 15 minutes. The above prediction task core input data depends on the output of global numerical weather forecast modes (e.g. ECMWF, GFS), however, the temporal resolution of the global numerical weather forecast modes is significantly lower than this requirement. For example, the resolution of the forecast data of the ECMWF is 1 hour in the forecast meteorological data within the first 90 hours from the time of the report, and the subsequent forecast period is further reduced to 3 hours or 6 hours. Therefore, in the technical process of wind power prediction, there is a problem that the inherent input and output resolutions are not matched, specifically how to generate high-frequency wind power prediction data with the time resolution up to 15 minutes based on the low-frequency numerical weather prediction data with the time resolution of only 1 hour or more. In the current weather forecast data resolution improvement method, the main stream thinking still takes interpolation technology as a core, but the methods have obvious defects in practical application. The following is described in terms of two general interpolation methods: The first method is to directly perform mathematical interpolation (such as spline interpolation) on the original low-resolution prediction data, and does not introduce any measured data to participate in correction. In addition, the interpolation method assumes that the meteorological elements change uniformly in time, and it is difficult to accurately reflect the rapid change process existing in the real atmosphere, such as the rapid rise or fall of wind speed, so that the forecast details under the critical weather phenomenon are lost. The second method is to introduce measured data to correct on the basis of interpolation, and is generally divided into two steps, namely, interpolation is carried out on low-resolution data, then a machine learning model is established to correct the interpolation result, and the two-stage processing mode has obvious defects. Firstly, the interpolation error of the first stage directly influences the input of the second stage model to cause error transmission, secondly, the targets of the two stages are inconsistent, the interpolation of the first stage pursues curve smoothness, the correction model of the second stage pursues fitting of measured values, the two stages are difficult to cooperatively optimize, and the improvement of the final effect is limited. In summary, the prior art processes the "data upscaling" and the "data correction" as two independent links, and lacks an end-to-end joint learning mechanism from the original low-resolution data to the high-resolution result, so that the problem of high-quality input data required in wind power prediction cannot be thoroughly solved. Disclosure of Invention Aiming at the problems, the invention provides a weather forecast data resolution improvement method for wind power prediction, which realizes reliable conversion from weather forecast data with low time resolution to wind power prediction data with high time resolution. In order to solve the technical problems, the technical scheme of the invention is as follows: a weather forecast data resolution improvement method for wind power prediction comprises the following steps: acquiring weather forecast data, site actual measurement weather data and analysis weather data, wherein the analysis weather data is consistent with weather variables of the weather forecast data; Constructing a collaborative training frame comprising an up-sampling model, a site weather capturing module, a characteristic fusion model and a denoising diffusion model, inputting the weather forecast data, the site actual measurement weather data and the analysis weather data into the collaborative training frame, generating real noise based on the training requirement of the denoising diffusion model, calculating training input data through the denoising diffusion model to obtain predicted noise, generating training stage site simulation characteristics through the site weather capturing module, calculating the difference between the predicted noise and the real noise through a loss function, and the difference between the training stage site simulation characteristics and the site actual measurement weather data, taking the minimized two differences as optimization targets, and updating th