Search

CN-121685269-B - Physical information guiding wind speed field data downscaling method integrating terrain and time perception

CN121685269BCN 121685269 BCN121685269 BCN 121685269BCN-121685269-B

Abstract

The invention discloses a physical information guided wind speed field data downscaling method integrating terrain and time perception, which comprises the steps of 1, multisource physical constraint modeling and data construction, 2, construction of a generated downscaling frame based on a conditional diffusion probability model, 3, design of a noise prediction network structure integrating terrain and time priori, and 4, model training and rapid sampling reasoning. The invention solves the problems that the wind speed field texture generated by the existing deep learning downscaling method is excessively smooth, high-frequency turbulence details are lost, and the terrain forcing effect and the wind speed season/day periodic physical rule are ignored, and provides a framework combining an explicit physical feature engineering and a conditional diffusion model, and the high fidelity, the physical consistency and the reasoning efficiency of the generated result are considered through the double deep fusion of the terrain and time information.

Inventors

  • YANG QINMIN
  • ZHAO YUHANG
  • JIANG XUEJUN
  • ZHOU YIHUAN
  • MENG WENCHAO
  • PAN YU

Assignees

  • 浙江大学
  • 湖州工业控制技术研究院

Dates

Publication Date
20260512
Application Date
20260209

Claims (7)

  1. 1. A method for downscaling wind speed field data guided by physical information integrating terrain and time perception, comprising the steps of: step 1, carrying out multisource physical constraint modeling and data construction, wherein the steps comprise: Step 1.1, constructing a wind field data pair; step 1.2, constructing multichannel topographic feature data, comprising: construction of high resolution static terrain data The method comprises the following four channel information (1) digital elevation model data, (2) horizontal topography gradient, and Sobel operator The data of the digital elevation model is obtained by convolution calculation, which reflects the steep degree of the terrain in the east-west direction, (3) the vertical terrain gradient is obtained by utilizing the Sobel operator The data of the digital elevation model is obtained by convolution calculation, and the steep degree of the terrain in the north-south direction is reflected; The four channel information is processed through a shallow convolution network to obtain a final topographic feature map; the step 1.3 includes: Constructing a time feature vector tau by adopting sine/cosine coding: , wherein doy represents the day of the year, Indicating the annual constant in days, the hour indicating the hours of the day, A daily cycle constant expressed in hours; step 1.3, performing time periodic feature coding; Step 2, constructing a generated downscaling frame based on a conditional diffusion probability model, which comprises the following steps of; Step 2.1, adding noise to the real wind field to build a forward diffusion process, comprising: the denoising diffusion probability model is used as a backbone frame, and the downscaling task is modeled as information under physical conditions The guided conditional probability distribution generation process, the forward process being a fixed Markov chain, by generating a direct response to the real high resolution data Gradually adding Gaussian noise until becoming pure Gaussian noise T is a preset parameter representing the total number of iterations required to completely convert real data into pure Gaussian noise, and at any diffusion time step T, noisy data Direct sampling is performed by: , Wherein, the The original signal is represented by a representation of the original signal, Representing the introduced random Gaussian noise term, sign Representation of Obeys standard normal distribution, the mean vector of the distribution is 0, and the covariance matrix is an identity matrix T represents a diffusion time step, is an integer discrete variable with a value ranging from 1 to T, Is a predefined noise variance scheduling parameter corresponding to the diffusion time step t, which monotonically decreases with the diffusion time step t; Step 2.2, constructing a reverse denoising model to predict added noise; Step 3, designing a noise prediction network structure integrating topography and time priori, wherein the noise prediction network structure comprises a residual block integrating channel-space attention, a topography information integration module and a time information integration module, and the design of the residual block integrating channel-space attention comprises the following steps: Introducing a channel attention module and a spatial attention module which are connected in series in a residual block; The channel attention module is implemented by first inputting feature graphs Global average pooling and global maximum pooling are carried out along the space dimension to aggregate space information, then two generated feature vectors are respectively input into a multi-layer perceptron sharing weight, and finally the outputs of the multi-layer perceptron are added and a channel weight graph is generated through a Sigmoid activation function ; The channel weight graph is acted on the input characteristics to obtain characteristics after channel refinement ; The spatial attention module is implemented by firstly refining the channel characteristics On the basis of the above, respectively carrying out global average pooling and global maximum pooling along the channel dimension to obtain two-dimensional feature images, then splicing the two feature images, and passing through one Fusion is carried out on convolution layers of (2), and finally a space weight graph is generated through Sigmoid function ; Finally, the space weight graph is acted on Resulting in a final feature enhanced by a dual attentive mechanism And the final feature Adding the residual connection with the original input as the input of the next layer; And 4, performing model training and rapid sampling reasoning, wherein the method comprises the following steps of: Step 4.1, designing a loss function and training a model; and 4.2, constructing a rapid sampling reasoning strategy.
  2. 2. A method of downscaling wind speed field data guided by physical information fusing terrain and time perception as claimed in claim 1, wherein said step 1.1 comprises: Defining low resolution wind farm data as The high-resolution wind field data is , Is the object of the model training and, Is one of the input conditions of the model, and in actual operation, the model is obtained by bicubic interpolation Upsampling to AND The same dimensions are then entered into the model.
  3. 3. A method of downscaling wind speed field data guided by physical information fusing terrain and time perception as claimed in claim 1, wherein said step 2.2 comprises: The reverse process is aimed at removing noise from pure gaussian Recovering the original signal Training a noise prediction network Its input is the current noisy data The diffusion time step t and all the physical condition information c are outputted as predictions of the noise added to the diffusion time step.
  4. 4. The method for downscaling wind speed field data guided by physical information fusing terrain and time perception according to claim 2, wherein the design of the terrain information fusing module in step 3.2 comprises the following steps: designing a double fusion strategy; shallow layer direct fusion, namely, the topographic information generated in the step 1.2 is characterized Noisy data input to a network Upsampling of Directly splicing in the channel dimension; The global multiscale topographic attention is characterized by designing a parallel multiscale topographic feature extraction module which consists of three parallel branches, adopting three paths of cavity convolutions with different expansion rates, respectively extracting microscopic topographic details, mesoscale topographic relief and macroscopic topography trend, merging the three paths of features and generating a global topographic attention map The global topography attention map is used to modulate the output characteristics of the encoder : , Wherein, the Representing multiplication element by element, modulated features Directly to the decoder via a long hop connection.
  5. 5. The method for downscaling wind speed field data guided by physical information fused with terrain and time perception according to claim 4, wherein the design of the time information fusion module in step 3.3 comprises: First, the position code vector of the diffusion time step t Embedding vector with physical time feature vector τ in step 1.3 Adding to obtain a combined time embedding : , Within each residual block in the network, the full connection layer is utilized to perform the following Mapping to scaling factor gamma and shifting factor beta, and mapping to characteristic diagram Affine transformation is carried out: , Wherein the method comprises the steps of By means of the mechanism, the noise prediction network can dynamically adjust the response of the convolution kernel according to the double conditions of the diffuse noise level determined by t and the moment of the physical world determined by tau, so that the wind field characteristics conforming to the seasons and the daily change rules can be accurately recovered while denoising.
  6. 6. A method of downscaling wind speed field data guided by physical information fusing terrain and time perception as claimed in claim 1, wherein said step 4.1 comprises: the loss function is constructed by using the L1 norm, and the specific formula is as follows: , Wherein, the The loss function is represented by a function of the loss, Representing the mathematical expectation that the data will be, Representing the random gaussian noise term introduced, Representing that noise prediction network is based on noisy data The noise predicted by the diffusion time step t and the physical condition information c, Represents an L1 norm; During training, a random gradient descent algorithm is used to determine the loss function Calculating gradients And updating the network parameter θ by back propagation until the loss function And (5) convergence.
  7. 7. A method of downscaling wind speed field data guided by physical information fusing terrain and time perception as claimed in claim 1, wherein said step 4.2 comprises: adopting a denoising diffusion implicit model sampling algorithm in an reasoning stage; setting the diffusion time step of the sampling sub-sequence as At any diffusion time step According to the corresponding diffusion time steps Is a noisy sample of (1) And noise Noisy samples from previous diffusion time steps are calculated The formula of (2) is as follows: , Wherein, the Respectively corresponding to diffusion time steps And Is used to determine the pre-defined noise variance scheduling parameters, Is the final high resolution wind field.

Description

Physical information guiding wind speed field data downscaling method integrating terrain and time perception Technical Field The invention relates to the field of wind energy resource evaluation and meteorological data processing, in particular to a wind speed field data downscaling method guided by physical information integrating topography and time perception. Background With global energy conversion, wind energy is increasingly important as a clean renewable energy source. The wind resource data set with high space-time resolution is the basis for macroscopic site selection, microscopic layout optimization and power grid dispatching operation of the wind power plant. Currently, meteorological data is mainly derived from numerical weather forecast models or analytical data (e.g., ERA 5). However, due to the computational resources and the grid spacing of the physical model, the spatial resolution of these data is generally low (e.g., ERA5 is 30 km), local wind field features (e.g., ridge acceleration effects, canyon wind, etc.) under complex terrain cannot be accurately characterized, and it is difficult to directly meet the requirements of refined wind energy development. In order to solve the problem of insufficient resolution, the existing methods are mainly divided into dynamic downscaling and statistical downscaling. The dynamic downscaling is operated by using the regional climate model in a nested way, and has definite physical meaning, extremely high calculation cost and long time consumption. The statistical downscaling is high in calculation efficiency by establishing a mapping relation from low resolution to high resolution. In recent years, super-resolution techniques (e.g., convolutional neural networks, transducers) based on deep learning have been applied to wind farm downscaling. However, existing deep learning-based downscaling methods suffer from the disadvantage that, on the one hand, the traditional convolutional neural network or transducer model is generally aimed at minimizing pixel-level errors (e.g. mean square error), which tend to produce averaged results, resulting in loss of vital high frequency details (e.g. gusts, turbulent texture) in the wind farm. Secondly, the physical mechanism is lost, and the wind speed field is mostly treated as common natural image processing by the existing method, so that the physical attribute of the wind speed field is ignored. In fact, local terrain is the primary driving force for wind speed variation, but most methods simply splice terrain as input, failing to effectively capture multi-scale terrain-wind field interactions. Thirdly, the time prior is not utilized, and the wind speed has obvious daily variation (such as sea and land wind) and seasonal variation rules. The existing model often ignores this strong prior knowledge, resulting in poor physical consistency of the generated wind field in the time dimension. Therefore, how to generate a high-resolution wind field with rich details and conforming to the topography and time physical laws while calculating with high efficiency is a technical problem to be solved currently. Disclosure of Invention The invention aims to provide a wind speed field data downscaling method guided by physical information integrating terrain and time perception, so as to solve the problems in the background technology. In order to achieve the above purpose, the present invention provides the following technical solutions: a method of merging terrain and time-aware physical information-guided downscaling of wind speed field data, comprising: step 1, carrying out multisource physical constraint modeling and data construction, wherein the steps comprise: Step 1.1, constructing a wind field data pair; Step 1.2, constructing multichannel topographic feature data; step 1.3, performing time periodic feature coding; Step 2, constructing a generated downscaling frame based on a conditional diffusion probability model, which comprises the following steps of; step 2.1, adding noise to a real wind field to construct a forward diffusion process; Step 2.2, constructing a reverse denoising model to predict added noise; Step 3, designing a noise prediction network structure fusing terrain and time priori, wherein the noise prediction network structure comprises a residual block integrating channel-space attention, a terrain information fusion module and a time information fusion module; And 4, performing model training and rapid sampling reasoning, wherein the method comprises the following steps of: Step 4.1, designing a loss function and training a model; and 4.2, constructing a rapid sampling reasoning strategy. Further, the step 1.1 includes: Defining low resolution wind farm data as The high-resolution wind field data is,Is the object of the model training and,Is one of the input conditions of the model, and in actual operation, the model is obtained by bicubic interpolationUpsampling to ANDThe same dimensions are then