CN-122021267-A - Residual error correction wind profile radar inversion method based on dynamic space-time field network

CN122021267ACN 122021267 ACN122021267 ACN 122021267ACN-122021267-A

Abstract

The invention discloses a residual correction wind profile radar inversion method based on a dynamic space-time field network, which relates to the technical field of meteorological radar data processing and comprises the steps of constructing a four-dimensional observation tensor and fusing Beidou sounding truth values, synchronously extracting wind field space-time characteristics by using mask time self-attention and a graph neural network through a residual learning mechanism guided by physical priori, and combining a composite loss function to realize high-precision and physical consistent inversion of the whole wind profile. The method can effectively solve the problems of insufficient precision, vertical fracture and time discontinuity caused by failure of 'wind field horizontal uniformity' assumption in the traditional method, and remarkably improves the wind field inversion quality under complex meteorological conditions.

Inventors

LIU HUI
TAO RUI
ZHANG XIANG
REN KE
WEN JIANWEI

Assignees

内蒙古自治区气象数据中心(内蒙古自治区气象探测中心、内蒙古自治区气象档案馆)

Dates

Publication Date: 20260512
Application Date: 20260112

Claims (6)

1. A residual error correction wind profile radar inversion method based on a dynamic space-time field network is characterized by comprising the following steps: S1, organizing observation data of wind profile radar at multiple moments, multiple heights and multiple beams into four-dimensional tensors, processing missing measurement values to generate a mask matrix, and providing structural input for space-time modeling; s2, constructing a basic wind field by adopting a traditional vector algorithm as a physical priori, and simultaneously calculating a full-profile residual error by utilizing a high-precision true value acquired by a Beidou sounding system, and taking the full-profile residual error as a training target label of a deep learning model; S3, extracting wind field evolution characteristics of a time dimension through a transducer, extracting interlayer coupling characteristics of a vertical space through a graph neural network, and adopting an attention mechanism to deeply fuse the space-time characteristics with a physical priori to generate a fused characteristic field with physical constraint; S4, inputting the fusion characteristics into a residual error prediction head, and outputting a full-height layer residual error field at one time; S5, ensuring accuracy of residual regression loss, restraining evolution stability of time smooth loss, strengthening vertical continuity of space smooth loss, and training by adopting a time sequence division strategy to ensure generalization capability of the model in unknown weather process.
2. The method for inverting a residual correction wind profile radar based on a dynamic space-time field network as set forth in claim 1, wherein said step S1 comprises the steps of forming K observables of the wind profile radar for T consecutive times, H heights, and B beams into a four-dimensional tensor D (T, z, B, K) E The observed quantity comprises a radial speed Vr, a signal-to-noise ratio SNR and a spectrum width W, wherein, Processing the missing values using time/height direction interpolation while generating a mask matrix ∈ Recording the availability of data, marking 1 indicates that the original data is available, and 0 indicates that the point is interpolated.
3. The method for inverting the residual correction wind profile radar based on the dynamic space-time field network according to claim 2, wherein the step S2 is specifically as follows: Generating a full-height layer foundation wind field matrix by using a traditional three-beam/five-beam inversion method , As a physical priori, the method is a final correction standard, performs space-time matching on the wind profile radar and the co-located Beidou sounding true value, and aligns the vertical height axis to generate a full-height layer real wind field label matrix ; Sample pair construction will be all-time, full-height radar observation tensor D, mask matrix M, and physical prior field Beidou sounding truth value tag aligned with space time Correlating to construct training sample pair of observation field-truth field Calculating the deviation between the true value of each layer of the full profile and the basic field to obtain a target residual field of model training , =[ ]。
4. The method for inverting the residual correction wind profile radar based on the dynamic space-time field network according to claim 3, wherein the step S3 is specifically as follows: (1) Feature dimension normalization: the preset feature embedding dimension is as follows Mapping original observations and physical priors to Space, which ensures that different source data are fused in a unified semantic space; (2) Time encoder: extracting time sequence features by using a transducer time attention mechanism; a full-height observation sequence D of past m times (t-m: t, B, K) and the corresponding mask matrix M (t-M: t, , B, K)。 The operation description comprises the steps of extracting historical evolution features in parallel for each height layer by adopting a mask time self-attention mechanism; output: time feature matrix: (3) Spatial encoder: extracting vertical coupling characteristics by using a Graph Neural Network (GNN) and graph convolution; the full-height observation data D at the present time is input (t, B, K) and the corresponding mask matrix M (t, , B, K) The operation description includes that a Graph Neural Network (GNN) is utilized to execute graph convolution operation, H height layers are abstracted into graph nodes, wind shear and physical coupling characteristics in the vertical direction are captured through a neighborhood aggregation mechanism, convolution kernel weights are dynamically adjusted through masks, and error diffusion in the vertical space is restrained; output: spatial feature matrix: (4) Multimode physical fusion layer: The method adopts an attention weighted fusion mechanism, and the mechanism utilizes As a query, guided model learning generation is directed to And Dynamic weight vector of (a) And Implementing physical constraint information Nonlinear spatiotemporal features 、 Is a depth fusion of (2); Input: , , Description of the operation: physical query field construction using a full connectivity layer (MLP) to Mapping into a query vector , Dynamic weight field generation to query matrix Respectively with time characteristic matrix And a spatial feature matrix Similarity calculation is carried out, and dynamic weight matrix aiming at the full-height layer is generated through normalization of Sigmoid activation function And The whole profile is synchronously fused, namely the generated dynamic weight is utilized to carry out element-level weighting on the time space characteristics and splice with the basic wind field characteristics; output full-height layer fusion feature matrix Formula (VI) 。
5. The method for performing residual correction wind profile radar inversion based on dynamic space-time field network as set forth in claim 4, wherein said step S4 is specifically implemented by combining a feature matrix with a full-height layer by using residual pre-measurement head composed of multiple perceptrons or convolution layers Performing nonlinear mapping, and directly outputting wind field residual matrixes of all height layers at the moment ; =[ ] Superposing the initial wind field and the prediction residual error to obtain a final wind field ; = 。
6. The method for inverting the residual correction wind profile radar based on the dynamic space-time field network according to claim 5, wherein the step S5 is specifically: For each sample, the full-height layer observation tensor D (t-M: t, 1:H, B, K), the quality mask matrix M (t-M: t, 1:H, B, K) and the base wind field matrix of the current time and its history period are input once ; Generating a target residual error matrix covering the full height layer by differencing the Beidou sounding true value and the basic wind field ; Model training employs a composite loss function including space-time smoothness constraints + Wherein, the , , The weight coefficient for each loss term is defined as follows: Residual regression loss Minimizing model prediction residuals using Mean Square Error (MSE) Residual with target The difference between the wind field inversion accuracy is ensured; wherein H is the total height layer number of wind profile radar detection, z is the vertical height layer index, z is E ; 2-Dimensional wind field residual vector obtained by predicting model at time t and height z [ ]; Time smoothing loss Constraining the evolution stability of the wind field on a time axis, calculating the total profile vector difference between adjacent moments in m continuous moments, I, a time sequence index, which is traversed from the second moment in a window to calculate adjacent difference values; Outputting a wind field matrix at the full height layer of the moment i; Space smoothing loss : Wherein: The final output wind field after z-th layer correction is calculated by the following way 。

Description

Residual error correction wind profile radar inversion method based on dynamic space-time field network Technical Field The invention relates to the technical field of meteorological observation, in particular to a residual error correction wind profile radar inversion method based on a dynamic space-time field network. Background Wind profile radar is used as a key device in a modern meteorological observation system, and measures Doppler radial velocity in each beam direction by emitting a plurality of directional beams (usually 3 beams or 5 beams, including oblique beams with perpendicular beams orthogonal to azimuth), and reflects wind field information such as atmospheric wind speed, wind direction and the like based on a vector synthesis principle. The technology can provide vertical wind profile data with high space-time resolution, and has important application value in the fields of numerical weather forecast, disaster weather early warning, aerodynamic research and the like. The wind field profile is accurate, simulation capability of a numerical mode on an atmospheric movement process can be remarkably improved, accuracy of weather forecast in short term to medium term is improved, meanwhile, real-time capturing capability of the wind field profile radar on wind field structures in extreme weather processes such as strong convection, strong wind and sand storm is provided for early recognition and early warning of meteorological disasters, and in addition, continuous observation characteristics of the wind field profile radar provide high-quality data bases for basic scientific researches such as evolution of an atmospheric boundary layer and a convection triggering mechanism. At present, inversion of wind measurement data of a wind profile radar mainly depends on a traditional vector calculation method, and typical schemes comprise a 3-beam method, a 4-beam method and a 5-beam joint verification method. The 3-beam method utilizes two adjacent oblique beams and a vertical beam to form a triangular relation, calculates horizontal wind components through a trigonometric function, is suitable for a fast airflow change scene to improve time resolution, 4-beam method utilizes east-west and north-south oblique beams to independently calculate U, V wind components, avoids the influence of ground clutter interference on the vertical beam, generally has higher wind measuring precision, and 5-beam method combines the advantages of the two, ensures data integrity and also gives consideration to inversion accuracy, and is a scheme widely adopted in current business application. However, the above methods are all based on the idealized assumption that the wind field is uniform in the horizontal direction, and the actual atmosphere often shows significant non-uniformity and three-dimensional structural characteristics especially under the influence of strong convection, wind shear or complex terrain, resulting in significant increase of inversion errors of the conventional algorithm under complex meteorological conditions. More critical is that wind profile radar wind measurement error sources are various and have strong coupling, including radar deployment position difference, inter-beam space separation effect, beam deflection caused by atmospheric refraction, ground clutter pollution, interference of extreme weather on signal quality and the like. The existing quality control means mostly adopt strategies such as consistency test, suspicious data elimination or simple average, and the like, and can inhibit abnormal values to a certain extent, but inevitably cause data missing measurement, destroy the vertical continuity and time sequence integrity of wind profiles, and are difficult to realize effective correction of error essence. In addition, the traditional inversion algorithm is highly dependent on the beam data of the radar, and lacks systematic fusion of external high-precision observation information. Although research has been attempted to introduce exploration data for post-evaluation or rough calibration, an error correction model with physical constraints based on high-precision truth-value guidance has not been constructed yet, limiting further improvement of inversion accuracy. In summary, the existing wind profile radar inversion technology has obvious limitations in theoretical assumption, error processing mechanism and multi-source data fusion, and is difficult to meet urgent requirements of high-precision and high-integrity wind field products in fine weather service. Therefore, development of a novel inversion method capable of breaking through the constraint of traditional vector algorithm assumptions, effectively fusing external high-precision observation and fully excavating the wind field space-time evolution law is needed to realize accurate, continuous and physically consistent dynamic reconstruction of wind profile data in a complex atmospheric environment. Disclosure of Invention T