Observation station data assimilation method based on deep learning

Abstract

The invention discloses an observation station data assimilation method based on deep learning, and belongs to the technical field of meteorological data assimilation, and the method comprises the steps: obtaining forecast background field data of a Fuhu meteorological big model, obtaining observation station data, the observation data comprises the actual measurement data of a ground meteorological observation station, a radar, a sounding station and a wind profile observation station, and the actual measurement data of the ground meteorological observation station, the radar, the sounding station and the wind profile observation station; performing spatial interpolation, time alignment and standardization on the background field and the observation data, and introducing position codes and time codes; constructing a deep learning assimilation model based on a sliding window Transform, inputting a background field and observation data, and outputting an assimilation analysis field result; and the effectiveness of the method is verified through comparison and evaluation with fifth-generation global atmosphere reanalysis data of the European mid-term weather forecast center. According to the method, the fineness and the physical consistency of local meteorological elements can be remarkably improved while the integrity of large-scale forecasting is kept, and high-precision assimilation of meteorological forecasting data is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of meteorological data assimilation, and particularly relates to an observation station data assimilation method based on deep learning. BACKGROUND

[0002] Data assimilation aims to combine observation data and numerical model background fields to obtain the optimal atmospheric state estimation. Although existing assimilation methods such as four-dimensional variation (4D-Var) and ensemble Kalman filter (EnKF) can effectively integrate information, they have problems such as high computational cost and strong error distribution assumption. In recent years, deep learning models, especially Transformers, have shown outstanding performance in processing large-scale spatio-temporal data. Swin Transformer uses window self-attention mechanism to realize efficient modeling of local and global features. However, existing research mainly focuses on bias correction and downscaling, and less on assimilation methods that integrate background fields and observation data. SUMMARY

[0003] The technical problem to be solved by the application is to provide an observation station data assimilation method based on deep learning to solve the problems of the background art.

[0004] To solve the above technical problems, the application adopts the following technical solutions: An observation station data assimilation method based on deep learning, specifically comprising the following steps: Step 1: Obtain the forecast background field data of the Fuxi meteorological model, with a spatial resolution of 0.25°, a time interval of 6 hours, and containing 88 meteorological variables; Step 2: Obtain observation data, including ground meteorological observation stations, radars, sounding stations, and wind profile observation stations; Step 3: Preprocess the background field data and observation data, including spatial interpolation, time alignment, standardization, and the introduction of position encoding and time encoding; Step 4: Construct a deep learning assimilation model based on sliding window Transformer, input the background field data and observation data jointly, and obtain the assimilation result; Step 5: Evaluate the assimilation result with the fifth generation global atmospheric reanalysis data ERA5 of the European Centre for Medium-Range Weather Forecasts (ECMWF).

[0005] As a further preferred scheme of the observation station data assimilation method based on deep learning, in step 1, the 88 meteorological variables include 88 variables of high-altitude variables on pressure levels levels = [50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000], surface variables, and sea wave variables.

[0006] As a further preferred scheme of the observation station data assimilation method based on deep learning of the present application, in step 2, the spatial distribution of the observation data is taken as the background of the gray terrain height, the national observation station is a blue dot, the radar is a cross, the sounding station is a hollow circle, and the wind profile station is a triangular symbol.

[0007] As a further preferred scheme of the observation station data assimilation method based on deep learning of the present application, in step 3, the spatial interpolation method is used to interpolate the observation data to the background field grid, and a mask is added to the position without matching observation, and the obtained observation field after processing is the same resolution and grid as the background; The observation data and the background field are time-aligned by weighted averaging all observation data within 1 hour before and after the specific analysis time point; The standardized method is used to unify the distribution range of each variable of the input data; The latitude and longitude encoding in the form of sine and cosine and the time encoding are used as additional observation channels to maintain the geographical and time characteristics; The background field and the observation data are spliced in the channel dimension to form a mixed information flow input.

[0008] As a further preferred scheme of the observation station data assimilation method based on deep learning of the present application, in step 4, the Swin Transformer structure is used, wherein the Swin Transformer is a sliding window Transformer; A three-branch structure composed of background flow, observation flow and mixed flow is constructed, and finally a residual form is used to write back the increment to realize the analysis of three-dimensional variation 3D-Var: ; wherein, is the state vector of the analysis field, is the state vector of the background field, is the observation vector, and H is the observation operator, is the tangent linear operator of H, B is the covariance matrix of the background field error, and R is the covariance matrix of the observation error; is the weight matrix of the analysis field; in the deep network, the weight and the increment are learned in an implicit manner, and then applied to the background features and the output through the residual form; The mixed information flow is used as the input vector and enters the Swin Transformer backbone network to complete the end-to-end assimilation and increment generation; The backbone feature extraction adopts a four-level hierarchical structure, and the Swin Transformer network is stacked level by level and the block Patch merging is performed between levels; In the deep feature space, end-to-end learning is performed incrementally, and the residual form is written back to the background branch, while the observation and mixing branches are updated for the next level of fusion; After encoding and multi-level fusion, the decoding end is up-sampled to the spatial resolution of 160x160, and is projected to 88 physical variable channels through 1x1 convolution; pixel interpolation is used to restore the original resolution, which is consistent with the input background grid; The training strategy uses AdamW optimization, where the AdamW regular parameters =0.9, =0.999, and the weight decay =10 -5 ; and a combination of linear warm-up and cosine annealing learning rate schedule is used: the learning rate is linearly increased from 10 -8 to 2x10 -3 in the first 500 steps, and the form is: ; Where, =500 is the first 500 steps; start = 10 -8 is the starting learning rate; stop = 2x10 -3 is the end learning rate; After the warm-up is over, it is gradually decayed according to the cosine annealing within the total iteration of 6000 steps; In terms of loss function, the analysis moment and multi-time effect prediction are optimized at the same time, and the basic term is the latitudinal weighted L1 aligned with ERA5: ; Where, C, H, and W are the number of channels and the number of latitude and longitude grid points, is the true value of channel c at latitude i and longitude j, is the predicted value of channel c at latitude i and longitude j, is the weight at latitude .

[0009] As a further preferred scheme of the observation station data assimilation method based on deep learning of the application, in step 5, the assimilation effect is evaluated by the root mean square error RMSE index; The effect evaluation index mainly uses global latitude weighted RMSE and regional RMSE; the global latitude weighted RMSE is defined as: ; Where, H and W are the number of channels and the number of latitude and longitude grid points, is the true value of channel c at latitude i and longitude j, is the predicted value of channel c at latitude i and longitude j, is the weight at latitude weighting of the grid points in the given latitude-longitude subdomain, and reporting the normalized difference ; the area RMSE on a given latitude-longitude subdomain is defined as follows: ; where, , , , denotes the range of grid points corresponding to the selected area, is the weight of the latitude .

[0010] Compared with the prior art, the above technical scheme has the following technical effects: The application can not only significantly improve the consistency and accuracy of assimilation and prediction, but also has the following advantages: (1) high efficiency in a large-scale parallel computing environment, and only a few seconds are needed for a single assimilation inference; (2) adaptive to different types and coverage ranges of observation data, and good scalability; (3) while ensuring the physical consistency of large-scale dynamic and thermal fields, the ability to depict local weather elements is significantly improved; the application provides a new deep learning assimilation scheme for the next generation of intelligent numerical weather prediction system, and has a wide application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is a flowchart of the observation station data assimilation method based on deep learning of the application; Figure 2 is a Swin Transformer assimilation model structure diagram of the application. DETAILED DESCRIPTION

[0012] To make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0013] As shown in Figure 1 , the embodiment of the observation station data assimilation method based on deep learning of the application comprises the following steps: (1) Data acquisition A background field containing 88 meteorological variables is obtained from the Fuxi meteorological model, with a resolution of 0.25° and a time interval of 6 hours. Observation data is obtained, including observation stations, radars, sounding stations and wind profile stations.

[0014] The forecast data variable summary of the Fuxi weather model is shown in Table 1.

[0015] Table 1

[0016] (2) Data preprocessing The observation station data and the grid background field type differ greatly. In order to match the grid alignment and reduce the observation noise, a spatial interpolation method is used to interpolate the observation data to the background field grid, and a mask is added to the position without matching observation. The observation field obtained after processing has the same resolution and grid as the background.

[0017] The model uses a ±1 hour assimilation window. For a specific analysis time point (for example, 00:00 UTC), the model will use all observation data within 1 hour before and after this time point (i.e. 1 hour before to 1 hour after, a total of 2 hours) to achieve time alignment of observation data and background field.

[0018] Standardization is used to unify the distribution range of each variable of the input data.

[0019] Considering the geographical correlation bias and the time sequence correlation of the input data, the latitude and longitude encoding in the form of sine and cosine and the time encoding are used as additional observation channels to maintain the geographical and time characteristics.

[0020] The background field and the observation data are spliced in the channel dimension to form a “mixed information flow” input.

[0021] Model structure: a deep learning assimilation model based on Swin Transformer is constructed, the input is the background field and the observation data, and the output is the assimilation analysis field result; the main model structure of the deep learning-based observation station data assimilation method is as shown in Figure 2 .

[0022] The present application adopts a hierarchical Swin Transformer as the core backbone network, utilizes its hierarchical window attention mechanism and relative position bias modeling capability, can maintain the overall structure of the global background field on a large scale, and at the same time, capture detailed features on a local scale, realize efficient interaction of background and observation information; the output of the model is a high-precision analysis field after assimilation, which not only retains the physical rationality of large-scale circulation, but also significantly improves the spatial resolution and structural fineness on small-scale meteorological elements.

[0023] The model is based on the Swin Transformer structure, and a three-branch structure composed of background flow, observation flow and mixed flow is constructed. The fusion module adopts a residual form to write back the increment, realizing an analysis similar to three-dimensional variation 3D-Var: ; in To analyze the state vector of the field, Let be the state vector of the background field. Let H be the observation vector, and H be the observation operator. Let H be the tangent linear operator, B be the covariance matrix of the background field error, and R be the covariance matrix of the observation error. This is typically referred to as the weight matrix of the analysis field. In deep networks, these weights and increments are learned implicitly and then applied to the background features and output as residuals.

[0024] This embodiment uses a hybrid information stream as the input tensor, which is fed into the Swin Transformer backbone network to complete end-to-end assimilation and increment generation. First defined as , These are the initial channel, latitude, and longitude, respectively. The model front-end... Block partitioning and linear embedding: The spatial dimension is divided into non-overlapping blocks of size 4×4. Each patch flattens out to 4×4×4. After the vector is obtained, it is projected onto the new channel C through linear mapping and then subjected to layer normalization to form a vector of size C. The vector.

[0025] The backbone feature extraction adopts a four-level hierarchical structure, stacking Swin Transformer network layers (SwinBlocks) level by level and performing block merging between layers. Specifically: In the first stage, the resolution is maintained at 160×160, the number of channels is C, the window side length of the window attention is set to M=8, and the number of SwinBlocks is N=2. In the second stage, every 2×2 neighborhood is spliced ​​in the channel dimension through block fusion, and reduced to 2C channels through layer normalization and linear layers, and the resolution is reduced to 80×80. This is repeated with 2 SwinBlocks. In the third stage, the resolution is further reduced to 40×40, the number of channels is expanded to 4C, and 6 SwinBlocks are stacked. In the fourth stage, a high-level feature of 20×20 and 8C channels is obtained, 2 SwinBlocks are stacked, and no merging is performed in the final stage. Each SwinBlock is implemented in two steps: "window attention + feedforward network". The window side length for window attention is set to M=8. First, multi-head self-attention (W-MSA) is performed within a local window of M×M. Then, a cyclic translation is performed with a sliding displacement of (M / 2, M / 2), and the shifted window attention (SW-MSA) is calculated within the new window. The attention mask maintains locality while enabling cross-window interaction. A relative position bias table is introduced into the attention. (h is the number of attention heads), indexed by relative displacement. Then a two-layer feedforward network (expansion ratio 4, activation function GELU) is used for nonlinear transformation, and pre-normalization and residual connection are used to ensure training stability. The implementation of block merging follows the method of "concatenate every 2x2 neighborhood in the channel dimension, and then reduce to 2xC through layer normalization + linear layer", which increases the channel capacity while reducing the sampling, ensuring the effectiveness of multi-scale hierarchical representation.

[0026] Considering the spatial sparsity and missing observations, the model injects a learnable mask vector into the missing block and masks these positions during attention calculation, so that attention relies more on the context information of the background and neighborhood to complete the completion in the missing area. To learn incrementally in the deep representation space, the assimilation fusion mechanism introduces an incremental assimilation fusion module ILFM at the end of each level output. This module receives background stream features , observed stream features and mixed stream features , first generates , two gating weights [0,1] through 1x1 convolution and Sigmoid function to characterize "background / observation trustworthiness", and then concatenates the three features to generate increments through Swin Transformer module : ; Finally, it is written back to the background branch in residual form: , and updates the observation and mixed branches for the next level of fusion.

[0027] After encoding and multi-level fusion, the decoding end uses the inverse process of block merging to upsample to 160x160 spatial resolution, and projects it to 88 physical variable channels through 1x1 convolution; pixel interpolation is used to restore the original resolution, consistent with the input background grid; (4) Training method The training strategy uses AdamW optimization, where AdamW (regular parameters =0.9, =0.999, weight decay =10 -5 ), and uses a combination of linear warm-up and cosine annealing learning rate schedule: in the first 500 steps ( =500), the learning rate is linearly increased from 10 -8 (start = 10 -8 ) to 2x10 -3 (stop = 2x10 -3 ), which is: ; After the preheating ends, cosine annealing is performed within a total iteration of 6000 steps with gradual attenuation; In terms of loss function, the analysis moment is optimized simultaneously with multi-time effect prediction, and the basic term is the zonal weighted L1 aligned with ERA5: ; Wherein, C, H and W are respectively the number of channels and the number of latitude and longitude grid points, is the true value of channel c at latitude i and longitude j, is the predicted value of channel c at latitude i and longitude j, is the weight at latitude .

[0028] (5) Effect evaluation The effect evaluation index is mainly global latitude weighted RMSE and regional RMSE. The global latitude weighted RMSE is defined as follows: ; Wherein, H and W are respectively the number of channels and the number of latitude and longitude grid points, is the true value of channel c at latitude i and longitude j, is the predicted value of channel c at latitude i and longitude j, is the weight at latitude , and the normalized difference is reported; the regional RMSE on a given latitude and longitude subdomain is defined as follows: ; Wherein, , , , indicate the range of grid points corresponding to the selected region, is the weight at latitude .

[0029] Through the above steps, the present application realizes high-precision and high-efficiency multi-source data assimilation, and provides a new scheme for numerical weather prediction and climate research.

[0030] The effectiveness of the method is verified by comparison and evaluation with ERA5 reanalysis data; the assimilation results are systematically compared and evaluated with ERA5 reanalysis data; the mean square error and bias of key meteorological variables such as humidity, temperature, wind field and potential height can be significantly reduced, and the improvement effect is maintained in multi-time forecast, which proves the stability and generalization of the proposed method; compared with traditional assimilation technologies such as three-dimensional variation 3D-Var, four-dimensional variation and ensemble Kalman filter, the present application does not rely on explicit error covariance matrix modeling, but implicitly learns background error, observation error and observation operator in an end-to-end feature space through a deep neural network, so it has obvious advantages in assimilation efficiency and adaptability.

[0031] Those skilled in the art can understand that the above description is only preferred examples of the application and is not used to limit the application, although the application is described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions recorded in the foregoing examples or replace some technical features equivalently. Any modification, equivalent replacement, etc. within the spirit and principles of the application should be included in the protection scope of the application. All technical features in the embodiments can be freely combined according to actual needs.

[0032] Finally, it should be pointed out that: the above description is only the preferred embodiment of the application and is not used to limit the application, although the application is described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions recorded in the foregoing examples or replace some technical features equivalently, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application should be included in the protection scope of the application.

Claims

1. A deep learning based observation station data assimilation method, characterized in that, Specifically comprising the following steps: Step 1, obtain the forecast background field data of the Fuxi weather model, with a spatial resolution of 0.25°, a time interval of 6 hours, and containing 88 meteorological variables; Step 2, obtain observation data, including ground meteorological observation stations, radars, sounding stations, and wind profile observation stations; Step 3, preprocess the background field data and observation data, including spatial interpolation, time alignment, standardization, and the introduction of position encoding and time encoding; Step 4, build a deep learning assimilation model based on sliding window Transformer, jointly input the background field data and observation data, and obtain the assimilation result; Step 5, evaluate the assimilation result and the fifth generation global atmospheric reanalysis data ERA5 of the European Centre for Medium-Range Weather Forecasts (ECMWF).

2. The deep learning based observation station data assimilation method according to claim 1, characterized in that, In step 1, the 88 meteorological variables include high-altitude variables on pressure levels levels = [50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000], surface variables, and sea wave variables, totaling 88 variables.

3. The method of claim 1, wherein, In step 2, the spatial distribution of observation data is shown in the background of gray terrain height, with national observation stations as blue dots, radars as cross signs, sounding stations as hollow circles, and wind profile stations as triangle symbols.

4. The method of claim 1, wherein, In step 3, spatial interpolation is used to interpolate the observation data to the background field grid, and a mask is added to the positions without matching observations. The processed observation field has the same resolution and grid as the background; For all observation data within 1 hour before and after a specific analysis time point, weighted average is performed to achieve time alignment of observation data and background field; Standardization is used to unify the distribution range of each variable in the input data; The latitude and longitude encoding in sine and cosine form and the time encoding are used as additional observation channels to maintain geographical and temporal characteristics; The background field and observation data are concatenated in the channel dimension to form a mixed information stream input.

5. The method of claim 1, wherein, In step 4, based on the Swin Transformer structure, wherein Swin Transformer is a sliding window Transformer; The three-branch structure composed of background flow, observation flow and mixed flow is constructed, and finally the residual form is used to write back the increment to realize the analysis of three-dimensional variation 3D-Var: ; wherein, is the state vector of the background field, is the state vector of the background field, is the observation vector, H is the observation operator, is the tangent linear operator of H, B is the covariance matrix of the background field error, R is the covariance matrix of the observation error; is the weight matrix of the analysis field; in the deep network, the weights and increments are learned in an implicit way and then applied to the background features and the output in the form of residuals; The mixed information stream is used as an input vector and enters the Swin Transformer backbone network to complete end-to-end assimilation and incremental generation; The backbone feature extraction adopts a four-level hierarchical structure, with Swin Transformer networks stacked at each level and block patch merging performed between levels; In the deep representation space, the increment is learned end-to-end and written back to the background branch in residual form, while updating the observation and mixed branches for further fusion at the next level; After completing encoding and multi-level fusion, the decoding end is upsampled back to a spatial resolution of 160x160, and projected to 88 physical variable channels through 1x1 convolution; pixel interpolation is used to restore the original resolution, consistent with the input background grid; The training strategy employs AdamW optimization with AdamW regular parameters = 0.9, = 0.999, weight decay = 10 -5 ; and a combined learning rate schedule of linear warmup and cosine annealing: linearly increase the learning rate from 10 -8 to 2x10 -3 over the first 500 steps, in the form of: ； wherein, = 500 for the first 500 steps; start = 10 -8 = starting learning rate; stop = 2 x 10 -3 = ending learning rate; After preheating, cosine annealing is used to gradually decay within a total of 6000 iterations; In terms of loss function, the analysis time is optimized simultaneously with multi-time forecast, and the basic term is the zonal weighted L1 aligned with ERA5: ； where C, H, W are the number of channels and the number of latitude and longitude grid points, respectively, is the true value for channel c at latitude i, longitude j, is the predicted value for channel c at latitude i, longitude j, is the weight at latitude .

6. The method of claim 1, wherein, In step 5, the assimilation effect is evaluated by the root mean square error RMSE index; The effect evaluation index is mainly global latitude weighted RMSE and regional RMSE; the global latitude weighted RMSE is defined as: ； where H, W are the number of channels and the number of latitude and longitude grid points, respectively, is the true value for channel c at latitude i, longitude j, is the predicted value for channel c at latitude i, longitude j, is the weight at latitude and reports the normalized difference ; the area RMSE over a given latitude-longitude subdomain is defined as follows: ； wherein, , , , represents the range of the number of grid points corresponding to the selected region, is the weight of the latitude .

Images