An urban low-altitude high-vertical-resolution element artificial intelligence correction method based on high tower observation

Abstract

The application discloses a kind of urban low altitude high vertical resolution element artificial intelligence revision method based on high tower observation, obtains meteorological observation tower meteorological element data, reanalysis material grid data, underlying surface spatial characteristic parameter, and is preprocessed;Physical constraint multivariate joint revision deep learning model is constructed;Design composite loss function to train model, including data-driven loss term for fitting observation true value and physical constraint loss term for constraining the physical relationship between output variables;The point revision ability of the location where meteorological observation tower is located is generalized to the predetermined urban area;Spatial consistency post-processing and physical consistency post-processing are carried out on the grid-by-grid inference result, to generate the multivariate coupled consistent low-altitude meteorological parameter field covering the predetermined urban area, with high vertical resolution and high time resolution.The present application can directly serve urban low-altitude flight support, city meteorological service and other real-time business scenarios.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to an artificial intelligence correction method for low-altitude, high-vertical-resolution elements in urban areas based on high-tower observations. Background Technology

[0002] The main ways to obtain low-altitude vertical meteorological parameters are: 1. Ground meteorological observation stations and conventional radiosonde stations: Conventional ground stations only provide meteorological observations at a single altitude (usually 10m or 2m) and cannot obtain gradient information in the vertical direction; although conventional radiosonde stations can provide vertical profiles, they are only released twice a day, and the minimum effective altitude is usually more than several hundred meters. The time and spatial resolution cannot meet the minute-level requirements of low-altitude flight support.

[0003] 2. Remote sensing methods: These include Doppler LiDAR, microwave radiometers, and wind profiler radar. These devices can acquire near-surface vertical profile information, but they suffer from high equipment costs, high maintenance costs, and limited spatial coverage. LiDAR data quality deteriorates significantly in low-visibility weather conditions such as rain and fog, while wind profiler radar typically has a minimum detection altitude of at least 100m, leaving a blind spot in the crucial ground-to-100m layer.

[0004] 3. Numerical Weather Prediction Models and Reanalysis Data: ERA5 has a horizontal resolution of approximately 0.25° × 0.25° (about 3 km), and its vertical interlayer spacing in the lower atmosphere is also relatively large (the near-surface layer thickness is typically between 20-60 m), resulting in a severe deficiency in characterizing the fine vertical structure below the urban canopy. More importantly, reanalysis data such as ERA5 are generated based on a global assimilation system, and the surface parameterization schemes and boundary layer parameterization schemes they use exhibit systematic biases when dealing with complex urban underlying surfaces.

[0005] 4. Meteorological Observation Towers: High-rise meteorological observation towers are typically equipped with multiple layers (e.g., 8 or more) of precision meteorological sensors, providing continuous, high-temporal-resolution (e.g., 10 minutes or even 1 minute) vertical profile data of temperature, air pressure, wind speed, wind direction, humidity, etc., from the ground to the top of the tower. This is the most reliable means of obtaining true near-surface vertical meteorological structure. However, observation towers are essentially single-point observation devices. The high-precision data they acquire only represents the meteorological conditions of the tower's location and a very limited surrounding area (on the order of hundreds of meters), and cannot be directly extrapolated to a larger spatial range.

[0006] In summary, the insufficient spatial coverage of high-precision observation data (observation towers) fundamentally contradicts the insufficient vertical resolution and accuracy of wide-coverage reanalysis data (ERA5, etc.) in the low-altitude urban area. How to fully integrate the advantages of both types of data sources to generate a low-altitude meteorological parameter field that possesses both high vertical resolution and coverage of a certain area within the urban area has become a pressing technical challenge. In the field of deep learning, recent advancements such as spatiotemporal attention mechanisms, Physics-Informed Neural Networks (PINN), and multi-task learning have provided new insights into solving this problem. However, existing research largely remains at the level of univariate post-processing or simple pattern output statistical downscaling (MOS), lacking a complete technical solution for multivariate joint correction, synchronous reconstruction of multi-layered vertical structures, and generalization to surrounding areas using observation towers as anchor points. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to provide an artificial intelligence correction method for urban low-altitude high vertical resolution elements based on high tower observations, which solves the technical problems of insufficient vertical resolution of urban low-altitude data, significant system bias, weak ability to capture extreme meteorological parameters, and the limitation of limited spatial coverage of high tower observation data in the existing technology.

[0008] Technical Solution: The present invention provides an artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations, comprising the following steps:

[0009] Step 1: Acquire meteorological element data, reanalysis data grid points, and underlying surface spatial characteristic parameters from meteorological observation towers at multiple vertical levels, and perform preprocessing.

[0010] Step 2: Construct a multivariate joint correction deep learning model for physical constraints, including a vertical feature encoder, a time-dependent modeler, a spatiotemporal cross-attention fusion module, an underlying surface condition injection module, a multi-task output head, and a physical information constraint layer;

[0011] Step 3: Design a composite loss function to train the model, including a data-driven loss term for fitting the observed true values ​​and a physical constraint loss term for constraining the physical relationships between output variables;

[0012] Step 4: Generalize the location correction capability of the meteorological observation tower to the predetermined urban area;

[0013] Step 5: Perform spatial consistency post-processing and physical consistency post-processing on the grid-by-grid inference results to generate a low-altitude meteorological parameter field with multivariate coupling consistency that covers the predetermined urban area and has high vertical and temporal resolution.

[0014] Furthermore, in step 1, the preprocessing is as follows: quality control is performed on the meteorological element data; the reanalysis data and the meteorological element data are strictly paired in time and space through spatiotemporal interpolation to obtain the virtual profile of the reanalysis data; the spatial characteristic parameters of the underlying surface of the predetermined urban area around the meteorological observation tower are obtained and constructed as the underlying surface characteristic tensor.

[0015] Furthermore, in step 2, the vertical feature encoder is used to extract local gradient features and long-range coupling features of the input data in the vertical dimension, including a one-dimensional convolutional stack and a vertical attention module; the temporal dependency modeler is used to capture the changing features and temporal dependencies of multiple time scales along the time dimension; the spatiotemporal cross-attention fusion module is used to realize the deep interaction and adaptive fusion of vertical dimension information and temporal dimension information; the underlying surface condition injection module is used to inject the spatial features of the underlying surface as conditional information into the model to modulate the feature expression of the model; wherein, the one-dimensional convolutional stack extracts the local vertical gradient features between adjacent layers in sequence along the vertical dimension through multiple one-dimensional convolutional layers; the vertical attention module, based on the output of the one-dimensional convolutional stack, uses a multi-head self-attention mechanism to calculate the association weights between each layer along the vertical dimension, and introduces a learnable vertical position bias matrix to encode the relative height information between each layer, thereby capturing the long-range coupling relationship between non-adjacent vertical layers.

[0016] Furthermore, the time-dependent modeler includes a causal dilated convolutional layer and a temporal self-attention module. The causal dilated convolutional layer uses causal convolution with different dilation rates to extract multi-timescale change features along the time dimension, and ensures that only information from the current and past moments is used to conform to physical causality. Based on the output of the causal dilated convolutional layer, the temporal self-attention module uses a multi-head self-attention mechanism with causal mask to calculate the dependency weights between each time step along the time dimension, and integrates auxiliary time features such as time period encoding and atmospheric stability parameters to enhance the model's ability to perceive diurnal cycles, seasonal changes, and atmospheric stability conditions.

[0017] The spatiotemporal cross-attention fusion module expands the outputs of the vertical feature encoder and the temporal dependency modeler into two-dimensional sequences, and then performs bidirectional cross-attention calculation: using vertical features as queries and temporal features as keys and values, a temporally enhanced vertical representation is calculated, and simultaneously using temporal features as queries and vertical features as keys and values, a vertically enhanced temporal representation is calculated; then, an adaptive weighted fusion of the two sets of enhanced representations is performed through a gating mechanism; wherein, the gating mechanism dynamically controls the contribution ratio of the two sets of features in the fusion through a learnable gating weight matrix.

[0018] Furthermore, in step 2, the underlying surface condition injection module maps the underlying surface feature vector into a condition vector through a fully connected network. Then, using a feature modulation mechanism, the condition vector is transformed by an affine transformation to generate scaling and offset factors. The scaling and offset factors are used to scale and offset the output features of the spatiotemporal cross-attention fusion module element by element. This allows for adaptive adjustment of the correction strategy and magnitude based on the difference in underlying surface conditions between the target grid point and the observation tower, providing a key capability for spatial generalization from a single point to a region.

[0019] Furthermore, in step 2, the multi-task output head includes a main task head and an auxiliary task head. The main task head is used to predict the correction bias of multiple meteorological variables at each vertical level, and adds the prediction bias to the input reanalysis data through residual connection to obtain the final correction value. The auxiliary task head includes a wind shear prediction branch and an extreme wind speed probability prediction branch. The wind shear prediction branch is used to directly predict the vertical wind shear value between adjacent vertical levels, and the extreme wind speed probability prediction branch is used to predict the probability that the wind speed at each level exceeds a predetermined threshold. The auxiliary task head and the main task head form a multi-task collaborative supervision mechanism to strengthen the model's learning of the vertical gradient of the wind field and the distribution characteristics of extreme events.

[0020] Furthermore, in step 2, the physical information constraint layer applies hard physical constraints to the output results of the multi-task output head during the model inference stage, including: static equilibrium constraint verification, which jointly fine-tunes the corrected temperature and pressure profiles to satisfy the static equilibrium relationship; wind shear consistency constraint verification, which compares and adjusts the consistency between the wind shear value implied by the wind field output by the main task head and the wind shear value directly predicted by the auxiliary task head; and variable physical range constraint verification, which applies physical feasibility truncation to each output variable to ensure that it is within a reasonable range.

[0021] Furthermore, in step 3, the data-driven loss term of the composite loss function includes the main task weighted multivariate correction loss, wind shear prediction loss, and extreme wind speed probability prediction loss. The main task weighted multivariate correction loss adopts an extreme value adaptive weighted loss function, which adaptively adjusts the sample weights according to the degree of deviation between the observed values ​​and the mean, so that extreme events with greater deviations receive higher loss weights. An additional weighting term is added to the wind speed variable under strong wind conditions to enhance the model's learning and optimization capabilities for extreme weather events. The physical constraint loss term of the composite loss function includes static equilibrium constraint loss, thermodynamic consistency constraint loss, and momentum conservation approximation constraint loss. The static equilibrium constraint loss is used to constrain the corrected temperature and pressure profiles to satisfy the static equations. The thermodynamic consistency constraint loss is used to constrain the corrected temperature, humidity, and pressure to satisfy thermodynamic consistency relationships. The momentum conservation approximation constraint loss is used to constrain the corrected wind shear value to not exceed a reasonable range determined based on boundary layer theory. During the model training phase, the physical constraint loss term is added to the total loss function as a soft constraint regularization term in a weighted manner.

[0022] Furthermore, during model training, a course learning strategy is adopted. In the early stage of training, the data-driven loss term is mainly optimized, and in the later stage of training, the weight of the physical constraint loss term is gradually increased to avoid the conflict between physical constraints and data fitting in the early stage of training. The weight coefficients of each loss term are adaptively adjusted during training as learnable parameters through an uncertainty weighting strategy. An early stopping strategy based on validation set performance is adopted to avoid overfitting.

[0023] Further, step 4 is as follows: A target spatial grid is established within the predetermined urban area. For each target grid point, corresponding reanalysis input data and underlying surface feature vectors are prepared. The trained model is used to perform grid-by-grid inference to obtain the multivariate meteorological element correction bias for each grid point. For the correction result of each target grid point, a distance-based attenuation weight is introduced. The attenuation weight decreases as the distance between the target grid point and the meteorological observation tower increases, so that the correction bias of grid points farther from the tower decreases towards zero, reflecting the objective law that the model's generalization ability decreases with increasing distance. A confidence level label is attached to the output of each target grid point. The confidence level is determined comprehensively by the distance between the grid point and the tower point, the underlying surface similarity, and the physical constraint test results.

[0024] Furthermore, the spatial consistency post-processing uses Gaussian filtering to spatially smooth the correction results of adjacent grid points, in order to eliminate non-physical spatial jumps caused by abrupt changes in the underlying surface parameters between grid points, and the kernel radius of the Gaussian filter is adaptively adjusted according to the local underlying surface uniformity.

[0025] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: The present invention takes the high-precision multi-layer observation of meteorological towers as the anchor point, and through the collaboration of deep learning models and underlying surface condition injection modules (FiLM), it effectively extends the high vertical resolution correction capability of tower points to the urban area (e.g., radius 10 km) for the first time, solving the fundamental contradiction between the weak precision of high tower observation points and the wide coverage and coarse quality of reanalysis data.

[0026] This invention uses high-density observations from eight tower levels (10m to 250m) as a benchmark. Through a vertical feature encoder and a spatiotemporal cross-attention mechanism, it accurately reconstructs the ERA5 low-altitude rough profile to a high vertical resolution consistent with the tower points. Furthermore, it achieves superior accuracy compared to traditional downscaling methods in multi-variable (temperature, pressure, humidity, wind) joint correction. This invention embeds physical constraints such as static equilibrium, thermodynamic consistency, and momentum conservation into the model architecture and loss function, effectively suppressing non-physical distortions that may arise from purely data-driven models and ensuring that the corrected meteorological field satisfies fundamental physical laws in the vertical direction. Through an extreme value adaptive weighted loss function and a multi-task auxiliary head, the model actively focuses on extreme samples with large deviations from the mean during training, demonstrating a significantly better ability to capture dangerous meteorological conditions such as low-altitude wind shear and strong winds than conventional models trained with mean square error loss. This invention introduces distance-based attenuation weights and confidence level labels to reasonably attenuate corrections for grid points far from the tower and label their reliability. Combined with adaptive Gaussian filtering to eliminate spatial jumps, the generated low-altitude meteorological field is spatially continuous and smooth. Furthermore, each grid point result is accompanied by a confidence level indicator, facilitating risk assessment in practical applications. This invention can directly serve real-time operational scenarios such as urban low-altitude flight support and urban meteorological services. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention;

[0028] Figure 2 This is the PhySTAM-Net model structure of the present invention. Detailed Implementation

[0029] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0030] like Figure 1 As shown, this embodiment of the invention provides an artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations, comprising the following steps:

[0031] This embodiment uses the application of a meteorological observation tower in low-altitude meteorological support in a city along the Yangtze River as an example for illustration.

[0032] A meteorological observation tower in a certain city (hereinafter referred to as the meteorological observation tower) is approximately 260m high. The tower is equipped with eight layers of precision meteorological sensors (10m, 30m, 50m, 80m, 120m, 160m, 200m, and 250m), with a time resolution of 10 minutes. It has been operating stably for over 17 years (2008 to 2025). The underlying surface of this area is a complex low-altitude environment typical of urban areas, characterized by alternating water, urban, and mountainous terrain. Therefore, the need for accurate forecasting of vertical wind shear and cross-river gusts is extremely urgent.

[0033] Step S1: Multi-source heterogeneous data acquisition and collaborative preprocessing.

[0034] Step S1.1: Data acquisition and quality control for multi-level observations on the high tower.

[0035] Data was collected from eight vertical levels of the meteorological observation tower (each level being 10, 30, 50, 80, 120, 160, 200, and 250 meters high). The data included temperature data for each level. air pressure relative humidity Horizontal wind speed Horizontal wind direction Vertical wind speed And calculate the vertical wind shear parameters based on the wind vector difference between adjacent layers. The time resolution of the data is no less than 10 minutes, and the calculation method is as follows:

[0036] ;

[0037] In the formula For the first Horizontal wind vector of the layer (by the layer) , (converted to latitudinal and longitudinal components according to meteorological conventions) The geometric height difference between two adjacent floors. It is a vector model. The time resolution of the data is no less than 10 minutes.

[0038] Three levels of quality control were performed on the raw data:

[0039] Level 1: Physical threshold test to remove outliers that exceed the climatological extreme range;

[0040] Level 2: Time consistency check, using a sliding window mid-range filter to remove isolated jump points;

[0041] Level 3: Vertical consistency check, marking suspicious data based on the reasonable gradient range of physical quantities at adjacent levels.

[0042] For missing data after quality control, an adaptive weighted interpolation method based on the time dimension is used for imputation:

[0043] Short-term missing ( For missing data of 30 minutes to 6 hours, cubic spline interpolation was used; for missing data of medium duration (30 minutes to 6 hours), Gaussian process regression interpolation of the same time series was used; for missing data of long duration (minutes ... cubic spline interpolation was used. If the data is less than 1 hour, it is marked as unavailable. In this embodiment, the observation data comes from the meteorological observation tower. 15 years of data from January 1, 2008 to December 31, 2022 are used as the training set (approximately 788,000 records), and 2 years of data from January 1, 2023 to December 31, 2024 are used as the independent validation set (approximately 105,000 records).

[0044] Step S1.2: Extraction, spatiotemporal alignment, and vertical interpolation of ERA5 reanalysis data. In this embodiment, hourly ERA5 pressure layer and model layer data for the region from 2008 to 2024 are downloaded from the ECMWFClimate Data Store. After spatial interpolation to the precise coordinates of the meteorological observation towers and vertical PCHIP interpolation to eight observation heights, an ERA5 virtual profile dataset with strict spatiotemporal matching to the tower data is established.

[0045] Extract gridded data from ERA5 reanalysis data around the location of the tower, with variables including geopotential height, temperature, specific humidity, and so on. Wind component, vertical velocity, and air pressure; where vertical velocity is determined by ω in the pressure coordinate system. The relative humidity, converted to geometric vertical velocity and obtained from temperature, air pressure, and specific humidity diagnostics, was used as the residual benchmark for relative humidity correction. The hourly ERA5 data was aligned to the observation tower's 10-minute time resolution via time-linear interpolation.

[0046] Spatial dimension alignment: A method combining inverse distance weighting (IDW) and bilinear interpolation is used to interpolate the data of four grid points around ERA5 to the precise latitude and longitude of the tower, thus obtaining the "virtual profile" of ERA5 at the tower point.

[0047] Vertical interpolation: Converting ERA5 model layer data into geometric height coordinates based on static equations.

[0048] in Air density (unit: kg / m³). The acceleration due to gravity is taken as 9.81 m / s²; P is the air pressure (unit: Pa), and z is the geometric height. Calculated from the ideal gas law:

[0049] ;

[0050] in The gas constant for dry air is taken as 287.05 J / (kg·K). This is a virtual temperature, measured in Kelvin (K).

[0051] Then, the conformal segmented cubic Hermite interpolation (PCHIP) method is used to accurately interpolate the irregular vertical layer data of ERA5 to eight heights consistent with the sensors of each layer of the observation tower, ensuring that the two sets of data are strictly matched in space.

[0052] Step S1.3: Multivariate joint normalization.

[0053] Calculate the mean of each paired meteorological variable over the 15-year training period. and standard deviation Z-score normalization is used:

[0054] ;

[0055] For variables that are sensitive to extreme values, such as wind speed and wind shear, a robust normalization method (based on the median and interquartile range, IQR) is used to reduce the sensitivity of extreme values ​​to the normalization parameters.

[0056] Step S1.4: Construction of Urban Underlying Surface Feature Parameter Set. This embodiment uses high-resolution remote sensing imagery and a 3D urban building database of a certain city from 2020 to generate an underlying surface feature raster dataset within a 10km radius centered on a meteorological observation tower. Typical underlying surface types in this area include: the Yangtze River surface (z0≈0.001m, FAR=0), low-rise urban areas along the river (z0≈0.5m, FAR≈1.2), the core urban area (z0≈1.2m, FAR≈2.5), and hills (with high vegetation coverage).

[0057] High-resolution underlying surface parameters were collected within a 10 km radius of the tower, including surface aerodynamic roughness length. Zero plane displacement height Urban building plot ratio Average building height Vegetation coverage impermeable surface ratio Topographic elevation .

[0058] The above parameters are rasterized to a uniform spatial resolution of 1000m×1000m to form the underlying surface feature tensor:

[0059] ;

[0060] in For grid dimensions, This represents the number of channels for the underlying surface parameters. The feature vector of the underlying surface at the tower's location is extracted. As auxiliary input during the training phase.

[0061] Step S2: Construct a physical constraint multivariate joint correction deep learning model.

[0062] This invention designs a deep learning architecture called PhySTAM-Net (Physics-constrained Spatio-TemporalAttention Multi-task Network). Figure 2 As shown in the figure, this architecture integrates spatiotemporal attention mechanism, physical information constraint and multi-task learning framework to realize physical coupling constraint and joint correction between multiple variables.

[0063] Step S2.1: Input feature construction.

[0064] For each training sample, with a time window length of... ( This represents the past 24 10-minute time steps, or 4 hours), constructing the input tensor. The model input is defined as triples. :

[0065] (1) ERA5 virtual profile sequence:

[0066] ;

[0067] in Number of vertical layers ( ), For ERA5 variables (including temperature, air pressure, specific humidity, etc.), Wind weight, (Six variables in total, including wind component and vertical velocity).

[0068] (2) Auxiliary time characteristics:

[0069] This includes time period coding (hourly sine / cosine coding, day number sine / cosine coding, monthly sine / cosine coding), solar zenith angle, and atmospheric stability parameters (Richardson number calculated based on ERA5).

[0070] (3) Feature vector of the underlying surface:

[0071] During the training phase, the values ​​are fixed for the tower positions, and during the generalization phase, they are the underlying surface features of the target grid points.

[0072] Step S2.2: Vertical Profile Encoder (VPE).

[0073] The vertical feature encoder is designed to extract features from the vertical profile at each time step. VPE includes:

[0074] (a) One-dimensional convolutional stack (1D-CNN Stack): Multiple one-dimensional convolutional layers are set along the vertical dimension (kernel size 3, stride 1, padding 1), extracting local vertical gradient features between adjacent layers sequentially. Each one-dimensional convolution is followed by batch normalization (BatchNorm) and the GELU activation function. The number of channels in the convolutional stack expands layer by layer ( The output feature map maintains the same spatial dimension as the input. layer).

[0075] (b) Vertical Attention Module (VAM): Following the one-dimensional convolutional stack, a multi-head self-attention mechanism is introduced to calculate the association weights between layers along the vertical dimension, enabling the model to capture long-range vertical coupling relationships between non-adjacent layers. The attention calculation formula is:

[0076] ;

[0077] in These are query, key, and value matrices, respectively. For the attention head dimension, A learnable vertical positional bias matrix is ​​used to encode the relative height information between layers. In this invention, the model feature dimension is d. model =256 (consistent with the number of channels in the last layer of a one-dimensional convolutional stack), the number of heads in a multi-head self-attention system is set to 8, therefore the dimension d of each attention head is... k = d model / 8 = 32.

[0078] VPE runs independently at each time step of the input sequence and outputs a vertical feature sequence. .

[0079] The input consists of an ERA5 virtual profile sequence spanning L=24 time steps (covering nearly 4 hours, with one time step every 10 minutes), where each time step contains K=8 vertical layers and each layer has a V... e =6 meteorological variables (temperature, air pressure, specific humidity, U-wind, V-wind, vertical velocity), i.e., a three-dimensional tensor X with shape 24×8×6. ERA5VPE processes the 8 profiles at each time step independently. Output: Encodes the 6 variables of each layer as d. model =256-dimensional feature vectors, resulting in a vertical feature sequence Z of shape 24×8×256. vert (recorded as) ).

[0080] Step S2.3: Temporal Dependency Modeler (TDM).

[0081] After obtaining the vertical feature sequences, temporal dependencies are independently modeled along the time dimension for each vertical layer. The TDM consists of the following components:

[0082] (c) Causal Dilated Convolution: Multiple layers of causal dilated convolution (with dilation rates of 1, 2, 4, and 8) are used along the time dimension to expand the receptive field without increasing the number of parameters, capturing the changing features at multiple time scales (10-minute, 30-minute, 1-hour, and 2-hour trends). Causal convolution ensures that only information from the current and past moments is used, which conforms to the physical causality law.

[0083] (d) Temporal Self-Attention Module (TSAM): Above the causal convolution output, a multi-head self-attention mechanism is introduced to calculate the dependency weights between each time step along the temporal dimension, and learnable temporal position encoding is used. A causal mask is applied to the attention matrix to ensure the causal direction of the information flow (8 heads in the multi-head self-attention mechanism, feature dimension d). model =256, consistent with VPE).

[0084] (e) Auxiliary feature fusion: Integrating auxiliary temporal features Mapped to via fully connected layer After that, the time attention output is injected element by element into the output, enabling the model to perceive the day-night cycle, seasonal changes and atmospheric stability conditions.

[0085] TDM runs independently for each of the eight vertical layers (with shared weights), outputting time feature sequences. .

[0086] Input: The vertical feature sequence Z obtained in the previous step vert (shape 24×8×256), and auxiliary time feature X aux (Shape 24×8, i.e., 24 time steps, each time step D) aux=8-dimensional auxiliary features: 6 dimensions of sine and cosine encoding for hourly / daily / monthly data, 1 dimension of solar zenith angle, and 1 dimension of atmospheric stability (Richardson number). Output: The time feature sequence Z after modeling the temporal dependencies along the time direction at each vertical layer. temp The shape remains 24×8×256 (denoted as...). ).

[0087] Step S2.4: Spatio-Temporal Cross-Attention Fusion (STCAF) module.

[0088] To enable deep interaction between vertical and temporal information, a spatiotemporal cross-attention fusion module is designed. This module will... and Expanded into A two-dimensional sequence is then processed, and bidirectional cross-attention is performed:

[0089] (f) Cross-attention is performed using vertical features as the Query and temporal features as the Key and Value to obtain a temporally enhanced vertical representation. ;

[0090] (g) Perform cross-attention calculation with time features as the Query and vertical features as the Key and Value to obtain a vertically enhanced time representation. ;

[0091] (h) Adaptively weightedly fuse the two sets of enhanced representations using a gated fusion mechanism:

[0092] ;

[0093] in For the Sigmoid function, For learnable weight matrix, This indicates a splicing operation. This is an element-wise product.

[0094] Input: Vertical feature Z vert With time feature Z temp (Both are 24×8×256). The module first flattens each into a 192×256 two-dimensional sequence (192=24×8, that is, merging 24 time steps and 8 vertical layers into 192 spatiotemporal locations, each location being 256-dimensional), and then performs bidirectional cross-attention. Output: The fused feature Z. fused The flattened shape is 192×256, and it is restored to 24×8×256 (denoted as...). ).

[0095] Step S2.5: Surface Feature Conditioning Module (SFCM).

[0096] The feature vector of the underlying surface The conditional vector is mapped through a two-layer fully connected network (including GELU activation). Then, it is injected into the fused features through a feature-wise linear modulation (FiLM) mechanism:

[0097] ;

[0098] in and For the condition vector The scaling and offset factors are generated through affine transformation. This mechanism enables the model to adaptively adjust feature representations according to different surface conditions, providing crucial capabilities for subsequent spatial generalization.

[0099] Input: Fusion feature Z fused (24×8×256), and the feature vector u0 (C) of the underlying surface at that point. u =7 dimensions (representing the dimension of the underlying surface feature vector), corresponding to 7 types of parameters: roughness length, zero-plane displacement height, building volume ratio, average building height, vegetation cover, impermeable surface ratio, and terrain elevation; u0 is first mapped to d through two fully connected layers. model =256-dimensional conditional vector c0 output: conditional modulated features Z after scaling and translating (FiLM) each feature channel using c. cond The shape remains 24×8×256.

[0100] Step S2.6: Multi-task Output Heads.

[0101] Extract the features from the last time step of the fused feature sequence. The data is sent to the multitasking output header. The multitasking output header includes:

[0102] (a1) Main task header: Multivariable profile correction: by It consists of multiple parallel multilayer perceptrons (MLPs) (one for each layer), with each MLP outputting the correction bias of that layer. (i.e., the predicted value of the difference between the true value and ERA5), correction variables include: temperature deviation. air pressure deviation relative humidity deviation , Wind deviation , Wind deviation Vertical wind speed deviation ;common Each output quantity. The final correction result is obtained through residual connections:

[0103] ;

[0104] (b1) Auxiliary Task 1: Direct Prediction of Vertical Wind Shear: Based on fused features, an independent MLP branch is set up to directly predict the vertical wind shear values ​​between adjacent layers. ,common One output. This branch provides complementary supervision to the main task head, enhancing the model's sensitivity to the vertical gradient of the wind field.

[0105] (c1) Auxiliary Task 2: Extreme Wind Speed ​​Probability Prediction: A separate MLP branch is set up to output the probability of wind speed exceeding a specific threshold (e.g., 10 m / s, 15 m / s, 20 m / s) for each layer, and the output is activated by Sigmoid. This branch enhances the model's ability to learn the distribution characteristics of extreme wind speed events during training.

[0106] Input: Take the feature Z from the last time step of the conditional modulation feature sequence. cond (L), with a shape of 8×256 (i.e., 8 vertical layers, each with 256 dimensions). Outputs three paths:

[0107] Main task: Correction deviations for 6 variables per floor (temperature, air pressure, relative humidity, U-wind, V-wind, vertical wind speed), totaling 8 × 6 = 48 values ​​(denoted as...). );

[0108] Auxiliary head 1: Vertical wind shear between adjacent layers, totaling K-1=7 values ​​(denoted as...) );

[0109] Second auxiliary factor: the probability that the wind speed at each layer exceeds J = 3 threshold values ​​(10, 15, 20 m / s), for a total of 8 × 3 = 24 values ​​(denoted as J). ).

[0110] Step S2.7: Physical Information Constraint Layer (PICL).

[0111] After the multi-task output head, a physical information constraint layer is set up to apply soft and hard physical constraints to the model output:

[0112] (a2) Soft constraint: Physical consistency regularization term: included in the loss function for training optimization (see step S3 for details).

[0113] (b2) Hard constraints: Post-processing physical verification, which enforces the following physical consistency adjustments on the output during the model inference stage:

[0114] Hard constraint B1: Static equilibrium constraint. Corrected temperature output from the main task head. and air pressure The profile is examined layer by layer to check whether its difference form satisfies the static equation. From the ideal gas law:

[0115] ;

[0116] Calculate the density of each layer ( (This is a false temperature, diagnosed by temperature and specific humidity), the test residual is:

[0117] ;

[0118] If the root mean square of the absolute value of the residuals in each layer exceeds the threshold (In this embodiment, 0.5 Pa / m is used), then least squares joint fine-tuning is performed with the main task correction value as the initial value, and the objective function is:

[0119] ;

[0120] in , To fine-tune the correction value for the k-th layer, , The standard deviation of temperature and pressure before normalization (used to balance dimensions) for training the layer. The constraint weight is set to 100 in this embodiment. This constrained least squares problem is transformed into a system of linear equations about ΔT and ΔP. These equations are solved in closed-form for each profile and then superimposed onto the main task output to obtain the fine-tuned temperature and pressure profiles. The fine-tuning increment is limited to no more than 15% of the absolute value of the original correction deviation to avoid over-correction that could disrupt the data fitting terms.

[0121] After this fine-tuning, the consistency of the wind shear auxiliary branch is recalculated (see hard constraint B2); relative humidity is still diagnosed by thermodynamic consistency but does not participate in this backpropagation update.

[0122] Hard constraint B2: Wind shear consistency constraint. Calculated from the output of the main task header. The vertical wind shear value implied by the wind component is compared with the wind shear value directly predicted by the auxiliary task head. If the difference exceeds a threshold, the weighted average of the two is taken as the final wind shear output, and the wind component of the main task is adjusted in reverse to maintain consistency: the magnitude of the wind vector difference between two adjacent layers is scaled proportionally according to the final wind shear value, while keeping its direction unchanged, and adjusted layer by layer from bottom to top.

[0123] Hard constraint B3: Physical range constraint for variables. Imposes physical feasibility cutoffs on all output variables: relative humidity is limited to 0% to 100%, and air pressure is within a reasonable range.

[0124] Input: The 8-layer profile after the main task correction (8×6) Interlayer wind shear with auxiliary head 1 (7 values). This layer only performs physical verification and fine-tuning on the values, without changing the shape. Output: Consistently corrected profile after adjustments based on three types of constraints: static equilibrium constraints, wind shear consistency constraints, and variable physical range constraints (such as relative humidity limited to 0~100%, air pressure falling within a reasonable range). The shape remains 8×6 (denoted as ). ).

[0125] Step S3: Design and training optimization of physical constraint composite loss function.

[0126] Step S3.1: Overall architecture of the composite loss function.

[0127] The composite loss function designed in this invention consists of two main parts: a data-driven loss term and a physical constraint loss term. Its overall form is as follows:

[0128] ;

[0129] in to The hyperparameter weight coefficients are learned adaptively through an uncertainty-weighted strategy.

[0130] Step S3.2: Weighted Multivariate Correction of Loss for the Main Task :

[0131] ;

[0132] in For vertical layer indexing, For meteorological variable index, Assign importance weights to each variable (set higher weights to wind speed variables). Assign weights to each vertical layer (with higher weights to the lower altitude layer, as it is more critical to flight safety). This represents the number of samples.

[0133] Step S3.3: Adaptive Weighted Huber Loss Function for Extreme Values :

[0134] Huber loss Defined as:

[0135] ;

[0136] in For predicted values, Let be the observed true value. To enhance attention to extreme observations, an observation correlation weighting function is defined:

[0137] ;

[0138] in To find the mean and standard deviation of this variable in the training set, The trigger threshold (default value is 1.5). For weighted intensity coefficients, Let be the weighted curvature parameter. Therefore, the extreme value adaptive weighted Huber loss is:

[0139] ;

[0140] This design ensures that when the observed value deviates from the mean by more than [a certain amount]... When the value reaches one standard deviation (i.e., enters the extreme value range), the loss weights increase exponentially, forcing the model to pay more attention to extreme events during training.

[0141] For wind speed variables, The superimposed weights are based on gusts (or high percentiles):

[0142] ;

[0143] in To train the 95th percentile of wind speed, For additional weighting coefficients, This is an indicator function.

[0144] Step S3.4: Wind Shear Prediction Loss :

[0145] ;

[0146] in The first calculation based on observation data Layer and First True value of vertical wind shear between layers Use the same extreme value weighting function as in step S3.3 (with added weights for strong wind shear events).

[0147] Step S3.5: Probabilistic prediction loss of extreme wind speeds :

[0148] ;

[0149] in The number of thresholds, For the first Does the wind speed on the first floor exceed the first floor? Binary labels with threshold values, To predict probabilities for the model, The positive class weighting coefficient (determined by the reciprocal of the positive and negative sample ratio) is used to alleviate the class imbalance problem in extreme events.

[0150] Step S3.6: Physical constraint loss .

[0151] The physical constraint loss includes the following three sub-items:

[0152] (a3) Static equilibrium constraint loss:

[0153] :

[0154] in For the first After the layer was corrected, the air pressure was... For the first The corrected air density is calculated using the ideal gas law based on the corrected temperature, pressure, and humidity. This is the acceleration due to gravity.

[0155] (b3) Loss of thermodynamic consistency constraint:

[0156] This requirement mandates that the corrected temperature, humidity, and air pressure satisfy the Clausius-Clapeyron relationship. The relationship between relative humidity and specific humidity is as follows:

[0157] ;

[0158] From this, the actual water vapor pressure can be obtained:

[0159] ;

[0160] Relative humidity is defined as ,in The saturated vapor pressure (which can be calculated using the Magnus formula):

[0161] In the formula, T is the temperature in Celsius (°C), and e s P is the saturated vapor pressure (hPa), P is the air pressure (hPa), and q is the dimensionless ratio.

[0162] Therefore, the thermodynamic consistency constraint loss is:

[0163] ;

[0164] In the formula, K is the number of vertical layers, q k The specific humidity of the k-th layer ERA5 obtained by interpolation in step S1.2 (this invention does not correct for specific humidity; this constraint relative humidity correction is coordinated with temperature and air pressure corrections under the premise that the water vapor content remains unchanged). For the first After layer correction, air pressure and e s All are expressed in hPa. It is calculated as a ratio of 0 to 1. The corrected temperature for the k-th layer; This is the saturated vapor pressure calculated from the corrected temperature.

[0165] (c3) Loss due to approximate constraint of momentum conservation:

[0166] Define the wind vector difference between adjacent layers Interlayer wind shear amplitude Let the maximum reasonable shear threshold for each layer be... This threshold is given by the Monin-Obukhov similarity theory under the near-surface stability function: when the atmospheric stability parameter Richardson number Ri ≥ 0 (stable / neutral stratification), We take the empirical upper limit of the boundary layer (0.15s for the near-surface layer in this embodiment). -1 When Ri < 0 (unstable stratification), the time eases linearly to 0.25 s as |Ri| increases. -1 The momentum approximation constraint loss is then:

[0167] ;

[0168] In other words, a penalty gradient is generated only when the inter-layer wind shear output by the model exceeds a reasonable upper limit on the stability of that layer, and the penalty increases with the square of the excess. This design avoids over-constraint interference within the reasonable shear range.

[0169] The overall physical constraint loss is:

[0170] in, , , Initially set to 0.5, 0.3, and 0.2 respectively, and adjusted according to the epoch during the course learning phase. Unified scaling is introduced (when epoch≤30, the physical loss weights increase linearly from 0 to avoid conflicts between physical constraints and data fitting in the early stages of training).

[0171] Step S3.7: Multi-task consistency loss :

[0172] ;

[0173] The output of this forced wind shear auxiliary task head is consistent with the wind shear value implied in the wind field output by the main task head, realizing cross-monitoring among multiple tasks.

[0174] Step S3.8: Uncertainty weighting and joint loss function.

[0175] An uncertainty-weighted strategy is adopted, with the weights of each loss term used as learnable parameters:

[0176] ;

[0177] This allows the model to adaptively adjust the weights of each task during training.

[0178] Step S3.9: Training strategy.

[0179] The following training strategy was adopted:

[0180] (a4) Data partitioning: The first 15 years of the 17 years of tower observation data were used as the training set, and the last 2 years were used as the independent validation set. The training set was further divided into a model training subset and a parameter tuning validation subset in an 8:2 ratio.

[0181] (b4) Optimizer selection: The AdamW optimizer is used, and the initial learning rate is set to The weight decays to .

[0182] (c4) Learning rate scheduling: Cosine Annealing with WarmRestarts is adopted, with an initial period of 20 epochs, and the period doubles after each restart.

[0183] (d4) Curriculum Learning Strategy: In the early stages of training (the first 30 epochs), the main focus is on optimizing the data-driven loss term. ), and then gradually increase the physical constraint loss term ( The weights of ) are determined by a linear growth factor:

[0184] ;

[0185] Control the pace of introducing physical constraints to avoid conflicts between physical constraints and data fitting in the early stages of training.

[0186] (e) Early stopping strategy: The weighted average of the multivariate combined RMSE on the validation set is used as the monitoring indicator. If there is no improvement after 15 consecutive epochs, training is stopped and the optimal model parameters are restored.

[0187] Step S4: Spatial generalization – scaling up from point locations to regions.

[0188] Step S4.1: Spatial gridding of the target area.

[0189] A target spatial grid is established within a 10 km radius of the urban area, centered on the tower. The horizontal resolution of the grid is set to... (Consistent with the grid parameters of the underlying surface), maintaining 8 layers in the vertical direction.

[0190] Step S4.2: Prepare for ERA5 input for each grid point.

[0191] For each grid point within the target grid Repeat the spatial interpolation method in step S1.2 to interpolate the ERA5 data from the original grid points to the precise position of the target grid point, and obtain the ERA5 virtual profile sequence of the grid point.

[0192] Step S4.3: Spatial generalization reasoning driven by underlying surface conditions.

[0193] Since the training data of this invention comes from a single tower, the underlying surface feature vector u0 is a fixed value for the tower position during the training phase. If trained directly with a constant u0, the FiLM mechanism will struggle to learn the mapping between "underlying surface differences and correction responses" from the data (constant conditional inputs do not produce effective gradient signals), and spatial generalization will degenerate into applying nearly the same correction to all grid points. To make this step feasible, this invention introduces a surface-feature perturbation augmentation strategy during the training phase:

[0194] To enable the FiLM mechanism of the Surface Conditional Injection Module (SFCM) during the single-tower training phase to learn the mapping from data gradients of "surface difference → corrected response" rather than degenerating into an identity transformation on a constant input, this invention introduces a surface feature perturbation enhancement strategy during the training phase, as follows:

[0195] In each training epoch, with a preset probability (Initially 0, linearly increasing from 0 to 0.3 over the first 20 epochs, then remaining constant at 0.3) Apply a random perturbation to the feature vectors of the underlying surface in the training batch to obtain the perturbed conditional input. Each channel is independently and randomly sampled within its physical interval: roughness length z0∈[0.01, 2.0] m, zero plane displacement height d0∈[0, 20] m, building volume ratio FAR∈[0.5, 5.0], and average building height. ∈[5, 80] m, vegetation coverage NDVI∈[0, 1], impermeable surface ratio ISA∈[0, 1], topographic elevation DEM superimposed with ±30m Gaussian noise.

[0196] The construction rules for the input side and the label side are as follows:

[0197] ERA5 Virtual Profile Sequence Auxiliary time characteristics The original tower point reanalysis field remains unchanged and is still used as the model input.

[0198] After the condition vector is fed into the perturbation Drive FiLM to generate scaling factors and offset factor ;

[0199] The target side applies pseudo-label adjustments based on the near-surface logarithmic law only to the horizontal wind speed components (U, V): the friction velocity is calculated by back-calculating the measured wind speed at a reference height (80 m) on the tower. The original tower location parameters (z0, d0) and the parameters after disturbance were used respectively. , Calculate the wind speed difference between each floor:

[0200] ;

[0201] The difference is superimposed on the observed true value as a pseudo-label for the wind speed of the perturbation sample; when a certain layer satisfies When the wind speed is within the post-disturbance canopy, no wind speed adjustment is applied to this layer; the true observation value is used directly.

[0202] The true values ​​of temperature, air pressure, relative humidity, and vertical velocity labels are retained from the original quality control values ​​of the observation tower without adjustment; the wind shear auxiliary label is recalculated synchronously according to the adjusted U / V pseudo label.

[0203] Through the above construction, the training batch simultaneously contains (original u0, observed true value) and (perturbation). (Logarithmic law-adjusted pseudo-labels) Two types of sample pairs, FiLM , Meaningful gradient updates can be generated under different underlying surface conditions, and physically interpretable responses to parameters such as roughness and zero-plane displacement can be learned. The range of underlying surface parameters covered by this perturbation enhancement is the effective working range for subsequent spatial generalization of the model; when the underlying surface features of the target grid point exceed this range, they are downgraded by the distance decay factor and confidence label in step S4.4.

[0204] In summary, given the introduction of enhanced perturbations of underlying surface features, the spatial generalization inference driven by underlying surface conditions described in this step is feasible.

[0205] Based on the above, for each target grid point, the model adaptively adjusts the correction magnitude and feature representation of each layer according to the underlying surface differences (differences in z0, d0, FAR, etc.) between the grid point and the tower point through the FiLM mechanism of the underlying surface condition injection module (SFCM). Specifically, this manifests as follows:

[0206] Single-lattice generalization inference process:

[0207] (1) Take the ERA5 virtual profile sequence of this grid point for the past 4 hours (L=24 time steps). Corresponding auxiliary time features (2) Take the feature vector of the underlying surface of the grid point. (3) Input the three into the trained PhySTAM-Net, and pass them sequentially through VPE→TDM→STCAF→SFCM→multi-task output head→PICL to obtain the 8-layer correction bias of the grid point. The corrected profile is obtained by connecting the residuals. (4) After performing the above reasoning in parallel on all grid points of the target region, proceed to the spatial consistency post-processing in step S4.4.

[0208] (a) When the underlying surface of the target grid point is close to the tower point (such as in a dense urban area, z0 and FAR are close), FiLM generates modulation parameters close to the training tower location, and the correction amplitude is comparable to that of the tower point;

[0209] (b) When the underlying surface of the target grid point differs significantly from that of the tower point, FiLM dynamically adjusts the modulation parameters based on the characteristics of the underlying surface: for example, the Yangtze River surface (z0 is very small and surface friction is weak), ERA5's overestimation of near-ground wind speed is milder than that in urban areas, and FiLM correspondingly reduces the wind speed correction magnitude; in densely built-up areas (z0 and d0 are relatively large), the effective wind profile is raised as a whole, and FiLM correspondingly increases the correction amount that varies with height to match the real wind and temperature structure of this type of underlying surface.

[0210] It should be noted that the generalization ability of FiLM is effective within the range of underlying surface parameters covered by the perturbation during the training phase. When the underlying surface features of the target grid point significantly exceed this range, the correction reliability decreases. This situation is constrained and identified by the distance decay factor and confidence label in step S4.4.

[0211] In this embodiment, only the wind speed is perturbed and enhanced, while the temperature and humidity remain at their original ERA5 interpolation values, because low-altitude wind shear is the core correction target of this invention.

[0212] Step S4.4: Spatial consistency post-processing.

[0213] Spatial consistency post-processing is applied to the correction results of all grid points within the target region:

[0214] (a6) Spatial smoothing: Gaussian filtering is used to smooth the correction results of adjacent grid points, eliminating non-physical spatial jumps caused by abrupt changes in the underlying surface parameters between grid points. The kernel radius of the Gaussian filter is adaptively adjusted according to the local underlying surface uniformity.

[0215] (b6) Distance attenuation factor: Introduce a distance attenuation weighting function centered on the tower to moderately attenuate the corrections for grid points farther from the tower:

[0216] ;

[0217] in This represents the distance from the grid point to the tower. The attenuation scale parameter is 5 km. The further the grid point is from the tower, the closer the correction bias decreases towards zero (i.e., the closer it is to the original ERA5 value), reflecting the objective law that the model's generalization ability decreases with increasing distance. The final correction result is:

[0218] ;

[0219] (c) Quality Label: Output the confidence percentage and its corresponding high / medium / low level for each grid point. The confidence level is determined by a combination of the distance between the grid point and the tower point, the similarity of the underlying surface, and the results of the physical constraint test, and is calculated as follows. First, calculate three factors with values ​​ranging from 0 to 1: (1) Distance factor f d Take the distance decay weight of this grid point in (b6). .

[0220] (2) Underlying surface factor f s: Normalize the underlying surface feature tensor U from step S1.4 to 0-1 channel by channel within the target region, and take the cosine similarity between the normalized grid point feature vector and the tower point feature vector u0; if any of z0, d0, or FAR of the grid point exceeds the value range set in step S4.3 for perturbation enhancement, f s Take 0.

[0221] (3) Physical factors f p Record whether the grid point was adjusted in the wind shear consistency constraint check (hard constraint B2) and whether any variable was truncated at any level in the variable physical range constraint check (hard constraint B3); if neither occurred, take 1; if one occurred, take 0.5; if both occurred, take 0.

[0222] Confidence percentage P = (0.4· f d +0.3· f s +0.3· f p The relative reliability score of the grid point correction result is calculated as P × 100%; P ≥ 75% is marked as high, 40% ≤ P < 75% as medium, and P < 40% as low. f d , f s It is independent of time and only needs to be calculated once during the meshing process in step S4.1; f p Updated with each output of step S4.5.

[0223] Step S4.5: Real-time process execution.

[0224] When the system is running in real time, it executes the following loop with an update cycle of 10 minutes:

[0225] (a7) Receive the latest numerical model analysis field or short-term forecast field in real time (such as regional model outputs like ECMWF HRES, CMA-MESO / GRAPES, etc., whose variables and vertical layer configurations are consistent with the ERA5 input used for training; ERA5 is only used for offline training due to a several-day delay in release), and perform spatiotemporal interpolation according to the method in step S1.2. Specifically, when the real-time input field comes from a numerical model from a different source than the ERA5 used for training (such as ECMWF HRES, CMA-GRAPES, etc.), first calculate the parameters based on the training set statistical parameters (…). , The model field is normalized and aligned, and the physical quantities are restored by inverse normalization after inference. If a channel of the real-time field variable is missing (such as the vertical velocity ω), it is replaced by the vertical velocity diagnostic field or zero value output by the same model and the confidence level is downgraded. It does not participate in hard constraint fine-tuning.

[0226] (b7) Extract the time series input of nearly 4 hours from the sliding window buffer.

[0227] (c7) Perform parallel inference on all grid points in the target area (accelerated by batch inference using GPU) and output the 8-layer meteorological parameter correction values ​​for each grid point.

[0228] (d7) Perform physical hard constraint post-processing and spatial consistency post-processing.

[0229] (e7) Output and write standardized structured data products (NetCDF or GRIB2 format).

[0230] It should be noted that the hyperparameters involved in the above embodiments, such as network depth, number of channels, number of attention heads, loss weight coefficient, physical constraint threshold, distance attenuation scale, and perturbation enhancement probability, are all exemplary optimized values ​​obtained by debugging based on meteorological observation towers and ERA5 reanalysis data. Those skilled in the art will know that, for different meteorological observation tower height configurations, different reanalysis / numerical model data sources, and different underlying surface complexities of target urban areas, the above parameters can be adaptively adjusted on the validation set, and similar technical effects can be obtained without creative effort, without affecting the implementation of the technical solution of the present invention.

Claims

1. An artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations, characterized in that: Includes the following steps: Step 1: Acquire meteorological element data, reanalysis data grid points, and underlying surface spatial characteristic parameters from meteorological observation towers at multiple vertical levels, and perform preprocessing. Step 2: Construct a multivariate joint correction deep learning model for physical constraints, including a vertical feature encoder, a time-dependent modeler, a spatiotemporal cross-attention fusion module, an underlying surface condition injection module, a multi-task output head, and a physical information constraint layer; Step 3: Design a composite loss function to train the model, including a data-driven loss term for fitting the observed true values ​​and a physical constraint loss term for constraining the physical relationships between output variables; Step 4: Generalize the location correction capability of the meteorological observation tower to the predetermined urban area; Step 5: Perform spatial consistency post-processing and physical consistency post-processing on the grid-by-grid inference results to generate a low-altitude meteorological parameter field.

2. The artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations as described in claim 1, characterized in that, In step 1, the preprocessing is as follows: quality control is performed on the meteorological element data; the reanalysis data and the meteorological element data are strictly paired in time and space through spatiotemporal interpolation to obtain the virtual profile of the reanalysis data; the spatial characteristic parameters of the underlying surface of the predetermined urban area around the meteorological observation tower are obtained and constructed as the underlying surface feature tensor.

3. The artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations as described in claim 1, characterized in that... In step 2, the vertical feature encoder is used to extract local gradient features and long-range coupling features of the input data in the vertical dimension, including a one-dimensional convolution stack and a vertical attention module; the temporal dependency modeler is used to capture the changing features and temporal dependencies across multiple time scales along the time dimension. The spatiotemporal cross-attention fusion module is used to achieve deep interaction and adaptive fusion of vertical and temporal dimension information; The underlying surface condition injection module is used to inject the spatial features of the underlying surface as conditional information into the model to modulate the feature expression of the model; among them, the one-dimensional convolution stack extracts the local vertical gradient features between adjacent layers sequentially along the vertical dimension through multiple one-dimensional convolutional layers. The vertical attention module, based on the output of the one-dimensional convolution stack, uses a multi-head self-attention mechanism to calculate the association weights between layers along the vertical dimension, and introduces a learnable vertical position bias matrix to encode the relative height information between layers, thereby capturing the long-range coupling relationship between non-adjacent vertical layers.

4. The artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations as described in claim 3, characterized in that... The time-dependent modeler includes a causal dilated convolutional layer and a temporal self-attention module. The causal dilated convolutional layer uses causal convolution with different dilation rates to extract multi-timescale variation features along the time dimension, ensuring that only information from the current and past moments is used to conform to physical causality. Based on the output of the causal dilated convolutional layer, the temporal self-attention module uses a multi-head self-attention mechanism with causal masking to calculate the dependency weights between each time step along the time dimension, and integrates the time periodic encoding and atmospheric stability enhancement model to perceive diurnal cycles, seasonal variations, and atmospheric stability conditions. The spatiotemporal cross-attention fusion module unfolds the output of the vertical feature encoder and the output of the time-dependent modeler into two-dimensional sequences, and then performs bidirectional cross-attention calculation. Specifically, it calculates a time-enhanced vertical representation using vertical features as queries and time features as keys and values, and simultaneously calculates a vertically enhanced temporal representation using time features as queries and vertical features as keys and values. Then, it performs adaptive weighted fusion of the two sets of enhanced representations through a gating mechanism. The gating mechanism dynamically controls the contribution ratio of the two sets of features in the fusion through a learnable gating weight matrix.

5. The artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations as described in claim 1, characterized in that, In step 2, the underlying surface condition injection module maps the underlying surface feature vector into a condition vector through a fully connected network. Then, using a feature modulation mechanism, the condition vector is transformed by an affine transformation to generate scaling and offset factors. The scaling and offset factors are used to scale and offset the output features of the spatiotemporal cross-attention fusion module element by element. This allows for adaptive adjustment of the correction strategy and magnitude based on the difference in underlying surface conditions between the target grid point and the observation tower, providing a key capability for spatial generalization from a single point to a region.

6. The artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations as described in claim 1, characterized in that, In step 2, the multi-task output head includes a main task head and an auxiliary task head. The main task head is used to predict the correction bias of multiple meteorological variables at each vertical level, and the prediction bias is added to the input reanalysis data through residual connection to obtain the final correction value. The auxiliary task head includes a wind shear prediction branch and an extreme wind speed probability prediction branch. The wind shear prediction branch is used to directly predict the vertical wind shear value between adjacent vertical levels, and the extreme wind speed probability prediction branch is used to predict the probability that the wind speed at each level exceeds a predetermined threshold. The auxiliary task head and the main task head form a multi-task collaborative supervision mechanism to strengthen the model's learning of the vertical gradient of the wind field and the distribution characteristics of extreme events.

7. The artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations as described in claim 1, characterized in that, In step 3, the composite loss function includes a data-driven loss term and a physical constraint loss term. The data-driven loss term includes the main task weighted multivariate correction loss, wind shear prediction loss, and extreme wind speed probability prediction loss. The main task weighted multivariate correction loss adopts an extreme value adaptive weighted loss function. The loss function adaptively adjusts the sample weights according to the degree of deviation between the observed values ​​and the mean, so that extreme events with greater deviations receive higher loss weights, and a weighted term is superimposed on the wind speed variable under strong wind conditions. The physical constraint loss term includes static equilibrium constraint loss, thermodynamic consistency constraint loss, and momentum conservation approximation constraint loss. The static equilibrium constraint loss is used to constrain the corrected temperature and pressure profiles to satisfy the static equations. The thermodynamic consistency constraint loss is used to constrain the corrected temperature, humidity, and pressure to satisfy the thermodynamic consistency relationship. The momentum conservation approximation constraint loss is used to constrain the corrected wind shear value to not exceed the range determined based on boundary layer theory. During the model training phase, the physical constraint loss term is added to the total loss function as a soft constraint regularization term in a weighted manner.

8. The artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations according to claim 1, characterized in that, Step 4 is as follows: During the model training phase, a surface feature perturbation enhancement strategy is introduced. Random perturbations are applied to the roughness length, zero-plane displacement height, and building volume ratio in the surface feature vector of the tower point within the range. Based on the logarithmic law of the near-surface wind profile in the boundary layer similarity theory, a profile adjustment amount consistent with the perturbation is generated to form an enhanced training sample. This allows the surface condition injection module to learn the mapping relationship between surface differences and correction responses. A target spatial grid is established within the predetermined urban area. For each target grid point, corresponding reanalysis data input data and surface feature vectors are prepared. The perturbation-enhanced model is used to perform grid-by-grid inference to obtain the multivariate meteorological element correction bias for each grid point. For the correction result of each target grid point, a distance-based attenuation weight is introduced. The attenuation weight decreases as the distance between the target grid point and the meteorological observation tower increases, causing the correction bias of grid points farther from the tower to decrease towards zero; and a confidence level label is attached to the output of each target grid point; the confidence level is determined by the distance between the grid point and the tower point, the similarity of the underlying surface, and the results of the physical constraint test.

9. The artificial intelligence correction method for low-altitude, high-vertical-resolution urban features based on high-tower observations according to claim 1, characterized in that, In step 5, the spatial consistency post-processing uses Gaussian filtering to spatially smooth the correction results of adjacent grid points; the kernel radius of the Gaussian filter is adaptively adjusted according to the local surface uniformity.

Images