Global-local weather prediction correction method and system based on variational data assimilation and physical conservation

Abstract

The application discloses a global-local weather prediction correction method and system based on variational data assimilation and physical conservation, and the method is as follows: S1. Obtain the prediction data of the Panggu meteorological large model in a set region and the measured data of the region, and select a set surface meteorological variable as a correction object; S2. Preprocess the meteorological data; S3. Establish the connection between the global model and the local observation based on the variational data assimilation, combine the physical conservation constraint, and construct an objective function; S4. According to the constructed objective function, design a meteorological prediction data correction model of the set region based on the Seq2Seq structure in the transformer architecture; S5. Divide the input data into a training set, a validation set and a test set in proportion; input the preprocessed meteorological data into the model, optimize the model through the training set and the validation set, output the prediction data of the Panggu meteorological large model in the set region corrected by the trained model, select the output results of multiple windows on the test set, and compare them with the measured data to verify the model effect.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, specifically relating to a global-local meteorological forecasting correction method and system based on the principle of variational data assimilation and physical conservation. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence technology, AI-powered meteorological big data models have been widely applied in the field of weather forecasting, with the Pangu meteorological big data model being a prime example. Huawei Cloud's Pangu meteorological big data model is a high-resolution global AI weather forecasting system, boasting a prediction speed more than 10,000 times faster than traditional numerical weather prediction. Currently, the Pangu meteorological big data model can complete global weather forecasts within seconds on a single GPU on a single server. Its weather forecast results include geopotential, humidity, wind speed, temperature, and sea-level pressure, and can be directly applied to various meteorological research scenarios.

[0003] Although the Pangu meteorological model can generate medium- and long-term weather forecasts at an extremely fast pace, its inference is based on ERA5 reanalysis data, and its spatial resolution is limited. Therefore, significant errors still exist between the model's predictions and observed data in certain local areas. This necessitates correction of the model's predictions. Traditional local correction methods, such as weighted regression, correct the predictions by establishing relationships between data within local regions. These methods focus on local correction but struggle to provide a global interpretation.

[0004] With the development of artificial intelligence technology, deep learning models are increasingly suitable for processing time series data correction. RNNs can effectively capture long-term dependencies in time series data and also have strong local feature capture capabilities. Transformers' parallel computing capabilities and self-attention mechanisms can better uncover autocorrelation between sequence elements, giving them a greater advantage in long-term time series tasks. However, purely data-driven AI correction methods are prone to disrupting the physical consistency of meteorological fields, affecting the compatibility of correction results with global models. Summary of the Invention

[0005] To address the issue that the accuracy of prediction data from the existing Pangu meteorological model is still insufficient compared to measured data in local areas, this invention provides a global-local meteorological prediction correction method and system based on variational data assimilation and physical conservation.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A global-local weather forecast correction method based on variational data assimilation and physical conservation includes the following steps:

[0008] S1. Acquire meteorological data, which includes the prediction data of the Pangu meteorological model in the set area and the measured data of the meteorological station in the area, and select the set surface meteorological variables as the correction objects;

[0009] S2. Preprocess the meteorological data obtained in step S1, including forecast error calculation, standardization, adding time coding features, and generating a sliding window;

[0010] S3. Based on variational data assimilation, establish the connection between the global model and local observations, and combine physical conservation constraints to construct the objective function to be optimized;

[0011] S4. Based on the objective function constructed in step S3, and using the Seq2Seq structure in the transformer architecture as a basis, design a meteorological forecast data correction model for the region specified in step S1.

[0012] S5. Divide the input data into training set, validation set and test set according to the proportion; input the meteorological data preprocessed in step S2 into the meteorological forecast data correction model in step S4, optimize the meteorological forecast data correction model through training set and validation set, output the corrected forecast data of the set area of ​​the Pangu meteorological big model, select the output results of multiple windows on the test set and compare them with the measured data to verify the model effect.

[0013] Furthermore, in step S1, the process of acquiring meteorological data is as follows:

[0014] The predicted data from the Pangu meteorological model for the designated area and the measured data from the meteorological stations in that area were obtained respectively. The 2m temperature (T) was used as the basis for the determination. 2m (Unit: K), wind speed in longitude and latitude at 10m (u 10 v 10 (Unit: m / s) is used as the surface meteorological variable to be corrected.

[0015] Furthermore, in step S2, the data preprocessing process is as follows:

[0016] (1) Calculate the difference between the measured data and the predicted data in step S1 to obtain the error data of the Pangu meteorological model prediction, which serves as the basis for correcting the meteorological prediction data of the set area.

[0017] (2) Select a time series meteorological dataset consisting of measured data, predicted data, and error data, and standardize the above three types of meteorological data separately according to the following formula:

[0018]

[0019] Among them, X stanX represents the data after standardization, and X represents the data before standardization. mean X represents the mean of the original data. std This represents the standard deviation of the original data;

[0020] (3) Add sine-cosine time coding features for hour and date, the specific process of which is as follows:

[0021]

[0022]

[0023]

[0024]

[0025] Among them, hour sin and hour cos These represent the periodic sine and cosine codes for hours, respectively. `hour` represents the hourly data in the original dataset, and `date` represents the hourly data. sin and date cos These represent the periodic sine and cosine codes for dates, respectively, with date representing the date data in the original data.

[0026] (4) Use the time series sliding windows generated from all meteorological data as training samples. The total length of each window is 95, the length of historical data is 25, the length of prediction is 70, and the step size is 1.

[0027] Furthermore, in step S3, the objective function J(x) is defined as follows:

[0028] J(x) = α·||y obs -H(x)|| 2 +β·||xx global || 2 +λ·φ(x)

[0029] Among them, y obs Let x represent the measured meteorological data for a specified region, x represent the corrected global meteorological field, and H(x) represent the projection of the corrected field x onto a meteorological station in the specified region after bilinear interpolation of a single grid by the observation operator H. global This represents the global weather forecast data from the Pangoal Meteorological Model, where φ(x) represents the physical conservation constraint term, and α, β, and λ represent the observation fitting term ||y||. obs -H(x)|| 2 Background field constraint terms ||xx global || 2 And the weight value of the physical conservation constraint term φ(x).

[0030] For the physical conservation constraint φ(x), this invention selects mass conservation as the basis for the physical constraint of the wind speed field, and takes φ(x) = ||P(x)|| 2 The construction process of its residual operator P(x) is as follows:

[0031]

[0032]

[0033] This is the continuous form of the mass conservation equation, where ρ represents the air density in the atmosphere, t is the time variable, v represents the wind speed vector in the longitude and latitude directions, and the divergence operator... This invention describes the spatial rate of change of mass flux ρv. It primarily deals with wind speeds along the longitude and latitude directions, transforming the equation into a discretized residual form for these two wind speeds, thus constructing the residual operator P(x), where i and j are the grid point coordinates of the correction field x along the longitude and latitude directions, respectively. Let represent the air density at grid coordinates (i,j) at time t, Δu and Δv represent the horizontal spatial resolution of the correction field x in the longitude and latitude directions, respectively, F and G represent the mass flux in the longitude and latitude directions, respectively, and i±1 / 2 and j±1 / 2 represent the mass flux defined at the center of each grid. Further, in step S4, based on the constructed objective function, the overall meteorological forecast data correction model for the region is designed based on the Seq2Seq structure in the transformer architecture, including an encoder and a decoder;

[0034] The encoder extracts features from the input data by mining the relationship between time features and meteorological data, as well as the relationship between three meteorological variables: temperature, longitude, and latitude wind speed. This feature sequence is then used as one of the inputs to the decoder to output the final local area prediction data correction result.

[0035] For each window of the training samples, the predicted data, measured data, error data, and periodic time-coded features from the previous 25 hours are used as the input data for the encoder. Before entering the encoder, the input data first passes through an embedding layer to transform the meteorological features and time-coded features into a high-dimensional vector representation, and then passes through a position encoding layer using linear transformations of sin and cos functions to provide the encoder with the temporal characteristics of the input data. The position encoding formula is as follows:

[0036] PE(pos,2i)=sin(pos / 10000 2i / d model )

[0037] PE(pos,2i+1)=cos(pos / 100002i / d model )

[0038] Where PE(g) represents the location code of the input data, pos represents the position of the meteorological data at a single time point in the time series, i represents the dimension number of the input data after the embedding layer, and d model Indicates the dimension size of the embedding layer;

[0039] The decoder generates corrected weather forecast data for a given area by focusing on the relevant information between the feature sequence output by the encoder and the weather forecast data input to the decoder.

[0040] For each window of the training samples, the predicted data from the last 70 hours is used as the input data for the decoder. The input data is processed by the embedding layer and the positional coding layer before entering the decoder.

[0041] Furthermore, step S5 is detailed as follows:

[0042] (1) Divide the input data into training set, validation set and test set in a 6:2:2 ratio;

[0043] (2) Input the preprocessed meteorological data into the meteorological forecast data correction model in step S4, optimize the meteorological forecast data correction model on the training set and validation set by the objective function, and adjust the model parameters until the objective function reaches the optimal;

[0044] (3) On the test set, the modified forecast data of the Pangu meteorological model for the set area is output by the trained meteorological forecast data correction model. The output results of multiple windows are selected and compared with the measured data to make a comparison chart to verify the model effect.

[0045] This invention also discloses a global-local weather forecast correction system based on variational data assimilation and physical conservation, used to execute the above method, comprising the following modules:

[0046] Meteorological data acquisition module: used to acquire meteorological data, including the prediction data of the Pangu meteorological model in the set area and the measured data of the meteorological station in the area, and select the set surface meteorological variables as the correction objects;

[0047] Preprocessing module: performs preprocessing on the acquired meteorological data;

[0048] Objective function construction module: Based on variational data assimilation, establish the connection between the global model and local observations, and combine physical conservation constraints to construct the objective function to be optimized;

[0049] Meteorological forecast data correction model construction module: Based on the constructed objective function, and using the Seq2Seq structure in the transformer architecture, a meteorological forecast data correction model for the specified region is constructed.

[0050] Model training and validation module: The input data is divided into training set, validation set and test set according to the proportion; the preprocessed meteorological data is input into the meteorological forecast data correction model, the meteorological forecast data correction model is optimized through the training set and validation set, the trained meteorological forecast data correction model outputs the corrected forecast data of the specified area of ​​the Pangu meteorological big model, and the output results of multiple windows are selected on the test set and compared with the measured data to verify the model effect.

[0051] The beneficial effects of this invention are:

[0052] This invention establishes a link between the global model and local observations through an objective function constructed based on variational data assimilation. Compared with traditional correction methods and purely data-driven AI correction methods, the meteorological forecast data correction model built on this objective function can effectively reduce the error of the Pangu meteorological model's prediction data in a specified region. Furthermore, due to the inclusion of physical conservation constraints, the model does not disrupt the physical consistency of the meteorological field, ensuring the compatibility of the correction results with the global model. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of a preferred embodiment of the present invention of a global-local weather forecast correction method based on variational data assimilation and physical conservation;

[0054] Figure 2 This is a flowchart illustrating the data acquisition and preprocessing process of a preferred embodiment of the present invention.

[0055] Figure 3 This is a diagram of the internal structure of the meteorological forecast data correction model according to a preferred embodiment of the present invention;

[0056] Figure 4 This is a comparison chart of the model performance verification curves of the present invention;

[0057] Figure 5 This is a comparison chart of the model performance verification curves of the present invention;

[0058] Figure 6 This is a comparison chart of the model performance verification curves of the present invention.

[0059] Figure 7 This is a block diagram of a global-local weather forecast correction system based on variational data assimilation and physical conservation, according to a preferred embodiment of the present invention.

[0060] Figure 8This is a schematic diagram of the Seq2Seq structure in the transformer architecture of a preferred embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] See Figures 1 to 6 A preferred embodiment of the present invention relates to a global-local weather forecast correction method based on variational data assimilation and physical conservation, comprising the following steps:

[0063] S1. Acquire meteorological data, including the prediction data of the Pangu meteorological model in the designated area and the measured data of the meteorological station in the area, and select the designated surface meteorological variables as the correction objects;

[0064] S2. Preprocess the meteorological data obtained in step S1, including forecast error calculation, standardization, adding time coding features, and generating a sliding window;

[0065] S3. Based on variational data assimilation, establish the connection between the global model and local observations, and combine physical conservation constraints to construct the objective function to be optimized;

[0066] S4. Based on the objective function constructed in step S3, use the Seq2Seq structure in the transformer architecture (e.g.) Figure 8 Based on the above, design step S1 to set the meteorological forecast data correction model for the region;

[0067] S5. Divide the input data into training set, validation set and test set according to the proportion; input the meteorological data preprocessed in step S2 into the meteorological forecast data correction model in step S4, optimize the meteorological forecast data correction model through training set and validation set, output the corrected forecast data of the set area of ​​the Pangu meteorological big model, select the output results of multiple windows on the test set and compare them with the measured data to verify the model effect.

[0068] In step S1 of this embodiment, the process of acquiring meteorological data is as follows:

[0069] (1) The horizontal spatial resolution of the Pangoal Meteorological Model reaches 0.25°×0.25°, and the temporal resolution is 1 hour. Downloading ERA5 reanalysis data at a specific moment is used as input to the Pangoal Meteorological Model. The model outputs hourly global weather forecasts for the next 4 days, including geopotential, humidity, wind speed, temperature, and sea-level pressure. Repeating this process several times yields global weather forecast data for a certain period (longer-term). Local forecast data at a set latitude and longitude is then extracted from this data, and the 2m temperature (T) is used as the basis for the forecast. 2m (Unit: K), wind speed in longitude and latitude at 10m (u 10 v 10 (Unit: m / s) is used as the surface meteorological variable to be corrected.

[0070] (2) Select data corresponding to the above-mentioned surface meteorological variables from meteorological stations in the same area at the same time to obtain the measured meteorological data for the designated area. In particular, since the meteorological station only provides data on wind speed (wv, unit: m / s) and wind direction (wd, unit: deg), convert the wind speed and wind direction into longitude and latitude wind speeds according to the following formula:

[0071]

[0072]

[0073] Where U represents the meridional wind speed, V represents the zonal wind speed, wv represents the wind speed, and wd represents the wind direction.

[0074] In step S2 of this embodiment, the data preprocessing process is as follows:

[0075] (1) Calculate the difference between the measured data and the predicted data to obtain the error data of the Pangu meteorological model prediction, which serves as the basis for correcting the meteorological prediction data of the set area;

[0076] (2) Select a time series meteorological dataset consisting of measured data, predicted data, and error data, and standardize the above three types of meteorological data separately according to the following formula:

[0077]

[0078] Among them, X stan X represents the data after standardization, and X represents the data before standardization. mean X represents the mean of the original data. std This represents the standard deviation of the original data;

[0079] (3) Add sine-cosine time coding features for hour and date, the specific process of which is as follows:

[0080]

[0081]

[0082]

[0083]

[0084] Among them, hour sin and hour cos These represent the periodic sine and cosine codes for hours, respectively. `hour` represents the hourly data in the original dataset, and `date` represents the hourly data. sin and date cos These represent the periodic sine and cosine codes for dates, respectively, with date representing the date data in the original data.

[0085] (4) Use the time series sliding windows generated from all meteorological data as training samples. The total length of each window is 95, the length of historical data is 25, the length of prediction is 70, and the step size is 1.

[0086] In step S3 of this embodiment, the objective function construction process based on variational data assimilation and physical conservation constraints is as follows:

[0087] (1) Variational data assimilation is a data fusion method that combines numerical models with actual observation data. It adjusts the model state through optimization algorithms to more accurately reflect physical reality. This invention establishes the connection between global predictions and local observations of the Pangoal Meteorological Model based on variational data assimilation, constructs an objective function to be optimized to quantify the difference between the Pangoal Meteorological Model's predicted data and the observed data, and uses an optimization algorithm to minimize the objective function value, thereby correcting the Pangoal Meteorological Model's predicted data. The objective function J(x) is specifically defined as follows:

[0088] J(x) = α·||y obs -H(x)|| 2 +β·||xx global || 2 +λ·φ(x)

[0089] Among them, y obs Let x represent the measured meteorological data for a specified region, x represent the corrected global meteorological field, and H(x) represent the projection of the corrected field x onto a meteorological station in the specified region after bilinear interpolation of a single grid by the observation operator H. global This represents the global weather forecast data from the Pangoal Meteorological Model, where φ(x) represents the physical conservation constraint term, and α, β, and λ represent the observation fitting term ||y||. obs -H(x)|| 2 Background field constraint terms ||xx global ||2 And the weight value of the physical conservation constraint term φ(x).

[0090] For the physical conservation constraint φ(x), this invention selects mass conservation as the basis for the physical constraint of the wind speed field, and takes φ(x) = ||P(x)|| 2 The construction process of its residual operator P(x) is as follows:

[0091]

[0092]

[0093] This is the continuous form of the mass conservation equation, where ρ represents the air density in the atmosphere, t is the time variable, v represents the wind speed vector in the longitude and latitude directions, and the divergence operator... This invention describes the spatial rate of change of mass flux ρv. It primarily deals with wind speeds along the longitude and latitude directions, transforming the equation into a discretized residual form for these two wind speeds, thus constructing the residual operator P(x), where i and j are the grid point coordinates of the correction field x along the longitude and latitude directions, respectively. Let represent the air density at grid coordinates (i,j) at time t, Δu and Δv represent the horizontal spatial resolution of the correction field x in the longitude and latitude directions, respectively, F and G represent the mass flux in the longitude and latitude directions, respectively, and i±1 / 2 and j±1 / 2 represent the mass flux defined at the center of each grid.

[0094] (2) For the objective function J(x), observe the fitted term ||H(x)-y obs || 2 Its function is to make the corrected field x approximate the measured data y of the set area after being projected by the observation operator H. obs Background field constraint terms ||xx global || 2 Its function is to constrain the modified field x and the global background field x. global To prevent the correction results from deviating excessively from the predictions of the Pangu meteorological model, the physical conservation constraint term φ(x) constrains the wind speed field based on mass conservation, ensuring the physical consistency of the global meteorological field.

[0095] In step S4 of this embodiment, the process of designing a local meteorological forecast data correction model based on the objective function is as follows:

[0096] Based on the constructed objective function, the regional meteorological forecast data correction model is designed based on the Seq2Seq structure in the transformer architecture, which includes two parts: encoder and decoder.

[0097] The encoder extracts features from the input data by mining the relationship between time features and meteorological data, as well as the relationship between three meteorological variables: temperature, longitude, and latitude wind speed. This feature sequence is then used as one of the inputs to the decoder to output the final local area prediction data correction result.

[0098] For each window of the training samples, the predicted data, measured data, error data, and periodic time-coded features (a total of 9 meteorological features + 4 time-coded features) from the previous 25 hours are used as the input data for the encoder. Before entering the encoder, the input data first passes through an embedding layer to transform the meteorological features and time-coded features into high-dimensional vector representations to capture the relationships of the input data in high-dimensional space. Then, it passes through a positional encoding layer using linear transformations of sin and cosine functions to provide the encoder with the temporal characteristics of the input data. The positional encoding formula is as follows:

[0099] PE(pos,2i)=sin(pos / 10000 2i / d model )

[0100] PE(pos,2i+1)=cos(pos / 10000 2i / d model )

[0101] Where PE(g) represents the location code of the input data, pos represents the position of the meteorological data at a single time point in the time series, i represents the dimension number of the input data after the embedding layer, and d model This indicates the dimension size of the embedding layer.

[0102] The encoder consists of two sub-layers: a multi-head attention layer and a feedforward fully connected layer. Compared to single-head attention, the multi-head attention mechanism can capture more diverse dependencies and improve the model's robustness. Through multi-head attention, the encoder uses multiple sets of Q, K, and V matrices to capture multi-dimensional relationships between meteorological time series, enhancing the richness of the input features. The output of the multi-head attention layer is passed through the feedforward fully connected layer and used as the encoder's output. Residual connections and layer normalization are applied after each sub-layer to enhance the model's learning ability. Finally, the encoder maps the input meteorological time series into hidden layers and outputs them to the decoder.

[0103] The decoder section focuses on the relevant information between the feature sequence output by the encoder and the meteorological forecast data input to the decoder, and finally gradually generates corrected local meteorological forecast data.

[0104] For each window of the training samples, the predicted data from the following 70 hours is used as the input data for the decoder. This input data is also processed through an embedding layer and a positional encoding layer before entering the decoder. The decoder also employs a multi-head attention mechanism, but unlike the encoder, to avoid exposing information from future time points, the decoder uses a mask to shield information after the current processing time, ensuring that the generated corrected meteorological time series only relies on the portions generated at previous time points. The residual connections, layer normalization, and feedforward fully connected layers in the decoder are the same as those in the encoder.

[0105] Based on the concept of residual learning, the decoder does not directly output the corrected prediction data. Instead, it outputs the error between the measured data and the prediction data for the next 70 hours to simplify model training. The decoder output is processed through fully connected layers and softmax layers, and then added to the prediction data for the corresponding time period to obtain the corrected weather prediction data for the specified area.

[0106] In step S5 of this embodiment, the training and validation process of the model is as follows:

[0107] (1) Divide the input data into training set, validation set and test set in a 6:2:2 ratio;

[0108] (2) Input the preprocessed meteorological data into the model, optimize the model on the training set and validation set using the objective function, and adjust the model parameters until the objective function reaches its optimum;

[0109] (3) On the test set, the corrected Pangu meteorological big model is used to output the predicted data of the set area through the trained model. The output results of multiple windows are selected and compared with the measured data to make a comparison chart to verify the model effect.

[0110] The following verification will be based on specific data.

[0111] Data from the WS Saaleaue meteorological station in Germany were selected as the measured data, and the prediction data of the Pangoal Meteorological Model at the same latitude and longitude (51°N, 11.5°E) were obtained. The selected time range was from 01:00:00 on December 31, 2023 to 23:00:00 on December 31, 2024.

[0112] During model training, the optimizer was AdamW. After hyperparameter tuning, the model had 5 encoder and decoder layers, 8 attention heads, and 512 embedding layers. The objective function's α, β, and λ were set to 1, 0.1, and 0.01, respectively. The corrected Pangu meteorological model prediction data was obtained on the test set. For each selected window, a set of curve comparisons was created between the predicted data, observed data, and the corrected prediction data, as shown in the figure. Figure 4-6As shown, the horizontal axis represents time, and the vertical axis represents meridional wind speed, zonal wind speed, and temperature, respectively. The model's effectiveness is verified by comparing the closeness of the corrected predicted data with the measured data. Figure 4-6 It can be seen that there is a certain error between the original forecast data of the Pangu meteorological model and the measured data of the meteorological stations in the designated area. After processing by the meteorological forecast data correction model, the corrected forecast data is basically consistent with the measured data. This result shows that the meteorological forecast data correction model designed in this invention can effectively correct the error between the forecast data of the Pangu meteorological model and the measured data of the meteorological stations in the designated area, and significantly improve the forecast accuracy of the Pangu meteorological model.

[0113] like Figure 7 As shown in the figure, this embodiment presents a global-local weather forecast correction system based on variational data assimilation and physical conservation, used to execute the above method, and includes the following modules:

[0114] Meteorological data acquisition module: used to acquire meteorological data, including the prediction data of the Pangu meteorological model in the set area and the measured data of the meteorological station in the area, and select the set surface meteorological variables as the correction objects;

[0115] Preprocessing module: performs preprocessing on the acquired meteorological data;

[0116] Objective function construction module: Based on variational data assimilation, establish the connection between the global model and local observations, and combine physical conservation constraints to construct the objective function to be optimized;

[0117] Meteorological forecast data correction model construction module: Based on the constructed objective function, and using the Seq2Seq structure in the transformer architecture, a meteorological forecast data correction model for the specified region is constructed.

[0118] Model training and validation module: The input data is divided into training set, validation set and test set according to the proportion; the preprocessed meteorological data is input into the meteorological forecast data correction model, the meteorological forecast data correction model is optimized through the training set and validation set, the trained meteorological forecast data correction model outputs the corrected forecast data of the specified area of ​​the Pangu meteorological big model, and the output results of multiple windows are selected on the test set and compared with the measured data to verify the model effect.

[0119] Other aspects of this embodiment can be found in the above method embodiments.

[0120] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A global-local weather forecast correction method based on variational data assimilation and physical conservation, characterized by: Includes the following steps: S1. Obtain meteorological data, which includes the prediction data of the Pangu meteorological model in the set area and the measured data of the meteorological station in the area. Select the set surface meteorological variables as the correction objects. S2. Preprocess the meteorological data obtained in step S1; S3. Based on variational data assimilation, establish the connection between the global model and local observations, and construct the objective function to be optimized by combining physical conservation constraints; in this step, the objective function... The definition is as follows: in, This represents the actual meteorological data for the designated area. This indicates the corrected global weather field. Indicates the modified field Observation Operator The projection of a single grid cell, calculated using bilinear interpolation, onto a weather station in a specified region. This represents global weather forecast data from the Pangoal Meteorological Model. Represents the physical conservation constraint term. , , They represent the observation fitting terms respectively. Background field constraints and physical conservation constraints The weight value; For physical conservation constraints The law of mass conservation is chosen as the basis for the physical constraints of the wind speed field, and... Its residual operator The construction process is as follows: Let be the continuous form of the mass conservation equation, where This indicates the density of air in the atmosphere. For time variables, The divergence operator represents the wind speed vector along the longitude and latitude directions. Used to describe mass flux The rate of spatial change; S4. Based on the objective function constructed in step S3, and using the Seq2Seq structure in the transformer architecture as a basis, design a meteorological forecast data correction model for the region specified in step S1. S5. Divide the input data into training set, validation set and test set according to the proportion; input the meteorological data preprocessed in step S2 into the meteorological forecast data correction model in step S4, optimize the meteorological forecast data correction model through training set and validation set, output the corrected forecast data of the set area of ​​the Pangu meteorological big model, select the output results of multiple windows on the test set and compare them with the measured data to verify the model effect.

2. The global-local weather forecast correction method based on variational data assimilation and physical conservation as described in claim 1, characterized in that, Step S1 is as follows: The predicted data of the Pangu meteorological model in the designated area and the measured data of the meteorological station in the area were obtained respectively. The 2m temperature, 10m longitude and latitude wind speed were used as the surface meteorological variables to be corrected.

3. The global-local weather forecast correction method based on variational data assimilation and physical conservation as described in claim 1, characterized in that, The preprocessing in step S2 includes prediction error calculation, standardization, adding time-coded features, and generating a sliding window.

4. The global-local weather forecast correction method based on variational data assimilation and physical conservation as described in any one of claims 1-3, characterized in that, Step 2 is as follows: (1) Calculate the difference between the measured data and the predicted data in step S1 to obtain the error data predicted by the Pangu meteorological model, which serves as the basis for correcting the meteorological prediction data of the set area. (2) Select a time series meteorological dataset consisting of measured data, predicted data, and error data, and standardize the three types of meteorological data separately according to the following formula: in, This represents the standardized data. This represents the data before standardization. This represents the mean of the original data. This represents the standard deviation of the original data; (3) Add sine-cosine time coding features for hour and date. The specific process is as follows: in, and These represent the periodic sine and cosine codes for each hour, respectively. This represents the hourly data in the original dataset. and These represent the periodic sine and cosine codes for dates, respectively. This represents the date data in the original data; (4) Use the time series sliding window generated from all meteorological data as training samples.

5. The global-local weather forecast correction method based on variational data assimilation and physical conservation as described in claim 4, characterized in that, In step S4, based on the constructed objective function, the overall design of the regional meteorological forecast data correction model is based on the Seq2Seq structure in the transformer architecture, including an encoder and a decoder. The encoder extracts features from the input data by mining the relationship between time features and meteorological data, as well as the relationship between three meteorological variables: temperature, longitude, and latitude wind speed. This feature sequence is then used as one of the inputs to the decoder to output the final local area prediction data correction result. For each window of the training samples, the predicted data, measured data, error data, and periodic time-coded features from the previous 25 hours are used as input data for the encoder. Before entering the encoder, the input data first passes through an embedding layer to transform the meteorological features and time-coded features into high-dimensional vector representations, and then... and The position coding layer, which involves linear transformation of a function, provides the encoder with the temporal characteristics of the input data; the position coding formula is as follows: in, Indicates the position code of the input data. This indicates the position of meteorological data at a single point in time within the time series. This indicates the dimension number of the input data after it has passed through the embedding layer. Indicates the dimension size of the embedding layer; The decoder generates corrected local weather forecast data step by step by focusing on the relevant information between the feature sequence output by the encoder and the weather forecast data input to the decoder. For each window of the training samples, the predicted data from the last 70 hours is used as the input data for the decoder. The input data is processed by the embedding layer and the positional coding layer before entering the decoder.

6. The global-local weather forecast correction method based on variational data assimilation and physical conservation as described in claim 5, characterized in that, Step S5 is as follows: (1) Divide the input data into training set, validation set and test set in a 6:2:2 ratio; (2) Input the preprocessed meteorological data into the meteorological forecast data correction model in step S4, optimize the meteorological forecast data correction model on the training set and validation set by the objective function, and adjust the model parameters until the objective function reaches the optimal. (3) On the test set, the modified forecast data of the Pangu meteorological model for the set area is output by the trained meteorological forecast data correction model. The output results of multiple windows are selected and compared with the measured data to make a comparison chart to verify the model effect.

7. A global-local weather forecast correction system based on variational data assimilation and physical conservation, used to perform the method as described in any one of claims 1-6, characterized in that, Includes the following modules: Meteorological data acquisition module: used to acquire meteorological data, including the prediction data of the Pangu meteorological model in the set area and the measured data of the meteorological station in the area, and select the set surface meteorological variables as the correction objects; Preprocessing module: performs preprocessing on the acquired meteorological data; Objective function construction module: Based on variational data assimilation, establish the connection between the global model and local observations, and combine physical conservation constraints to construct the objective function to be optimized; Meteorological forecast data correction model construction module: Based on the constructed objective function, and using the Seq2Seq structure in the transformer architecture, a meteorological forecast data correction model for the specified region is constructed. Model training and validation module: The input data is divided into training set, validation set and test set according to the proportion; the preprocessed meteorological data is input into the meteorological forecast data correction model, the meteorological forecast data correction model is optimized through the training set and validation set, the trained meteorological forecast data correction model outputs the corrected forecast data of the specified area of ​​the Pangu meteorological big model, and the output results of multiple windows are selected on the test set and compared with the measured data to verify the model effect.

Images