Meteorological element assimilation method and system based on encoder-decoder and graph neural network
By constructing a meteorological element assimilation model based on an encoder-decoder and graph neural network, the problems of insufficient computing speed and resources in existing technologies are solved, and efficient meteorological data assimilation and rapid generation of numerical models are achieved.
Patent Information
- Application Number
- CN202311129888.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-01
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2043-09-01
AI Technical Summary
Existing meteorological observation data and gridded meteorological data assimilation methods cannot meet the needs of practical applications, especially in terms of computational speed and resources, and cannot effectively drive numerical weather forecasting.
A meteorological element assimilation method based on encoder-decoder and graph neural network is adopted to construct a meteorological element assimilation model. The encoder encodes gridded meteorological forecast information and uses the decoder to merge with graph structured observation data information for convolution calculation to generate a gridded meteorological field.
It significantly improved the assimilation speed of meteorological elements, reduced the demand for computing resources, enhanced the gridding accuracy of meteorological data and the application level of ground observation data, while accelerating the generation of assimilation fields in numerical models and improving the operating efficiency of numerical models.
Smart Images

Figure CN117112844B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a meteorological element assimilation method and system, in particular to a meteorological element assimilation method and system based on encoder-decoder and graph neural network. Background Art
[0002] Assimilation is a method that organically integrates meteorological observation data and two- or three-dimensional gridded meteorological data. It can be applied to the fields of initial field generation for numerical weather forecasting and application of real-time observation data. In a broad sense, it is a type of optimal interpolation method.
[0003] Generally speaking, if you want to integrate meteorological observation data and gridded meteorological data of numerical models, the optimal interpolation method or variational assimilation method is currently used. Specifically, the optimal interpolation method can be the Cressman interpolation method, and the variational assimilation method can be three- / four-dimensional variational assimilation, ensemble Kalman filtering, etc. Among them, the optimal interpolation method has the characteristics of fast calculation speed, but the interpolation effect is not good, and it cannot be used as an initial field to drive numerical weather forecasting; the variational assimilation method can obtain a better assimilation structure, but the calculation result is extremely slow and requires huge supercomputing resources.
[0004] In summary, the current methods used to assimilate meteorological observation data and gridded meteorological data cannot meet the actual application needs. Summary of the invention
[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a meteorological element assimilation method and system based on encoder-decoder and graph neural network, which can greatly improve the assimilation speed of meteorological elements, reduce computing resources, and reduce assimilation costs.
[0006] According to the technical solution provided by the present invention, a meteorological element assimilation method based on an encoder-decoder and a graph neural network, the meteorological element assimilation method comprising:
[0007] For a target area and a target meteorological element, a meteorological element assimilation model adapted to the target area and the target meteorological element is constructed, wherein:
[0008] The meteorological element assimilation model includes an encoder and a decoder adaptively connected to the encoder;
[0009] During assimilation, gridded meteorological forecast information of target meteorological elements in the target area and graph structure observation data information constructed based on target observation stations in the target area are provided, among which:
[0010] For the constructed graph structure observation data information, each target observation station in the target area is used as a node of the graph structure data, the observation value obtained by the target observation station on the target meteorological element is used as the node feature, the longitude and latitude of the target observation station is used as the coordinate of the node, all target observation stations are linked to each other, and the spatial distance between two target observation stations is used as the edge feature;
[0011] Encoding the gridded meteorological forecast information by using an encoder to generate a forecast information feature encoding vector after encoding;
[0012] The graph structure observation data information and the forecast information feature coding vector generated by the encoder are loaded into the decoder, and the decoder is used to merge the feature coding vector with the graph structure observation data information. After the merging, the graph neural network is used to perform convolution calculation to generate a gridded meteorological field after the convolution calculation.
[0013] When constructing a meteorological element assimilation model, the construction method includes:
[0014] A meteorological element assimilation basic model corresponding to the meteorological element assimilation model and a model training data set for training the meteorological element assimilation basic model are constructed, wherein:
[0015] The model training data set includes a number of training samples. For any training sample, it includes gridded meteorological training information of target meteorological elements and graph structure training data information constructed based on the gridded meteorological training information, wherein:
[0016] For the constructed graph structure training data information, the target observation sites in the target area are used as nodes of the graph structure data, the meteorological element values generated based on the grid meteorological training information are used as node features, the longitude and latitude of the target observation sites are used as node coordinates, all target observation sites are linked to each other, and the spatial distance between two target observation sites is used as the edge feature;
[0017] During training, for any training sample, the gridded meteorological training information is encoded using the encoder in the meteorological element assimilation basic model to generate a training information feature encoding vector after encoding;
[0018] Loading graph structure training data information and training information feature encoding vectors into a decoder in a meteorological element assimilation basic model to generate a gridded training meteorological field;
[0019] Calculate training loss based on gridded training meteorological fields and gridded meteorological training information;
[0020] When the calculated training loss matches the training loss threshold, the currently trained meteorological element assimilation basic model is configured as the meteorological element assimilation model; otherwise, the parameters of the meteorological element assimilation basic model are updated until the calculated training loss matches the training loss threshold.
[0021] When preparing training samples, the generation method of gridded meteorological training information in the training samples includes:
[0022] Acquire a reanalysis numerical model data set, and download reanalysis meteorological data corresponding to a target area and a target meteorological element from the reanalysis numerical model data set;
[0023] The downloaded reanalysis meteorological data are spatially downscaled to generate gridded meteorological training information after spatial downscaling.
[0024] When constructing graph structure training data information, it includes:
[0025] Acquire location information of a target observation site in a target area, wherein the location information of the target observation site includes the latitude and longitude of the target observation site;
[0026] Based on the location information of the target observation site, the gridded meteorological training information is interpolated to the corresponding target observation site to generate gridded interpolated meteorological data after interpolation, and meteorological element values as node features are generated based on the gridded interpolated meteorological data.
[0027] When generating meteorological element values as node features, it also includes:
[0028] Acquire site observation history data of each target observation site, and generate an observation list based on the acquired site observation history data;
[0029] Based on the gridded interpolated meteorological data of each target observation station, an interpolation list of the target observation station is generated, wherein the length of the interpolation list is the same as the length of the corresponding data in the observation list;
[0030] For the interpolation list and observation list of the same target observation station, the corresponding elements in the interpolation list and the observation list are subtracted to generate an error list;
[0031] determining an error probability distribution based on the error list;
[0032] Random disturbance is added to the gridded interpolated meteorological data based on the error probability distribution to generate meteorological element values as node features after adding the random disturbance.
[0033] When calculating the loss, the loss function includes a mean square error loss function.
[0034] When performing spatial downscaling on reanalysis meteorological data, it includes:
[0035] Determine the target spatial resolution;
[0036] Processing the reanalysis meteorological data into a data volume recognized by WRF, and generating initial fields and boundary conditions based on the data volume;
[0037] Based on the generated initial field and boundary conditions, WRF is used to generate gridded meteorological training information of target spatial resolution.
[0038] When determining the error probability distribution, include:
[0039] Based on the error list, count the maximum error value and the minimum error value;
[0040] Between the maximum error value and the minimum error value, the error frequency is calculated based on the preset element interval threshold;
[0041] Divide the calculated error frequency by the data length of the error list to obtain the frequency of each error;
[0042] Based on the frequency of each error, the probability distribution of the error over all time is determined.
[0043] The target meteorological elements include temperature, humidity, air pressure or wind speed.
[0044] A meteorological element assimilation system based on encoder-decoder and graph neural network, comprising a meteorological element assimilation device, wherein:
[0045] The meteorological element assimilation device assimilates the gridded meteorological forecast information of the target meteorological element in the target area and the graph structure observation data information constructed based on the target observation station in the target area based on the above-mentioned method.
[0046] The advantages of the present invention are assimilation using a constructed meteorological element assimilation model, which includes an encoder and a decoder. Compared with the traditional variational assimilation method, the assimilation speed can be greatly improved, the computing resources can be greatly reduced, and the model can be deployed on a computer. Compared with the traditional Cressman and other simple optimal interpolation methods, the model can greatly improve the longitude interpolation accuracy, which is beneficial to the gridding level of ground meteorological data and the application level of ground observation data. At the same time, it helps to accelerate the generation of the assimilation field of the numerical model and improve the operation efficiency of the numerical model. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 A schematic diagram of an embodiment of the meteorological element assimilation of the present invention.
[0048] Figure 2An embodiment of constructing graph structure observation data for the present invention. DETAILED DESCRIPTION
[0049] The present invention will be further described below in conjunction with specific drawings and embodiments.
[0050] In order to greatly improve the assimilation speed of meteorological elements, reduce computing resources, and lower assimilation costs, a meteorological element assimilation method based on an encoder-decoder and a graph neural network is provided. In one embodiment of the present invention, the meteorological element assimilation method includes:
[0051] For a target area and a target meteorological element, a meteorological element assimilation model adapted to the target area and the target meteorological element is constructed, wherein:
[0052] The meteorological element assimilation model includes an encoder and a decoder adaptively connected to the encoder;
[0053] During assimilation, gridded meteorological forecast information of target meteorological elements in the target area and graph structure observation data information constructed based on target observation stations in the target area are provided, among which:
[0054] For the constructed graph structure observation data information, each target observation station in the target area is used as a node of the graph structure data, the observation value obtained by the target observation station on the target meteorological element is used as the node feature, the longitude and latitude of the target observation station is used as the coordinate of the node, all target observation stations are linked to each other, and the spatial distance between two target observation stations is used as the edge feature;
[0055] Encoding the gridded meteorological forecast information by using an encoder to generate a forecast information feature encoding vector after encoding;
[0056] The graph structure observation data information and the forecast information feature coding vector generated by the encoder are loaded into the decoder, and the decoder is used to merge the feature coding vector with the graph structure observation data information. After the merging, the graph neural network is used to perform convolution calculation to generate a gridded meteorological field after the convolution calculation.
[0057] Specifically, the target area specifically refers to the area in the meteorological observation, such as the target area can be a provincial area, a municipal area or other regional conditions, and the specific conditions of the target area can be selected as needed. The target meteorological elements can include temperature, humidity, air pressure or wind speed, that is, it can be one of temperature, humidity, air pressure or wind speed. Of course, the target meteorological element can also be other common meteorological elements, which can be selected as needed. When the target meteorological element is wind speed, the wind speed can be radial wind speed or latitudinal wind speed.
[0058] Figure 1An embodiment of meteorological element assimilation is shown in the figure. After determining the target area and the target meteorological element, a meteorological element assimilation model needs to be constructed. The meteorological element assimilation model is adapted to the target area and the target meteorological element. Specifically, the meteorological element assimilation model assimilates / integrates the target meteorological elements of the target area. When the target area and / or the target meteorological element are different, the corresponding meteorological element assimilation model needs to be reconstructed.
[0059] In one embodiment of the present invention, the meteorological element assimilation model includes an encoder and a decoder connected in sequence. The specific working conditions of the encoder and the decoder are described in detail below in conjunction with the process of meteorological element assimilation.
[0060] After determining the target area, the distribution of target observation stations in the target area can be determined according to the technical means commonly used in this technical field. The target observation stations are specifically stations that can observe the target meteorological elements in the target area. Specifically, all observation stations in the target area can be configured as target observation stations, or some of the observation stations in the target area can be configured as target observation stations. The specific selection can be based on needs.
[0061] From the above description of assimilation, it can be known that during assimilation, it is necessary to provide gridded meteorological forecast information of target meteorological elements in the target area and graph structure observation data information constructed based on target observation sites in the target area. The gridded meteorological forecast information specifically refers to information about target meteorological elements generated based on multiple target observation sites. For example, when the target meteorological element information is temperature, the gridded meteorological forecast information is forecast information about the temperature in the target area. During specific implementation, the gridded meteorological forecast information is the gridded meteorological data used in assimilation mentioned in the background technology, and the graph structure observation data information is the meteorological observation data used in assimilation mentioned in the background technology.
[0062] The following is an explanation of the generation or provision of gridded meteorological forecast information of target meteorological elements in the target area, specifically:
[0063] For the data of gridded weather forecast information, the public global forecast model data (such as GFS (Global Forecast System) etc.) can be directly obtained, or the mesoscale forecast model can be driven by the global forecast model, or a self-owned numerical model system can be used. In one embodiment of the present invention, the process is described by taking ECMWF data driving the WRF (Weather Research and Forecasting Model) model to obtain the target meteorological elements of the target area as an example; specifically:
[0064] Get the latest ECMWF data;
[0065] Set the latitude and longitude range of WRF operation to the latitude and longitude range of the target area, the spatial resolution of WRF to 4 kilometers (target resolution), and set the corresponding parameterization scheme options;
[0066] Use WPS (WRF Preprocessing System) to process the latest ECMWF data into data that can be recognized by WRF. Then run the real.exe module to generate the initial field and boundary conditions, and then run WRF to obtain the forecast results with a resolution of 4 km from the running results.
[0067] The above description gives a way to generate gridded weather forecast information from ECMWF data through spatial downscaling, that is, the forecast result obtained is gridded weather forecast information, which is generally expressed as a two-dimensional data matrix. WRF and WPS are both commonly used data processing tools, and the real.exe module is a module in WRF. The specific conditions of WRF, WPS and real.exe can be consistent with the existing ones, and will not be repeated here.
[0068] From the above description, it can be seen that after determining the target observation site in the target area, the corresponding network structure data can be constructed. Figure 2 An embodiment of the constructed graph structure data is given in FIG. In the figure, V1, V2, V3 and V4 respectively represent the target observation sites in the target area, d 12 is the distance between the target observation site V1 and the target observation site V2. For other situations, the distance d can be used as a reference. 12 For the target observation sites in the target area, the longitude and latitude of each target observation site can be determined by the commonly used technical means in the technical field, and thus the distance between two target observation sites can be determined according to the longitude and latitude of the target observation sites.
[0069] Depend on Figure 2 It can be seen from the illustrated embodiment that when constructing graph structure data, each target observation site is used as a node of the graph structure data, the longitude and latitude of the target observation site are used as the coordinates of the node, all target observation sites are linked to each other, and the spatial distance between two target observation sites is used as the edge feature.
[0070] In addition, the observation values obtained by the target observation station from observing the target meteorological elements are used as node features. During specific implementation, the real-time meteorological station observation data of the target observation station can be obtained from the Meteorological Bureau and other channels, and the obtained meteorological station observation data can be quality controlled to eliminate the station data with observation errors. The specific method can use a fixed-rule quality control method, such as the method introduced in "QX / T 118-2020 Meteorological Observation Data Quality Control Ground", etc., or other existing methods can be used, depending on what can meet actual needs. The meteorological station observation data that has undergone quality control is configured with the observation values obtained by the target observation station from observing the target meteorological elements.
[0071] The above provides an embodiment of providing gridded meteorological forecast information of target meteorological elements in a target area and graph structure observation data information constructed based on target observation sites in the target area. When assimilating using the meteorological element assimilation model, the gridded meteorological forecast information is loaded into an encoder, and the gridded meteorological forecast information is encoded using the encoder to generate a forecast information feature encoding vector after encoding, wherein the forecast information feature encoding vector is a one-dimensional feature vector, and the dimension of the forecast information feature vector can be selected as needed, such as a length of 4096 or other lengths.
[0072] The graph structure observation data information and the forecast information feature coding vector generated by the encoder are loaded into the decoder, the decoder is used to merge the feature coding vector with the graph structure observation data information, and the graph neural network is used to perform convolution calculation after the merging to generate a gridded meteorological field after the convolution calculation; specifically, after the gridded meteorological field is generated, the assimilation is completed, and the generated gridded meteorological field is a two-dimensional matrix, and the size of the gridded meteorological field is consistent with the size of the gridded meteorological forecast information.
[0073] In one embodiment of the present invention, when constructing a meteorological element assimilation model, the construction method includes:
[0074] A meteorological element assimilation basic model corresponding to the meteorological element assimilation model and a model training data set for training the meteorological element assimilation basic model are constructed, wherein:
[0075] The model training data set includes a number of training samples. For any training sample, it includes gridded meteorological training information of target meteorological elements and graph structure training data information constructed based on the gridded meteorological training information, wherein:
[0076] For the constructed graph structure training data information, the target observation sites in the target area are used as nodes of the graph structure data, the meteorological element values generated based on the grid meteorological training information are used as node features, the longitude and latitude of the target observation sites are used as node coordinates, all target observation sites are linked to each other, and the spatial distance between two target observation sites is used as the edge feature;
[0077] During training, for any training sample, the gridded meteorological training information is encoded using the encoder in the meteorological element assimilation basic model to generate a training information feature encoding vector after encoding;
[0078] Loading graph structure training data information and training information feature encoding vectors into a decoder in a meteorological element assimilation basic model to generate a gridded training meteorological field;
[0079] Calculate training loss based on gridded training meteorological fields and gridded meteorological training information;
[0080] When the calculated training loss matches the training loss threshold, the currently trained meteorological element assimilation basic model is configured as the meteorological element assimilation model; otherwise, the parameters of the meteorological element assimilation basic model are updated until the calculated training loss matches the training loss threshold.
[0081] The constructed meteorological element assimilation basic model is an untrained model. Therefore, it can be seen from the above description that the meteorological element assimilation basic model includes an encoder and a decoder adapted to be connected to the encoder. In order to form a meteorological element assimilation model after training the meteorological element assimilation basic model, that is, to enable the encoder and decoder to complete the above processing, the encoder and decoder need to be selected and configured accordingly.
[0082] Specifically, the encoder can adopt a commonly used encoder form, such as a Swin Transformer model with the classification head removed as an encoder. Of course, the encoder can also use other structures, such as other classification models with the classification head removed, but the Transformer structure can relatively better obtain global information and has relatively better effects. Therefore, in specific implementations, the encoder preferably uses the Swin Transformer model with the classification head removed as the encoder.
[0083] For the decoder, we can use the NNConv model. NNConv is a convolutional layer in the PyTorch Geometric library that is used to process graph structure data. It is based on the convolution operation of the neural network and can merge graph data with matrix data to perform convolution operations on the graph neural network. For more information about the working principle of the NNConv model in PyTorch Geometric, please refer to the corresponding description in https: / / blog.csdn.net / qq_41987033 / article / details / 103497749.
[0084] In order to form the required decoder, in one embodiment of the present invention, it is necessary to customize a neural network for edge feature conversion, and the structure of the customized neural network can be the same as the conventional decoder structure; the customized neural network is used as a parameter, input into the NNConv model, and a customized NNConv layer is generated. At this time, the construction of the encoder is completed.
[0085] In the above, the custom neural network is the same as the conventional decoder structure. The conventional decoder structure specifically refers to the general architecture or model structure commonly used for decoders, such as the recurrent neural network (RNN), long short-term memory network (LSTM), or Transformer models commonly used in natural language processing tasks. These structures are usually composed of different layers and are used to generate outputs from the features generated by the encoder. Different from the encoder-decoder structure of the graph network, their common feature is that the input and output of all layers are matrix data rather than graph data. The "conventional" here refers to the "matrix data" rather than "graph data". In specific implementation, the custom neural network can be a recurrent neural network (RNN), a long short-term memory network (LSTM), or one of the Transformer models.
[0086] In the specific implementation, the defined neural network is a neural network with matrix data as input and output, and the NNConv model uses the defined neural network structure as its parameter. After the NNConv model is instantiated, it becomes a graph neural network; during assimilation, the graph neural network is used for convolution calculation after merging, specifically referring to the convolution calculation using the formed graph neural network.
[0087] When working, node features and edge features are extracted from the graph structure data, the node features are combined with the feature encoding vector, and passed into the previous custom NNConv layer to complete the decoding process of the decoder.
[0088] After building the meteorological element assimilation basic model, it is necessary to make / build a model training data set, and use the model training data set to train the meteorological element assimilation basic model, so that after reaching the target training state, the meteorological element assimilation basic model that has reached the target training state can be configured as a meteorological element assimilation model. Generally, the model training data set includes a number of training samples, and the number of training samples can be selected as needed to meet the training of the meteorological element assimilation basic model. In specific implementation, each training sample includes gridded meteorological training information and graph structure training data information. The situation of each training sample is described in detail below.
[0089] In one embodiment of the present invention, when preparing training samples, a method for generating gridded meteorological training information in the training samples includes:
[0090] Acquire a reanalysis numerical model data set, and download reanalysis meteorological data corresponding to a target area and a target meteorological element from the reanalysis numerical model data set;
[0091] The downloaded reanalysis meteorological data are spatially downscaled to generate gridded meteorological training information after spatial downscaling.
[0092] Specifically, the reanalysis numerical model dataset can generally be the ERA5 dataset, the FNL dataset or the CRA dataset. The ERA5 dataset is a global reanalysis dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), which provides high-resolution (25km) atmospheric, ground and ocean variables, covering data from 1979 to the present; the FNL dataset is a global reanalysis dataset (50km) provided by the National Center for Environmental Information (NCEP) of the United States, including atmospheric variables and ground fields, with a time range from 1999; the CRA dataset is a Chinese regional reanalysis dataset provided by the China Meteorological Administration, which provides high-resolution (30km) atmospheric and ground observation data, with a time range from 1979. Of course, the reanalysis numerical model dataset can also be of other types, which can generally be obtained by downloading from the Internet.
[0093] From the above description, it can be seen that the constructed meteorological element assimilation model is related to the target area and the target meteorological element. Therefore, it is necessary to download the reanalysis meteorological data related to the target area and the target meteorological element. For example, when the target meteorological element is temperature, the reanalysis meteorological data is the meteorological data of the temperature in the target area.
[0094] Generally speaking, in meteorology, reanalysis meteorological data belongs to large-scale data, which is generally used to drive medium / small / micro-scale numerical models and provide initial fields and boundary conditions for meteorological models. In meteorology and weather forecasting, meteorological models are mathematical models used to simulate physical processes in the atmosphere and earth system. These models use a set of equations to describe atmospheric motion, thermodynamics, radiation transfer, water cycle and other processes to predict future weather conditions.
[0095] In meteorological models, "initial field" refers to the starting point of the simulation, that is, the starting state of the simulation. In order to perform weather forecasting or climate simulation, the model requires an initial state, which contains various meteorological elements in the atmosphere and on the ground, such as temperature, humidity, wind speed, air pressure, etc.
[0096] "Boundary conditions" refer to the conditions used to limit the model around the simulation area during the simulation process. The meteorological model needs to define the boundaries of the simulation area and set boundary conditions on these boundaries to simulate the interaction between the atmosphere and the outside of the simulation area. These boundary conditions can be output from global meteorological datasets or other meteorological models.
[0097] Taken together, initial fields and boundary conditions are the input data required for the operation of meteorological models, providing the model with the initial state of the simulation and external constraints around the simulation area, thus enabling weather forecasting or climate simulation.
[0098] It can be seen from the characteristics of the above-mentioned reanalysis meteorological data that it is necessary to perform spatial downscaling processing on the reanalysis meteorological data with a lower resolution (generally 10 to 50 km). In one embodiment of the present invention, when performing spatial downscaling processing on the reanalysis meteorological data, it includes:
[0099] Determine the target spatial resolution;
[0100] Processing the reanalysis meteorological data into a data volume recognized by WRF, and generating initial fields and boundary conditions based on the data volume;
[0101] Based on the generated initial field and boundary conditions, WRF is used to generate gridded meteorological training information of target spatial resolution.
[0102] Specifically, the target spatial resolution can be selected as needed, such as the target spatial resolution is configured as 4 km as mentioned above. When performing spatial downscaling on the reanalysis meteorological data, the specific process can refer to the above description of generating gridded meteorological forecast information from ECMWF data by spatial downscaling, which will not be repeated here.
[0103] In one embodiment of the present invention, the constructed graph structure training data information includes:
[0104] Acquire location information of a target observation site in a target area, wherein the location information of the target observation site includes the latitude and longitude of the target observation site;
[0105] Based on the location information of the target observation site, the gridded meteorological training information is interpolated to the corresponding target observation site to generate gridded interpolated meteorological data after interpolation, and meteorological element values as node features are generated based on the gridded interpolated meteorological data.
[0106] Specifically, the constructed graph structure training data information can refer to the above-mentioned graph structure observation data information. The difference from the graph structure observation data information is that the values of the node features are different. During the specific implementation, for the graph structure training data information, the gridded meteorological training information is interpolated to the corresponding target observation site by interpolation, so as to generate gridded interpolated meteorological data after interpolation, and the meteorological element values as node features are generated based on the gridded interpolated meteorological data.
[0107] The interpolation method can select linear difference. Generally, there are many mature implementation methods of linear interpolation methods in the meteorological field. In one embodiment of the present invention, the simplest and most commonly used wrf_user_getvar function in the ncl language can be used to implement linear interpolation. The method of using the wrf_user_getvar function for linear interpolation can refer to the corresponding description in: https: / / blog.csdn.net / islandowner2017 / article / details / 119302952, which will not be repeated here.
[0108] In one embodiment of the present invention, when generating a meteorological element value as a node feature, it also includes:
[0109] Acquire site observation history data of each target observation site, and generate an observation list based on the acquired site observation history data;
[0110] Based on the gridded interpolated meteorological data of each target observation station, an interpolation list of the target observation station is generated, wherein the length of the interpolation list is the same as the length of the corresponding data in the observation list;
[0111] For the interpolation list and observation list of the same target observation station, the corresponding elements in the interpolation list and the observation list are subtracted to generate an error list;
[0112] determining an error probability distribution based on the error list;
[0113] Random disturbance is added to the gridded interpolated meteorological data based on the error probability distribution to generate meteorological element values as node features after adding the random disturbance.
[0114] Specifically, the site observation history data of each target observation site can be obtained through technical means commonly used in this technical field, such as through channels such as the Meteorological Bureau. Of course, the site observation history data obtained at this time specifically refers to the site observation history data of the target meteorological elements, and the time span corresponding to the observation history data is consistent with the time information of the gridded meteorological training information.
[0115] During specific implementation, an observation list is generated based on the acquired historical observation data of the site, and the elements in the observation list are all observation values corresponding to the target meteorological elements, and the observation values are true values. Based on the gridded interpolated meteorological data of each target observation site, an interpolation list of the target observation site is generated, and the elements in the interpolation list are all interpolations corresponding to the target meteorological elements, and the interpolations are not observation values. In one embodiment of the present invention, the interpolation list and the observation list have the same data length, so each element in the interpolation list corresponds one-to-one to the element in the observation list, and the corresponding elements all represent the values of the target meteorological elements in the same time period.
[0116] From the above description, it can be seen that for the same target observation site, an interpolation list and an observation list can be obtained. At this time, the corresponding elements in the interpolation list and the observation list are subtracted to generate an error list; based on the error list, the error probability distribution can be determined.
[0117] In one embodiment of the present invention, determining the error probability distribution includes:
[0118] Based on the error list, count the maximum error value and the minimum error value;
[0119] Between the maximum error value and the minimum error value, the error frequency is calculated based on the preset element interval threshold;
[0120] Divide the calculated error frequency by the data length of the error list to obtain the frequency of each error;
[0121] Based on the frequency of each error, the probability distribution of the error over all time is determined.
[0122] Specifically, the maximum error value and the minimum error value are statistically defined, which specifically refers to the maximum value and the minimum value in the statistical error list. The preset element interval threshold can be selected as needed. For example, when the meteorological element is a temperature value, the preset element interval threshold can be 0.1°C. When the meteorological element is other conditions, the preset element interval threshold can be selected as needed. The data length of the error list is the number of errors in the error list.
[0123] An embodiment of determining the probability distribution of errors is given below. Specifically, assuming that the error list is: [0.2, 0.5, 0.3, 0.1, 0.4, 0.2, 0.6, 0.3, 0.2, 0.4]. For the error list, the maximum error value is determined to be 0.6, and the minimum error value is 0.1. Assuming that the preset element interval threshold is 0.2, between the maximum error value and the minimum error value, based on the preset element interval threshold, when calculating the error frequency, we have:
[0124] 0.1~0.3: 4 times;
[0125] 0.3~0.5:4 times:
[0126] 0.5~0.7: 2 times.
[0127] The data length of the above error list is 10, so the frequency of each error is:
[0128] 0.1~0.3:4 / 10=0.4;
[0129] 0.3~0.5:4 / 10=0.4;
[0130] 0.5~0.7:2 / 10=0.2.
[0131] Based on the frequency of each error, the probability distribution of the error over all time is determined as:
[0132] 0.1~0.3:0.4;
[0133] 0.3~0.5:0.4;
[0134] 0.5~0.7:0.2.
[0135] After determining the probability distribution of errors at all times, random errors can be generated in a random manner based on the error probability distribution, that is, the distribution of the generated random errors is consistent with the probability distribution of the errors, and the random errors are added to the gridded interpolated meteorological data, that is, random disturbances are added to the gridded interpolated meteorological data based on the error probability distribution to update or adjust the gridded difference meteorological data and use them as the meteorological element values of the node characteristics.
[0136] In one embodiment of the present invention, random disturbance is added to the gridded interpolated meteorological data based on the error probability distribution to generate meteorological element values as node features after adding the random disturbance. Specifically, by adding random disturbance, when the meteorological element assimilation basic model is assimilated using the formed graph structure training data information, the meteorological element assimilation model obtained by training has better robustness.
[0137] An embodiment of training sample production is given above. After the training samples are produced in the same way and the training data set is formed, the meteorological element assimilation basic model is trained. In specific implementation, similar to the assimilation method, during training, for any training sample, the gridded meteorological training information is encoded using the encoder in the meteorological element assimilation basic model to generate a training information feature encoding vector after encoding; the graph structure training data information and the training information feature encoding vector are loaded into the decoder in the meteorological element assimilation basic model to generate a gridded training meteorological field.
[0138] Based on the gridded training meteorological field and the gridded meteorological training information, the training loss is calculated. In one embodiment of the present invention, when calculating the loss, the loss function includes a mean square error loss function (MSELoss). Of course, the loss function can also take other forms, which can be selected according to actual needs.
[0139] In specific implementation, matrix A is used to represent the gridded training meteorological field, and matrix B is used to represent the gridded meteorological training information. When calculating the training loss, the error matrix is calculated first: matrix B is subtracted element by element from matrix A to obtain a new matrix, namely the error matrix. At this time, each element of the error matrix represents the difference between the element value of matrix A and the element value of matrix B at the corresponding position, that is, error matrix = AB.
[0140] Calculate the square error matrix: Square each element in the error matrix to get a new matrix, the square error matrix. Each element of this matrix represents the square of the error at the corresponding position, so: Square error matrix = error matrix^2.
[0141] Calculate the mean square error loss: add the values of all elements in the square error matrix and divide it by the total number of elements in the matrix to get the mean square error loss. The mean square error loss measures the overall difference between matrix A and matrix B, then:
[0142] Mean squared error loss = Σ(all elements in the squared error matrix) / (total number of elements in A and B).
[0143] Through this process, the mean square error loss function is used to measure the difference between the gridded training meteorological field and the gridded meteorological training information, and a numerical value representing the degree of difference is obtained, that is, the computational loss is obtained.
[0144] Specifically, after the training loss is calculated, the calculated loss is compared with the training loss threshold, and the specific size of the training loss threshold can be selected as needed to meet the needs of meteorological element assimilation. In one embodiment of the present invention, when the calculated training loss matches the training loss threshold, the current trained meteorological element assimilation basic model is configured as the meteorological element assimilation model, otherwise, the parameters of the meteorological element assimilation basic model are updated until the calculated training loss matches the training loss threshold.
[0145] From the above, a meteorological element assimilation system based on an encoder-decoder and a graph neural network is obtained. In one embodiment of the present invention, a meteorological element assimilation device is included, wherein:
[0146] The meteorological element assimilation device assimilates the gridded meteorological forecast information of the target meteorological element in the target area and the graph structure observation data information constructed based on the target observation station in the target area based on the above-mentioned method.
[0147] Specifically, the meteorological element assimilation device can be a common computer device, such as a PC. An embodiment of constructing a meteorological element assimilation model and an embodiment of assimilating a target meteorological element using the meteorological element assimilation model are given above. Therefore, the specific assimilation method and process of the meteorological element assimilation device can refer to the above description and will not be repeated here.
[0148] The present invention uses a constructed meteorological element assimilation model for assimilation. The meteorological element assimilation model includes an encoder and a decoder. Compared with the traditional variational assimilation method, it can greatly improve the assimilation speed, greatly reduce computing resources, and can be deployed on a computer; compared with the traditional Cressman and other simple optimal interpolation methods, it can greatly improve the longitude interpolation accuracy, which is beneficial to the gridding level of high-level ground meteorological data and the application level of ground observation data; at the same time, it helps to accelerate the generation of numerical model assimilation fields and improve the operating efficiency of numerical models.
Claims
1. A meteorological element assimilation method based on encoder-decoder and graph neural network, characterized in that: The meteorological element assimilation method comprises: For a target area and a target meteorological element, a meteorological element assimilation model adapted to the target area and the target meteorological element is constructed, wherein: The meteorological element assimilation model includes an encoder and a decoder adaptively connected to the encoder; During assimilation, gridded meteorological forecast information of target meteorological elements in the target area and graph structure observation data information constructed based on target observation stations in the target area are provided, among which: For the constructed graph structure observation data information, each target observation station in the target area is used as a node of the graph structure data, the observation value obtained by the target observation station on the target meteorological element is used as the node feature, the longitude and latitude of the target observation station is used as the coordinate of the node, all target observation stations are linked to each other, and the spatial distance between two target observation stations is used as the edge feature; Encoding the gridded meteorological forecast information by using an encoder to generate a forecast information feature encoding vector after encoding; The graph structure observation data information and the forecast information feature coding vector generated by the encoder are loaded into the decoder, the decoder is used to merge the feature coding vector with the graph structure observation data information, and after the merging, the graph neural network is used to perform convolution calculation to generate a gridded meteorological field after the convolution calculation; When building a meteorological element assimilation model, it includes: A meteorological element assimilation basic model corresponding to the meteorological element assimilation model and a model training data set for training the meteorological element assimilation basic model are constructed, wherein: The model training data set includes a number of training samples. For any training sample, it includes gridded meteorological training information of target meteorological elements and graph structure training data information constructed based on the gridded meteorological training information, wherein: For the constructed graph structure training data information, the target observation sites in the target area are used as nodes of the graph structure data, the meteorological element values generated based on the grid meteorological training information are used as node features, the longitude and latitude of the target observation sites are used as node coordinates, all target observation sites are linked to each other, and the spatial distance between two target observation sites is used as the edge feature; During training, for any training sample, the gridded meteorological training information is encoded using the encoder in the meteorological element assimilation basic model to generate a training information feature encoding vector after encoding; Loading graph structure training data information and training information feature encoding vectors into a decoder in a meteorological element assimilation basic model to generate a gridded training meteorological field; Calculate training loss based on gridded training meteorological fields and gridded meteorological training information; When the calculated training loss matches the training loss threshold, the currently trained meteorological element assimilation basic model is configured as the meteorological element assimilation model; otherwise, the parameters of the meteorological element assimilation basic model are updated until the calculated training loss matches the training loss threshold.
2. The meteorological element assimilation method based on encoder-decoder and graph neural network according to claim 1, characterized in that: When preparing training samples, the generation method of gridded meteorological training information in the training samples includes: Acquire a reanalysis numerical model data set, and download reanalysis meteorological data corresponding to a target area and a target meteorological element from the reanalysis numerical model data set; The downloaded reanalysis meteorological data are spatially downscaled to generate gridded meteorological training information after spatial downscaling.
3. The meteorological element assimilation method based on encoder-decoder and graph neural network according to claim 2 is characterized in that: When constructing graph structure training data information, it includes: Acquire location information of a target observation site in a target area, wherein the location information of the target observation site includes the latitude and longitude of the target observation site; Based on the location information of the target observation site, the gridded meteorological training information is interpolated to the corresponding target observation site to generate gridded interpolated meteorological data after interpolation, and meteorological element values as node features are generated based on the gridded interpolated meteorological data.
4. The meteorological element assimilation method based on encoder-decoder and graph neural network according to claim 3 is characterized in that: When generating meteorological element values as node features, it also includes: Acquire site observation history data of each target observation site, and generate an observation list based on the acquired site observation history data; Based on the gridded interpolated meteorological data of each target observation station, an interpolation list of the target observation station is generated, wherein the length of the data corresponding to the interpolation list and the observation list is the same; For the interpolation list and observation list of the same target observation station, the corresponding elements in the interpolation list and the observation list are subtracted to generate an error list; determining an error probability distribution based on the error list; Random disturbance is added to the gridded interpolated meteorological data based on the error probability distribution to generate meteorological element values as node features after adding the random disturbance.
5. The meteorological element assimilation method based on encoder-decoder and graph neural network according to claim 1, characterized in that: When calculating the training loss, the loss functions used include the mean square error loss function.
6. The meteorological element assimilation method based on encoder-decoder and graph neural network according to claim 2, characterized in that: When performing spatial downscaling on reanalysis meteorological data, it includes: Determine the target spatial resolution; Processing the reanalysis meteorological data into a data volume recognized by the Weather Research and Forecasting Model WRF, and generating initial fields and boundary conditions based on the data volume; Based on the generated initial field and boundary conditions, the Weather Research and Forecasting Model (WRF) is used to generate gridded meteorological training information of the target spatial resolution.
7. The meteorological element assimilation method based on encoder-decoder and graph neural network according to claim 2 is characterized in that: When determining the error probability distribution, include: Based on the error list, count the maximum error value and the minimum error value; Between the maximum error value and the minimum error value, the error frequency is calculated based on the preset element interval threshold; Divide the calculated error frequency by the data length of the error list to obtain the frequency of each error; Based on the frequency of each error, the probability distribution of the error over all time is determined.
8. The meteorological element assimilation method based on encoder-decoder and graph neural network according to any one of claims 1 to 7, characterized in that: The target meteorological elements include temperature, humidity, air pressure or wind speed.
9. A meteorological element assimilation system based on encoder-decoder and graph neural network, characterized in that: Includes meteorological element assimilation equipment, among which, The meteorological element assimilation device assimilates the gridded meteorological forecast information of the target meteorological element in the target area and the graph structure observation data information constructed based on the target observation station in the target area based on the method described in any one of claims 1 to 8 above.
