Multi-meteorological element variable forecasting method based on space-time adaptive kernel feature interaction network

Abstract

The invention discloses a multi-meteorological element variable forecasting method based on a space-time adaptive kernel feature interaction network, and belongs to the technical field of atmospheric science. The method specifically comprises the following steps: providing a branch cross-attention fusion strategy aiming at complex nonlinear interaction among a plurality of meteorological elements, and constructing a multi-element fusion network MFFN; secondly, in order to capture the relevance of different meteorological element features on a long distance and the difference between a large-scale region and a small-scale region, a novel large-kernel convolution (NLKC) is proposed to replace a conventional convolution operation in an original RNN, and a novel large-kernel recurrent neural network NLKRNN is constructed based on an NLK-LSTM unit; and finally, the MFFN and the NLKRNN jointly form a multi-meteorological element variable space-time prediction model. A multi-element meteorological data set is constructed by adopting four typical meteorological element data of temperature, dew point temperature, weft wind and warp wind, and a model is trained for space-time prediction of meteorological element variables of a target area. According to the invention, a more accurate temperature space-time prediction result can be obtained.

Description

Technical Field

[0001] The present invention belongs to the field of atmospheric science, and in particular, relates to a multi-meteorological element variable prediction method based on a spatio-temporal adaptive kernel feature interaction network. Background Art

[0002] With the continuous change of the global weather pattern, weather forecasting has become an important part of people's daily life.

[0003] Abnormal weather changes may trigger natural disasters such as droughts or typhoons, which have a serious impact on the ecological environment and social development. Global warming has led to an increase in temperature fluctuations and a significant change in the atmospheric circulation pattern, thus triggering regional temperature rises and falls. High-temperature heat wave weather has many adverse effects on people's lives, health, crop growth, energy utilization, economic development, and the ecological environment. In the wind energy industry, understanding future wind speeds helps optimize wind power generation. For aviation and shipping, accurate wind speed information can improve safety and efficiency. In addition, wind speed prediction also plays a role in disaster management, urban planning, and agriculture, providing basic data for decision-making in various fields and ensuring the operation of society and the benefits of people's lives.

[0004] Dew point temperature prediction is crucial for multiple fields; in agriculture, understanding the dew point temperature helps determine the optimal irrigation strategy. In meteorology, the dew point temperature is a key factor in predicting precipitation and heavy rain, and is crucial for disaster management.

[0005] The current mainstream weather forecasting method is based on traditional numerical weather prediction (NWP), which uses mathematical equations to describe atmospheric motion and physical processes and solves them through numerical simulation. However, this method has high computational resource requirements and mainly relies on parameterized numerical models, which are easily affected by the initial conditions and parameter selection, resulting in uncertainty and an increase in errors.

[0006] In recent years, with the rapid development of data-driven deep learning methods, deep learning-based weather forecasting methods are considered a powerful supplement to traditional methods. Related models for spatio-temporal sequence prediction based on deep learning are applied to the spatio-temporal prediction of meteorological elements, which can capture the spatio-temporal correlation and non-linear characteristics in meteorological data. By learning historical meteorological data, hidden patterns and laws in the climate system are discovered, and then the future climate change trend is predicted.

[0007] However, most current meteorological spatio-temporal predictions only identify it as a spatio-temporal sequence prediction problem and only perform autoregressive prediction for a single meteorological element, without considering the spatio-temporal characteristics of meteorological data in model construction. Therefore, there is an urgent need to invent a method that breaks through the limitation of the current meteorological prediction model's weak feature capture ability in capturing meteorological uniqueness from three aspects: spatio-temporal change difference, long-distance dependence relationship, and multi-element correlation. Summary of the Invention

[0008] In view of the above problems, the present invention proposes a multi-meteorological element variable forecasting method based on a spatio-temporal adaptive kernel feature interaction network, which couples multiple elements such as air temperature, dew point temperature, and wind to form a meteorological forecasting model, effectively captures the potential coupling relationships and spatio-temporal variation differences among various meteorological elements, thereby realizing the joint forecasting of multiple meteorological elements and improving the forecasting accuracy.

[0009] The multi-meteorological element variable forecasting method based on the spatio-temporal adaptive kernel feature interaction network includes the following steps:

[0010] Step 1: Collect ERA5-Land temperature, dew point temperature, zonal wind, and meridional wind data, construct a multi-meteorological element variable data set, perform normalization and set the forecasting duration, and divide the training set, validation set, and test set.

[0011] Step 2: Use the spatio-temporal adaptive kernel feature interaction network to construct a multi-meteorological element variable spatio-temporal prediction model;

[0012] The spatio-temporal adaptive kernel feature interaction network is divided into: a multi-element fusion network and a large kernel recurrent neural network NLKRNN;

[0013] First, at time T, given the original multi-element meteorological data X with a sequence length of T p use the multi-element feature fusion network to extract features and model the feature coupling relationships among the elements to obtain a fused feature sequence Then input it into the large kernel recurrent neural network NLKRNN for spatio-temporal dynamic modeling to generate a spatio-temporal sequence with a future time step of T f to complete the forecasting work. Among them, in the meteorological data

[0014] H represents the height of the multi-element meteorological data, and W represents the width of the multi-element meteorological data. C is the number of channels, and H f and W f are the height and width of the fused features. The sequence length T f and the time step T p take the same value. f The multi-element fusion network introduces a multi-branch cross-attention fusion strategy and is composed of N multi-element fusion blocks connected in series. Each multi-element fusion block is composed of 4 element branches and one fusion branch.

[0015] Each element branch inputs the corresponding meteorological observation data respectively:

[0016] Feature extraction is performed through a Layernorm layer, a convolutional layer, and two fully connected layers to obtain four different meteorological features.

[0017] The calculation formula for the feature branch is:

[0018]

[0019] and are both the original multi-element meteorological data input to the first (layer) multi-element fusion block; Layernorm is the layer normalization operation; Conv is the ordinary convolution operation; Linear is the fully connected operation; and are the corresponding meteorological features obtained;

[0020] The input to the fusion branch is: Four different meteorological features are concatenated in the channel dimension, combined with multi-branch cross-attention fusion, and finally the fusion feature is obtained

[0021] The calculation formula for the fusion branch is:

[0022]

[0023] where is the input to the fusion branch of the first (layer) multi-element fusion block, is the hidden feature extracted by the multi-elements through the network, represents dynamic attention, Silu is the activation function, is the final output after the first (layer) multi-element fusion block.

[0024] Take the output as the input data of the second (layer) multi-element fusion block. After the calculations of the feature branch and the fusion branch, the output feature is used as the input data of the third (layer) multi-element fusion block, and so on; at the same time, the fusion branch additionally introduces a gating mechanism, and each feature branch and the fusion branch are guided by cross-attention until the output feature F of the Nth (layer) multi-element fusion block is finally obtained t .

[0025] The overall architecture of the large kernel recurrent neural network NLKRNN follows the encoder-decoder architecture and is an autoregressive prediction model composed of multiple stacked NLK-LSTM recurrent units.

[0026] The NLK-LSTM recurrent unit introduces a new large-kernel convolutional module NLKC. The NLKC module consists of two channel convolutions (Conv1×1), one spatial local convolution (DWConv), and one spatial remote convolution (DW-D-Conv).

[0027] At the top is the channel convolution Conv1×1, which is used to expand the channel dimension and increase the channel capacity of the features, providing a richer feature representation for the subsequent DWConv and DW-D-Conv connected in sequence. DWConv is used to capture local spatial features, and DW-D-Conv introduces a dilation operation to expand the receptive field and capture remote spatial correlations. At the bottom is another channel convolution Conv1×1, which is used to restore the expanded channel dimension to the number of channels of the original input, ensuring that the features output by the module are consistent with the input data. This module combines local and global feature extraction capabilities, can model remote spatial dependencies at the same time, and realizes the efficient expression of features through multi-level convolution operations.

[0028] Step 3: Use the multi-meteorological element variable dataset to train the multi-meteorological element variable spatio-temporal prediction model.

[0029] Step 4: Input the multi-meteorological element variable data of the target area collected in real time into the trained multi-meteorological element variable spatio-temporal prediction model for meteorological prediction at future times.

[0030] First, normalize the temperature, dew point temperature, zonal wind, and meridional wind data of the target area collected in real time respectively, then fuse the four meteorological element data to construct a multi-element fusion dataset. Finally, perform an inverse normalization operation on the prediction results to obtain the prediction results of the four meteorological element variables in the final target area.

[0031] The advantages of the present invention are as follows:

[0032] (1) A multi-meteorological element variable forecasting method based on a spatio-temporal adaptive kernel feature interaction network according to the present invention, based on ERA5-Land historical temperature, dew point temperature, zonal wind, and meridional wind data, uses deep learning to learn the non-linear spatio-temporal variation law of historical meteorological element features and the coupling relationship between multi-meteorological element variables, constructs a meteorological forecasting model that can effectively learn the historical spatio-temporal variation law, and realizes the forecasting result of future multi-meteorological element variables based on historical multi-meteorological element variable data.

[0033] (2) A multi-meteorological element variable forecasting method based on a spatio-temporal adaptive kernel feature interaction network according to the present invention, the proposed meteorological forecasting model aims to effectively capture the potential coupling relationship between each meteorological element and its spatio-temporal variation difference, so as to realize the joint forecasting of multiple meteorological elements and improve the forecasting accuracy. Brief Description of the Drawings

[0034] Figure 1 It is a flowchart of the multi - meteorological element variable forecasting method based on the spatio - temporal adaptive kernel feature interaction network of the present invention;

[0035] Figure 2 It is the overall structure diagram of the spatio - temporal adaptive kernel feature interaction network of the present invention;

[0036] Figure 3 It is the structure diagram of the multi - element fusion network in the spatio - temporal adaptive kernel feature interaction network of the present invention;

[0037] Figure 4 It is the NLK - LSTM recurrent unit diagram of the present invention. Detailed implementation manners

[0038] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

[0039] In the prior art, accurately predicting air temperature, dew - point temperature, and wind speed is crucial for many aspects of society. It can help people make scientific decisions, reduce losses, improve efficiency, protect the environment, and enhance people's quality of life. It is crucial for mitigating weather - related risks and optimizing resource utilization.

[0040] Based on this, the present invention proposes a multi - meteorological element variable forecasting method based on the spatio - temporal adaptive kernel feature interaction network. First, collect ERA5 - Land reanalysis temperature, dew - point temperature, zonal wind, and meridional wind data. Then, construct a spatio - temporal adaptive kernel feature interaction network to extract features from multi - element meteorological data and learn the feature coupling relationships between elements. A multi - element fusion network (MFFN) is constructed. In this network, a multi - branch cross - attention fusion strategy is introduced, which consists of four element branches and one fusion branch. The four element features guide the generation of multi - element features through the cross - attention mechanism. Then, considering the characteristics of spatio - temporal variation differences and long - distance dependence relationships in meteorological data, a new recurrent unit NLK - LSTM is designed. This unit replaces the original conventional convolution operation in Predrnnv2 with the constructed new large - kernel convolution (NLKC), pays attention to spatio - temporal information and long - distance feature dependence relationships in large - scale and small - scale regions, and constructs a new large - kernel recurrent neural network NLKRNN based on this unit. The NLKRNN is used to model multi - meteorological element information. The present invention considers the coupling relationships between multi - meteorological element variables, and at the same time pays attention to the characteristics of spatio - temporal variation differences and long - distance dependence relationships in meteorological data, constructs a neural network that can fully learn the spatio - temporal variations of multi - meteorological element variables, and obtains more accurate meteorological prediction results.

[0041] Such as Figure 1As shown, the multi - meteorological element variable forecasting method based on the spatio - temporal adaptive kernel feature interaction network includes the following steps:

[0042] Step 1: Collect ERA5 - Land temperature, dew - point temperature, zonal wind, and meridional wind data, construct a multi - meteorological element variable data set, and perform normalization and set the forecasting duration. Then divide it into a training set, a validation set, and a test set.

[0043] Step 2: Use the spatio - temporal adaptive kernel feature interaction network to construct a multi - meteorological element variable spatio - temporal prediction model;

[0044] The spatio - temporal adaptive kernel feature interaction network is divided into: a multi - element fusion network and a large - kernel recurrent neural network NLKRNN;

[0045] First, at time T, given the original multi - element meteorological data X with a sequence length of T p use the multi - element feature fusion network to extract features and model the feature coupling relationship between elements to obtain a fused feature sequence Then input it into the large - kernel recurrent neural network NLKRNN for spatio - temporal dynamic modeling to generate a spatio - temporal sequence with a future time step of T f to complete the forecasting work. Complete the forecasting work.

[0046] Among them, for the meteorological data H represents the height of the multi - element meteorological data, and W represents the width of the multi - element meteorological data. C f is the number of channels, H f and W f are the height and width of the fused features. The sequence length T p and the time step T f take the same value.

[0047] The multi - element fusion network introduces a multi - branch cross - attention fusion strategy and is composed of N multi - element fusion blocks connected in series. Each multi - element fusion block consists of 4 element branches and one fusion branch.

[0048] Each element branch inputs the corresponding meteorological observation data respectively: All pass through a Layernorm layer, a convolutional layer, and two fully - connected layers for element feature extraction to obtain 4 different meteorological features.

[0049] The calculation formula of the element branch is:

[0050]

[0051] and All are the original multi-element meteorological data input to the first (layer) multi-element fusion block; Layernorm is the layer normalization operation; Conv is the ordinary convolution operation; Linear is the fully connected operation; and are the corresponding meteorological features obtained;

[0052] The input of the fusion branch is: 4 different meteorological features are concatenated in the channel dimension, combined with multi-branch cross-attention fusion, and finally the fusion feature is obtained

[0053] The calculation formula of the fusion branch is:

[0054]

[0055] where is the input of the fusion branch of the first (layer) multi-element fusion block, is the hidden feature extracted by the multi-elements through the network, represents dynamic attention, and Silu is the activation function, whose full name is Sigmoid gated linear unit. is the final output after the first (layer) multi-element fusion block.

[0056] Take the output as the input data of the second (layer) multi-element fusion block. After the calculation of the element branch and the fusion branch, the output feature obtained is used as the input data of the third (layer) multi-element fusion block, and so on; at the same time, a gating mechanism is additionally introduced between the fusion branch and the element branch, and each element branch is guided by cross-attention with the fusion branch until the output feature F of the Nth (layer) multi-element fusion block is finally obtained t , which enhances the information interaction between elements.

[0057] The overall architecture of the large kernel recurrent neural network NLKRNN follows the encoder-decoder architecture. In the design, it continues the basic architecture of PredRNN and is an autoregressive prediction model composed of multiple layers of stacked NLK-LSTM recurrent units.

[0058] This design makes full use of the advantages of NLK-LSTM, organically combines the learning of time and space features, thereby improving the model's ability to model complex spatio-temporal sequence data.

[0059] The NLK-LSTM recurrent unit introduces a new large kernel convolution module NLKC. The NLKC module consists of two-channel convolution (Conv1×1), one spatial local convolution (DWConv), and one spatial remote convolution (DW-D-Conv).

[0060] The top is a channel convolution Conv1×1 used to expand the channel dimension, increasing the channel capacity of the features and providing richer feature expressions for the subsequent DWConv and DW-D-Conv connected in sequence. DWConv is used to capture local spatial features. It focuses on extracting local detailed information from the input data and is suitable for processing local spatio-temporal dependencies. DW-D-Conv introduces a dilation operation to expand the receptive field and capture long-range spatial correlations, enabling the learning of feature connections within the global range. The bottom is another channel convolution Conv1×1 used to restore the expanded channel dimension to the number of channels of the original input, ensuring that the features output by the module are consistent with the input data. This module combines local and global feature extraction capabilities, can model long-range spatial dependencies at the same time, and achieves efficient feature expression through multi-level convolution operations, solving the problem of insufficient expression of local and global information in spatio-temporal feature extraction.

[0061] This design makes up for the deficiency of grouped convolution in extracting channel information interaction ability, and at the same time significantly enhances the ability to extract spatial features. Through the combined application of DWConv and DW-D-Conv, the NLKC module can establish a balance between local features and long-range features, comprehensively capturing the feature expressions in the data. In addition, the design of the NLKC module effectively avoids the problem of information loss. By preprocessing and postprocessing the features through the Conv1×1 convolutions at the top and bottom, the interaction ability of the features inside the module is enhanced.

[0062] Step 3: Use the multi-meteorological element variable dataset to train the spatio-temporal prediction model for multi-meteorological element variables;

[0063] Step 4: Input the multi-meteorological element variable data of the target area collected in real time into the trained spatio-temporal prediction model for multi-meteorological element variables to perform meteorological predictions for future moments.

[0064] First, normalize the temperature, dew point temperature, zonal wind, and meridional wind data of the target area collected in real time respectively, and then fuse the four meteorological element data to construct a multi-element fusion dataset. At the same time, when training the model, input the four meteorological element data simultaneously to complete the forecasting work of the four meteorological element variables at the same time. Finally, perform an inverse normalization operation on the prediction results to obtain the prediction results of the four meteorological element variables in the final target area.

[0065] Example:

[0066] In the embodiment of the present invention, ERA5-Land temperature, dew point temperature, zonal wind, and meridional wind data are collected to construct a multi-meteorological element variable dataset.

[0067] Data collection: The spatial resolution of the ground data in the ERA5 dataset is 0.1°×0.1°, and the time resolution is 1 hour. The data area is selected as east longitude (115.7° - 103.0°) and north latitude (36.7° - 24.0°). At each moment, a frame of 128*128 grid data is formed. The meteorological element variables are selected as temperature, dew point temperature, zonal wind, and meridional wind. The data with a time span from 2017 to 2023 is selected.

[0068] Data processing: To accelerate the training speed of the model and improve the accuracy of the model, the data is standardized to follow the standard normal distribution. At the same time, in the embodiment of the present invention, the data of the first 72 hours is used as the input of the model, and the data of the next 72 hours is used as the label data to complete the construction of the overall dataset.

[0069] Dataset division: The data from 2017 to 2021 is selected as the training set, the data in 2022 is used as the validation set, and the data in 2023 is used as the test dataset for evaluation.

[0070] Step 2, use the spatio-temporal adaptive kernel feature interaction network to construct a spatio-temporal prediction model for multiple meteorological element variables.

[0071] As Figure 2 shown, the spatio-temporal adaptive kernel feature interaction network of the present invention is mainly divided into two parts: a multi-element fusion network and a new large-kernel recurrent neural network - NLKRNN whose overall architecture follows the encoder-decoder architecture.

[0072] The multi-element fusion network introduces a multi-branch cross-attention fusion strategy and consists of N multi-element fusion blocks. As Figure 3 shown, each multi-element fusion block consists of 4 element branches and one fusion branch. Each element branch inputs the meteorological observation data of the corresponding element, and the input in the fusion branch is the multi-element meteorological data obtained by splicing the meteorological data of the four elements in the channel dimension At the same time, the fusion branch additionally introduces a gating mechanism, and each element branch and the fusion branch use cross-attention to guide the generation of the final fusion features.

[0073] The NLK-LSTM recurrent unit adopted by the new large-kernel recurrent neural network NLKRNN is as Figure 4As shown, the NLKC module is designed. After extracting features through NLKC, the spatio-temporal memory state M and cell state C of NLK-LSTM will first receive the state information of the previous moment, and selectively update or discard some information through the forget gate and input gate. Subsequently, NLK-LSTM concatenates the cell state and spatio-temporal memory state, and generates the final output through matrix operations. Through this mechanism, NLK-LSTM can effectively extract local features and long-range spatial correlations, realize the learning of global features, and dynamically adjust the retention and update of information through the gate structure. During the state transfer process, the cell state propagates along the time dimension, while the spatio-temporal memory state not only transfers in the spatial dimension, but also can perform information interaction across adjacent time steps and different spatial positions.

[0074] This bidirectional transfer mechanism enables the network to make full use of the dynamic adjustment ability of the spatio-temporal memory state in space and time, thereby enhancing the fusion of spatio-temporal features. The final output flexibly adjusts the spatio-temporal information to ensure that the model can capture long-term dependencies and comprehensively extract spatial features, thus realizing the accurate modeling of complex spatio-temporal dynamic changes.

[0075] The calculation formula of the NLK-LSTM module is as follows:

[0076]

[0077]

[0078] Among them, C and M represent the cell state and spatio-temporal memory state respectively, c, i, f represent the gate structures affecting the cell state, and c`, i`, f` are the gate structures affecting the spatio-temporal memory state. o represents the output value of the output gate, W is the weight matrix, X represents the input data at each moment, H is the final output value of the recurrent unit, and b is the bias value.

[0079] At the same time, NLK-LSTM introduces the decoupling loss of PredRNNv2, which is defined based on cosine similarity, and encourages the increments of the cell state and spatio-temporal memory state to remain orthogonal at any time step. This mechanism can give play to the unique advantages of both, improve the model's ability to model long-term and short-term dynamics, and avoid the cell state and spatio-temporal memory state from learning redundant features. The calculation formula of the decoupling loss is as follows:

[0080]

[0081] Where W decouple represents a 1×1 convolution shared by all NLK-LSTM units, represents calculating the dot product between the spatio-temporal memory unit and cell unit of each channel C, and They respectively represent calculating the L2 norms of the spatio-temporal memory unit and the cell unit for each channel C.

[0082] Step 3: Use the multi-meteorological element variable dataset in Step 1 to train the model.

[0083] Selection of the loss function during training: The loss function used is the cross-entropy loss function J, and the formula is as follows:

[0084]

[0085] where n is the total grid data volume of the input training data sequence, p i represents the predicted air temperature value at time i, and y i is the air temperature label value at time i. The loss function is used to evaluate the difference between the model prediction value and the true label, update the gradient through backpropagation, and optimize the model to make the prediction value approach the true value.

[0086] Training hyperparameters and optimization: The model is trained for 40 epochs on the training set using the Adam optimizer, the learning rate is set to 0.0001, the CosineLRScheduler learning rate adjustment strategy is used to adjust the learning rate, and the batch size is set to 16.

[0087] Step 4: Input the multi-meteorological element variable data constructed by fusing the currently collected temperature, dew point temperature, zonal wind, and meridional wind data into the trained multi-meteorological element variable prediction model to predict the temperature, dew point temperature, zonal wind, and meridional wind in each region at future times.

[0088] In the embodiment of the present invention, the collected data is standardized by mean and standard deviation, and then input into the trained multi-meteorological element variable prediction model to output the temperature, dew point temperature, zonal wind, and meridional wind of each grid in the target region at future times. Similarly, inverse standardization operations are performed on the predicted temperature, dew point temperature, zonal wind, and meridional wind to obtain the final spatio-temporal prediction results of temperature, dew point temperature, zonal wind, and meridional wind.

[0089] Evaluate the multi-meteorological element variable prediction model constructed by the present invention to prove the technical effect of the present invention.

[0090] Among them, the MAE (Mean Absolute Error), MSE (Mean Squared Error), and RMSE (Root Mean Squared Error) indicators are selected to evaluate the prediction results. The smaller the values of the three indicators, the better the prediction effect. At the same time, because the data input into the model is after standardized processing, therefore, after performing inverse standardization operations on the prediction results, the three indicators are used to evaluate the prediction results again.

[0091] The experimental data selects the temperature, dew point temperature, zonal wind and meridional wind in the target area of (115.7° - 103.0° E longitude, 36.7° - 24.0° N latitude) throughout the year 2023 for experiments.

[0092] Most of the previous methods simplify weather forecasting to the autoregressive prediction of single meteorological elements, failing to fully consider the complex coupling relationships among multi-element features, resulting in a significant bottleneck in forecasting accuracy and not considering the spatio-temporal variation differences and long-distance dependence relationships existing in meteorological data. Based on these problems, the multi-element feature fusion network proposed in the present invention introduces a multi-branch cross-fusion strategy, in which four element branches supplement the fused features through a cross-attention mechanism to guide the generation of the fused features, so as to fully learn the coupling relationships among meteorological elements. In addition, a recurrent neural network NLKRNN is constructed to model the spatio-temporal dynamic changes of meteorological elements. The designed new RNN recurrent unit NLKLSTM introduces a new large-kernel convolution NLKC, enabling NLKRNN to effectively model the complex spatio-temporal variation characteristics of meteorology, taking into account the variation differences between large-scale and small-scale regional features, and fully learning the long-distance dependence relationships of meteorological elements.

[0093] Except for the technical features described in the specification, they are all well-known technologies to those skilled in the art. The present invention omits the description of well-known components and well-known technologies to avoid redundancy and unnecessarily limit the present invention. The implementation manners described in the above embodiments do not represent all implementation manners consistent with the present application. On the basis of the technical solution of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A multi-meteorological element variable forecasting method based on a spatio-temporal adaptive kernel feature interaction network, characterized in that, It includes the following steps: Step 1: Collect ERA5-Land temperature, dew point temperature, zonal wind, and meridional wind data, construct a multi-meteorological element variable dataset, perform normalization and set the forecast duration, and divide the training set, validation set, and test set; Step 2: Use a spatio-temporal adaptive kernel feature interaction network to construct a spatio-temporal prediction model for multi-meteorological element variables; The spatio-temporal adaptive kernel feature interaction network is divided into: a multi-element fusion network and a large kernel recurrent neural network NLKRNN; First, at time T, given the original multi-element meteorological data X with a sequence length of T p perform feature extraction using a multi-element feature fusion network, model the feature coupling relationship between elements, and obtain a fused feature sequence Then input it into the large kernel recurrent neural network NLKRNN for spatio-temporal dynamic modeling to generate a spatio-temporal sequence with a future time step of T f to complete the forecasting work; ​ Among them, meteorological data wherein, H represents the height of the multi-element meteorological data, and W represents the width of the multi-element meteorological data; C f is the number of channels, H f and W f are the height and width of the fusion feature; Step 3: Use the multi-meteorological element variable dataset to train the spatio-temporal prediction model for multi-meteorological element variables; Step 4: Input the multi-meteorological element variable data of the target area collected in real time into the trained spatio-temporal prediction model for multi-meteorological element variables to perform meteorological prediction for future moments.

2. The method according to claim 1, wherein The multi-element fusion network introduces a multi-branch cross-attention fusion strategy and is composed of N multi-element fusion blocks connected in series. Each multi-element fusion block consists of 4 element branches and one fusion branch; Each element branch inputs corresponding meteorological observation data respectively: Element feature extraction is carried out through a Layernorm layer, a convolutional layer and two fully connected layers, and 4 different meteorological features are obtained; The input of the fusion branch is: 4 different meteorological features are concatenated in the channel dimension, combined with multi-branch cross-attention fusion, and finally the fusion features are obtained.

3. The method according to claim 2, characterized in that The calculation formula of the element branch is: T t 1 = Linear(Conv(Linear(Layernorm(T t 0 )))) V t 1 = Linear(Conv(Linear(Layernorm(V t 0 )))) T t 0 、 V t 0 and All are the original multi-factor meteorological data of the first (layer) multi-factor fusion block; Layernorm is the layer normalization operation; Conv is the ordinary convolution operation; Linear is the full connection operation; T t 1 、 V t 1 and is the corresponding meteorological characteristics obtained.

4. The method according to claim 2, wherein The calculation formula of the fusion branch is: F t h = Linear(Conv(Linear(Layernorm(F t 0 )))) SA = Silu(Linear(Layernorm(F t 0 ))) Among them, F t 0 is the input of the fusion branch of the first multi-element fusion block, is the hidden feature extracted by the multi-elements through the network, represents dynamic attention, and Silu is the activation function, is the final output after passing through the first multi-element fusion block; The output is used as the input data of the second multi-element fusion block. Through the calculations of the element branch and the fusion branch, the output features obtained are used as the input data of the third multi-element fusion block, and so on. At the same time, a gating mechanism is additionally introduced in the fusion branch, and each element branch and the fusion branch are guided by cross-attention until the output features F of the Nth multi-element fusion block are finally obtained t .

5. The method according to claim 1, wherein The large kernel recurrent neural network NLKRNN is composed of multiple layers of stacked NLK-LSTM recurrent units; The NLK-LSTM recurrent unit introduces a new large kernel convolution module NLKC. The NLKC module consists of two-channel convolution Conv1×1, one spatial local convolution DWConv, and one spatial remote convolution DW-D-Conv. At the top is the channel convolution Conv1×1 used to expand the channel dimension and increase the channel capacity of the features, providing a richer feature expression for the subsequent sequentially connected DWConv and DW-D-Conv; DWConv is used to capture local spatial features, and DW-D-Conv introduces a dilation operation to expand the receptive field and capture long-range spatial correlations; at the bottom is another channel convolution Conv1×1 used to restore the expanded channel dimension to the original input channel number to ensure that the features output by the module are consistent with the input data.

6. The method according to claim 1, wherein Step 4 refers to: first, perform normalization processing on the temperature, dew point temperature, zonal wind, and meridional wind data of the target area collected in real time, then fuse the four meteorological element data to construct a multi-element fusion dataset, and finally, perform an inverse normalization operation on the prediction result to obtain the prediction results of the four meteorological element variables in the final target area.

Images