Multi-scale spatial-temporal feature fusion and wind direction discriminator-based strong wind back calculation and prediction method
The method integrates multi-source meteorological data with a wind direction discriminator to enhance high wind prediction accuracy in complex terrain and monsoon climate areas by dynamically selecting relevant terrain elements, addressing the limitations of existing forecasting technologies.
Patent Information
- Application Number
- CN202411032500.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-12-31
AI Technical Summary
Existing gale forecasting technologies struggle to accurately predict high winds in complex terrain and monsoon climate areas due to low spatial resolution and dependence on historical data, failing to effectively integrate terrain information and adapt to climate change.
A high wind back-calculation and prediction method using multi-scale spatiotemporal feature fusion and a wind direction discriminator, integrating ECMWF forecast products, DEM data, ERA5 reanalysis data, and monsoon system data through a spatiotemporal memory network and graph convolution network to dynamically select relevant terrain elements and improve prediction accuracy.
Enhances the precision and efficiency of high wind prediction by dynamically selecting terrain elements based on wind direction, reducing noise and improving model generalization, thus providing accurate and reliable high wind forecasts.
Smart Images
Figure CN119227853A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence intelligent meteorological technology, and in particular to a high wind back-calculation and prediction method of a multi-scale spatiotemporal feature fusion and a wind direction discriminator. Background Art
[0002] In recent years, with the rapid improvement of computing power and the rapid development of deep learning technology, artificial intelligence has made significant progress in the fields of perception, recognition and prediction. Although existing gale forecasting technologies, such as numerical weather forecast models and statistical forecasting methods, can predict gale events to a certain extent, they still have some limitations in complex terrain and monsoon climate areas: traditional numerical weather forecast models usually use low spatial resolution and it is difficult to accurately capture the details of the impact of terrain on gale; existing statistical forecasting methods are too dependent on historical observation data and it is difficult to cope with the evolution of gale characteristics caused by climate change; at the same time, existing methods usually cannot effectively utilize the impact of spatial information such as terrain on gale forecasting. Therefore, there is an urgent need for a model that can integrate multi-source dynamic meteorological data, and can dynamically select terrain elements related to prediction according to the characteristics of the gale meteorological field, remove the noise influence of irrelevant terrain, speed up the gale prediction efficiency of the entire study area, and improve the generalization ability of the model, so as to fill the gap in high-precision back calculation and prediction of gale in Guizhou, and bring new breakthroughs to meteorological prediction and climate research. Summary of the invention
[0003] In view of the above existing problems, the present invention is proposed.
[0004] Therefore, the technical problem to be solved by the present invention is to provide a high wind back-calculation and prediction method based on multi-scale spatiotemporal feature fusion and wind direction discriminator in view of the shortcomings of the existing technology, aiming to solve the limitations of the existing high wind forecasting technology in complex terrain and monsoon climate areas. The method provides a method of using advanced deep learning technology to dynamically integrate multi-source meteorological data, optimize the accuracy and generalization ability of the high wind prediction model, and provide an effective solution for high-precision back-calculation and prediction of high winds.
[0005] In order to solve the above technical problems, the present invention provides the following technical solutions, a high wind back calculation and prediction method of multi-scale spatiotemporal feature fusion and wind direction discriminator, comprising:
[0006] Based on the ECMWF forecast product in the area to be predicted, obtain the DEM data in the area to be predicted, the strong wind meteorological data of ERA5 reanalysis within the preset historical range, and the monsoon system data; construct a spatiotemporal memory network layer to extract the spatial characteristics of the area around the grid point to be predicted, learn the change law of the previous time step of the grid point to be predicted, pass the spatiotemporal characteristics to the graph convolution network, construct a graph convolution layer, capture the seasonal and long-term wind direction change trend, and output high-dimensional meteorological field characteristics; construct a wind direction discriminator, calculate the wind direction angle according to the meteorological field characteristics, and classify the wind direction angles, dynamically select appropriate DEM element data according to the wind direction category, and use the graph convolution network to extract the characteristics of the terrain elements; construct a feature splicing module to splice the meteorological field features obtained by the graph convolution layer and the terrain features obtained by the wind direction discriminator Then, a multi-scale spatiotemporal feature map is generated as the input feature of the HRNet network; a prediction module is constructed to reconstruct the strong wind meteorological field image of the grid point to be predicted in the future. The DEM data, the strong wind meteorological grid data of ERA5 reanalysis from T-15 to T, and the monsoon system grid data of ERA5 reanalysis from T-45 to T-30 are used as input data, and the ERA5 actual data at T+5 is used as the label to build the training data set of the IMSTFFWD-GRP model. According to the training data set of the IMSTFFWD-GRP model, the IMSTFFWD-GRP model loss is calculated, the network parameters are optimized, and the architecture model is obtained. The trained model is applied to the grid points in the area for sliding prediction to obtain the strong wind prediction results for each grid point in the future.
[0007] As a preferred solution of the high wind back calculation and prediction method of the multi-scale spatiotemporal feature fusion and wind direction discriminator of the present invention, wherein: the monsoon system data includes selecting the area within the range of 100° to 145° east longitude and 15° to 50° north latitude as the monsoon system, and the area covering the East Asian monsoon zone is defined as the monsoon system to be considered;
[0008] The resolution of the DEM data is 1 km, including altitude, slope, slope curvature, slope aspect and flow direction; the spatial resolution of the ERA5 reanalysis data is 0.25°×0.25°, and the temporal resolution is 1 hour. The strong wind meteorological data include 2m dew point temperature, 2m temperature, cloud area fraction, 10m meridional wind component, 10m neutral meridional wind component, 10m zonal wind component, 10m neutral zonal wind component, post-processed 10m gust, instantaneous 10m gust, relative humidity, absolute humidity, post-processed 2m minimum temperature, post-processed 2m maximum temperature, average surface latent heat flux, average surface sensible heat flux, surface temperature and surface pressure; the monsoon system data include 10m meridional wind component, 10m zonal wind component, 100m meridional wind component, 100m zonal wind component, 2m temperature, relative humidity, absolute humidity, instantaneous large-scale surface precipitation, convective precipitation, cloud height, total cloud cover, surface pressure, and surface downward thermal radiation.
[0009] As an optimal solution for the high wind back calculation and prediction method of the multi-scale spatiotemporal feature fusion and wind direction discriminator described in the present invention, wherein: the construction of the spatiotemporal memory network layer includes constructing a spatiotemporal memory network layer, the network layer is composed of an extended convolutional memory network, which is used to process meteorological data in the East Asian monsoon region, extract the spatial characteristics of the area around the grid point to be predicted, learn the change law and spatiotemporal characteristics of the time step before the grid point to be predicted, and pass the learned features downward to the graph convolutional network; the extended convolutional memory network includes extended convolution, long short-term memory network and attention mechanism.
[0010] As a preferred solution of the high wind back calculation and prediction method of the multi-scale spatiotemporal feature fusion and wind direction discriminator described in the present invention, wherein: the dilated convolutional memory network includes, the input data is the meteorological grid data X∈R from T-15 to T in the East Asian monsoon region T×H×W×C , H and W are the spatial dimensions, and C is the number of features;
[0011] Spatial features are extracted by dilated convolution, using three different dilation rates r 1 、r 2 and r 3 Perform an expansion convolution operation. For each expansion rate, the output of the convolution kernel is expressed as:
[0012] For each time step t, a dilated convolution operation is performed on the feature map, and multi-scale spatial features are extracted using different dilation rates. The size of the convolution kernel is 3×3, and the convolution output of the dilation rate r is expressed as:
[0013]
[0014] Among them, X (t) is the meteorological data input at the tth time step, i and j are the current convolution operation in X(t) The position in the convolution kernel; k and l are the indices of the convolution kernel, traversing from 0 to 2 respectively; c is the feature channel index; r·k and r·l are the convolution kernel positions under the influence of the current expansion rate r; X (t) (i+r·k,j+r·l,c) represents the input feature map index; W r (k,l) is the convolution kernel weight with dilation rate r at the corresponding position, and the size is 3×3; is the output feature value of the dilated convolution operation at position (i, j); for different dilation rates, when r = 1, the convolution output is When r=2, the convolution output is When r=3, the convolution output is
[0015] The time features are extracted by dilated convolution, and the feature map after spatial convolution is dilated and convolved in the time dimension. The convolution output with dilation rate d is expressed as:
[0016]
[0017] in, is the input feature value at position (i, j) at time step td·p; W d (p) is the weight of the temporal convolution kernel with a dilation rate of d; p is the index in the temporal dimension, from 0 to P-1; for different dilation rates: when d = 1, the convolution output is Z 1 (t,i,j,c); when d=2, the convolution output is Z 2 (t,i,j,c); when d=3, the convolution output is Z 3 (t,i,j,c);
[0018] Concatenate the outputs of spatial and temporal dilated convolutions to form a comprehensive feature representation:
[0019]
[0020] The concatenated features are used as the input of LSTM, and the temporal variation patterns are captured through LSTM units. The specific formula is:
[0021] h t ,c t =LSTM(F (t) ,h t-1 ,c t-1 )
[0022] The attention mechanism is used to adaptively focus on spatiotemporal features. The specific formula is:
[0023] a t =softmax(W a ·ht +b a )
[0024] o t =a t ⊙h t
[0025] Among them, a t represents the attention weight, which is the weight of each feature at time step t; W a Represents the weight matrix of the attention mechanism, which is used to t Perform linear transformation; b a Represents the bias vector of the attention mechanism, which is used to translate the result after linear transformation; o t represents the output feature, which is the hidden state after combining the attention weight; ⊙ represents the element-level multiplication, which means that the attention weight a t Applied element-wise to the hidden state h t ; Features of different scales are integrated and passed downward. The specific formula is:
[0026]
[0027] Among them, F ST It represents the fused features and the result after fusion of features of different scales.
[0028] As a preferred solution of the high wind back calculation and prediction method of the multi-scale spatiotemporal feature fusion and wind direction discriminator described in the present invention, wherein: the construction of the graph convolution layer includes a graph convolution network GCN with different sizes of convolution kernels, which is used to process the meteorological grid data and transmitted features of the monsoon system, capture seasonal and long-term wind direction change trends, learn the long-term spatiotemporal influence of the historical data of the monsoon system on the current features, adaptively pay attention to the spatiotemporal features and perform feature fusion, and output high-dimensional meteorological field features;
[0029] Assume that the input is the meteorological grid data X of the monsoon system t And the transferred feature F ST , using convolution kernels of different sizes k i , k i =3, 5, 7; perform graph convolution operation, the specific formula is:
[0030]
[0031] Among them, H t (k i ) is the convolution kernel k i The extracted features; σ is the activation function; N(i) represents the set of neighbor nodes of node i; c ij is the normalized coefficient between nodes i and j; and are weight matrices and bias terms respectively; the feature F ST And the feature H extracted by multi-scale graph convolution t (k i ) for fusion, the specific formula is:
[0032]
[0033] Among them, ⊕ represents the feature fusion operation, which uses the attention mechanism to adaptively focus on spatiotemporal features. The specific formula is:
[0034]
[0035]
[0036] in, is the attention weight, It is the output feature after applying the attention weight, which combines the features extracted by different convolution kernel sizes. The specific formula is:
[0037]
[0038] F GCN Represents the fused features.
[0039] As a preferred scheme of the high wind back-calculation and prediction method of the multi-scale spatiotemporal feature fusion and wind direction discriminator described in the present invention, wherein: the construction of the wind direction discriminator includes, the wind direction discriminator converts the longitude U and latitudinal V wind components in the meteorological field characteristics into wind direction angles, and classifies the wind direction angles into specific wind direction categories, the specific wind direction categories include N, NE, E, SE, S, SW, W, NW, constructs a custom wind direction selection layer, dynamically selects the corresponding windward slope terrain element data according to the wind direction category, removes the terrain data of the leeward slope, and extracts the characteristics of the terrain elements selected each time through a graph convolutional network.
[0040] As a preferred solution of the high wind back calculation and prediction method of the multi-scale spatiotemporal feature fusion and wind direction discriminator described in the present invention, wherein: the wind direction discriminator includes: GCN The longitude U and zonal V wind components are used to calculate the wind direction angle. The specific formula is:
[0041] θ=arctan2(V,U)
[0042] Among them, θ is the wind direction angle. The calculated wind direction angle θ is classified into a specific wind direction category, and a custom wind direction selection layer is constructed. The corresponding windward slope terrain element data is dynamically selected according to the wind direction category, and the leeward slope terrain data is removed. According to the wind direction category, the corresponding windward slope terrain element data is selected from the acquired terrain data, and the non-corresponding leeward slope terrain data is removed. The features of the terrain elements selected each time are extracted through the graph convolutional network.
[0043] As a preferred solution of the high wind back-calculation and prediction method of the multi-scale spatiotemporal feature fusion and wind direction discriminator described in the present invention, the T moment corresponds to a time span of 1 day, the original time resolution of the acquired ERA5 reanalysis high wind meteorological data and monsoon system data is 1 hour, which is converted into 1 day's data after aggregation processing.
[0044] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of a high wind back-calculation and prediction method of multi-scale spatiotemporal feature fusion and wind direction discriminator are implemented.
[0045] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of a high wind back-calculation and prediction method of multi-scale spatiotemporal feature fusion and wind direction discriminator are implemented.
[0046] Beneficial effects of the present invention: The present invention designs a high wind back-calculation and prediction method based on multi-scale spatiotemporal feature fusion and wind direction discriminator. The method can integrate the multi-scale spatiotemporal features of the high wind meteorological field, and based on the learned related element features. Through the wind direction discriminator, the wind direction change trend of each grid point at each moment is learned, and the windward slope terrain element data with a greater impact on the wind speed is dynamically selected, and the leeward slope terrain element data with a small impact is removed. Traditional high wind prediction models usually blindly introduce terrain element data. Although the terrain has a certain impact on the meteorology, these elements cannot be used accurately, which will only increase the training time and complexity of the model. The wind direction discriminator in the present invention effectively improves the high wind prediction accuracy of the grid points to be predicted and the interpretability of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0048] Figure 1A schematic flow chart of a method for back-calculating and predicting strong winds using a multi-scale spatiotemporal feature fusion and a wind direction discriminator provided in accordance with an embodiment of the present invention.
[0049] Figure 2 A schematic diagram of the overall structure of a network model of a high wind back-calculation and prediction method for multi-scale spatiotemporal feature fusion and wind direction discriminator provided in one embodiment of the present invention.
[0050] Figure 3 A schematic diagram of the dilated convolutional memory network layer structure of a high wind back-calculation and prediction method for multi-scale spatiotemporal feature fusion and wind direction discriminator provided for one embodiment of the present invention.
[0051] Figure 4 A schematic diagram of the graph convolutional layer structure of a high wind back calculation and prediction method for multi-scale spatiotemporal feature fusion and wind direction discriminator provided in one embodiment of the present invention.
[0052] Figure 5 A prediction diagram of a high wind back calculation and prediction method that combines multi-scale spatiotemporal features with a wind direction discriminator provided by an embodiment of the present invention.
[0053] Figure 6 An ERA5 real-time data diagram of a high wind back-calculation and prediction method using multi-scale spatiotemporal feature fusion and wind direction discriminator provided in one embodiment of the present invention.
[0054] Figure 7 A forecast chart of a CMA (China Meteorological Administration) wind speed forecast product of a high wind back calculation and prediction method that combines multi-scale spatiotemporal features with a wind direction discriminator provided by an embodiment of the present invention.
[0055] (Note: The length of the black arrow in the wind field diagram represents the wind speed, and the direction of the arrow represents the wind direction) DETAILED DESCRIPTION
[0056] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.
[0057] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0058] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.
[0059] The present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.
[0060] At the same time, in the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "upper, lower, inner and outer" are based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention. In addition, the terms "first, second or third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.
[0061] In the present invention, unless otherwise clearly specified and limited, the terms "install, connect, connect" should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection or an integral connection; it can also be a mechanical connection, an electrical connection or a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0062] Example 1, reference Figure 1-Figure 7 , which is the first embodiment of the present invention, provides a high wind back-calculation and prediction method using multi-scale spatiotemporal feature fusion and wind direction discriminator, including:
[0063] A method for gale recalculation and prediction based on multi-scale spatiotemporal feature fusion and wind direction discriminator (IMSTFFWD-GRP) is disclosed. The network structure of the method is as follows: Figure 2 As shown in the figure, the network structure includes input data, feature fusion module (expanded convolutional memory network layer, graph convolution layer, wind direction discriminator and feature splicing module), prediction module and output data.
[0064] See also Figure 1 and Figure 2 The high wind back-calculation and prediction method based on multi-scale spatiotemporal feature fusion and wind direction discriminator comprises the following steps:
[0065] S1, based on the ECMWF forecast products in the area to be predicted, obtain the DEM data in the area to be predicted, the ERA5 reanalysis wind meteorological data within the preset historical range, and the monsoon system data.
[0066] The monsoon system is defined as follows: Guizhou is located in the East Asian monsoon region, and the area within the range of 100° to 145° east longitude and 15° to 50° north latitude is selected as the monsoon system, that is, the area covering Guizhou is defined as the monsoon system that needs to be considered.
[0067] The DEM data, ERA5 reanalysis strong wind meteorological data and monsoon system data have the following characteristics: the resolution of the DEM data is 1 km, and it contains five elements, namely, altitude, slope, slope curvature, slope direction and flow direction. The spatial resolution of ERA5 reanalysis data is 0.25°×0.25°, and the temporal resolution is 1 hour. The strong wind meteorological data include 2m dew point temperature, 2m temperature, cloud area fraction, 10m meridional wind component, 10m neutral meridional wind component, 10m zonal wind component, 10m neutral zonal wind component, post-processed 10m gust, instantaneous 10m gust, relative humidity, absolute humidity, post-processed 2m minimum temperature, post-processed 2m maximum temperature, average surface latent heat flux, average surface sensible heat flux, surface temperature and surface pressure, a total of 17 elements. The monsoon system data include 10m meridional wind component, 10m zonal wind component, 100m meridional wind component, 100m zonal wind component, 2-meter temperature, relative humidity, absolute humidity, instantaneous large-scale surface precipitation, convective precipitation, cloud height, total cloud cover, surface pressure, and surface downward thermal radiation, a total of 13 elements.
[0068] S2, constructs a spatiotemporal memory network layer, which is composed of a dilated convolutional memory network (DCLSTM-Attention) and is used to process meteorological data in the East Asian monsoon region, extract the spatial features of the area around the grid point to be predicted, learn the change law of the time step before the grid point to be predicted, adaptively focus on the spatiotemporal features that are important to the prediction results, and pass the learned features down to the graph convolutional network in step 3.
[0069] The dilated convolutional memory network (DCLSTM-Attention) combines dilated convolution, long short-term memory network and attention mechanism. The specific process of the network is as follows: First, the meteorological grid data from T-15 to T in the East Asian monsoon region is input, and the wind direction change pattern at different scales is captured through dilated convolution. Then, combining time and space information, the temporal change pattern is captured through LSTM units. Then, the attention mechanism is used to adaptively focus on important spatiotemporal features. Finally, the features of different scales are fused and passed down. The specific operations of the DCLSTM-Attention are as follows:
[0070] The input data is the meteorological grid data from T-15 to T in the East Asian monsoon region. H and W are spatial dimensions, and C is the number of features (17 elements). The specific operation of the dilated convolution is as follows:
[0071] Dilated convolution extracts spatial features: using three different dilation rates r 1 、r 2 and r 3 To perform dilated convolution operations, they are 1, 2, and 3 respectively. For each dilation rate, the output of the convolution kernel can be expressed as:
[0072] For each time step t, the feature map is dilated and convolved, using different dilation rates to extract multi-scale spatial features. The size of the convolution kernel is 3×3, so the convolution output for the dilation rate r can be expressed as:
[0073]
[0074] Among them, X (t) is the meteorological data input at the tth time step, i and j are the current convolution operation in X (t) k and l are the indices of the convolution kernel, traversing from 0 to 2 respectively; c is the feature channel index; r·k and r·l are the positions of the convolution kernel under the influence of the current expansion rate r; X (t) (i+r·k,j+r·l,c) can represent the input feature map index; W r (k,l) is the convolution kernel weight with a dilation rate of r at the corresponding position and a size of 3×3. is the output feature value of the dilated convolution operation at position (i, j). For different dilation rates: when r = 1, the convolution output is When r=2, the convolution output is When r=3, the convolution output is
[0075] Dilated convolution extracts temporal features: Dilated convolution is performed on the feature map after spatial convolution in the time dimension. The convolution output with a dilation rate of d can be expressed as:
[0076]
[0077] in, W is the input feature value at position (i, j) at time step td·p, where Yr is the feature map obtained by spatial dilation convolution. d (p) is the weight of the temporal convolution kernel with a dilation rate of d. p is an index in the temporal dimension, from 0 to P-1, where P(=3) is the size of the temporal convolution kernel. For different dilation rates: When d=1, the convolution output is Z 1 (t,i,j,c); when d=2, the convolution output is Z 2 (t,i,j,c); when d=3, the convolution output is Z 3 (t,i,j,c).
[0078] Concatenate the outputs of spatial and temporal dilated convolutions to form a comprehensive feature representation:
[0079]
[0080] The concatenated features are used as the input of LSTM, and the temporal variation patterns are captured through LSTM units. The specific formula is:
[0081] h t ,c t =LSTM(F (t) ,h t-1 ,c t-1 )
[0082] The attention mechanism is used to adaptively focus on important spatiotemporal features. The specific formula is:
[0083] a t =softmax(W a ·h t +b a )
[0084] o t =a t ⊙h t
[0085] Among them, a t represents the attention weights, which are the weights of each feature at time step t; W a Represents the weight matrix (learnableweightmatrix) of the attention mechanism, which is used to tPerform linear transformation; b a Represents the bias vector of the attention mechanism, which is used to translate the result after linear transformation; o t represents the output features, the hidden state after combining the attention weights; ⊙ represents element-wise multiplication, which means that the attention weights a t Applied element-wise to the hidden state h t . Features of different scales are fused and passed downward. The specific formula is:
[0086]
[0087] Among them, F ST It represents the fused features and the result after fusion of features of different scales.
[0088] S3, build a graph convolution layer, which contains graph convolutional networks (GCN) with different convolution kernel sizes, to process the meteorological grid data of the monsoon system and the features passed down in step 2, capture seasonal and long-term wind direction change trends, learn the long-term spatiotemporal impact of the monsoon system data one month ago on the current features, adaptively focus on key spatiotemporal features and perform feature fusion, and output high-dimensional meteorological field features;
[0089] The operation of the GCN is as follows:
[0090] Assume that the input is the meteorological grid data Xt of the monsoon system and the feature F passed down in step 2 ST , using convolution kernels of different sizes k i (=3, 5, 7) to perform graph convolution operation. The specific formula is:
[0091]
[0092] Among them, H t (k i ) is the feature extracted using convolution kernel ki. σ is the activation function. N(i) represents the set of neighbor nodes of node i. ij is the normalized coefficient between nodes i and j. and are the weight matrix and the bias term respectively. The feature F passed down from step 2 ST And the feature H extracted by multi-scale graph convolution t (k i ) for fusion, the specific formula is:
[0093]
[0094] Among them, ⊕ represents the feature fusion operation. The attention mechanism is used to adaptively focus on key spatiotemporal features. The specific formula is:
[0095]
[0096]
[0097] in, is the attention weight, is the output feature after applying the attention weight. The features extracted by different convolution kernel sizes are fused together. The specific formula is:
[0098]
[0099] F GCN Represents the fused features.
[0100] S4, builds a wind direction discriminator, calculates the wind direction angle according to the characteristics of the meteorological field, and classifies the wind direction angle into a specific wind direction category. Dynamically select appropriate DEM feature data according to the wind direction category, and then use the graph convolutional network to extract the features of the terrain elements.
[0101] The wind direction discriminator can reduce data redundancy and speed up model training. The wind direction discriminator converts the longitude (U) and latitudinal (V) wind components in the meteorological field characteristics into wind direction angles, and classifies the wind direction angles into specific wind direction categories (such as north wind, south wind, northwest wind, etc.). Next, a custom wind direction selection layer (WindDirectionDEMSelector) is constructed to dynamically select the corresponding windward slope terrain element data according to the wind direction category and remove the terrain data of the leeward slope. Finally, the features of the terrain elements selected each time are extracted through a graph convolutional network. The details of the wind direction discriminator are as follows:
[0102] According to F GCN The wind direction angle is calculated by using the longitude (U) and latitudinal (V) wind components. The specific formula is:
[0103] θ=arctan2(V,U)
[0104] Among them, θ is the wind direction angle. The calculated wind direction angle θ is classified into 8 specific main wind direction categories (N, NE, E, SE, S, SW, W, NW). Construct a custom wind direction selection layer (WindDirectionDEMSelector) to dynamically select the corresponding windward slope terrain feature data according to the wind direction category and remove the leeward slope terrain data. The specific steps are as follows: First, according to the wind direction category, the corresponding windward slope terrain feature data is selected from the acquired terrain data, and irrelevant leeward slope terrain data is removed to reduce data redundancy. Then, the feature F of each selected terrain element is extracted through the graph convolutional network. DEM .
[0105] S5, construct a feature splicing module, splice the meteorological field features in step 3 and the terrain features in step 4, and generate a multi-scale spatiotemporal feature map as the input feature of the HRNet network.
[0106] The features of the splicing module are as follows:
[0107] The extracted terrain feature F DEM and high-dimensional meteorological field characteristics F GCN Combined to generate the comprehensive feature representation F for high wind forecasting combined .
[0108] S6, constructs a prediction module, which uses a high-resolution neural network (HRNet). The network maintains parallel convolution branches at multiple scales and exchanges information at the same time, and finally reconstructs the strong wind meteorological field image at the future moment of the grid point to be predicted.
[0109] The details of the HRNet are as follows:
[0110] The comprehensive feature is represented by F combined Input to HRNet, HRNet contains multiple stages, each stage has multiple parallel convolution paths, which process features of different resolutions. The output features of each stage are communicated through a cross-scale feature fusion mechanism to ensure that features of different resolutions can pass information to each other. In each parallel branch of each stage, standard convolution operations (3×3 convolutions) are applied to extract features. Features of different resolutions are fused through bilinear interpolation and downsampling. At the end of each stage, multi-scale feature fusion is used to obtain richer and more diverse feature representations. The multi-scale features output by all stages are upsampled and fused to generate a high-resolution feature map F. prediction .
[0111] S7, using DEM data, ERA5 reanalysis gale weather grid data from T-15 to T, and ERA5 reanalysis monsoon system grid data from T-45 to T-30 as input data, and ERA5 actual data at T+5 as labels, build the training data set of the IMSTFFWD-GRP model.
[0112] The T moment corresponds to a time span of 1 day. The original time resolution of the ERA5 reanalysis wind weather data and monsoon system data is 1 hour, which is converted into 1 day data after aggregation processing.
[0113] S8, based on the data set constructed in step 7, calculate the IMSTFFWD-GRP model loss, optimize the network parameters, and obtain a model that conforms to the architecture of step 7.
[0114] S9, apply the model trained in step 8 to each grid point in the East Asian monsoon region for sliding prediction, and obtain the strong wind prediction result at each grid point at the future moment.
[0115] The experimental results presented in the present invention rely on a study area located between 24.5° and 29.5° north latitude and 103.5° and 109.5° east longitude. This rectangular area roughly covers the main part of Guizhou Province, China, including complex karst terrain and diverse climatic conditions. The selection of this area is based on its unique geographical and meteorological characteristics, making it an ideal place to study wind speed prediction models. The experimental results show that the method of the present invention has high accuracy and reliability in back-calculating and predicting high wind weather in this area.
[0116] The present invention uses data from the time period from February 2021 to February 2024 as the training set, and data from the time period from February 2023 to July 2023 as the test set.
[0117] In a specific embodiment of the present invention, the method of the present invention is used for gale prediction in Guizhou region, such as Figure 5 , Figure 6 , Figure 7 As shown in Figure 2, it can be observed that the predicted values of the existing CMA (China Meteorological Administration) wind speed forecast products (such as Figure 7 ) in comparison with ERA5 live data (as shown Figure 6 However, the predicted value of the high wind back calculation and prediction method based on multi-scale spatiotemporal feature fusion and wind direction discriminator of the present invention (as shown in Figure 2) is different in some areas of the wind speed and wind direction. Figure 5 The results show that the proposed gale prediction method can better fit the ERA5 real-time data and show a clear similar trend in temporal and spatial distribution. These observations further verify the advantages of the gale prediction method in Guizhou gale back-calculation and prediction, and provide support for its reliability in practical applications.
[0118] The present invention provides a high wind back calculation and prediction method based on multi-scale spatiotemporal feature fusion and wind direction discriminator. There are many methods and ways to implement the technical solution. The above is only a preferred implementation of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.
[0119] Embodiment 2, the second embodiment of the present invention, is different from the previous embodiment in that:
[0120] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.
[0121] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0122] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0123] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.
[0124] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.
[0125] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.
Claims
1. A high wind back-calculation and prediction method based on multi-scale spatiotemporal feature fusion and wind direction discriminator, characterized by: include, Based on the ECMWF forecast products in the area to be predicted, obtain DEM data in the area to be predicted, ERA5 reanalysis wind meteorological data within the preset historical range, and monsoon system data; Construct a spatiotemporal memory network layer to extract the spatial features of the area around the grid point to be predicted, learn the change law of the previous time step of the grid point to be predicted, pass the spatiotemporal features to the graph convolution network, construct a graph convolution layer, capture the seasonal and long-term wind direction change trends, and output high-dimensional meteorological field features; Construct a wind direction discriminator, calculate the wind direction angle according to the characteristics of the meteorological field, classify the wind direction angle, dynamically select DEM feature data according to the wind direction category, and use the graph convolutional network to extract the characteristics of terrain elements; Construct a feature splicing module to splice the meteorological field features obtained by the graph convolution layer and the terrain features obtained by the wind direction discriminator to generate a multi-scale spatiotemporal feature map as the input feature of the HRNet network; A prediction module was constructed to reconstruct the gale meteorological field image of the grid point to be predicted in the future. The DEM data, the gale meteorological grid data of ERA5 reanalysis from T-15 to T, and the monsoon system grid data of ERA5 reanalysis from T-45 to T-30 were used as input data. The ERA5 actual data at T+5 was used as the label. The training data set of the IMSTFFWD-GRP model was built. According to the training data set of the IMSTFFWD-GRP model, the IMSTFFWD-GRP model loss was calculated, the network parameters were optimized, and the architecture model was obtained. The trained model was applied to the grid points in the area for sliding prediction to obtain the gale prediction results for each grid point in the future.
2. The high wind back-calculation and prediction method of multi-scale spatiotemporal feature fusion and wind direction discriminator as claimed in claim 1, characterized in that: The monsoon system data includes selecting the area within the range of 100° to 145° east longitude and 15° to 50° north latitude as the monsoon system, and the area covering the East Asian monsoon region is defined as the monsoon system to be considered; The resolution of the DEM data is 1 km, including altitude, slope, slope curvature, slope aspect and flow direction; the spatial resolution of the ERA5 reanalysis data is 0.25°×0.25°, and the temporal resolution is 1 hour. The strong wind meteorological data include 2m dew point temperature, 2m temperature, cloud area fraction, 10m meridional wind component, 10m neutral meridional wind component, 10m zonal wind component, 10m neutral zonal wind component, post-processed 10m gust, instantaneous 10m gust, relative humidity, absolute humidity, post-processed 2m minimum temperature, post-processed 2m maximum temperature, average surface latent heat flux, average surface sensible heat flux, surface temperature and surface pressure; the monsoon system data include 10m meridional wind component, 10m zonal wind component, 100m meridional wind component, 100m zonal wind component, 2m temperature, relative humidity, absolute humidity, instantaneous large-scale surface precipitation, convective precipitation, cloud height, total cloud cover, surface pressure, and surface downward thermal radiation.
3. The high wind back-calculation and prediction method of multi-scale spatiotemporal feature fusion and wind direction discriminator as claimed in claim 2 is characterized by: The construction of the spatiotemporal memory network layer includes constructing a spatiotemporal memory network layer, the network layer is composed of a dilated convolutional memory network, which is used to process meteorological data in the East Asian monsoon region, extract spatial features of the area around the grid point to be predicted, learn the change law and spatiotemporal features of the time step before the grid point to be predicted, and pass the learned features downward to the graph convolutional network; the dilated convolutional memory network includes dilated convolution, long short-term memory network and attention mechanism.
4. The method for high wind back-calculation and prediction using multi-scale spatiotemporal feature fusion and wind direction discriminator as claimed in claim 3, characterized in that: The dilated convolutional memory network includes: the input data is the meteorological grid data X∈R from time T-15 to time T in the East Asian monsoon region T×H×W×C , H and W are the spatial dimensions, and C is the number of features; The spatial features are extracted by dilated convolution. Three different dilation rates r1, r2 and r3 are used for dilated convolution operation. For each dilation rate, the output of the convolution kernel is expressed as: For each time step t, a dilated convolution operation is performed on the feature map, and multi-scale spatial features are extracted using different dilation rates. The size of the convolution kernel is 3×3, and the convolution output of the dilation rate r is expressed as: Among them, X (t) is the meteorological data input at the tth time step, i and j are the current convolution operation in X (t) The position in the convolution kernel; k and l are the indices of the convolution kernel, traversing from 0 to 2 respectively; c is the feature channel index; r·k and r·l are the convolution kernel positions under the influence of the current expansion rate r; X (t) (i+r·k,j+r·l,c ) represents the input feature map index; W r (k,l) is the convolution kernel weight with dilation rate r at the corresponding position, and the size is 3×3; is the output feature value of the dilated convolution operation at position (i, j); for different dilation rates, when r = 1, the convolution output is When r=2, the convolution output is When r=3, the convolution output is The time features are extracted by dilated convolution, and the feature map after spatial convolution is dilated and convolved in the time dimension. The convolution output with dilation rate d is expressed as: in, is the input feature value at position (i, j) at time step td·p; W d (p) is the weight of the temporal convolution kernel with a dilation rate of d; p is an index in the temporal dimension, from 0 to P-1; for different dilation rates: when d=1, the convolution output is Z1(t,i,j,c); when d=2, the convolution output is Z2(t,i,j,c); when d=3, the convolution output is Z3(t,i,j,c); Concatenate the outputs of spatial and temporal dilated convolutions to form a comprehensive feature representation: The concatenated features are used as the input of LSTM, and the temporal variation patterns are captured through LSTM units. The specific formula is: h t ,c t =LSTM(F (t) ,h t-1 ,c t-1 ) The attention mechanism is used to adaptively focus on spatiotemporal features. The specific formula is: a t =softmax(W a ·h t +b a ) the t =a t ⊙h t Among them, a t represents the attention weight, which is the weight of each feature at time step t; W a Represents the weight matrix of the attention mechanism, which is used to t Perform linear transformation; b a Represents the bias vector of the attention mechanism, which is used to translate the result after linear transformation; o t represents the output feature, which is the hidden state after combining the attention weight; ⊙ represents the element-level multiplication, which means that the attention weight a t Applied element-wise to the hidden state h t ; Features of different scales are integrated and passed downward. The specific formula is: Among them, F ST It represents the fused features and the result after fusion of features of different scales.
5. The method for high wind back-calculation and prediction using multi-scale spatiotemporal feature fusion and wind direction discriminator as claimed in claim 4, characterized in that: The constructed graph convolution layer includes a graph convolution network GCN with different sizes of convolution kernels, which is used to process the meteorological grid data of the monsoon system and the transmitted features, capture the seasonal and long-term wind direction change trends, learn the long-term spatiotemporal impact of the historical data of the monsoon system on the current features, adaptively focus on the spatiotemporal features and perform feature fusion, and output high-dimensional meteorological field features; Assume that the input is the meteorological grid data X of the monsoon system t And the transferred feature F ST , using convolution kernels of different sizes k i , k i =3, 5, 7; perform graph convolution operation, the specific formula is: Among them, H t (k i ) is the convolution kernel k i The extracted features; σ is the activation function; N(i) represents the set of neighbor nodes of node i; c ij is the normalized coefficient between nodes i and j; and are weight matrices and bias terms respectively; the feature F ST And the feature H extracted by multi-scale graph convolution t (k i ) are fused, and the specific formula is: in, It represents the feature fusion operation, which uses the attention mechanism to adaptively focus on spatiotemporal features. The specific formula is: in, is the attention weight, It is the output feature after applying the attention weight, which combines the features extracted by different convolution kernel sizes. The specific formula is: F GCN Represents the fused features.
6. The high wind back-calculation and prediction method of multi-scale spatiotemporal feature fusion and wind direction discriminator as claimed in claim 5, characterized in that: The construction of the wind direction discriminator includes converting the longitude U and latitudinal V wind components in the meteorological field characteristics into wind direction angles, and classifying the wind direction angles into specific wind direction categories, wherein the specific wind direction categories include N, NE, E, SE, S, SW, W, and NW; constructing a custom wind direction selection layer; dynamically selecting corresponding windward slope terrain element data according to the wind direction category, removing the terrain data of the leeward slope, and extracting the characteristics of the terrain elements selected each time through a graph convolutional network.
7. The method for high wind back-calculation and prediction using multi-scale spatiotemporal feature fusion and wind direction discriminator as claimed in claim 6, characterized in that: The wind direction discriminator comprises: GCN The longitude U and zonal V wind components are used to calculate the wind direction angle. The specific formula is: θ=arctan2(V,U) Among them, θ is the wind direction angle. The calculated wind direction angle θ is classified into a specific wind direction category, and a custom wind direction selection layer is constructed. The corresponding windward slope terrain element data is dynamically selected according to the wind direction category, and the leeward slope terrain data is removed. According to the wind direction category, the corresponding windward slope terrain element data is selected from the acquired terrain data, and the non-corresponding leeward slope terrain data is removed. The features of the terrain elements selected each time are extracted through the graph convolutional network.
8. The high wind back-calculation and prediction method of multi-scale spatiotemporal feature fusion and wind direction discriminator as claimed in claim 7, characterized in that: The T moment corresponds to a time span of 1 day. The original time resolution of the ERA5 reanalysis gale meteorological data and monsoon system data is 1 hour, which is converted into 1 day's data after aggregation processing.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the high wind back-calculation and prediction method of the multi-scale spatiotemporal feature fusion and wind direction discriminator described in any one of claims 1 to 8 are implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the high wind back-calculation and prediction method of multi-scale spatiotemporal feature fusion and wind direction discriminator according to any one of claims 1 to 8 are implemented.