Meteorological data simulation method and system based on generative adversarial network

Abstract

The invention provides a meteorological data simulation method and system based on a generative adversarial network, and the method comprises the steps: firstly obtaining a historical meteorological data set containing a time sequence relation and geographic region identification information, inputting the historical meteorological data set into a time sequence rule learning model, and extracting a meteorological evolution rule feature set; the method comprises the following steps: constructing a meteorological evolution rule feature set which covers time sequence relevance features of meteorological elements and geographical area relevance features, then constructing a conditional generative adversarial network model which comprises a generator module and a discriminator module, inputting the meteorological evolution rule feature set into the generator module, generating a virtual meteorological field data set with geographical area identification information, the virtual meteorological field data set comprises meteorological element distribution information consistent with the spatial-temporal characteristics of the historical meteorological data set, and finally outputting the virtual meteorological field data set to simulate the meteorological scene evolution process of the specific geographic area, so that the meteorological scene evolution of the specific geographic area can be effectively simulated. And the accuracy and the practicability of meteorological simulation are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a meteorological data simulation method and system based on a generative adversarial network. BACKGROUND

[0002] In the field of meteorological research and application, accurate meteorological data is crucial for weather prediction, disaster warning, climate research, etc. The traditional way of obtaining meteorological data mainly relies on real-time monitoring of meteorological observation sites. However, the distribution of observation sites is limited and the observation cost is high, making it difficult to fully cover all geographical areas, especially in some remote or complex terrain areas, where it is extremely difficult to obtain complete and continuous meteorological data. At the same time, although historical meteorological data is abundant, how to mine the inherent laws of meteorological evolution from it and generate virtual meteorological data with spatio-temporal consistency based on these laws to simulate the evolution process of meteorological scenarios in different geographical areas is a challenge faced by current meteorological data simulation technology. The existing meteorological data simulation methods have deficiencies in handling the temporal sequence correlation and geographical area correlation of meteorological elements, making it difficult to generate accurate and realistic meteorological simulation data. SUMMARY

[0003] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a meteorological data simulation method based on a generative adversarial network, which comprises: obtaining a historical meteorological data set, the historical meteorological data set containing meteorological evolution records with time sequence relationship and corresponding geographical area identification information; inputting the historical meteorological data set into a preset time sequence law learning model for meteorological evolution law extraction processing to obtain a meteorological evolution law feature set, the meteorological evolution law feature set containing temporal sequence correlation features and geographical area correlation features of meteorological elements; constructing a conditional generative adversarial network model, the conditional generative adversarial network model containing a generator module and a discriminator module, the generator module being used to receive the meteorological evolution law feature set as a conditional input, and the discriminator module being used to distinguish between generated meteorological data and real meteorological data; inputting the meteorological evolution law feature set into the generator module of the conditional generative adversarial network model to generate a virtual meteorological field data set with geographical area identification information, the virtual meteorological field data set containing meteorological element distribution information consistent with the spatio-temporal features of the historical meteorological data set; outputting the virtual meteorological field data set, the virtual meteorological field data set being used to simulate the evolution process of meteorological scenarios in a specific geographical area.

[0004] In still another aspect, the embodiments of the present application also provide a weather data simulation system based on a generative adversarial network, comprising a processor, a machine readable storage medium, the machine readable storage medium being connected with the processor, the machine readable storage medium being used for storing programs, instructions or codes, and the processor being used for executing the programs, instructions or codes in the machine readable storage medium to realize the above-mentioned method.

[0005] Based on the above aspects, the embodiments of the present application comprehensively capture the complex correlation of weather elements in time and geographical space by acquiring a historical weather data set containing time series relationship and geographical area identification information, and extracting weather evolution law features by using a time series law learning model. The conditional generative adversarial network model is constructed, the generator module can generate a virtual weather field data set with geographical area identification information according to the extracted weather evolution law features, and the discriminator module effectively ensures the consistency of the generated data and the real weather data in distribution. The overall method realizes the accurate simulation of the evolution process of the weather scene in a specific geographical area, and improves the accuracy and practicability of weather simulation. BRIEF DESCRIPTION OF DRAWINGS

[0006] Figure 1 is the execution flow diagram of the weather data simulation method based on the generative adversarial network provided by the embodiments of the present application.

[0007] Figure 2 is the schematic diagram of exemplary hardware and software components of the weather data simulation system based on the generative adversarial network provided by the embodiments of the present application. DETAILED DESCRIPTION

[0008] The present application will be described in detail below with reference to the accompanying drawings, Figure 1 is the flow diagram of the weather data simulation method based on the generative adversarial network provided by an embodiment of the present application, and the weather data simulation method based on the generative adversarial network will be described in detail below.

[0009] Step S110: Acquire a historical weather data set, the historical weather data set containing weather evolution records with time series relationship and corresponding geographical area identification information.

[0010] In the air traffic control scene, the acquisition of the historical weather data set needs to cover all the key airspaces involved in the flight of the aircraft. These key airspaces include the airport runway surrounding area, the approach route coverage area, the high-altitude air route area, etc., each area corresponds to a unique geographical area identification information, which is composed of letters and numbers, and is used to accurately distinguish different control airspaces.

[0011] The meteorological evolution record contains continuous observation values of various meteorological elements, which can specifically include ground wind speed, ground wind direction, wind speed and direction at different altitudes, visibility, cloud base height, cloud cover, temperature, dew point temperature, air pressure, precipitation, thunderstorm location and moving direction, etc. The observation data of these elements are collected by devices such as meteorological radars, meteorological satellites, ground observation stations, sounding balloons, and meteorological sensors carried by aircraft distributed in the airspace.

[0012] During the collection process, the time interval of the data is set according to the variation characteristics of the elements, for example, the observation interval of ground wind speed and direction is once every minute, while the observation interval of temperature and air pressure is once every five minutes, and the thunderstorm related data is updated in real time. All collected data will be transmitted to the air traffic control meteorological data center, and encryption protocol is used for encryption during transmission to prevent data leakage.

[0013] After the data center receives the data, it first performs integrity check to check whether each meteorological element at each time point is missing. For missing data, an interpolation method based on adjacent time point observation values is used to fill in the missing data to ensure the continuity of the time series. At the same time, the data is checked for consistency and data that obviously does not conform to physical laws, such as wind speed exceeding the extreme value that can occur in nature, is removed.

[0014] The above processed historical meteorological data set is stored according to geographical region identification information and time sequence to form a structured database for subsequent step calling.

[0015] Step S120: input the historical meteorological data set into a preset time sequence rule learning model for meteorological evolution rule extraction processing to obtain a meteorological evolution rule feature set, the meteorological evolution rule feature set containing time sequence correlation features and geographical region correlation features of meteorological elements.

[0016] The historical meteorological data set stored in the database is grouped according to geographical region identification information, and each group of data corresponds to a specific airspace. Then, each group of data is input into a preset time sequence rule learning model.

[0017] The time sequence rule learning model is composed of a long short-term memory network layer and a Transformer encoder layer, which can extract the rules of meteorological elements changing over time and the correlation between different geographical regions from historical meteorological data.

[0018] During processing, the model first performs standardization processing on the input meteorological data, converting meteorological element values of different magnitudes into the same numerical range. The conversion is achieved by subtracting the mean value of the element and then dividing by the standard deviation, ensuring that each element has the same weight influence in subsequent processing.

[0019] The obtained weather evolution rule feature set after processing, wherein the time sequence correlation feature is embodied as the mutual influence of different weather elements in the same airspace in the time dimension, for example, the influence rule of temperature change on air pressure; and the geographical area correlation feature is embodied as the propagation relationship of weather elements between adjacent airspaces, for example, the rule of thunderstorm movement from a certain airspace to an adjacent airspace.

[0020] Step S121: performing time sequence division processing on the weather evolution records in the historical weather data set to obtain a plurality of continuous weather time sequence segments, each weather time sequence segment containing a weather element observation value sequence in a preset time length.

[0021] For the weather evolution record of each geographical area, division is performed according to a preset time length. In the air traffic control scene, the preset time length is determined according to the time period of flight operation, for example, six hours, which is consistent with the typical time of aircraft crossing areas.

[0022] In the division, the first six hours of weather data are intercepted as the first weather time sequence segment from the start time of the historical data in chronological order, and then the next six hours of data are intercepted as the second segment from the end time of the first segment, and so on, until all historical data are processed.

[0023] Each weather time sequence segment contains an observation value sequence of all weather elements in the time period, and the length of each sequence is determined by the observation interval and the time length. For example, for wind speed data with an observation interval of one minute, a six-hour time sequence segment contains 360 observation values, forming a continuous sequence.

[0024] In the division process, if the time length of the last segment is less than the preset time length, the last observation value in the segment is copied to supplement it, so that the time length of all weather time sequence segments is consistent, facilitating subsequent model processing.

[0025] Step S122: inputting the weather time sequence segment into the long short-term memory network layer of the time sequence rule learning model to perform time dimension feature extraction processing and generating a short-term dependence feature vector of the weather element, wherein the short-term dependence feature vector is used to represent the correlation of adjacent time step changes of the weather element.

[0026] Before the weather time sequence segment is input into the long short-term memory network layer, it needs to be converted into a tensor form recognizable by the model, and each observation value sequence of the weather element corresponds to a channel of the tensor, and the dimension of the tensor is (number of time steps, number of weather element categories).

[0027] The LSTM layer contains multiple memory cells, each of which is composed of an input gate, a forget gate, a cell state, and an output gate. These memory cells are arranged in time order, and each cell processes the meteorological data of one time step.

[0028] During processing, the meteorological data of each time step first enters the input gate, which calculates a weight value based on the current data and the hidden state of the previous time step. This weight value determines the degree of update of the cell state by the current data. Next, the forget gate calculates another weight value based on the same input, which determines the information to be retained and forgotten in the cell state.

[0029] After the cell state is updated according to the weight value of the forget gate, it is further updated in combination with the weight value of the input gate and the processing result of the current data to form a new cell state. Finally, the output gate calculates an output weight based on the new cell state and the current data to determine which information in the cell state will be output as the hidden state of the current time step.

[0030] In this way, the LSTM layer can capture the correlation between adjacent time steps of meteorological elements, such as the influence of the increase in wind speed in the previous minute on the wind speed in the next minute. After integrating these correlation information, a short-term dependence feature vector is formed, the dimension of which is related to the number of meteorological element types and the number of time steps, and contains the change characteristics of each element in adjacent time steps.

[0031] Step S1221: Perform sequence alignment processing on the meteorological element observation value sequence in the meteorological time sequence segment to make the number of time steps consistent for each meteorological time sequence segment.

[0032] Since the observation intervals of different meteorological elements may be different, the lengths of the observation value sequences of different elements in the same meteorological time sequence segment may differ. For example, the temperature observation interval is five minutes, the sequence length of six hours is 72, and the wind speed sequence length is 360. The above differences will affect the processing effect of the model.

[0033] Therefore, it is necessary to perform sequence alignment processing on the observation value sequences of all meteorological elements. The processing method is to increase the length of the sequence with longer observation interval through interpolation method to make it consistent with the length of the sequence with shortest observation interval.

[0034] The interpolation method adopts linear interpolation. For a temperature sequence, four values are inserted between every two adjacent observation values, and each value is uniformly distributed according to the difference between the adjacent observation values. For example, if two adjacent temperature observation values are T1 and T2, and the interval is five minutes, then the four values inserted in the middle are T1+(T2-T1) / 5, T1+2*(T2-T1) / 5, T1+3*(T2-T1) / 5, and T1+4*(T2-T1) / 5. In this way, the length of the temperature sequence is expanded to 360.

[0035] For sequences with the same observation interval but different lengths due to data missing, the same linear interpolation method is used to fill in the missing values, so as to ensure that the number of time steps of all sequences is consistent.

[0036] Step S1222: input the aligned meteorological element observation value sequence into the input gate unit of the long short-term memory network layer, calculate the fusion weight of the input information and the historical memory information of the current time step, and generate the input gate activation value.

[0037] The aligned meteorological element observation value sequence is sequentially input into the input gate unit according to the time steps. At each time step, the input gate unit receives the meteorological element observation value vector of the current time step and the hidden state vector of the previous time step.

[0038] The input gate unit first splices the two vectors to form a joint vector, and then processes the joint vector through a fully connected layer. The weight matrix of the fully connected layer has a dimension of (joint vector dimension, 1), and the bias term is a constant. The processed result is converted through a sigmoid activation function to obtain the input gate activation value, which ranges from 0 to 1.

[0039] The size of the input gate activation value represents the fusion proportion of the input information and the historical memory information of the current time step. The closer the value is to 1, the greater the proportion of the current input information in the fusion result; the closer the value is to 0, the greater the proportion of the historical memory information. For example, when a sudden wind speed surge occurs, the input gate activation value will tend to 1, so that the current wind speed change information is more integrated into the cell state.

[0040] Step S1223: input the aligned meteorological element observation value sequence into the forget gate unit of the long short-term memory network layer, calculate the retention weight of the historical memory information, and generate the forget gate activation value.

[0041] Similar to the input gate unit, the forget gate unit receives the meteorological element observation value vector of the current time step and the hidden state vector of the previous time step at each time step.

[0042] The forget gate unit also splices the two vectors into a joint vector, which is processed by another fully connected layer whose weight matrix and bias term are different from those of the input gate unit. The processed result is converted by a sigmoid activation function to generate a forget gate activation value, which is also in the range of 0 to 1.

[0043] The forget gate activation value is used to determine the degree of retention of historical memory information. The closer the value is to 1, the higher the proportion of historical memory information that is retained. The closer the value is to 0, the higher the proportion of historical memory information that is forgotten. For example, when the meteorological elements are in a stable change phase, the forget gate activation value will remain at a high level, allowing the historical change trend information to be retained.

[0044] Step S1224: updating the cell state vector based on the forget gate activation value, retaining the key part of the historical memory information and discarding the redundant part.

[0045] The cell state vector is a carrier for storing historical memory information in the long short-term memory network layer, and its dimension is the same as the number of meteorological element types. At each time step, the cell state vector is updated according to the forget gate activation value.

[0046] The update method is to multiply the cell state vector at the previous time step and the forget gate activation value element by element. For the part with a larger value in the multiplication result, it means that the corresponding historical memory information is the key part and is retained. The part with a smaller value is redundant information and is discarded.

[0047] For example, under the condition of continuous stable wind direction, the cell state vector elements related to wind direction will be retained due to the high forget gate activation value, while some elements corresponding to minor wind direction fluctuations may be weakened due to the low activation value. Through the above update method, the cell state vector can continuously retain historical information valuable for current meteorological analysis.

[0048] Step S1225: inputting the updated cell state vector into the output gate unit of the long short-term memory network layer, calculating the output weight of the cell state vector, and generating an output gate activation value.

[0049] The updated cell state vector, the meteorological element observation value vector at the current time step, and the hidden state vector at the previous time step are input into the output gate unit.

[0050] The output gate unit first splices the meteorological element observation value vector at the current time step and the hidden state vector at the previous time step into a joint vector, which is processed by a fully connected layer, and then interacts with the updated cell state vector. Finally, the output gate activation value is generated by a sigmoid activation function, which is in the range of 0 to 1.

[0051] The output gate activation value determines which information in the cell state vector can be passed into the hidden state vector of the current time step. The closer the value is to 1, the easier the corresponding cell state information is to be output; the closer the value is to 0, the more difficult the corresponding cell state information is to be output. For example, when analyzing the visibility change that is crucial for aircraft take-off and landing, the cell state information related to visibility will be output more due to the high output gate activation value.

[0052] Step S1226: generating the hidden state vector of the current time step according to the output gate activation value and the tanh activation result of the updated cell state vector.

[0053] First, the tanh activation processing is performed on the updated cell state vector to convert each element value in the vector to the range of -1 to 1, which can suppress the influence of excessively large values on subsequent processing.

[0054] Then, the result of the tanh activation processing is multiplied element by element with the output gate activation value to obtain the hidden state vector of the current time step. The vector has the same dimension as the cell state vector and contains the meteorological element information of the current time step after screening and processing, which integrates the key content of the current input and historical memory.

[0055] For example, at the initial stage of thunderstorm formation, the hidden state vector contains information about the thunderstorm location, intensity, and features similar to the historical thunderstorm formation process, which will be used for subsequent time step processing.

[0056] Step S1227: concatenating the hidden state vectors of all time steps in the meteorological time series segment in chronological order to generate a short-term dependency feature vector representing the correlation of meteorological element changes between adjacent time steps.

[0057] After all time steps of the meteorological time series segment are processed by the long short-term memory network layer, the hidden state vector corresponding to each time step can be obtained. These vectors are concatenated in chronological order to form a matrix, which has the number of time steps as the number of rows and the dimension of the hidden state vector as the number of columns. This matrix is the short-term dependency feature vector, which completely records the correlation of meteorological element changes between adjacent time steps in the entire meteorological time series segment. For example, the difference between a row and the next row in the matrix can reflect the change amplitude and trend of meteorological elements in adjacent time steps, such as the increase or decrease of wind speed, the clockwise or counterclockwise change of wind direction, etc. The short-term dependency feature vector will be stored as input data for the subsequent Transformer encoder layer.

[0058] Step S123: input the short-term dependency feature vector into the Transformer encoder layer of the time series rule learning model for global time sequence correlation modeling processing to generate a long-term dependency feature vector of the meteorological element, which is used to represent the correlation of non-adjacent time step changes in the meteorological element.

[0059] After the short-term dependency feature vector is input into the Transformer encoder layer, position encoding processing is first performed to add time position information to the vector. Position encoding is generated by sine and cosine functions, and different time steps correspond to different encoding values to ensure that the model can distinguish the time information at different positions in the vector.

[0060] The Transformer encoder layer includes multiple encoding blocks, and each encoding block is composed of a multi-head self-attention mechanism and a feedforward neural network. In the multi-head self-attention mechanism, the short-term dependency feature vector is linearly projected into multiple different feature spaces, and each space corresponds to an attention head.

[0061] Each attention head calculates the attention weight between any two time steps in the short-term dependency feature vector. The calculation method is to perform dot product operation on the feature vectors of the two time steps, and then convert the weight value through the softmax function. The size of the weight value represents the correlation strength between the two time steps, and the greater the value, the tighter the correlation.

[0062] The multi-head self-attention mechanism splices the outputs of all attention heads and integrates them through a linear layer to obtain a feature vector containing global time correlation information. Subsequently, the feature vector enters the feedforward neural network and is processed through two linear transformations and ReLU activation functions to enhance the non-linear expression ability of the model.

[0063] After processing by multiple encoding blocks, a long-term dependency feature vector is finally generated, which can capture the correlation of non-adjacent time step changes in the meteorological element, such as the correlation between humidity changes in the morning and precipitation probability in the evening.

[0064] Step S124: extract the geographical region identification information from the historical meteorological data set, and perform association mapping processing on the geographical region identification information and the long-term dependency feature vector to generate a region correlation feature vector containing geographical region attributes.

[0065] The geographical region identification information is extracted from the metadata of the historical meteorological data set, and each identification information is a string. These strings are one-hot encoded to convert them into binary vectors. The dimension of the vector is the same as the number of all possible geographical region identification information, and only one element in each vector is 1 and the rest are 0, corresponding to a specific geographical region.

[0066] The one-hot encoded geographical region identification vector is associatedly mapped with the long-term dependence feature vector. The mapping is realized through a fully connected layer, the input of which is the spliced vector of the long-term dependence feature vector and the geographical region identification vector, and the output is a region association feature vector with the same dimension as the long-term dependence feature vector.

[0067] The weight matrix of the fully connected layer is initialized according to the spatial adjacent relationship between geographical regions, and the weight values corresponding to adjacent spatial domains are set to a higher initial value, and the weight values of non-adjacent spatial domains are set to a lower initial value. During the model training process, the weight matrix will be adjusted according to the training data, so that the region association feature vector can accurately reflect the meteorological element association properties between different geographical regions.

[0068] For example, for adjacent approach control zones and regional control zones, the region association feature vector will contain the feature information of wind speed transmission between the two, and when the wind speed of one region changes, the trend of the wind speed change of the other region after a period of time can be reflected through the vector.

[0069] Step S125: fuse the short-term dependence feature vector, the long-term dependence feature vector and the region association feature vector to generate a meteorological evolution rule feature set containing time sequence association features and geographical region association features, the time sequence association features being composed of the short-term dependence feature vector and the long-term dependence feature vector, and the geographical region association features being composed of the region association feature vector.

[0070] The fusion process is carried out in the form of feature splicing. First, the short-term dependence feature vector and the long-term dependence feature vector are spliced along the time dimension to form time sequence association features, the dimension of which is (number of time steps, short-term feature dimension + long-term feature dimension), containing all the association information of meteorological elements in the time dimension.

[0071] Then, the region association feature vector is spliced with the time sequence association features along the feature dimension to form a meteorological evolution rule feature set. The dimension of the meteorological evolution rule feature set is (number of time steps, short-term feature dimension + long-term feature dimension + region feature dimension), and it contains both the time sequence association features and the geographical region association features.

[0072] In the spliced feature set, the features of each time step contain the association information of the time step with adjacent time steps, non-adjacent time steps, and other geographical regions. For example, the features of a certain time step contain both the association of meteorological elements with the same time of the previous hour and the previous day, and the association of meteorological elements with adjacent spatial domains.

[0073] After the generation of the meteorological evolution law feature set, standardization processing can be performed to ensure that the numerical ranges of all features are consistent, thereby preparing for the subsequent input condition generative adversarial network model.

[0074] Step S130: constructing a condition generative adversarial network model, the condition generative adversarial network model comprising a generator module and a discriminator module, the generator module being configured to receive the meteorological evolution law feature set as a conditional input, and the discriminator module being configured to distinguish between generated meteorological data and real meteorological data.

[0075] In the air traffic control scenario, the constructed condition generative adversarial network model needs to meet the requirement of generating high-precision virtual meteorological data to support aircraft flight simulation and air traffic control training. Both the generator module and the discriminator module adopt a deep learning network structure, and the model performance is improved through adversarial training.

[0076] In the initialization stage, the network parameters of the generator module and the discriminator module are set according to the characteristics of the air traffic control scenario meteorological data to ensure that the model can quickly adapt to the distribution characteristics of the air traffic control meteorological data.

[0077] Step S131: configuring the network structure of the generator module, the generator module comprising an input embedding layer, a feature expansion layer, a spatial feature generation layer and an output mapping layer connected in sequence, the input embedding layer being configured to receive the meteorological evolution law feature set and perform dimension mapping processing.

[0078] The core function of the input embedding layer is to map the meteorological evolution law feature set to a feature space suitable for subsequent processing of the generator module. This layer includes an embedding matrix, the number of rows of which is equal to the feature dimension of the meteorological evolution law feature set, and the number of columns is determined according to the complexity of the air traffic control meteorological data to ensure that the information in the feature set can be fully expressed.

[0079] In the air traffic control scenario, due to the large number of meteorological elements and the complex correlation, the number of columns of the embedding matrix needs to be set to a larger value to avoid information loss. The embedding process is realized by matrix multiplication operation between the meteorological evolution law feature set and the embedding matrix, and the operation result is the embedded feature vector after mapping, the dimension of which is consistent with the number of columns of the embedding matrix.

[0080] The main function of the feature expansion layer is to expand the dimension of the embedded feature vector and increase the spatial dimension information of the feature. This layer adopts a stacked structure of transpose convolution layers, each of which includes a transpose convolution operation and a batch normalization operation. The transpose convolution operation gradually increases the spatial size of the feature vector through specific convolution kernel and step size settings, so that it can correspond to a specific spatial grid.

[0081] The batch normalization operation is used to normalize the features after the transpose convolution, so that the input data distribution of each layer remains stable, speeds up the training of the model, and reduces the occurrence of overfitting. In the air traffic control scenario, the division of the spatial grid needs to match the actual air traffic control sector division, so the transpose convolution parameters of the feature expansion layer need to be adjusted according to the size and number of sectors.

[0082] The spatial feature generation layer is used to further refine and optimize the feature maps output by the feature expansion layer, and enhance the spatial correlation of the features. This layer contains a convolution block attention module that can automatically learn the importance weights of different meteorological elements in space, so that the model pays more attention to meteorological regions that have a greater impact on aircraft flight safety, such as airspace near airports and route intersections.

[0083] The output mapping layer is used to convert the feature maps output by the spatial feature generation layer into actual meteorological data format. This layer adjusts the number of channels of the feature maps to the number of types of meteorological elements through a convolution layer, and then maps the feature values to the actual physical range of each meteorological element through an activation function, such as the value range of wind speed and the value range of temperature.

[0084] Step S1311: setting the mapping parameters of the input embedding layer, the mapping parameters including an embedding matrix, the number of rows of the embedding matrix consistent with the feature dimension of the meteorological evolution law feature set, and the number of columns of the embedding matrix consistent with the preset embedding dimension.

[0085] The initialization of the embedding matrix is performed in a random normal distribution manner, and each element in the matrix is sampled from a normal distribution with a mean of 0 and a standard deviation of a certain value. In the air traffic control scenario, the setting of this standard deviation needs to consider the variance of each feature in the meteorological evolution law feature set to ensure that the initialized embedding matrix can better adapt to the distribution of the features.

[0086] The preset of the embedding dimension needs to consider the complexity of the meteorological data and the computational efficiency of the model. If the embedding dimension is too small, it may lead to insufficient expression of feature information, affecting the quality of the generated virtual meteorological data; if the embedding dimension is too large, it will increase the computational load of the model and reduce the training and inference speed.

[0087] In actual settings, the appropriate embedding dimension can be determined through multiple tests, and the error between the generated virtual meteorological data and the real meteorological data is used as an evaluation index during the test process. The embedding dimension with the smallest error is selected as the preset value. The embedding matrix is continuously updated during the model training process, and the element values in the matrix are adjusted through the backpropagation algorithm, so that the mapped embedding feature vector can better serve the subsequent feature expansion and meteorological data generation process.

[0088] Step S1312: configure the feature expansion layer as a transposed convolution layer stack structure, the transposed convolution layer stack structure comprising a plurality of transposed convolution units connected in sequence, each transposed convolution unit comprising a transposed convolution operation and a batch normalization operation, the number of transposed convolution units being determined according to the spatial resolution requirement of the target virtual meteorological field data set.

[0089] The number of transposed convolution units is closely related to the spatial resolution of the target virtual meteorological field data set. The higher the spatial resolution, the more transposed convolution units are needed. In the air traffic control scenario, the spatial resolution of the target virtual meteorological field data set needs to be determined according to different application scenarios. For example, in the airport runway surrounding area, a higher spatial resolution is needed to accurately simulate the meteorological changes near the runway; while in the high-altitude air route area, the spatial resolution can be appropriately reduced.

[0090] The transposed convolution operation in each transposed convolution unit has a specific convolution kernel size, step and padding method. The selection of the convolution kernel size needs to consider the correlation range of meteorological elements in space, the step determines the spatial size magnification multiple of the feature map after each transposed convolution operation, and the padding method is used to maintain the information integrity of the feature map edge.

[0091] The batch normalization operation is performed after each transposed convolution operation, and its parameters include moving mean and moving variance, which are calculated by sliding window during model training, and are used to standardize the features after transposed convolution, so that the distribution of feature values is more stable.

[0092] In the configuration process, the parameters of adjacent transposed convolution units need to be reasonably designed to ensure that the spatial size of the feature map can be gradually magnified to the spatial resolution of the target virtual meteorological field data set. For example, the first transposed convolution unit magnifies the spatial size of the feature map by a certain multiple, and the second transposed convolution unit further magnifies it, until the target size is reached.

[0093] Step S1313: set the network parameters of the spatial feature generation layer, the spatial feature generation layer comprising a convolution block attention module, the convolution block attention module being used for calculating channel attention weights and spatial attention weights of the feature map output by the feature expansion layer to generate a weighted spatial feature map.

[0094] The network parameters of the convolution block attention module include the convolution layer weights and bias terms in the channel attention submodule and the spatial attention submodule. The channel attention submodule comprises two convolution layers connected in sequence, the first convolution layer being used for dimension reduction processing of the feature map to reduce the calculation amount, and the second convolution layer being used for mapping the dimension-reduced feature back to the original channel number.

[0095] The spatial attention sub-module includes a convolution layer, and a size of a convolution kernel of the convolution layer is determined according to a range of association of meteorological elements in space, and is used for extracting spatial attention weight information from a feature map. Weight parameters of the convolution layers are all set in a random initialization manner, and are optimized through back propagation in a model training process.

[0096] The calculation process of the channel attention weight is as follows: first, global average pooling and global maximum pooling operations are performed on the feature map output by the feature expansion layer to obtain two different channel description vectors; then, the two vectors are input into the convolution layer of the channel attention sub-module for processing, and finally, the channel attention weight of each channel is generated through a sigmoid activation function.

[0097] The calculation process of the spatial attention weight is as follows: average pooling and maximum pooling operations in the channel dimension are performed on the feature map after channel attention weighting to obtain two spatial description maps; the two maps are spliced and input into the convolution layer of the spatial attention sub-module for processing, and then a spatial attention weight map is generated through a sigmoid activation function.

[0098] By multiplying the feature map with the channel attention weight and the spatial attention weight respectively, a weighted spatial feature map is obtained, which can highlight important meteorological element and spatial region information.

[0099] Step S1314: The output mapping layer is configured as a combination structure of a convolution layer and an activation function, the output channel number of the convolution layer is consistent with the number of meteorological element types included in the virtual meteorological field data set, and the activation function is used to map the output feature value to a preset meteorological element numerical range.

[0100] The input of the convolution layer is the spatial weighted feature map output by the spatial feature generation layer, and a size of a convolution kernel of the convolution layer is usually set to 1x1 to adjust the channel number without changing the spatial size of the feature map. The weight parameters of the convolution layer are obtained through random initialization and are continuously optimized in the training process to ensure that the output feature map can accurately reflect the distribution of each meteorological element.

[0101] The selection of the activation function is determined according to the numerical range characteristics of the meteorological elements. For meteorological elements with a value range between 0 and a maximum value, such as visibility and cloud cover, a sigmoid activation function is used to map the feature value to a range between 0 and 1, and then the actual value is obtained by multiplying the maximum value of the element.

[0102] For meteorological elements with a value range spanning positive and negative, such as temperature (which can appear below zero), a tanh activation function is used to map the feature values to between -1 and 1, and then converted to the actual temperature value range through linear transformation. In the air traffic control scene, it is necessary to accurately set the activation function parameters corresponding to each meteorological element to ensure that the generated virtual meteorological data has the correct numerical range and meets the needs of flight simulation and air traffic control training.

[0103] Step S132: configuring the network structure of the discriminator module, the discriminator module comprising a multi-scale feature extraction layer, a conditional feature fusion layer, and a discriminant decision layer connected in sequence, the multi-scale feature extraction layer being configured to extract different spatial scale features of the input meteorological data, and the conditional feature fusion layer being configured to fuse the set of meteorological evolution rule features and the extracted spatial scale features.

[0104] The multi-scale feature extraction layer can extract features from the input meteorological data at different spatial scales, capturing the distribution characteristics of meteorological elements in different spatial ranges. In the air traffic control scene, meteorological data behaves differently at different spatial scales, for example, a thunderstorm behaves as a whole moving trend at a large scale, and behaves as a local intensity change at a small scale.

[0105] The conditional feature fusion layer fuses the set of meteorological evolution rule features with the features extracted by the multi-scale feature extraction layer, so that the discriminator module can combine the regularity information of meteorological evolution when distinguishing between real meteorological data and generated meteorological data, and improve the accuracy of discrimination.

[0106] The discriminant decision layer outputs a probability value based on the fused features, which represents the likelihood that the input meteorological data is real data.

[0107] Step S1321: setting the multi-scale feature extraction layer to a convolutional layer pyramid structure, the convolutional layer pyramid structure comprising a plurality of convolutional units with different convolution kernel sizes, each convolutional unit being configured to extract a local feature map of the input meteorological data at a corresponding spatial scale, and the convolution kernel size gradually increasing from the first layer to the last layer.

[0108] Each convolutional unit in the convolutional layer pyramid structure includes a convolution operation, an activation function, and a pooling operation. The size of the convolution kernel is determined according to the spatial scale to be extracted, with a smaller convolution kernel (such as 3x3) used to extract small-scale local features such as subtle changes in meteorological elements, and a larger convolution kernel (such as 7x7, 11x11) used to extract large-scale global features such as large-scale pressure system distribution.

[0109] In the air traffic control scenario, the increment amplitude of the convolution kernel size needs to be adjusted according to the size of the spatial domain and the characteristics of the meteorological phenomenon. The number of output channels of each convolution unit is determined according to the complexity of the scale feature, and is usually increased with the increase of the convolution kernel size to accommodate more large-scale feature information.

[0110] The activation function adopts LeakyReLU to solve the gradient disappearance problem and improve the training effect of the model. The pooling operation adopts maximum pooling to reduce the spatial size of the feature map, reduce the calculation amount, and retain important feature information.

[0111] Step S1322: performing global average pooling processing on the local feature map output by each convolution unit to generate a global feature vector corresponding to the spatial scale, the dimension of the global feature vector being consistent with the number of output channels of the convolution unit.

[0112] The global average pooling processing converts the two-dimensional feature map into a one-dimensional global feature vector by calculating the average value of each local feature map in the spatial dimension. The above processing method can reduce the dimension of the feature vector while retaining the global feature information at this spatial scale.

[0113] In the air traffic control scenario, the global feature vector can reflect the overall distribution characteristics of the meteorological elements at this spatial scale, such as the average wind speed, average temperature, etc. at a certain scale. The global average pooling operation avoids a large number of parameters brought by the fully connected layer and reduces the risk of overfitting of the model.

[0114] Step S1323: configuring a fusion strategy of the conditional feature fusion layer, the fusion strategy including a feature concatenation operation and an attention weighting operation, the feature concatenation operation being used to concatenate the global feature vectors of different spatial scales along the channel dimension to generate a multi-scale concatenated feature vector.

[0115] The global feature vectors of different spatial scales contain different levels of meteorological information. Concatenating them along the channel dimension can integrate the feature information of each scale to form a multi-scale concatenated feature vector. The dimension of the vector is the sum of the dimensions of each global feature vector, and contains meteorological features of different spatial scales from fine to coarse.

[0116] In the air traffic control scenario, the integration of the above multi-scale features helps the discriminator module to fully understand the characteristics of the meteorological data and improve the ability to distinguish between real meteorological data and generated meteorological data.

[0117] Step S1324: mapping the set of meteorological evolution law features to the same dimension as the multi-scale concatenated feature vector through a fully connected layer to generate a conditional mapping feature vector.

[0118] The input of the full connection layer is the meteorological evolution law feature set, which is converted into a conditional mapping feature vector with the same dimension as the multi-scale spliced feature vector by setting appropriate weight matrix and bias term. The parameters of the full connection layer are continuously optimized during the model training process to ensure that the mapped conditional features can accurately reflect the meteorological evolution law and match the multi-scale features.

[0119] In the air traffic control scenario, the meteorological evolution law feature set contains time and spatial correlation information of meteorological elements. After being mapped into a conditional mapping feature vector, it can be effectively fused with the multi-scale features.

[0120] Step S1325: Perform attention weighting operation on the multi-scale spliced feature vector and the conditional mapping feature vector to calculate the fusion weight of each feature channel and generate a weighted feature vector of fused conditional information.

[0121] The attention weighting operation first calculates the correlation between the multi-scale spliced feature vector and the conditional mapping feature vector, and determines the fusion weight of each feature channel according to the correlation size. The higher the correlation of the feature channel, the greater the fusion weight, and the greater the proportion in the final weighted feature vector.

[0122] The correlation can be calculated by dot product operation, which performs dot product on the corresponding channels of the multi-scale spliced feature vector and the conditional mapping feature vector to obtain the correlation value, and then normalizes these values by softmax function to obtain the fusion weight.

[0123] After multiplying the multi-scale spliced feature vector and the conditional mapping feature vector by the fusion weight and adding them, a weighted feature vector of fused conditional information is obtained. In the air traffic control scenario, the above fusion method can make the discriminator module fully consider the meteorological evolution law when distinguishing meteorological data, and improve the accuracy of discrimination.

[0124] Step S1326: Set the discriminant decision layer as a combination structure of full connection layer and sigmoid activation function, the input dimension of the full connection layer is consistent with the dimension of the weighted feature vector, the output dimension of the full connection layer is 1, and the sigmoid activation function is used to output the probability value of the input meteorological data being real data.

[0125] The full connection layer converts the weighted feature vector into a one-dimensional output value, which is processed by the sigmoid activation function to obtain a probability value between 0 and 1. If the probability value is close to 1, it means that the input meteorological data is more likely to be real data; if the probability value is close to 0, it means that the input meteorological data is more likely to be generated data.

[0126] In the air traffic control scenario, the output of the decision layer is used to guide the parameter update of the generator module, and the ability of the generator module to generate real meteorological data is continuously improved through adversarial training.

[0127] Step S133: Set the loss function of the generator module and the discriminator module, which includes a generator loss term and a discriminator loss term. The generator loss term is used to measure the difference between the generated meteorological data and the real meteorological data, and the discriminator loss term is used to measure the discrimination ability of the discriminator module for real meteorological data and generated meteorological data.

[0128] The generator loss term adopts a mean square error loss function to calculate the average of the square differences between corresponding elements of the generated meteorological data and the real meteorological data. This loss term can reflect the closeness of the generated data to the real data in terms of numerical value, and the smaller the loss value, the closer the generated data to the real data.

[0129] In the air traffic control scenario, for some meteorological elements that are crucial to flight safety, such as wind speed and visibility, a higher weight can be set in the loss function to make the generator module pay more attention to the generation accuracy of these elements.

[0130] The discriminator loss term adopts a binary cross-entropy loss function to calculate the loss value according to the predicted probability of the discriminator module for real meteorological data and generated meteorological data and the actual label (real data label is 1, generated data label is 0). This loss term can reflect the discrimination ability of the discriminator module, and the smaller the loss value, the stronger the discrimination ability of the discriminator module.

[0131] Step S134: Initialize the network parameters of the generator module and the discriminator module, which include the weight matrix and bias vector of each layer. The initialization method adopts random normal distribution sampling.

[0132] The initialization of the weight matrix and the bias vector is crucial to the training effect of the model. The random normal distribution sampling method assigns a value to each parameter from a normal distribution with a mean of 0 and a standard deviation of a certain value. In the air traffic control scenario, the setting of this standard deviation needs to be adjusted according to the type of network layer and the size of input features.

[0133] For convolutional layers, a small standard deviation is usually set to avoid the problem of gradient explosion caused by large initial weights; for fully connected layers, the standard deviation can be appropriately increased. The initialized network parameters will be continuously updated through the backpropagation algorithm during the model training process, gradually optimizing the performance of the model.

[0134] Step S135: connecting the generator module and the discriminator module to form an end-to-end conditional generative adversarial network model, wherein the output of the generator module is as one of the inputs of the discriminator module, and the meteorological evolution law feature set is input into the conditional feature fusion layers of the generator module and the discriminator module.

[0135] The output of the generator module is a set of virtual meteorological field data, which is input into the discriminator module as an input together with real meteorological data. The meteorological evolution law feature set is input into both modules, so that the generator module can generate reasonable virtual meteorological data according to the meteorological evolution law, and the discriminator module can discriminate the input data in combination with the meteorological evolution law.

[0136] In the air traffic control scenario, the above-mentioned end-to-end connection mode ensures that the model can be trained under the constraint of the meteorological evolution law to generate virtual data that conforms to the actual meteorological change law, meeting the needs of flight simulation and air traffic control training.

[0137] Step S140: inputting the meteorological evolution law feature set into the generator module of the conditional generative adversarial network model to generate a set of virtual meteorological field data with geographical region identification information, wherein the set of virtual meteorological field data contains meteorological element distribution information consistent with the spatiotemporal features of the set of historical meteorological data.

[0138] After the meteorological evolution law feature set is input into the generator module, it is mapped in dimension by the input embedding layer, expanded in dimension by the feature expansion layer, optimized in feature by the spatial feature generation layer, and converted in format by the output mapping layer, and finally a set of virtual meteorological field data is generated.

[0139] In the air traffic control scenario, the generated set of virtual meteorological field data needs to be consistent with the set of historical meteorological data in terms of time and space features, such as the change trend of meteorological elements and the meteorological correlation between different regions, to ensure that it can truly reflect the actual meteorological conditions.

[0140] Step S141: extracting geographical region correlation features in the meteorological evolution law feature set as geographical condition constraint information.

[0141] The geographical region correlation features contain the correlation information of meteorological elements between different geographical regions, and when they are used as geographical condition constraint information, they can guide the generator module to generate virtual meteorological data with correct regional correlation. In the air traffic control scenario, the geographical region correlation features can reflect the propagation law of meteorological elements between different air traffic control sectors, such as the gradient change of wind speed between adjacent sectors, the time interval of thunderstorm moving from one sector to another, etc.

[0142] When extracting the geographical area correlation feature, the corresponding data part needs to be separated from the meteorological evolution law feature set. Since the meteorological evolution law feature set is organized according to time steps and feature dimensions, and the geographical area correlation feature occupies a specific feature dimension interval, it can be directly extracted by dimension indexing. The extracted geographical area correlation feature retains its inherent time series structure, and each time step feature corresponds to the regional correlation information at that time.

[0143] After extraction, the geographical area correlation feature is verified to check whether it accurately reflects the meteorological element correlation between different geographical areas. The verification method is to compare the extracted feature with the actual regional meteorological element correlation in the historical data. If the deviation is within an acceptable range, it is determined as the geographical condition constraint information; if the deviation is large, it returns to the feature extraction stage for reprocessing.

[0144] Step S142: input the meteorological evolution law feature set into the input embedding layer of the generator module, and perform dimension mapping processing through an embedding matrix to generate an embedding feature vector.

[0145] Before inputting the meteorological evolution law feature set into the input embedding layer, it needs to be converted into a tensor form matching the input dimension of the embedding layer. The core of the input embedding layer is the embedding matrix, which has a number of rows equal to the feature dimension of the meteorological evolution law feature set and a number of columns equal to the preset embedding dimension. The setting of the embedding dimension needs to consider the complexity of the feature and the calculation efficiency of the model.

[0146] In the processing process, each feature value in the meteorological evolution law feature set will be operated with the corresponding row in the embedding matrix. Specifically, for each time step feature vector in the feature set, matrix multiplication operation is performed with the embedding matrix to map it from the original feature dimension to the preset embedding dimension, generating an embedding feature vector.

[0147] The process of matrix multiplication operation is that each element in the feature vector is multiplied by the corresponding row element in the embedding matrix and summed to obtain an element in the embedding feature vector. In this way, the high-dimensional meteorological evolution law feature set is converted into a low-dimensional embedding feature vector while retaining the key feature information.

[0148] The generated embedding feature vector has the same dimension as the preset embedding dimension and retains the original time series structure, with each time step corresponding to an embedding feature vector. These vectors will be used as input for the feature expansion layer for further feature processing.

[0149] Step S143: input the embedding feature vector into the feature expansion layer of the generator module, perform feature dimension expansion processing through the transpose convolution layer stack structure, and generate a high-dimensional feature map, wherein the spatial dimensions of the high-dimensional feature map match the spatial resolution of the target virtual meteorological field data set.

[0150] After the embedding feature vector is input into the feature expansion layer, dimension adjustment is first needed to convert it into a three-dimensional tensor form suitable for transpose convolution operation, i.e., (time step, spatial height, spatial width, embedding dimension), wherein the initial values of the spatial height and the spatial width are determined according to the embedding dimension and the target spatial resolution.

[0151] The transpose convolution layer stack structure of the feature expansion layer includes multiple transpose convolution units connected in sequence, and each transpose convolution unit is composed of a transpose convolution operation and a batch normalization operation. The transpose convolution operation performs up-sampling processing on the input tensor by setting appropriate convolution kernel size, step and padding mode, thereby increasing the spatial dimensions.

[0152] The first transpose convolution unit receives the adjusted three-dimensional tensor, and through the transpose convolution operation, the spatial height and the spatial width are expanded by a set multiple while keeping the number of channels unchanged. Then, the batch normalization operation is performed to standardize the feature values of each channel, so that the mean of the feature values is zero and the variance is one, thereby accelerating the training convergence of the model.

[0153] The feature map processed by the first transpose convolution unit is input into the next transpose convolution unit, and the above-mentioned transpose convolution and batch normalization operations are repeated. The convolution kernel size, step and padding mode of each transpose convolution unit are set according to the requirements of the target spatial resolution, so that the spatial height and width of the high-dimensional feature map generated after processing by all transpose convolution units are consistent with the spatial resolution of the target virtual meteorological field data set.

[0154] The number of channels of the high-dimensional feature map is the same as the output channel number of the last transpose convolution unit, and it contains rich spatial distribution characteristics of meteorological elements.

[0155] Step S144: input the high-dimensional feature map into the spatial feature generation layer of the generator module, combine the geographical condition constraint information, calculate the channel attention weight and the spatial attention weight through the convolution block attention module, and perform weighted processing on the high-dimensional feature map to generate a spatial weighted feature map with geographical regional characteristics.

[0156] After the high-dimensional feature map is input into the spatial feature generation layer, it first enters the channel attention submodule of the convolution block attention module. At the same time, the geographical condition constraint information is converted into a feature vector with the same number of channels as the high-dimensional feature map through a fully connected layer, which serves as auxiliary information for channel attention calculation.

[0157] In the channel attention submodule, the feature vector generated in combination with the geographical condition constraint information is used to process the high-dimensional feature map to calculate the channel attention weight. Then, the high-dimensional feature map is multiplied by the channel attention weight to obtain a channel weighted feature map.

[0158] The channel weighted feature map then enters the spatial attention submodule, also in combination with the geographical condition constraint information, to calculate a spatial attention weight map through spatial dimension processing. The channel weighted feature map is multiplied by the spatial attention weight map to obtain a preliminary spatial weighted feature map.

[0159] Finally, the geographical condition constraint information is converted into a geographical weight map with the same spatial dimension as the preliminary spatial weighted feature map through another fully connected layer, and the preliminary spatial weighted feature map is added to the geographical weight map to generate a spatial weighted feature map with geographical area characteristics.

[0160] Step S1441: input the high-dimensional feature map into the channel attention submodule of the convolution block attention module, and perform global average pooling and global maximum pooling processing on the high-dimensional feature map to generate a channel average feature vector and a channel maximum feature vector.

[0161] After the high-dimensional feature map is input into the channel attention submodule, global average pooling and global maximum pooling operations are simultaneously performed. Global average pooling calculates the average value of all elements in the entire spatial dimension for each channel of the high-dimensional feature map to obtain a scalar value, and the scalar values of all channels form a channel average feature vector, which has the same dimension as the number of channels of the high-dimensional feature map.

[0162] Global maximum pooling finds the maximum value of all elements in the entire spatial dimension for each channel of the high-dimensional feature map to obtain a scalar value, and the scalar values of all channels form a channel maximum feature vector, which also has the same dimension as the number of channels of the high-dimensional feature map.

[0163] These two pooling operations extract statistical information of channel features from different angles, respectively. Global average pooling can reflect the overall strength of channel features, and global maximum pooling can highlight the significant change part of channel features. The generated channel average feature vector and channel maximum feature vector will be used for subsequent channel attention weight calculation.

[0164] Step S1442: input the channel average feature vector and the channel maximum feature vector into a double-channel fully connected network, and perform feature fusion processing through a shared hidden layer to generate a channel fusion feature vector.

[0165] The channel average feature vector and the channel maximum feature vector are respectively input into two branches of the dual-channel fully connected network. Each branch first compresses the dimension of the input vector to a lower dimension, which is half of the number of high-dimensional feature map channels, through a fully connected layer.

[0166] The two compressed vectors are input into a shared hidden layer composed of multiple neurons, which performs nonlinear processing on the input vectors through a ReLU activation function to realize feature fusion. During the fusion process, the hidden layer learns the correlation between the two kinds of pooled features and integrates useful information together.

[0167] After processing by the hidden layer, the output vector restores the dimension to the same as the number of high-dimensional feature map channels through another fully connected layer to generate a channel fusion feature vector. This vector integrates the information of both channel average and channel maximum features and can more comprehensively reflect the importance of each channel of the high-dimensional feature map.

[0168] Step S1443: Perform sigmoid activation processing on the channel fusion feature vector to generate a channel attention weight for each feature channel.

[0169] Each element in the channel fusion feature vector is processed by a sigmoid activation function, which converts the input value to a range of 0 to 1. The vector obtained after processing is the channel attention weight, where each element corresponds to the weight value of a channel in the high-dimensional feature map.

[0170] The size of the weight value represents the importance of the channel in generating the virtual meteorological field data. The closer the value is to 1, the greater the influence of the channel's feature information on the generation result; the closer the value is to 0, the smaller the influence. In this way, the channel features that are crucial to meteorological field generation can be highlighted, and the secondary channel features can be suppressed.

[0171] Step S1444: Perform channel-by-channel multiplication processing on the high-dimensional feature map and the channel attention weight to generate a channel weighted feature map.

[0172] Each channel of the high-dimensional feature map is multiplied element-by-element with the corresponding weight value in the channel attention weight. For each element in the spatial dimension of the high-dimensional feature map, it is multiplied by the weight value corresponding to the channel, so that the feature values of important channels are enhanced and the feature values of secondary channels are weakened.

[0173] The result of the multiplication processing is the channel weighted feature map, which has the same dimension as the high-dimensional feature map, but the feature strengths of each channel are adjusted according to their importance. For example, for a channel related to thunderstorms, if its weight value is high, the feature values of this channel at each spatial location will be amplified, making it easier to be captured in subsequent processing.

[0174] Step S1445: input the channel weighted feature map into the spatial attention submodule of the convolution block attention module, perform average pooling and maximum pooling processing on the channel dimension of the channel weighted feature map, and generate a spatial average feature map and a spatial maximum feature map.

[0175] After the channel weighted feature map is input into the spatial attention submodule, average pooling and maximum pooling operations are respectively performed on the channel dimension. The average pooling on the channel dimension is to calculate the average value of all channels at each spatial position to generate a spatial average feature map. Each pixel value of the map represents the average feature intensity of all channels at the corresponding position.

[0176] The maximum pooling on the channel dimension is to find the maximum value of all channels at each spatial position to generate a spatial maximum feature map. Each pixel value of the map represents the maximum feature intensity of all channels at the corresponding position.

[0177] The spatial dimensions of the spatial average feature map and the spatial maximum feature map are the same as those of the channel weighted feature map, both being (time step, spatial height, spatial width), but the number of channels is 1. They reflect the feature importance of each spatial position from different angles.

[0178] Step S1446: concatenate the spatial average feature map and the spatial maximum feature map along the channel dimension to generate a spatial concatenated feature map.

[0179] The spatial average feature map and the spatial maximum feature map are concatenated along the channel dimension, i.e., two single-channel feature maps are combined into a double-channel feature map to generate a spatial concatenated feature map. The dimension of the feature map is (time step, spatial height, spatial width, 2), which contains both the average feature intensity and the maximum feature intensity information of each spatial position.

[0180] Step S1447: perform convolution operation and sigmoid activation processing on the spatial concatenated feature map to generate a spatial attention weight map.

[0181] The spatial concatenated feature map is first processed by a convolution layer with a preset convolution kernel size. The convolution operation converts the double-channel feature map into a single-channel feature map. This process is achieved by sliding window calculation of the convolution kernel and the feature map, and extracts the local correlation features in the space.

[0182] The feature map after convolution processing is processed by a sigmoid activation function to convert each pixel value to a range of 0 to 1, generating a spatial attention weight map. The dimension of the map is the same as the spatial dimension of the spatial concatenated feature map. Each pixel value represents the attention weight of the corresponding spatial position. The greater the value, the more important the feature information of the position in the generated result.

[0183] Step S1448: pixel-wise multiplication is performed between the channel- weighted feature map and the spatial attention weight map to generate a preliminary spatial weighted feature map.

[0184] The channel-weighted feature map is pixel-wise multiplied with the spatial attention weight map, that is, each spatial pixel value of each channel in the channel-weighted feature map is multiplied with the pixel value at the corresponding position in the spatial attention weight map. This operation enhances the feature values at important positions in space and weakens the feature values at unimportant positions, while retaining the weight adjustment results at the channel level.

[0185] The preliminary spatial weighted feature map generated after multiplication has been adjusted in terms of feature importance in both the channel and spatial dimensions, and can highlight the feature information that is crucial for generating the virtual meteorological field.

[0186] Step S1449: the geographical condition constraint information is mapped to a geographical weight map consistent with the spatial dimension of the preliminary spatial weighted feature map through a fully connected layer.

[0187] The geographical condition constraint information is first processed by a fully connected layer, and the input dimension of the fully connected layer is the feature dimension of the geographical condition constraint information, and the output dimension is the product of the spatial height and the spatial width of the preliminary spatial weighted feature map.

[0188] The processed result is converted into a geographical weight map with the same spatial dimension as the preliminary spatial weighted feature map, that is, (time step, spatial height, spatial width), through dimension reshaping. Each pixel value in the geographical weight map reflects the association strength of the corresponding spatial position and geographical area feature, and the stronger the association, the larger the pixel value.

[0189] Step S14410: pixel-wise addition is performed between the preliminary spatial weighted feature map and the geographical weight map to generate a spatial weighted feature map with geographical area features.

[0190] The preliminary spatial weighted feature map is pixel-wise added with the geographical weight map, that is, each spatial pixel value of each channel in the preliminary spatial weighted feature map is added with the pixel value at the corresponding position in the geographical weight map. This operation integrates the geographical area features into the spatial weighted feature map, so that the generated feature map can reflect the meteorological features of a specific geographical area.

[0191] For example, for a mountainous airspace, the pixel value at the position corresponding to the mountainous area in the geographical weight map is large, and after addition, the feature value at this position is enhanced, so that in the subsequent generation of the virtual meteorological field, the unique meteorological element distribution of the mountainous area, such as the change of wind speed with the terrain, is more easily reflected.

[0192] The spatially weighted feature map with geographical area characteristics generated after addition will be input as the output mapping layer, and the final virtual meteorological field data will be generated.

[0193] Step S145: input the spatially weighted feature map into the output mapping layer of the generator module, map the feature values into the meteorological element numerical range through the convolution layer and the activation function, and generate preliminary virtual meteorological field data.

[0194] After the spatially weighted feature map is input into the output mapping layer, it is first processed by a convolution layer. The number of convolution kernels of the convolution layer is the same as the number of meteorological element types contained in the virtual meteorological field data, and the channel number of the spatially weighted feature map is converted into the number of meteorological element types through convolution operation, and each channel corresponds to a meteorological element.

[0195] The process of convolution operation is that the convolution kernel slides on the spatially weighted feature map, multiplies and sums the feature values of the corresponding region to obtain the feature values of each meteorological element at each spatial position. The size and step length of the convolution kernel should be set to ensure that the spatial dimension of the output feature map is consistent with the target virtual meteorological field data.

[0196] The feature map after convolution processing is then processed by an activation function. The selection of the activation function should be determined according to the numerical range of the meteorological element. For example, for elements such as wind speed and temperature that have a clear numerical range, linear activation function or ReLU function with parameter adjustment can be used to map the feature values into the corresponding actual numerical range; for elements such as cloud cover in percentage form, sigmoid function can be used to map the feature values between 0 and 1.

[0197] After the activation function processing, preliminary virtual meteorological field data is generated, which contains the numerical values of each meteorological element at different time steps and spatial positions, and has the same structure and numerical range as the target virtual meteorological field data.

[0198] Step S146: perform spatiotemporal feature comparison processing on the preliminary virtual meteorological field data and the meteorological evolution records in the historical meteorological data set, adjust the output mapping layer parameters of the generator module, and make the spatiotemporal features of the preliminary virtual meteorological field data consistent with those of the historical meteorological data set.

[0199] The spatiotemporal feature comparison between the preliminary virtual meteorological field data and the meteorological evolution records in the historical meteorological data set includes comparison in time dimension and space dimension.

[0200] In the time dimension, the change trend of the meteorological elements in the same time step interval is compared, such as the increase / decrease rate of wind speed, the daily variation law of temperature, etc. The difference between the two in time series is calculated, and the difference is measured by the deviation of the difference value of the meteorological element value at each time step and the adjacent time step.

[0201] In the spatial dimension, the distribution characteristics of meteorological elements at the same spatial position are compared, such as the gradient change of wind speed in different airspaces and the spatial distribution uniformity of temperature, etc. The difference between the two in space is calculated, and the difference is measured by the deviation of the meteorological element values at the same spatial position.

[0202] According to the comparison results of the spatio-temporal characteristics, the parameters of the output mapping layer are adjusted, mainly the weights and biases of the convolution layer. The adjustment method is to calculate the gradient of the parameters according to the difference value through the back propagation algorithm, and then update the parameters in the direction of gradient descent, so that the spatio-temporal characteristics difference between the preliminary virtual meteorological field data and the historical meteorological data gradually decreases until they are consistent.

[0203] This adjustment process may need to be iterated several times, each iteration regenerates the preliminary virtual meteorological field data and performs comparison until the preset difference threshold is met.

[0204] Step S147: output the adjusted preliminary virtual meteorological field data as a virtual meteorological field data set with geographical area identification information.

[0205] After adjustment, the spatio-temporal characteristics of the preliminary virtual meteorological field data are consistent with those of the historical meteorological data set, and at this time it is determined as a virtual meteorological field data set.

[0206] Add the corresponding geographical area identification information to the data set, which corresponds to the geographical area identification information in the historical meteorological data set, ensuring that each virtual meteorological field data can be accurately associated with a specific airspace.

[0207] The virtual meteorological field data set is organized according to time steps and spatial positions to form a structured data, which is convenient for subsequent output and application.

[0208] Step S150: output the virtual meteorological field data set, which is used to simulate the evolution process of the meteorological scene in a specific geographical area.

[0209] Before outputting the virtual meteorological field data set, format conversion is required to convert it into a format that meets the data interface requirements of the air traffic control system, such as a specific binary format or XML format, to ensure that the data can be correctly read by the air traffic control simulation system.

[0210] During the output process, the data is compressed to reduce the bandwidth occupation of data transmission. The compression method uses a lossless compression algorithm to ensure the accuracy of the data. The compressed data is sent to the air traffic control simulation system through an encrypted transmission channel, which is used to simulate the evolution process of the weather scene in a specific geographical area, such as simulating the formation, development and movement of thunderstorms, real-time changes in wind speed and direction, etc. to provide realistic weather environment data for aircraft flight simulation and air traffic controller training.

[0211] Step S151: Perform spatiotemporal consistency verification processing on the meteorological element distribution information in the virtual weather field data set. Check whether the change of meteorological elements between adjacent time steps conforms to the time sequence correlation characteristics in the set of meteorological evolution regularity characteristics.

[0212] The spatiotemporal consistency verification processing first extracts the time series and spatial distribution information of each meteorological element from the virtual weather field data set. For temporal consistency verification, the change amount of meteorological elements between adjacent time steps is calculated and compared with the corresponding time sequence correlation characteristics in the set of meteorological evolution regularity characteristics to check whether the change amount is within a reasonable range and conforms to the change regularity presented in historical data.

[0213] For example, if the set of meteorological evolution regularity characteristics shows that the maximum change in wind speed in a certain airspace between adjacent time steps is a certain value, then check whether the change in wind speed in that airspace between adjacent time steps in the virtual weather field data exceeds this range. If there is an exceeding case, it is marked as a temporal consistency anomaly point.

[0214] For spatial consistency verification, analyze whether the distribution of meteorological elements in different geographical regions within the same time step conforms to the geographical region correlation characteristics in the set of meteorological evolution regularity characteristics. For example, if the characteristic set indicates that the temperature change trend of adjacent airspace should be consistent, then check whether the temperature change direction of adjacent airspace in the virtual weather field data is the same and whether there is a reasonable proportional relationship between the change amplitudes. If the temperature change trend of adjacent airspace is completely opposite and there is no reasonable meteorological explanation, it is marked as a spatial consistency anomaly point.

[0215] After completing the spatiotemporal consistency verification of all meteorological elements, the number and distribution of anomaly points are counted. For anomaly points, they are corrected according to the corresponding meteorological evolution regularity characteristics. The correction method is to adjust the abnormal meteorological element value by referring to the element change under similar meteorological conditions in historical data, so that it meets the spatiotemporal consistency requirements.

[0216] Step S152: Extract the geographical region identification information in the virtual weather field data set, and group the virtual weather field data set according to the geographical region identification information to obtain multiple virtual weather field sub-sets dedicated to geographical regions.

[0217] The geographical area identification information corresponding to each data unit is extracted from the metadata of the virtual weather field data set, which is consistent with the geographical area identification information in the historical weather data set, and can accurately correspond to each control airspace in the air traffic control scene.

[0218] The virtual weather field data set is grouped according to the geographical area identification information, and all virtual weather data units belonging to the same geographical area are grouped into a group to form a virtual weather field sub-set exclusive to the geographical area. For example, the sub-set corresponding to the geographical area with identification information "R1" contains all virtual weather data of time steps in the airspace, including the distribution information of wind speed, wind direction, temperature and other elements.

[0219] During the grouping process, a mapping relationship table between the geographical area identification information and the sub-set is established, which records the storage path and data range of each sub-set corresponding to the identification information, facilitating subsequent quick search and call. At the same time, integrity check is performed on each sub-set to ensure that it contains all weather data of time steps in the geographical area and there is no data loss.

[0220] Step S153: Calculate the statistical characteristics of each virtual weather field sub-set exclusive to the geographical area to generate statistical characteristic descriptors containing mean, variance and extreme value.

[0221] For each virtual weather field sub-set exclusive to the geographical area, the statistical characteristics of each meteorological element are calculated. For mean calculation, the observation values of a meteorological element in the sub-set at all time steps are added up and then divided by the total number of time steps to obtain the mean of the element, reflecting the average level of the element in the entire time period.

[0222] The variance calculation is to first calculate the difference between the value of each time step and the mean, then square the difference and add it up, and then divide by the total number of time steps to obtain the variance, which is used to describe the dispersion degree of the element value.

[0223] The extreme value calculation includes maximum and minimum, which is obtained by traversing the observation values of a meteorological element in the sub-set at all time steps, and finding the maximum and minimum values, reflecting the extreme situation of the element in the entire time period.

[0224] The mean, variance, maximum and minimum of each meteorological element are integrated together to form the statistical characteristic descriptor of the geographical area. For example, the statistical characteristic descriptor of the "R1" geographical area will contain the mean of wind speed, the variance of wind speed, the maximum of wind speed, the minimum of wind speed, and the corresponding statistical characteristics of temperature, pressure and other elements.

[0225] The statistical characteristic descriptor is stored in the form of structured data, and is associated with the corresponding virtual meteorological field sub-set, facilitating subsequent comparison processing.

[0226] Step S154: Comparing the statistical characteristic descriptor with the statistical characteristics of the corresponding geographical area in the historical meteorological data set, adjusting the abnormal value points in the virtual meteorological field sub-set to make the statistical characteristic descriptor consistent with the statistical characteristics of the historical meteorological data set.

[0227] The statistical characteristics of the corresponding geographical area are extracted from the historical meteorological data set, which are obtained by the same calculation method as step S153, ensuring comparability between the two.

[0228] The statistical characteristic descriptor of the virtual meteorological field sub-set is compared with the statistical characteristics of the historical meteorological data one by one. When comparing, the difference value between the two is calculated, for example, the mean difference is the virtual mean minus the historical mean, the variance difference is the virtual variance minus the historical variance, etc.

[0229] Set a difference threshold value, when the difference value of a certain statistical characteristic exceeds the threshold value, it indicates that there is an abnormal value point in the virtual meteorological field sub-set that causes the difference. For example, if the difference value of the wind speed mean exceeds the threshold value, the wind speed value point that causes the mean abnormality needs to be found in the virtual meteorological field sub-set.

[0230] The method of finding abnormal value points is to calculate the deviation of each time step meteorological element value and the corresponding time step element value in the historical data, and determine the value point with larger deviation as the abnormal value point. The adjustment method of abnormal value points is to refer to the element change trend of similar time period in historical data, combined with the characteristics of meteorological evolution law, to correct it, so that the statistical characteristic descriptor of the adjusted virtual meteorological field sub-set is consistent with the statistical characteristics of the historical meteorological data.

[0231] After adjustment, the statistical characteristic descriptor of the virtual meteorological field sub-set is recalculated, and compared with the historical statistical characteristics again until the difference values of all statistical characteristics are within the set threshold value range.

[0232] Step S155: Adding time stamp markers and geographical area identification labels to each geographical area specific virtual meteorological field sub-set after adjustment to generate annotated virtual meteorological field data units.

[0233] The time stamp marker is generated according to the time step corresponding to each data unit in the virtual meteorological field sub-set. The format of the time stamp contains year, month, day, hour, minute and second, and is accurate to the same time granularity as the historical meteorological data, ensuring that it can accurately reflect the time information of the data.

[0234] The geographical area identification tag adopts geographical area identification information corresponding to the virtual weather field sub-set, and is directly added to the metadata of each data unit.

[0235] The timestamp label and the geographical area identification tag are associated with each data unit in the virtual weather field sub-set to form a labeled virtual weather field data unit. Each data unit contains both the distribution information of meteorological elements and the corresponding time and geographical area information, facilitating accurate calling and analysis in subsequent air traffic control scenario applications.

[0236] For example, a labeled virtual weather field data unit may contain a timestamp of "2023-10-01 08:00:00", a geographical area identification of "R1", and meteorological element data such as wind speed, wind direction, and temperature in that time and area.

[0237] Step S156: Arrange all labeled virtual weather field data units in chronological order to form a complete virtual weather field data set.

[0238] Collect all labeled virtual weather field data units generated in all geographical areas and sort them in chronological order according to the timestamp label. During the sorting process, for data units of different geographical areas with the same timestamp, arrange them in alphabetical and numerical order according to the geographical area identification information to form an ordered data set.

[0239] After the arrangement is completed, the complete virtual weather field data set is checked as a whole to ensure that the time sequence of the data units is continuous, there are no time breakpoints, the geographical area coverage is comprehensive, and there is no omission of key airspace. At the same time, verify whether the labeling information of the data units is accurate, and whether the timestamp and geographical area identification are consistent with the data content.

[0240] Step S157: Output the arranged virtual weather field data set for simulating the evolution process of the meteorological scene in a specific geographical area.

[0241] The output virtual weather field data set adopts a data format compatible with the air traffic control system, ensuring that it can be directly read and used by air traffic control simulation training systems, aircraft flight simulation software, and other application programs.

[0242] During the output process, the data is compressed to reduce the overhead of data storage and transmission, and encryption measures are used to protect the security of the data to prevent it from being illegally obtained or tampered with during transmission and use.

[0243] The output virtual weather field data set can be used in various air traffic control scenarios, such as simulating the flight operation of an airspace under different weather conditions to provide a realistic weather environment simulation for controller training, and testing the effectiveness of new air traffic flow management strategies under complex weather conditions.

[0244] In use, virtual meteorological data of a specific geographical area and a specific time period can be extracted as needed, combined with other data of the air traffic control system, such as flight plans, aircraft position information, etc., for comprehensive analysis and simulation, to improve the efficiency and safety of air traffic control.

[0245] Figure 2 A schematic diagram of exemplary hardware and software components of the meteorological data simulation system 100 based on a generative adversarial network that can implement the idea of the present application is shown. For example, the processor 120 can be used on the meteorological data simulation system 100 based on a generative adversarial network and used to perform the functions in the present application.

[0246] For example, the meteorological data simulation system 100 based on a generative adversarial network can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, a ROM, or a RAM, or any combination thereof. Exemplarily, the meteorological data simulation system 100 based on a generative adversarial network can also include program instructions stored in a ROM, a RAM, or other types of non-transitory storage media, or any combination thereof. The methods of the present application can be implemented according to these program instructions. The meteorological data simulation system 100 based on a generative adversarial network also includes an I / O interface 150 between the computer and other input / output devices.

[0247] It should be noted that, in order to simplify the description of the present application and to help understand one or more embodiments of the present application, in the foregoing description of embodiments of the present application, various features are sometimes combined into one embodiment, drawing, or description thereof.

Claims

1.A method for weather data simulation based on a generative adversarial network, characterized in that, The method comprises: acquiring a historical meteorological data set containing meteorological evolution records with time sequence relationship and corresponding geographical area identification information; inputting the historical meteorological data set into a preset time sequence rule learning model for meteorological evolution rule extraction processing to obtain a meteorological evolution rule feature set containing time sequence correlation features and geographical area correlation features of meteorological elements; constructing a conditional generative adversarial network model containing a generator module and a discriminator module, the generator module being configured to receive the meteorological evolution rule feature set as conditional input, and the discriminator module being configured to distinguish between generated meteorological data and real meteorological data; inputting the meteorological evolution rule feature set into the generator module of the conditional generative adversarial network model to generate a virtual meteorological field data set with geographical area identification information, the virtual meteorological field data set containing meteorological element distribution information consistent with the spatiotemporal features of the historical meteorological data set; outputting the virtual meteorological field data set, which is used to simulate the meteorological scenario evolution process of a specific geographical area. 2.The weather data simulation method based on a generative adversarial network according to claim 1, wherein, The method comprises: performing time sequence division processing on the meteorological evolution records in the historical meteorological data set to obtain a plurality of continuous meteorological time sequence segments, each meteorological time sequence segment containing a meteorological element observation value sequence within a preset time length; inputting the meteorological time sequence segments into a long short-term memory network layer of the time sequence rule learning model for time dimension feature extraction processing to generate short-term dependency feature vectors of meteorological elements, the short-term dependency feature vectors being used to represent the meteorological element change correlation of adjacent time steps; inputting the short-term dependency feature vectors into a Transformer encoder layer of the time sequence rule learning model for global time sequence correlation modeling processing to generate long-term dependency feature vectors of meteorological elements, the long-term dependency feature vectors being used to represent the meteorological element change correlation of non-adjacent time steps; extracting geographical area identification information from the historical meteorological data set and performing associated mapping processing on the geographical area identification information and the long-term dependency feature vectors to generate a region correlation feature vector containing geographical area attributes; fusing the short-term dependency feature vectors, the long-term dependency feature vectors and the region correlation feature vector to generate a meteorological evolution rule feature set containing time sequence correlation features and geographical area correlation features, the time sequence correlation features being composed of the short-term dependency feature vectors and the long-term dependency feature vectors, and the geographical area correlation features being composed of the region correlation feature vector. 3.The weather data simulation method based on a generative adversarial network according to claim 1, wherein, The method comprises: The network structure of the generator module is configured, and the generator module comprises an input embedding layer, a feature expansion layer, a spatial feature generation layer and an output mapping layer connected in sequence, the input embedding layer is used for receiving the meteorological evolution law feature set and performing dimension mapping processing; The network structure of the discriminator module is configured, and the discriminator module comprises a multi-scale feature extraction layer, a conditional feature fusion layer and a discrimination decision layer connected in sequence, the multi-scale feature extraction layer is used for extracting different spatial scale features of input meteorological data, and the conditional feature fusion layer is used for fusing the meteorological evolution law feature set and the extracted spatial scale features; The loss function of the generator module and the discriminator module is set, the loss function comprises a generator loss term and a discriminator loss term, the generator loss term is used for measuring the difference between the generated meteorological data and the real meteorological data, and the discriminator loss term is used for measuring the discrimination ability of the discriminator module to the real meteorological data and the generated meteorological data; The network parameters of the generator module and the discriminator module are initialized, the network parameters comprise weight matrices and bias vectors of each layer, and the initialization mode adopts random normal distribution sampling; The generator module and the discriminator module are connected to form an end-to-end conditional generative adversarial network model, wherein the output of the generator module is used as one of the inputs of the discriminator module, and the meteorological evolution law feature set is simultaneously input into the conditional feature fusion layer of the generator module and the discriminator module. 4.The weather data simulation method based on a generative adversarial network according to claim 2, wherein, The meteorological time series segment is input into the long short-term memory network layer of the time series law learning model to perform time dimension feature extraction processing and generate a short-term dependence feature vector of a meteorological element, including: The sequence alignment processing is performed on the meteorological element observation value sequence in the meteorological time series segment, so that the number of time steps of each meteorological time series segment is consistent; The aligned meteorological element observation value sequence is input into the input gate unit of the long short-term memory network layer to calculate the fusion weight of the input information and the historical memory information of the current time step and generate an input gate activation value; The aligned meteorological element observation value sequence is input into the forget gate unit of the long short-term memory network layer to calculate the retention weight of the historical memory information and generate a forget gate activation value; The cell state vector is updated based on the forget gate activation value, and the key part of the historical memory information is retained and the redundant part is discarded; The updated cell state vector is input into the output gate unit of the long short-term memory network layer to calculate the output weight of the cell state vector and generate an output gate activation value; According to the output gate activation value and the tanh activation result of the updated cell state vector, a hidden state vector of the current time step is generated. The hidden state vectors of all time steps in the meteorological time series segment are spliced in time sequence to generate a short-term dependence feature vector representing the change correlation of adjacent time steps. 5.The weather data simulation method based on a generative adversarial network according to claim 3, wherein, The network structure of the generator module comprises: The mapping parameters of the input embedding layer are set, and the mapping parameters include an embedding matrix, a number of rows of the embedding matrix is consistent with a feature dimension of the meteorological evolution law feature set, and a number of columns of the embedding matrix is consistent with a preset embedding dimension; The feature expansion layer is configured as a transposed convolution layer stack structure, the transposed convolution layer stack structure includes a plurality of transposed convolution units connected in sequence, each transposed convolution unit includes a transposed convolution operation and a batch normalization operation, and a number of the transposed convolution units is determined according to a spatial resolution requirement of the target virtual meteorological field data set; Network parameters of the spatial feature generation layer are set, the spatial feature generation layer includes a convolution block attention module, the convolution block attention module is used for calculating channel attention weights and spatial attention weights of feature maps output by the feature expansion layer, and a weighted spatial feature map is generated; The output mapping layer is configured as a combination structure of a convolution layer and an activation function, a number of output channels of the convolution layer is consistent with a number of meteorological element types included in the virtual meteorological field data set, and the activation function is used for mapping output feature values to a preset meteorological element numerical range; The input embedding layer, the feature expansion layer, the spatial feature generation layer and the output mapping layer are connected to form a forward propagation path of the generator module, wherein an output of the input embedding layer is used as an input of the feature expansion layer, an output of the feature expansion layer is used as an input of the spatial feature generation layer, and an output of the spatial feature generation layer is used as an input of the output mapping layer. 6.The weather data simulation method based on a generative adversarial network according to claim 3, wherein, The network structure of the discriminator module is configured, including: The multi-scale feature extraction layer is set as a convolution layer pyramid structure, the convolution layer pyramid structure includes a plurality of convolution units with different convolution kernel sizes, each convolution unit is used for extracting a local feature map of input meteorological data at a corresponding spatial scale, and the convolution kernel size gradually increases from a first layer to a last layer; Global average pooling processing is performed on the local feature map output by each convolution unit to generate a global feature vector at a corresponding spatial scale, and a dimension of the global feature vector is consistent with an output channel number of the convolution unit; A fusion strategy of the conditional feature fusion layer is configured, the fusion strategy includes a feature splicing operation and an attention weighting operation, the feature splicing operation is used for splicing global feature vectors at different spatial scales along a channel dimension to generate a multi-scale spliced feature vector; The meteorological evolution law feature set is mapped to the same dimension as the multi-scale spliced feature vector through a fully connected layer to generate a conditional mapping feature vector; Attention weighting operations are performed on the multi-scale spliced feature vector and the conditional mapping feature vector to calculate a fusion weight of each feature channel, and a weighted feature vector of fusion conditional information is generated; The discriminant decision layer is set as a combination structure of a fully connected layer and a sigmoid activation function, an input dimension of the fully connected layer is consistent with a dimension of the weighted feature vector, an output dimension of the fully connected layer is 1, and the sigmoid activation function is used for outputting a probability value of the input meteorological data being real data. The multi-scale feature extraction layer, the conditional feature fusion layer and the discriminative decision layer are connected to form a forward propagation path of the discriminator module, wherein the output of the multi-scale feature extraction layer is taken as the input of the conditional feature fusion layer, and the output of the conditional feature fusion layer is taken as the input of the discriminative decision layer. 7.The weather data simulation method based on a generative adversarial network according to claim 1, wherein, The meteorological evolution rule feature set is input into the generator module of the conditional generative adversarial network model to generate a virtual meteorological field data set with geographical area identification information, including: extracting geographical area correlation features in the meteorological evolution rule feature set as geographical condition constraint information; inputting the meteorological evolution rule feature set into the input embedding layer of the generator module for dimension mapping processing through an embedding matrix to generate an embedding feature vector; inputting the embedding feature vector into the feature expansion layer of the generator module for feature dimension expansion processing through a transposed convolution layer stacking structure to generate a high-dimensional feature map, wherein the spatial dimension of the high-dimensional feature map matches the spatial resolution of the target virtual meteorological field data set; inputting the high-dimensional feature map into the spatial feature generation layer of the generator module, combining the geographical condition constraint information, calculating channel attention weights and spatial attention weights through a convolution block attention module, and performing weighting processing on the high-dimensional feature map to generate a spatial weighted feature map with geographical area features; inputting the spatial weighted feature map into the output mapping layer of the generator module, mapping the feature values into the meteorological element numerical range through a convolution layer and an activation function to generate preliminary virtual meteorological field data; performing spatio-temporal feature comparison processing on the preliminary virtual meteorological field data and the meteorological evolution records in the historical meteorological data set to adjust the output mapping layer parameters of the generator module, so that the spatio-temporal features of the preliminary virtual meteorological field data are consistent with the spatio-temporal features of the historical meteorological data set; outputting the adjusted preliminary virtual meteorological field data as the virtual meteorological field data set with geographical area identification information. 8.The weather data simulation method based on a generative adversarial network according to claim 7, characterized in that, The high-dimensional feature map is input into the spatial feature generation layer of the generator module, combining the geographical condition constraint information, calculating channel attention weights and spatial attention weights through a convolution block attention module, and performing weighting processing on the high-dimensional feature map to generate a spatial weighted feature map with geographical area features, including: inputting the high-dimensional feature map into the channel attention submodule of the convolution block attention module to perform global average pooling and global maximum pooling processing on the high-dimensional feature map to generate a channel average feature vector and a channel maximum feature vector; inputting the channel average feature vector and the channel maximum feature vector into a double-channel fully connected network to perform feature fusion processing through a shared hidden layer to generate a channel fusion feature vector; performing sigmoid activation processing on the channel fusion feature vector to generate channel attention weights for each feature channel; performing channel-by-channel multiplication processing on the high-dimensional feature map and the channel attention weights to generate a channel weighted feature map; The channel-weighted feature map is input into a spatial attention submodule of the convolution block attention module, and average pooling and maximum pooling processing are performed on the channel-weighted feature map in the channel dimension to generate a spatial average feature map and a spatial maximum feature map; The spatial average feature map and the spatial maximum feature map are spliced along the channel dimension to generate a spatial spliced feature map; The spatial spliced feature map is subjected to convolution operation and sigmoid activation processing to generate a spatial attention weight map; The channel-weighted feature map and the spatial attention weight map are subjected to pixel-by-pixel multiplication processing to generate a preliminary spatial weighted feature map; The geographic condition constraint information is mapped into a geographic weight map consistent with the spatial dimension of the preliminary spatial weighted feature map through a fully connected layer; The preliminary spatial weighted feature map and the geographic weight map are subjected to pixel-by-pixel addition processing to generate a spatial weighted feature map with geographic area features. 9.The weather data simulation method based on a generative adversarial network according to claim 1, wherein, The output of the virtual meteorological field data set includes: The spatial and temporal consistency of the meteorological element distribution information in the virtual meteorological field data set is checked to check whether the change of the meteorological element at adjacent time steps conforms to the time sequence correlation feature in the meteorological evolution rule feature set; Geographic area identification information in the virtual meteorological field data set is extracted, and the virtual meteorological field data set is grouped according to the geographic area identification information to obtain a plurality of virtual meteorological field sub-sets dedicated to geographic areas; Meteorological element statistical feature calculation is performed on each virtual meteorological field sub-set dedicated to a geographic area to generate statistical feature descriptors including mean, variance and extreme value; The statistical feature descriptors are compared with the statistical features of the corresponding geographic area in the historical meteorological data set to adjust the abnormal value points in the virtual meteorological field sub-set, so that the statistical feature descriptors are consistent with the statistical features of the historical meteorological data set; A timestamp label and a geographic area identification tag are added to each adjusted virtual meteorological field sub-set dedicated to a geographic area to generate a labeled virtual meteorological field data unit; All labeled virtual meteorological field data units are arranged in time sequence order to form a complete virtual meteorological field data set; The arranged virtual meteorological field data set is output for simulating the meteorological scene evolution process of a specific geographic area. 10.A weather data simulation system based on a generative adversarial network, characterized in that, The device includes a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to realize the meteorological data simulation method based on the generative adversarial network in any one of claims 1-9.

Images