Method for predicting content of supercooled water in cloud based on deep learning and spatio-temporal data driving

Abstract

The invention discloses a cloud supercooled water content prediction method based on deep learning and spatio-temporal data driving, and the method comprises the steps: obtaining satellite multi-channel radiation transmission data, carrying out the channel screening and feature extraction, obtaining radar reflectivity data, and carrying out the data interpolation and feature extraction; the method comprises the following steps: acquiring conventional meteorological data of a ground automatic meteorological station, performing feature extraction, fusing the features by using a multi-head attention mechanism to obtain fused features, establishing a deep learning neural network model, and predicting the content of supercooled water in cloud in a future time period. The method can more comprehensively and accurately predict the content and the distribution condition of the supercooled water in the cloud, and has important scientific and practical significance for relieving water resource shortage, improving the ecological environment and guaranteeing low-altitude economic safety.

Description

Technical Field

[0001] The present invention relates to the field of supercooled water prediction, and particularly to a method for predicting the content of supercooled water in clouds driven by deep learning and spatio-temporal data. Background Art

[0002] Supercooled water refers to liquid water with a temperature below 0°C but not frozen. Accurately grasping and precisely forecasting the content of supercooled water in clouds is of great significance for meteorological forecasting, agricultural production, ecological environment restoration, power safety, and aviation safety. With the continuous progress of meteorological observation technology, the amount of data generated by various observation means has increased explosively. These data contain rich spatio-temporal information, providing a new opportunity for predicting the content of supercooled water in clouds.

[0003] Currently, the means for forecasting supercooled water mainly rely on numerical models. Due to reasons such as the setting of the initial field and parameter settings in the models, there are some uncertainties in the forecasting results. Especially in some severe convective weather conditions, it is difficult for the models to accurately grasp the development and evolution of clouds. At the same time, although aircraft onboard detection equipment and ground-based observations can provide real-time detection of the supercooled water content along the route or at a single point, they cannot achieve extrapolation forecasting over a horizontal range, resulting in deviations in the evaluation results and being unable to accurately forecast the content of supercooled water in clouds. Therefore, the present invention proposes a method for predicting the content of supercooled water in clouds driven by deep learning and spatio-temporal data. It uses a convolutional neural network to extract features from multi-dimensional real-time meteorological data, and utilizes a multi-head attention mechanism for feature fusion to establish a deep learning neural network. The bitter fish algorithm is used to optimize the parameters during the network training process to achieve the prediction of the supercooled water distribution in clouds for the next 3 hours. This method effectively integrates satellite, radar, and conventional meteorological data, extracts and fuses the features of these data, efficiently captures the supercooled water information contained in multi-source observation data, and effectively improves the accuracy of supercooled water content prediction, which has important scientific and practical significance for alleviating water resource shortages, improving the ecological environment, and ensuring agricultural production safety, power safety, and aviation safety. Summary of the Invention

[0004] The object of the present invention is to provide a method for predicting the content of supercooled water in clouds driven by deep learning and spatio-temporal data.

[0005] To achieve the above object, the present invention is implemented according to the following technical solution: The present invention includes the following steps: Obtain satellite multi-channel radiation data for the past 6 hours, perform calibration, projection conversion, and channel screening on the multi-channel radiation data, and use a first feature extraction model to extract features to obtain satellite features; the dimensions of the multi-channel radiation data include the number of samples, the number of channels, the number of meridional grid points, the number of latitudinal grid points, and the time step; Retrieve radar data for the past 6 hours and use the second feature extraction model to extract features to obtain radar features; the dimensions of the radar data include the number of features, the number of grid points in the meridional direction, the number of grid points in the zonal direction, and the time step; Retrieve the conventional meteorological data observed by the surface automatic weather station for the past 6 hours and use the third feature extraction model to extract features to obtain conventional meteorological features; the conventional meteorological data includes temperature, pressure, wind direction, wind speed, precipitation, and relative humidity; the dimensions of the conventional meteorological data include the number of features, the number of stations, and the time step; Use the multi-head attention mechanism to fuse the satellite features, the radar features, and the conventional meteorological features to obtain fused features; Construct a deep learning network based on the fused features and the corresponding supercooled water content at future times, determine the optimization objective function, and optimize the deep learning network according to the optimization objective function; the deep learning network is a ConvLSTM-Transformer hybrid network; Input the fused features to be predicted into the optimized deep learning network to obtain the predicted value of the supercooled water content in clouds for the next 3 hours; The method for fusing features to obtain fused features includes the following steps: Use the multi-head attention mechanism to fuse the satellite features, the radar features, and the conventional meteorological features to obtain fused features; the attention mechanism module uses the multi-head attention mechanism for feature fusion, and the expression is:

[0006] where is the attention mechanism, is the query vector, is the key vector, is the value vector, is the dimension of the key vector for scaling, is the multi-head attention, is the th hidden state of the head, , is the weight matrix used to combine the outputs of each head, , , are learnable weight matrices that map the inputs , , to different subspaces respectively; The method for determining the optimization objective function is based on the input feature category and the number of features Determine the optimization objective function of the deep learning network: where is the optimization objective function, is the threshold stability weight, is the eigenvalue deviation weight, is the time constraint weight, is the space constraint weight, is the threshold type weight, is the predicted value of the class dynamic recognition threshold, is the historical mean of the class threshold, calculated using a sliding window, is the grid point in the weather feature value of class the time penalty coefficient, is the space penalty coefficient, is the th input feature class, is the input feature class index.

[0007] Furthermore, the method for obtaining satellite features includes the following steps: Calibrate, project and channel filter the multi-channel radiation data to obtain the first radiation data; the data dimensions of the multi-channel radiation data include the number of samples, the number of channels, the number of longitude grid points, the number of latitude grid points and the time step; Input the first radiation data into the first feature extraction model for feature extraction to obtain satellite features; the first feature extraction model includes an input layer, a convolutional layer, a pooling layer, an activation layer, a fully connected layer and an output layer; The input layer is used to receive the first radiation data; The convolutional layer uses a 4D convolutional neural network to perform convolutional operations on the input data to capture the local features of the data in space and time to obtain the first spatio-temporal radiation features; the convolutional kernel of the 4D convolutional neural network uses 2×3×3×3 or 2×5×5×3, and the stride is 2×2×2×2; The pooling layer downsamples the first spatio-temporal radiation features to simplify the dimensions of the first spatio-temporal radiation features to obtain the second spatio-temporal radiation features; the pooling layer performs pooling operations in the space and time dimensions; The activation layer uses an activation function to perform a non-linear transformation on the output of the convolutional layer or the pooling layer; the activation layer includes a first activation layer and a second activation layer; the first activation layer receives the output of the convolutional layer and outputs it to the pooling layer; the second activation layer receives the output of the pooling layer and outputs it to the fully connected layer; The fully connected layer maps the second spatio-temporal radiation feature after convolution, pooling, and activation operations to the output space. The fully connected layer contains multiple neurons, performs a linear transformation on the second spatio-temporal radiation feature through a weight matrix and a bias term, and obtains a satellite feature through non-linear processing using an activation function. The output layer is connected to the output of the fully connected layer to output the satellite feature.

[0008] Further, the method for obtaining the radar feature includes the following steps: Input the radar data into the second feature extraction model for feature extraction to obtain the radar feature; the dimensions of the radar data include the number of samples, the number of features, the number of longitude grid points, the number of latitude grid points, and the number of time steps; the second feature extraction model includes an input layer, a convolutional layer, a pooling layer, an activation layer, a fully connected layer, and an output layer. The input layer is used to receive the radar data. The convolutional layer uses a 4D convolutional neural network to perform a convolution operation on the input data to capture the local features of the data in space and time to obtain the first spatio-temporal radar feature; the convolutional kernel of the 4D convolutional neural network uses 1×3×3×3 or 1×5×5×3, and the stride is 2×2×2×2. The pooling layer downsamples the first spatio-temporal radar feature to simplify the dimensions of the first spatio-temporal radar feature to obtain the second spatio-temporal radar feature; the pooling layer performs a pooling operation in the spatial and temporal dimensions. The activation layer uses an activation function to perform a non-linear transformation on the output of the convolutional layer or the pooling layer; the activation layer includes a first activation layer and a second activation layer; the first activation layer receives the output of the convolutional layer and outputs it to the pooling layer; the second activation layer receives the output of the pooling layer and outputs it to the fully connected layer. The fully connected layer maps the second spatio-temporal radar feature after convolution, pooling, and activation operations to the output space. The fully connected layer contains multiple neurons, performs a linear transformation on the second spatio-temporal radar feature through a weight matrix and a bias term, and obtains a satellite feature through non-linear processing using an activation function. The output layer is connected to the fully connected layer to output the radar feature.

[0009] Further, the method for obtaining the conventional meteorological feature includes the following steps: Input the conventional meteorological data into the third feature extraction model for feature extraction to obtain the conventional meteorological feature; the conventional meteorological data includes temperature, pressure, wind direction, wind speed, precipitation, and relative humidity; the dimensions of the conventional meteorological data include the number of features, the number of stations, and the number of time steps; the third feature extraction model includes an input layer, a convolutional layer, a pooling layer, an activation layer, a fully connected layer, and an output layer. The input layer is used to receive the conventional meteorological data. The convolutional layer uses a 3D convolutional neural network to perform a convolution operation on the input data, capturing the local features of the data in space and time to obtain the first spatio-temporal conventional meteorological features; the convolutional kernel of the 3D convolutional neural network uses a convolutional kernel of 2×3×3 or 2×5×3, and the stride is 2×2×2; The pooling layer downsamples the first spatio-temporal conventional meteorological data features, simplifying the dimensions of the first spatio-temporal conventional meteorological features to obtain the second spatio-temporal conventional meteorological features; the pooling layer performs pooling operations in the spatial and temporal dimensions; The activation layer uses an activation function to perform a non-linear transformation on the output of the convolutional layer or the pooling layer; the activation layer includes a first activation layer and a second activation layer; the first activation layer receives the output of the convolutional layer and outputs it to the pooling layer; the second activation layer receives the output of the pooling layer and outputs it to the fully connected layer; The fully connected layer maps the second spatio-temporal conventional meteorological features after convolution, pooling, and activation operations to the output space. The fully connected layer contains multiple neurons, performs a linear transformation on the second spatio-temporal conventional meteorological features through a weight matrix and a bias term, and performs non-linear processing through an activation function to obtain the conventional meteorological features; The output layer is connected to the fully connected layer to output radar features.

[0010] Furthermore, the method for optimizing the deep learning network according to the optimization objective function includes: Using the bitter fish optimization algorithm to determine the optimal network model parameters, define as the th network model parameter, corresponding to the bitter fish spawning point, as the input feature dimension within the historical 3 hours, and a bitter fish population is composed of multiple bitter fish individuals , as the population size, and at the same time initialize the population with the standard network model parameters to obtain , where the boundary is , is the random perturbation number; Search for suitable oysters to determine the bitter fish spawning point and update the bitter fish position (update the dynamic network model parameters). When an oyster is successfully captured, the bitter fish position update expression is: where is the updated position of the th fish in the th iteration, is the current position of the th fish in the th iteration, is the dynamic inertia weight, is the number of steps that the bitter fish moves to the escaping oyster at the -th iteration, , is a random number, is the best oyster, that is, the optimal spawning point attracting the bitter fish, is the most valuable oyster randomly selected from the population, , are the maximum inertia weight and the minimum inertia weight, is the population variance at the -th iteration, is the initial population variance, is the maximum number of iterations, is a random function; When the oyster escapes, re-exploration is carried out, and the expression for updating the position of the bitter fish is: where is the escape reset probability, taking ; According to the position of the successfully captured oyster, the female fish lays eggs at this position to generate a new individual. The expression for updating the position of the new individual bitter fish is: where is the position of the newly generated bitter fish at the -th iteration, is the distribution radius of the newly generated bitter fish inside the oyster shell, is Gaussian noise with a mean of 0 and a variance of ; Calculate the population fitness and determine the best oyster and the most valuable oyster in the population. The expression is:

[0011]

[0012] where is the population fitness, is the dynamic weight of the fitness, is the position of the -th iteration and the -th dynamic network model parameter, is the position of the -th standard network model parameter, is the position corresponding to the maximum value of the -th dynamic network model parameter, The The position corresponding to the minimum value of the dynamic network model parameter, is the weight of the coordination entropy, is the coordination entropy of the th dynamic recognition threshold, is the time-varying penalty weight, is the time-varying penalty decay rate, is the th iteration of the th position of the dynamic network model parameter and the correlation with the target position ; Adopt the strategy of oysters hunting newborn bitter fish to eliminate population individuals. The elimination probability of newborn bitter fish is: where is the elimination probability of newborn bitter fish, is the population selection pressure, is the fitness of the th iteration of the newborn bitter fish population, is the fitness of the new bitter fish population containing newborn bitter fish and the original bitter fish at the th iteration, is the diversity penalty weight, is the th iteration of the population diversity, population historical average position; Repeat the above operations until the optimization objective function is minimized or the maximum number of iterations is reached, then stop the iteration, output the optimal network model parameters, and obtain the optimized deep learning network.

[0013] Furthermore, the satellite features include satellite multi-channel radiation data; the radar features include composite reflectivity and VIL; the conventional meteorological features include temperature, relative humidity, wind direction, wind speed, and precipitation.

[0014] The beneficial effects of the present invention are: The present invention is a method for predicting the supercooled water content in clouds based on deep learning and spatio-temporal data-driven. Compared with the prior art, the present invention has the following technical effects: Through steps of feature extraction, feature fusion, model construction, and parameter optimization, the present invention can enhance the data preprocessing ability in the dynamic evaluation of the potential area for cold cloud artificial rainfall augmentation, improve the effective extraction and integration of the information on the supercooled water content in clouds contained in spatio-temporal data, increase the data utilization rate and reduce the computational amount, and improve the accuracy and timeliness of the prediction of the supercooled water content in clouds. It provides scientific and intelligent decision-making support for the analysis of operation conditions and the dynamic identification of potential areas during cold cloud artificial rainfall augmentation operations, and has important application value for alleviating water resource protection, improving the ecological environment, and ensuring agricultural production safety, power safety, and aviation safety. Brief Description of the Drawings

[0015] Figure 1 It is a flowchart of the steps of the method for predicting the supercooled water content in clouds based on deep learning and spatio-temporal data driving of the present invention. Detailed Embodiments

[0016] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but do not limit the present invention.

[0017] The method for predicting the supercooled water content in clouds based on deep learning and spatio-temporal data driving of the present invention includes the following steps: As Figure 1 shown, in this embodiment, it includes the following steps: Obtain satellite multi-channel radiation data in the past 6 hours, perform calibration, projection conversion, and channel screening on the multi-channel radiation data, and use a first feature extraction model to perform feature extraction to obtain satellite features; the dimensions of the multi-channel radiation data include the number of samples, the number of channels, the number of meridional grid points, the number of zonal grid points, and the time step; Obtain radar data in the past 6 hours, and use a second feature extraction model to perform feature extraction to obtain radar features; the dimensions of the radar data include the number of features, the number of meridional grid points, the number of zonal grid points, and the time step; Obtain conventional meteorological data observed by a ground automatic station in the past 6 hours, and use a third feature extraction model to perform feature extraction to obtain conventional meteorological features; the conventional meteorological data includes temperature, pressure, wind direction, wind speed, precipitation, and relative humidity; the dimensions of the conventional meteorological data include the number of features, the number of stations, and the time step; Use a multi-head attention mechanism to perform feature fusion on the satellite features, the radar features, and the conventional meteorological features to obtain fused features; Construct a deep learning network according to the fused features and the corresponding supercooled water content at a future time, determine an optimization objective function, and optimize the deep learning network according to the optimization objective function; the deep learning network is a ConvLSTM-Transformer hybrid network; Input the fused features to be predicted into the optimized deep learning network to obtain the predicted value of the supercooled water content in clouds for the next 3 hours; The method for performing feature fusion to obtain fused features includes the following steps: Use the multi-head attention mechanism to perform feature fusion on the satellite features, the radar features, and the conventional meteorological features to obtain fused features; the attention mechanism module uses the multi-head attention mechanism for feature fusion, and the expression is:

[0018] where is the attention mechanism, is the query vector, is the key vector, is the value vector, is the dimension of the key vector for scaling, is the multi-head attention, is the hidden state of the -th head, is the weight matrix used to combine the outputs of each head, and and are learnable weight matrices that map the inputs and and to different subspaces respectively; The method for determining the optimization objective function is to determine the optimization objective function of the deep learning network according to the input feature category and the number of features : where is the optimization objective function, is the threshold stability weight, is the eigenvalue deviation weight, is the time constraint weight, is the space constraint weight, is the threshold type weight, is the predicted value of the dynamic recognition threshold of the -th class, is the historical mean of the threshold of the -th class of grid calculated using a sliding window, is the weather feature value of the -th class in the grid is the spatial penalty coefficient, is the th input feature category, is the input feature category index.

[0019] In this embodiment, the method for obtaining satellite features includes: Performing calibration, projection conversion, and channel screening on multi-channel radiation data to obtain first radiation data; the data dimensions of the multi-channel radiation data include the number of samples, the number of channels, the number of grid points in the meridional direction, the number of grid points in the zonal direction, and the time step.

[0020] In this embodiment, the method for obtaining radar features includes: Inputting radar data into a second feature extraction model for feature extraction to obtain radar features; the data dimensions of the radar data include the number of samples, the number of features, the number of grid points in the meridional direction, the number of grid points in the zonal direction, and the time step; the second feature extraction model includes an input layer, a convolutional layer, a pooling layer, an activation layer, a fully connected layer, and an output layer.

[0021] In this embodiment, the method for obtaining conventional meteorological features includes: Inputting conventional meteorological data into a third feature extraction model for feature extraction to obtain conventional meteorological features; the conventional meteorological data includes temperature, pressure, wind direction, wind speed, precipitation, and relative humidity; the data dimensions of the conventional meteorological data include the number of features, the number of stations, and the time step; the third feature extraction model includes an input layer, a convolutional layer, a pooling layer, an activation layer, a fully connected layer, and an output layer.

[0022] In this embodiment, the method for constructing a deep learning network includes: Taking the fused features and the corresponding supercooled water content at future times as a comprehensive set, dividing the comprehensive set into a training set and a test set according to 6:4, training the deep learning network using the training set, and evaluating the performance of the deep learning network using the test set.

[0023] In this embodiment, the method for optimizing the deep learning network according to the optimization objective function includes: Using the bitter fish optimization algorithm to determine the optimal network model parameters, defining as the th network model parameter, corresponding to the bitter fish spawning point, is the input feature dimension within the past 3 hours, forming a bitter fish population consisting of multiple bitter fish individuals , is the population size, and at the same time initializing the population with the standard network model parameters to obtain , where the boundary is , is the random perturbation number; Search for suitable oysters to determine the spawning points of bitter fish and update the positions of bitter fish (update the parameters of the dynamic network model). When an oyster is successfully captured, the expression for updating the position of the bitter fish is: where is the updated position of the -th fish in the -th iteration, is the current position of the -th fish in the -th iteration, is the dynamic inertia weight, is the number of steps for the bitter fish to move to the escaping oyster at the -th iteration, and are random numbers, is the best oyster, i.e., the optimal spawning point that attracts the bitter fish, is the most valuable oyster randomly selected from the population, and are the maximum inertia weight and the minimum inertia weight, is the population variance at the -th iteration, is the initial population variance, is the maximum number of iterations, is the random function; When the oyster escapes, re-exploration is carried out, and the expression for updating the position of the bitter fish is: where is the escape reset probability, taking ; According to the position of the successfully captured oyster, the female fish spawns new individuals at this position. The expression for updating the position of the new individual bitter fish is: where is the position of the newly generated bitter fish at the -th iteration, is the distribution radius of the newly generated bitter fish inside the oyster shell, is Gaussian noise with a mean of 0 and a variance of ; Calculate the population fitness to determine the best oyster and the most valuable oyster in the population, and the expression is:

[0024]

[0025] where is the population fitness, is the dynamic weight of fitness, is the th iteration and the th position of the dynamic network model parameter, is the th position of the standard network model parameter, is the th position corresponding to the maximum value of the dynamic network model parameter, The th position corresponding to the minimum value of the dynamic network model parameter, is the weight of the coordination entropy, is the th coordination entropy of the dynamic recognition threshold, is the time-varying penalty weight, is the time-varying penalty decay rate, is the th iteration and the th correlation between the position of the dynamic network model parameter and the target position ; Adopt the strategy of oyster hunting and killing newly born bitter fish to eliminate individuals in the population. The elimination probability of newly born bitter fish is: where is the elimination probability of newly born bitter fish, is the population selection pressure, is the fitness of the newly born bitter fish population at the th iteration, is the fitness of the new bitter fish population including newly born bitter fish and original bitter fish at the th iteration, is the diversity penalty weight, is the th population diversity at the th iteration; Repeat the above operations until the optimization objective function is minimized or the maximum number of iterations is reached, then stop the iteration, output the optimal network model parameters, and obtain the optimized deep learning network; In this embodiment, the satellite features include satellite multi-channel radiation data; the radar features include composite reflectivity and VIL; the conventional meteorological features include temperature, relative humidity, wind direction, wind speed, and precipitation.

[0026] In actual evaluation, to predict the supercooled water distribution in a certain area at a certain moment, it is necessary to first use historical data over the years to train and model a deep learning network model. Obtain the hourly satellite radiation data (15 channels), weather radar network mosaic data (composite reflectivity and VIL), and conventional meteorological data (temperature, pressure, relative humidity, wind direction, wind speed, precipitation) from 20:00 on July 21, 2024 to 23:00 on July 21, 2024. The cumulative duration of the concerned time period is 3 hours; Use the first feature extraction model to extract satellite multi-channel radiation data to obtain satellite features, use the second feature extraction model to extract radar data to obtain radar features, and use the third feature extraction model to extract conventional meteorological data to obtain conventional meteorological features; Use the multi-head attention mechanism to fuse satellite features, radar features, and conventional meteorological features; The size of the feature image (height * width) obtained by processing with the multi-branch CNN network is 4 * 4, and the number of channels is 4. The channel attention mechanism divides the feature map into two groups (height + temperature, wind speed + humidity). The lightweight multi-layer perceptron can learn the importance weights between groups, which are 0.6 and 0.4 respectively. The weights of color, texture, and shape features after being weighted by the channel attention are 0.40, 0.30, and 0.30 respectively; The size convolution kernel weights corresponding to the multi-head attention mechanism are 3 * 3 / 0.4, 5 * 5 / 0.3, and 7 * 7 / 0.7 respectively. The polar coordinate convolution (radial distance takes 1, the angle takes 45°) weight takes 0.1, the variance on the position takes 0.2, the variance adjustment coefficient takes 0.5. The weights of satellite, radar, and conventional meteorological features after being weighted by attention are 0.30, 0.35, and 0.35 respectively; Adopt a fully connected module to splice the weighted features and output the field features corresponding to 3 moments; Optimization objective function The threshold stability weight takes 0.3, the eigenvalue deviation weight takes 0.5, the time constraint weight takes 0.2, the space constraint weight takes 0.1, the threshold type weight takes (satellite feature 0.30, radar feature 0.35, conventional meteorological feature 0.35), the time penalty coefficient takes 0.1, the space penalty coefficient takes 0.05; Initial population size takes 10, initialize the network weight parameters, , the maximum inertia weight and the minimum inertia weight are taken as and , the maximum number of iterations is taken as 100 times; Taking the optimization of the prediction model parameters in the first prediction time period (21:00 - 23:00 on the 21st) as an example, the prediction step is 3h, and the lag time step is 6h (15:00 - 20:00 on the 21st). When the fish-swarm search algorithm is iterated to the 56th, 57th, 58th, and 59th times, the optimization objective function is 0.11, 0.09, 0.09, and 0.09 respectively. The optimal position of the population at the 57th iteration is taken as the optimized total population position, and the optimized network weight parameters are output; the fused features of the past 6h are input into the constructed prediction network to obtain the prediction result of the supercooled water content in the next 3h.

[0027] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for predicting the supercooled water content in clouds based on deep learning and spatio-temporal data-driven, characterized in that, It includes the following steps: S1. Obtain satellite multi-channel radiation data in the past 6 hours, perform calibration, projection transformation, and channel screening on the multi-channel radiation data, and use a first feature extraction model to extract features to obtain satellite features; The dimensions of the multi-channel radiation data include the number of samples, the number of channels, the number of meridional grid points, the number of latitudinal grid points, and the time step; S2. Obtain radar data in the past 6 hours, and use a second feature extraction model to extract features to obtain radar features; the dimensions of the radar data include the number of features, the number of meridional grid points, the number of latitudinal grid points, and the time step; S3. Obtain conventional meteorological data observed by a ground automatic station in the past 6 hours, and use a third feature extraction model to extract features to obtain conventional meteorological features; The conventional meteorological data includes temperature, air pressure, wind direction, wind speed, precipitation, and relative humidity; the dimensions of the conventional meteorological data include the number of features, the number of stations, and the time step; S4. Use a multi-head attention mechanism to fuse the satellite features, the radar features, and the conventional meteorological features to obtain fused features; S5. Construct a deep learning network based on the fused features and the corresponding supercooled water content in the future time, determine an optimization objective function, and optimize the deep learning network according to the optimization objective function; the deep learning network is a ConvLSTM-Transformer hybrid network; S6. Input the fused features to be predicted into the optimized deep learning network to obtain a predicted value of the supercooled water content in clouds in the next 3 hours; The method for fusing features to obtain fused features includes the following steps: Use a multi-head attention mechanism to fuse the satellite features, the radar features, and the conventional meteorological features to obtain fused features; The attention mechanism module uses the multi-head attention mechanism for feature fusion, and the expression is: Among them is the attention mechanism, is the query vector, is the key vector, is the value vector, is the dimension of the key vector for scaling, is the multi - head attention, is the th hidden state of the head, , is the weight matrix used to combine the outputs of each head, , , are learnable weight matrices that map the inputs , , to different sub - spaces respectively; The method for determining the optimization objective function is to determine the optimization objective function of the deep learning network according to the input feature category and the number of features : Among them is the optimization objective function is the threshold stability weight is the eigenvalue deviation weight is the time constraint weight is the space constraint weight is the threshold type weight is the predicted value of the class dynamic recognition threshold is the historical mean of the class threshold, calculated using a sliding window is the grid in class weather eigenvalue time penalty coefficient is the space penalty coefficient is the th input feature category is the input feature category index 2. The method for predicting the supercooled water content in clouds based on deep learning and spatio-temporal data driving according to claim 1, wherein The method for obtaining satellite features includes: Perform calibration, projection transformation, and channel screening on the multi-channel radiation data to obtain first radiation data; the data dimensions of the multi-channel radiation data include the number of samples, the number of channels, the number of meridional grid points, the number of latitudinal grid points, and the time step; Input the first radiation data into a first feature extraction model to extract features to obtain satellite features; the first feature extraction model includes an input layer, a convolutional layer, a pooling layer, an activation layer, a fully connected layer, and an output layer; The input layer is used to receive the first radiation data; The convolutional layer uses a 4D convolutional neural network to perform a convolution operation on the input data to capture local features of the data in space and time to obtain first spatio-temporal radiation features; the convolutional kernel of the 4D convolutional neural network uses 2×3×3×3 or 2×5×5×3, and the stride is 2×2×2×2; The pooling layer downsamples the first spatio-temporal radiation features to simplify the dimensions of the first spatio-temporal radiation features to obtain second spatio-temporal radiation features; the pooling layer performs pooling operations in the spatial and temporal dimensions; The activation layer uses an activation function to perform a non-linear transformation on the output of the convolutional layer or the pooling layer; the activation layer includes a first activation layer and a second activation layer; the first activation layer receives the output of the convolutional layer and outputs it to the pooling layer; the second activation layer receives the output of the pooling layer and outputs it to the fully connected layer; The fully connected layer maps the second spatio-temporal radiation feature after convolution, pooling, and activation operations to the output space. The fully connected layer contains multiple neurons, performs a linear transformation on the second spatio-temporal radiation feature through a weight matrix and a bias term, and obtains a satellite feature through non-linear processing using an activation function. The output layer is connected to the fully connected layer to output the satellite feature.

3. The method for predicting the supercooled water content in clouds based on deep learning and spatio-temporal data driving according to claim 1, wherein The method for obtaining the radar feature includes: Inputting radar data into a second feature extraction model for feature extraction to obtain a radar feature; the dimensions of the radar data include the number of samples, the number of features, the number of longitude grid points, the number of latitude grid points, and the number of time steps; the second feature extraction model includes an input layer, a convolutional layer, a pooling layer, an activation layer, a fully connected layer, and an output layer. The input layer is used to receive radar data. The convolutional layer uses a 4D convolutional neural network to perform a convolution operation on the input data to capture local features in space and time to obtain a first spatio-temporal radar feature; the convolutional kernel of the 4D convolutional neural network uses 1×3×3×3 or 1×5×5×3, and the stride is 2×2×2×2. The pooling layer downsamples the first spatio-temporal radar feature to simplify the dimensions of the first spatio-temporal radar feature to obtain a second spatio-temporal radar feature; the pooling layer performs pooling operations in the spatial and temporal dimensions. The activation layer uses an activation function to perform a non-linear transformation on the output of the convolutional layer or the pooling layer; the activation layer includes a first activation layer and a second activation layer; the first activation layer receives the output of the convolutional layer and outputs it to the pooling layer; the second activation layer receives the output of the pooling layer and outputs it to the fully connected layer. The fully connected layer maps the second spatio-temporal radar feature after convolution, pooling, and activation operations to the output space. The fully connected layer contains multiple neurons, performs a linear transformation on the second spatio-temporal radar feature through a weight matrix and a bias term, and obtains a satellite feature through non-linear processing using an activation function. The output layer is connected to the fully connected layer to output the radar feature.

4. The method for predicting supercooled water content in cloud based on deep learning and spatio-temporal data driving according to claim 1, wherein The method for obtaining the conventional meteorological feature includes: Inputting conventional meteorological data into a third feature extraction model for feature extraction to obtain a conventional meteorological feature; the conventional meteorological data includes temperature, pressure, wind direction, wind speed, precipitation, and relative humidity; the dimensions of the conventional meteorological data include the number of features, the number of stations, and the number of time steps; the third feature extraction model includes an input layer, a convolutional layer, a pooling layer, an activation layer, a fully connected layer, and an output layer. The input layer is used to receive conventional meteorological data. The convolutional layer uses a 3D convolutional neural network to perform a convolution operation on the input data to capture local features in space and time to obtain a first spatio-temporal conventional meteorological feature; the convolutional kernel of the 3D convolutional neural network uses 2×3×3 or 2×5×3, and the stride is 2×2×2. The pooling layer downsamples the first spatio-temporal conventional meteorological data feature to simplify the dimensions of the first spatio-temporal conventional meteorological feature to obtain a second spatio-temporal conventional meteorological feature; the pooling layer performs pooling operations in the spatial and temporal dimensions. The activation layer uses an activation function to perform a non-linear transformation on the output of the convolutional layer or the pooling layer; the activation layer includes a first activation layer and a second activation layer; the first activation layer receives the output of the convolutional layer and outputs it to the pooling layer; the second activation layer receives the output of the pooling layer and outputs it to the fully connected layer; The fully connected layer maps the second spatio-temporal conventional meteorological features after convolution, pooling, and activation operations to the output space. The fully connected layer contains multiple neurons, performs a linear transformation on the second spatio-temporal conventional meteorological features through a weight matrix and a bias term, and obtains the conventional meteorological features through non-linear processing by an activation function; The output layer is connected to the output of the fully connected layer to output radar features.

5. The method for predicting the supercooled water content in cloud based on deep learning and spatio-temporal data driving according to claim 1, wherein The method for optimizing the deep learning network according to the optimization objective function includes the following steps: Use the bitter fish optimization algorithm to determine the optimal network model parameters, and define as the th network model parameter, corresponding to the bitter fish spawning point, is the input feature dimension within the past 3 hours, and a bitter fish population is composed of multiple bitter fish individuals , is the population size. At the same time, the population is initialized with the standard network model parameters to obtain , where the boundary is , is the random perturbation number; Search for suitable oysters to determine the bitter fish spawning point, update the bitter fish position (update the dynamic network model parameters). When an oyster is successfully captured, the bitter fish position update expression is: Among them is the updated position of the th fish in the th iteration, is the current position of the th fish in the th iteration, is the dynamic inertia weight, is the number of steps for the bitter fish to move to the escape oyster at the th iteration, , are random numbers, is the best oyster, that is, the optimal spawning point that attracts the bitter fish, is the most valuable oyster randomly selected from the population, , are the maximum inertia weight and the minimum inertia weight, is the population variance at the th iteration, is the initial population variance, is the maximum number of iterations, is the random function; When the oyster escapes, re-exploration is carried out, and the bitter fish position update expression is: wherein is the escape reset probability, taking ; According to the position of the successfully captured oyster, the female fish lays eggs at this position to produce new individuals. The bitter fish position update expression for the new individual is: Among them is the position of the newly generated bitter fish at the iteration, is the distribution radius of the newly generated bitter fish inside the oyster shell, is the Gaussian noise with a mean of 0 and a variance of ; Calculate the fitness of the population and determine the best oyster and the most valuable oyster in the population , the expression is: ； ； Among them is the population fitness is the dynamic weight of fitness is the th iteration, the th position of the dynamic network model parameters is the th position of the standard network model parameters is the th position corresponding to the maximum value of the dynamic network model parameters The th position corresponding to the minimum value of the dynamic network model parameters is the weight of the coordination entropy is the th coordination entropy of the dynamic recognition threshold is the time-varying penalty weight is the time-varying penalty decay rate is the th iteration, the th correlation between the position of the dynamic network model parameters and the target position ; Adopt the strategy of oyster hunting and killing new bitter fish to eliminate population individuals. The elimination probability of new bitter fish is: where is the elimination probability of newly born bitter fish, is the population selection pressure, is the fitness of the newly born bitter fish population at the -th iteration, is the fitness of the new bitter fish population containing newly born bitter fish and original bitter fish at the -th iteration, is the diversity penalty weight, is the population diversity at the -th iteration, population historical average position; Repeat the above operations until the optimization objective function is minimized or the maximum number of iterations is reached to stop the iteration, output the optimal network model parameters, and obtain the optimized deep learning network.

6. The method for predicting supercooled water content in clouds based on deep learning and spatio-temporal data driving according to claim 1, wherein: The satellite features include satellite multi-channel radiation data; the radar features include composite reflectivity and VIL; the conventional meteorological features include temperature, relative humidity, wind direction, wind speed, precipitation.

Images