Cloud cluster trajectory prediction method based on lightweight convolutional neural network

Abstract

The invention discloses a cloud cluster trajectory prediction method based on a lightweight convolutional neural network, and belongs to the field of cloud picture trajectory prediction. According to the method, a one-dimensional convolutional neural network and a two-dimensional convolutional neural network are combined to extract key features in data, time sequence data of a meteorological station are processed through the one-dimensional convolutional neural network, space-time coupling features can be captured through the two-dimensional convolutional neural network, the cloud form and texture features of a satellite or radar image can be analyzed, and finally, the cloud form and texture features of the satellite or radar image can be obtained. After spatio-temporal features are integrated through GRU, 24-hour data are input, the cloud cluster moving path in the next 12-24 hours can be predicted, and the processing capacity of the model for time sequence data is further enhanced. According to the method, the calculation complexity is greatly reduced, and the memory occupation is remarkably reduced; in addition, the method improves the prediction accuracy, and has a wide application prospect in the fields of meteorological analysis, environment monitoring and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of cloud track prediction, in particular to a cloud cluster track prediction method based on a light-weight convolutional neural network. BACKGROUND

[0002] The cloud-based photovoltaic power prediction technology can perform high-precision photovoltaic power prediction, reduce power generation fluctuations caused by cloud shading, improve the power grid's ability to accommodate photovoltaic power, and has the advantages of multi-step prediction and effective response to complex weather such as cloud shading. It has become the main technology for photovoltaic power prediction. With the rapid development of neural network technology, it can realize the combination of multi-source data fusion to improve the prediction accuracy, and the prediction of cloud cluster trajectory is the key to cloud-based photovoltaic power prediction technology.

[0003] At present, most cloud track prediction uses a method based on a three-dimensional convolutional neural network (3D-CNN), but 3D-CNN has a large amount of calculation when extracting data features, resulting in delays in the prediction process. The model predicts future trajectories based on historical frames, and the lack of real-time performance will cause errors to rapidly magnify over time, making it difficult to meet the needs of photovoltaic power station edge deployment. SUMMARY

[0004] The present application aims to provide a cloud cluster track prediction method based on a light-weight convolutional neural network, which processes time series data from a meteorological station using a one-dimensional convolutional neural network, captures spatial features using a two-dimensional convolutional neural network, and predicts the moving path of the cloud cluster after integrating the spatio-temporal features using a GRU.

[0005] The cloud cluster track prediction method based on a light-weight convolutional neural network of the present application includes the following steps:

[0006] Obtain satellite cloud image data and meteorological data using a meteorological feature sensor, preprocess the meteorological data and satellite cloud image data to obtain time input samples and spatial input samples, and divide the data set into a training set and a test set according to a proportion.

[0007] Construct a cloud cluster track prediction model and train the cloud cluster track prediction model using the training set; the cloud cluster track prediction model includes a one-dimensional convolutional neural network, a two-dimensional convolutional neural network, and a gated recurrent unit.

[0008] Use the trained cloud cluster track prediction model to perform track prediction.

[0009] Further, the step of training the cloud cluster track prediction model using the training set includes:

[0010] input the spatial input sample xi The spatial input sample is first subjected to convolution operation by a first convolution layer, then subjected to pooling operation by a first pooling layer, then subjected to convolution operation by a second convolution layer, and finally subjected to pooling operation by a second pooling layer in the two-dimensional convolution neural network. The feature data after the pooling is connected into a feature set by a full connection layer. All neuron nodes are connected with the output of the previous layer. The feature array is finally obtained by the full connection layer, that is, the feature vector y of the final output of the two-dimensional convolution neural network is obtained. The feature vector y is used to represent the abstract spatial features of the cloud image.

[0011] The calculation formula when the spatial input sample is subjected to convolution operation is:

[0012]

[0013] In the formula, j represents the jth output channel, i represents the ith input channel, Cin represents the number of input channels, ω ij represents the weight of the jth convolution kernel, * represents two-dimensional convolution operation, b j represents the jth output bias term, f(·) is a ReLU activation function, h j represents the feature map obtained after convolution h j .

[0014] The calculation formula when the pooling operation is performed is:

[0015]

[0016] In the formula, W is the width of the pooling region, s is the pooling step, m is the moving variable, k is a one-dimensional index of the spatial position, and the spatial feature map y j is obtained through the pooling layer.

[0017] The spatial feature map is flattened into a one-dimensional feature vector x, which is then input into the full connection layer to obtain the feature vector y. The formula is:

[0018] y = σ(W A · x + b),

[0019] In the formula, W A is the weight matrix, x is the input feature vector, b is the bias vector, and σ is the activation function.

[0020] Further, the step of training the cloud cluster trajectory prediction model using the training set further comprises:

[0021] The sensor obtained data is arranged in time sequence into sequence data, the sorted data is input according to a three-dimensional tensor, recorded as a time input sample, the shape is recorded as (N, C, L), wherein N represents the data amount, C represents the feature quantity, and L represents the number of data points, and the input is input into a one-dimensional convolutional neural network;

[0022] In the one-dimensional convolutional neural network, the time input sample is first input through a multi-channel, and then convolution operation is performed through a convolution layer to extract local time features Y u,v,c , and the formula is as follows:

[0023]

[0024] In the formula, K1 is the size of the convolution kernel, u is the u-th sample, v is the v-th time step, tau is a moving variable, w 1,τ,c represents the weight of the first layer of tau convolution kernels in the c-th output channel, x u,v-τ+1,c is the time input sample, and b 1,c is a bias term;

[0025] The local time features Y u,v,c are reduced in dimension and main features are extracted by using a pooling layer, maximum pooling is adopted, main time features Z u,v,c are extracted, and the formula is as follows:

[0026]

[0027] In the formula, k p is the size of the pooling window, and s p is the pooling stride;

[0028] Finally, the one-dimensional convolutional neural network outputs a three-dimensional tensor feature map Z u,v,c , the three-dimensional tensor feature map is flattened into a one-dimensional vector through a fully connected layer and connected to a neural output layer to obtain features X; then linear transformation and mapping are performed through a weight matrix W and a bias b to output the results of each meteorological factor in the data set, that is, the time sequence feature represents the meteorological factor, and the formula is as follows:

[0029] Y=XW+b.

[0030] Further, the step of training the cloud cluster trajectory prediction model by using the training set further includes:

[0031] The feature vector y extracted by the two-dimensional convolutional neural network and the time sequence feature vector Y extracted by the one-dimensional convolutional neural network are input into a gated recurrent unit, the update gate, the reset gate, the candidate hidden state, and the hidden state of the current time step are calculated, and an output vector h t of the prediction result is obtained, and then a fully connected layer is used for prediction and output of the prediction result y t , that is, the coordinates of the cloud cluster.

[0032] Further, the step of pre-processing the meteorological data and the satellite cloud image data respectively comprises:

[0033] normalizing the temperature, humidity and pressure data obtained by the meteorological feature sensors respectively;

[0034] normalizing the obtained wind speed data x1, and the formula is:

[0035]

[0036] wherein μ is the mean value, and σ0 is the standard deviation;

[0037] taking the pre-processed temperature, humidity, pressure and wind speed data as time input samples.

[0038] Further, the step of pre-processing the meteorological data and the satellite cloud image data respectively comprises:

[0039] converting the row and column numbers of the stationary orbit nominal projection in the satellite cloud image data into corresponding longitude and latitude coordinates, first, calculating variables ξ and

[0040]

[0041] wherein c0 and l are the nominal column and row numbers, COFF and LOFF are the column offset and row offset respectively, and CAFC and LAFC are the column and row scale factors respectively;

[0042] then, calculating an intermediate variable S according to the following formula: d n xy

[0043]

[0044] wherein h is the distance between the center of the earth and the center of the satellite, ea is the semi-major axis of the earth, and eb is the semi-minor axis of the earth;

[0045] finally, calculating the longitude coordinate lon and the latitude coordinate lat, and the formulas are respectively:

[0046]

[0047] wherein λ D is the sub-satellite point longitude corresponding to different satellites;

[0048] taking the cloud image processed into a standard size as a space input sample.

[0049] ​​​Beneficial effects: Compared with the prior art, the present application has the following advantages:

[0050] 1、The present application adopts the technology of combining one-dimensional convolutional neural network and two-dimensional convolutional neural network to extract key features in data, which not only greatly reduces the computational complexity, but also significantly reduces the memory occupation, which is of great significance for those resource-constrained edge devices.

[0051] 2、In addition, the present application further enhances the processing ability of the cloud cluster trajectory prediction model for time series data by introducing GRU; the introduction of GRU makes the model perform well in capturing the dynamic change characteristics of cloud clusters, thereby significantly improving the accuracy of prediction in cloud cluster trajectory prediction. This improvement not only improves the accuracy of prediction, but also has wide application prospects in the fields of meteorological analysis, environmental monitoring, etc.

[0052] 3、In addition, the present application also considers the real-time performance of data processing, ensuring that in the case of large data volume and high processing demand, accurate processing results can still be provided quickly. Through this innovative data processing method, the present application provides a new solution for data-intensive applications such as real-time traffic monitoring and intelligent medical diagnosis. These application fields have very high requirements for the efficiency and accuracy of data processing, and the present application can exactly meet these needs, thus showing great potential and value in practical applications.

[0053] 4、The application range of the present application is more extensive, and even in more complex situations, it can maintain high accuracy. This innovative solution not only optimizes the prediction process, but also provides strong support for the efficient operation of photovoltaic power stations. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 is the flow chart of the present application;

[0055] Figure 2 is the structure diagram of one-dimensional convolutional neural network;

[0056] Figure 3 is the structure diagram of two-dimensional convolutional neural network;

[0057] Figure 4 is the structure diagram of GRU. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples.

[0059] The overall flow chart of the cloud cluster trajectory prediction method based on lightweight convolutional neural network of the present application is as follows: Figure 1As shown, the method specifically includes the following steps:

[0060] Step 1, obtaining satellite cloud image data, obtaining meteorological data using a meteorological feature sensor, pre-processing the meteorological data and the satellite cloud image data respectively to obtain time input samples and space input samples, combining the time input samples and the space input samples into a data set, and dividing the data set into a training set and a test set according to a proportion.

[0061] Obtain relevant meteorological data using the sensor of the weather station, including but not limited to data such as temperature, humidity, air pressure, wind speed, short-term trend of cloud movement, periodic fluctuation, etc. Obtain satellite cloud image data after satellite spatial distortion improvement through Japanese Himawari-8. In order to meet the requirements of standard data for convolutional neural network input, the data needs to be pre-processed.

[0062] Further, the step of pre-processing the meteorological data and the satellite cloud image data respectively includes:

[0063] The pre-processing of the meteorological data includes:

[0064] The temperature, humidity and air pressure data obtained by the meteorological feature sensor are normalized respectively, and the formula is:

[0065]

[0066] In the formula, x θ represents the original meteorological data, x min is the minimum value of the feature dimension, and x max is the maximum value of the feature dimension.

[0067] After normalization, all features are in a similar numerical range, the gradient update direction is more consistent, and the training speed is significantly improved.

[0068] The obtained wind speed data x1 is normalized, and the formula is:

[0069]

[0070] In the formula, since the wind speed has no obvious boundary, μ is used for mean value and σ0 is used for standard deviation after normalization.

[0071] After processing the wind speed, its amplitude can be weakened to avoid the model being dominated by outliers, and the data is closer to the normal distribution.

[0072] The pre-processed temperature, humidity, air pressure and wind speed data are used as time input samples.

[0073] In order to obtain the cloud image meeting the input standard of two-dimensional convolutional neural network, the original cloud image needs to be geometrically corrected and geographically calibrated. Because Himawari-8 is a geostationary satellite, although it is fixed in a position relative to the ground, because the earth is a sphere, the shooting angle of the satellite will change with the change of the observation site. Therefore, the picture taken by the satellite will have the phenomenon of spatial distortion. Accurately corresponding the satellite shot cloud image data to the actual geographical coordinates of the earth's surface can effectively solve the distortion problem caused by the angle of view, light refraction and earth rotation. According to the explanation of the user manual of Himawari-8 satellite, the row and column numbers of the nominal projection of the geostationary orbit need to be converted into the corresponding latitude and longitude coordinates.

[0074] Further, the step of pre-processing the meteorological data and the satellite cloud image data respectively further comprises:

[0075] The row and column numbers of the nominal projection of the geostationary orbit in the satellite cloud image data are converted into the corresponding latitude and longitude coordinates. First, the variables ξ and η are calculated according to the following formula:

[0076]

[0077] In the formula, c0 and l are the nominal column number and row number, COFF and LOFF are the column offset and row offset, CAFC and LAFC are the column scale factor and row scale factor respectively, and different spatial resolutions correspond to different values;

[0078] Then, the intermediate variable S is calculated according to the following formula: d n xy

[0079]

[0080] Wherein, h is the distance between the center of the earth and the center of the satellite, ea is the semi-major axis of the earth, and eb is the semi-minor axis of the earth;

[0081] Finally, the longitude coordinate lon and the latitude coordinate lat are calculated, and the formulas are respectively:

[0082]

[0083] Wherein, λ D is the sub-satellite point longitude corresponding to different satellites;

[0084] After the cloud image is processed into a standard size, it is used as a spatial input sample.

[0085] ​​​The original satellite cloud image is associated with the longitude and latitude coordinates by the above method for calibration and geometric correction, and the selected cloud image is cropped according to the longitude and latitude. At the same time, the visible light satellite cloud image is mainly made by using the principle that different thickness of cloud clusters have different degrees of sunlight reflection. Due to the influence of the solar elevation angle, the time of day will affect the identification of the cloud cluster pixel points. The target area needs to be image enhanced to better distinguish and better identify the cloud cluster pixel points to eliminate the diurnal variation of the cloud cluster. Finally, the cloud image is processed into a standard size of 224x224x3 and sent into the model.

[0086] Step 2, constructing a cloud cluster trajectory prediction model, and training the cloud cluster trajectory prediction model using the training set; wherein the cloud cluster trajectory prediction model includes a one-dimensional convolutional neural network (1D-CNN), a two-dimensional convolutional neural network (2D-CNN), and a gated recurrent unit (GRU).

[0087] 1D-CNN convolution is selected as a means of extracting one-dimensional time data, and the basic network structure is as shown in Figure 2 It uses a one-dimensional convolution kernel to slide on the time axis, which can capture local patterns in the sequence. 2D-CNN extracts and processes data converted from two-dimensional images, and the framework is as shown in Figure 3 Through a two-dimensional convolution kernel to extract spatial local features (such as edges, textures), combined with a dimension reduction pooling layer, to capture spatial or frequency domain features. The two-dimensional convolutional neural network is composed of an input layer, a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer, which extracts local spatial features in the input data, and classifies the features to complete fault diagnosis,

[0088] In combination with Figure 3 Further, the step of training the cloud cluster trajectory prediction model using the training set includes:

[0089] The spatial input sample x i is input into the two-dimensional convolutional neural network. In the two-dimensional convolutional neural network, the spatial input sample is first convolved by the first convolutional layer, then pooled by the first pooling layer, then convolved by the second convolutional layer, and finally pooled by the second pooling layer. Finally, the pooled feature data is connected into a feature set by the fully connected layer, all neuron nodes are connected to the output of the previous layer, and finally the feature array is obtained by the fully connected layer, that is, the feature vector y of the final output of the two-dimensional convolutional neural network is obtained, and the feature vector y is used to represent the abstract spatial features of the cloud image;

[0090] The calculation formula when the spatial input sample is convolved is:

[0091]

[0092] where j represents the jth output channel, i represents the ith input channel, Cin represents the number of input channels, ω ij represents the weight of the jth convolution kernel, * represents a two-dimensional convolution operation, b j represents the jth output bias term, f(·) is a ReLU activation function, h j represents the feature map obtained after convolution j ;

[0093] When the pooling operation is performed, the calculation formula is:

[0094]

[0095] where W is the width of the pooling region, s is the pooling step, m is the moving variable, k is a one-dimensional index of the spatial position, and the spatial feature map y j ;

[0096] The input sample is repeatedly subjected to two convolution layers and pooling layers, abstract features are gradually extracted and the dimension is reduced, a simplified feature map is obtained, the spatial feature map is flattened into a one-dimensional feature vector x, in order to let the computer recognize the image, the one-dimensional feature vector x is input to the fully connected layer, the pooled feature data is connected into a feature set by using the fully connected layer, all neuron nodes of the fully connected layer are connected with the output of the previous layer, finally the feature array is obtained through the fully connected layer and compared with the training set to obtain the feature vector y, and the formula is:

[0097] y = σ(W A · x + b),

[0098] where W A is a weight matrix, x is an input feature vector, b is a bias vector, and σ is an activation function.

[0099] The feature vector y output by the two-dimensional convolutional neural network is a high-dimensional feature vector, which is a parameter, representing the abstract spatial features of the cloud image, and the cloud cluster movement trend can be obtained.

[0100] As shown in Figure 2 , further, the step of training the cloud cluster trajectory prediction model using the training set further comprises:

[0101] The data obtained by the sensor is arranged in sequence according to time sequence, and the sorted data is input according to three-dimensional tensor, denoted as time input sample, the shape is denoted as (N, C, L), where N represents the data amount, C represents the feature quantity, and L represents the number of data points, and the time input sample is input into the one-dimensional convolutional neural network;

[0102] In the one-dimensional convolutional neural network, the multi-channel input time input sample is first input, then the convolution layer is used for convolution operation, and the local time feature Y u,v,c is extracted, and the formula is as follows:

[0103]

[0104] where K1 is the size of the convolution kernel, u is the u-th sample, v is the v-th time step, τ is the shift variable, w 1,τ,c represents the weight of the first layer of τ convolution kernels in the c-th output channel, x u,v-τ+1,c is the time input sample, and b 1,c is the bias term.

[0105] The local time feature Y u,v,c is reduced in dimension and the main feature is extracted using a pooling layer, maximum pooling is used, and the main time feature Z u,v,c is extracted, and the formula is:

[0106]

[0107] where k p is the size of the pooling window, s p is the pooling stride.

[0108] Finally, the one-dimensional convolutional neural network outputs a three-dimensional tensor feature map Z u,v,c , which is flattened into a one-dimensional vector by a fully connected layer and connected to the neural output layer to obtain the feature X; then the linear transformation mapping is performed through the weight matrix W and the bias b, and the results of each meteorological factor in the data set, i.e., the time series feature representing the meteorological factor, such as the pressure change rule, are output. If a meteorological factor data, i.e., a parameter, is input, the feature vector Y is represented, and the formula is as follows:

[0109] Y = XW + b.

[0110] Further, the step of training the cloud cluster trajectory prediction model using the training set further includes:

[0111] The feature vector y extracted by the two-dimensional convolutional neural network and the time series feature vector Y extracted by the one-dimensional convolutional neural network are input into the gated recurrent unit, the update gate, the reset gate, the candidate hidden state, and the hidden state of the current time step are calculated, and the output vector h t of the prediction result is obtained, and then the prediction result y t , i.e., the coordinates of the cloud cluster, is output through the fully connected layer.

[0112] For how to fuse and predict multiple source data, the gated recurrent unit (GRU) is used in this example, which is simply referred to as GRU. It is a neuron used in recurrent neural networks. The main components of GRU include the update gate and the reset gate, which together determine how much past information the neural network should remember and how much new information it should receive.

[0113] Combine Figure 4 In this example, the input data of the current time step and the hidden state of the previous time step are first input into the GRU unit, and then the update gate, reset gate, candidate hidden state and hidden state of the current time step are calculated by the following formula. Finally, the update gate is used to decide when to update the old hidden state, which represents how much past information is remembered.

[0114] The calculation formula of the update gate is:

[0115] z t =σ(W z ×[h t-1 ,x t ]+b z ),

[0116] In the formula, σ represents the sigmoid activation function, W z and b z are the weights and biases that need to be learned during training.

[0117] The remake gate is used to determine how much of the past hidden state will be forgotten. The remake gate calculation formula is:

[0118] r t =σ(W r ×[h t-1 ,x t ]+b r ),

[0119] This is similar to the calculation of the update gate, W z and b z These are the weights and biases that need to be learned during training, but the weight and bias parameters are different.

[0120] The candidate hidden state is generated based on the current input and the information of the past hidden state, and the calculation formula is:

[0121]

[0122] Where ⊙ represents element-wise multiplication, is the candidate hidden state, W h is the weight matrix input to the candidate hidden state, b h is the bias term of the candidate hidden state.

[0123] The current hidden state is obtained by combining the past hidden state and the candidate hidden state. The update gate controls the information that needs to be fused. The calculation formula is:

[0124]

[0125] Where ht The hidden state for the current time.

[0126] Finally, the output value is calculated, and the algorithm is as follows:

[0127] y t = W hy h t + b y ,

[0128] In the formula, y t is the output value of the current time step, W ht is the weight matrix of the hidden state to the output, and b y is the bias term of the output.

[0129] Step 3, using the cloud cluster trajectory prediction model after training to perform trajectory prediction.

[0130] The time series data of the weather station (such as temperature, humidity, air pressure, short-term trend of cloud cluster movement, periodic fluctuation, etc.) is processed by 1D-CNN; 2D-CNN can capture spatial-temporal coupling features, analyze the cloud layer morphology and texture features of satellite / radar images, and finally, through GRU integration of space-time features, 24-hour data can be input to predict the cloud cluster movement path in the next 12-24 hours.

[0131] The cloud cluster trajectory prediction method based on the lightweight convolutional neural network provided by the application simplifies the entire data processing process through an efficient data processing method. Specifically, the scheme adopts the technology of combining one-dimensional convolutional neural network and two-dimensional convolutional neural network to extract key features in the data. This method not only greatly reduces the computational complexity, but also significantly reduces the memory occupation, which is of great significance for those resource-constrained edge devices.

[0132] In addition, by introducing GRU, the application further enhances the processing capability of the model for time series data. The introduction of GRU makes the model perform well in capturing the dynamic change characteristics of cloud clusters, thereby significantly improving the prediction accuracy in cloud cluster trajectory prediction. This improvement not only improves the prediction accuracy, but also has wide application prospects in the fields of meteorological analysis and environmental monitoring. The application range of the application is more extensive, and even in more complex situations, it can maintain high accuracy. This innovative scheme not only optimizes the prediction process, but also provides strong support for the efficient operation of photovoltaic power stations.

[0133] In addition, the present application also considers the real-time performance of data processing, ensuring that in the case of large data volume and high processing demand, it can still respond quickly and provide accurate processing results. Through this innovative data processing method, the present application provides a new solution for data-intensive applications such as real-time traffic monitoring and intelligent medical diagnosis. These application fields have very high requirements for the efficiency and accuracy of data processing, and the present application can exactly meet these needs, thus showing great potential and value in practical applications.

[0134] In evaluating the prediction results, evaluation indicators usually include prediction accuracy, error rate, etc., to measure the accuracy of the model prediction results. In order to comprehensively evaluate the model performance, other factors such as computing efficiency and robustness should also be considered. Through comprehensive analysis of these indicators, the performance of the model in actual application can be judged, and the model can be optimized and improved accordingly. For result accuracy, the RMSE (Root Mean Square Error) is used to evaluate the model test results, and the formula is as follows:

[0135]

[0136] In the formula, is the predicted value of the i-th sample, is the true value of the i-th sample, and N is the total number of samples. The prediction is obtained by RMSE, which can intuitively reflect the deviation between the model prediction value and the true value. RMSE is sensitive to outliers. If there are individual extreme values in the data set, it may have a greater impact on the calculation result of RMSE. When the RMSE value is small, it means that the prediction accuracy of the model is high, and the prediction result is close to the true situation.

[0137] For the stability of the results, the standard deviation (Standard Deviation) is used to measure, and its core formula is as follows:

[0138]

[0139] In the formula, σ γ is the standard deviation, n is the number of data points, x i is the i-th data point, and μ is the average value of the data, and the calculation formula is:

[0140] By calculating the standard deviation between the predicted trajectory and the actual trajectory, the stability of the model prediction result can be evaluated. The smaller the standard deviation, the more stable the prediction result. By analyzing the standard deviation of the prediction result, we can find out where the model performs poorly and optimize the model accordingly. For computational efficiency evaluation, the time required for the model to complete a trajectory prediction can be recorded. The shorter the time, the better the real-time performance. At the same time, pay attention to the occupation of computer memory, CPU and other resources during the model running process.

Claims

1. A cloud trajectory prediction method based on a lightweight convolutional neural network, characterized by: The following steps are involved: Obtain satellite cloud image data, use meteorological characteristic sensors to obtain meteorological data, preprocess the meteorological data and satellite cloud image data separately to obtain time input samples and space input samples, form a data set with the time input samples and space input samples, and divide the data set into a training set and a test set in proportion; Construct a cloud trajectory prediction model and train it using the training set; the cloud trajectory prediction model includes a one-dimensional convolutional neural network, a two-dimensional convolutional neural network, and a gated recurrent unit. The trained cloud trajectory prediction model is used for trajectory prediction.

2. The cloud trajectory prediction method based on lightweight convolutional neural network according to claim 1 is characterized in that: The steps for training the cloud trajectory prediction model using the training set include: Input the sample x into the space i Input into the two-dimensional convolutional neural network. In the two-dimensional convolutional neural network, the spatial input samples are first convolved by the first convolutional layer, then pooled by the first pooling layer, and then convolved by the second convolutional layer. Finally, the pooled data are connected into a feature set through the fully connected layer. All neuron nodes are connected to the output of the previous layer. Finally, the feature array is obtained through the fully connected layer, that is, the feature vector y of the final output of the two-dimensional convolutional neural network is obtained. The feature vector y is used to represent the abstract spatial features of the cloud map. When the spatial input sample is convolved, the calculation formula is: Where j represents the jth output channel, i represents the ith input channel, Cin represents the number of input channels, ω ij represents the weight of the jth convolution kernel, * represents the two-dimensional convolution operation, b j represents the j-th output bias term, f(·) is the ReLU activation function, and h j Represents the feature map h obtained after convolution j ; The calculation formula when performing pooling operation is: Where W is the width of the pooling area, s is the pooling stride, m is the moving variable, and k is the one-dimensional index of the spatial position. The spatial feature map y is obtained through this pooling layer. j ; Flatten the spatial feature map into a one-dimensional feature vector x, and then input it into the fully connected layer to obtain the feature vector y. The formula is: y=σ(W A ·x+b), Where W A is the weight matrix, x is the input feature vector, b is the bias vector, and σ is the activation function.

3. The cloud trajectory prediction method based on lightweight convolutional neural network according to claim 2 is characterized in that: The steps of training the cloud trajectory prediction model using the training set also include: The data obtained by the sensor is arranged into sequence data in chronological order, and the sorted data is input as a three-dimensional tensor, recorded as a time input sample, with a shape of (N, C, L), where N represents the amount of data, C represents the number of features, and L represents the number of data points, and is input into a one-dimensional convolutional neural network; In a one-dimensional convolutional neural network, the time input sample is first input through multiple channels, and then the convolution operation is performed through the convolution layer to extract the local time feature Y u,v,c , the formula is as follows: Where K1 is the size of the convolution kernel, u is the u-th sample, v is the v-th time step, τ is the moving variable, and w 1,τ,c represents the weight of the first layer τ convolution kernel in the cth output channel, x u,v-τ+1,c is the time input sample, b 1,c is the bias term; Use the pooling layer to transform the local temporal features Y u,v,c Perform dimensionality reduction and extract the main features, use maximum pooling to extract the main time feature Z u,v,c , the formula is: Where k p is the size of the pooling window, s p is the pooling stride; Finally, the one-dimensional convolutional neural network outputs a three-dimensional tensor feature map Z u,v,c , the three-dimensional tensor feature map is flattened into a one-dimensional vector through the fully connected layer and connected to the neural output layer to obtain the feature X; then the linear transformation mapping is performed through the weight matrix W and the bias b, and the results of each meteorological factor in the data set are output, that is, the time series feature represents the meteorological factor. The formula is as follows: Y=XW+b.

4. The cloud trajectory prediction method based on lightweight convolutional neural network according to claim 3 is characterized in that: The steps of training the cloud trajectory prediction model using the training set also include: The feature vector y extracted by the two-dimensional convolutional neural network and the time series feature vector Y extracted by the one-dimensional convolutional neural network are input into the gated recurrent unit, and the update gate, reset gate, candidate hidden state and hidden state of the current time step are calculated to obtain the output vector h of the prediction result. t , and then predict through the fully connected layer and output the prediction result y t , which are the coordinates of the cloud.

5. The cloud trajectory prediction method based on a lightweight convolutional neural network according to any one of claims 1 to 4, characterized in that: The steps of preprocessing meteorological data and satellite cloud image data respectively include: Normalize the temperature, humidity and air pressure data obtained by the meteorological characteristic sensor respectively; The obtained wind speed data x1 is normalized, and the formula is: Where μ is the mean and σ0 is the standard deviation; The preprocessed temperature, humidity, air pressure and wind speed data are used as time input samples.

6. The cloud trajectory prediction method based on a lightweight convolutional neural network according to any one of claims 1 to 4, characterized in that: The steps of pre-processing meteorological data and satellite cloud image data separately also include: The row and column numbers of the geostationary orbit nominal projection in the satellite cloud image data are converted into the corresponding longitude and latitude coordinates. First, the variables ξ and ξ are calculated according to the following formula: Where c0 and l are the nominal column and row numbers, COFF and LOFF are the column offset and row offset, CAFC and LAFC are the column scale factor and row scale factor, respectively. Then, the intermediate variable S is calculated according to the following formula: d 、S n 、S1、S2、S3、S xy , the formula is: Where h is the distance between the center of the Earth and the center of mass of the satellite, ea is the semi-major axis of the Earth, and eb is the semi-minor axis of the Earth; Finally, calculate the longitude coordinate lon and latitude coordinate lat, the formulas are: Among them, λ D is the longitude of the sub-satellite point corresponding to different satellites; After processing the cloud image into a standard size, it is used as a spatial input sample.

Images