A method, device, electronic device and storage medium for water vapor map prediction

Abstract

This application belongs to the field of weather forecasting technology. Specifically, it relates to a water vapor map prediction method, device, electronic device, and storage medium. First, water vapor map data is collected and preprocessed; a generative adversarial network is constructed, trained, and verified to obtain the optimal generative adversarial network parameters; the water vapor map data is input into the generative adversarial network to obtain the prediction result of the water vapor map. This application uses the generative adversarial network to extract cloud movement information from the water vapor map sequence, thereby predicting the water vapor map. This application is for short-term prediction of cloud maps, and the prediction period is generally within 1 hour. As the prediction period extends, the prediction accuracy shows a declining trend, and the clarity of the predicted image gradually decreases. The embodiments of the present disclosure address the problems of low prediction accuracy and blurred predicted images in the water vapor map prediction method for a long prediction period, and propose a water vapor map prediction method with a time step of 1 hour and a prediction period of 6 hours.

Description

Technical Field

[0001] The present application belongs to the technical field of meteorological forecasting, and specifically relates to a water vapor map prediction method, device, electronic device and storage medium. Background Art

[0002] In the existing technical solutions, cloud image prediction mostly adopts feature matching methods. An important part of this method is to obtain cloud motion vector (CMV: Cloud Motion Vector), and then use the cloud motion vector to predict the cloud image at the next moment. Cloud motion vector can be obtained from images collected by the whole sky imager (WSI: Whole Sky Imager) or from satellite cloud images. Images collected by the whole sky imager are usually used for short-term forecasts in local areas, while satellite cloud images can be used to track global atmospheric motion and climate conditions in large areas. Feature matching methods usually first locate some significant features in the image, such as brightness temperature gradient and cloud edge. Since these features will not change significantly in a short time, these features can be used to obtain cloud motion vector. One typical method was proposed by Brad et al. (2002). This method combines the block matching algorithm (BMA: Block Matching Algorithm), the best candidate block search and the vector median regularization method to calculate the cloud motion vector. This method divides continuous images into blocks, pre-sets the number and range of candidate blocks, and selects the best matching block from the candidate blocks. Compared with the full search block algorithm that finds the best match between continuous frames in the entire domain, it has a significant improvement in computing speed. Gong Ke et al. (2000) used a sequence similarity detection algorithm to search for the best matching block, and used the average template and median filtering method to remove the "mosaic" effect. Based on the idea of ​​block matching algorithm, Jamaly et al. (2018) proposed a cross-correlation method (CCM: Cross-correlation Method) and used cross-spectral analysis (CSA: Cross-spectral Analysis) as the matching criterion for cloud motion vector estimation. In addition, this method also uses quality control measures to eliminate data with low variability and low correlation, so as to improve the accuracy of cloud motion vector estimation. However, algorithms based on the idea of ​​block matching often have large calculation errors. Because such methods assume that the local movement of clouds is uniform, in fact, due to cloud deformation, wind speed changes caused by terrain, changes in optical perspective, etc., it is difficult to characterize cloud motion with simple linear changes. The longer the forecast period, the more serious this limitation. Chow et al. (2015) proposed a variational optical flow technique (VOF) to calculate cloud motion vectors at pixel accuracy. However, their method was only used to calculate cloud motion vectors and was not extended to cloud image prediction. Summary of the Invention

[0003] In view of this, the present disclosure proposes a water vapor map prediction method, apparatus, electronic device, and computer-readable storage medium to solve the technical problems in the related art.

[0004] According to the first aspect of the present disclosure, a water vapor map prediction method is proposed, including:

[0005] Collect water vapor map data and preprocess the water vapor map data;

[0006] Construct a generative adversarial network, train and validate the generative adversarial network to obtain optimal generative adversarial network parameters;

[0007] Input the water vapor map data into the generative adversarial network to obtain a prediction result of the water vapor map.

[0008] Optionally, the preprocessing of the water vapor map data includes: data standardization, spatial division, and temporal division.

[0009] Optionally, the generative adversarial network consists of a generator and a discriminator:

[0010] The generator consists of an encoding module and a prediction module. The encoding module consists of 4 layers of gated recurrent unit-convolutional neural network with residual units, and the prediction module consists of 4 layers of stacked gated recurrent unit-convolutional neural network;

[0011] The discriminator is a three-dimensional convolutional neural network, including four layers of three-dimensional convolutional layers and two layers of fully connected layers. The convolutional kernel sizes of the first three layers of the four layers of three-dimensional convolutional layers are 3×3×3, and the convolutional kernel size of the last layer of the three-dimensional convolutional layer is 1×1×1. In the three-dimensional convolutional layer, k, n, and s are used to represent the convolutional kernel size, the number of feature maps, and the stride; the activation function of each layer of the three-dimensional convolutional layer is the Leaky ReLU function; in the two layers of fully connected layers, the feature dimension output by the first layer of the fully connected layer is 1024, and the feature dimension output by the second layer of the fully connected layer is 1. The input of the discriminator is the multi-hour water vapor map predicted by the generator, and the output of the discriminator obtains the probability that the water vapor map predicted by the generator belongs to the water vapor map collected by the satellite.

[0012] Optionally, the training and validation of the generative adversarial network to obtain optimal generative adversarial network parameters includes:

[0013] (1) Randomly divide the preprocessed water vapor map data to obtain a training set, a validation set, and a test set. Among them, 80% of the data is used as the training set, 10% of the data is used as the validation set, and 10% of the data is used as the test set;

[0014] (2) Input the water vapor map data of the prediction region and the reference region in the water vapor map data training set into the generator in the generative adversarial network. The generator outputs the predicted water vapor map of the prediction region.

[0015] (3) Keep the parameters of the generator in the generative adversarial network unchanged. Input the measured water vapor map at the prediction moment and the predicted water vapor map of the generator into the discriminator. The discriminator obtains the probability that the input water vapor map belongs to the measured water vapor map, calculates the loss function of the discriminator based on the predicted probability, and updates the parameters of the discriminator with the goal of minimizing the loss function.

[0016] (4) Keep the parameters of the discriminator in the generative adversarial network unchanged. Calculate the loss function of the generator based on the difference value between the predicted water vapor map of the generator and the measured water vapor map at the prediction moment, as well as the predicted probability of the discriminator for the predicted water vapor map of the generator. Update the parameters of the generator with the goal of minimizing the loss function value to obtain the generative adversarial network.

[0017] (5) Verify the generative adversarial network. Input the water vapor map data of the prediction region and the reference region in the water vapor map data validation set into the generative adversarial network, and the generator outputs the predicted water vapor map.

[0018] (6) Use the peak signal-to-noise ratio PSNR between the measured image f and the predicted image g and the structural similarity index SSIM between the measured image f and the predicted image g as evaluation metrics. If the evaluation metrics decline, the training is overfitting, and the training is terminated, using the model parameters at the previous moment as the optimal generative adversarial network parameters; if the training is not overfitting, go to step (7).

[0019] (7) Repeat the above steps (1)-(6) until overfitting occurs in the training or the preset maximum number of iteration steps is reached to obtain the optimal generative adversarial network parameters.

[0020] The generative adversarial network of the present disclosure uses Multi-GRU (Gated Recurrent Unit)-RCN (Recurrent Convolutional Network) as the generator. The input of the model adopts multi-scale input composed of large-scale input and small-scale input, and the output of the generator is the predicted sequence. A three-dimensional convolutional neural network is used as the discriminator in the network. The generator predicts the next sequence through the input sequence, and then uses the predicted sequence as a negative example and the measured sequence at the prediction moment as a positive example to be input into the discriminator for judgment by the discriminator.

[0021] According to the second aspect of the present disclosure, a water vapor map prediction device is proposed, including:

[0022] An acquisition module, configured to acquire water vapor map data and preprocess the water vapor map data.

[0023] A network construction module for constructing a generative adversarial network, training and validating the generative adversarial network to obtain optimal generative adversarial network parameters;

[0024] An application module for inputting water vapor map data into the generative adversarial network to obtain a prediction result of the water vapor map.

[0025] According to a third aspect of the present disclosure, there is provided an electronic device, comprising:

[0026] A memory for storing computer-executable instructions;

[0027] A processor configured to execute:

[0028] Collect water vapor map data and preprocess the water vapor map data;

[0029] Construct a generative adversarial network, train and validate the generative adversarial network to obtain optimal generative adversarial network parameters;

[0030] Input water vapor map data into the generative adversarial network to obtain a prediction result of the water vapor map.

[0031] According to a fourth aspect of the present disclosure, there is provided a computer-readable storage medium having a computer program stored thereon, the computer program being configured to cause the computer to execute:

[0032] Collect water vapor map data and preprocess the water vapor map data;

[0033] Construct a generative adversarial network, train and validate the generative adversarial network to obtain optimal generative adversarial network parameters;

[0034] Input water vapor map data into the generative adversarial network to obtain a prediction result of the water vapor map.

[0035] According to an embodiment of the present disclosure, a generative adversarial network is used to extract cloud movement information from a water vapor map sequence, thereby predicting the water vapor map. Existing research methods are mostly applied to the field of solar power generation, mainly for short-term prediction of cloud maps, and the prediction period is generally within 1 hour. As the prediction period extends, the prediction accuracy shows a declining trend, and the clarity of the prediction image gradually decreases. The embodiment of the present disclosure addresses the problems of low prediction accuracy and blurred prediction images in the water vapor map prediction method for a long prediction period, and proposes a water vapor map prediction method with a time step of 1 hour and a prediction period of 6 hours.

[0036] Additional aspects and advantages of the present disclosure will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. Description of the Drawings

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0038] Figure 1 It is a flowchart of a water vapor map prediction method shown in an embodiment of the present disclosure.

[0039] Figure 2 It is a schematic diagram of the structure of a generative adversarial network in the water vapor map prediction method shown in an embodiment of the present disclosure.

[0040] Figure 3 It is a schematic diagram of the structure of a generator in the water vapor map prediction method shown in an embodiment of the present disclosure.

[0041] Figure 4 It is a schematic diagram of the structure of a discriminator in the water vapor map prediction method shown in an embodiment of the present disclosure.

[0042] Figure 5 It is a flowchart of the training of a generative adversarial network shown in an embodiment of the present disclosure.

[0043] Figure 6 It is a block diagram of the structure of a water vapor map prediction device shown in an embodiment of the present disclosure. Detailed implementation manners

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts fall within the scope of protection of the present application.

[0045] Figure 1 It is a flowchart of a water vapor map prediction method shown in an embodiment of the present disclosure. The water vapor map prediction method shown in this embodiment can be applied to user equipment, such as a personal computer, a tablet computer, etc.

[0046] As Figure 1 shown, the water vapor map prediction method may include the following steps:

[0047] In step 1, collect water vapor map data and preprocess the water vapor map data;

[0048] In one embodiment, preprocessing the water vapor map data includes: data standardization, spatial partitioning, and temporal partitioning.

[0049] In one embodiment of the present disclosure, for data standardization, the value range of the pixel values of the original water vapor map data is 0 - 255. For the convenience of calculation, the data is standardized by dividing the actual data pixel value by 255, thereby converting the data range to 0 to 1.

[0050] Spatial division: Each water vapor map is cut into tiles of 64×64 pixels (water vapor map prediction area) and 128×128 pixels (reference area). The water vapor map prediction area and the reference area correspond one by one, and their central positions are the same. The area of the reference area can be multiple times that of the water vapor map prediction area, for example, four times in the embodiment of the present disclosure. That is, the size of the water vapor map prediction area is 64×64×1, and the size of the reference area is 128×128×1. In each continuous cloud map sequence, the positions of the cut tiles are fixed, so as to ensure the consistency of the spatial information of the tiles in the same sequence.

[0051] Temporal division: According to the time, the daily water vapor maps are divided into two groups. One group is from 0 to 11 o'clock, and the other group is from 12 to 23 o'clock. For each group of water vapor maps, the first 6 hours (i.e., 0 to 5 o'clock in the first group and 12 to 17 o'clock in the second group) are set as the input time, and the water vapor maps at these times are used as the input data of the generator; the last 6 hours (i.e., 6 to 11 o'clock in the first group and 18 to 23 o'clock in the second group) are set as the prediction time, and the water vapor maps at these times are used as the measured data.

[0052] In one embodiment of the present disclosure, the water vapor map is divided into a prediction area and a reference area, and the water vapor maps of the prediction area and the reference area are used as the prediction basis respectively. In the existing water vapor map prediction methods, only a single-scale input is adopted. The multi-scale input of the present disclosure introduces the surrounding spatial information and increases the extraction of different-scale features, adding multi-scale input to the model, and improving the prediction index, prediction image quality, and prediction data distribution of the water vapor map. Especially in the case of a longer prediction period, the improvement amplitude is more obvious.

[0053] In step 2, a generative adversarial network is constructed, trained and verified to obtain the optimal generative adversarial network parameters; the generator consists of an encoding module and a prediction module. The encoding module consists of 4 layers of gated recurrent unit-recurrent convolutional neural networks (abbreviated as GRU-RCN) containing residual units, and the prediction module consists of 4 layers of stacked gated recurrent unit-recurrent convolutional neural networks.

[0054] Advantages of the generative adversarial network: Add more details to the prediction image and improve the clarity of the prediction image.

[0055] Advantages of the generator: The multi-scale recurrent convolutional neural network is used as the generator of the generative adversarial network. In the existing solutions, the generator in the generative adversarial network structure uses a convolutional neural network or a recurrent neural network, so it can only extract spatial or temporal information. As a spatio-temporal sequence prediction problem, the water vapor map prediction problem requires information extraction in both the time and space dimensions. Therefore, using a recurrent convolutional neural network can improve the prediction accuracy.

[0056] The encoding module and the prediction module respectively introduce batch normalization processing to accelerate the training process of the generative adversarial network and avoid overfitting of the model.

[0057] In one embodiment, the structure of the generator is as Figure 3 shown. The input of the generator includes the water vapor map data of the prediction region and the water vapor map data of the reference region. The water vapor map prediction region and the reference region are in one-to-one correspondence, and their central positions are the same. The area of the reference region is four times that of the water vapor map prediction region. That is, the size of the water vapor map prediction region is 64×64×1, and the size of the reference region is 128×128×1.

[0058] For both the reference region input and the water vapor map prediction region input, the number of feature maps after passing through the encoding model is 32. When using the reference region input, a max pooling layer is used to reduce the size of the feature maps, improving the model's computational efficiency and the ability to capture invariant information of objects in the image. The initial state of the prediction model is copied from the final state of the encoding model. The output feature maps of the reference region input and the water vapor map prediction region input are concatenated together to form a hidden layer state with 64 feature channels, which is then input into the prediction module and finally into a 1×1 convolutional layer, and the final prediction result is obtained through the ReLU activation function.

[0059] In the existing solutions, the generator in the generative adversarial network structure uses a convolutional neural network or a recurrent neural network, so it can only extract spatial or temporal information. As spatio-temporal sequence prediction problems, the cloud map and water vapor map prediction problems require information extraction in both the time and space dimensions. Therefore, using a recurrent convolutional neural network can improve the prediction accuracy.

[0060] In one embodiment, the structure of the discriminator is as Figure 4Shown is a three-dimensional convolutional neural network, including four three-dimensional convolutional layers and two fully-connected layers. The convolutional kernel sizes of the first three of the four three-dimensional convolutional layers are 3×3×3, and the convolutional kernel size of the last three-dimensional convolutional layer is 1×1×1. The purpose of using 1×1×1 convolution is mainly to reduce the number of parameters while ensuring the feature extraction ability. In the three-dimensional convolutional layer, k, n, and s are used to represent the convolutional kernel size, the number of feature maps, and the stride; since the stride s is different in the time and space dimensions, it is represented by a three-dimensional array. The first dimension is the stride in the time dimension, and the second and third dimensions are the strides in the space dimension. For example Figure 3 in Figure 3 , k3n128s(2,1,1) in the first three-dimensional convolutional layer means that the convolutional kernel size of this layer is 3×3×3, the number of output features is 128, the stride in the time dimension is 2, and the stride in the space dimension is 1. The activation function of each three-dimensional convolutional layer is the Leaky ReLU function; in the two fully-connected layers, the feature dimension output by the first fully-connected layer is 1024, and the feature dimension output by the second fully-connected layer is 1. The input of the discriminator is the multi-hour water vapor map predicted by the generator. After passing through the two fully-connected layers, the output of the discriminator gives the probability that the water vapor map predicted by the generator belongs to the water vapor map collected by the satellite.

[0061] Advantages of the discriminator: Using a three-dimensional convolutional neural network as the discriminator of the generative adversarial network. In previous technical solutions, the discriminator of the generative adversarial network usually uses a two-dimensional convolutional neural network to discriminate each predicted image. In an embodiment of the present disclosure, a three-dimensional convolutional neural network is used to discriminate each predicted sequence, so that the network can better learn the time information in the sequence and improve the prediction accuracy.

[0062] In step 3, the generative adversarial network is trained and verified, including the generator and the discriminator. The water vapor map data in the training set is used to alternately train the generator and the discriminator in the generative adversarial network. After training, the parameters of the generator and the discriminator are obtained. The water vapor map data in the validation set is used to verify the prediction effect of the trained generative adversarial network. The above training steps and validation steps are iterated until the optimal generative adversarial network parameters are obtained.

[0063] In one embodiment, the process of obtaining the optimal generative adversarial network parameters is as Figure 5 shown, including:[[]]

[0064] (1) Randomly divide the preprocessed water vapor map data to obtain a training set, a validation set, and a test set. Among them, 80% of the data is used as the training set, 10% of the data is used as the validation set, and 10% of the data is used as the test set;

[0065] (2) Input the water vapor map data of the prediction area and the reference area in the water vapor map data training set into the generator in the generative adversarial network. The generator outputs the predicted water vapor map of the prediction area;

[0066] (3) Keep the parameters of the generator in the generative adversarial network unchanged. Input the measured water vapor map at the prediction moment and the predicted water vapor map of the generator into the discriminator. The discriminator obtains the probability that the input water vapor map belongs to the measured water vapor map, calculates the loss function of the discriminator according to the prediction probability, and updates the discriminator parameters with the goal of minimizing the loss function;

[0067] The formula of the loss function of the discriminator is as follows:

[0068]

[0069] where \(x=(x_1,\ldots,x_{ t )\) is the input data of the generator, and \(y=(y_1,\ldots,y_{ t )\) is the measured water vapor map at the prediction moment; \(p_{ data}\) is the distribution of the measured water vapor map, \(G\) represents the generator, and \(D\) represents the discriminator.

[0070] When minimizing the loss function, the backpropagation method is used. This is a gradient conduction method, which is used here to update the model parameters and is a common method in deep learning models.

[0071] (4) Keep the parameters of the discriminator in the generative adversarial network unchanged. Calculate the loss function of the generator according to the difference value between the predicted water vapor map of the generator and the measured water vapor map at the prediction moment, and the prediction probability of the discriminator for the predicted water vapor map of the generator. Update the generator parameters with the goal of minimizing the loss function value to obtain the generative adversarial network;

[0072] The formula of the loss function of the generator is as follows:

[0073]

[0074] where \(\alpha\) is a constant greater than zero. In an embodiment of the present disclosure, the value of \(\alpha\) is 1.

[0075] (5) Verify the generative adversarial network. Input the water vapor map data of the prediction area and the reference area in the water vapor map data validation set into the generative adversarial network, and the generator outputs the predicted water vapor map;

[0076] (6) Use the peak signal-to-noise ratio PSNR between the measured image \(f\) and the predicted image \(g\) and the structural similarity index SSIM between the measured image \(f\) and the predicted image \(g\) as evaluation metrics. If the evaluation metrics decline, the training is overfitting, and the training is terminated, and the model parameters at the previous moment are used as the optimal generative adversarial network parameters; if the training is not overfitting, go to step (7);

[0077] The peak signal-to-noise ratio PSNR between the measured image f and the predicted image g is defined as:

[0078] PSNR(f,g) = 10log 10 (I max 2 / MSE(f,g))

[0079] where I max = 255 is the maximum pixel intensity in the image. MSE(f,g) is the mean squared error between the measured image f and the predicted image g:

[0080]

[0081] N is the number of pixel points in the image.

[0082] The structural similarity index SSIM between the measured image f and the predicted image g is defined as:

[0083] SSIM(f,g) = l(f,g)c(f,g)s(f,g)

[0084] where l(f,g), c(f,g), and s(f,g) correspond to the luminance contrast, contrast contrast, and structural contrast between the measured image f and the predicted image g, respectively. The corresponding formulas are as follows:

[0085]

[0086]

[0087]

[0088] where μ f and μ g are the pixel averages of the measured image f and the predicted image g, respectively, σ f and σ g are the pixel value variances of the measured image f and the predicted image g, respectively, σ fg is the pixel value covariance of f and g, and c1, c2, and c3 are positive constants. In an embodiment of the present disclosure, c1 = 10 -4 , c2 = 9×10 -4 , c3 = 4.5×10 -4 .

[0089] Since the generator and the discriminator are in a mutually adversarial state during training, it may cause the loss function to exhibit an oscillatory state. Therefore, when determining whether overfitting occurs, it is not directly judged by the loss function, but according to the evaluation index of the predicted data generated by the generator. If the evaluation index decreases, the training is overfitted and the training is terminated prematurely.

[0090] (7) Repeat the above steps (1)-(6) until overfitting occurs during training or the preset maximum number of iteration steps is reached, to obtain the optimal parameters of the generative adversarial network.

[0091] According to the optimal parameters of the generative adversarial network, obtain the generative adversarial network, and input the water vapor map data in the test set into the generative adversarial network to achieve water vapor map prediction.

[0092] In terms of prediction accuracy, the water vapor map prediction method proposed in the embodiments of the present disclosure introduces peripheral spatial information by using multi-scale input, and has a greater improvement in prediction accuracy compared with other methods. Especially when the prediction period is relatively long, the improvement amplitude is more obvious. In terms of prediction clarity, the embodiments of the present disclosure use a generative adversarial network structure with a multi-scale input recurrent convolutional neural network as the generator and a three-dimensional convolutional neural network as the discriminator to add more details to the predicted image and improve the clarity of the predicted image.

[0093] Corresponding to the embodiments of the above water vapor map prediction method, the present disclosure also proposes embodiments of a water vapor map prediction device.

[0094] Figure 6 It is a structural block diagram of a water vapor map prediction device shown in an embodiment of the present disclosure. As Figure 6 shown, the water vapor map prediction device includes:

[0095] An acquisition module, configured to acquire water vapor map data and preprocess the water vapor map data;

[0096] A network construction module, configured to construct a generative adversarial network, train and verify the generative adversarial network, and obtain the optimal parameters of the generative adversarial network;

[0097] An application module, configured to input water vapor map data into the generative adversarial network to obtain the prediction result of the water vapor map.

[0098] The embodiments of the present disclosure also propose an electronic device, including:

[0099] A memory, configured to store computer-executable instructions;

[0100] A processor, where the processor is configured to execute:

[0101] Acquire water vapor map data and preprocess the water vapor map data;

[0102] Construct a generative adversarial network, train and verify the generative adversarial network, and obtain the optimal parameters of the generative adversarial network;

[0103] Input water vapor map data into the generative adversarial network to obtain the prediction result of the water vapor map.

[0104] It should be noted that in the embodiments of the present disclosure, the so-called processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The memory can be used to store the computer program and / or module. The processor realizes various functions of the automotive parts picture dataset production device by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one magnetic disk storage device, a flash device, or other volatile solid-state storage devices. If the modules / units of the device for constructing the stable operation domain of the wind power system are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-mentioned method embodiments of the present disclosure, it can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disc, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided by the embodiments of the present disclosure, the connection relationships between the modules indicate that there is a communication connection between them, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement this without creative effort.

[0105] The above is the preferred embodiment of the present disclosure. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A method for predicting a water vapor map, characterized in that, Including: Collect water vapor map data and preprocess the water vapor map data. The preprocessing of the water vapor map data includes: data standardization, spatial division, and time division. Among them, the spatial division includes: dividing each water vapor map into a prediction area and a reference area, and using the water vapor maps of the prediction area and the reference area as the prediction basis. The prediction area and the reference area correspond one by one, and the central positions of the prediction area and the reference area are the same. The time division includes: dividing the water vapor maps of each day into two groups according to the time of day, one group from 0 to 11 o'clock, and the other group from 12 to 23 o'clock. For each group of water vapor maps, the first 6 hours are set as the input time, and the water vapor map at the input time is used as the input data of the generator. Among them, the first 6 hours are from 0 to 5 o'clock in the first group and from 12 to 17 o'clock in the second group; the next 6 hours are set as the prediction time, and the water vapor map at the prediction time is used as the measured data. Among them, the next 6 hours are from 6 to 11 o'clock in the first group and from 18 to 23 o'clock in the second group. Construct a generative adversarial network, train and validate the generative adversarial network to obtain the optimal generative adversarial network parameters. The generative adversarial network consists of a generator and a discriminator; the generator consists of an encoding module and a prediction module. The encoding module consists of 4 layers of gated recurrent unit-convolutional neural networks containing residual units, and the prediction module consists of 4 layers of stacked gated recurrent unit-convolutional neural networks. Input the water vapor map data into the generative adversarial network to obtain the prediction result of the water vapor map. The training and validation of the generative adversarial network to obtain the optimal generative adversarial network parameters include: (1) Randomly divide the preprocessed water vapor map data into a training set, a validation set, and a test set. Among them, 80% of the data is used as the training set, 10% of the data is used as the validation set, and 10% of the data is used as the test set. (2) Input the water vapor map data of the prediction area and the reference area in the water vapor map data training set into the generator in the generative adversarial network, and the generator outputs the predicted water vapor map of the prediction area. (3) Keep the parameters of the generator in the generative adversarial network unchanged, input the measured water vapor map at the prediction time and the predicted water vapor map of the generator into the discriminator, the discriminator obtains the probability that the input water vapor map belongs to the measured water vapor map, and calculates the loss function of the discriminator according to the prediction probability. With the goal of minimizing the loss function, update the discriminator parameters. (4) Keep the parameters of the discriminator in the generative adversarial network unchanged, calculate the loss function of the generator according to the difference value between the predicted water vapor map of the generator and the measured water vapor map at the prediction time, and the prediction probability of the discriminator for the predicted water vapor map of the generator. With the goal of minimizing the loss function value, update the generator parameters to obtain the generative adversarial network. (5) Validate the generative adversarial network, input the water vapor map data of the prediction area and the reference area in the water vapor map data validation set into the generative adversarial network, and the generator outputs the predicted water vapor map. (6) The peak signal-to-noise ratio PSNR between the measured image f and the predicted image g and the structural similarity index SSIM between the measured image f and the predicted image g are used as evaluation metrics. If the evaluation metrics decline, the training is overfitting, then the training is terminated, and the model parameters at the previous moment are used as the optimal generative adversarial network parameters; if the training is not overfitting, then step (7) is entered; (7) Repeat the above steps (1)-(6) until overfitting occurs during training or the preset maximum number of iteration steps is reached, and the optimal generative adversarial network parameters are obtained.

2. The water vapor map prediction method according to claim 1, wherein The discriminator is a three-dimensional convolutional neural network, including four three-dimensional convolutional layers and two fully connected layers. The convolutional kernel sizes of the first three three-dimensional convolutional layers in the four three-dimensional convolutional layers are 3×3×3, and the convolutional kernel size of the last three-dimensional convolutional layer is 1×1×1. In the three-dimensional convolutional layer, k, n, and s are used to represent the convolutional kernel size, the number of feature maps, and the stride; the activation function of each three-dimensional convolutional layer is the LeakyReLU function; in the two fully connected layers, the feature dimension output by the first fully connected layer is 1024, and the feature dimension output by the second fully connected layer is 1. The input of the discriminator is the multi-hour water vapor map predicted by the generator, and the output of the discriminator obtains the probability that the water vapor map predicted by the generator belongs to the water vapor map collected by the satellite.

3. A water vapor map prediction device for performing the water vapor map prediction method according to claim 1 or 2, characterized in that, The device includes: An acquisition module, configured to acquire water vapor map data and preprocess the water vapor map data; A network construction module, configured to construct a generative adversarial network, train and validate the generative adversarial network, and obtain the optimal generative adversarial network parameters; An application module, configured to input the water vapor map data into the generative adversarial network to obtain the prediction result of the water vapor map.

4. An electronic device for predicting water vapor images, characterized in that, It includes: A memory, configured to store computer-executable instructions; A processor, where the processor is configured to execute the water vapor map prediction method according to claim 1 or 2.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that The computer program is used to cause the computer to execute the water vapor map prediction method according to claim 1 or 2.

Images