Radar Echo Extrapolation Cycle Generative Adversarial Forecasting Method Based on Edge-Cloud Collaboration

Abstract

The present invention discloses a radar echo extrapolation cyclic generative adversarial prediction method based on edge-cloud collaboration, which relates to the technical field of meteorological forecasting. First, an edge-cloud collaborative hierarchical architecture is constructed for the radar echo extrapolation task, including: cloud layer, edge layer, and sensor layer. Then, a cyclic generative adversarial network is constructed in the edge layer. A cyclic generative model is deployed on each edge server using a cyclic generative architecture to establish a distributed learning model based on federated learning for the radar echo sequence data processed by the sensor layer, and to predict subsequent radar echo sequences. Finally, a dual cross-attention long short-term memory unit is constructed, and a stacked architecture is used to form a cyclic unit based on the long short-term memory network for radar echo extrapolation, by mining the potential connections between radar echo data at different time points. The method of the present invention can effectively utilize the computing power of cloud servers and improve the performance of radar echo prediction at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of weather forecasting, and particularly to a radar echo extrapolation cyclic generative adversarial forecasting method based on edge-cloud collaboration. Background Art

[0002] Excessive precipitation caused by weather often triggers various disasters. Due to the great uncertainty inherent in meteorological events, accurately predicting precipitation is inherently extremely challenging. Forecasting short-term (usually the next few hours) precipitation patterns requires extrapolating radar echoes based on historical data.

[0003] In recent decades, deep learning models have been widely used in various fields. Assuming that radar echo extrapolation belongs to a sequence prediction task, some researchers have started using deep learning models to complete this task. However, although deep learning-based models perform reasonably well when using historical data for prediction, they usually do not work well when predicting high-intensity radar echo regions.

[0004] On the other hand, the operation of meteorological services requires a large amount of computing resources and access to numerous meteorological data sources. The ever-growing data volume and huge computing requirements exceed the capabilities of current meteorological platform technologies. Therefore, it is crucial to develop a meteorological cloud platform centered around cloud computing. Cloud computing platforms have powerful processing capabilities, can manage massive amounts of meteorological data, and perform advanced statistical analyses. At the same time, they are flexible and available, allowing data to be used for decision-making at different times and locations. However, meteorological data sources are scattered, while cloud computing centers are centralized and far away. Offloading computing tasks from distributed meteorological data stations to cloud computing centers will lead to problems such as overly long delays and low efficiency in the computing process. Due to many challenges in cloud computing models, such as network latency, unstable data transmission, and system crashes, it is difficult to ensure the fast and reliable processing and storage of meteorological data.

[0005] In response to these problems, edge computing has emerged. By separating application programs and services, edge computing can process data at the sensor and network edge, thereby improving the system's response speed and efficiency. Therefore, an edge computing framework integrated with a meteorological cloud platform can provide high-performance processing capabilities for meteorological data stations. Although sharing meteorological data among edge servers helps in developing a meteorological cloud platform, there are some unexpected risks in terms of energy consumption that need to be considered. Federated learning (FL) is introduced to address this issue. Essentially, federated learning is a distributed machine learning technique.

[0006] In an edge computing environment, local data processing and model training help reduce the energy consumption associated with data transmission and central server computing, thereby reducing the overall energy consumption of the system. This decentralized processing method not only reduces energy consumption but also decreases the dependence on traditional cloud computing centers, thus reducing carbon emissions. By effectively utilizing the processing capabilities of edge devices, energy consumption can be minimized without sacrificing performance, promoting energy conservation and emission reduction efforts. Therefore, integrating federated learning into the edge computing framework for short-term precipitation pattern forecasting can better understand and predict natural meteorological phenomena. Summary of the Invention

[0007] The problem to be solved by the present invention is to provide a radar echo extrapolation cyclic generative adversarial forecasting method based on edge-cloud collaboration, which effectively utilizes the computing power of cloud servers while improving the performance of radar echo prediction.

[0008] The present invention adopts the following technical solutions: A radar echo extrapolation cyclic generative adversarial forecasting method based on edge-cloud collaboration, comprising the following steps:

[0009] Step 1: Construct an edge-cloud collaborative hierarchical architecture for the radar echo extrapolation task, including: cloud layer, edge layer, and sensor layer. Collect original radar data through the sensor layer to extract radar echo sequences, perform local data processing and transmission through the edge layer, and utilize the computing power of the cloud service center of the cloud layer through the federated learning architecture;

[0010] Step 2: Construct a cyclic generative adversarial network in the edge layer. Adopt a cyclic generative architecture to deploy a cyclic generative model on each edge server, establish a distributed learning model based on federated learning for the radar echo sequence data processed by the sensor layer, and predict subsequent radar echo sequences;

[0011] Step 3: Construct a double-cross attention long short-term memory unit, adopt a stacked architecture to form a cyclic unit based on the long short-term memory network for radar echo extrapolation, and improve the performance of the edge-cloud collaborative hierarchical architecture in generating radar echo predictions by mining the potential connections between radar echo data at different time points.

[0012] Preferably, in the edge-cloud collaborative hierarchical architecture, the cloud layer, edge layer, and sensor layer are interconnected through the Internet;

[0013] The sensor layer includes multiple radar networks, each radar network consists of multiple local radars, which collect original radar data in real time, extract radar echo sequences, and transmit them to the nearest edge server; the radar echo sequence consists of multiple radar echo maps captured at continuous time intervals, and is used to depict the weather conditions in a specific area for a period of time;

[0014] The edge layer contains multiple edge servers that receive radar echo sequence data. As local clients, they perform calculations and transmissions on local data through a cyclic generative adversarial network, update local parameters, and transmit them to the cloud service center in the cloud layer.

[0015] The cloud service center includes storage devices, network devices, and homogeneous or heterogeneous computing devices. After receiving the local parameters from the edge side, the cloud service center performs global aggregation through a global model, refreshes the global model using global parameters, and distributes the global parameters to each edge server.

[0016] After receiving the global parameters sent by the cloud service center, the edge server trains through a local model, calculates the local loss gradient, and updates the local parameters.

[0017] Preferably, in step 2, the cyclic generation model deployed on each edge server includes two identical generators and discriminators; in the edge layer, the input radar echo sequence is split into two subsequences and respectively input into the generators, and the time resolution of each subsequence is half of the original radar echo sequence.

[0018] The generator, constructed based on a convolutional neural network (CNN), includes several layers, and the generated latent sequence results depend on the information of the previous time step and the previous layer, as well as the current input features and hidden states.

[0019] The discriminator, constructed based on a recurrent generative long short-term memory network, obtains the latent sequence results generated by the generator, combines them with the actual sequence, and obtains a scalar to describe the probability that the input radar echo sequence comes from real data or the generator, and evaluates the authenticity of the radar echo sequence.

[0020] The cyclic generation model predicts the next radar echo map based on the received radar echo sequence.

[0021] Preferably, in the distributed learning model, each edge server trains two generators based on the dataset to learn the generated data distribution and , and the method is as follows:

[0022] Step 2.1: Based on the random noise from the probability distribution , a false radar echo sequence is obtained through the generator;

[0023] Step 2.2: Establish discriminators and to distinguish between false images from the data distributions and and real images from the distribution .

[0024] Step 2.3. On each edge server represent the objective function of the cyclic generative adversarial network as a value function .

[0025] Preferably, in step 3, the dual cross-attention long short-term memory unit consists of a forget gate, an input gate, and an output gate, and high-intensity radar echoes are extracted through the memory state, the hidden state, and convolution. The method is as follows;

[0026] Step 3.1. For the initial dual cross-attention long short-term memory unit, the input data is the radar echo map at the current time step and the hidden state at the previous time step of the same layer , and it is updated through the first group of modulation gates , the input gate and the forget gate to obtain the updated long-term memory information ;

[0027] Step 3.2. Combine and process the radar echo map at time step with the spatio-temporal unit , and update it through the second group of modulation gates , the input gate and the forget gate at the current time step to obtain the updated spatio-temporal memory information ;

[0028] Step 3.3. Fuse the long-term memory information and the spatio-temporal memory information through convolution operation to obtain the output control gate ;

[0029] Step 3.4. Concatenate the long-term memory information and the spatio-temporal memory information along the channel dimension, perform convolution operation and then process it through an activation function to obtain the current hidden state .

[0030] Preferably, the cross-cross attention mechanism is implemented through a cross-cross attention unit. The method is as follows:

[0031] Step 3.2.1. For the radar echo map generated by the generator, use two convolutional layers with filters to generate a query map and a key map;

[0032] Step 3.2.2. Perform an affinity operation on the query map and the key map to obtain an attention map;

[0033] Step 3.2.3: Apply another filter convolutional layer to the input data to create values for feature adaptation, perform an aggregation operation between the attention map and the values for feature adaptation, and collect key information;

[0034] Step 3.2.4: Transpose the width and height and feed the output back to the cross - attention unit.

[0035] Compared with the prior art, the present invention adopts the above technical solutions and has the following technical effects:

[0036] 1. The method of the present invention constructs a federated - learning - based edge - cloud collaborative hierarchical architecture for the radar echo extrapolation task to effectively utilize the computing power of the cloud server.

[0037] 2. The method of the present invention constructs a recurrent generative adversarial network, performs distributed learning based on the generative adversarial network (GAN), establishes better short - term dependencies, and improves the training efficiency while ensuring the model accuracy.

[0038] 3. The method of the present invention also incorporates a dual - cross - attention long - short - term memory unit based on the long - short - term memory network to mine the potential connections between radar echo data at different time points and improve the performance of the model in generating radar echo predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a block diagram of the federated - learning - based edge - cloud collaborative hierarchical architecture of the present invention;

[0040] Figure 2 It is a general structure block diagram of the recurrent generation architecture of the present invention;

[0041] Figure 3 It is a structure diagram of the generator of the present invention;

[0042] Figure 4 It is a structure diagram of the dual - cross - attention long - short - term memory unit of the present invention;

[0043] Figure 5 It is a schematic diagram of the process of the cross - attention mechanism of the present invention;

[0044] Figure 6 It is a visualization display diagram of a set of radar echo sequences in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the following further elaborates on the technical solutions of the application in detail with reference to the accompanying drawings. The described embodiments are only a part of the embodiments related to the present invention. All non-innovative embodiments made by other researchers in this field fall within the protection scope of the present invention. At the same time, for the step numbers in the embodiments of the present invention, they are only set for convenience of explanation and illustration, and no limitation is imposed on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0046] In an embodiment of the present invention, a hierarchical edge-cloud collaboration model based on federated learning is proposed, including: a cloud layer, an edge layer, and a sensor layer, which are interconnected through the Internet.

[0047] As Figure 1 shown, the local radar stations in the sensor layer collect raw radar data. After processing the raw radar data in the sensor layer, radar echo sequences are extracted from the raw radar data. To reduce data transmission loss, these radar sequences are transmitted to the nearest edge device.

[0048] Suppose represents the set of local radar stations, defined as ; suppose represents the set of edge devices, defined as .

[0049] The edge layer contains multiple edge servers. To build a federated learning framework, the edge servers act as local clients to calculate local data, while the cloud server in the cloud layer acts as a trusted third party.

[0050] After completing the parameter configuration, the cloud server distributes the global model to each edge server. After receiving the global parameters, the local models of the edge devices are trained to calculate the local loss gradients, thereby updating the local parameters in the model.

[0051] Specifically, the sensor layer contains several radar networks, and each radar network consists of a large number of radars. The sensor layer captures the current state information of the monitored area and collects various required data in real time. The transmission cost of radar data is high, and a large amount of storage space is required. There may be multiple radars running on multiple meteorological data platforms simultaneously.

[0052] In this embodiment, the upload delay of the raw radar data is expressed as:

[0053] ;

[0054] where is the radar network The number of mid-radar echo maps, is the radar network bandwidth, is the signal-to-noise ratio, is the number of local radars.

[0055] Meanwhile, it is not feasible to directly use the original data, and it is necessary to perform unified data access, analysis, and vectorization processing on the original radar data. This process takes a certain amount of time, and the time required for different radar networks will vary.

[0056] In this embodiment, the processing delay of different radar networks is expressed as:

[0057] ;

[0058] where, is the data transmission speed of the radar network and depends on the performance of the local network.

[0059] Let be the transmission power of the device, be the power for processing radar data. Then, the energy consumption of data transmission in the sensor layer is expressed as:

[0060] ;

[0061] where, , respectively represent the data upload time and data processing time, , respectively represent the data upload energy and data processing energy.

[0062] Furthermore, a distributed learning model based on federated learning is constructed in the edge layer.

[0063] In the sensor layer, the local site processes the radar data and transmits it to the edge layer. In the edge layer, each edge device n collects the radar echo sequences from the corresponding radar network and records them as . These sequences are composed of multiple radar echo maps captured at consecutive time intervals and depict the weather conditions in a specific area over a period of time.

[0064] The loss function of the model deployed on each edge device n for the local radar network is expressed as:

[0065] ;

[0066] where, represents the model parameters, is the number of radar echo sequences, is the loss function for each radar echo sequence on it.

[0067] During the global iteration the distributed model on the edge server uses the gradient descent method to update its local parameters according to the global parameters of the previous iteration as follows: The formula is as follows:

[0068] ;

[0069] where represents the learning rate, is the sign of the gradient, represents the loss value calculated based on its local data by the model corresponding to the local radar network on the edge server at the round of global iteration. During the local model training, computing resources are needed to meet the computing requirements, which will cause computing delays and energy overheads.

[0070] Each edge server needs to perform local model training and parameter updates. The computing process, including gradient calculation, model optimization, and parameter updates, requires a large number of floating-point operations and memory resources to support.

[0071] In this embodiment, the usage of computing resources on the edge server is quantified by analyzing the number of floating-point operations performed by the distributed learning model. The computing delay formula for the edge server in one global iteration is:

[0072] ;

[0073] where represents the number of floating-point operations of the distributed model in the edge server , represents the number of floating-point operations per GPU cycle, is the computing resource allocated by the edge server for the radar network .

[0074] Transmission bandwidth limitations or network delays may cause delays when the edge server uploads parameters to the cloud server, thus affecting the speed and effect of global model updates. The upload delay formula is:

[0075] ;

[0076] where is the size of the local model parameters, represents from the radar network The bandwidth allocated for the training task.

[0077] The total energy consumption can be determined by using the computing time of the edge server and the parameter upload delay, and its calculation formula is:

[0078] ;

[0079] Wherein, is the transmission power of the edge server, is the computing power of the edge server, , respectively represent the energy consumption of data upload in the edge layer and the energy consumption of data calculation in the edge layer.

[0080] Furthermore, in the cloud layer, a cloud aggregation model is constructed.

[0081] The components of the cloud layer are the cloud service center, which is composed of various hardware including storage, network, homogeneous or heterogeneous computing, etc. After receiving the local parameters provided by each edge server, the service center will perform a global aggregation operation and refresh the global model with the global parameters. Subsequently, all edge servers will receive the global parameters from the cloud data center.

[0082] In this embodiment, for global iteration, the local parameters are sent to the global model in the cloud layer for model aggregation, and the formula is:

[0083] ;

[0084] Wherein, is all radar data sequences from the radar network in the sensor layer, is the number of radar data sequences, represents the number of edge server devices, represents the global model parameters generated after the

[0085] To accelerate the calculation of global parameters, the cloud data center usually has more powerful processing capabilities and rich computing resources. Therefore, compared with edge devices, the processing delay of the cloud data center is extremely small.

[0086] Then, a distributed learning model is constructed based on the cyclic generation architecture.

[0087] In this embodiment, the proposed cyclic generation model is deployed on each edge device, and this model predicts the next radar echo map according to the radar echo sequence, and the formula is:

[0088] ;

[0089] Wherein, is the number of input radar echo images, is the number of predicted radar echo images, represents the actual radar echo image at time t, represents the predicted radar echo image, The function represents the conditional probability distribution function, represents that based on the echo images of the past time steps (from to ), the probability distribution of the echo images in the next time steps (from to ) is predicted.

[0090] Then a cyclic generative adversarial network (CG-GAN) model is proposed and deployed on edge devices, which consists of two identical generators and discriminators.

[0091] To generate reasonable results for subsequent sequences, the role of the generator is to model the distribution of radar echo data processed by the sensor layer and learn its performance characteristics.

[0092] The discriminator combines the potential sequence results generated by the generator with the actual sequence to obtain a scalar to evaluate the authenticity of the radar echo sequence. This scalar describes the probability that the input radar echo sequence comes from real data or the generator.

[0093] To help the cyclic generative adversarial network model more accurately capture the latent features of the original data, this embodiment also proposes a cyclic generative long short-term memory network, which is embedded in the distributed learning model to improve the quality of the generated data.

[0094] At the edge layer, as Figure 2 shown, the original radar echo sequence is divided into two subsequences, and the time resolution of each subsequence is half of the original sequence. These subsequences are respectively input into the generator. The discriminator will obtain the results generated by the generator to determine whether the input data comes from real data or data generated by the generator.

[0095] The structure of the generator, as Figure 3 shown, each layer has four inputs from the previous time step and the previous layer; The time conversion path of is represented by black arrows, while the spatio-temporal conversion path of

[0096] is highlighted by orange arrows. Therefore, the update of the model depends on the information of the previous time step and the previous layer, as well as the current input features and hidden states.

[0097] The recurrent unit proposed in this embodiment adopts a stacked architecture and consists of a forget gate, an input gate, and an output gate, which is used for radar echo extrapolation. These gates endow the model with the ability to forget past time series features.

[0098] Therefore, the model using the recurrent unit can accurately convey the potential subtle features in the radar echo sequence. The traditional LSTM unit extracts high-intensity radar echoes through the memory state, hidden state, and simple convolution. However, due to the small number of high-intensity echo regions, the model based on the traditional LSTM unit will miss these regions and is difficult to characterize their features. This embodiment proposes a DCA-LSTM unit, which is equipped with a novel attention mechanism to enhance the ability to capture high-intensity echo features.

[0099] The structure of the DCA-LSTM unit proposed in this embodiment is shown in Figure 5 Figure. Long-term memory represents the state of the time memory unit at time step and is used to store long-term information when processing time series data. The hidden state stores the data processed by the model at the previous time step and contains the knowledge and memory of the input sequence to determine the behavior of the model. Spatiotemporal memory is the part of the model used to store and understand the spatiotemporal characteristics of the input sequence. It retains the memory state and stores and updates long-term dependencies.

[0100] Specifically, the DCA-LSTM unit takes the radar echo map at time as the input, and also inputs the hidden state of the previous time step in the same layer. The long-term memory unit is updated through an initial set of modulation gates , input gates , and forget gates . These gates constitute a gating mechanism for .

[0101] The formula for this process is:

[0102] ;

[0103] ;

[0104] ;

[0105] ;

[0106] where represents the Sigmoid activation function, , , respectively represent the spatio-temporal memory input gate at time step t, the spatio-temporal memory modulation gate output at time step t, and the spatio-temporal memory forget gate output at time step , represents the radar echo map (input data) at a time step, which includes spatial dimensions (height, width) and feature channels (such as reflectivity, radial velocity), , , respectively represent the spatial weight matrix of the input to the input gate, the spatial weight matrix of the input to the modulation gate, and the spatial weight matrix of the input to the forget gate, , , respectively represent the time weight matrix of the hidden state at the previous time step to the input gate, the time weight matrix of the hidden state at the previous time step to the modulation gate, and the time weight matrix of the hidden state at the previous time step to the forget gate, , , respectively represent the bias term of the input gate, the bias term of the modulation gate, and the bias term representing the forget gate; represents the interactive dual attention mechanism, is the number of previous radar echo maps, represents the long-term memory sequence of the past time steps, represents the updated long-term memory information of the th layer at the current time step, represents the matrix product, represents the convolution process.

[0107] Meanwhile, the radar echo map at time will be processed together with the spatio-temporal unit . These combined information will then be updated through the second set of modulation gates , input gates and forget gates at the current time step, so as to obtain the updated spatio-temporal memory unit .

[0108] The formula for this process is:

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] Among them, , , respectively represent the spatial weight matrix input from to the input gate of the spatio-temporal memory unit , the spatial weight matrix input from to the modulation gate of the spatio-temporal memory unit , and the spatial weight matrix input from to the forget gate of the spatio-temporal memory unit . , , respectively represent the temporal weight matrix from the hidden state at the previous time step to the input gate of the spatio-temporal memory unit , the temporal weight matrix from the hidden state at the previous time step to the modulation gate of the spatio-temporal memory unit , and the temporal weight matrix from the hidden state at the previous time step to the forget gate of the spatio-temporal memory unit . represents the number of layers of the network, , , respectively represent the outputs of the input gate of the spatio-temporal memory unit at time step , the outputs of the modulation gate of the spatio-temporal memory unit at time step , and the outputs of the forget gate of the spatio-temporal memory unit at time step . represents the cross-attention mechanism, , respectively represent the spatio-temporal memory units of the th layer and the th layer.

[0114] Finally, the long-term memory information and the spatio-temporal memory information are fused through a convolution operation to obtain the output control gate .

[0115] In addition, after the long-term memory information and the spatio-temporal memory information are concatenated along the channel dimension, a convolution operation is performed, and then after being processed by an activation function, the current hidden state is obtained.

[0116] The formula for this process is:

[0117] ;

[0118] ;

[0119] Among them, represents matrix multiplication, represents the convolution process, represents the tensor splicing operation.

[0120] In this embodiment, the cross-shaped attention unit is as Figure 4 shown.

[0121] Assume that in the generator, the radar echo map is processed into size. First, two convolution layers with filters are used to generate the query map and the key map .

[0122] Then, by performing an affinity operation on the query map and the key map , the attention map is obtained. At the same time, another filter convolution layer is applied to the input data to create a value for feature adaptation; since an aggregation operation is performed between and , key information can be collected.

[0123] Finally, in order to feedback the output back to the unit, the last two dimensions (width and height) are transposed.

[0124] Specifically, this embodiment also introduces a loss function with cyclic consistency loss. Each edge device trains two generators based on the dataset to learn the generated data distributions and . Based on the random noise from the probability distribution , a fake radar echo sequence is obtained through the generator, denoted as .

[0125] At the same time, discriminators and are designed to distinguish between fake images from the distributions and and real images from the distribution .

[0126] The objective function of the cyclic generative adversarial network (CG-GAN) on each edge server , through the value function Denote:

[0127] ;

[0128] ;

[0129] ;

[0130] Wherein, respectively denote the discriminator set, which contains two discriminators and , the generator set, which contains two generators and ; respectively denote the odd-time-step sub-sequence discriminator and the odd-time-step sub-sequence generator, respectively denote the even-time-step sub-sequence discriminator and the even-time-step sub-sequence generator, , , respectively denote the distribution of the real radar echo sequence at odd time steps, the distribution of the real radar echo sequence at even time steps, and the noise distribution of the generator input; denotes expectation, and respectively denote the probabilities that the discriminators and determine as a real data sample, while and respectively denote the probabilities that the discriminators and determine the data generated by and as true.

[0131] Furthermore, in this embodiment, the CIKM-2017 dataset is adopted, and the proposed model is evaluated experimentally using meteorological indicators. The radar maps in the dataset show the radar reflectivity in the area of 101 km × 101 km adjacent to the central station. The radar echo sequence in the dataset contains 15 radar echo maps, and the interval between each radar echo image is 6 minutes.

[0132] In this experiment, the constructed edge-cloud collaborative hierarchical architecture involves experiments with 5 edge nodes, 1 cloud service node, and 100 local radar networks. The parameters used in the experiment are shown in Table 1. The stochastic gradient descent method (SGD) is selected as the optimizer, and the batch size is set to 8, which can not only reduce the risk of crossing the optimal parameters but also ensure the convergence speed. The learning rate of the model is set to 0.003, and the loss function is as described above.

[0133] Table 1. Experimental settings

[0134]

[0135] The evaluation metrics used in this experiment are set as follows:

[0136] 1. Z-R relationship: The Z-R relationship is often used in radar echo tasks to convert echo readings into precipitation amounts. It describes the connection between the radar reflectivity factor Z and the precipitation rate R, and the formula is:

[0137] ;

[0138] where and depend on specific climatic conditions and geographical locations. Therefore, once the values of and are determined, the radar reflectivity value can be converted into the precipitation rate, and vice versa. The three radar reflectivity thresholds selected in this paper are 10 dBZ, 20 dBZ, and 30 dBZ respectively.

[0139] 2. Probability of Detection (POD) score: The Probability of Detection score is used to evaluate the reliability of the model in accurately predicting the occurrence of precipitation events. Conversely, the False Alarm Rate (FAR) score is used to evaluate the situation where the model predicts precipitation events that actually do not occur.

[0140] 3. Critical Success Index (CSI) score: The Critical Success Index score measures the precipitation prediction probability provided by the model.

[0141] 4. Structural Similarity Index (SSIM): The Structural Similarity Index is used to quantify the visual similarity between the prediction result and the actual precipitation pattern.

[0142] The calculation formulas for the values of the above evaluation metrics used in this embodiment are:

[0143] ;

[0144] ;

[0145] ;

[0146] where TP is the true positive, representing the number of precipitation events correctly predicted by the model; FN is the false negative, representing the number of precipitation events missed by the model; and FP is the false positive, representing the number of precipitation events misreported by the model.

[0147] The experimental results are as follows:

[0148] This embodiment comprehensively evaluates the performance of the proposed cycle generative adversarial network model in radar echo extrapolation. To achieve this goal, several representative models are selected for comparison. These models have shown good performance in previous studies, but their performance in high-intensity regions is only at an average level.

[0149] The detection probability (POD) scores and false alarm rate (FAR) scores of the model proposed in this embodiment and the models used for comparison on different threshold datasets are shown in Table 2.

[0150] Table 2. Comparison of detection probability and false alarm rate under three threshold conditions

[0151]

[0152] Except for the cycle generative adversarial network (CG-GAN) of the present invention, the interactive dual-attention long short-term memory network (IDA-LSTM) has higher POD scores than other models at 10 dBZ, 20 dBZ, and 30 dBZ thresholds.

[0153] The CG-GAN proposed in the present invention is improved by 3.77% and 4.17% compared with IDA-LSTM at 20 dBZ and 30 dBZ thresholds respectively. However, the proposed CG-GAN is not the most effective at 10 dBZ threshold and is not an ideal choice at lower thresholds.

[0154] In terms of false alarm rate, CG-GAN has obtained the lowest scores at all selected thresholds, which indicates that the proposed model has significant advantages in prediction accuracy. Compared with the second-ranked scores, the scores of CG-GAN at 10 dBZ, 20 dBZ, and 30 dBZ thresholds are reduced by 1.1%, 0.87%, and 3.38% respectively.

[0155] The following Table 3 shows the comparison based on the critical success index and the comparison based on the structural similarity index under three threshold conditions, as well as the CSI and SSIM scores of all models.

[0156] Table 3. CSI and SSIM scores of all models

[0157]

[0158] It can be seen that compared with other comparison models, the CG-GAN model proposed by the present invention has achieved higher scores. Specifically, in terms of CSI, at the 10 dBZ threshold, CG-GAN is 4.49% higher than the gated attention convolutional gated recurrent unit (GA-ConvGRU) and 2.09% higher than the interactive dual attention long short-term memory network (IDA-LSTM). At the 20 dBZ threshold, the score is increased by 5.24% compared with the convolutional long short-term memory network (ConvLSTM); and it is increased by 4.46% compared with the upgraded version of the predictive recurrent neural network (PredRNN++). In addition, at the 30 dBZ threshold, the proposed CG-GAN is 1.97% and 2.55% higher than GA-ConvGRU and PredRNN++ in terms of scores respectively.

[0159] The SSIM score of CG-GAN is the highest, which indicates that it has better visual quality and is more similar to the actually observed radar map, thus improving its precipitation prediction ability.

[0160] The visualization display of the radar echo sequence in this embodiment, as Figure 6 shown, reflects the extrapolation results of different models for a standard radar echo sequence in the dataset. The colored areas in the echo map correspond to specific reflection intensity levels, as shown in the color scale. This sequence shows the process of the radar echo moving from the upper right corner to the center.

[0161] Based on the basic structure of the radar echo, the convolutional long short-term memory network (ConvLSTM) and the gated attention convolutional gated recurrent unit (GA-ConvGRU) models can only predict the total reflectivity value of the echo area and cannot accurately predict the high-intensity echo area. The upgraded version of the predictive recurrent neural network (PredRNN++) improves the ability to extract spatio-temporal feature information by adopting a gating method based on the ConvLSTM unit and the temporal memory unit. However, due to the inability to naturally fuse temporal and feature data, its global spatio-temporal feature prediction ability is very weak. The interactive dual attention long short-term memory network (IDA-LSTM) model introduces an attention mechanism and increases the prediction weight of the high-reflectivity area on the basis of the typical spatio-temporal memory unit, so as to accurately predict the high-intensity echo area while ignoring the low-reflectivity value area.

[0162] Furthermore, in order to verify the effectiveness of each design component in the cyclic generative adversarial network (CG-GAN) of the present invention in improving the model performance, and further enhance the interpretability of the model, an ablation experiment was conducted on CG-GAN.

[0163] Table 4 shows the quantitative evaluation results of these experiments on each component of the CG-GAN. The bold numbers indicate the best performance of the model at the current threshold. The spatio-temporal long short-term memory network (ST-LSTM) represents the baseline model without additional modules. CG-GAN (without DCC) represents the extrapolation network composed of double cross-attention long short-term memory (DCA-LSTM) units but without the double cross-attention mechanism. CG-GAN (without ConLoss) represents the extrapolation network without using the cycle consistency loss function internally.

[0164] Table 4. Quantitative Evaluation Results of Ablation Experiments

[0165]

[0166] Removing any one component from the CG-GAN will lead to a continuous decline in the performance of the model on all evaluation metrics. Specifically, when only relying on the ST-LSTM model to extract spatio-temporal features from radar echoes, the critical success index (CSI) scores of the model decrease by 5.39%, 2.96%, and 6.65% at the 10dBZ, 20dBZ, and 30dBZ thresholds, respectively. Correspondingly, the probability of detection (POD) scores decrease by 5.53%, 4.92%, and 4.25%, respectively. Similarly, after removing the double cross-attention mechanism, the CSI scores decrease by 2.92%, 0.67%, and 1.02% at the 10dBZ, 20dBZ, and 30dBZ thresholds, respectively, and the POD scores decrease by 1.99%, 1.66%, and 1.77%, respectively.

[0167] In addition, after removing the cycle consistency loss function from the extrapolation model, the CSI scores decrease by 2.18%, 1.89%, and 2.87% at the 10dBZ, 20dBZ, and 30dBZ thresholds, respectively, and the POD scores decrease by 0.75%, 3.03%, and 0.89%, respectively.

[0168] In an embodiment of the present invention, an electronic device is further provided, including: one or more processors; a storage device on which one or more programs are stored; when the one or more programs are executed by the one or more processors, the one or more processors implement the radar echo extrapolation cyclic generative adversarial prediction method based on edge-cloud collaboration described in the above embodiments.

[0169] In an embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored, and when the program is executed by a processor, the steps in the radar echo extrapolation cyclic generative adversarial prediction method based on edge-cloud collaboration in the above embodiments are implemented.

[0170] In summary, the present invention promotes the collaborative design of generative adversarial networks and federated learning, constructs an edge-cloud collaborative hierarchical model for radar echo extrapolation, and designs a federated learning architecture under edge-cloud collaboration to effectively utilize the computing power of cloud servers. At the same time, a GAN-based cyclic generative adversarial network inspired by CycleGAN is constructed, which consists of two generators based on long short-term memory networks deployed on edge servers and two discriminators based on convolutional neural networks deployed in the cloud service center, with the aim of establishing better short-term dependencies. A new dual cross-attention long short-term memory unit derived from LSTM is also incorporated into the GAN-based cyclic generative adversarial network architecture to explore the potential connections between radar echo data at different time points and improve the performance of the model in generating radar echo predictions.

[0171] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A radar echo extrapolation cyclic generative adversarial prediction method based on edge-cloud collaboration, characterized in that It includes the following steps: Step 1: Build an edge-cloud collaborative hierarchical architecture for radar echo extrapolation tasks, including: cloud layer, edge layer, and sensor layer. Collect original radar data through the sensor layer to extract radar echo sequences, perform local data processing and transmission through the edge layer, and utilize the computing power of the cloud service center in the cloud layer through the federated learning architecture; In the edge-cloud collaborative hierarchical architecture, the cloud layer, edge layer, and sensor layer are interconnected through the Internet; The sensor layer contains multiple radar networks, each radar network consists of multiple local radars, which collect original radar data in real time, extract radar echo sequences, and transmit them to the nearest edge server; the radar echo sequence consists of multiple radar echo maps captured at continuous time intervals, and is used to depict the weather conditions in a specific area for a period of time; The edge layer contains multiple edge servers, which receive radar echo sequence data, and as local clients, perform calculations and transmissions on local data through a cyclic generative adversarial network, update local parameters, and transmit them to the cloud service center in the cloud layer; The cloud service center contains storage devices, network devices, and homogeneous or heterogeneous computing devices; after receiving the local parameters from the edge layer, the cloud service center performs global aggregation through a global model, refreshes the global model with global parameters, and distributes the global parameters to each edge server; After receiving the global parameters sent by the cloud service center, the edge server performs training through the local model on the edge side, calculates the local loss gradient, and updates the local parameters; Step 2: Build a cyclic generative adversarial network in the edge layer, adopt a cyclic generative architecture to deploy a cyclic generative model on each edge server, establish a distributed learning model based on federated learning for the radar echo sequence data processed by the sensor layer, and predict the subsequent radar echo sequence; In the distributed learning model, it is deployed on the edge server for the radar network loss function , which is expressed as: ； Among them, are the parameters of the distributed model, represents the edge server collects the radar echo sequences from the corresponding local radar network and is the number of radar echo sequences, represents the loss function on each radar echo sequence ; In the global iteration the distributed model on the edge server uses the gradient descent method to update the local parameters according to the global parameters of the previous iteration as follows: The formula is as follows: ； Among them, is the learning rate, is the sign of the gradient, indicates that at the th round of global iteration, the corresponding local radar network on the edge server, the loss value calculated by the model based on local data; The cyclic generative model deployed on each edge server contains two groups of identical generators and discriminators; In the edge layer, the input radar echo sequence is split into two subsequences and input into the generator respectively, and the time resolution of each subsequence is half of the original radar echo sequence; The generator is built based on a convolutional neural network, includes several layers, and the generated latent sequence result depends on the information of the previous time step and the previous layer, as well as the current input features and hidden states; The discriminator is built based on a cyclic generative long short-term memory network, obtains the latent sequence result generated by the generator, combines it with the actual sequence, and gets a scalar, which is used to describe the probability that the input radar echo sequence comes from real data or the generator, and evaluates the authenticity of the radar echo sequence; The cyclic generative model predicts the next radar echo map according to the received radar echo sequence, and the formula is: ； Among them, is the number of input radar echo maps, is the number of predicted radar echo maps, represents the actual radar echo map, represents the predicted radar echo map, The function represents the conditional probability distribution function, represents that based on the echo maps of the past time steps, the probability distribution of the echo maps of the future time steps is predicted; In the distributed learning model, each edge server trains two generators based on the dataset to learn the generated data distribution and , and the method is as follows: Step 2.1: Based on the random noise from the probability distribution , a false radar echo sequence is obtained through the generator ; Step 2.2, establish a discriminator and , which is used to distinguish between fake images from the data distribution and and real images from the distribution ; Step 2.3, on each edge server The objective function of the generative adversarial network is expressed as a value function : ; ； ； Among them, , respectively represent the discriminator set and the generator set, respectively represent the odd-time-step sub-sequence discriminator and generator, respectively represent the even-time-step sub-sequence discriminator and generator, , , respectively represent the distributions of the real radar echo sequences at odd time steps, the real radar echo sequences at even time steps, and the noise distribution of the generator input; represents the expectation, and respectively represent the discriminator and the probability of determining as a real data sample, and respectively represent the discriminator and the probability of determining the data generated by and as true; Step 3: Build a double cross-attention long short-term memory unit, adopt a stacked architecture to form a cyclic unit based on the long short-term memory network, perform radar echo extrapolation, and improve the performance of the edge-cloud collaborative hierarchical architecture in generating radar echo predictions by mining the potential connections between radar echo data at different time points; The double-cross attention long short-term memory unit consists of a forgetting gate, a modulation gate, an input gate, and an output gate, and extracts high-intensity radar echoes through the memory state, the hidden state, and convolution.

2. The radar echo extrapolation cyclic generative adversarial prediction method based on edge-cloud collaboration according to claim 1, wherein The sensor layer collects raw radar data through multiple local radars on multiple meteorological data platforms simultaneously, and the upload delay of the raw data of the local radars , which is expressed as: ； Among them, is the number of radar echo maps in the radar network, is the bandwidth of the radar network ; is the signal-to-noise ratio, is the number of local radars; Processing delays of different radar networks , expressed as: ； Among them, is the data transmission speed of the radar network ; Data transmission energy consumption of the sensor layer , expressed as: ； Among them, is the transmission power of the device, is the power for processing radar data, , respectively represent the data upload time and the data processing time, , respectively represent the data upload energy and the data processing energy.

3. The radar echo extrapolation cyclic generative adversarial prediction method based on edge-cloud collaboration according to claim 2, wherein Local model training and local parameter updates are performed in each edge server, including: gradient calculation, local model optimization, and parameter updates. The method is as follows: The number of floating-point operations performed by the distributed learning model quantifies the usage of computing resources on the edge server, and the computing latency of the edge server in one global iteration , expressed as: ； Among them, represents the floating-point operation count of the distributed model in the edge server, represents the floating-point operation count per GPU cycle, represents the computing resources allocated by the edge server for the radar network; Parameter upload delay that occurs when the edge server uploads parameters to the cloud server , expressed as: ； Among them, represents the size of the local model parameters, represents the bandwidth allocated for the training task from the radar network ; Determine the total energy consumption based on the computing delay and parameter upload delay of the edge server , and the calculation formula is as follows: ； Among them, represents the transmission power of the edge server, represents the computing power of the edge server, , respectively represent the energy consumption of data uploading in the edge layer and the energy consumption of data computing in the edge layer.

4. The radar echo extrapolation cyclic generative adversarial prediction method based on edge-cloud collaboration according to claim 3, wherein The cloud service center receives the local parameters uploaded by each edge server and performs model aggregation through the global model: ； Among them, is all the radar data sequences from the radar network of the sensor layer, is the number of radar data sequences, represents the number of edge server devices, represents the global model parameters generated after the -th round of global iteration.

5. The radar echo extrapolation cyclic generative adversarial prediction method based on edge-cloud collaboration according to claim 1, wherein The double-cross attention long short-term memory unit extracts high-intensity radar echoes through the memory state, the hidden state, and convolution. The method is as follows; Step 3.

1. The initial dual-cross attention long short-term memory unit takes the radar echo map at the current time step and the hidden state at the same layer and previous time step as input, and updates them through the first set of modulation gates , input gates , and forget gates to obtain the updated long-term memory information . The update process is expressed as: ​ ； ； ； ； Among them, represents the Sigmoid activation function; , , respectively represent the spatio-temporal memory input gate, spatio-temporal memory modulation gate output, and spatio-temporal memory forgetting gate output at time step ; represents the radar echo map at time step , which contains height, width spatial dimensions, and feature channels. , , respectively represent the spatial weight matrices of the input data to the input gate, modulation gate, and forgetting gate. , , respectively represent the temporal weight matrices of the hidden state at the previous time step to the input gate, modulation gate, and forgetting gate. , , respectively represent the bias term of the input gate, the bias term of the modulation gate, and the bias term of the forgetting gate. represents the interactive dual attention mechanism. is the number of previous radar echo maps. represents the long-term memory sequence of the past time steps. represents the long-term memory information of the th layer after update at the time step. represents the matrix product. represents the convolution process; Step 3.2: Combine the radar echo map at time step with the spatio-temporal unit and perform an update at the current time step through the second set of modulation gates , input gates and forget gates to obtain the updated spatio-temporal memory information . The update process is expressed as: ； ； ； ； Among them, , , respectively represent the outputs of the input gate, modulation gate, and forget gate of the spatio-temporal memory unit at time step . , , respectively represent the spatial weight matrices of the input data to the input gate , modulation gate , and forget gate of the spatio-temporal memory unit. , , respectively represent the time weight matrices of the hidden state at the previous time step to the input gate , modulation gate , and forget gate of the spatio-temporal memory unit. represents the number of layers of the network. represents the cross-attention mechanism. , respectively represent the spatio-temporal memory units of the th layer and the th layer at time step ; Step 3.

3. Fuse the long-term memory information and the spatio-temporal memory information through convolution operation to obtain the output control gate , and the formula is as follows: ； Among them, , , , respectively represent the current input data , the hidden state at the previous time step , the long-term memory unit , the spatio-temporal memory unit to the weight matrix of the output gate ; represents the bias term of the output gate, represents the time step the output of the output gate; Step 3.4, use the long-term memory information and the spatio-temporal memory information to splice them along the channel dimension and perform convolution operation, followed by processing with an activation function, to obtain the current hidden state , and the formula is as follows: ； Among them, represents convolution operation, represents tensor concatenation operation.

6. The radar echo extrapolation cyclic generative adversarial prediction method based on edge-cloud collaboration according to claim 5, characterized in that The cross-cross attention mechanism is implemented through the cross-cross attention unit. The method is as follows: Step 3.2.

1. For the radar echo map generated by the generator, use two convolutional layers with filters to generate a query map and a key map ; Step 3.2.2, perform an affinity operation on the query graph and the key graph to obtain the attention graph ; Step 3.2.3, apply another filter convolutional layer to create values for feature adaptation , and perform an aggregation operation between and to collect key information; Step 3.2.4: Transpose the width and height, and output the feedback back to the cross-cross attention unit.

Images