Radar echo extrapolation cycle generative adversarial forecasting method based on edge cloud cooperation

Abstract

The invention discloses a radar echo extrapolation cycle generative adversarial forecasting method based on edge cloud cooperation, and relates to the technical field of weather forecasting. The method comprises the following steps: firstly, constructing an edge-cloud cooperation layered architecture for a radar echo extrapolation task, which comprises a cloud layer, an edge layer and a sensor layer; then, constructing a cyclic generative adversarial network on the edge layer, deploying a cyclic generation model on each edge server by adopting a cyclic generation architecture, establishing a distributed learning model based on federated learning for the radar echo sequence data processed by the sensor layer, and predicting a subsequent radar echo sequence; and finally, constructing a double-cross attention long-short-term memory unit, forming a circulating unit based on a long-short-term memory network by adopting a stacked architecture, carrying out radar echo extrapolation, and mining a potential relation between radar echo data at different time points. According to the method, the computing power of the cloud server can be effectively utilized, and meanwhile, the radar echo prediction performance is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a radar echo extrapolation cycle generation adversarial forecasting method based on edge cloud collaboration. Background Art

[0002] Excessive precipitation brought by weather often causes various disasters. Due to the great uncertainty of meteorological events themselves, accurate precipitation prediction is inherently challenging. Forecasting precipitation patterns in the short term (usually the next few hours) 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 is a sequence prediction task, some researchers have begun to use deep learning models to complete this task. However, although deep learning-based models perform well when using historical data for prediction, they usually do not perform well when predicting high-intensity radar echo areas.

[0004] On the other hand, the operation of meteorological services requires a lot of computing resources and access to many meteorological data sources. The growing amount of data and huge computing requirements are beyond the capabilities of current meteorological platform technologies. Therefore, it is crucial to develop a meteorological cloud platform with cloud computing as the core. Cloud computing platforms have powerful processing capabilities, can manage massive amounts of meteorological data, and perform advanced statistical analysis, while having flexibility and availability, 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. Distributed meteorological data stations offload computing tasks to cloud computing centers, which will cause long delays and inefficiencies in the computing process. Due to the many challenges of cloud computing models, such as network delays, unstable data transmission, and system crashes, it is difficult to ensure that meteorological data processing and storage are fast and reliable.

[0005] In response to these problems, edge computing came into being. By separating applications and services, edge computing can process data at the edge of sensors and networks, thereby improving the response speed and efficiency of the system. Therefore, the edge computing framework integrated with the meteorological cloud platform can provide high-performance processing capabilities for meteorological data sites. Although sharing meteorological data between edge servers helps to develop meteorological cloud platforms, there are some unexpected risks in terms of energy consumption that need to be considered. Federated learning (FL) was introduced to solve this problem. In essence, federated learning is a distributed machine learning technology.

[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 reduces dependence on traditional cloud computing centers, thereby reducing carbon emissions. By effectively utilizing the processing power of edge devices, energy consumption can be minimized without affecting performance, promoting energy conservation and emission reduction. 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 cycle generation adversarial prediction method based on edge cloud collaboration, which effectively utilizes the computing power of the cloud server and improves the performance of radar echo prediction.

[0008] The present invention adopts the following technical solution: a radar echo extrapolation cycle generation countermeasure prediction method based on edge cloud collaboration, comprising the following steps: Step 1: Build an edge-cloud collaborative layered architecture for the radar echo extrapolation task, including: cloud layer, edge layer, and sensor layer. The sensor layer collects raw radar data to extract radar echo sequences, and the edge layer performs local data processing and transmission. The federated learning architecture utilizes the computing power of the cloud service center in the cloud layer. Step 2: Build a recurrent generative adversarial network at the edge layer. Use a recurrent generative architecture to deploy a recurrent 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 to predict the subsequent radar echo sequence. Step 3: Construct a double cross-attention long short-term memory unit and adopt a stacked architecture to form a recurrent unit based on the long short-term memory network to perform radar echo extrapolation. By mining the potential connections between radar echo data at different time points, the performance of the edge-cloud collaborative hierarchical architecture in generating radar echo predictions is improved.

[0009] Preferably, in the edge-cloud collaborative layered architecture, the cloud layer, edge layer, and sensor layer are interconnected via the Internet; The sensor layer includes multiple radar networks, each of which is composed of multiple local radars, which collect raw radar data in real time, extract radar echo sequences, and transmit them to the nearest edge server; the radar echo sequence is composed of multiple radar echo images captured in continuous time intervals, which are used to describe the weather conditions in a specific area over a period of time; The edge layer includes multiple edge servers, which receive radar echo sequence data and act as local clients to calculate and transmit 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 includes storage devices, network devices, and homogeneous or heterogeneous computing devices. After receiving local parameters on the edge side, the cloud service center performs global aggregation through the global model, refreshes the global model using global parameters, and distributes the global parameters to each edge server. After receiving the global parameters from the cloud service center, the edge server trains the local model, calculates the local loss gradient, and updates the local parameters.

[0010] Preferably, in step 2, the cyclic generation model deployed on each edge server includes two sets of identical generators and discriminators; in the edge layer, the input radar echo sequence is divided 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 on a convolutional neural network (CNN) and consists of several layers. The generated potential sequence results depend on the information of the previous time step and the previous layer, as well as the current input features and hidden state; The discriminator is constructed based on the long short-term memory network generated by the recurrent generator. It obtains the potential sequence results generated by the generator and combines them with the actual sequence to obtain a scalar, which is used to describe the probability that the input radar echo sequence comes from the real data or the generator, and evaluates the authenticity of the radar echo sequence. The cyclic generative model predicts the next radar echo pattern based on the received radar echo sequence.

[0011] Preferably, in the distributed learning model, each edge server is based on the data set Train two generators to learn to generate data distribution and , the method is as follows: Step 2.1: Based on the probability distribution Random noise , the false radar echo sequence is obtained through the generator ; Step 2.2: Build a discriminator and , used to distinguish from data distribution and The false images are from the distribution The real image of Step 2.3: On each edge server The objective function of the cyclic generative adversarial network is expressed as the value function .

[0012] Preferably, in step 3, the double cross attention long short-term memory unit is composed of a forget gate, an input gate and an output gate, and the high-intensity radar echo is extracted through the memory state, the hidden state and the convolution, and the method is as follows; Step 3.1, initial double cross attention long short-term memory unit, input data is the current time step The radar echo map and the previous time step of the same layer The hidden state of , through the first set of modulation gates , Input Gate and forget gate Update to obtain updated long-term memory information ; Step 3.2: Set the time step Radar echogram and space-time unit Combined processing, passing through the second set of modulation gates at the current time step , Input Gate and forget gate Update to obtain updated spatiotemporal memory information ; Step 3.3: Long-term memory information and spatiotemporal memory information Through convolution operation, the output control gate is obtained. ; Step 3.4: Store information in long-term memory and spatiotemporal memory information Splicing along the channel dimension, After the convolution operation, the activation function is processed to obtain the current hidden state .

[0013] Preferably, the cross attention mechanism is implemented by a cross attention unit, and the method is as follows: Step 3.2.1: For the radar echo map generated by the generator, use two The convolutional layer of filters generates query graph and key graph; Step 3.2.2, perform affinity operation on the query graph and the key graph to obtain the attention graph; Step 3.2.3: Apply another The filter convolution layer creates values ​​for feature adaptation, performs aggregation operations between the attention map and the values ​​for feature adaptation, and collects key information; Step 3.2.4, transpose the width and height, and feed the output back to the cross attention unit.

[0014] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects: 1. The method of the present invention constructs an edge-cloud collaborative hierarchical architecture based on federated learning for radar echo extrapolation tasks to effectively utilize the computing power of cloud servers.

[0015] 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 training efficiency while ensuring model accuracy.

[0016] 3. The method of the present invention also incorporates a double cross-attention long short-term memory unit based on a long short-term memory network 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. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a block diagram of the edge-cloud collaboration layered architecture based on federated learning of the present invention; Figure 2 The overall structural block diagram of the loop generation architecture of the present invention; Figure 3 This is a structural diagram of the generator of the present invention; Figure 4 This is a structural diagram of the double cross attention long short-term memory unit of the present invention; Figure 5 It is a schematic diagram of the cross attention mechanism flow of the present invention; Figure 6 This is a visualization diagram of a set of radar echo sequences according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0019] In one embodiment of the present invention, a federated learning-based edge-cloud collaborative hierarchical model is proposed, including: a cloud layer, an edge layer, and a sensor layer, and these layers are interconnected through the Internet.

[0020] like Figure 1 As shown in Figure 1, the local radar station at the sensor layer collects raw radar data, and after the raw radar data is processed at 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.

[0021] Assumptions represents the set of local radar stations, defined as Assumptions Represents a collection of edge devices, defined as .

[0022] The edge layer contains multiple edge servers. To build a federated learning framework, the edge servers act as local clients to perform calculations on local data, while the cloud servers in the cloud layer act as trusted third parties.

[0023] After completing the parameter configuration, the cloud server will distribute the global model to each edge server. After the edge device receives the global parameters, its local model will be trained to calculate the local loss gradient, thereby updating the local parameters in the model.

[0024] Specifically, the sensor layer contains several radar networks, each of which consists of a large number of radars. The sensor layer captures the current status information of the monitored area and collects various required data in real time. The transmission cost of radar data is high and requires a lot of storage space. There may be multiple radars running at the same time on multiple meteorological data platforms.

[0025] In this embodiment, the radar raw data upload delay is expressed as: ; in, For radar networks The number of radar echograms in For radar networks bandwidth, is the signal-to-noise ratio, is the number of local radars.

[0026] At the same time, it is not feasible to use the raw data directly. It is necessary to uniformly access, analyze and vectorize the raw radar data. This process takes a certain amount of time, and the time required for different radar networks will vary.

[0027] In this embodiment, the processing delays of different radar networks are expressed as: ; in, It's a radar network The data transmission speed depends on the performance of the local network.

[0028] set up is the transmission power of the device, is the power for processing radar data. Then, the data transmission energy consumption of the sensor layer is expressed as: ; in, , They represent data upload time and data processing time respectively. , They represent data uploading energy and data processing energy respectively.

[0029] Furthermore, a distributed learning model based on federated learning is constructed at the edge layer.

[0030] At the sensor layer, the local site processes the radar data and transmits it to the edge layer. At the edge layer, each edge device n receives data from the corresponding radar network. Collect radar echo sequence, denoted as These sequences consist of multiple radar echograms captured at consecutive time intervals and depict weather conditions over a specific area over time.

[0031] Deployed on each edge device n for the local radar network The loss function of the model is expressed as: ; in, represents the model parameters, is the number of radar echo sequences, is each radar echo sequence The loss function on .

[0032] In global iteration In the distributed model on the edge server, the gradient descent method is used to calculate the global parameters of the previous iteration. Update its local parameters , the formula is as follows: ; in, represents the learning rate, is the sign of the gradient, Indicated in During the global iteration, the edge server Corresponding to the local radar network The loss value calculated based on the local data of the model. During local model training, computing resources need to be used to meet the computing needs, which will cause computing delays and energy overhead.

[0033] Each edge server needs to perform local model training and parameter updates. The calculation process, including gradient calculation, model optimization, and parameter updates, requires a large amount of floating-point operations and memory resources to support.

[0034] 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 calculation delay formula of the edge server in one global iteration is: ; in, Represents edge server The number of floating point operations in the distributed model, Indicates the number of floating-point operations per GPU cycle, It is an edge server For radar networks Allocated computing resources.

[0035] Transmission bandwidth limitations or network delays may cause delays when edge servers upload parameters to cloud servers, thus affecting the speed and effect of global model updates. The formula for calculating upload delay is: ; in, is the size of the local model parameters, Represented as from the radar network The bandwidth allocated to the training tasks.

[0036] The total energy consumption can be determined by using the edge server's computation time and parameter upload delay, and the calculation formula is: ; in, is the transmission power of the edge server, is the computing power of the edge server, , They represent the energy consumption of edge layer data uploading and edge layer data calculation respectively.

[0037] Furthermore, at the cloud layer, a cloud aggregation model is constructed.

[0038] The cloud layer consists of a 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.

[0039] In this embodiment, for global iteration, local parameters are sent to the global model of the cloud layer for model aggregation, and the formula is: ; in, are all radar data sequences from the sensor layer radar network, is the number of radar data sequences, Indicates the number of edge server devices, Indicates The global model parameters generated after rounds of global iterations.

[0040] In order to accelerate the global parameter calculation, cloud data centers usually have more powerful processing capabilities and abundant computing resources. Therefore, compared with edge devices, the processing delay of cloud data centers is extremely small.

[0041] Then, a distributed learning model is constructed based on the recurrent generation architecture.

[0042] In this embodiment, each edge device deploys the proposed cyclic generation model, which predicts the next radar echo map based on the radar echo sequence. The formula is: ; in, is the number of input radar echograms, is the number of predicted radar echograms, represents the actual radar echo map at time t, represents the predicted radar echogram, The function represents the conditional probability distribution function, Indicates that it will be based on the past The echogram of time steps (from arrive ), predicting the future time step echogram (from arrive ) probability distribution.

[0043] Then a Cyclic Generative Adversarial Network (CG-GAN) model is proposed and deployed on edge devices, which contains two sets of identical generators and discriminators.

[0044] In order to generate reasonable results for subsequent sequences, the role of the generator is to model the distribution of radar echo data after being processed by the sensor layer and learn its performance characteristics.

[0045] The discriminator evaluates the authenticity of the radar echo sequence by combining the potential sequence results generated by the generator with the actual sequence to obtain a scalar that describes the probability that the input radar echo sequence comes from real data or the generator.

[0046] In order to help the cyclic generative adversarial network model to more accurately obtain the potential features of the original data, this embodiment also proposes a long short-term memory network based on cyclic generation, which is embedded in the distributed learning model model to improve the quality of generated data.

[0047] At the edge layer, such as Figure 2As shown in Figure 1, the original radar echo sequence is divided into two subsequences, each with half the time resolution of the original sequence. These subsequences are input into the generator separately. The discriminator obtains the results produced by the generator to determine whether the input data comes from real data or data generated by the generator.

[0048] The structure of the generator is as follows: Figure 3 As shown, each layer has four inputs from the previous time step and the previous layer; The time conversion path of is indicated by black arrows, and The spatiotemporal transformation path is highlighted with 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 state.

[0049] Furthermore, a DCA-LSTM unit is constructed.

[0050] The recurrent unit proposed in this embodiment adopts a stacked architecture, which consists of a forget gate, an input gate, and an output gate, and is used for radar echo extrapolation. These gates give the model the ability to forget past time series features.

[0051] 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 memory states, hidden states, and simple convolutions; however, due to the small number of high-intensity echo areas, the model based on the traditional LSTM unit will miss these areas and have difficulty characterizing their features. This embodiment proposes a DCA-LSTM unit equipped with a novel attention mechanism to enhance the ability to capture high-intensity echo features.

[0052] The structure of the DCA-LSTM unit proposed in this embodiment is as follows: Figure 5 Long-term memory Represents the time step Hidden state is the state of the temporal memory unit, which is used to store long-term information when processing time series data. Stores data processed by the model at previous time steps, containing knowledge and memory of the input sequence that determines how the model behaves. It is the part of the model that stores and understands the spatiotemporal characteristics of the input sequence. It retains the memory state and stores and updates long-term dependencies.

[0053] Specifically, the DCA-LSTM unit converts the time The radar echo map at the time is taken as input, and the hidden state of the previous time step of the same layer is also input . Long-term memory unit Through an initial set of modulation gates , Input Gate and forget gate For update, these gates constitute gating mechanism.

[0054] The formula for this process is: ; ; ; ; in, represents the Sigmoid activation function, , , They represent the spatiotemporal memory input gate at time step t, the spatiotemporal memory modulation gate output at time step t, and the spatiotemporal memory modulation gate output at time step t. The spatiotemporal memory forget gate output, The radar echo map (input data) representing the time step contains spatial dimensions (height, width) and characteristic channels (such as reflectivity, radial velocity), , , Respectively represent input Spatial weight matrix to the input gate, input Spatial weight matrix to the modulation gate, input To the spatial weight matrix of the forget gate, , , Represent the hidden state of the previous time step Time weight matrix to the input gate, hidden state of the previous time step To the time weight matrix of the modulation gate, the hidden state of the previous time step The time weight matrix to the forget gate, , , The bias term of the input gate, the bias term of the modulation gate, and the bias term of the forget gate are respectively; represents the interactive dual attention mechanism, is the number of previous radar echograms, Indicates the past time steps of long-term memory sequence, Indicates the updated interval No. The long-term memory information of the layer, represents matrix product, Represents the convolution process.

[0055] At the same time, time The radar echo map will be related to the time and space unit These combined information are then passed through a second set of modulation gates at the current time step. , Input Gate and forget gate Update to obtain the updated spatiotemporal memory unit .

[0056] The formula for this process is: ; ; ; ; in, , , Respectively represent input To the spatiotemporal memory unit input gate The spatial weight matrix, input To the spatiotemporal memory unit modulation gate The spatial weight matrix, input To the forget gate of the spatiotemporal memory unit The spatial weight matrix, , , Represent the hidden state of the previous time step To the spatiotemporal memory unit input gate The time weight matrix of the previous time step, the hidden state To the spatiotemporal memory unit modulation gate The time weight matrix of the previous time step, the hidden state To the forget gate of the spatiotemporal memory unit The time weight matrix, represents the number of layers in the network, , , Represents the time step The output of the input gate of the spatiotemporal memory unit, the time step The output of the spatiotemporal memory unit modulation gate, the time step The output of the forget gate of the spatiotemporal memory unit, represents the cross attention mechanism, , Represents the time step No. Layer, The spatiotemporal memory unit of the layer.

[0057] Finally, long-term memory information and spatiotemporal memory information The output control gate is obtained by fusion through convolution operation. .

[0058] In addition, long-term memory information and spatiotemporal memory information After splicing along the channel dimension, Convolution operation, and then activation function processing, get the current hidden state .

[0059] The formula for this process is: ; ; in, represents matrix product, represents the convolution process, Represents a tensor concatenation operation.

[0060] In this embodiment, the cross attention unit is as follows: Figure 4 shown.

[0061] Assume that the radar echo map is processed into , first use two The convolutional layer of filters is used to generate the query graph and bond graph .

[0062] Then, by querying the graph and bond graph Perform affinity operation to get the attention map , and at the same time, another Filter convolutional layer to create numerical values ​​for feature adaptation ; Because in and Aggregation operations are performed between them so that key information can be collected.

[0063] Finally, to feed the output back into the unit, the last two dimensions (width and height) are transposed.

[0064] In particular, this embodiment also introduces a loss function with cycle consistency loss, and each edge device is based on the data set Train two generators to learn to generate data distributions and Based on the probability distribution Random noise , the false radar echo sequence is obtained through the generator, denoted as .

[0065] At the same time, design the discriminator and To distinguish from the distribution and The false images are from the distribution The real image.

[0066] Each edge server The objective function of the cyclic generative adversarial network (CG-GAN) on express: ; ; ; in, Respectively represent the discriminator set, including two discriminators and , a generator set, containing two generators and ; They represent the odd-numbered time-step subsequence discriminator and the odd-numbered time-step subsequence generator, respectively. They represent the even time step subsequence discriminator and the even time step subsequence generator respectively. , , They represent 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 respectively; Express expectations, and Respectively represent the discriminator and Will The probability of being judged as a real data sample, and and Respectively represent the discriminator and Will and The probability that the generated data is true.

[0067] Furthermore, this embodiment uses the CIKM-2017 dataset and meteorological indicators to evaluate the proposed model. The radar map in the dataset shows the radar reflectivity of the 101 km × 101 km area adjacent to the central station. The radar echo sequence in the dataset contains 15 radar echo images, and the interval between each radar echo image is 6 minutes.

[0068] 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. Stochastic gradient descent (SGD) is selected as the optimizer, and the batch size is set to 8, which can reduce the risk of crossing the optimal parameters and ensure the convergence speed. The learning rate of the model is set to 0.003, and the loss function is as described above.

[0069] Table 1. Experimental settings

[0070] The evaluation indicators used in this experiment are set as follows: 1. ZR relationship: ZR relationship is often used in radar echo tasks to convert echo readings into precipitation. It describes the relationship between the radar reflectivity factor Z and the precipitation rate R. The formula is: ; in, and depends on the specific climatic conditions and geographical location. Therefore, once the and The radar reflectivity value can be converted into precipitation rate and vice versa by using the value of . The three radar reflectivity thresholds selected in this paper are 10dBZ, 20dBZ and 30dBZ respectively.

[0071] 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. In contrast, the false alarm rate (FAR) score is used to evaluate the situation where the model predicts precipitation events that did not actually occur.

[0072] 3. Critical Success Index (CSI): The critical success index score measures the probability of precipitation prediction provided by the model.

[0073] 4. Structural Similarity Index (SSIM): The structural similarity index is used to quantify the visual similarity between the forecast results and the actual precipitation patterns.

[0074] The calculation formula of the value of the above evaluation index used in this embodiment is: ; ; ; Among them, TP is true positive, which indicates the number of precipitation events correctly predicted by the model; FN is false negative, which indicates the number of precipitation events missed by the model; FP is false positive, which indicates the number of precipitation events misreported by the model.

[0075] The experimental results are as follows: This example comprehensively evaluates the performance of the proposed recurrent 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 areas is only average.

[0076] The detection probability (POD) scores and false alarm rate (FAR) scores of the model proposed in this embodiment and the model used for comparison on different threshold data sets are shown in Table 2.

[0077] Table 2. Comparison of detection probability and false alarm rate under three threshold conditions

[0078] Except for the proposed CG-GAN, the interactive dual attention long short-term memory network (IDA-LSTM) has higher POD scores than other models at 10dBZ, 20dBZ and 30dBZ thresholds.

[0079] The proposed CG-GAN is 3.77% and 4.17% better than IDA-LSTM at 20dBZ and 30dBZ thresholds, respectively. However, the proposed CG-GAN is not the most effective at 10dBZ threshold and is not an ideal choice at lower thresholds.

[0080] In terms of false alarm rate, CG-GAN achieved the lowest score under all selected thresholds, which shows that the proposed model has a significant advantage in prediction accuracy. Compared with the second-ranked score, CG-GAN's scores at 10dBZ, 20dBZ, and 30dBZ thresholds are reduced by 1.1%, 0.87%, and 3.38%, respectively.

[0081] Table 3 below shows the CSI and SSIM scores of all models based on the comparison of critical success index and structural similarity index under three threshold conditions.

[0082] Table 3. CSI and SSIM scores of all models

[0083] It can be seen that the CG-GAN model proposed in the present invention achieves higher scores compared with other comparison models. Specifically, in terms of CSI, at the 10dBZ 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 20dBZ threshold, the score is improved by 5.24% compared with the convolutional long short-term memory network (ConvLSTM); compared with the prediction recurrent neural network upgrade version (PredRNN++), it is improved by 4.46%. In addition, at the 30dBZ threshold, the proposed CG-GAN scores 1.97% and 2.55% higher than GA-ConvGRU and PredRNN++, respectively.

[0084] CG-GAN has the highest SSIM score, which indicates that it has better visual quality and is more similar to the actual observed radar images, thereby improving its precipitation prediction ability.

[0085] The visual display of the radar echo sequence in this embodiment is as follows: Figure 6 The figure shows the extrapolation results of different models for a standard radar echo sequence in the data set. The colored areas in the echo map correspond to specific reflection intensity levels, as shown in the color scale. The sequence shows the process of radar echoes moving from the upper right corner to the center.

[0086] Based on the basic structure of radar echoes, the Convolutional Long Short-Term Memory (ConvLSTM) and 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 Prediction Recurrent Neural Network Upgrade (PredRNN++) improves the ability to extract spatiotemporal feature information by adopting a gating method based on ConvLSTM units and temporal memory units. However, due to the inability to naturally fuse time and feature data, its global spatiotemporal feature prediction ability is very weak. Based on the typical spatiotemporal memory unit, the Interactive Dual Attention Long Short-Term Memory (IDA-LSTM) model introduces an attention mechanism and increases the prediction weight of high-reflectivity areas, so that it can accurately predict high-intensity echo areas while ignoring low-reflectivity value areas.

[0087] 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 thus enhancing the interpretability of the model, an ablation experiment was conducted on CG-GAN.

[0088] Table 4 shows the quantitative evaluation results of these experiments on the components of CG-GAN. The bold numbers indicate the best performance of the model under the current threshold. The spatiotemporal long short-term memory network (ST-LSTM) represents the baseline model without additional modules. CG-GAN (without DCC) represents the extrapolated 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 extrapolated network without the cycle consistency loss function.

[0089] Table 4. Quantitative evaluation results of ablation experiments

[0090] Removing any one component from CG-GAN results in a consistent drop in the model's performance on all evaluation metrics. Specifically, when relying solely on the ST-LSTM model to extract spatiotemporal features from radar echoes, the model's critical success index (CSI) score decreases by 5.39%, 2.96%, and 6.65% at 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 10dBZ, 20dBZ, and 30dBZ thresholds, and the POD scores decrease by 1.99%, 1.66%, and 1.77%, respectively.

[0091] In addition, after removing the cycle consistency loss function from the extrapolated model, the CSI scores are reduced by 2.18%, 1.89% and 2.87% at 10dBZ, 20dBZ and 30dBZ thresholds, and the POD scores are reduced by 0.75%, 3.03% and 0.89%, respectively.

[0092] In an embodiment of the present invention, an electronic device is also 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 cycle generation countermeasure prediction method based on edge-cloud collaboration described in the above embodiment.

[0093] In an embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored. When the program is executed by a processor, the steps in the radar echo extrapolation cycle generation countermeasure prediction method based on edge-cloud collaboration in the above embodiment are implemented.

[0094] 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.

[0095] 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 cycle generation adversarial prediction method based on edge cloud collaboration, characterized in that: The steps include: Step 1: Build an edge-cloud collaborative layered architecture for the radar echo extrapolation task, including: cloud layer, edge layer, and sensor layer. The sensor layer collects raw radar data to extract radar echo sequences, and the edge layer performs local data processing and transmission. The federated learning architecture utilizes the computing power of the cloud service center in the cloud layer. Step 2: Build a recurrent generative adversarial network at the edge layer. Use a recurrent generative architecture to deploy a recurrent 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 to predict the subsequent radar echo sequence. Step 3: Construct a double cross-attention long short-term memory unit and adopt a stacked architecture to form a recurrent unit based on the long short-term memory network to perform radar echo extrapolation. By mining the potential connections between radar echo data at different time points, the performance of the edge-cloud collaborative hierarchical architecture in generating radar echo predictions is improved.

2. The radar echo extrapolation cycle generation countermeasure prediction method based on edge cloud collaboration according to claim 1 is characterized in that: In the edge-cloud collaborative layered architecture, the cloud layer, edge layer, and sensor layer are interconnected via the Internet; The sensor layer includes multiple radar networks, each of which is composed of multiple local radars, which collect raw radar data in real time, extract radar echo sequences, and transmit them to the nearest edge server; the radar echo sequence is composed of multiple radar echo images captured in continuous time intervals, which are used to describe the weather conditions in a specific area over a period of time; The edge layer includes multiple edge servers, which receive radar echo sequence data, and as local clients, calculate and transmit local data through a cyclic generative adversarial network, update local parameters and transmit them to the cloud service center of the cloud layer; The cloud service center includes storage devices, network devices, and homogeneous or heterogeneous computing devices; after receiving the local parameters of the edge layer, the cloud service center performs global aggregation through the global model, refreshes the global model using the 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 edge-side local model, calculates the local loss gradient, and updates the local parameters.

3. The radar echo extrapolation cycle generation countermeasure prediction method based on edge cloud collaboration according to claim 2 is characterized in that: The sensor layer collects raw radar data through multiple local radars on multiple meteorological data platforms at the same time, and the local radar raw data upload delay , expressed as: ； in, It's a radar network The number of radar echograms in It's a radar network bandwidth, is the signal-to-noise ratio, is the number of local radars; Different radar network processing delays , expressed as: ； in, It's a radar network Data transfer speed; Data transmission energy consumption at the sensor layer , expressed as: ； in, is the transmission power of the device, is the power for processing radar data, , They represent data upload time and data processing time respectively. , They represent data uploading energy and data processing energy respectively.

4. The radar echo extrapolation cycle generation countermeasure prediction method based on edge cloud collaboration according to claim 2 is characterized in that: In the distributed learning model, the edge server is deployed Radar Network The loss function , expressed as: ； in, are the parameters of the distributed model, Represents edge server From the corresponding local radar network The collected radar echo sequence, is the number of radar echo sequences, Represents each radar echo sequence The loss function on ; In global iteration In the distributed model on the edge server, the gradient descent method is used to calculate the global parameters of the previous iteration. Update local parameters , the formula is as follows: ； in, is the learning rate, is the sign of the gradient, Indicated in During the global iteration, the edge server Corresponding to the local radar network The loss value calculated by the model based on local data.

5. The radar echo extrapolation cycle generation countermeasure prediction method based on edge cloud collaboration according to claim 4 is characterized in that: Perform local model training and local parameter update in each edge server, including gradient calculation, local model optimization and parameter update, as follows: The number of floating-point operations performed by the distributed learning model is used to quantify the usage of computing resources on the edge server and the computational latency of the edge server in one global iteration. , expressed as: ； in, Represents edge server The number of floating point operations in the distributed model, Indicates the number of floating-point operations per GPU cycle, Represents edge server For radar networks Allocated computing resources; Parameter upload delay when edge server uploads parameters to cloud server , expressed as: ； in, represents the size of the local model parameters, Indicates from the radar network The bandwidth allocated to the training task; Determine the total energy consumption based on the edge server’s computation latency and parameter upload latency , the calculation formula is: ； in, represents the transmission power of the edge server, represents the computing power of the edge server, , They represent the energy consumption of edge layer data uploading and edge layer data calculation respectively.

6. The radar echo extrapolation cycle generation countermeasure prediction method based on edge cloud collaboration according to claim 4 is characterized in that: The cloud service center receives local parameters uploaded by each edge server and performs model aggregation through the global model: ； in, are all radar data sequences from the sensor layer radar network, is the number of radar data sequences, Indicates the number of edge server devices, Indicates The global model parameters generated after rounds of global iterations.

7. The radar echo extrapolation cycle generation countermeasure prediction method based on edge cloud collaboration according to claim 4 is characterized in that: In step 2, the recurrent generation model deployed on each edge server contains two identical sets of generators and discriminators; In the edge layer, the input radar echo sequence is divided 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 and includes several layers. The generated potential sequence results depend on the information of the previous time step and the previous layer, as well as the current input features and hidden state; The discriminator is constructed based on a cyclically generated long short-term memory network, obtains the potential sequence results generated by the generator, and combines them with the actual sequence to obtain a scalar for describing 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 generation model predicts the next radar echo map based on the received radar echo sequence. The formula is: ； in, is the number of input radar echograms, is the number of predicted radar echograms, represents the actual radar echo map, represents the predicted radar echogram, The function represents the conditional probability distribution function, Indicates that it will be based on the past Echograph of time steps, predicting the future The probability distribution of the echogram for each time step.

8. The radar echo extrapolation cycle generation countermeasure prediction method based on edge cloud collaboration according to claim 7 is characterized in that: In the distributed learning model, each edge server is based on the data set Train two generators to learn to generate data distribution and , the method is as follows: Step 2.1: Based on the probability distribution Random noise , the false radar echo sequence is obtained through the generator ; Step 2.2: Build a discriminator and , used to distinguish from data distribution and The false images are from the distribution The real image of Step 2.3: On each edge server The objective function of the recurrent generative adversarial network is expressed as the value function : ; ; ; in, , They represent the discriminator set and the generator set respectively. Respectively represent the odd time step subsequence discriminator and generator, Respectively represent the even time step subsequence discriminator and generator, , , They represent 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 respectively; Express expectations, and Respectively represent the discriminator and Will The probability of being judged as a real data sample, and Respectively represent the discriminator and Will and The probability that the generated data is true.

9. The radar echo extrapolation cycle generation countermeasure prediction method based on edge cloud collaboration according to claim 1 is characterized in that: The dual cross-attention long short-term memory unit is composed of a forget gate, a modulation gate, an input gate and an output gate, and extracts high-intensity radar echoes through memory states, hidden states and convolutions, and the method is as follows; Step 3.1, initial double cross attention long short-term memory unit, input data is the current time step The radar echo map and the previous time step of the same layer The hidden state , through the first set of modulation gates , Input Gate and the forget gate Update to obtain updated long-term memory information , the update process is expressed as: ； ； ； ； in, Represents the Sigmoid activation function; , , Represents the time step The spatiotemporal memory input gate, spatiotemporal memory modulation gate output, and spatiotemporal memory forgetting gate output; Represents the time step The radar echo map contains the height, width spatial dimensions and characteristic channels. , , Respectively represent input data To the spatial weight matrix of the input gate, modulation gate, and forget gate, , , Represent the hidden state of the previous time step To the time weight matrix of the input gate, modulation gate, and forget gate, , , The bias term of the input gate, the bias term of the modulation gate, and the bias term of the forget gate are respectively; represents the interactive dual attention mechanism, is the number of previous radar echograms, Indicates the past time steps of long-term memory sequence, Indicates the updated interval No. The long-term memory information of the layer, represents matrix product, Represents the convolution process; Step 3.2: Set the time step Radar echogram and space-time unit Combined processing, passing through the second set of modulation gates at the current time step , Input Gate and the forget gate Update to obtain updated spatiotemporal memory information , the update process is expressed as: ； ； ； ； in, , , Represents the time step The output of the spatiotemporal memory unit input gate, modulation gate, and forget gate, , , Respectively represent input data To the spatiotemporal memory unit input gate , Modulation Gate , Forget Gate The spatial weight matrix of , , Represent the hidden state of the previous time step To the spatiotemporal memory unit input gate , Modulation Gate , Forget Gate The time weight matrix, represents the number of layers in the network, represents the cross attention mechanism, , Represents the time step No. Layer, Layers of spatiotemporal memory units; Step 3.3: Store information in long-term memory and spatiotemporal memory information Through convolution operation, the output control gate is obtained. , the formula is: ； in, , , , Respectively represent the current input data , the hidden state at the previous time step , long-term memory unit , space-time memory unit To output gate The weight matrix of represents the bias term of the output gate, Represents the time step The output of the output gate; Step 3.4: Store information in long-term memory and spatiotemporal memory information Splicing along the channel dimension, After the convolution operation, the activation function is processed to obtain the current hidden state , the formula is: ； in, express Convolution operation, Represents a tensor concatenation operation.

10. The radar echo extrapolation cycle generation countermeasure prediction method based on edge cloud collaboration according to claim 9 is characterized in that: The cross attention mechanism is implemented by the cross attention unit as follows: Step 3.2.1: For the radar echo map generated by the generator, use two The convolutional layer of filters generates the query graph and bond graph ; Step 3.2.2: Query graph and bond graph Perform affinity operation to obtain the attention map ; Step 3.2.3: Apply another Filter convolution layer, creating numerical values ​​for feature adaptation ,exist and Perform aggregation operations between them to collect key information; Step 3.2.4, transpose the width and height, and feed the output back to the cross attention unit.

Images