A data processing method and system of a virtual power plant edge computing gateway

Abstract

The application discloses a kind of data processing method and system of virtual power plant edge computing gateway, it is related to the field of industrial internet, the present application can effectively utilize mass multi-modal sample data and position sample data, train model, obtain task scheduling instruction model, the trained task scheduling instruction model is deployed in edge gateway node, and edge gateway node can be quickly identified task scheduling instruction according to model prediction and analysis, and the identification efficiency and accuracy of different task scheduling instructions in heterogeneous network are higher, beneficial to improve edge collaborative effect, to decompose the task scheduling instruction transmitted by cloud computing platform / network node into the task scheduling instruction to terminal equipment in subsequent process, accurately issue to terminal equipment for equipment adjustment.

Description

Technical Field

[0001] This invention relates to the technical field of industrial internet, and in particular to a data processing method and system for a virtual power plant edge computing gateway. Background Technology

[0002] Edge computing is a computing model that uses network edge detection, physically close to the data source, to integrate core capabilities of networking, computing, storage, and applications into an open platform, providing edge intelligent services locally. With the continuous improvement of system informatization and intelligence, edge computing has become increasingly important. On the one hand, it has advantages in big data processing and analysis; on the other hand, edge computing has irreplaceable advantages in low-latency and high-reliability scenarios; furthermore, edge computing can provide elastically scalable cloud service capabilities, featuring fast response, low latency, and lightweight computing. Edge computing and systems can be widely applied in various application scenarios such as content delivery networks, e-commerce, gaming, audio and video, IoT, logistics, industrial AI, and city AI, and are widely used by cloud vendors, telecom operators, software / industry solution providers, system integrators, manufacturers, and infrastructure builders. Edge computing plays a positive and effective role in various industries.

[0003] The management and control of edge gateway nodes in edge computing are key to improving data processing efficiency. In addition to computing, storage, and communication functions, they also have functions such as task management, data management, and data analysis. In complex scenarios with various factors such as different application scenarios, different access terminals, different access numbers, different times, different links, and different access requests, how to process data from heterogeneous networks and accurately distribute it to the corresponding network terminal devices to ensure the efficiency of edge computing services is an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a data processing method and system for a virtual power plant edge computing gateway, which solves the technical problems of low efficiency and accuracy of edge gateway nodes in recognizing different task scheduling instructions in heterogeneous networks, poor edge collaboration effect, and impact on service performance.

[0005] The first aspect of this invention provides a data processing method for a virtual power plant edge computing gateway, comprising:

[0006] Obtain multiple sample sets; the sample sets include location sample data and multimodal sample data, wherein the location sample data is the location of effective information of heterogeneous networks in different scenarios in the transmitted data, and the multimodal sample data is multimodal data in different formats;

[0007] Based on the task training model, the task training model is pre-trained using the multimodal sample data of the sample set;

[0008] The pre-trained task training model is perceptually trained based on the location sample data of the heterogeneous network in the sample set.

[0009] The task training model trained after perception is used as the first model, and the first multimodal sample data is obtained from the sample set;

[0010] The first model is trained based on the first multimodal sample data, and the prediction result is output.

[0011] Perform parameter index analysis on the current output prediction results and determine whether the analyzed parameter indexes have converged;

[0012] If convergence is achieved, the current output prediction result is used as the pseudo-label result of the first multimodal sample data to obtain pseudo-label sample data and supplement it to the sample set.

[0013] If convergence fails, the first model is updated based on the current output prediction result, and the process of training the updated first model using the first multimodal sample data is returned.

[0014] Based on the sample set and the trained task training model, transfer learning is performed to form a task scheduling instruction model.

[0015] Optionally, the step of using the current output prediction result as the pseudo-label result of the first multimodal sample data if convergence is achieved, obtaining pseudo-label sample data and supplementing it to the sample set, includes:

[0016] The current output prediction result is used as the pseudo-label result of the first multimodal sample data, and the pseudo-label result is filtered according to a preset rule;

[0017] The filtered multimodal sample data and the pseudo-label results are identified as pseudo-label sample data and added to the sample set.

[0018] Optionally, the step of filtering the pseudo-label results according to preset rules includes:

[0019] Verify the consistency of task labeling results from different heterogeneous networks in the pseudo-label results;

[0020] Filter pseudo-label results based on consistency results.

[0021] Optionally, before the step of using the perception-trained task training model as the first model to obtain the first multimodal sample data from the sample set, the method further includes:

[0022] The first and second models are initialized based on the task training model after perception training, and the second multimodal sample data is obtained from the sample set;

[0023] The second multimodal sample data is perturbed using two different perturbation methods to form the first training data and the second training data, respectively.

[0024] Input the first training data into the first model and output the first prediction result;

[0025] The second training data is input into the second model, and the second prediction result is output.

[0026] Calculate the loss between the first prediction result and the second prediction result, and use it as the pseudo-label loss;

[0027] Based on the features output from the first model and the second model, feature consistency calculation is performed to obtain feature consistency loss;

[0028] The gradient is determined based on the pseudo-label loss and the feature consistency loss;

[0029] The parameters of the second model are updated based on the gradient and the pseudo-label loss.

[0030] The updated parameters of the second model guide the updating of the parameters of the first model.

[0031] Optionally, the task training model includes a preprocessing network, a main network, and a multi-task output network.

[0032] A second aspect of the present invention provides a data processing system for a virtual power plant edge computing gateway, comprising:

[0033] The sample set module is used to acquire multiple sample sets; the sample sets include location sample data and multimodal sample data, wherein the location sample data is the location of effective information of heterogeneous networks in different scenarios in the transmitted data, and the multimodal sample data is multimodal data in different formats;

[0034] The task training model module is used to pre-train the task training model based on the task training model using the multimodal sample data of the sample set.

[0035] The perception training module is used to perform perception training on the pre-trained task training model based on the location sample data of the heterogeneous network set in the sample set.

[0036] The sample data module is used to obtain first multimodal sample data from the sample set by using the task training model after perception training as the first model;

[0037] The prediction result module is used to train the first model based on the first multimodal sample data and output the prediction result;

[0038] The convergence judgment module is used to analyze the parameter indicators of the current output prediction results and determine whether the analyzed parameter indicators have converged.

[0039] The first data processing module is used to, if convergence is achieved, take the current output prediction result as the pseudo-label result of the first multimodal sample data, obtain pseudo-label sample data, and supplement it to the sample set.

[0040] The second data processing module is used to update the first model based on the current output prediction result if the model does not converge, and then return to execute the training process of the updated first model using the first multimodal sample data.

[0041] The task scheduling instruction model module is used to perform transfer learning based on the sample set and the trained task training model to form a task scheduling instruction model.

[0042] Optionally, the first data processing module includes:

[0043] The filtering submodule is used to use the current output prediction result as the pseudo-label result of the first multimodal sample data, and to filter the pseudo-label result according to a preset rule;

[0044] The supplementary submodule is used to determine the filtered multimodal sample data and the pseudo-label results as pseudo-label sample data and supplement them into the sample set.

[0045] Optionally, the filtering submodule includes:

[0046] The verification unit is used to verify the consistency of task labeling results from different heterogeneous networks in the pseudo-label results;

[0047] The pseudo-label result unit is used to filter pseudo-label results based on consistency results.

[0048] Optionally, it also includes:

[0049] An initialization module is used to initialize the first model and the second model based on the task training model after perception training, and to obtain the second multimodal sample data from the sample set.

[0050] The perturbation processing module is used to perturb the second multimodal sample data using two different perturbation methods to form the first training data and the second training data respectively.

[0051] The first prediction result module is used to input the first training data into the first model and output the first prediction result.

[0052] The second prediction result module is used to input the second training data into the second model and output the second prediction result;

[0053] The pseudo-label loss module is used to calculate the loss between the first prediction result and the second prediction result, which is the pseudo-label loss.

[0054] The feature consistency loss module is used to perform feature consistency calculation based on the features output from the first model and the second model to obtain the feature consistency loss.

[0055] The gradient module is used to determine the gradient based on the pseudo-label loss and the feature consistency loss;

[0056] The first update module is used to update the parameters of the second model based on the gradient and the pseudo-label loss.

[0057] The second update module is used to guide the updating of the parameters of the first model based on the updated parameters of the second model.

[0058] Optionally, the task training model includes a preprocessing network, a main network, and a multi-task output network.

[0059] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the data processing method of the virtual power plant edge computing gateway as described in any of the preceding claims.

[0060] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the data processing method of the virtual power plant edge computing gateway as described in any of the preceding claims.

[0061] As can be seen from the above technical solutions, the present invention has the following advantages:

[0062] Obtain multiple sample sets; the sample sets include location sample data and multimodal sample data. Based on the task training model, the task training model is pre-trained using the multimodal sample data from the sample sets. The pre-trained task training model is then perceptually trained using location sample data from a heterogeneous network defined in the sample sets. This perceptually trained task training model is used as the first model. First multimodal sample data is obtained from the sample sets, and the first model is trained based on this first multimodal sample data. Prediction results are output, and parameter index analysis is performed on the current output prediction results. The convergence of the analyzed parameter indexes is then determined. If convergence is achieved... If convergence occurs, the current output prediction result is used as the pseudo-label result of the first multimodal sample data to obtain pseudo-label sample data and supplement it to the sample set. If convergence does not occur, the first model is updated according to the current output prediction result, and the process is returned to execute the training of the updated first model using the first multimodal sample data. Based on the sample set and the trained task training model, transfer learning is performed to form a task scheduling instruction model. This solves the technical problem of low efficiency and accuracy of edge gateway nodes in recognizing different task scheduling instructions in heterogeneous networks, poor edge collaboration effect, and impact on service performance in edge computing. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart illustrating the steps of a data processing method for a virtual power plant edge computing gateway provided in Embodiment 1 of the present invention.

[0065] Figure 2 This is a structural block diagram of a data processing system for a virtual power plant edge computing gateway provided in Embodiment 3 of the present invention. Detailed Implementation

[0066] This invention provides a data processing method and system for a virtual power plant edge computing gateway, which addresses the technical problems of low efficiency and accuracy in identifying different task scheduling instructions in heterogeneous networks, poor edge collaboration, and impact on service performance in edge computing.

[0067] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0068] Please see Figure 1 , Figure 1 This is a flowchart illustrating the steps of a data processing method for a virtual power plant edge computing gateway provided in Embodiment 1 of the present invention.

[0069] This invention provides a data processing method for a virtual power plant edge computing gateway, comprising:

[0070] Step 101: Obtain multiple sample sets; the sample sets include location sample data and multimodal sample data. The location sample data represents the location of effective information of heterogeneous networks in different scenarios within the transmitted data, while the multimodal sample data consists of multimodal data in different formats.

[0071] In this invention embodiment, the heterogeneous network in different scenarios is a network composed of terminal devices (terminal nodes), gateway devices (edge ​​gateway nodes), network devices (network nodes), and cloud computing platforms from different manufacturers. In most cases, the devices in the heterogeneous network run on different protocols and support different functions or applications. The sample set involved in this invention includes location sample data and multimodal sample data. Location sample data refers to the initial data of task scheduling instructions from different network nodes collected by the edge gateway node in different heterogeneous networks and different scenarios, labeled with the results to be analyzed, which is labeled sample data. For example, location sample data can include the location of effective information in the transmitted data in scenarios such as games / e-commerce / IoT in heterogeneous networks. An initial data set can include one or more location sample data. Multimodal sample data can include multimodal data in formats such as text, images, videos, or audio, which is unlabeled sample data. Accurate location sample data requires a lot of time to label. Gateway devices (edge ​​gateway nodes) mostly receive multimodal sample data. This invention can effectively utilize massive amounts of multimodal sample data and location sample data to train the model. Based on model prediction and analysis, the task scheduling instructions transmitted by the cloud computing platform / network nodes are decomposed into task scheduling instructions for terminal devices, which are then sent to the terminal devices for device adjustment.

[0072] Step 102: Based on the task training model, pre-train the task training model using multimodal sample data from the sample set.

[0073] In this embodiment of the invention, the invention relates to task training models and task scheduling models. The task training model is a perceptual model of a large model, meaning it has a model structure with a large number of parameters, typically on the order of tens or hundreds of millions. Regarding the topology of the large model, topologies such as UFO (Unified Feature Optimization) or DWNet (Data Warehouse Network) are used. Using a perceptual model of a large model as the task training model can fully leverage the powerful feature representation capabilities of the large model and improve the prediction accuracy of the task training model.

[0074] Since the sample set includes a large amount of multimodal sample data, in order to effectively utilize this data and improve training efficiency, it is advisable to pre-train the task training model using multimodal sample data. This enhances the feature extraction capability of the task training model. Pre-training of the task training model can primarily be used to train the feature extraction network, and self-supervised learning can be achieved through two stages: 1) intra-modal self-supervised learning; 2) cross-modal inter-modal self-supervised comparative learning. The methods for implementing self-supervised learning are existing technologies.

[0075] Step 103: Perform perceptual training on the pre-trained task training model based on the position sample data of the heterogeneous network in the sample set.

[0076] In this embodiment of the invention, perceptual training is performed on task training models, whether pre-trained or not, based on location sample data from a defined heterogeneous network. Optionally, the labeled results of the location sample data from the defined heterogeneous network correspond to the scheduling prediction results of at least one sub-task. The defined heterogeneous network can be a heterogeneous network in scenarios such as gaming, e-commerce, or the Internet of Things. Therefore, different heterogeneous networks may have different labeled results, and the labeled results correspond to the scheduling prediction results of at least one sub-task. By utilizing various location sample data with labeled results, perceptual training can be performed on the task training model, improving the model's accuracy.

[0077] Step 104: Use the task training model trained after perception training as the first model, and obtain the first multimodal sample data from the sample set.

[0078] In this embodiment of the invention, the task training model can be trained using a self-iterative (Self-tr first multimodal sample data ining) approach. A new version of the task training model is trained using a previously trained task training model (which can be referred to as the first model), location sample data, and multimodal sample data as input. The initial trained task training model used for self-iterative training can be a task training model that has undergone perceptual training; subsequent trained models can be task training models that have been updated again after iterative training. During one self-iterative training process, this task training model can be referred to as the first model or the teacher model.

[0079] Step 105: Train the first model based on the first multimodal sample data and output the prediction results.

[0080] In this embodiment of the invention, an appropriate number of multimodal sample data are obtained from the sample set as the first multimodal sample data for this round of training. The first multimodal sample data is input into the first model, and a prediction result is output. The parameter indicators of the prediction result are calculated. The first output prediction result is recorded, and then the parameters of the first model are updated based on the prediction result. The first multimodal sample data is input into the updated first model again, and a prediction result is output.

[0081] Step 106: Perform parameter index analysis on the current output prediction results and determine whether the analyzed parameter indexes have converged.

[0082] In this embodiment of the invention, step 107 or step 108 is executed depending on the convergence status.

[0083] Step 107: If convergence is achieved, the current output prediction result is used as the pseudo-label result of the first multimodal sample data to obtain pseudo-label sample data and supplement it to the sample set.

[0084] In this embodiment of the invention, the prediction results output by the first model based on the first multimodal sample data are subject to normalization processing by the multi-task output network and processing based on classification probability values, resulting in changes in the prediction results for each calculation process. This change retains value for updating and training the first model. Therefore, a self-iterative training method using the first multimodal sample data is adopted. In step 103, the task training model is trained using location sample data, serving as the first model. Multimodal sample data is then continuously input into the first model for training. When the prediction results of the first model meet the convergence requirements of the parameter indices, they are used as pseudo-label results for the multimodal sample data, and pseudo-label sample data is added to the sample set. After mixing the pseudo-label sample data with the location sample data, the task training model can be trained. After training, it is updated as the first model to begin a new round of training. If the analyzed parameter indices show convergence, the self-iterative training using the first multimodal sample data can be stopped.

[0085] Step 108: If convergence is not achieved, update the first model based on the current output prediction results, and return to the step of training the updated first model using the first multimodal sample data.

[0086] In this embodiment of the invention, if the parameter index has not yet converged, the first multimodal sample data in this round still has training value, so the process returns to continue updating the first model and continues to use the first multimodal sample data for training.

[0087] It is worth mentioning that in steps 106-108, the parameter metrics can be performance metrics used to evaluate the prediction results on the instruction sample sets of each scheduling task. Specific metrics can be set according to requirements. For example, parameter metrics can include, but are not limited to, detection mean, average precision, precision, recall, segmentation mean, intersection, union ratio, etc. Parameter metrics can also be called performance metrics. Based on the current output prediction results and the recorded historical prediction results, the parameter metrics can be calculated, and it can be determined whether the parameter metrics have converged, for example, whether the parameter metrics have stopped increasing. If the parameter metrics converge, it indicates that the training of the task training model using the first multimodal sample data in this round has been completed. The current output prediction results can then be used as pseudo-label results for the first multimodal sample data, generating new pseudo-label sample data. This pseudo-label sample data can be added to the sample set as a sample set for subsequent iterative training or other training. If the parameter metrics have not yet converged, the first multimodal sample data in this round still has training value. Therefore, the process returns to continue updating the first model and continuing to use the first multimodal sample data for training.

[0088] Step 109: Based on the sample set and the trained task training model, perform transfer learning to form a task scheduling instruction model.

[0089] In this embodiment of the invention, when performing transfer learning on the task scheduling instruction model, one can choose a set of sample sets applicable before the update, or a set of sample sets after the update, or a subset can be selected from the sample sets for transfer learning. When the task training model reaches the conditions for completion of training, transfer learning can be performed as needed to generate a task scheduling instruction model with a small-scale model structure as the main component. The task scheduling instruction model is a perceptual model with a small-scale model structure, and the number of model parameters of the task scheduling instruction model is smaller than the number of parameters of the perceptual model, which is a large model.

[0090] In this invention, multiple sample sets are obtained; these sample sets include location sample data and multimodal sample data. Based on a task training model, the task training model is pre-trained using the multimodal sample data from the sample sets. The pre-trained task training model is then perceptually trained using location sample data from a heterogeneous network defined in the sample sets. This perceptually trained task training model is used as a first model. First multimodal sample data is obtained from the sample sets, and the first model is trained based on this data. Prediction results are output, and parameter index analysis is performed on the current output prediction results. Finally, it is determined whether the analyzed parameter indexes have converged. If convergence occurs, the current output prediction result is used as the pseudo-label result of the first multimodal sample data to obtain pseudo-label sample data and supplement it to the sample set. If convergence does not occur, the first model is updated according to the current output prediction result, and the process returns to execute the training of the updated first model using the first multimodal sample data. Based on the sample set and the trained task training model, transfer learning is performed to form a task scheduling instruction model. This solves the technical problem of low efficiency and accuracy of edge gateway nodes in recognizing different task scheduling instructions in heterogeneous networks, poor edge collaboration effect, and impact on service performance in edge computing.

[0091] Example 2: A data processing method for a virtual power plant edge computing gateway provided by the present invention includes:

[0092] Step 201: Obtain multiple sample sets; the sample sets include location sample data and multimodal sample data. The location sample data represents the location of effective information of heterogeneous networks in different scenarios within the transmitted data, while the multimodal sample data consists of multimodal data in different formats.

[0093] In this embodiment of the invention, the specific implementation process of step 201 is similar to that of step 101, and will not be repeated here.

[0094] Step 202: Based on the task training model, pre-train the task training model using multimodal sample data from the sample set.

[0095] Furthermore, the task training model includes a preprocessing network, a main network, and a multi-task output network.

[0096] In this embodiment of the invention, the input data is preprocessed through a preprocessing network to adapt to the input requirements of the main network. The main network can jointly process single-modal or multi-modal data to extract features. The multi-task output network may include multiple sub-task branch modules, which can be selectively used to adapt to various task prediction requirements of heterogeneous networks. In this invention, multiple versions of task training models can be pre-trained to form a task training model library. In the transfer learning and data annotation stages, the advantages of multiple versions of task training models can be effectively utilized. The number of versions of the task training model can be multiple, and each version of the task training model includes at least one of the following differences: the topology of the preprocessing network, the topology of the main network, the training hyperparameters of the task training model, and the sample set used by the task training model.

[0097] To create multiple versions of the task training model, different preprocessing network topologies and / or main network topologies can be preset and trained separately, resulting in different versions of the task training model. Additionally, different training hyperparameters can be set, such as the number of iterations, loss target, or parameter update gradients, to obtain different versions of the task training model. Alternatively, different subsets of the entire sample set can be selected to train the task training model, resulting in different versions. In addition to including various versions of the task training model, the task training model library can further collect other related models, such as task training models for processing single-modal data, and other trained small-scale models, thus playing a supporting role in semi-supervised training, semi-automated data labeling, and transfer learning.

[0098] Step 203: Perform perceptual training on the pre-trained task training model based on the position sample data of the heterogeneous network in the sample set.

[0099] In this embodiment of the invention, the specific implementation process of step 203 is similar to that of step 103, and will not be repeated here.

[0100] Furthermore, the following steps are included before step 204:

[0101] 11. Initialize the first and second models based on the task training model after perception training, and obtain the second multimodal sample data from the sample set.

[0102] The first multimodal sample data 12, and the second multimodal sample data are perturbed using two different perturbation methods to form the first training data and the second training data respectively.

[0103] 13. Input the first training data into the first model and output the first prediction result.

[0104] 14. Input the second training data into the second model and output the second prediction result.

[0105] 15. Calculate the loss between the first and second prediction results for the first multimodal sample data, as the pseudo-label loss.

[0106] 16. Based on the features output from the first and second models, perform feature consistency calculation to obtain feature consistency loss.

[0107] 17. Based on the pseudo-label loss and feature consistency loss, determine the gradient of the first multimodal sample data.

[0108] 18. Update the parameters of the second model based on the gradient and pseudo-label loss of the first multimodal sample data.

[0109] 19. The parameters of the first model are updated based on the updated parameters of the second model.

[0110] In this embodiment of the invention, firstly, the first model and the second model are initialized using a task training model trained on perception. The first model can be a teacher model, and the second model can be a student model. The parameters of the second model (student model) in this invention can guide the parameter updates of the first model (teacher model), further reducing the fluctuation of the first model (teacher model) parameters. Then, the input second multimodal sample data undergoes two different perturbation processes, such as different random perturbations. The perturbed data is then input into the first model and the second model respectively, and prediction results are output. Finally, the first prediction result output by the first model is used as pseudo-labeled data for supervised training of the second model. Simultaneously, pseudo-labeled data can also be added to the sample set. The feature consistency loss between the first model and the second model is calculated. The features of the first model and the second model can be features output from any stage of the model, such as features extracted or transformed in the main network. The gradient is calculated using the backpropagation method employed in neural network training, specifically by backpropagating the gradient based on the constructed pseudo-label loss and the feature consistency loss. The gradient can be used to control the update gradient of the parameters in the second model. In this invention, the parameters of the second model can guide the parameter updates of the first model, further reducing the fluctuation of the first model's parameters. During training, the second model may update more frequently, leading to larger fluctuations. Therefore, the parameters of the second model can be used to guide the parameter updates of the first model, ensuring that the overall trend of the first model's parameter changes is consistent with that of the second model. An adjustment strategy can be used to reduce the fluctuation of the first model's parameters. This adjustment strategy could be, for example, using a moving average, where the moving average of the second model's parameters over training epochs is used as the parameters of the first model, thus reducing the fluctuation of the first model's parameters. The first and second models are updated together until the model converges using the second multimodal sample data. Either the first or second model can be used as the perceptual model after semi-supervised training. This invention enables iterative training of the task training model using a semi-supervised training method, fully utilizing the advantage of a large amount of second multimodal sample data to optimize the task training model.

[0111] It is worth mentioning that the "first" and "second" in the first and second multimodal sample data are mainly for differentiation. The first and second multimodal sample data can be the same or selected based on certain rules. Both the first and second multimodal sample data can be multimodal data including formats such as text, images, videos or audio, and are unlabeled sample data.

[0112] The first multimodal sample data mainly consists of unlabeled sample data used during the training process of the task training model, that is, unlabeled sample data used during training in the self-iterative method; the second multimodal sample data mainly consists of unlabeled sample data used during the initialization of the task training model in the semi-supervised training method.

[0113] Step 204: Use the task training model trained after perception training as the first model, and obtain the first multimodal sample data from the sample set.

[0114] In this embodiment of the invention, the specific implementation process of step 204 is similar to that of step 104, and will not be repeated here.

[0115] Step 205: Train the first model based on the first multimodal sample data and output the prediction results.

[0116] In this embodiment of the invention, the specific implementation process of step 205 is similar to that of step 105, and will not be repeated here.

[0117] Step 206: Perform parameter index analysis on the current output prediction results and determine whether the analyzed parameter indexes have converged.

[0118] In this embodiment of the invention, the specific implementation process of step 206 is similar to that of step 106, and will not be repeated here.

[0119] Step 207: If convergence is achieved, the current output prediction result is used as the pseudo-label result of the first multimodal sample data to obtain pseudo-label sample data and supplement it to the sample set.

[0120] Furthermore, step 207 may also include the following sub-steps:

[0121] S11. Use the current output prediction result as the pseudo-label result of the first multimodal sample data, and filter the pseudo-label result according to the preset rules.

[0122] S12. The filtered multimodal sample data and pseudo-label results are identified as pseudo-label sample data and added to the sample set.

[0123] In this embodiment of the invention, a task training model is trained using location sample data as the first model. Multimodal sample data is then continuously input into the first model for training. When the prediction result of the first model meets the convergence requirements of the parameter indicators, it is used as the pseudo-label result of the multimodal sample data, and pseudo-labeled sample data is added to the sample set. After mixing the pseudo-labeled sample data with the location sample data, the task training model can be trained. After training is completed, it is updated as the first model to start a new round of training. When performing pseudo-labeling using the current task training model that meets the parameter indicator requirements, the pseudo-labeled sample data can be further refined and filtered.

[0124] Furthermore, S11 may also include the following sub-steps:

[0125] S111. Verify the consistency of task labeling results from different heterogeneous networks in the pseudo-label results.

[0126] S112. Filter pseudo-label results based on consistency results.

[0127] In this embodiment of the invention, the filtering rules (preset rules) can be implemented based on the consistency of the pseudo-label results. Consistency can be reflected in the consistency between the task labeling results output by the task training model from different heterogeneous networks. For example, the multi-task output network of the task training model can process task scheduling instructions from the same network node in different heterogeneous networks. The location of valid information in the same task scheduling instruction issued by the same network node is theoretically consistent. Therefore, the gateway device (edge ​​gateway node) determines whether the output prediction result is relatively accurate by receiving the consistency of the same task scheduling instruction from the same network node in different heterogeneous networks, and filters out the prediction results with higher consistency matching as pseudo-label results.

[0128] Step 208: If convergence is not achieved, update the first model based on the current output prediction results, and return to the step of training the updated first model using the first multimodal sample data.

[0129] In this embodiment of the invention, if the parameter index has not yet converged, the first multimodal sample data in this round still has training value, so the process returns to continue updating the first model and continues to use the first multimodal sample data for training.

[0130] Step 209: Based on the sample set and the trained task training model, perform transfer learning to form a task scheduling instruction model.

[0131] In this embodiment of the invention, the specific implementation process of step 209 is similar to that of step 109, and will not be repeated here.

[0132] It should be noted that in the existing heterogeneous networks, which are formed by various factors such as different application scenarios, different access terminals, different access numbers, different times, different links, and different access requests, the efficiency and accuracy of edge gateway nodes in edge computing in recognizing different task scheduling instructions in heterogeneous networks are low. This is because terminal devices, terminal nodes, gateway devices (edge ​​gateway nodes), network devices (network nodes), and cloud computing platforms produced by different manufacturers run on different protocols and support different functions or applications.

[0133] In addition to receiving various task scheduling instructions, edge gateway nodes also receive massive amounts of multimodal data in formats such as text, images, videos, and audio. By deploying a task scheduling instruction model on the edge gateway node, after receiving data transmitted from other devices, the edge gateway node can quickly identify task scheduling instructions based on prediction and analysis using the task scheduling instruction model. This allows it to decompose the task scheduling instructions transmitted by the cloud computing platform / network nodes into task scheduling instructions for terminal devices and send them to the terminal devices for device adjustment.

[0134] The training process of the task scheduling instruction model of the present invention includes: acquiring multiple sample sets; the sample sets include location sample data and multimodal sample data; pre-training the task training model based on the task training model using the multimodal sample data of the sample sets; performing perceptual training on the pre-trained task training model according to the location sample data of the heterogeneous network set in the sample sets; using the perceptually trained task training model as the first model, acquiring the first multimodal sample data from the sample sets; training the first model based on the first multimodal sample data and outputting the prediction result; performing parameter index analysis on the currently output prediction result and determining whether the analyzed parameter index has converged. Based on the convergence status, if convergence occurs, the currently output prediction result is used as the pseudo-label result of the first multimodal sample data to obtain pseudo-label sample data and supplement it to the sample set; if convergence does not occur, the first model is updated according to the currently output prediction result, and the process returns to execute the training process using the first multimodal sample data to the updated first model; based on the sample set and the trained task training model, transfer learning is performed to form a task scheduling instruction model. This invention can effectively utilize massive multimodal sample data and location sample data to train a model and obtain a task scheduling instruction model. The trained task scheduling instruction model is deployed on edge gateway nodes. Based on the model's prediction and analysis, the edge gateway nodes can quickly identify task scheduling instructions. The identification efficiency and accuracy of different task scheduling instructions in heterogeneous networks are higher, which is conducive to improving the edge collaboration effect. In subsequent processes, the task scheduling instructions transmitted by the cloud computing platform / network nodes are decomposed into task scheduling instructions for terminal devices and accurately sent to the terminal devices for device adjustment.

[0135] In this invention, multiple sample sets are obtained; these sample sets include location sample data and multimodal sample data. Based on a task training model, the task training model is pre-trained using the multimodal sample data from the sample sets. The pre-trained task training model is then perceptually trained using location sample data from a heterogeneous network defined in the sample sets. This perceptually trained task training model is used as a first model. First multimodal sample data is obtained from the sample sets, and the first model is trained based on this data. Prediction results are output, and parameter index analysis is performed on the current output prediction results. Finally, it is determined whether the analyzed parameter indexes have converged. If convergence occurs, the current output prediction result is used as the pseudo-label result of the first multimodal sample data to obtain pseudo-label sample data and supplement it to the sample set. If convergence does not occur, the first model is updated according to the current output prediction result, and the process returns to execute the training of the updated first model using the first multimodal sample data. Based on the sample set and the trained task training model, transfer learning is performed to form a task scheduling instruction model. This solves the technical problem of low efficiency and accuracy of edge gateway nodes in recognizing different task scheduling instructions in heterogeneous networks, poor edge collaboration effect, and impact on service performance in edge computing.

[0136] Please see Figure 2 , Figure 2 This is a structural block diagram of a data processing system for a virtual power plant edge computing gateway provided in Embodiment 3 of the present invention.

[0137] This invention provides a data processing system for a virtual power plant edge computing gateway, comprising:

[0138] The sample set module 301 is used to acquire multiple sample sets; the sample sets include location sample data and multimodal sample data, wherein the location sample data is the location of effective information of heterogeneous networks in different scenarios in the transmitted data, and the multimodal sample data is multimodal data in different formats;

[0139] The task training model module 302 is used to pre-train the task training model based on the task training model using multimodal sample data from the sample set;

[0140] The perception training module 303 is used to perform perception training on the pre-trained task training model based on the position sample data of the heterogeneous network in the sample set.

[0141] The sample data module 304 is used to obtain the first multimodal sample data from the sample set by using the task training model after perception training as the first model;

[0142] The prediction result module 305 is used to train the first model based on the first multimodal sample data and output the prediction result;

[0143] The convergence judgment module 306 is used to perform parameter index analysis on the current output prediction results and determine whether the analyzed parameter index has converged.

[0144] The first data processing module 307 is used to, if convergence is achieved, take the current output prediction result as the pseudo-label result of the first multimodal sample data, obtain pseudo-label sample data and supplement it to the sample set;

[0145] The second data processing module 308 is used to update the first model based on the current output prediction result if the convergence fails, and then return to execute the training process of the updated first model using the first multimodal sample data.

[0146] The task scheduling instruction model module 309 is used to perform transfer learning based on the sample set and the trained task training model to form a task scheduling instruction model.

[0147] Furthermore, the first data processing module 307 includes:

[0148] The filtering submodule is used to use the current output prediction result as the pseudo-label result of the first multimodal sample data, and to filter the pseudo-label result according to preset rules.

[0149] The supplementary submodule is used to identify the filtered multimodal sample data and pseudo-label results as pseudo-label sample data and add them to the sample set.

[0150] Furthermore, the filtering submodule includes:

[0151] The verification unit is used to verify the consistency of task labeling results from different heterogeneous networks in the pseudo-label results;

[0152] The pseudo-label result unit is used to filter pseudo-label results based on consistency results.

[0153] Furthermore, it also includes:

[0154] The initialization module is used to initialize the first and second models based on the task-trained model after perception training, and to obtain the second multimodal sample data from the sample set.

[0155] The perturbation processing module is used to perturb the second multimodal sample data using two different perturbation methods to form the first training data and the second training data, respectively.

[0156] The first prediction result module is used to input the first training data into the first model and output the first prediction result.

[0157] The second prediction result module is used to input the second training data into the second model and output the second prediction result.

[0158] The pseudo-label loss module is used to calculate the loss between the first prediction result and the second prediction result, which is the pseudo-label loss.

[0159] The feature consistency loss module is used to calculate feature consistency based on the features output from the first model and the second model to obtain the feature consistency loss.

[0160] The gradient module is used to determine the gradient based on the pseudo-label loss and feature consistency loss;

[0161] The first update module is used to update the parameters of the second model based on the gradient and pseudo-label loss.

[0162] The second update module is used to guide the updating of the parameters of the first model based on the updated parameters of the second model.

[0163] Furthermore, the task training model includes a preprocessing network, a main network, and a multi-task output network.

[0164] In this invention, multiple sample sets are obtained; these sample sets include location sample data and multimodal sample data. Based on a task training model, the task training model is pre-trained using the multimodal sample data from the sample sets. The pre-trained task training model is then perceptually trained using location sample data from a heterogeneous network defined in the sample sets. This perceptually trained task training model is used as a first model. First multimodal sample data is obtained from the sample sets, and the first model is trained based on this data. Prediction results are output, and parameter index analysis is performed on the current output prediction results. Finally, it is determined whether the analyzed parameter indexes have converged. If convergence occurs, the current output prediction result is used as the pseudo-label result of the first multimodal sample data to obtain pseudo-label sample data and supplement it to the sample set. If convergence does not occur, the first model is updated according to the current output prediction result, and the process returns to execute the training of the updated first model using the first multimodal sample data. Based on the sample set and the trained task training model, transfer learning is performed to form a task scheduling instruction model. This solves the technical problem of low efficiency and accuracy of edge gateway nodes in recognizing different task scheduling instructions in heterogeneous networks, poor edge collaboration effect, and impact on service performance in edge computing.

[0165] An electronic device according to an embodiment of the present invention includes: a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs a data processing method for a virtual power plant edge computing gateway as described in any of the above embodiments.

[0166] The memory can be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. The memory has storage space for program code used to perform any of the method steps described above. For example, the storage space for program code may include individual program codes for implementing the various steps in the methods described above. This program code can be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, compact discs (CDs), memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the methods described above.

[0167] This invention provides a computer-readable storage medium storing a computer program thereon, which, when executed, implements a data processing method for a virtual power plant edge computing gateway as described in any embodiment of this invention.

[0168] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0169] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0170] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0171] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0172] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0173] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data processing method for a virtual power plant edge computing gateway, characterized in that, include: Obtain multiple sample sets; the sample sets include location sample data and multimodal sample data, wherein the location sample data is the location of effective information of heterogeneous networks in different scenarios in the transmitted data, and the multimodal sample data is multimodal data in different formats; Based on the task training model, the task training model is pre-trained using the multimodal sample data of the sample set; The pre-trained task training model is perceptually trained based on the location sample data of the heterogeneous network in the sample set. The task training model trained after perception is used as the first model, and the first multimodal sample data is obtained from the sample set; The first model is trained based on the first multimodal sample data, and the prediction result is output. Perform parameter index analysis on the current output prediction results and determine whether the analyzed parameter indexes have converged; If convergence is achieved, the current output prediction result is used as the pseudo-label result of the first multimodal sample data to obtain pseudo-label sample data and supplement it to the sample set. If convergence fails, the first model is updated based on the current output prediction result, and the process of training the updated first model using the first multimodal sample data is returned. Based on the sample set and the trained task training model, transfer learning is performed to form a task scheduling instruction model.

2. The data processing method for the virtual power plant edge computing gateway according to claim 1, characterized in that, If convergence is achieved, the step of using the current output prediction result as the pseudo-label result of the first multimodal sample data, obtaining pseudo-label sample data, and supplementing it to the sample set includes: The current output prediction result is used as the pseudo-label result of the first multimodal sample data, and the pseudo-label result is filtered according to a preset rule; The filtered multimodal sample data and the pseudo-label results are identified as pseudo-label sample data and added to the sample set.

3. The data processing method for the virtual power plant edge computing gateway according to claim 2, characterized in that, The step of filtering the pseudo-label results according to preset rules includes: Verify the consistency of task labeling results from different heterogeneous networks in the pseudo-label results; Filter pseudo-label results based on consistency results.

4. The data processing method for the virtual power plant edge computing gateway according to claim 1, characterized in that, Before the step of using the task training model trained by perception as the first model to obtain the first multimodal sample data from the sample set, the method further includes: The first and second models are initialized based on the task training model after perception training, and the second multimodal sample data is obtained from the sample set; The second multimodal sample data is perturbed using two different perturbation methods to form the first training data and the second training data, respectively. Input the first training data into the first model and output the first prediction result; The second training data is input into the second model, and the second prediction result is output. Calculate the loss between the first prediction result and the second prediction result, and use it as the pseudo-label loss; Based on the features output from the first model and the second model, feature consistency calculation is performed to obtain feature consistency loss; The gradient is determined based on the pseudo-label loss and the feature consistency loss; The parameters of the second model are updated based on the gradient and the pseudo-label loss. The updated parameters of the second model guide the updating of the parameters of the first model.

5. The data processing method for the virtual power plant edge computing gateway according to claim 1, characterized in that, The task training model includes a preprocessing network, a main network, and a multi-task output network.

6. A data processing system for a virtual power plant edge computing gateway, characterized in that, include: The sample set module is used to acquire multiple sample sets; the sample sets include location sample data and multimodal sample data, wherein the location sample data is the location of effective information of heterogeneous networks in different scenarios in the transmitted data, and the multimodal sample data is multimodal data in different formats; The task training model module is used to pre-train the task training model based on the task training model using the multimodal sample data of the sample set. The perception training module is used to perform perception training on the pre-trained task training model based on the location sample data of the heterogeneous network set in the sample set. The sample data module is used to obtain first multimodal sample data from the sample set by using the task training model after perception training as the first model; The prediction result module is used to train the first model based on the first multimodal sample data and output the prediction result; The convergence judgment module is used to analyze the parameter indicators of the current output prediction results and determine whether the analyzed parameter indicators have converged. The first data processing module is used to, if convergence is achieved, take the current output prediction result as the pseudo-label result of the first multimodal sample data, obtain pseudo-label sample data, and supplement it to the sample set. The second data processing module is used to update the first model based on the current output prediction result if the model does not converge, and then return to execute the training process of the updated first model using the first multimodal sample data. The task scheduling instruction model module is used to perform transfer learning based on the sample set and the trained task training model to form a task scheduling instruction model.

7. The data processing system for the virtual power plant edge computing gateway according to claim 6, characterized in that, The first data processing module includes: The filtering submodule is used to use the current output prediction result as the pseudo-label result of the first multimodal sample data, and to filter the pseudo-label result according to a preset rule; The supplementary submodule is used to determine the filtered multimodal sample data and the pseudo-label results as pseudo-label sample data and supplement them into the sample set.

8. The data processing system for the virtual power plant edge computing gateway according to claim 7, characterized in that, The filtering submodule includes: The verification unit is used to verify the consistency of task labeling results from different heterogeneous networks in the pseudo-label results; The pseudo-label result unit is used to filter pseudo-label results based on consistency results.

9. The data processing system for the virtual power plant edge computing gateway according to claim 6, characterized in that, Also includes: An initialization module is used to initialize the first model and the second model based on the task training model after perception training, and to obtain the second multimodal sample data from the sample set. The perturbation processing module is used to perturb the second multimodal sample data using two different perturbation methods to form the first training data and the second training data respectively. The first prediction result module is used to input the first training data into the first model and output the first prediction result. The second prediction result module is used to input the second training data into the second model and output the second prediction result; The pseudo-label loss module is used to calculate the loss between the first prediction result and the second prediction result, which is the pseudo-label loss. The feature consistency loss module is used to perform feature consistency calculation based on the features output from the first model and the second model to obtain the feature consistency loss. The gradient module is used to determine the gradient based on the pseudo-label loss and the feature consistency loss; The first update module is used to update the parameters of the second model based on the gradient and the pseudo-label loss. The second update module is used to guide the updating of the parameters of the first model based on the updated parameters of the second model.

10. The data processing system for the virtual power plant edge computing gateway according to claim 6, characterized in that, The task training model includes a preprocessing network, a main network, and a multi-task output network.

11. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the data processing method for a virtual power plant edge computing gateway as described in any one of claims 1-5.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the data processing method of the virtual power plant edge computing gateway as described in any one of claims 1-5.

Images