Task scheduling method and apparatus based on cloud-edge collaboration

Abstract

Provided in the present invention are a task scheduling method and apparatus based on cloud-edge collaboration. The method comprises: acquiring, from an edge end, a task to be allocated; acquiring time queues of resource load values of the edge end and a cloud end; extracting resource load time-series features of the edge end and the cloud end; on the basis of the resource load time-series features, performing refined feature interaction based on a spatial distance measurement, so as to obtain a resource load time-series collaborative feature; and on the basis of the resource load time-series collaborative feature, splitting said task, and respectively scheduling same to the edge end and the cloud end, such that the edge end and the cloud end respectively execute said task. In the present invention, a resource load time-series feature of an edge end and a resource load time-series feature of a cloud end are extracted, resource load conditions of the cloud end and the edge end are taken into consideration, and refined feature interaction based on a spatial distance measurement is then performed to obtain a resource load time-series collaborative feature, so as to realize the flexible splitting and scheduling of a task, such that the overall operation stability of a system can be improved, thereby improving the task processing efficiency.

Description

Cloud-edge collaborative task scheduling method and device Technical Field

[0001] This invention relates to the field of task scheduling, and in particular to a task scheduling method and apparatus based on cloud-edge collaboration. Background Technology

[0002] With the integration of new energy sources and the widespread application of distributed energy, adjustable loads, and energy storage facilities in the power grid, the dispatch and control mode of the power system is shifting from the traditional "source follows load" to a multi-level coordinated control mode of "source-grid-load-storage". This shift requires the power grid to have higher real-time sensing capabilities and more precise control capabilities over distributed energy, loads, and energy storage resources.

[0003] To address this challenge, a two-tiered collaborative power dispatching system based on "cloud + edge" has emerged. This system provides data services by integrating network, storage, and computing technologies at the edge, aiming to improve system efficiency. However, when handling large-scale power system computational analysis tasks, on the one hand, while edge devices can respond quickly, their limited computing power and storage space may make them inadequate for complex tasks; on the other hand, while cloud computing centers possess powerful computing capabilities, they may encounter latency issues when processing large amounts of real-time data. Therefore, existing systems often face an imbalance in load between the edge network and the cloud, affecting the overall stability of the system. How to effectively integrate the advantages of edge computing and cloud computing to create synergy has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a task scheduling method and apparatus based on cloud-edge collaboration to solve the technical problem of uneven task allocation between the edge and the cloud, and improve the overall stability of the system operation.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a task scheduling method based on cloud-edge collaboration, comprising:

[0006] Obtain tasks to be assigned from the edge, the tasks to be assigned being power calculation tasks;

[0007] The time queue for obtaining the resource load values ​​at the edge and the time queue for obtaining the resource load values ​​in the cloud;

[0008] Based on the time queue of resource load values ​​at the edge, extract the time-series features of resource load at the edge; based on the time queue of resource load values ​​at the cloud, extract the time-series features of resource load at the cloud.

[0009] Based on the time-series characteristics of resource load at the edge and in the cloud, detailed feature interaction based on spatial distance metric is performed to obtain resource load time-series collaborative characteristics.

[0010] Based on the resource load timing coordination characteristics, the task to be allocated is split into a first power calculation subtask and a second power calculation subtask. The first power calculation subtask and the second power calculation subtask are scheduled to the edge and the cloud respectively, so that the edge and the cloud can execute the first power calculation subtask and the second power calculation subtask respectively.

[0011] Accordingly, this invention application also provides a task scheduling device based on cloud-edge collaboration, including a task acquisition module, a queue acquisition module, a feature extraction module, a feature interaction module, and a scheduling module; wherein,

[0012] The task acquisition module is used to acquire tasks to be assigned from the edge, and the tasks to be assigned are power calculation tasks;

[0013] The queue acquisition module is used to acquire the time queue of resource load values ​​at the edge and the time queue of resource load values ​​in the cloud.

[0014] The feature extraction module is used to extract time-series features of resource load at the edge end based on the time-series of resource load values ​​at the edge end; and to extract time-series features of resource load at the cloud end based on the time-series of resource load values ​​at the cloud end.

[0015] The feature interaction module is used to perform detailed feature interaction based on spatial distance metric based on the time-series features of resource load at the edge and the time-series features of resource load in the cloud, so as to obtain resource load time-series collaborative features.

[0016] The scheduling module is used to split the task to be allocated according to the resource load timing coordination characteristics to obtain a first power calculation subtask to be allocated and a second power calculation subtask to be allocated, and to schedule the first power calculation subtask to be allocated and the second power calculation subtask to be allocated to the edge and the cloud respectively, so that the edge and the cloud respectively execute the first power calculation subtask to be allocated and the second power calculation subtask to be allocated.

[0017] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0018] This invention provides a task scheduling method and apparatus based on cloud-edge collaboration. By extracting the time-series characteristics of resource load at the edge and the cloud, and taking into account the resource load conditions at both the cloud and the edge, it performs detailed feature interaction based on spatial distance metrics to obtain resource load time-series collaborative characteristics. This enables flexible task splitting and scheduling. Compared with existing technologies, it can achieve load balancing of power computing tasks between the edge and the cloud, effectively integrate the advantages of the edge and the cloud, improve the overall stability of the system and the intelligence of task scheduling, and improve task processing efficiency. Attached Figure Description

[0019] Figure 1: A flowchart illustrating an embodiment of the cloud-edge collaborative task scheduling method provided in this application.

[0020] Figure 2: A schematic diagram of an embodiment of the cloud-edge collaborative task scheduling device provided in this application.

[0021] Figure 3: A structural block diagram of an electronic device based on a cloud-edge collaborative task scheduling method provided in this invention application.

[0022] Figure 4: A schematic diagram of an application scenario of an embodiment of the cloud-edge collaborative task scheduling method provided in this application. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1:

[0025] Please refer to Figure 1, which illustrates a cloud-edge collaborative task scheduling method according to an embodiment of the present invention, including steps S101 to S105; wherein...

[0026] Step S101: Obtain the task to be assigned from the edge, wherein the task to be assigned is a power calculation task.

[0027] Step S102: Obtain the time queue of resource load values ​​at the edge and the time queue of resource load values ​​in the cloud.

[0028] Step S103: Extract the time-series features of the resource load at the edge based on the time-series queue of the resource load values ​​at the edge; extract the time-series features of the resource load at the cloud based on the time-series queue of the resource load values ​​at the cloud.

[0029] Step S104: Based on the time-series characteristics of resource load at the edge and the time-series characteristics of resource load in the cloud, perform detailed feature interaction based on spatial distance metric to obtain resource load time-series collaborative characteristics.

[0030] Step S105: Based on the resource load timing coordination characteristics, the task to be allocated is split into a first power calculation subtask to be allocated and a second power calculation subtask to be allocated. The first power calculation subtask to be allocated and the second power calculation subtask to be allocated are scheduled to the edge and the cloud respectively, so that the edge and the cloud respectively execute the first power calculation subtask to be allocated and the second power calculation subtask to be allocated.

[0031] In this embodiment, the edge terminal can be an edge gateway or an edge device. Regarding step S103 above, extracting the time-series features of edge terminal resource load based on the time queue of the resource load values ​​of the edge terminal; and extracting the time-series features of cloud resource load based on the time queue of the resource load values ​​of the cloud, specifically:

[0032] The temporal features of edge resource load and cloud resource load are input into the 1D-CNN model to obtain the sequence of local temporal correlation feature vectors of edge resource load and the sequence of local temporal correlation feature vectors of cloud resource load.

[0033] The edge resource load temporal characteristics are obtained by performing cluster analysis on the sequence of local temporal correlation feature vectors of resource load; the cloud resource load temporal characteristics are obtained by performing cluster analysis on the sequence of local temporal correlation feature vectors of resource load.

[0034] Specifically, the feature enhancement aggregation representation processing, which involves clustering the sequence of local temporal correlation feature vectors of edge resource load to obtain the temporal features of edge resource load, is as follows:

[0035] Cluster analysis is performed on the sequence of local temporal correlation feature vectors of edge resource load to obtain a self-supervised clustering representation vector of local temporal correlation features of edge resource load; based on the self-supervised clustering representation vector of local temporal correlation features of edge resource load, feature clustering field modulation aggregation is performed on the sequence of local temporal correlation feature vectors of edge resource load to obtain a temporal significant aggregation representation vector of edge resource load, which is used as the temporal feature of edge resource load.

[0036] It is understood that the sequences of local temporal correlation feature vectors of edge resource load and cloud resource load respectively contain local temporal feature information about the resource load values ​​at the edge and cloud. These local temporal features have contextual correlation information in the entire time domain, which is of great significance for subsequent task allocation. Therefore, this embodiment can further perform clustering analysis on the sequences of local temporal correlation feature vectors of edge resource load and cloud resource load to obtain the temporal features of edge resource load and cloud resource load respectively. Furthermore, the clustering analysis feature enhancement aggregation representation aims to use pseudo-field analysis theory to perform field-based explicit modeling of the clustering contribution of each local temporal correlation feature vector of edge resource load and each local temporal correlation feature vector of cloud resource load, thereby improving the accuracy of feature aggregation analysis and improving the performance and generalization ability of the model.

[0037] Further, based on the self-supervised clustering representation vector of the local temporal correlation features of the edge resource load, the sequence of the local temporal correlation feature vector of the edge resource load is subjected to feature clustering field modulation aggregation to obtain the temporal significant aggregation representation vector of the edge resource load, which serves as the temporal feature of the edge resource load. Specifically, this involves:

[0038] Calculate the implicit clustering contribution factor of each edge resource load local temporal correlation feature vector relative to the edge resource load local temporal correlation feature self-supervised clustering representation vector in the sequence of edge resource load local temporal correlation feature vectors, so as to obtain the sequence of implicit clustering contribution factors of edge resource load local temporal correlation feature.

[0039] The sequence of latent clustering contribution factors of the local temporal correlation features of edge resource load is vectorized to obtain the distribution vector of the clustering contribution field of the local temporal correlation features of edge resource load.

[0040] The distribution vector of the cluster contribution field of the local temporal correlation feature of the edge resource load is input into the explicit modeling module of the cluster contribution field based on the self-attention mechanism to obtain the domain modulation weight vector of the cluster contribution field of the local temporal correlation feature of the edge resource load.

[0041] Using each feature value in the field modulation weight vector of the local temporal correlation feature clustering of edge resource load as a weight, the weighted sum of the corresponding positions of the sequence of the local temporal correlation feature vector of edge resource load is calculated to obtain the temporal significant aggregation representation vector of edge resource load.

[0042] The feature enhancement aggregation representation step in cluster analysis first utilizes unsupervised learning methods to perform cluster analysis on the sequence of edge resource load local temporal correlation feature vectors. This automatically discovers the feature patterns and structures of edge resource load local temporal context correlation in the set, generating a self-supervised clustering representation vector for edge resource load local temporal correlation features. It can be understood that self-supervised clustering reveals the inherent structure of each edge resource load local temporal context correlation feature, helping the model learn meaningful feature representations even without explicit labels, thus enhancing the model's understanding of the distribution of edge resource load local temporal context correlation features. Furthermore, by calculating the implicit clustering contribution factor between the edge resource load local temporal correlation feature vector and the self-supervised clustering representation vector, the degree of contribution of each edge resource load local temporal correlation feature vector to its respective cluster is reflected. Accordingly, calculating the latent clustering contribution factor helps identify and quantify the role of each edge resource load local temporal correlation feature vector in the clustering process, providing a basis for subsequent feature selection and weight allocation, enabling the model to focus more on features that significantly contribute to clustering. Next, the calculated sequence of the latent clustering contribution factors of the edge resource load local temporal correlation features is converted into a distribution vector of the edge resource load local temporal correlation feature clustering contribution field. This involves integrating the sequence of the latent clustering contribution factors of the edge resource load local temporal correlation features into a vector field to construct a clustering contribution field. Subsequently, a self-attention mechanism is used to explicitly model the distribution vector of the edge resource load local temporal correlation feature clustering contribution field to generate a modulated weight vector for the edge resource load local temporal correlation feature clustering contribution field. In other words, a self-attention mechanism is used to explicitly model the clustering contribution of the implicit clustering contribution factors of the local temporal correlation features of each edge resource load in the clustering contribution field based on the field autocorrelation global domain analysis. This strengthens the focus on feature vectors with large clustering contributions while suppressing unimportant features, improving the quality of feature representation and the accuracy of subsequent power calculations and task scheduling. Finally, the field modulation weight vector of the clustering contribution features of the edge resource load is used to weight and fuse the sequence of the local temporal correlation feature vectors of the edge resource load to generate a spatial domain edge resource load significant aggregation representation vector. This spatial domain edge resource load significant aggregation representation vector provides a comprehensive feature representation that not only contains the original local temporal correlation feature information of the edge resource load but also integrates the knowledge learned during the clustering process. This enables the model to more accurately capture the key local temporal semantics of the edge resource load when dealing with complex tasks, providing powerful feature support for the application of deep learning models in various complex tasks, thereby improving the model's performance and generalization ability.

[0043] In some preferred embodiments, the implicit clustering contribution factor of each edge resource load local temporal correlation feature vector relative to the edge resource load local temporal correlation feature self-supervised clustering representation vector in the sequence of calculating the edge resource load local temporal correlation feature vector is used to obtain a sequence of implicit clustering contribution factors of edge resource load local temporal correlation features, specifically:

[0044] Calculate the absolute value of the division between the corresponding positions of the edge resource load local temporal correlation feature vector and the edge resource load local temporal correlation feature self-supervised clustering representation vector to obtain the edge resource load local temporal self-supervised clustering interaction correlation feature vector;

[0045] Calculate the logarithmic function value of each position feature value in the local temporal self-supervised clustering interaction correlation feature vector of the edge resource load with base 2, so as to obtain the local temporal self-supervised clustering interaction logarithmic representation vector of the edge resource load;

[0046] Using the feature values ​​at each position in the local temporal self-supervised clustering interaction logarithm representation vector of the edge resource load as weighting coefficients, the weighted sum of the feature values ​​at each position in the local temporal correlation feature vector of the edge resource load is calculated to obtain the clustering contribution representation factor of the local temporal correlation feature of the edge resource load.

[0047] The exponential function value of the clustering contribution factor of the local temporal correlation feature of the edge resource load, with the natural constant e as the base, is calculated to obtain the latent clustering contribution factor of the local temporal correlation feature of the edge resource load.

[0048] Further, the step of inputting the distribution vector of the clustering contribution field of the local temporal correlation features of the edge resource load into the explicit modeling module of the clustering contribution field based on the self-attention mechanism to obtain the modulation weight vector of the clustering contribution field of the local temporal correlation features of the edge resource load is specifically as follows:

[0049] Calculate the matrix multiplications between the distribution vector of the local temporal correlation feature clustering contribution field of the edge resource load and the query weight matrix, key weight matrix and value weight matrix, respectively, to obtain the query vector of the distribution vector of the clustering contribution field of the edge resource load, the key vector of the distribution vector of the clustering contribution field of the edge resource load and the value vector of the distribution vector of the clustering contribution field of the edge resource load;

[0050] Calculate the multiplication between the edge resource load clustering distribution query vector and the transpose of the edge resource load clustering contribution field distribution key vector. Use the result of the multiplication as the dividend and the square root of the length of the edge resource load clustering contribution field distribution key vector as the divisor. Perform outward processing according to the corresponding position to obtain the edge resource load clustering contribution field semantic interaction representation matrix.

[0051] The semantic interaction representation matrix of the edge resource load clustering contribution field is normalized by the Softmax function, and the product between the distribution value vector of the edge resource load clustering contribution field and the normalization result is calculated to obtain the domain modulation weight vector of the edge resource load local temporal correlation feature clustering contribution field.

[0052] In some preferred embodiments, the feature enhancement aggregation representation processing of the sequence of local temporal correlation feature vectors of edge resource load can be performed according to the following feature aggregation formula to obtain the temporal features of edge resource load. The formula is specifically: X={x1,x2,...,x k ,...,x n}; v d ={d1; d2; ...; d j ;...;d n}; v q =W q v d ; v k =W k v d ; v v =W b v d ;

[0053] Where X is a sequence of local temporal correlation feature vectors of the edge resource load, x1, x2, x... k ,x n These are the 1st, 2nd, kth, and nth edge resource load local temporal correlation feature vectors in the sequence of edge resource load local temporal correlation feature vectors, respectively. i Let x be the i-th local temporal correlation feature vector of the edge resource load in the sequence of the edge resource load local temporal correlation feature vectors, and N be the number of feature vectors in the sequence of the edge resource load local temporal correlation feature vectors. c x is the self-supervised clustering representation vector of local temporal correlation features of edge resource load. jLet j be the j-th edge resource load local temporal correlation feature vector in the sequence of edge resource load local temporal correlation feature vectors. Let k be the feature value at the k-th position in the local temporal correlation feature vector of the j-th edge resource load. Let be the feature value at the k-th position in the self-supervised clustering representation vector of the local temporal correlation features of the edge resource load, where log represents the logarithmic function value to the base 2, exp represents the natural exponential function value, and d j v is the implicit clustering contribution factor of the local temporal correlation feature of the edge resource load corresponding to the j-th edge resource load local temporal correlation feature vector. d W is the distribution vector of the clustering contribution field of the local temporal correlation features of edge resource load, obtained by vectorizing and arranging the latent clustering contribution factors of multiple edge resource load local temporal correlation features. q W k and W b These are the query weight matrix, key weight matrix, and value weight matrix, respectively. q v k and v v These are, respectively, the edge resource load clustering contribution field distribution query vector, the edge resource load clustering contribution field distribution key vector, and the edge resource load clustering contribution field distribution value vector, where L is the length of the edge resource load clustering contribution field distribution key vector, and Softmax(·) is the Softmax function. For vector multiplication, v w To contribute a field modulation weight vector to the clustering of local temporal correlation features of edge resource load, x f The temporal significant aggregation representation vector of the resource load at the edge.

[0054] In some implementations, cluster analysis is performed on the sequence of local temporal correlation feature vectors of edge resource load to obtain a self-supervised clustering representation vector of local temporal correlation features of edge resource load. This includes: calculating the mean vector of the sequence of local temporal correlation feature vectors of edge resource load to obtain the self-supervised clustering representation vector of local temporal correlation features of edge resource load.

[0055] In some implementations, for step S104 above, the detailed feature interaction based on spatial distance metric, based on the time-series characteristics of edge resource load and cloud resource load, to obtain resource load time-series collaborative characteristics, specifically includes:

[0056] Fine-grained feature decoupling is performed on the temporal saliency aggregation representation vector of edge resource load according to a preset scale to obtain a set of feature vectors of temporal sub-components of edge resource load; fine-grained feature decoupling is also performed on the temporal saliency aggregation representation vector of cloud resource load according to the preset scale to obtain a set of feature vectors of temporal sub-components of cloud resource load. This allows for feature comparison at a lower level, thereby capturing more subtle feature differences. This means that the system not only considers the overall load level in the edge and cloud, but also pays attention to minute temporal changes in load characteristics, which is crucial for dynamically adjusting task scheduling.

[0057] The set of feature vectors of temporal sub-components of edge resource load and the set of feature vectors of temporal sub-components of cloud resource load are subjected to (optimal) feature sub-component pairing and screening and interactive fusion processing to obtain the (optimal) temporal collaborative features of edge-cloud resource load.

[0058] In this embodiment, a detailed feature interaction method based on spatial distance metrics can enhance the effectiveness of feature representation, thereby improving the accuracy of task allocation and dynamic scheduling decisions, and better integrating and interacting the temporal features of edge resource load and cloud resource load. In practical applications, the resource load status of the edge and cloud continuously changes over time, and these changes may be very subtle and easily overlooked, leading to a decrease in the accuracy of task allocation and dynamic scheduling decisions. Therefore, in the technical solution of this application, a detailed feature interaction based on spatial distance metrics is further performed on the temporal features of edge resource load and cloud resource load to obtain edge-cloud resource load temporal collaborative features.

[0059] Furthermore, the process of performing feature component pairing, filtering, and interactive fusion on the set of temporal sub-component feature vectors of the edge resource load and the set of temporal sub-component feature vectors of the cloud resource load to obtain resource load temporal collaborative features specifically involves:

[0060] Calculate the hyperbolic space distance metric between the feature vectors of the temporal sub-components of resource load at each edge and the feature vectors of the temporal sub-components of resource load at each cloud.

[0061] Based on the hyperbolic space distance metric factor, the (optimal) feature component pairing and screening of the feature vectors of the time-series sub-components of resource load at each edge end and the feature vectors of the time-series sub-components of resource load at each cloud end are performed to obtain the optimal pairing set of {feature vectors of the time-series sub-components of resource load at the edge end and feature vectors of the time-series sub-components of resource load at the cloud end}.

[0062] Each optimal {edge-end resource load time-series sub-component feature vector, cloud resource load time-series sub-component feature vector} in the pairing set is input into the feature multi-dimensional interaction module to obtain a set of (edge-end-cloud optimal) load time-series sub-component feature pairing interaction fusion feature vectors;

[0063] The set of feature vectors of the load time-series sub-components are concatenated to obtain the resource load time-series collaborative features (edge-cloud resources).

[0064] In this embodiment, after feature decoupling, a hyperbolic space distance metric is calculated between the feature vectors of different sub-components. Using hyperbolic space distance, rather than traditional Euclidean distance or other metrics, more effectively represents hierarchical or high-dimensional sparse data. Since hyperbolic space grows faster away from the origin, this metric better preserves the hierarchical structure of the data and is more suitable for handling data distributions with exponential growth characteristics. This is crucial for capturing the fine-grained relationship between the time-series features of cloud resource load and edge resource load, especially when dealing with complex or nonlinear relationships, thus more accurately reflecting the actual load coordination and dynamic changes between the edge and cloud. Subsequently, based on the hyperbolic space distance metric, optimal pairing is performed between the time-series feature vectors of edge resource load and cloud resource load. The aim is to find the pair of feature vectors with the smallest distance in hyperbolic space, thereby achieving optimal matching. The key to this step is ensuring that the matching of each pair of cloud resource load temporal sub-components and edge resource load temporal sub-components best represents the similarity between the two feature vector sets. This ensures that even very subtle feature differences are not ignored, guaranteeing that the subsequent fusion process retains the most valuable information. After optimal pairing, these pairings are input into the feature multi-dimensional interaction module. This module is responsible for enabling interactions between paired feature vectors, generating feature pairing interaction fusion feature vectors. The interaction process may include, but is not limited to, addition, multiplication, or other forms of combination operations, with the goal of creating new feature representations that better reflect the complex relationship between cloud resource load temporal and edge resource load temporal in the original data. This step enhances the expressive power of the features, enabling the model to learn deeper feature associations. Finally, all the edge-cloud resource load optimal sub-component feature pairing interaction fusion feature vectors generated by the interactions are concatenated, that is, merged into a single edge-cloud resource load fine-grained optimal aligned fusion feature vector according to a certain order or rule. The purpose of concatenating fusion is to integrate the information after the interaction of all edge-cloud resource load sub-components, forming a comprehensive and optimized feature representation. The resulting feature vector not only contains information from the original feature vector but also incorporates new information generated through fine-grained matching and multi-dimensional interaction, thus enabling a more comprehensive representation of the data and improving the accuracy and efficiency of task scheduling.

[0065] Furthermore, the hyperbolic space distance metric factor for calculating the temporal sub-component feature vectors of resource load at each edge and the temporal sub-component feature vectors of resource load at each cloud specifically includes:

[0066] Extract the edge resource load time-series sub-component feature vectors at a preset position from the set of edge resource load time-series sub-component feature vectors;

[0067] Extract the cloud resource load time-series sub-component feature vectors at preset positions from the set of cloud resource load time-series sub-component feature vectors;

[0068] Calculate the difference vector between the edge resource load time-series sub-component feature vector at the preset location and the cloud resource load time-series sub-component feature vector at the preset location, and perform squaring based on the first norm of the difference vector to obtain the first norm representation of the semantic difference of the load time-series sub-component.

[0069] Calculate the square of the first norm of the feature vector of the temporal sub-component of the edge resource load at the preset position to obtain the first squared result; subtract the first constant (which can be a constant 1) from the first squared result according to the corresponding position to obtain the first norm representation of the temporal sub-component of the edge resource load.

[0070] Calculate the square of the first norm of the feature vector of the time series sub-component of cloud resource load at the preset position to obtain the second square result; subtract the second constant (which can be a constant 1) from the second square result according to the corresponding position to obtain the first norm representation of the time series sub-component of cloud resource load.

[0071] Calculate the multiplication between the first-norm representation of the temporal component of the edge resource load and the first-norm representation of the temporal component of the cloud resource load. Use the multiplication result as the divisor and the first-norm of the semantic difference of the temporal component of the load as the dividend. Perform division to obtain the semantic implicit association representation of the first-norm of the load temporal load.

[0072] The load time-series norm semantic implicit association representation is multiplied by the third constant (which can be constant 2), the product is added to the fourth constant (which can be constant 1), and the result of the addition is processed by the inverse hyperbolic cosine function to obtain the hyperbolic space distance metric factor.

[0073] In some implementations, detailed feature interaction based on spatial distance metrics is performed on the temporal characteristics of edge-cloud resource load to obtain the temporal collaborative characteristics of edge-cloud resource load. This can be achieved according to the following formula: Decouple(v1)={v 11 ,v 12 ,...,v 1i ,...,v 1n}; Decouple(v2)={v 21 ,v 22 ,...,v 2i ,...,v2n}; v f =[v p1 ;v p2 ;...v pn ];

[0074] Wherein, v1 is the temporal significant aggregated representation vector of the edge resource load, v2 is the temporal significant aggregated representation vector of the cloud resource load, and Decouple is the fine-grained feature decoupling operation. 11 ,v 12 ,v 1i ,v 1n These are the 1st, 2nd, ith, and nth edge resource load temporal component feature vectors in the set of edge resource load temporal component feature vectors, respectively. 21 ,v 22 ,v 2i ,v 2n Let be the 1st, 2nd, 1st, and 1st time-series sub-component feature vectors of cloud resource load, respectively, in the set of cloud resource load time-series sub-component feature vectors. ||·|| denotes the first norm of the vector, and arccosh is the inverse hyperbolic cosine function. d P (v 1i ,v 2j ) for v 1i and v 2j Hyperbolic space distance metric between them To obtain d P (v 1i ,v 2j The minimum value corresponds to parameter j, k is the optimal pairing position value, and v 2k Let k be the k-th cloud resource load time-series sub-component feature vector in the set of cloud resource load time-series sub-component feature vectors. ⊙ and These represent subtraction by position, dot product by position, and addition by position, respectively. α, β, and γ are weighting hyperparameters, and v is the value of the expression. pi Let be the i-th feature vector in the set of edge-cloud resource load time-series optimal sub-component feature pairing and interactive fusion feature vectors, where [·;·;·] represents the vector concatenation operation, v f This is the temporal collaborative alignment representation vector for the edge-cloud resource load.

[0075] In some implementations, based on the hyperbolic spatial distance metric, the (optimal) feature component pairing of each edge resource load temporal sub-component feature vector and each cloud resource load temporal sub-component feature vector is screened to obtain an optimal {edge resource load temporal sub-component feature vector, cloud resource load temporal sub-component feature vector} pairing set. Specifically, the position of the cloud resource load temporal sub-component feature vector corresponding to the minimum value among multiple hyperbolic spatial distance metric factors is selected as the optimal pairing position value of the edge resource load temporal sub-component feature vector at the preset position; the {edge resource load temporal sub-component feature vector, cloud resource load temporal sub-component feature vector} corresponding to the optimal pairing position value is paired to obtain the optimal {edge resource load temporal sub-component feature vector, cloud resource load temporal sub-component feature vector} pairing.

[0076] In some implementations, the step of obtaining a set of load time-series sub-component feature pairing interaction fusion feature vectors by multi-dimensional interaction modules of each paired input feature in the pairing set specifically involves:

[0077] The edge-end resource load time-series sub-component feature vector and the cloud-end resource load time-series sub-component feature vector in the optimal {edge-end resource load time-series sub-component feature vector, cloud-end resource load time-series sub-component feature vector} pair are respectively processed by positional difference, positional dot product, and positional summation to obtain edge-end-cloud resource load time-series sub-component difference feature vector, edge-end-cloud resource load time-series sub-component dot product feature vector, and edge-end-cloud resource load time-series sub-component summation feature vector; the weighted sum of corresponding positions between the edge-end-cloud resource load time-series sub-component difference feature vector, the edge-end-cloud resource load time-series sub-component dot product feature vector, and the edge-end-cloud resource load time-series sub-component summation feature vector is calculated to obtain the edge-end-cloud resource load time-series optimal sub-component feature pairing interactive fusion feature vector.

[0078] Furthermore, regarding step S105 above, which involves splitting the task to be allocated based on the resource load timing coordination characteristics to obtain a first power calculation subtask to be allocated and a second power calculation subtask to be allocated, and scheduling the first power calculation subtask to be allocated and the second power calculation subtask to the edge and the cloud respectively, the specific steps are as follows:

[0079] The resource load time-series collaborative features are input into the classifier-based task splitting module to obtain the first power calculation subtask to be allocated and the second power calculation subtask to be allocated.

[0080] The first and second power calculation subtasks to be assigned are distributed to the edge and the cloud, respectively.

[0081] For example, the first power calculation subtask to be allocated and the second power calculation subtask to be allocated can correspond to a first ratio and a second ratio, respectively, and the sum of the first ratio and the second ratio is one. In this way, this implementation can achieve more intelligent dynamic task scheduling, so as to ensure load balancing and system stability between the edge network and the cloud in the cloud-edge system, and help optimize the task processing efficiency of the entire system.

[0082] In some implementations, the resource load time-series collaborative features are input into a classifier-based task splitting module to obtain a first power calculation subtask to be allocated and a second power calculation subtask to be allocated, specifically:

[0083] Determine the maximum and minimum eigenvalues ​​of the temporal collaborative alignment of the edge-cloud resource load temporal collaborative alignment representation vector, and calculate the mean and standard deviation of the temporal collaborative alignment of the feature set of the edge-cloud resource load temporal collaborative alignment representation vector.

[0084] Calculate the quotient of the temporal co-alignment mean and the temporal co-alignment standard deviation to obtain the statistically standardized value of temporal co-alignment;

[0085] Calculate the reciprocal of each eigenvalue of the edge-cloud resource load time-series collaborative alignment representation vector, multiply it by the difference between the largest and smallest eigenvalues ​​of the time-series collaborative alignment, and then subtract it from the statistically standardized value of the time-series collaborative alignment to obtain the time-series collaborative alignment distribution approximation vector.

[0086] Calculate an exponential function of each eigenvalue of the temporal co-aligned distribution approximation vector, with the natural constant as the base, to obtain a temporal co-aligned distribution class approximation vector;

[0087] The approximate vector of the temporal collaborative alignment distribution class is added to the statistically normalized value of the temporal collaborative alignment, and the base-2 logarithm of the absolute value of each feature value of the added vector is calculated to obtain the optimized edge-cloud resource load temporal collaborative alignment representation vector; and the optimized edge-cloud resource load temporal collaborative alignment representation vector is input into the task segmentation module based on the classifier to obtain the segmentation result, namely the first power calculation subtask to be assigned and the second power calculation subtask to be assigned.

[0088] Wherein, the optimized representation of the resource load time-series collaborative alignment vector, denoted as V, is: δ=v max -v min ;

[0089] η is the statistically normalized value of temporal co-alignment, δ is the difference between the maximum and minimum eigenvalues ​​of temporal co-alignment, log is the logarithm to the base 2, and v max It is the largest eigenvalue of the temporal collaborative alignment of the edge-cloud resource load temporal collaborative alignment representation vector, v min V is the minimum eigenvalue of the temporal collaborative alignment of the edge-cloud resource load temporal collaborative alignment representation vector, μ is the mean of the temporal collaborative alignment, σ is the standard deviation of the temporal collaborative alignment, and V is the minimum eigenvalue of the temporal collaborative alignment. ⊙-1 It is the reciprocal of each eigenvalue of the edge-cloud resource load time-series collaborative alignment representation vector, exp(·) is an exponential function calculated with the natural constant as the base and each eigenvalue of the vector as the exponent, V e It is an approximate vector of the temporal collaborative alignment distribution, where log is the base-2 logarithmic value, and V' is the optimized edge-cloud resource load temporal collaborative alignment representation vector. It is added based on the position point. ⊙ is subtraction by position, and ⊙ is multiplication by position.

[0090] In implementing this preferred embodiment, considering the distribution differences between the time queues of resource load values ​​at the edge and the time queues of resource load values ​​in the temporal source domain, the fine-grained differences in local features between the edge-end resource load temporal significant aggregation representation vector and the cloud-end resource load temporal significant aggregation representation vector will be amplified after local temporal feature extraction and local feature temporal aggregation. Thus, when input into a feature interaction network based on fine-grained feature optimal matching, the resulting edge-cloud resource load temporal co-aligned representation vector will also have a diverse set expression distribution due to the offset of local fine-grained feature matching. Therefore, it is desirable to improve the balance between the accuracy and completeness of the regression mapping when the edge-cloud resource load temporal co-aligned representation vector is input into a task segmentation module based on a classifier for class regression, thereby improving the accuracy of the obtained segmentation results.

[0091] Therefore, by randomly statistically standardizing the diverse feature set of the edge-cloud resource load temporal co-alignment representation vector, the confidence space constructed based on the overall feature values ​​of the edge-cloud resource load temporal co-alignment representation vector is approximated by a standardized continuous probability density distribution relative to the response hypothesis test of each feature value of the edge-cloud resource load temporal co-alignment representation vector. This establishes the target reachability of the diverse feature distribution of the edge-cloud resource load temporal co-alignment representation vector to a unified regression target, achieving a balance between mapping accuracy and mapping completeness in the class regression process based on the diverse feature distribution of the edge-cloud resource load temporal co-alignment representation vector, and improving the accuracy of the segmentation results obtained by the task segmentation module based on the classifier when the edge-cloud resource load temporal co-alignment representation vector is input. In this way, large tasks can be more accurately divided into multiple sub-tasks based on the collaborative correlation and dynamic changes between the resource load values ​​of the edge and the cloud, thereby achieving more intelligent dynamic task scheduling and ensuring load balancing and system stability in the cloud-edge system.

[0092] In summary, the above-described scheme involves real-time monitoring and collection of resource load values ​​at both the edge and cloud ends. Then, artificial intelligence and deep learning-based data processing and analysis algorithms are introduced at the backend to perform temporal collaborative analysis and feature interaction on the resource load values ​​of the edge and cloud. This allows the scheme to learn and capture the temporal collaborative changes and dynamic correlations between the resource loads of the edge and cloud, thereby determining the proportion allocated to the edge and cloud. Based on this proportion, the total unallocated power computing tasks are then split and allocated. This enables more intelligent dynamic task scheduling, ensuring load balancing and system stability in the cloud-edge system, and contributing to the optimization of the overall system's task processing efficiency.

[0093] Accordingly, referring to Figure 2, this embodiment of the invention also provides a cloud-edge collaborative task scheduling device 200, including a task acquisition module 201, a queue acquisition module 202, a feature extraction module 203, a feature interaction module 204, and a scheduling module 205; wherein,

[0094] The task acquisition module 201 is used to acquire tasks to be assigned from the edge, and the tasks to be assigned are power calculation tasks.

[0095] The queue acquisition module 202 is used to acquire the time queue of the resource load value at the edge and the time queue of the resource load value in the cloud.

[0096] The feature extraction module 203 is used to extract time-series features of resource load at the edge end based on the time-series of resource load values ​​at the edge end; and to extract time-series features of resource load at the cloud end based on the time-series of resource load values ​​at the cloud end.

[0097] The feature interaction module 204 is used to perform detailed feature interaction based on spatial distance metric based on the time-series features of resource load at the edge and the time-series features of resource load in the cloud, so as to obtain resource load time-series collaborative features.

[0098] The scheduling module 205 is used to split the task to be allocated according to the resource load timing coordination characteristics to obtain a first power calculation subtask to be allocated and a second power calculation subtask to be allocated, and to schedule the first power calculation subtask to be allocated and the second power calculation subtask to be allocated to the edge and the cloud respectively, so that the edge and the cloud respectively execute the first power calculation subtask to be allocated and the second power calculation subtask to be allocated.

[0099] As a preferred embodiment, the feature extraction module 203 extracts time-series features of edge resource load based on the time queue of resource load values ​​at the edge; and extracts time-series features of cloud resource load based on the time queue of resource load values ​​at the cloud, including:

[0100] The feature extraction module 203 inputs the temporal features of edge resource load and cloud resource load into the 1D-CNN model to obtain a sequence of local temporal correlation feature vectors of edge resource load and a sequence of local temporal correlation feature vectors of cloud resource load;

[0101] The edge resource load temporal characteristics are obtained by performing cluster analysis on the sequence of local temporal correlation feature vectors of resource load; the cloud resource load temporal characteristics are obtained by performing cluster analysis on the sequence of local temporal correlation feature vectors of resource load.

[0102] Specifically, the feature enhancement aggregation representation processing, which involves clustering the sequence of local temporal correlation feature vectors of edge resource load to obtain the temporal features of edge resource load, is as follows:

[0103] Cluster analysis is performed on the sequence of local temporal correlation feature vectors of edge resource load to obtain a self-supervised clustering representation vector of local temporal correlation features of edge resource load; based on the self-supervised clustering representation vector of local temporal correlation features of edge resource load, feature clustering field modulation aggregation is performed on the sequence of local temporal correlation feature vectors of edge resource load to obtain a temporal significant aggregation representation vector of edge resource load, which is used as the temporal feature of edge resource load.

[0104] As a preferred embodiment, the feature extraction module 203 performs feature clustering field modulation aggregation on the sequence of the edge resource load local temporal correlation feature vector based on the self-supervised clustering representation vector of the edge resource load local temporal correlation feature, so as to obtain the edge resource load temporal significant aggregation representation vector, which serves as the edge resource load temporal feature, specifically:

[0105] The feature extraction module 203 calculates the implicit clustering contribution factor of each edge resource load local temporal correlation feature vector relative to the edge resource load local temporal correlation feature self-supervised clustering representation vector in the sequence of edge resource load local temporal correlation feature vectors, so as to obtain the sequence of implicit clustering contribution factors of edge resource load local temporal correlation features.

[0106] The sequence of latent clustering contribution factors of the local temporal correlation features of edge resource load is vectorized to obtain the distribution vector of the clustering contribution field of the local temporal correlation features of edge resource load.

[0107] The distribution vector of the cluster contribution field of the local temporal correlation feature of the edge resource load is input into the explicit modeling module of the cluster contribution field based on the self-attention mechanism to obtain the domain modulation weight vector of the cluster contribution field of the local temporal correlation feature of the edge resource load.

[0108] Using each feature value in the field modulation weight vector of the local temporal correlation feature clustering of edge resource load as a weight, the weighted sum of the corresponding positions of the sequence of the local temporal correlation feature vector of edge resource load is calculated to obtain the temporal significant aggregation representation vector of edge resource load.

[0109] As a preferred embodiment, the feature extraction module 203 calculates the implicit clustering contribution factor of each edge resource load local temporal correlation feature vector relative to the edge resource load local temporal correlation feature self-supervised clustering representation vector in the sequence of edge resource load local temporal correlation feature vectors, to obtain a sequence of implicit clustering contribution factors of edge resource load local temporal correlation features, specifically as follows:

[0110] The feature extraction module 203 calculates the absolute value of the division between the corresponding positions of the edge resource load local temporal correlation feature vector and the edge resource load local temporal correlation feature self-supervised clustering representation vector to obtain the edge resource load local temporal self-supervised clustering interaction correlation feature vector.

[0111] Calculate the logarithmic function value of each position feature value in the local temporal self-supervised clustering interaction association feature vector of the edge resource load with base 2, so as to obtain the local temporal self-supervised clustering interaction logarithmic representation vector of the edge resource load;

[0112] Using the feature values ​​at each position in the local temporal self-supervised clustering interaction logarithm representation vector of the edge resource load as weighting coefficients, the weighted sum of the feature values ​​at each position in the local temporal correlation feature vector of the edge resource load is calculated to obtain the clustering contribution representation factor of the local temporal correlation feature of the edge resource load.

[0113] The exponential function value of the clustering contribution factor of the local temporal correlation feature of the edge resource load, with the natural constant e as the base, is calculated to obtain the latent clustering contribution factor of the local temporal correlation feature of the edge resource load.

[0114] As a preferred embodiment, the feature extraction module 203 inputs the distribution vector of the clustering contribution field of the local temporal correlation feature of the edge resource load into the explicit modeling module of the clustering contribution field based on the self-attention mechanism to obtain the modulation weight vector of the local temporal correlation feature clustering contribution field of the edge resource load, specifically:

[0115] The feature extraction module 203 calculates the matrix multiplications between the edge resource load local temporal correlation feature clustering contribution field distribution vector and the query weight matrix, key weight matrix and value weight matrix, respectively, to obtain the edge resource load clustering contribution field distribution query vector, the edge resource load clustering contribution field distribution key vector and the edge resource load clustering contribution field distribution value vector;

[0116] Calculate the multiplication between the edge resource load clustering distribution query vector and the transpose of the edge resource load clustering contribution field distribution key vector. Use the result of the multiplication as the dividend and the square root of the length of the edge resource load clustering contribution field distribution key vector as the divisor. Perform outward processing according to the corresponding position to obtain the edge resource load clustering contribution field semantic interaction representation matrix.

[0117] The semantic interaction representation matrix of the edge resource load clustering contribution field is normalized by the Softmax function, and the product between the distribution value vector of the edge resource load clustering contribution field and the normalization result is calculated to obtain the domain modulation weight vector of the edge resource load local temporal correlation feature clustering contribution field.

[0118] As a preferred embodiment, the feature interaction module 204 performs detailed feature interaction based on spatial distance metrics, based on the time-series features of resource load at the edge and the time-series features of resource load in the cloud, to obtain resource load time-series collaborative features, specifically:

[0119] The feature interaction module 204 performs fine-grained feature decoupling on the temporal saliency aggregation representation vector of the edge resource load according to a preset scale to obtain a set of feature vectors of the temporal sub-components of the edge resource load; and performs fine-grained feature decoupling on the temporal saliency aggregation representation vector of the cloud resource load according to the preset scale to obtain a set of feature vectors of the temporal sub-components of the cloud resource load.

[0120] The set of feature vectors of temporal sub-components of resource load at the edge and the set of feature vectors of temporal sub-components of resource load in the cloud are subjected to feature sub-component pairing, filtering and interactive fusion processing to obtain resource load temporal collaborative features.

[0121] As a preferred embodiment, the feature interaction module 204 performs feature component pairing, filtering, and interactive fusion processing on the set of feature vectors of temporal sub-components of resource load at the edge and the set of feature vectors of temporal sub-components of resource load in the cloud to obtain resource load temporal collaborative features, specifically:

[0122] The feature interaction module 204 calculates the hyperbolic space distance metric between the feature vectors of the temporal sub-components of resource load at each edge end and the feature vectors of the temporal sub-components of resource load at each cloud end.

[0123] Based on the hyperbolic spatial distance metric factor, feature component pairing and filtering are performed on the feature vectors of temporal sub-components of resource load at each edge end and the feature vectors of temporal sub-components of resource load at each cloud end to obtain a pairing set;

[0124] The multi-dimensional interaction module of each paired input feature in the pairing set is used to obtain a set of feature vectors of load time series sub-component feature pairing interaction fusion.

[0125] The set of feature vectors of the load time-series sub-components are concatenated to obtain resource load time-series collaborative features.

[0126] As a preferred embodiment, the feature interaction module 204 calculates the hyperbolic space distance metric between the feature vectors of the temporal sub-components of resource load at each edge and the feature vectors of the temporal sub-components of resource load at each cloud, specifically as follows:

[0127] The feature interaction module 204 extracts the edge resource load time-series sub-component feature vectors at preset positions from the set of edge resource load time-series sub-component feature vectors;

[0128] Extract the cloud resource load time-series sub-component feature vectors at preset positions from the set of cloud resource load time-series sub-component feature vectors;

[0129] Calculate the difference vector between the edge resource load time-series sub-component feature vector at the preset location and the cloud resource load time-series sub-component feature vector at the preset location, and perform squaring based on the first norm of the difference vector to obtain the first norm representation of the semantic difference of the load time-series sub-component.

[0130] Calculate the square of the first norm of the feature vector of the temporal sub-component of the edge resource load at the preset position to obtain the first squared result; subtract the first constant from the first squared result according to the corresponding position to obtain the first norm representation of the temporal sub-component of the edge resource load.

[0131] Calculate the square of the first norm of the feature vector of the time series sub-component of cloud resource load at the preset location to obtain the second squared result; subtract the second constant from the second squared result according to the corresponding position to obtain the first norm representation of the time series sub-component of cloud resource load.

[0132] Calculate the multiplication between the first-norm representation of the temporal component of the edge resource load and the first-norm representation of the temporal component of the cloud resource load. Use the multiplication result as the divisor and the first-norm of the semantic difference of the temporal component of the load as the dividend. Perform division to obtain the semantic implicit association representation of the first-norm of the load temporal load.

[0133] The load time-series norm semantic implicit association representation is multiplied by the third constant, the product is added to the fourth constant, and the result of the addition is processed by the inverse hyperbolic cosine function to obtain the hyperbolic space distance metric factor.

[0134] As a preferred embodiment, the scheduling module 205 splits the task to be allocated according to the resource load timing coordination characteristics to obtain a first power calculation subtask to be allocated and a second power calculation subtask to be allocated, and schedules the first power calculation subtask to be allocated and the second power calculation subtask to the edge and the cloud respectively, specifically as follows:

[0135] The scheduling module 205 inputs the resource load time-series collaborative features into the classifier-based task splitting module to obtain the first power calculation subtask to be allocated and the second power calculation subtask to be allocated.

[0136] The first and second power calculation subtasks to be assigned are distributed to the edge and the cloud, respectively.

[0137] Additionally, please refer to FIG3, which shows a schematic diagram of the structure of an electronic device 600 suitable for implementing the embodiments of this application. The terminal devices in the embodiments of this application may include, but are not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. The electronic device shown in FIG3 is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0138] As shown in Figure 3, the electronic device 600 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. The RAM 603 also stores various programs and data required for the operation of the electronic device 600. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0139] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic device 600 to communicate wirelessly or wiredly with other devices to exchange data. Although FIG3 shows electronic device 600 with various devices, it should be understood that it is not required to implement or possess all the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0140] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a storage device 608, or installed from a ROM 602. When the computer program is executed by the processing device 601, it performs the functions defined in the methods of the embodiments of this application.

[0141] It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium, a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0142] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol, such as HTTP (Hypertext Transfer Protocol), and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0143] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0144] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0145] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0146] In the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0147] Figure 4 illustrates an application scenario of a cloud-edge collaborative task scheduling method according to an exemplary embodiment. As shown in Figure 4, in this application scenario, firstly, a power computing task to be allocated is obtained at the edge (shown as an edge gateway in the figure) (e.g., C1 as shown in Figure 4); a time queue of resource load values ​​at the edge is obtained (e.g., C2 as shown in Figure 4); a time queue of resource load values ​​at the cloud is obtained (e.g., C3 as shown in Figure 4); then, the obtained power computing task to be allocated, the time queue of resource load values ​​at the edge, and the time queue of resource load values ​​at the cloud are input into a server (e.g., S as shown in Figure 4) deployed with a cloud-edge collaborative task scheduling mechanism. The server can split the power computing task to be allocated based on the cloud-edge collaborative task scheduling to obtain a first power computing subtask and a second power computing subtask, and schedule the first and second power computing subtasks to the edge and the cloud respectively, so that the edge and the cloud execute the first and second power computing subtasks respectively.

[0148] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0149] This invention provides a task scheduling method and apparatus based on cloud-edge collaboration. By extracting the time-series characteristics of resource load at the edge and the cloud, taking into account the resource load conditions at both the cloud and the edge, and then performing detailed feature interaction based on spatial distance metrics, resource load time-series collaborative characteristics are obtained. This enables flexible task splitting and scheduling. Compared with existing technologies, this method can achieve load balancing of power computing tasks between the edge and the cloud, effectively integrate the advantages of the edge and the cloud, improve the overall stability of the system and the intelligence of task scheduling, and improve task processing efficiency.

[0150] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A task scheduling method based on cloud-edge collaboration, characterized in that, include: Obtain tasks to be assigned from the edge, the tasks to be assigned being power calculation tasks; The time queue for obtaining the resource load values ​​at the edge and the time queue for obtaining the resource load values ​​in the cloud; Based on the time queue of the resource load values ​​at the edge, extract the temporal features of the resource load at the edge. Based on the time queue of the resource load values ​​in the cloud, extract the time-series features of the cloud resource load; Based on the time-series characteristics of resource load at the edge and in the cloud, detailed feature interaction based on spatial distance metric is performed to obtain resource load time-series collaborative characteristics. Based on the resource load timing coordination characteristics, the task to be allocated is split into a first power calculation subtask and a second power calculation subtask. The first power calculation subtask and the second power calculation subtask are scheduled to the edge and the cloud respectively, so that the edge and the cloud can execute the first power calculation subtask and the second power calculation subtask respectively.

2. The task scheduling method based on cloud-edge collaboration as described in claim 1, characterized in that, The time queue based on the resource load value of the edge terminal is used to extract the temporal features of the resource load at the edge terminal; Based on the time-series queue of resource load values ​​in the cloud, extract the time-series features of cloud resource load, including: The temporal features of edge resource load and cloud resource load are input into the 1D-CNN model to obtain the sequence of local temporal correlation feature vectors of edge resource load and the sequence of local temporal correlation feature vectors of cloud resource load. The edge resource load temporal characteristics are obtained by performing cluster analysis on the sequence of local temporal correlation feature vectors of resource load; the cloud resource load temporal characteristics are obtained by performing cluster analysis on the sequence of local temporal correlation feature vectors of resource load. Specifically, the feature enhancement aggregation representation processing, which involves clustering the sequence of local temporal correlation feature vectors of edge resource load to obtain the temporal features of edge resource load, is as follows: Cluster analysis is performed on the sequence of local temporal correlation feature vectors of edge resource load to obtain a self-supervised clustering representation vector of local temporal correlation features of edge resource load; based on the self-supervised clustering representation vector of local temporal correlation features of edge resource load, feature clustering field modulation aggregation is performed on the sequence of local temporal correlation feature vectors of edge resource load to obtain a temporal significant aggregation representation vector of edge resource load, which is used as the temporal feature of edge resource load.

3. The task scheduling method based on cloud-edge collaboration as described in claim 2, characterized in that, The step of performing feature clustering field modulation aggregation on the sequence of the edge resource load local temporal correlation feature vector based on the self-supervised clustering representation vector of the edge resource load local temporal correlation feature vector to obtain the edge resource load temporal significant aggregation representation vector, which serves as the edge resource load temporal feature, specifically involves: Calculate the implicit clustering contribution factor of each edge resource load local temporal correlation feature vector relative to the edge resource load local temporal correlation feature self-supervised clustering representation vector in the sequence of edge resource load local temporal correlation feature vectors, so as to obtain the sequence of implicit clustering contribution factors of edge resource load local temporal correlation feature. The sequence of latent clustering contribution factors of the local temporal correlation features of edge resource load is vectorized to obtain the distribution vector of the clustering contribution field of the local temporal correlation features of edge resource load. The distribution vector of the cluster contribution field of the local temporal correlation feature of the edge resource load is input into the explicit modeling module of the cluster contribution field based on the self-attention mechanism to obtain the domain modulation weight vector of the cluster contribution field of the local temporal correlation feature of the edge resource load. Using each feature value in the field modulation weight vector of the local temporal correlation feature clustering of edge resource load as a weight, the weighted sum of the corresponding positions of the sequence of the local temporal correlation feature vector of edge resource load is calculated to obtain the temporal significant aggregation representation vector of edge resource load.

4. The task scheduling method based on cloud-edge collaboration as described in claim 3, characterized in that, The sequence of calculating the local temporal correlation feature vectors of edge resource loads, specifically, involves calculating the latent clustering contribution factor of each local temporal correlation feature vector of edge resource loads relative to the self-supervised clustering representation vector of the local temporal correlation feature of edge resource loads, to obtain the sequence of latent clustering contribution factors of the local temporal correlation feature of edge resource loads. Calculate the absolute value of the division between the corresponding positions of the edge resource load local temporal correlation feature vector and the edge resource load local temporal correlation feature self-supervised clustering representation vector to obtain the edge resource load local temporal self-supervised clustering interaction correlation feature vector; Calculate the logarithmic function value of each position feature value in the local temporal self-supervised clustering interaction association feature vector of the edge resource load with base 2, so as to obtain the local temporal self-supervised clustering interaction logarithmic representation vector of the edge resource load; Using the feature values ​​at each position in the local temporal self-supervised clustering interaction logarithm representation vector of the edge resource load as weighting coefficients, the weighted sum of the feature values ​​at each position in the local temporal correlation feature vector of the edge resource load is calculated to obtain the clustering contribution representation factor of the local temporal correlation feature of the edge resource load. The exponential function value of the clustering contribution factor of the local temporal correlation feature of the edge resource load, with the natural constant e as the base, is calculated to obtain the latent clustering contribution factor of the local temporal correlation feature of the edge resource load.

5. The task scheduling method based on cloud-edge collaboration as described in claim 4, characterized in that, The step of inputting the distribution vector of the clustering contribution field of the local temporal correlation features of the edge resource load into the explicit modeling module of the clustering contribution field based on the self-attention mechanism to obtain the modulation weight vector of the local temporal correlation features of the edge resource load clustering contribution field is as follows: Calculate the matrix multiplications between the distribution vector of the local temporal correlation feature clustering contribution field of the edge resource load and the query weight matrix, key weight matrix and value weight matrix, respectively, to obtain the query vector of the distribution vector of the clustering contribution field of the edge resource load, the key vector of the distribution vector of the clustering contribution field of the edge resource load and the value vector of the distribution vector of the clustering contribution field of the edge resource load; Calculate the multiplication between the edge resource load clustering distribution query vector and the transpose of the edge resource load clustering contribution field distribution key vector. Use the result of the multiplication as the dividend and the square root of the length of the edge resource load clustering contribution field distribution key vector as the divisor. Perform outward processing according to the corresponding position to obtain the edge resource load clustering contribution field semantic interaction representation matrix. The semantic interaction representation matrix of the edge resource load clustering contribution field is normalized by the Softmax function, and the product between the distribution value vector of the edge resource load clustering contribution field and the normalization result is calculated to obtain the domain modulation weight vector of the edge resource load local temporal correlation feature clustering contribution field.

6. The task scheduling method based on cloud-edge collaboration as described in claim 5, characterized in that, The method involves performing detailed feature interaction based on spatial distance metrics, using both edge resource load time-series characteristics and cloud resource load time-series characteristics, to obtain resource load time-series collaborative characteristics. Specifically: Fine-grained feature decoupling is performed on the temporal salient aggregation representation vector of edge resource load according to a preset scale to obtain a set of temporal sub-component feature vectors of edge resource load; fine-grained feature decoupling is also performed on the temporal salient aggregation representation vector of cloud resource load according to the preset scale to obtain a set of temporal sub-component feature vectors of cloud resource load. The set of feature vectors of temporal sub-components of resource load at the edge and the set of feature vectors of temporal sub-components of resource load in the cloud are subjected to feature sub-component pairing, filtering and interactive fusion processing to obtain resource load temporal collaborative features.

7. The task scheduling method based on cloud-edge collaboration as described in claim 6, characterized in that, The process of performing feature component pairing, filtering, and interactive fusion on the set of temporal sub-component feature vectors of resource load at the edge and the set of temporal sub-component feature vectors of resource load in the cloud to obtain temporal collaborative features of resource load is as follows: Calculate the hyperbolic space distance metric between the feature vectors of the temporal sub-components of resource load at each edge and the feature vectors of the temporal sub-components of resource load at each cloud. Based on the hyperbolic spatial distance metric factor, feature component pairing and filtering are performed on the feature vectors of temporal sub-components of resource load at each edge end and the feature vectors of temporal sub-components of resource load at each cloud end to obtain a pairing set; The multi-dimensional interaction module of each paired input feature in the pairing set is used to obtain a set of feature vectors of load time series sub-component feature pairing interaction fusion. The set of feature vectors of the load time-series sub-components are concatenated to obtain resource load time-series collaborative features.

8. The task scheduling method based on cloud-edge collaboration as described in claim 7, characterized in that, The hyperbolic space distance metric factor for calculating the time-series sub-component feature vectors of resource load at each edge and the time-series sub-component feature vectors of resource load at each cloud is specifically as follows: Extract the edge resource load time-series sub-component feature vectors at a preset position from the set of edge resource load time-series sub-component feature vectors; Extract the cloud resource load time-series sub-component feature vectors at preset positions from the set of cloud resource load time-series sub-component feature vectors; Calculate the difference vector between the edge resource load time-series sub-component feature vector at the preset location and the cloud resource load time-series sub-component feature vector at the preset location, and perform squaring based on the first norm of the difference vector to obtain the first norm representation of the semantic difference of the load time-series sub-component. Calculate the square of the first norm of the feature vector of the temporal sub-component of the edge resource load at the preset position to obtain the first squared result; subtract the first constant from the first squared result according to the corresponding position to obtain the first norm representation of the temporal sub-component of the edge resource load. Calculate the square of the first norm of the feature vector of the time series sub-component of cloud resource load at the preset location to obtain the second squared result; subtract the second constant from the second squared result according to the corresponding position to obtain the first norm representation of the time series sub-component of cloud resource load. Calculate the multiplication between the first-norm representation of the temporal component of the edge resource load and the first-norm representation of the temporal component of the cloud resource load. Use the multiplication result as the divisor and the first-norm of the semantic difference of the temporal component of the load as the dividend. Perform division to obtain the semantic implicit association representation of the first-norm of the load temporal load. The load time-series norm semantic implicit association representation is multiplied by the third constant, the product is added to the fourth constant, and the result of the addition is processed by the inverse hyperbolic cosine function to obtain the hyperbolic space distance metric factor.

9. A task scheduling method based on cloud-edge collaboration as described in claim 8, characterized in that, The step of splitting the task to be allocated based on the resource load timing coordination characteristics to obtain a first power calculation subtask and a second power calculation subtask to be allocated, and scheduling the first power calculation subtask and the second power calculation subtask to the edge and the cloud respectively, specifically: The resource load time-series collaborative features are input into the classifier-based task splitting module to obtain the first power calculation subtask to be allocated and the second power calculation subtask to be allocated. The first and second power calculation subtasks to be assigned are distributed to the edge and the cloud, respectively.

10. A task scheduling device based on cloud-edge collaboration, characterized in that, It includes a task acquisition module, a queue acquisition module, a feature extraction module, a feature interaction module, and a scheduling module; among which, The task acquisition module is used to acquire tasks to be assigned from the edge, and the tasks to be assigned are power calculation tasks; The queue acquisition module is used to acquire the time queue of resource load values ​​at the edge and the time queue of resource load values ​​in the cloud. The feature extraction module is used to extract time-series features of resource load at the edge end based on the time-series of resource load values ​​at the edge end; and to extract time-series features of resource load at the cloud end based on the time-series of resource load values ​​at the cloud end. The feature interaction module is used to perform detailed feature interaction based on spatial distance metric based on the time-series features of resource load at the edge and the time-series features of resource load in the cloud, so as to obtain resource load time-series collaborative features. The scheduling module is used to split the task to be allocated according to the resource load timing coordination characteristics to obtain a first power calculation subtask to be allocated and a second power calculation subtask to be allocated, and to schedule the first power calculation subtask to be allocated and the second power calculation subtask to be allocated to the edge and the cloud respectively, so that the edge and the cloud respectively execute the first power calculation subtask to be allocated and the second power calculation subtask to be allocated.

Images