Business load prediction method based on dual-channel heterogeneous graph convolution and related equipment

Abstract

The invention provides a service load prediction method based on dual-channel heterogeneous graph convolution and related equipment, and relates to the technical field of service load prediction, and the method comprises the steps: training a constructed dual-channel heterogeneous graph convolution network model through employing a cloud computing load data set for training, and obtaining a service load prediction model; inputting historical load data of all nodes of the target data center into a service load prediction model for prediction to obtain a service load grade prediction result of the target data center; the dual-channel heterogeneous graph convolution network model comprises a time causal adaptive convolution module, an information fusion module and a plurality of dual-channel heterogeneous graph convolution modules connected in sequence; and the service load prediction precision is improved.

Description

Technical Field

[0001] This invention relates to the field of load prediction technology, and in particular to a service load prediction method and related equipment based on dual-channel heterogeneous graph convolution. Background Technology

[0002] In data center and cloud computing environments, accurate workload forecasting is a core technological prerequisite for achieving efficient resource scheduling, optimizing energy consumption, and ensuring Service-Level Agreements (SLAs). However, the workload characteristics of modern data centers exhibit complex multidimensional spatiotemporal coupling features, posing a significant challenge to traditional forecasting methods.

[0003] In the spatial dimension, workloads heavily rely on the strong physical connections formed between computing nodes and network nodes due to shared physical resources (such as top-of-the-line switches). Traditional time-series models (such as Long Short-Term Memory networks, LSTM) completely ignore this spatial topology. Traditional graph neural networks, when modeling this topology, face a "precision-scale" dilemma: on the one hand, using simplified, heuristic adjacency matrices (such as rack-level connections) prevents the model from capturing precise node-level physical constraints; on the other hand, attempting to construct a precise, globally node-level "physical cost matrix" results in a dense affinity matrix, making its computational complexity extremely high, which is computationally infeasible in large data center environments.

[0004] Meanwhile, logically, a business process (such as a microservice request chain) typically requires collaborative processing across multiple heterogeneous nodes in multiple server racks, forming a high-order dependency relationship of "multi-point participation." However, traditional graph neural network architectures are limited by the representation paradigm of paired edges and lack a mechanism to directly model the many-to-many high-order dependencies between a "business load (hyperedge)" and its associated "multiple heterogeneous node sets." Even existing hypergraph modeling methods often use static hypergraph Laplacian operators, which are limited by the inability to treat the "hyperedge," i.e., the business load itself, as a dynamically learnable and updatable feature carrier.

[0005] Furthermore, most existing prediction models treat this problem as a standard regression task and use loss functions such as mean squared error. The fundamental flaw of this approach lies in its business insensitivity. From the perspective of resource scheduling business logic, the "level" of the prediction (e.g., "warning zone" or "danger zone") is far more important than the precise numerical value. Standard regression loss treats a harmless error (e.g., 10%~20%) as equally as a fatal error (e.g., 85%~95%) that leads to SLA violations. This fundamental mismatch between the optimization objective and business needs causes the model to perform poorly near key decision thresholds. Summary of the Invention

[0006] This invention provides a service load prediction method and related equipment based on dual-channel heterogeneous graph convolution, with the aim of improving the accuracy of service load prediction.

[0007] To achieve the above objectives, this invention provides a service load prediction method based on dual-channel heterogeneous graph convolution, applied to data centers. The service load prediction method includes: Step 1: Obtain the training cloud computing workload dataset; Step 2: Train the constructed dual-channel heterogeneous graph convolutional network model using the training cloud computing load dataset to obtain the business load prediction model. Step 3: Input the historical load data of all nodes in the target data center into the business load prediction model for prediction, and obtain the business load level prediction result of the target data center. The dual-channel heterogeneous graph convolutional network model includes a temporal causal adaptive convolutional module for extracting temporal features at different scales, an information fusion module for fusing spatial features and temporal features at different scales, and multiple sequentially connected dual-channel heterogeneous graph convolutional modules for extracting multi-scale spatial features. The input terminals of the first dual-channel heterogeneous graph convolutional module and the temporal causal adaptive convolutional module are both input terminals of the dual-channel heterogeneous graph convolutional network model. The output terminals of the last dual-channel heterogeneous graph convolutional module and the temporal causal adaptive convolutional module are both connected to the input terminal of the information fusion module, and the output terminal of the information fusion module is the output terminal of the dual-channel heterogeneous graph convolutional network model.

[0008] Furthermore, a dual-channel heterogeneous graph convolutional network model includes at least two sequentially connected dual-channel heterogeneous graph convolutional modules.

[0009] Furthermore, each dual-channel heterogeneous graph convolutional module includes a business logic channel, a physical topology channel, and a node-level attention fusion unit; The business logic channel dynamically updates the characteristics of hyperedges and nodes through a bipartite graph mechanism; Physical topological channels propagate the physical neighbor features of the nucleus polymer through sparse propagation; The node-level attention fusion unit fuses the features updated by the business logic channel and the features aggregated by the physical topology channel.

[0010] Furthermore, the business logic channel dynamically updates the characteristics of hyperedges and nodes through a bipartite graph mechanism, including: For each hyperedge, the features of all nodes contained in the hyperedge are aggregated to obtain node aggregated features, and the node aggregated features are used as new features of the hyperedge. For each node, the node aggregates the new features of the hyperedges containing that node to obtain the hyperedge aggregated features, and uses the hyperedge aggregated features as the new features of the node.

[0011] Furthermore, physical topological channels propagate core-polymer physical neighbor features through sparse propagation, including: Construct a sparse propagation kernel using a global physical cost matrix; For each node, the node aggregates the features of multiple nodes that are physically nearest to it through a sparse propagation kernel to obtain the physical neighbor features.

[0012] Furthermore, a sparse propagation kernel is constructed using the global physical cost matrix, including: Construct a global physical cost matrix based on the global cost vector of each node; The global physical cost matrix is ​​transformed into an affinity matrix using a Gaussian kernel. Perform a k-nearest neighbor search on each node to obtain multiple neighbor nodes with the lowest communication cost; Construct a sparse affinity matrix using the global cost vectors of all neighboring nodes with the lowest communication cost; Symmetric normalization of the sparse affinity matrix yields the sparse propagation kernel.

[0013] Furthermore, the temporal causal adaptive convolution module includes: A set of convolutional layers with different dilation rates is used to extract temporal features from the historical load data of each node, resulting in temporal features at multiple scales. For the temporal features at each scale, a multilayer perceptron is used to map the temporal features to raw scores, resulting in multiple raw scores. For each raw score, the softmax function is used to transform the raw score to obtain dynamic weights; The temporal features of each node are obtained by aggregating the temporal features of multiple scales with dynamic weights.

[0014] This invention also provides a service load prediction device based on dual-channel heterogeneous graph convolution, applied in a data center. The service load prediction device includes: The acquisition module is used to acquire the cloud computing workload dataset for training. The training module is used to train the constructed dual-channel heterogeneous graph convolutional network model using the training cloud computing load dataset to obtain the business load prediction model. The prediction module is used to input the historical load data of all nodes in the target data center into the business load prediction model for prediction, and obtain the business load level prediction result of the target data center. The dual-channel heterogeneous graph convolutional network model includes a temporal causal adaptive convolutional module for extracting temporal features at different scales, an information fusion module for fusing spatial features and temporal features at different scales, and multiple sequentially connected dual-channel heterogeneous graph convolutional modules for extracting multi-scale spatial features. The input terminals of the first dual-channel heterogeneous graph convolutional module and the temporal causal adaptive convolutional module are both input terminals of the dual-channel heterogeneous graph convolutional network model. The output terminals of the last dual-channel heterogeneous graph convolutional module and the temporal causal adaptive convolutional module are both connected to the input terminal of the information fusion module, and the output terminal of the information fusion module is the output terminal of the dual-channel heterogeneous graph convolutional network model.

[0015] The present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a service load prediction method based on dual-channel heterogeneous graph convolution.

[0016] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a service load prediction method based on dual-channel heterogeneous graph convolution.

[0017] The above-described solution of the present invention has the following beneficial effects: This invention trains a dual-channel heterogeneous graph convolutional network model using a training cloud computing load dataset to obtain a business load prediction model. Historical load data from all nodes in the target data center are then input into the business load prediction model for prediction, yielding a business load level prediction result for the target data center. The dual-channel heterogeneous graph convolutional network model includes a temporal causal adaptive convolutional module, an information fusion module, and multiple sequentially connected dual-channel heterogeneous graph convolutional modules. Compared to existing technologies, this invention utilizes multiple dual-channel heterogeneous graph convolutional modules to achieve multi-hop propagation of information on the heterogeneous graph to extract multi-scale spatial features. It also utilizes the temporal causal adaptive convolutional module to extract temporal features of different scales from historical load data. Finally, the information fusion module fuses the spatial and temporal features of different scales to obtain the business load level prediction result for the target data center, thus improving the accuracy of business load prediction.

[0018] Other beneficial effects of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the dual-channel heterogeneous graph convolutional network model in an embodiment of the present invention; Figure 3This is a schematic diagram of the structure of the dual-channel heterogeneous graph convolution module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the temporal causal adaptive convolution module in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the service load prediction device in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of the terminal device in an embodiment of the present invention. Detailed Implementation

[0020] To make the technical problems, solutions, and advantages of this invention clearer, a detailed description will be provided below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0022] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a locking connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0024] This invention addresses existing problems by providing a service load prediction method and related equipment based on dual-channel heterogeneous graph convolution.

[0025] like Figure 1 As shown, embodiments of the present invention provide a service load prediction method based on dual-channel heterogeneous graph convolution, applied to data centers. The service load prediction method includes: Step 1: Obtain the training cloud computing workload dataset; Step 2: Train the constructed dual-channel heterogeneous graph convolutional network model using the training cloud computing load dataset to obtain the business load prediction model. Step 3: Input the historical load data of all nodes in the target data center into the business load prediction model for prediction, and obtain the business load level prediction result of the target data center. The business load level prediction result of the target data center is the probability distribution of the business load in a discrete state interval at future time. For example, when the historical load data of all nodes in the target data center is the historical load data of all physical nodes involved in the online payment business, the business load level prediction result is the probability that the online payment business is in the safe zone, the warning zone, and the danger zone respectively.

[0026] In this embodiment of the invention, the training cloud computing workload dataset includes CPU utilization, memory usage, network throughput, and disk I / O, which are simulated and synthesized by scripts.

[0027] Specifically, such as Figure 2 As shown, the dual-channel heterogeneous graph convolutional network model includes a temporal causal adaptive convolutional module for extracting temporal features at different scales, an information fusion module for fusing spatial features and temporal features at different scales, and multiple sequentially connected dual-channel heterogeneous graph convolutional modules for extracting multi-scale spatial features. The input terminals of the first dual-channel heterogeneous graph convolutional module and the temporal causal adaptive convolutional module are both input terminals of the dual-channel heterogeneous graph convolutional network model. The output terminals of the last dual-channel heterogeneous graph convolutional module and the temporal causal adaptive convolutional module are both connected to the input terminal of the information fusion module, and the output terminal of the information fusion module is the output terminal of the dual-channel heterogeneous graph convolutional network model.

[0028] The optimal choice is, such as Figure 3 As shown, each dual-channel heterogeneous graph convolutional module includes a business logic channel, a physical topology channel, and a node-level attention fusion unit; The business logic channel dynamically updates the characteristics of hyperedges and nodes through a bipartite graph mechanism; Physical topological channels propagate the physical neighbor features of the nucleus polymer through sparse propagation; The node-level attention fusion unit fuses the features updated by the business logic channel and the features aggregated by the physical topology channel.

[0029] In cloud computing scenarios, resource scheduling must consider two types of relationships simultaneously: one is the topological coupling between nodes due to physical adjacency and shared resources, i.e., pairing relationships; the other is the logical association of cross-node collaborative processing of a business load, i.e., multi-point participation relationships. Therefore, the heterogeneous hypergraph structure constructed in this embodiment of the invention includes hyperedges and pairing edges to model these two relationships within the same framework.

[0030] In this embodiment of the invention, a business logic channel is used in the dual-channel heterogeneous graph convolution module to model the multi-point participation relationship. The business logic channel dynamically updates the features of hyperedges and nodes through a bipartite graph mechanism, including: For each hyperedge, the features of all nodes contained in the hyperedge are aggregated to obtain node aggregated features, and the node aggregated features are used as new features of the hyperedge. For each node, the node aggregates the new features of the hyperedges containing that node to obtain the hyperedge aggregated features, and uses the hyperedge aggregated features as the new features of the node.

[0031] Specifically, the process of the bipartite graph mechanism is as follows: Using two typed correlation matrices and Let's record the membership relationships between computing power nodes, network nodes, and hyperedges respectively. Then, stack the two correlation matrices vertically to form a complete node-hyperedge correlation matrix, expressed as: ; SuperMap Viewed as a bipartite graph, its two sides are... Each node and Each hyperedge will form a node hyperedge association matrix. As the adjacency matrix of a bipartite graph; To support subsequent dynamic bipartite graph convolution, this embodiment of the invention defines two additional structure matrices, namely: Node degree matrix : ; The node degree matrix is a diagonal matrix; Hypermarginal matrix : ; The hyperedge degree matrix is A diagonal matrix.

[0032] In this embodiment of the invention, for each hyperedge, the features of all nodes contained within the hyperedge are aggregated to obtain the expression for the aggregated node features: ; in, The node aggregation feature, i.e., the new feature of the hyperedge, has a dimension of . , Representing feature dimension, This indicates the activation function for updating the superedge. This indicates that the hyperedge degree matrix is ​​normalized. Represents the learnable weights used for feature transformation. The features of a node are represented by the dimension . , Represents the feature dimension of a node.

[0033] In this embodiment of the invention, for each node, the features of the hyperedges containing that node are aggregated to obtain the expression for the aggregated hyperedge features: ; in, This represents the hyperedge aggregation feature, i.e., the new feature of the node. This indicates that the node degree matrix is ​​normalized.

[0034] Specifically, physical topological channels propagate core-polymer physical neighbor features through sparse propagation, including: Construct a sparse propagation kernel using a global physical cost matrix; For each node, the features of multiple physically nearest neighbors are aggregated using a sparse propagation kernel to obtain the expression for the physical neighbor features: ; in, Represents physical neighbor features, with dimension 1. , This indicates a sparsely propagating nucleus.

[0035] Specifically, a sparse propagation kernel is constructed using the global physical cost matrix, including: Construct a global physical cost matrix based on the global cost vector of each node; The global physical cost matrix is ​​transformed into an affinity matrix using a Gaussian kernel. Perform a k-nearest neighbor search on each node to obtain multiple neighbor nodes with the lowest communication cost; Construct a sparse affinity matrix using the global cost vectors of all neighboring nodes with the lowest communication cost; Symmetric normalization of the sparse affinity matrix yields the sparse propagation kernel.

[0036] In this embodiment of the invention, before constructing the global physical cost matrix based on the global cost vector of each node, the method further includes: Define basic parameters: Total number of racks The number of computing nodes per rack is In this embodiment of the invention, a tail switch is configured as a central network node in the data center. Network node set Size is Therefore, the total number of nodes ; Define basic cost input: Inter-rack cost matrix Dimensions are , Indicates from the rack to the rack Macro-level communication costs, cost vector per rack Dimensions are , Indicates server rack The Middle The cost of accessing each node; Constructing a global auxiliary input: a global cost vector :Will indivual vector( Vertically stacked, output dimension is Global all-1 column vector Dimensions are A matrix of all 1s Dimensions are .

[0037] In this embodiment of the invention, the global physical cost matrix is ​​a block matrix, which is constructed as follows: ; in, The communication cost between computing nodes is expressed as follows: ; in, Indicates leaving the node Communication costs, Indicates arrival at node Communication costs, Indicates the Kronecker product. Used to create a cost matrix for server racks Zoom in Dimension This indicates a correction term used to ensure that when hour, ; Used to represent all computing power nodes to a unique network node. The cost, in this embodiment of the invention, is... This cost is defined using vectors because It itself represents the access cost of each computing node; for transpose; The cost from a network node to itself.

[0038] In this embodiment of the invention, the expression for converting the global physical cost matrix into an affinity matrix using a Gaussian kernel is as follows: ; in, Represents the affinity matrix. Indicates to Perform element-wise squaring. This represents hyperparameters.

[0039] Due to the affinity matrix It is The dense matrix is ​​used because the Gaussian kernel converts all non-infinite costs into non-zero affinity. Therefore, directly using this dense matrix for graph convolution in GNNs will lead to... The computational and storage complexity is prohibitive in large data center environments. To address this scalability bottleneck, this invention introduces sparsification of the k-nearest neighbor affinity matrix, as detailed below: Perform a k-nearest neighbor search on each node to obtain multiple neighbor nodes with the lowest communication cost; Construct a sparse affinity matrix using the global cost vectors of all neighboring nodes with the lowest communication cost; Symmetric normalization of the sparse affinity matrix yields the sparse propagation kernel, expressed as: ; in, Indicates sparsely propagating nuclei, The degree matrix represents the sparse affinity matrix.

[0040] It should be noted that when the node It is a node of If one of the nearest neighbors is a given, then the sparse affinity matrix is: ; otherwise, .

[0041] In this embodiment of the invention, the node-level attention fusion unit is used to perform dynamic attention fusion on the features output by the business logic channel and the physical topology channel. The specific process is as follows: Using two independent sets of learnable weights , The features from the two channels are transformed to a unified node feature dimension to obtain the transformed features. , The expression is: ; ; An attention network is used to calculate the importance scores of the two channels based on the transformed features. , The expression is: ; ; Then, the importance score is converted into a normalized attention weight using the softmax function. and ensure The conversion expression is: ; ; Finally, the transformed features are fused using normalized attention weights to obtain the spatial features of the historical business load data. The calculation expression is as follows: ; in, This indicates the node update activation function.

[0042] Ideally, the dual-channel heterogeneous graph convolutional network model includes at least two sequentially connected dual-channel heterogeneous graph convolutional modules.

[0043] This invention adopts the standard paradigm of modern deep graph neural networks: spatial multi-hop expansion is achieved by stacking at least two dual-channel heterogeneous graph convolutional modules, as formalized as follows: ; in, and These are the initial node and hyperedge features (e.g., features from historical time steps or learnable embeddings). Represents a two-channel heterogeneous convolution operation (containing the independent learnable weights of this layer). ); Stacking at least two layers allows features to propagate at least two hops across a heterogeneous graph (including sparse physical pairs and logical hyperedges), resulting in the final multi-scale spatial features. It can be the second layer The output, or all intermediate layer features Aggregation (e.g., splicing or summation): or .

[0044] In order to capture multi-scale dynamic features over time, embodiments of the present invention have designed, as follows: Figure 4 The temporal causal adaptive convolution module shown includes: A set of convolutional layers with different dilation rates is used to extract temporal features from the historical load data of each node, resulting in temporal features at multiple scales. The calculation expression is: ; in, It is the kernel size. Indicates the first One convolutional kernel, It is the internal index of the convolution kernel; It is the core weight. This ensures the causality (depending only on the past) and extensibility (skip sampling) of convolutions. For each scale of temporal features, a multilayer perceptron is used to map the temporal features to raw scores, resulting in multiple raw scores. The expression is: ; For each raw score, the raw score is transformed using the softmax function to obtain dynamic weights. The transformation expression is as follows: ; The aggregation layer aggregates the temporal features of multiple scales with dynamic weights to obtain the temporal features of each node. The aggregation expression is as follows: .

[0045] Specifically, the information fusion module is used to aggregate node features and output them probabilistically. The specific process is as follows: First, the multi-scale spatial features output by the dual-channel heterogeneous graph convolution module and the multi-scale temporal features output by the temporal causal adaptive convolution module are concatenated and fused. After linear transformation and layer normalization, the node features are obtained, and the calculation expression is as follows: ; Subsequently, for each hyperedge A pooling layer is used to aggregate all nodes (sets) participating in the hyperedge. ) characteristics Thus, the superedge is obtained. Features : ; The pooling function mentioned above can be either mean pooling or weighted mean pooling, and the node set... That is, participate in the super-edge All nodes; Features after obtaining each hyperedge aggregation Then, using a multilayer perceptron and a softmax function, the probability of belonging to each ordered load level is output, assuming the load is divided into... An ordered hierarchy ( Its central value is ; ; Based on this probability distribution Calculate the expected value of the prediction. As a continuous predicted value: ; This design enables information to be transferred from the physical space of the underlying nodes. Mapped to the abstract space of top-level business .

[0046] In order for the model to fit numerical values ​​and perceive the sequential relationship and differences between different load levels, a loss function that is more in line with business logic than the standard regression loss is needed. Therefore, this embodiment of the invention uses cross-entropy as the loss function, and the specific steps are as follows: First, the actual continuous load value is determined based on a preset threshold. Discretization, determining its target level. ; Then, construct a one-hot encoded target distribution vector. ; The main objective of model optimization is to minimize the predicted probability distribution of the output. With the true distribution Cross-entropy loss between : ; In certain scenarios where extremely high accuracy in predicting continuous values ​​is required, an auxiliary L1 loss term can be introduced. This loss term serves as the expected value of the prediction. and true continuous values Flexible constraints between them: ; ; in, It is a trade-off factor, which can be considered an optional configuration. Embodiments of this invention allow the model to focus on the classification task (i.e., ).

[0047] To verify the effectiveness of the method provided in the embodiments of the present invention, the method provided in the embodiments of the present invention is compared with an existing baseline model, as follows: Setting up the experimental environment Hardware: GPU-1 nvidia-rtx-4090-24GB; Experimental platform: Bitahub cloud server; System version: PyTorch: 2.5.1-cuda12.4-cudnn9-py311-ubuntu22.04; Experimental data: dataset.npz (visualization can be found in cloudworkload.csv), a cloud computing load-related dataset synthesized by script simulation. Its features include four types of features: CPU utilization, memory utilization, network throughput, and disk I / O. It is generated on a physical topology consisting of 8×16 compute nodes and 1 network node, with a total of 200 time steps.

[0048] The experimental results are compared with the baseline model, as shown in Table 1 below: Table 1. Comparison of experimental results with baseline model ; The experimental results are compared and analyzed as follows: Table 1 shows the performance comparison of the model provided in this embodiment of the invention with six mainstream baseline models on the test set. The experimental results show that different categories of models exhibit obvious performance stratification, as detailed below: The significant advantages of the model provided in this embodiment of the invention: The model provided in this embodiment of the invention achieves optimal results in both the mean absolute error and root mean square error (RMSE). Specifically, the mean absolute error of the model provided in this embodiment of the invention is 1.0009, and the RMSE is 1.2003. Compared with the best-performing baseline model (gated recurrent neural network), the model provided in this embodiment of the invention reduces the mean absolute error by approximately 25.8% and the RMSE by a remarkable 48.9%. The significant reduction in RMSE indicates that the model provided in this embodiment of the invention has extremely strong robustness in handling extreme load changes and effectively avoids drastic fluctuations in predicted values.

[0049] Effectiveness of Deep Learning Models: Observation of baseline models shows that gated recurrent neural networks and long short-term memory networks significantly outperform temporal convolutional networks and gradient boosting regression models. This indicates that the dataset contains complex temporal dependencies, requiring the memory mechanism of recurrent neural networks to capture them. However, the model provided in this embodiment of the invention further mines deep features by modeling in accordance with actual physical scenes and introducing hypergraphs, breaking through the performance bottleneck of standard RNNs.

[0050] Limitations of traditional methods: Autoregressive moving average models and gradient boosting regression models perform relatively poorly. The root mean square error of the autoregressive moving average model is as high as 3.1581, indicating that the load data has high nonlinearity and non-stationarity, and the traditional linear assumption is no longer applicable.

[0051] This invention utilizes a training cloud computing load dataset to train a constructed dual-channel heterogeneous graph convolutional network model, obtaining a business load prediction model. Historical load data from all nodes of the target data center are input into the business load prediction model for prediction, yielding a business load level prediction result for the target data center. The dual-channel heterogeneous graph convolutional network model includes a temporal causal adaptive convolutional module, an information fusion module, and multiple sequentially connected dual-channel heterogeneous graph convolutional modules. Compared to existing technologies, this invention utilizes multiple dual-channel heterogeneous graph convolutional modules to achieve multi-hop propagation of information on the heterogeneous graph to extract multi-scale spatial features. It also utilizes the temporal causal adaptive convolutional module to extract temporal features of different scales from historical load data. Finally, the information fusion module fuses the spatial and temporal features of different scales to obtain the business load level prediction result for the target data center, thus improving the accuracy of business load prediction.

[0052] Corresponding to the service load prediction method based on dual-channel heterogeneous graph convolution described in the above embodiments, such as Figure 5 As shown, this embodiment of the invention also provides a service load prediction device 100 based on dual-channel heterogeneous graph convolution, applied in a data center. The service load prediction device 100 includes: Module 101 is used to acquire the cloud computing workload dataset for training. Training module 102 is used to train the constructed dual-channel heterogeneous graph convolutional network model using the training cloud computing load dataset to obtain the business load prediction model. Prediction module 103 is used to input the historical load data of all nodes in the target data center into the business load prediction model for prediction, and obtain the business load level prediction result of the target data center. The dual-channel heterogeneous graph convolutional network model includes a temporal causal adaptive convolutional module for extracting temporal features at different scales, an information fusion module for fusing spatial features and temporal features at different scales, and multiple sequentially connected dual-channel heterogeneous graph convolutional modules for extracting multi-scale spatial features. The input terminals of the first dual-channel heterogeneous graph convolutional module and the temporal causal adaptive convolutional module are both input terminals of the dual-channel heterogeneous graph convolutional network model. The output terminals of the last dual-channel heterogeneous graph convolutional module and the temporal causal adaptive convolutional module are both connected to the input terminal of the information fusion module, and the output terminal of the information fusion module is the output terminal of the dual-channel heterogeneous graph convolutional network model.

[0053] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0054] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0055] This invention also provides a terminal device, such as... Figure 6 As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 6 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100. When the processor D100 executes the computer program D102, it implements the above-described service load prediction method based on dual-channel heterogeneous graph convolution.

[0056] The terminal device D10 can be a desktop computer, laptop, handheld computer, server, server cluster, or cloud server, etc. This terminal device may include, but is not limited to, a processor D100 and a memory D101. Those skilled in the art will understand that... Figure 6 This is merely an example of terminal device D10 and does not constitute a limitation on terminal device D10. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0057] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0058] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0059] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0060] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0061] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a service load prediction method based on dual-channel heterogeneous graph convolution.

[0062] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a building device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0063] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for service load prediction based on dual-channel heterogeneous graph convolution, characterized in that, The business load prediction method is applied to a data center and comprises the following steps: Step 1, obtaining a training cloud computing load data set; Step 2, training a double-channel heterogeneous graph convolution network model constructed by using the training cloud computing load data set to obtain a business load prediction model; Step 3, inputting historical load data of all nodes of a target data center into the business load prediction model to obtain a business load level prediction result of the target data center. The double-channel heterogeneous graph convolution network model comprises a time causal adaptive convolution module for extracting time features of different scales, an information fusion module for fusing spatial features and time features of different scales, and a plurality of double-channel heterogeneous graph convolution modules connected in sequence for extracting multi-scale spatial features, wherein an input end of the first double-channel heterogeneous graph convolution module and an input end of the time causal adaptive convolution module are input ends of the double-channel heterogeneous graph convolution network model, an output end of the last double-channel heterogeneous graph convolution module and an output end of the time causal adaptive convolution module are connected to an input end of the information fusion module, and an output end of the information fusion module is an output end of the double-channel heterogeneous graph convolution network model.

2. The method of claim 1, wherein the method is based on a dual-channel heterogeneous graph convolution. The double-channel heterogeneous graph convolution network model comprises at least two double-channel heterogeneous graph convolution modules connected in sequence.

3. The method of claim 2, wherein the method is based on a dual-channel heterogeneous graph convolution. Each double-channel heterogeneous graph convolution module comprises a business logic channel, a physical topology channel and a node-level attention fusion unit. The business logic channel dynamically updates features of hyper-edges and nodes through a two-part graph mechanism. The physical topology channel aggregates physical neighbor features through a sparse propagation kernel. The node-level attention fusion unit fuses the updated features of the business logic channel and the aggregated features of the physical topology channel.

4. The method of claim 3, wherein the method is based on a dual-channel heterogeneous graph convolution. The business logic channel dynamically updates features of hyper-edges and nodes through a two-part graph mechanism, comprising: For each hyper-edge, the hyper-edge aggregates features of all nodes contained in the hyper-edge to obtain node aggregated features, and the node aggregated features are taken as new features of the hyper-edge; For each node, the node aggregates new features of hyper-edges containing the node to obtain hyper-edge aggregated features, and the hyper-edge aggregated features are taken as new features of the node.

5. The method of claim 3, wherein the method is based on a dual-channel heterogeneous graph convolution. The physical topology channel aggregates physical neighbor features through a sparse propagation kernel, comprising: A sparse propagation kernel is constructed through a global physical cost matrix; For each node, the node aggregates node features of a plurality of physically nearest neighbors of the node through the sparse propagation kernel to obtain physical neighbor features.

6. The method of claim 5, wherein the method is based on a dual-channel heterogeneous graph convolution. The sparse propagation kernel is constructed through a global physical cost matrix, comprising: A global physical cost matrix is constructed based on a global cost vector of each node; The global physical cost matrix is converted into an affinity matrix through a Gaussian kernel; k-nearest neighbor search is performed on each node to obtain a plurality of neighbor nodes with the lowest communication cost; A sparse affinity matrix is constructed using global cost vectors of all neighbor nodes with the lowest communication cost; The sparse affinity matrix is symmetrically normalized to obtain the sparse propagation kernel.

7. The method of claim 2, wherein the method is based on a dual-channel heterogeneous graph convolution. The time-causal adaptive convolution module comprises: extracting time features from historical load data of each node by using a set of convolution layers with different expansion rates to obtain time features of multiple scales; respectively for each scale of time features, mapping the time features to original scores by a multi-layer perception to obtain multiple original scores; respectively for each original score, converting the original score by a softmax function to obtain a dynamic weight; aggregating the time features of multiple scales and the dynamic weight to obtain the time features of each node.

8. A service load prediction device based on dual-channel heterogeneous graph convolution, characterized by, The business load prediction device applied to a data center comprises: an acquisition module configured to acquire a training cloud computing load dataset; a training module configured to train a constructed dual-channel heterogeneous graph convolution network model by using the training cloud computing load dataset to obtain a business load prediction model; a prediction module configured to input historical load data of all nodes of a target data center into the business load prediction model for prediction to obtain a business load level prediction result of the target data center. The dual-channel heterogeneous graph convolution network model comprises a time-causal adaptive convolution module for extracting time features of different scales, an information fusion module for fusing spatial features and time features of different scales, and multiple dual-channel heterogeneous graph convolution modules connected in sequence for extracting multi-scale spatial features, wherein an input end of the first dual-channel heterogeneous graph convolution module and an input end of the time-causal adaptive convolution module are both input ends of the dual-channel heterogeneous graph convolution network model, an output end of the last dual-channel heterogeneous graph convolution module and an output end of the time-causal adaptive convolution module are both connected to an input end of the information fusion module, and an output end of the information fusion module is an output end of the dual-channel heterogeneous graph convolution network model.

9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the business load prediction method based on dual-channel heterogeneous graph convolution as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the business load prediction method based on dual-channel heterogeneous graph convolution as claimed in any one of claims 1 to 7.

Images