Method and device for predicting computing power, nonvolatile storage medium and electronic equipment

Abstract

The invention discloses a computing power prediction method and device, a nonvolatile storage medium and electronic equipment. The method comprises the steps that historical computing power data of a target node within a preset historical time period is acquired, and the target node is an equipment node in a computing power network; performing feature extraction on the historical computing power data of the target node to obtain spatial features of the historical computing power data of the target node and time features of the historical computing power data of the target node; fusing the spatial feature of the target node and the time feature of the target node into a spatial-temporal feature of the target node; and predicting the congestion condition of the computing power network at the target node according to the spatial-temporal characteristics of the target node. The technical problem that the equipment congestion condition of the computing power network cannot be accurately predicted in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the field of electronic information technology, and in particular to a method, device, non-volatile storage medium and electronic device for predicting computing power. Background Art

[0002] With the rapid development of cloud computing and edge computing, mobile devices, IoT terminals and emerging applications are also increasing rapidly. Data demand is dynamic and diversified, which brings huge challenges to network management. Computing network achieves deep integration of computing and network resources by deploying distributed computing and storage resources in the network, providing low latency, high bandwidth and flexible computing service support for end users. However, with the increase in the number of users and the complexity of computing tasks, the volatility and burstiness of network resource demand are more significant, resulting in increased pressure on resource allocation, which may cause network congestion and service quality degradation. The traffic and business demand in the computing network have significant spatiotemporal characteristics. These changes are not only affected by time factors (such as day and night cycles and peak fluctuations), but also closely related to spatial factors (such as concentrated resource demand in hot spots). Traditional network management methods fail to effectively utilize these spatiotemporal characteristics for computing power prediction, which makes it difficult to effectively solve the congestion problem during peak traffic or sudden demand.

[0003] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0004] The embodiments of the present invention provide a method, device, non-volatile storage medium and electronic device for predicting computing power, so as to at least solve the technical problem that the prior art cannot accurately predict the equipment congestion of the computing power network.

[0005] According to one aspect of an embodiment of the present invention, a method for predicting computing power is provided, comprising: obtaining historical computing power data of a target node within a predetermined historical time period, wherein the target node is a device node in a computing power network; performing feature extraction on the historical computing power data of the target node to obtain spatial features of the historical computing power data of the target node, and temporal features of the historical computing power data of the target node; fusing the spatial features of the target node and the temporal features of the target node into spatiotemporal features of the target node; and predicting congestion of the computing power network at the target node based on the spatiotemporal features of the target node.

[0006] Optionally, feature extraction is performed on the historical computing power data of the target node to obtain spatial features of the historical computing power data of the target node, including: obtaining the connection relationship between device nodes in the computing power network; determining multiple neighbor nodes adjacent to the target node based on the connection relationship; obtaining the historical computing power data of each of the multiple neighbor nodes within a historical time period; inputting the historical computing power data of the target node and the historical computing power data of each of the multiple neighbor nodes into a target spatial attention model, and the target spatial attention model outputs the spatial features of the target node, wherein the target spatial attention model is trained by a first training sample, and the first training sample is the historical computing power data of the nodes in the computing power network, and the spatial features corresponding to the nodes in the computing power network.

[0007] Optionally, the historical computing power data of the target node and the historical computing power data of multiple neighboring nodes are input into a target space attention model, and the target space attention model outputs the spatial characteristics of the target node, including: inputting the historical computing power data of the target node and the historical computing power data of multiple neighboring nodes into the target space attention model, and the target space attention model determines the attention weights of the multiple neighboring nodes, wherein the attention weights represent the degree of influence of the neighboring nodes on the computing power of the target node; determining the spatial characteristics of the target node according to the attention weights of the multiple neighboring nodes, combined with the historical computing power data of the multiple neighboring nodes and the historical computing power data of the target node.

[0008] Optionally, feature extraction is performed on the historical computing power data of the target node to obtain the time features of the historical computing power data of the target node, including: inputting the historical computing power data of the target node into a target long short-term memory network, and outputting the time features of the historical computing power data of the target node by the target long short-term memory network, wherein the target convolutional long short-term memory network is obtained by training with a second training sample, and the second training sample is the historical computing power data of the nodes in the computing power network that change over time, and the time features corresponding to the nodes in the computing power network.

[0009] Optionally, the spatial features of the target node and the temporal features of the target node are fused into the spatiotemporal features of the target node, including: inputting the spatial features of the target node and the temporal features of the target node into a target fusion model, and the target fusion model divides the spatial features of the target node into multiple subspaces to obtain respective subspace features of the multiple subspaces; mapping the temporal features of the target node to multiple subspaces to obtain respective sub-temporal features of the multiple subspaces; and fusing the respective subspace features of the multiple subspaces with the respective sub-temporal features of the multiple subspaces to obtain the spatiotemporal features of the target node.

[0010] Optionally, the subspace features of each of the multiple subspaces and the sub-time features of each of the multiple subspaces are fused to obtain the spatiotemporal features of the target node, including: multiplying the subspace features and the sub-time features of each of the multiple subspaces in the multiple subspaces respectively to obtain the sub-spatiotemporal features of each of the multiple subspaces; and fusing the sub-spatiotemporal features of each of the multiple subspaces to obtain the spatiotemporal features of the target node.

[0011] According to another aspect of an embodiment of the present invention, there is provided a device for predicting computing power, comprising: an acquisition module for acquiring historical computing power data of a target node within a predetermined historical time period, wherein the target node is a device node in a computing power network; an extraction module for performing feature extraction on the historical computing power data of the target node to obtain spatial features of the historical computing power data of the target node, and temporal features of the historical computing power data of the target node; a fusion module for fusing the spatial features of the target node and the temporal features of the target node into spatiotemporal features of the target node; and a prediction module for predicting the congestion of the computing power network at the target node based on the spatiotemporal features of the target node.

[0012] According to another aspect of an embodiment of the present invention, a non-volatile storage medium is provided, wherein the non-volatile storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing any one of the methods for predicting computing power.

[0013] According to another aspect of an embodiment of the present invention, there is provided an electronic device, comprising: one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors are enabled to implement any one of the predictive computing power methods.

[0014] According to yet another aspect of an embodiment of the present invention, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the computer program implements any one of the above-mentioned methods for predicting computing power.

[0015] In an embodiment of the present invention, historical computing power data of a target node within a predetermined historical time period is obtained, wherein the target node is a device node in a computing power network; feature extraction is performed on the historical computing power data of the target node to obtain spatial features of the historical computing power data of the target node and temporal features of the historical computing power data of the target node; the spatial features of the target node and the temporal features of the target node are merged into the spatiotemporal features of the target node; and the congestion of the computing power network at the target node is predicted based on the spatiotemporal features of the target node, thereby solving the technical problem that the prior art cannot accurately predict the device congestion of the computing power network, achieving the purpose of comprehensively considering the factors of the computing power network in the spatial dimension and the temporal dimension, and then predicting the congestion of the device nodes of the computing power network, thereby achieving the technical effect of improving the accuracy of the prediction of the congestion of the device nodes of the computing network. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0017] Figure 1 is a flowchart of a method for predicting computing power provided according to an embodiment of the present invention;

[0018] Figure 2 is a schematic diagram of an optional computing power network structure provided according to an embodiment of the present invention;

[0019] Figure 3 is a schematic diagram of deployment of an optional decision algorithm provided in an embodiment of the present invention on a computing power network;

[0020] Figure 4 is an optional flowchart of predicting computing power through an attention mechanism according to an embodiment of the present invention;

[0021] Figure 5 is a schematic diagram of a device for predicting computing power provided according to an embodiment of the present invention;

[0022] Figure 6 is a schematic diagram of an electronic device provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] According to an embodiment of the present invention, a method embodiment for predicting computing power is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0026] Figure 1 is a flowchart of a method for predicting computing power according to an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0027] Step S102, obtaining historical computing power data of a target node within a predetermined historical time period, wherein the target node is a device node in the computing power network;

[0028] In this step, if Figure 2As shown in the figure, the computing network architecture includes the network foundation layer, the network prediction layer and the network decision layer. The network foundation layer is the bottom layer of the architecture, which is composed of various devices of the computing network. According to factors such as the geographical location of the nodes in the network, the computing network is divided into multiple regions (only the target region is shown in the architecture diagram). Each region can be regarded as an autonomous network, consisting of multiple network nodes. Each domain is managed by the Non-RT RIC (non-real-time RAN intelligent controller) and Near-RT RIC (near real-time RAN intelligent controller) of the O-RAN (Open Radio Access Network) architecture. The division of regions not only facilitates network management, but also helps to reduce the input space dimension of the prediction algorithm and improve the prediction speed. The network prediction layer is the middle layer of the computing network architecture. The near-real-time RAN intelligent controller in the domain is responsible for data collection, modeling and prediction in the target area, including the prediction of network traffic at the micro level and the prediction of business traffic at the macro level. The two types of traffic correspond to the basic status of the network and the basic needs of the business, respectively, and can provide effective guidance for the network decision layer of the architecture. The network decision layer is the highest layer of the computing network architecture, and appropriate decision algorithms are used to make decisions according to the actual business needs. This application mainly focuses on the prediction algorithm of the pre-network prediction layer.

[0029] like Figure 3 As shown in the figure, the algorithm deployment component in the computing network uses the Near-RT RIC (Near Real-time RAN Intelligent Controller) of O-RAN (Open Radio Access Network), which is specifically used to enhance the intelligence and automation capabilities of the network so that the algorithm can be executed more efficiently. Each area is managed by the Near-RT RIC (Near Real-time RAN Intelligent Controller) of the O-RAN architecture, on which the decision algorithm is deployed, responsible for the selection and execution of the decision algorithm, and issued through the O1 interface; the Non-RT RIC (Non-real-time RAN Intelligent Controller) communicates with the Near-RT RIC (Near Real-time RAN Intelligent Controller) through the A1 interface, and the Near-RT RIC (Near Real-time RAN Intelligent Controller) obtains the real-time data of each network node through the E2 interface, deploys the prediction algorithm on it, and uses the collected data for prediction. This division of regions helps to reduce the dimension of the environment space in the prediction, reduce the complexity of the prediction, and can also improve the response speed of the decision and reduce the algorithm deployment and execution delay.

[0030] In the target area of ​​the computing network architecture, a device node is selected as the target node. The computing power data of the target node, i.e., the device node in the computing network, within a predetermined historical time period is collected and analyzed, including but not limited to network traffic data, business traffic data, network feature data, and node resource data.

[0031] The network traffic data represents the traffic situation of multiple nodes in the computing network in a time series T = {1, 2, ..., T}, using the traffic matrix (Traffic Matrix) represents the i-th domain G i Network traffic, where C q Indicates the network traffic characteristics, which is set to 3 in this application, respectively, the inflow traffic size, outflow traffic size and traffic rate of each node in the target area, and represents the traffic situation of node v in time slot t, then the traffic matrix at time slot t is

[0032] Service traffic data refers to the application x traffic from user u received through Near-RT RIC (Near Real-Time RAN Intelligent Controller). represents the business traffic model, C x Set to 3, which are the size of the traffic packet, the type of traffic, and the rate at which the traffic packet arrives at the node. represents the service traffic of domain i in time slot t, and the traffic matrix of user u in time slot t can be expressed as The traffic is divided into packet-level traffic and data size-level traffic. For business traffic, the packet-level processing method is adopted. It indicates the number of traffic packets of the response transmitted in a certain time slot. The traffic type of the node is represented by a numerical value. is the video traffic belonging to application X. is the flow rate of other modes.

[0033] Network characteristic data is used to reflect network load conditions, using the matrix Represents the network node V i t The feature matrix at time slot t. In which, the node feature attribute C is set f =4, including the node's packet loss rate, failure rate, throughput and CPU (Central Processing Unit) utilization, Represents the characteristic situation of node v in time slot t.

[0034] Node resource data is used to measure whether it can carry the transmission of business flows, using the matrix Represents the network node V i t The state matrix at time slot t. Where C r =3, respectively, storage resource status, computing resource status and data resource status, Represents the state characteristics of the node in the time slot, Represents storage resources, which refers to the size of short-term storage allowed by each node. Represents computing resources, which are composed of CPU (Central Processing Unit) resources and GPU (Graphics Processing Unit) resources. Indicates data resources, which refers to the maximum data size that each node can process.

[0035] Step S104, extracting features from the historical computing power data of the target node to obtain spatial features of the historical computing power data of the target node and temporal features of the historical computing power data of the target node;

[0036] Alternatively, if Figure 4 As shown in the figure, the Encoder-Decoder architecture is adopted. The historical computing power data is converted into a high-dimensional representation through the input layer and input into the Encoder. The Encoder part is composed of an attention module that uses the attention mechanism to extract spatiotemporal features. Among them, the temporal features are extracted from short and long periods through ConvLSTM (Convolutional Long Short-Term Memory Network), and the spatial features are extracted through GAT (Graph Attention Network).

[0037] In an optional embodiment, feature extraction is performed on the historical computing power data of the target node to obtain spatial features of the historical computing power data of the target node, including: obtaining the connection relationship between device nodes in the computing power network; determining multiple neighbor nodes adjacent to the target node based on the connection relationship; obtaining the historical computing power data of each of the multiple neighbor nodes in the historical time period; inputting the historical computing power data of the target node and the historical computing power data of each of the multiple neighbor nodes into a target spatial attention model, and the target spatial attention model outputs the spatial features of the target node, wherein the target spatial attention model is trained by a first training sample, and the first training sample is the historical computing power data of the nodes in the computing power network, and the spatial features corresponding to the nodes in the computing power network.

[0038] Optionally, in order to extract the characteristics of the network from the space, a topology model is used to represent the connection between each forwarding node in the ith domain and the available bandwidth resources between links. Since the traffic is time-dependent, a dynamic topology model is established here.

[0039] For a v forwarding nodes, N e The network space is divided into domains composed of edges, using an undirected multi-weighted graph represents the i-th domain G in time slot t i The network topology model is: Represents each node in the domain, Represents the link status of the network connecting each node. If there is a physical link connection between nodes v and v′, then The attribute of the edge is the bandwidth between nodes, that is, The larger the value, the greater the link bandwidth. , it means there is no link between the nodes or the available bandwidth is 0.

[0040] In order to capture the spatial correlation between nodes, the GAT (Graph Attention Network) model is used to dynamically calculate the contribution of each neighbor to the target node feature update, and give higher weights to more important neighbor nodes. The historical computing power data of the target node, that is, the topology of each node in the predetermined time slot, is input into the spatial attention model with 3 GAT layers, and the GAT model can output the spatial features of the target node.

[0041] In an optional embodiment, the historical computing power data of the target node and the historical computing power data of each of the multiple neighboring nodes are input into a target space attention model, and the target space attention model outputs the spatial features of the target node, including: inputting the historical computing power data of the target node and the historical computing power data of each of the multiple neighboring nodes into the target space attention model, and the target space attention model determines the attention weights of each of the multiple neighboring nodes, wherein the attention weights represent the degree of influence of the neighboring nodes on the computing power of the target node; determining the spatial features of the target node based on the attention weights of each of the multiple neighboring nodes and the historical computing power data of the target node.

[0042] Optionally, in each graph attention layer, node v∈V i t The eigenvector of It will be updated through the attention mechanism according to the characteristics of its neighboring nodes. For the kth (1≤k≤3) graph attention layer, the attention coefficient between nodes v and v′, i.e., the attention weight, is

[0043]

[0044] in, represents the set of neighbor nodes of node v, u T and W (k) are all learnable parameters, W (k) is the weight matrix of the kth layer, Represents the eigenvalue of node v at the k-1th graph attention layer. Each node feature uses the graph attention coefficient to aggregate information of neighboring nodes with different focuses, and records the initial eigenvalue of node v Complete the graph convolution operations in sequence:

[0045]

[0046] After three layers of graph attention mechanism, the output result of the spatial attention model is obtained:

[0047] In an optional embodiment, feature extraction is performed on the historical computing power data of the target node to obtain the time features of the historical computing power data of the target node, including: inputting the historical computing power data of the target node into a target long short-term memory network, and outputting the time features of the historical computing power data of the target node by the target long short-term memory network, wherein the target convolutional long short-term memory network is obtained by training with a second training sample, and the second training sample is the historical computing power data of the nodes in the computing power network that change with time, and the time features corresponding to the nodes in the computing power network.

[0048] Optionally, the convolutional long short-term memory network model is an optimization model proposed based on the fully connected long short-term memory network. Given the short-term and medium-term time series flow matrix Q divided in the computing network closeness and Q middle ,A convolutional long short-term memory network is used to train each branch to extract local spatial features.

[0049] For the traffic matrix under one of the timing branches, it goes through the following two stages to enter the input of the next time state.

[0050] a. Forgetting and remembering stages

[0051] Decide which previous network conditions and traffic information changes over time and need to be ignored, and which new information will be saved in the unit state. Take a short cycle sequence as an example:

[0052] forget g(c) =σ(W qf *Q g(c) +W hf *h g(c-1) +W cf *C g(c-1) +b f )

[0053] memory g(c) =σ(W qm *Q g(c) +W hm *h g(c-1) +Wcm *C g(c-1) +b m )

[0054] Among them, forget g(c) Represents the information that needs to be forgotten at time step g(c), memory g(c) represents the information to be remembered, W and b are the convolution kernel and bias terms that can be trained, and h g(c-1) represents the output of the previous time step, the initial h0=0, and σ(·) is the sigmoid activation function.

[0055] b. Update and output phase

[0056] The cell state at update time g(c) is:

[0057] in,

[0058] According to h g(c) =o g(c) *tanh(C g(c) ) to get the output at time g(c), where o g(c) =σ(W qo *Q g(c) +W ho *h g(c-1) +W co *C g(c-1) +b o ).

[0059] Finally, the two branches each obtain a feature extraction time series h g(c) and h g(m) , through the gated fusion method g·h g(c) +(1-g)·h g(m) The output of the long short-term memory network (temporal attention module) is obtained by fusion Among them, g is the trainable weight parameter of the gating unit.

[0060] Step S106, fusing the spatial features of the target node and the temporal features of the target node into the spatiotemporal features of the target node;

[0061] Optionally, the traffic and computing power demand of the computing power network not only change over time, but are also affected by spatial factors. Single-dimensional features (such as only time or only space features) cannot fully reflect the true state and demand of the nodes, so fusing time factors and space factors into spatiotemporal features can provide a more complete and dynamic perspective, enabling the prediction model to capture the complex spatiotemporal patterns in computing power demand, thereby obtaining more accurate prediction results.

[0062] In an optional embodiment, the spatial features of the target node and the temporal features of the target node are fused into the spatiotemporal features of the target node, including: inputting the spatial features of the target node and the temporal features of the target node into a target fusion model, dividing the spatial features of the target node into multiple subspaces by the target fusion model to obtain respective subspace features of the multiple subspaces; mapping the temporal features of the target node to multiple subspaces to obtain respective sub-temporal features of the multiple subspaces; fusing the respective subspace features of the multiple subspaces with the respective sub-temporal features of the multiple subspaces to obtain the spatiotemporal features of the target node.

[0063] Optionally, considering the complementarity of spatiotemporal features in information representation, and that temporal and spatial features cannot be simply averaged or weighted together, a multi-head attention fusion method is used to adaptively and dynamically fuse spatiotemporal features to further capture their correlation.

[0064] The output of the spatial attention model and the output of the long short-term memory network (temporal attention module) are concatenated as the input of the fusion model:

[0065]

[0066] The input Divide into multiple subspaces and calculate the scaled dot product attention on these subspaces to obtain Connecting multiple attention results to obtain fused features.

[0067] In an optional embodiment, the subspace features of each of the multiple subspaces and the sub-time features of each of the multiple subspaces are fused to obtain the spatiotemporal features of the target node, including: multiplying the subspace features and the sub-time features of each of the multiple subspaces in the multiple subspaces to obtain the sub-spatiotemporal features of each of the multiple subspaces; and fusing the sub-spatiotemporal features of each of the multiple subspaces to obtain the spatiotemporal features of the target node.

[0068] Optionally, the fusion features obtained by connecting the attention results of multiple heads include:

[0069]

[0070] in, is the trainable weight matrix, is the number of fused attention heads, is the dimension of the fused attention head.

[0071] Step S108, predicting the congestion of the computing power network at the target node based on the spatiotemporal characteristics of the target node.

[0072] Optionally, get the output of the Encoder module through the Add&Norm module

[0073]

[0074] Among them, μ and θ are the outputs of the residual connection The mean and standard deviation in the feature dimension, ψ and ξ are learnable parameters. Add is a residual connection, which is usually used to solve the problem of multi-layer network training, allowing the network to focus only on the current difference. Norm converts the input of each layer of neurons into the same mean and variance to speed up convergence. By analyzing the output of the Encoder module, it is determined whether the node will become congested due to computing power demand or traffic surge at a specific future time point or time period, including but not limited to identifying whether the target node will have a traffic peak, whether the computing or storage resources have reached a bottleneck, and whether the bandwidth allocation is sufficient to support current business needs.

[0075] The historical computing power data of the target node within a predetermined historical time period is obtained through the above method, wherein the target node is a device node in the computing power network; feature extraction is performed on the historical computing power data of the target node to obtain the spatial features of the historical computing power data of the target node and the time features of the historical computing power data of the target node; the spatial features of the target node and the time features of the target node are merged into the spatiotemporal features of the target node; according to the spatiotemporal features of the target node, the congestion of the computing power network at the target node is predicted, which solves the technical problem that the prior art cannot accurately predict the device congestion of the computing power network, achieves the purpose of comprehensively considering the factors of the computing power network in the spatial dimension and the temporal dimension, and then predicting the congestion of the device nodes of the computing power network, thereby achieving the technical effect of improving the accuracy of the prediction of the congestion of the device nodes of the example network.

[0076] Based on the above embodiments and optional embodiments, the present invention proposes an optional implementation manner.

[0077] Step S1, obtaining historical computing power data of a target node within a predetermined historical time period, wherein the target node is a device node in a computing power network;

[0078] In this step, if Figure 2As shown in the figure, the computing network architecture includes the network foundation layer, the network prediction layer and the network decision layer. The network foundation layer is the bottom layer of the architecture, which is composed of various devices of the computing network. According to factors such as the geographical location of the nodes in the network, the computing network is divided into multiple regions (only the target region is shown in the architecture diagram). Each region can be regarded as an autonomous network, consisting of multiple network nodes. Each domain is managed by the Non-RT RIC (non-real-time RAN intelligent controller) and Near-RT RIC (near real-time RAN intelligent controller) of the O-RAN (Open Radio Access Network) architecture. The division of regions not only facilitates network management, but also helps to reduce the input space dimension of the prediction algorithm and improve the prediction speed. The network prediction layer is the middle layer of the computing network architecture. The near-real-time RAN intelligent controller in the domain is responsible for data collection, modeling and prediction in the target area, including the prediction of network traffic at the micro level and the prediction of business traffic at the macro level. The two types of traffic correspond to the basic status of the network and the basic needs of the business, respectively, and can provide effective guidance for the network decision layer of the architecture. The network decision layer is the highest layer of the computing network architecture, and appropriate decision algorithms are used to make decisions according to the actual business needs. This application mainly focuses on the prediction algorithm of the pre-network prediction layer.

[0079] like Figure 3 As shown in the figure, the algorithm deployment component in the computing network uses the Near-RT RIC (Near Real-time RAN Intelligent Controller) of O-RAN (Open Radio Access Network), which is specifically used to enhance the intelligence and automation capabilities of the network so that the algorithm can be executed more efficiently. Each area is managed by the Near-RT RIC (Near Real-time RAN Intelligent Controller) of the O-RAN architecture, on which the decision algorithm is deployed, responsible for the selection and execution of the decision algorithm, and issued through the O1 interface; the Non-RT RIC (Non-real-time RAN Intelligent Controller) communicates with the Near-RT RIC (Near Real-time RAN Intelligent Controller) through the A1 interface, and the Near-RT RIC (Near Real-time RAN Intelligent Controller) obtains the real-time data of each network node through the E2 interface, deploys the prediction algorithm on it, and uses the collected data for prediction. This division of regions helps to reduce the dimension of the environment space in the prediction, reduce the complexity of the prediction, and can also improve the response speed of the decision and reduce the algorithm deployment and execution delay.

[0080] In the target area of ​​the computing network architecture, a device node is selected as the target node. The computing power data of the target node, i.e., the device node in the computing network, within a predetermined historical time period is collected and analyzed, including but not limited to network traffic data, business traffic data, network feature data, and node resource data.

[0081] The network traffic data represents the traffic situation of multiple nodes in the computing network in a time series T = {1, 2, ..., T}, using the traffic matrix (Traffic Matrix) represents the i-th domain G i Network traffic, where C q Indicates the network traffic characteristics, which is set to 3 in this application, respectively, the inflow traffic size, outflow traffic size and traffic rate of each node in the target area, and represents the traffic situation of node v in time slot t, then the traffic matrix at time slot t is

[0082] Business traffic data refers to the traffic from users received through Near-RT RIC (Near Real-time RAN Intelligent Controller).

[0083] Applications Traffic, use represents the business traffic model, C x Set to 3, which are the size of the traffic packet, the type of traffic, and the rate at which the traffic packet arrives at the node. represents the traffic flow of domain i in time slot t, user The traffic matrix at time slot t can be expressed as The traffic is divided into packet-level traffic and data size-level traffic. For business traffic, the packet-level processing method is adopted. It indicates the number of traffic packets of the response transmitted in a certain time slot. The traffic type of the node is represented by a numerical value. Application If the video traffic is the flow rate of other modes.

[0084] Network characteristic data is used to reflect network load conditions, using the matrix Represents the network node V i t The feature matrix at time slot t. In which, the node feature attribute C is set f =4, including the node's packet loss rate, failure rate, throughput and CPU (Central Processing Unit) utilization, Represents the characteristic situation of node v in time slot t.

[0085] Node resource data is used to measure whether it can carry the transmission of business flows, using the matrix Represents the network node V i t The state matrix at time slot t. Where C r=3, respectively, storage resource status, computing resource status and data resource status, Represents the state characteristics of the node in the time slot, Represents storage resources, which refers to the size of short-term storage allowed by each node. Represents computing resources, which are composed of CPU (Central Processing Unit) resources and GPU (Graphics Processing Unit) resources. Indicates data resources, which refers to the maximum data size that each node can process.

[0086] Step S2, extracting features from the historical computing power data of the target node to obtain spatial features of the historical computing power data of the target node, including: obtaining the connection relationship between device nodes in the computing power network; determining multiple neighbor nodes adjacent to the target node based on the connection relationship; obtaining the historical computing power data of each of the multiple neighbor nodes in the historical time period; inputting the historical computing power data of the target node and the historical computing power data of each of the multiple neighbor nodes into the target spatial attention model, and the target spatial attention model outputs the spatial features of the target node, wherein the target spatial attention model is trained by a first training sample, and the first training sample is the historical computing power data of the nodes in the computing power network, and the spatial features corresponding to the nodes in the computing power network.

[0087] Optionally, the computing network G is divided into N regions {G1, G2, ..., G N}, in order to extract the characteristics of the network from the space, the topology model is used to represent the connection between each forwarding node in the ith domain and the available bandwidth resources between links. And the traffic is time-dependent, so a dynamic topology model is established here.

[0088] For a v forwarding nodes, N e The network space is divided into domains composed of edges, using an undirected multi-weighted graph represents the i-th domain G in time slot t i The network topology model is: Represents each node in the domain, Represents the link status of the network connecting each node. If there is a physical link connection between nodes v and v′, then The attribute of the edge is the bandwidth between nodes, that is, The larger the value, the greater the link bandwidth. , it means there is no link between the nodes or the available bandwidth is 0.

[0089] In order to capture the spatial correlation between nodes, the GAT (Graph Attention Network) model is used to dynamically calculate the contribution of each neighbor to the target node feature update, and give higher weights to more important neighbor nodes. The historical computing power data of the target node, i.e., the topology of each node in the predetermined time slot, the feature matrix and the resource matrix, are input into the spatial attention model with 3 GAT layers. The GAT model can output the spatial features of the target node.

[0090] Step S21, inputting the historical computing power data of the target node and the historical computing power data of each of the multiple neighboring nodes into the target spatial attention model, and the target spatial attention model outputs the spatial features of the target node, including:

[0091] Input the historical computing power data of the target node and the historical computing power data of each of the multiple neighboring nodes into the target space attention model, and determine the attention weights of each of the multiple neighboring nodes by the target space attention model, where the attention weights represent the degree of influence of the neighboring nodes on the computing power of the target node;

[0092] According to the attention weights of multiple neighbor nodes, the spatial characteristics of the target node are determined by combining the historical computing power data of multiple neighbor nodes and the historical computing power data of the target node.

[0093] Optionally, in each graph attention layer, node v∈V i t The eigenvector of It will be updated through the attention mechanism according to the characteristics of its neighboring nodes. For the kth (1≤k≤3) graph attention layer, the attention coefficient between nodes v and v′, i.e., the attention weight, is

[0094]

[0095] in, represents the set of neighbor nodes of node v, u T and W (k) are all learnable parameters, W (k) is the weight matrix of the kth layer, Represents the eigenvalue of node v at the k-1th graph attention layer. Each node feature uses the graph attention coefficient to aggregate information of neighboring nodes with different focuses, and records the initial eigenvalue of node v Complete the graph convolution operations in sequence:

[0096]

[0097] After three layers of graph attention mechanism, the output result of the spatial attention model is obtained:

[0098] Step S3, extracting features from the historical computing power data of the target node to obtain the time features of the historical computing power data of the target node, including: inputting the historical computing power data of the target node into the target long short-term memory network, and outputting the time features of the historical computing power data of the target node from the target long short-term memory network, wherein the target convolutional long short-term memory network is trained by a second training sample, and the second training sample is the historical computing power data of the nodes in the computing power network that change over time, and the time features corresponding to the nodes in the computing power network.

[0099] Optionally, the convolutional long short-term memory network model is an optimization model proposed based on the fully connected long short-term memory network. Given the short-term and medium-term time series flow matrix Q divided in the computing network closeness and Q middle ,A convolutional long short-term memory network is used to train each branch to extract local spatial features.

[0100] For the traffic matrix under one of the timing branches, it goes through the following two stages to enter the input of the next time state.

[0101] a. Forgetting and remembering stages

[0102] Decide which previous network conditions and traffic information changes over time and need to be ignored, and which new information will be saved in the unit state. Take a short cycle sequence as an example:

[0103] forget g(c) =σ(W qf *Q g(c) +W hf *h g(c-1) +W cf *C g(c-1) +b f )

[0104] memory g(c) =σ(W qm *Q g(c) +W hm *h g(c-1) +W cm *C g(c-1) +b m )

[0105] Among them, forget g(c) Represents the information that needs to be forgotten at time step g(c), memory g(c) represents the information to be remembered, W and b are the convolution kernel and bias terms that can be trained, and h g(c-1)represents the output of the previous time step, the initial h0=0, and σ(·) is the sigmoid activation function.

[0106] b. Update and output phase

[0107] The cell state at update time g(c) is:

[0108] in,

[0109] According to h g(c) =o g(c) *tanh(C g(c) ) to get the output at time g(c), where o g(c) =σ(W qo *Q g(c) +W ho *h g(c-1) +W co *C g(c-1) +b o ).

[0110] Finally, the two branches each obtain a feature extraction time series h g(c) and h g(m) , through the gated fusion method g·h g(c) +(1-g)·h g(m) The output of the long short-term memory network (temporal attention module) is obtained by fusion Among them, g is the trainable weight parameter of the gating unit.

[0111] Among them, forget g(c) Represents the information that needs to be forgotten at time step g(c), memory g(c) represents the information to be remembered, W and b are the convolution kernel and bias terms that can be trained, and h g(c-1) represents the output of the previous time step, the initial h0=0, and σ(·) is the sigmoid activation function.

[0112] Step S4, fusing the spatial features of the target node and the temporal features of the target node into the spatiotemporal features of the target node, including: inputting the spatial features of the target node and the temporal features of the target node into a target fusion model, and using the target fusion model to divide the spatial features of the target node into multiple subspaces to obtain respective subspace features of the multiple subspaces; mapping the temporal features of the target node to multiple subspaces to obtain respective sub-temporal features of the multiple subspaces; fusing the respective subspace features of the multiple subspaces with the respective sub-temporal features of the multiple subspaces to obtain the spatiotemporal features of the target node.

[0113] Optionally, considering the complementarity of spatiotemporal features in information representation, and that temporal and spatial features cannot be simply averaged or weighted together, a multi-head attention fusion method is used to adaptively and dynamically fuse spatiotemporal features to further capture their correlation.

[0114] The output of the spatial attention model and the output of the long short-term memory network (temporal attention module) are concatenated as the input of the fusion model:

[0115] The input Divide into multiple subspaces and calculate the scaled dot product attention on these subspaces to obtain

[0116] Connecting multiple attention results to obtain fused features.

[0117] Step S41, fusing the subspace features of each of the multiple subspaces and the sub-time features of each of the multiple subspaces to obtain the spatiotemporal features of the target node, including: multiplying the subspace features and the sub-time features of each of the multiple subspaces in the multiple subspaces respectively to obtain the sub-spatiotemporal features of each of the multiple subspaces; fusing the sub-spatiotemporal features of each of the multiple subspaces to obtain the spatiotemporal features of the target node.

[0118] The fusion features obtained by connecting multiple attention results include:

[0119]

[0120] in, is the trainable weight matrix, is the number of fused attention heads,

[0121] is the dimension of the fused attention head.

[0122] Step S5: predict the congestion of the computing network at the target node based on the spatiotemporal characteristics of the target node.

[0123] Optionally, get the output of the Encoder module through the Add&Norm module

[0124]

[0125] Among them, μ and θ are the outputs of the residual connection The mean and standard deviation in the feature dimension, ψ and ξ are learnable parameters. Add is a residual connection, which is usually used to solve the problem of multi-layer network training, allowing the network to focus only on the current difference. Norm converts the input of each layer of neurons into the same mean and variance to speed up convergence. By analyzing the output of the Encoder module, it is determined whether the node will become congested due to computing power demand or traffic surge at a specific future time point or time period, including but not limited to identifying whether the target node will have a traffic peak, whether the computing or storage resources have reached a bottleneck, and whether the bandwidth allocation is sufficient to support current business needs.

[0126] The above optional implementation methods achieve at least the following effects: (1) A more efficient computing network management architecture: The present application predicts and manages the computing network in a layered and regional manner, which can effectively reduce the latency of data collection, prediction, and decision-making algorithm deployment and execution.

[0127] (2) More accurate and comprehensive traffic prediction: This application designs a traffic prediction algorithm that uses an attention mechanism to assist in the fusion of spatiotemporal features to predict the network traffic and business traffic of the computing network. It fully considers the temporal characteristics of the traffic and the spatial characteristics of the network environment, and comprehensively measures the impact of various factors on the traffic, thereby achieving more accurate traffic prediction.

[0128] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0129] In this embodiment, a device for predicting computing power is also provided, which is used to implement the above embodiments and preferred implementation modes, and will not be repeated here. As used below, the terms "module" and "device" can implement a combination of software and / or hardware for a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0130] According to an embodiment of the present invention, a device embodiment for implementing a method for predicting computing power is also provided. Figure 5 is a schematic diagram of a device for predicting computing power according to an embodiment of the present invention. Figure 5 As shown, the above-mentioned device for predicting computing power includes an acquisition module 51, an extraction module 52, a fusion module 53, and a prediction module 54. The device is described below.

[0131] An acquisition module 51 is used to acquire historical computing power data of a target node within a predetermined historical time period, wherein the target node is a device node in the computing power network;

[0132] The extraction module 52 is connected to the acquisition module 51 and is used to extract features from the historical computing power data of the target node to obtain the spatial features of the historical computing power data of the target node and the temporal features of the historical computing power data of the target node;

[0133] A fusion module 53, connected to the extraction module 52, is used to fuse the spatial features of the target node and the temporal features of the target node into the spatiotemporal features of the target node;

[0134] The prediction module 54 is connected to the fusion module 53 and is used to predict the congestion of the computing power network at the target node according to the spatiotemporal characteristics of the target node.

[0135] In a computing power prediction device provided by an embodiment of the present invention, an acquisition module is set up to obtain historical computing power data of a target node within a predetermined historical time period, wherein the target node is a device node in a computing power network; an extraction module is used to perform feature extraction on the historical computing power data of the target node to obtain the spatial features of the historical computing power data of the target node and the time features of the historical computing power data of the target node; a fusion module is used to fuse the spatial features of the target node and the time features of the target node into the spatiotemporal features of the target node; a prediction module is used to predict the congestion of the computing power network at the target node based on the spatiotemporal features of the target node, thereby solving the technical problem that the prior art cannot accurately predict the device congestion of the computing power network, achieving the purpose of comprehensively considering the factors of the computing power network in the spatial dimension and the temporal dimension, and then predicting the congestion of the device nodes of the computing power network, thereby achieving the technical effect of improving the accuracy of predicting the congestion of the device nodes of the computing power network.

[0136] It should be noted that the above modules can be implemented by software or hardware. For example, for the latter, it can be implemented in the following ways: the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0137] It should be noted that the acquisition module 51, extraction module 52, fusion module 53, and prediction module 54 correspond to steps S102 to S108 in the embodiment, and the examples and application scenarios implemented by the modules and the corresponding steps are the same, but are not limited to the contents disclosed in the embodiment. It should be noted that the modules as part of the device can be run in a computer terminal.

[0138] It should be noted that the optional or preferred implementation of this embodiment can refer to the relevant description in the embodiment, which will not be repeated here.

[0139] The above-mentioned device for predicting computing power may also include a processor and a memory. The acquisition module 51, the extraction module 52, the fusion module 53, the prediction module 54, etc. are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize the corresponding functions.

[0140] The processor includes a kernel, which retrieves the corresponding program unit from the memory. There can be one or more kernels. The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one storage chip.

[0141] An embodiment of the present invention provides a non-volatile storage medium having a program stored thereon, which implements a method for predicting computing power when the program is executed by a processor.

[0142] like Figure 6 As shown, an embodiment of the present invention provides an electronic device, the electronic device 10 includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, the following steps are implemented: the memory is used to store a computer program, wherein when the computer program is executed by the processor, the processor implements the above-mentioned method for predicting computing power. The device in this article may be a server, a PC, etc.

[0143] The present invention also provides a computer program product, which, when executed on a data processing device, is suitable for executing a program initialized with the following method steps: computer instructions are executed by a processor to perform the above-mentioned method of predicting computing power.

[0144] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0146] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0147] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0148] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0149] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0150] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0151] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0152] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0153] The above are only embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for predicting computing power, characterized in that: include: Obtaining historical computing power data of a target node within a predetermined historical time period, wherein the target node is a device node in a computing power network; Performing feature extraction on the historical computing power data of the target node to obtain spatial features of the historical computing power data of the target node and temporal features of the historical computing power data of the target node; fusing the spatial feature of the target node and the temporal feature of the target node into the spatiotemporal feature of the target node; The congestion situation of the computing power network at the target node is predicted based on the spatiotemporal characteristics of the target node.

2. The method according to claim 1, characterized in that The extracting features of the historical computing power data of the target node to obtain the spatial features of the historical computing power data of the target node includes: Obtaining the connection relationship between device nodes in the computing power network; Determine, according to the connection relationship, a plurality of neighbor nodes adjacent to the target node; Obtaining historical computing power data of each of the plurality of neighbor nodes within the historical time period; The historical computing power data of the target node and the historical computing power data of each of the multiple neighboring nodes are input into a target spatial attention model, and the target spatial attention model outputs the spatial features of the target node, wherein the target spatial attention model is trained by a first training sample, and the first training sample is the historical computing power data of the nodes in the computing power network, and the spatial features corresponding to the nodes in the computing power network.

3. The method according to claim 2, inputting the historical computing power data of the target node and the historical computing power data of each of the plurality of neighboring nodes into a target spatial attention model, and outputting the spatial features of the target node by the target spatial attention model, comprising: Inputting the historical computing power data of the target node and the historical computing power data of each of the plurality of neighboring nodes into the target space attention model, and determining the attention weights of each of the plurality of neighboring nodes by the target space attention model, wherein the attention weights represent the influence of the neighboring nodes on the computing power of the target node; According to the attention weights of the multiple neighbor nodes, the spatial features of the target node are determined in combination with the historical computing power data of the multiple neighbor nodes and the historical computing power data of the target node.

4. The method according to claim 1, characterized in that: The extracting features of the historical computing power data of the target node to obtain the time features of the historical computing power data of the target node includes: The historical computing power data of the target node is input into a target long short-term memory network, and the target long short-term memory network outputs the time characteristics of the historical computing power data of the target node, wherein the target convolutional long short-term memory network is trained by a second training sample, and the second training sample is the historical computing power data of the nodes in the computing power network that change over time, and the time characteristics corresponding to the nodes in the computing power network.

5. The method according to claim 1, characterized in that: The step of fusing the spatial feature of the target node and the temporal feature of the target node into the spatiotemporal feature of the target node includes: Inputting the spatial feature of the target node and the temporal feature of the target node into a target fusion model, and using the target fusion model to divide the spatial feature of the target node into multiple subspaces to obtain subspace features of each of the multiple subspaces; Mapping the time feature of the target node to the multiple subspaces to obtain the sub-time features of each of the multiple subspaces; The subspace features of each of the multiple subspaces and the subtime features of each of the multiple subspaces are fused to obtain the spatiotemporal features of the target node.

6. The method according to claim 5, characterized in that The fusing the subspace features of each of the multiple subspaces and the subtime features of each of the multiple subspaces to obtain the spatiotemporal features of the target node includes: In the plurality of subspaces, respectively, multiplying the subspace features and the sub-time features of the plurality of subspaces to obtain the sub-space-time features of the plurality of subspaces; The sub-spatiotemporal features of each of the multiple subspaces are fused to obtain the spatiotemporal features of the target node.

7. A device for predicting computing power, characterized in that: include: An acquisition module, used to acquire historical computing power data of a target node within a predetermined historical time period, wherein the target node is a device node in the computing power network; An extraction module, used to perform feature extraction on the historical computing power data of the target node to obtain the spatial features of the historical computing power data of the target node and the temporal features of the historical computing power data of the target node; A fusion module, used for fusing the spatial feature of the target node and the temporal feature of the target node into the spatiotemporal feature of the target node; The prediction module is used to predict the congestion of the computing power network at the target node according to the spatiotemporal characteristics of the target node.

8. A non-volatile storage medium, characterized in that: The non-volatile storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the method for predicting computing power as described in any one of claims 1 to 6.

9. An electronic device, characterized in that: include: One or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method for predicting computing power as described in any one of claims 1 to 6.

10. A computer program product comprising computer instructions, characterized in that: The computer instructions are executed by the processor to implement the method for predicting computing power as described in any one of claims 1 to 6.

Images