Computing power network demand evaluation method and device, and computer readable storage medium

Abstract

The embodiment of the invention provides a computing power network demand evaluation method and device, and a computer readable storage medium, and the method comprises the steps: constructing a spatial correlation network corresponding to a computing power node based on a preset spatial relationship type of the computing power node; obtaining multiple types of time series data corresponding to the computing power node; the multiple types of time sequence data correspond to different time period attributes; extracting multi-dimensional time correlation information corresponding to the computing power node based on the multiple types of time sequence data; and based on the spatial correlation network and the multi-dimensional time correlation information, through an intelligent calculation demand model, obtaining a calculation power demand prediction value. Multi-dimensional space relevance, demand period volatility and space node influence difference are deeply fused, and the accuracy of intelligent calculation demand prediction is remarkably improved. By designing a supply-demand matching, economic benefit and operation efficiency three-dimensional evaluation system, the system can pre-judge the layout rationality of the computing power nodes, and ensures the high efficiency and accuracy of computing power network resource planning.

Description

A method, apparatus, and computer-readable storage medium for assessing computing power network requirements. Technical Field

[0001] This application relates to the field of computing power network technology, and in particular to a computing power network demand assessment method, apparatus, and computer-readable storage medium. Background Technology

[0002] Intelligent computing power refers to computing capabilities specifically designed for artificial intelligence (AI) tasks, focusing on processing complex nonlinear data and supporting AI model training and inference. Correspondingly, computing power demand refers to the computing capacity requirements put forward by users or industries to complete computing tasks (especially AI-related tasks), encompassing two major dimensions: "demand scale" and "demand characteristics." Computing power demand is the core driving force behind the development of intelligent computing power. With the iterative innovation of large-scale models, the demand for intelligent computing power has grown rapidly, gradually shifting from general-purpose computing to intelligent computing. How to support the adjustment and optimization of computing network resource layout has become an urgent problem to be solved. Summary of the Invention

[0003] This application provides a method, apparatus, and computer-readable storage medium for assessing computing network demand, which can solve the technical problem of unreasonable computing network resource layout in related technologies.

[0004] In a first aspect, embodiments of this application provide a method for assessing computing power network demand. The method includes: constructing a spatial association network corresponding to computing power nodes based on preset spatial relationship types of computing power nodes; acquiring multiple types of time-series data corresponding to computing power nodes; the multiple types of time-series data corresponding to different time period attributes; extracting multi-dimensional time association information corresponding to computing power nodes based on the multiple types of time-series data; and obtaining a predicted value of computing power demand based on the spatial association network and the multi-dimensional time association information through an intelligent computing demand model.

[0005] In one possible implementation of the first aspect, the preset spatial relationship type includes at least a first spatial relationship and a second spatial relationship; based on the preset spatial relationship type, a spatial association network corresponding to the computing power nodes is constructed, including: determining a local computing power node network graph based on the first spatial relationship; determining the first spatial relationship based on the geographical adjacency attribute of the computing power nodes; determining a global computing power node network graph based on the second spatial relationship; determining the second spatial relationship based on the demand similarity attribute of the computing power nodes; and constructing a spatial dimension attention matrix based on the local computing power node network graph and the global computing power node network graph to obtain the spatial association network.

[0006] In one possible implementation of the first aspect, the multiple types of time series data include instantaneous time series data, daily periodic time series data, and weekly periodic time series data that are adjacent to the prediction time period; both the daily periodic time series data and the weekly periodic time series data are extracted from the basic standardized data of the computing power nodes.

[0007] In one possible implementation of the first aspect, the multiple types of time-series data include first time-series data and second time-series data; the first time-series data and the second time-series data have different time period attributes; based on the multiple types of time-series data, multi-dimensional time-related information corresponding to the computing power nodes is extracted, including: using the first time-series data as the first input data to obtain a first time-related feature; using the second time-series data as the second input data to obtain a second time-related feature; and fusing the first time-related feature and the second time-related feature to jointly constitute multi-dimensional time-related information.

[0008] In one possible implementation of the first aspect, multi-dimensional time-related information corresponding to computing power nodes is extracted based on multiple types of time-series data, including: taking time-series data as the first input data to obtain a first time-related feature; taking daily periodic time-series data as the second input data to obtain a second time-related feature; taking weekly periodic time-series data as the third input data to obtain a third time-related feature; and fusing the first time-related feature, the second time-related feature, and the third time-related feature to jointly constitute multi-dimensional time-related information.

[0009] In one possible implementation of the first aspect, the method further includes: inputting the predicted computing power demand value into the trained random forest model to obtain the rationality judgment result of the computing power node layout; and generating a corresponding computing power layout unreasonable warning and optimization request in response to the rationality judgment result of the layout being unreasonable.

[0010] In one possible implementation of the first aspect, the method further includes: constructing a three-dimensional index framework of supply and demand, benefits and efficiency; collecting raw data of computing nodes based on the three-dimensional index framework, performing data preprocessing and data format conversion on the raw data in sequence to generate standardized basic data; dividing the standardized basic data into training set and test set, constructing an initial random forest model based on the training set, and using the test set to evaluate and adjust the parameters of the initial random forest model to obtain the trained random forest model.

[0011] The computing power network demand assessment method provided in this application has the following beneficial effects: It constructs a spatial correlation network corresponding to computing power nodes based on preset spatial relationship types; it obtains multiple types of time-series data corresponding to computing power nodes; these multiple types of time-series data correspond to different time period attributes; based on these multiple types of time-series data, it extracts multi-dimensional time correlation information corresponding to computing power nodes; and based on the spatial correlation network and multi-dimensional time correlation information, it obtains predicted computing power demand values ​​through an intelligent computing demand model. This method deeply integrates multi-dimensional spatial correlation, demand period fluctuations, and differences in the impact of spatial nodes, significantly improving the accuracy of intelligent computing demand prediction. Furthermore, by designing a three-dimensional evaluation system encompassing supply and demand matching, economic benefits, and operational efficiency, the system can predict the rationality of computing power node layout, ensuring efficient and accurate computing power network resource planning.

[0012] Secondly, embodiments of this application provide a computing power network demand assessment device, which is used to perform the computing power network demand assessment method of the first aspect described above.

[0013] In one possible implementation of the second aspect, the computing power network demand assessment device further includes a distributed data asset management terminal and a computing power node data asset think tank management platform. The distributed data asset management terminal is used to manage its own data assets based on distributed blockchain technology. The computing power node data asset think tank management platform is used to collect all data of computing power nodes based on distributed blockchain technology, generate and revise the computing power node data catalog, formulate and classify computing power node data tags, and store the data in a distributed ledger.

[0014] The computing power network demand assessment device of the second aspect mentioned above can refer to the beneficial effects of the first aspect and any of its possible design methods, which will not be elaborated here.

[0015] Thirdly, embodiments of this application provide an electronic device for performing the computing power network demand assessment method described in the first aspect.

[0016] The electronic device described in the third aspect above can refer to the beneficial effects of the first aspect above and any of its possible design methods, which will not be repeated here.

[0017] Fourthly, embodiments of this application provide a non-transient computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the computing power network demand assessment method of the first aspect described above.

[0018] The non-transient computer-readable storage medium of the fourth aspect described above can be referenced to the beneficial effects of the first aspect and any of its possible design methods, which will not be repeated here.

[0019] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the computing power network demand assessment method described in the first aspect.

[0020] The computer program product described in the fifth aspect above can refer to the beneficial effects of the first aspect above and any of its possible design methods, which will not be repeated here. Attached Figure Description

[0021] The accompanying drawings are for better understanding of this solution and do not constitute a limitation on the embodiments of this application. Specifically: Figure 1 is a flowchart illustrating a computing power network demand assessment method provided in some embodiments of this application; Figure 2 is a schematic diagram of spatiotemporal feature fusion of the intelligent computing demand model in a computing power network demand assessment method provided in some embodiments of this application; Figure 3 is a schematic diagram of a local computing power node network diagram in a computing power network demand assessment method provided in some embodiments of this application; Figure 4 is a schematic diagram of a global computing power node network diagram in a computing power network demand assessment method provided in some embodiments of this application; Figure 5 is an example diagram of periodic data in a computing power network demand assessment method provided in some embodiments of this application; Figure 6 is a flowchart illustrating a computing power network demand assessment method provided in some embodiments of this application; Figure 7a is a structural schematic diagram of a computing power network demand assessment device provided in some embodiments of this application; Figure 7b is a structural schematic diagram of a computing power network demand assessment device provided in some embodiments of this application; Figure 8 is a structural schematic diagram of a computing power network demand assessment device provided in some embodiments of this application; Figure 9 is a structural schematic diagram of an electronic device provided in some embodiments of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings. In the description of this application, unless otherwise stated, " / " indicates that the objects before and after are in an "or" relationship. For example, A / B can represent A or B. "And / or" in this application is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone, where A and B can be singular or plural. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple. Furthermore, to facilitate a clear description of the technical solutions in the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish identical or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" are not necessarily different. Meanwhile, in the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is being used as an example, illustration, or description. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of terms such as "exemplary" or "for example" is intended to present related concepts in a concrete manner for ease of understanding.

[0023] Furthermore, the network architecture and business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0024] In recent years, the parameter scale and training data volume of large-scale artificial intelligence models (AI models) have continued to expand, driving an explosive growth in the demand for intelligent computing power. According to industry data, since 2012, the computational load required for AI model training has doubled every 3.5 months. Simultaneously, the application penetration of large-scale models across various industries is deepening, and application scenarios are diversifying, further driving up the intelligent computing power required for model inference. Under this trend, computing power demand is gradually shifting from general-purpose computing to intelligent computing. IDC predicts that the scale of intelligent computing power will reach 1117.4 EFLOPS in 2027, with a compound annual growth rate of 33.9% from 2022 to 2027.

[0025] The rapid increase in demand for intelligent computing power places higher demands on the accuracy of demand forecasting and the rationality of resource deployment for computing power operators. If the supply and demand are not matched accurately, it can easily lead to resource surplus (causing computing power waste) or resource shortage (forming performance bottlenecks). Therefore, accurately predicting computing power business demand has become the key for computing power operators to plan resource allocation in advance, flexibly adjust deployment, improve customer satisfaction and market competitiveness.

[0026] However, computing power demands are characterized by diversity in type, temporal randomness, and varying magnitude. Statistical prediction methods, based on linear relationship assumptions and lacking adaptability and learning capabilities, struggle to meet prediction accuracy requirements. Currently, several well-known neural network technologies suitable for complex data processing have emerged in the industry. Among them, Graph Convolutional Networks (GCNs) are specifically designed for processing graph-structured data. Taking the node feature matrix and adjacency matrix as input, they extract node feature representations through graph convolution operations, supporting tasks such as classification and prediction. Gated Recurrent Units (GRUs), a variant of Recurrent Neural Networks (RNNs), selectively retain or update historical information through reset and update gates, solving the gradient vanishing problem of standard RNNs and effectively capturing long-term and short-term dependencies in time series. Attention mechanisms enable models to focus on key parts when processing information, improving model accuracy and efficiency by assigning differentiated weights to different parts of the input data. Random Forests, an ensemble learning method based on the Bagging strategy, reduce overfitting of individual decision trees and improve prediction accuracy by constructing multiple decision trees and fusing their results (voting or weighted calculation). Currently, the proposed method for predicting computing power demand based on cross-regional data involves marking plots with known computing power demand as prior plots, collecting demand feature sequences between the target region and prior plots, calculating the feature similarity between the two, and obtaining the prediction result for the target region by weighted combination of computing power demand from similar prior plots. This method aims to solve the "data silo effect caused by the reliance on local data in computing power demand prediction." However, it only uses a fixed monitoring period and does not consider the impact of time-point demand fluctuations on the accuracy of feature similarity judgment, thus failing to adapt to the temporal randomness of demand.

[0027] The computational demand prediction based on local historical data uses prediction accuracy and prediction efficiency as loss functions. It constructs an LSTM and BP dual neural network to make predictions by learning local historical computational demand. It relies solely on local historical data to ensure the timeliness of predictions, but it does not characterize the regional dependence of computational demand in the context of the computational network and cannot improve prediction accuracy by utilizing cross-regional correlations.

[0028] By collecting and preprocessing business behavior data, edge computing power prediction is achieved with the help of feature extraction models and computing power resource idle prediction models. It only considers the temporal correlation of edge computing power demand, aiming to improve the prediction robustness and generalization ability in scenarios with "insufficient business data and low scenario universality". However, it also does not characterize the regional dependence of computing power demand, and the applicable scenarios are limited to edge computing.

[0029] Deep learning algorithms based on local historical time-series data, deep learning algorithms based on cross-regional historical time-series data, or improvements to traditional deep learning algorithms through combination of algorithms and superimposed data processing, still have the following limitations: they do not consider the multidimensional spatial dependencies between computing power nodes, and cannot capture the correlation between nodes that are geographically far apart but have strong demand correlations; they do not consider the periodic fluctuations in computing power demand (such as the differences in demand at different times of the day or different days of the week), and relying solely on adjacent time-series data leads to biased predictions; they do not consider the differences in the degree of influence of different spatial nodes on demand prediction, and cannot focus on key node features to improve accuracy.

[0030] Furthermore, existing technologies cannot integrate the three-dimensional factors of "supply and demand matching, economic benefits (operating revenue, construction costs, maintenance costs, energy consumption, etc.), and service efficiency (network transmission rate, network latency, market response speed, etc.)" to determine the rationality of the layout of existing computing power nodes, thereby guiding actual operational decisions such as shutting down, migrating, or building new computing power resources. This application provides a computing power network demand assessment method that deeply integrates multi-dimensional spatial correlation, demand cycle fluctuations, and differences in the impact of spatial nodes, significantly improving the accuracy of intelligent computing demand prediction. By designing a three-dimensional evaluation system of supply and demand matching, economic benefits, and operational efficiency, the system can predict the rationality of the layout of computing power nodes, ensuring efficient and accurate planning of computing power network resources.

[0031] Figure 1 is a flowchart illustrating a computing power network demand assessment method provided in an embodiment of this application. As shown in Figure 1, in some embodiments, the method includes the following steps: S101, constructing a spatial association network corresponding to the computing power nodes based on the preset spatial relationship types of the computing power nodes.

[0032] In some embodiments, the preset spatial relationship types include at least a first spatial relationship and a second spatial relationship. The first spatial relationship is determined based on the geographical adjacency attribute of the computing power nodes. The second spatial relationship is determined based on the demand similarity attribute of the computing power nodes.

[0033] In some embodiments, the raw data of computing nodes (also called raw computing node data) can be extracted from the computing network management system and the economic analysis system. For example, the raw data of each computing node is extracted based on its identity identifier (ID) and field names. In one implementation, the raw data of the computing nodes undergoes preprocessing and format unification to transform it into a standardized dataset. Based on the standardized dataset, a preset spatial relationship type for the computing nodes is obtained. Preprocessing may include handling missing values, outliers, and duplicate values. Format unification may include data conversion, such as converting the data format into a numeric type. The converted numeric data is then normalized to obtain a standardized dataset.

[0034] In some embodiments, constructing a spatial association network corresponding to computing power nodes includes constructing a local computing power network graph and a global computing power network graph. In one implementation, the local computing power network graph and the global computing power network graph are obtained based on the basic computing power node network graph.

[0035] For example, the basic computing power node network graph can be defined as an undirected graph. .

[0036] Any computing node It is a point in the basic computing power node network graph, that is and Any two computing nodes and The edges of the basic computing power node network graph are denoted as the connection relationships. ,Right now The adjacency relationship between any two computing power nodes in the basic computing power node network graph forms an adjacency matrix. If there is an edge connecting two vertices, the corresponding matrix element is 1; otherwise, it is 0.

[0037] Define nodes In time The The feature values ​​are .in, That is, each node collects F-dimensional feature indicators at each time point, then Represents a node In time All eigenvalues. For example, let F = 5. Wherein, This refers to the scale of computing power demand. CPU utilization Refers to GPU utilization. Refers to memory usage. This refers to hard drive usage. I / O utilization rate This refers to network bandwidth utilization.

[0038] use Indicates a point in time The eigenvalues ​​of all nodes, using Indicates time period The feature values ​​of all nodes within the dataset. Based on historical data of all node feature values, predict future time periods. The computing power requirements of each node .

[0039] S102, obtain multiple types of time-series data corresponding to the computing power nodes.

[0040] Different types of time series data correspond to different time period attributes.

[0041] In some embodiments, the multiple types of time-series data include first time-series data and second time-series data. The first time-series data and the second time-series data have different time period attributes. Based on the multiple types of time-series data, multi-dimensional time correlation information corresponding to computing power nodes is extracted, including: using the first time-series data as first input data to obtain a first time correlation feature; using the second time-series data as second input data to obtain a second time correlation feature; and fusing the first time correlation feature and the second time correlation feature to jointly constitute multi-dimensional time correlation information. For example, the first time-series data can be adjacent data, and the second time-series data can be daily cycle time-series data or weekly cycle time-series data. Another example is that the first time-series data can be daily cycle time-series data or weekly cycle time-series data, and the second time-series data can be monthly cycle time-series data. It is understood that the multiple types of time-series data can include time-series data with multiple different time period attributes, such as the exemplary two different time period attributes, and can also include time-series data with more than three different time period attributes; this application does not limit this. In addition, the "time period attribute" in time series data with different time period attributes can be, for example, "adjacent", "day" or "week", or other time periods such as "month", and this application does not limit it.

[0042] In some embodiments, the multiple types of time-series data may include at least one of instantaneous time-series data (i.e., adjacent data) that is adjacent to the prediction time period, daily periodic time-series data, and weekly periodic time-series data. Both the daily periodic time-series data and the weekly periodic time-series data are extracted from the basic standardized data of the computing power nodes.

[0043] In some embodiments, a local computing power network graph and a global computing power network graph are constructed to extract three types of periodic time-series data: instantaneous time-series data, daily periodic time-series data, and weekly periodic time-series data. Spatial attention mechanisms and spatiotemporal convolutional networks are employed to deeply mine spatial and temporal features. For example, based on the feature dimensions of computing power nodes, three types of periodic time-series data—instantaneous time-series data, daily periodic time-series data, and weekly periodic time-series data—are extracted.

[0044] S103 extracts multi-dimensional time-related information corresponding to computing power nodes based on multiple types of time-series data.

[0045] In some embodiments, for local network structures and global network structures, attention matrices are calculated using Softmax based on three types of periodic time-series data.

[0046] S104, based on spatial correlation networks and multidimensional temporal correlation information, obtains the predicted value of computing power demand through an intelligent computing demand model.

[0047] In some embodiments, a local computing power network graph and a global computing power network graph are constructed, and three types of periodic time-series data—adjacency, daily, and weekly—are extracted. Spatial attention mechanisms and spatiotemporal convolutional networks are used to deeply mine spatial and temporal features, and the prediction results are fused through a fully connected layer to accurately output the predicted value of intelligent computing demand. In one implementation, based on the computing power network graph, attention matrix, and periodic time-series data, spatial and temporal features are extracted using GCN and GRU based on the three types of periodic time-series data, respectively. Based on the spatiotemporal features (spatial and temporal features), the predicted results of intelligent computing demand based on the three types of periodic time-series data are obtained respectively. The prediction results are then fused through weighted averaging to output the final predicted value of intelligent computing demand.

[0048] Among them, the spatiotemporal features are obtained by inputting the computing power network node graph and spatial attention matrix into the spatiotemporal convolution module, extracting the spatiotemporal features of the local and global computing power network node graphs based on three types of periodic time series data: adjacent, daily, and weekly, and then fusing the two spatiotemporal features.

[0049] As shown in Figure 2, an exemplary intelligent computing demand model may include a GCN model for extracting spatial features: the Laplace matrix of the computing power network node graph is defined as follows. ( ; After normalization, it becomes ,in It is an adjacency matrix. It is the identity matrix. It is a diagonal matrix composed of the degrees of each node in the graph.

[0050] Based on three types of periodic time-series data, Chebyshev polynomials are used to extract graph space features based on the normalized Laplacian matrix. The corresponding graph convolution formula is as follows: ,in It is a graph convolution kernel. These are the Chebyshev polynomial coefficients. Meanwhile... ,in It is the largest eigenvalue of the Laplace matrix. ,and , .

[0051] Considering the differences in influence between nodes in the computing power network node graph, a spatial attention mechanism is introduced to calculate... With normalized attention matrix The Hadamard product, the above graph convolution formula is adjusted to... .

[0052] After computation by the GCN module, new feature sequences of the local and global computing power node graphs can be obtained. .

[0053] Spatial feature extraction based on GRU model: using the new feature sequence updated by the GCN module. The GRU model is used to capture the long-term and short-term dependencies of computing power requirements. Key components of the GRU model include the update gate, reset gate, candidate hidden state, and hidden state, calculated using the following formulas: , , and ,in , and For learnable weights, , and Learnable bias terms.

[0054] Local and global spatiotemporal feature fusion: This proposal employs an attention mechanism to fuse local and global spatiotemporal features, combining the output of the GRU model... As input, a nonlinear transformation is performed using the tanh activation function, and the local and global attention values ​​are obtained from the attention vector. , Then, softmax is used for normalization to obtain the weights of local and global spatiotemporal features. , Finally, the output of the spatiotemporal convolution module is obtained. .

[0055] In some implementations, the output of the spatiotemporal feature extraction module is used. Input a fully connected layer and output computing power demand prediction results for three types of time-series historical data: recent, daily periodic, and weekly periodic. , The three prediction results are then fused to obtain the final predicted computing power requirement. The output of the spatiotemporal feature extraction module is then used. As input, the predicted computing power of the fully connected layer is: ,in and These represent the learnable weights and biases, respectively. The activation function is sigmoid. The demand prediction results of the fully connected layer are then used. As input, the result of the demand fusion is: ;in, , and These are learnable weights.

[0056] In some embodiments, when performing step S101, the computing power network demand assessment method includes the following steps: S201, determining a local computing power node network graph based on a first spatial relationship.

[0057] The first spatial relationship is determined based on the geographical adjacency attribute of the computing power nodes.

[0058] As shown in Figure 3, for example, the node structure comprises multiple physically connected computing power nodes (identified by dots, including computing power node E, computing power node A, computing power node F, computing power node D, computing power node C, and computing power node B), and a star-shaped computing power node I. Computing power node I represents the target computing power node. The connection relationships among the multiple computing power nodes include physical connections established only between adjacent computing power nodes. For example, computing power node E is connected to computing power node A, computing power node A is connected to computing power node F, computing power node A is connected to computing power node I, computing power node I is connected to computing power nodes C and B respectively, and computing power node C is connected to computing power nodes D and B respectively. The F-dimensional feature value of node F represents a multi-dimensional feature vector used to quantitatively describe the core attributes of the computing power node (such as computing power performance, memory capacity, current load rate, and remaining storage resources). Each color block corresponds to a dimension (i.e., a specific attribute indicator) in the feature vector. These feature values ​​are combined with the connection relationships between nodes (adjacency matrix) to support subsequent operations such as local computing power scheduling, resource matching, and node performance analysis. For example, these feature values ​​can be used to quickly determine whether a node has idle computing power that can be used by neighboring nodes.

[0059] S202, based on the second spatial relationship, determines the global computing power node network graph.

[0060] The second spatial relationship is determined based on the similarity attribute of the computing power nodes' needs.

[0061] In some embodiments, under the basic computing power node network graph structure, Euclidean distance is used to measure the similarity of computing power demand between nodes, which is used to characterize the connectivity of other computing power nodes that are geographically distant from the computing power nodes.

[0062] Calculate the global feature similarity matrix based on Euclidean distance. : Perform global feature similarity matrix analysis K-Nearest Neighbor (KNN) sparsification yields the adjacency matrix of the global computing power node network graph. ;in, for The Row vectors exist for each node in the global computing power node network graph. A neighboring node with similar business needs.

[0063] Figure 4 illustrates an exemplary implementation of logical connectivity between nodes that are geographically distant but have similar computing power requirements. Solid dot nodes (E, A, and F) represent the corresponding "physical connection nodes" (originally only physically linked to nodes geographically close). Hollow dot nodes (G, H, and J) represent "non-physical connection nodes" (originally without physical links to other nodes). Star node I represents the target computing power node. Auxiliary nodes (a, b, c, and J) represent related sub-units of computing power features, or intermediate identifier nodes participating in similarity calculations. The connections in the figure are not "physical links of geographical proximity," but rather logical connections based on the similarity of computing power requirements. The computing power feature vectors of each node are calculated using "Euclidean distance" to obtain a "global feature similarity matrix." Then... After sparsification using k-Nearest Neighbor (KNN), each node is matched with k neighbor nodes whose computing power requirements are most similar. The connections between nodes in the graph (such as the connection between core node I and non-physically connected node H, and physically connected node A) represent this "logical connectivity driven by similar requirements," corresponding to the edges in the adjacency matrix of the global computing power node network graph. The connectivity relationships between nodes in the graph correspond to the row vectors of the adjacency matrix after KNN sparsification (each row vector represents k similar neighbors of a node).

[0064] S203, based on the local computing power node network graph and the global computing power node network graph, constructs a spatial dimension attention matrix to obtain a spatial correlation network.

[0065] In some embodiments, the attention matrix of the local computing power node network graph is calculated by taking three types of periodic time-series data as input, applying a linear transformation to the data using the learned weight matrix, introducing a non-linear factor through a sigmoid activation function, and then calculating the attention matrix. And the attention matrix is ​​processed using the softmax function. Normalization is performed to obtain the normalized attention matrix. .

[0066] ; ;in , , , These represent adjacent, daily, and weekly data, respectively. , and Indicates learnable weights, This represents the learnable bias term. This represents the sigmoid activation function. , These are attention matrices and The element represents a node. and The strength of the correlation between them.

[0067] In some embodiments, the attention matrix of the global computing power node network graph is calculated by taking three types of periodic time-series data as input, applying a linear transformation to the data using the learned weight matrix, introducing a non-linear factor through a sigmoid activation function, and then calculating the attention matrix. And the attention matrix is ​​processed using the softmax function. Normalization is performed to obtain the normalized attention matrix. .

[0068]

[0069]

[0070] in , , , These represent adjacent, daily, and weekly data, respectively. , and Indicates learnable weights, This represents the learnable bias term. This represents the sigmoid activation function. , These are attention matrices and The element represents a node. and The strength of the correlation between them.

[0071] In some embodiments, based on multiple types of time-series data, multi-dimensional time-related information corresponding to computing power nodes is extracted. The computing power network demand assessment method includes the following steps: S301, taking time-series data as the first input data to obtain the first time-related features.

[0072] For example, suppose the current time is The predicted time span is And the sampling frequency of the feature value of each computing node is The time series data directly adjacent to the forecast period time window is defined as follows: .

[0073] S302, using daily periodic time series data as the second input data, obtains the second time-related feature.

[0074] For example, the definition and the past The time series data for the same time period as the forecast period are as follows: This is used to characterize the daily periodicity of computing power demand.

[0075] S303 uses the periodic time series data as the third input data to obtain the third time correlation feature.

[0076] The computing power demand on weekdays typically shares some similarities with historical weekday computing power demand, but differs significantly from weekend computing power demand. (Defining past...) The time-series data for the same time period and the same working day as the midweek and forecast periods are as follows:

[0077]

[0078] It is used to characterize the periodicity of computing power demand.

[0079] For example, as shown in Figure 5, the horizontal axis is a time axis, covering the period from August 4, 2026 (Wednesday) to August 18, 2026 (Tuesday), and the corresponding date is marked with "weekday" + "specific time period" (such as "Monday 8:00-10:00"). This represents the current time (around 8:00 AM on Tuesday, August 18, 2026). This includes Sunday, August 16, 2026, from 8:00 to 10:00, and Monday, August 17, 2026, from 8:00 to 10:00. This includes Tuesday, August 4, 2026, from 8:00 to 10:00, and Tuesday, August 11, 2026, from 8:00 to 10:00. This includes Tuesday, August 18, 2026, from 0:00 to 8:00. This is indicated as Tuesday, August 18, 2026, from 8:00 to 10:00.

[0080] S304 integrates the first time-related features, the second time-related features, and the third time-related features to form multi-dimensional time-related information.

[0081] In some embodiments, considering that any node in the computing power node network graph is affected by other different nodes to varying degrees, this proposal introduces a spatial attention mechanism to characterize the degree of mutual influence between nodes based on three types of periodic time-series data: adjacent, daily, and weekly.

[0082] By executing steps S301 to S304, and taking into account the similarity of computing node business needs at different times of the day and different workdays of the week, i.e., exhibiting certain periodic characteristics, the periodic factor of time series data is introduced.

[0083] In some embodiments, the method further includes the following step: S401, inputting the predicted computing power demand value into the trained random forest model to obtain the rationality judgment result of the layout of computing power nodes.

[0084] S402, in response to the layout rationality judgment result being that the layout is unreasonable, generates a corresponding computing power layout unreasonable warning and optimization request.

[0085] In some embodiments, the system automatically triggers warnings and optimization requests for computing power nodes that are deemed unreasonable. Based on these warnings and requests, targeted computing power layout adjustment strategies are executed to optimize resource allocation and improve overall performance.

[0086] In some embodiments, the method further includes: S501, constructing a three-dimensional indicator framework of supply and demand, benefits and efficiency.

[0087] S502 collects raw data from computing nodes based on a three-dimensional index framework, performs data preprocessing and data format conversion on the raw data in sequence, and generates standardized basic data.

[0088] In some embodiments, the raw data of computing nodes is collected. Influencing factors can be identified based on the rationality of the computing network resource layout, and a multi-dimensional indicator system of supply, demand, efficiency, and effectiveness can be constructed. This system includes: supply and demand side (A1 - supply and demand matching degree), effectiveness side (A2 - revenue-cost ratio), and efficiency side (A3 - CPU utilization rate, A4 - GPU utilization rate, A5 - network bandwidth utilization rate, and A6 - response time, etc.). The supply and demand matching degree refers to whether the supply and demand of computing nodes are balanced. If the supply exceeds the demand, there is oversupply, and the supply and demand matching degree = 1; otherwise, it = 0. The revenue-cost ratio = computing node operating revenue / computing node operating cost. The computing node operating cost includes construction costs, maintenance costs, and energy consumption. CPU utilization rate = computing node CPU running time / total running time; GPU utilization rate = computing node GPU working time / total running time; and network bandwidth utilization rate = computing node bandwidth occupied / total network bandwidth.

[0089] The computing network management system and the economic analysis system automatically extract the raw data of each computing node based on the computing node ID and field name, perform data preprocessing to handle issues such as missing values, outliers, and duplicate values, and perform data conversion to unify the data format into a numerical type. Finally, the data is normalized to output standardized data.

[0090] S503 divides the standardized basic data into training and test sets, builds an initial random forest model based on the training set, and uses the test set to evaluate and adjust the parameters of the initial random forest model to obtain the trained random forest model.

[0091] In some embodiments, a random partitioning method is used to divide the standardized data output by the data preparation module into a training set and a test set. For example, the initial proportion of the training set is set to 80%, and this training set is input into the model training module to train the random forest model. The 20% of standardized data is used as the test set and input into the model testing module to evaluate the performance of the random forest model. Parameters are then adjusted to optimize the model and generate the optimal random forest model.

[0092] In some embodiments, obtaining the trained random forest model includes the following steps: S5031, setting the parameters of the random forest model.

[0093] Set the relevant parameters of the random forest model, including the number of decision trees, the maximum number of features in the decision tree, the decision tree node splitting criteria (information gain, Gini coefficient, etc.), and the maximum depth of the decision tree.

[0094] S5032, Extract training samples.

[0095] Based on the training dataset, N samples are drawn with replacement, one sample at a time, resulting in N samples forming a subset D. This subset is used as the root node of a decision tree to train the tree. The number of subsets k is equal to the number of decision trees set in the random forest model parameters. Simultaneously, for each sample, m features are randomly selected from six features for training each decision tree.

[0096] S5033, training a decision tree model.

[0097] Based on the set model parameters and sampled subsets, multiple decision trees are constructed using the CART algorithm. The specific technical solution is as follows: S50331, calculate the Gini coefficient of the root node.

[0098] CART trees are binary trees with two node classification categories: Gini coefficient. ,in Let be the probability that a sample belongs to class 1. Given a training subset D, based on the features... A certain value 'a', based on sample points Is it equal to a? Divide D into (satisfy )and (Not satisfied) ),calculate Gini coefficient at time .

[0099] S50332, split the root node. Then traverse each feature of the dataset D and each of their split points, select the feature with the smallest Gini coefficient and its corresponding split point as the optimal feature and optimal split point. Based on the optimal feature and optimal split point, generate two child nodes from the current node, and distribute the dataset to the two child nodes to generate two subsets.

[0100] S50333, iteratively calculates child nodes.

[0101] Based on the classification at the root node, repeat steps ① and ② for the two generated child nodes respectively, traversing all features in the subset except for the optimal feature of the root node, searching for the optimal feature with the smallest Gini coefficient and the optimal split point, and continuing to generate two child nodes. Repeat the above process iteratively until the maximum depth condition of the decision tree is met.

[0102] S50334 forms a random forest.

[0103] For k subsets of data, repeat steps S50331 to S50333 to generate k decision trees. Combine them to form a random forest. After calculating the classification results of each tree, use a voting method to determine the final classification result.

[0104] S5034, Test decision tree model.

[0105] Input the test set into the random forest model described above, and compare the model's output with the actual results on the test set. Evaluate the performance of the random forest model using metrics such as accuracy, precision, recall, and F1 score. Based on the evaluation results, use cross-validation to adjust the parameters of the random forest model, such as the number of decision trees and the maximum number of features, to optimize the model's performance and generate a finely tuned random forest model.

[0106] S5035 predicts unreasonable layout of computing nodes.

[0107] The demand forecast results of the aforementioned intelligent computing demand forecasting module Input the optimized random forest model, determine whether the layout of each computing node is reasonable at future time steps, and output the determination result.

[0108] S5036 sends an alert and optimization request for unreasonable automated computing power layout.

[0109] Using a monthly time window, for computing power nodes that are predicted to have an unreasonable layout within the time period, an early warning is sent to the computing network brain through the node-side AI device, automatically starting a low-energy operation mode and triggering a computing power layout optimization request.

[0110] Thus, the standardized dataset in step S101 is divided into a training set and a test set, and the random forest algorithm is used for model training and optimization to obtain the optimal model. Subsequently, the predicted demand values ​​obtained in step S104 are input into this random forest model to achieve intelligent prediction of the rationality of the computing node layout.

[0111] The following is a feasible implementation of the computing power network demand assessment method.

[0112] As shown in Figure 6, in some embodiments, the computing power network demand assessment method includes the following steps: S1, data preparation stage.

[0113] Raw computing node data is extracted from the computing network management system and economic analysis system, and after preprocessing and format unification, it is transformed into a standardized dataset.

[0114] In some embodiments, performing step S1 includes the following steps: S11, collecting data.

[0115] Raw computing node data is extracted from the computing network management system and the economic analysis system.

[0116] S12, Preprocessed data.

[0117] Raw computing node data is extracted from the computing network management system and economic analysis system and preprocessed.

[0118] S13, convert the data format, and finally output standardized data.

[0119] Raw computing node data is extracted from the computing network management system and economic analysis system, and after preprocessing and format unification, it is transformed into a standardized dataset.

[0120] By executing steps S11 to S13, we provide the standardized data required for computing power analysis in the intelligent demand prediction stage and the standardized dataset for model training in the layout rationality prediction stage.

[0121] S2, Intelligent Demand Forecasting Stage.

[0122] Based on the standardized data output from the data preparation phase, intelligent demand prediction is completed. A local and global computing power network graph is constructed, and three types of periodic time-series data—adjacency, daily, and weekly—are extracted. Spatial attention mechanisms and spatiotemporal convolutional networks are used to deeply mine spatial and temporal features, and the prediction results are fused through fully connected layers to accurately output predicted values ​​for intelligent computing demand.

[0123] In some embodiments, when performing step S2, the following step is further included: S21, constructing a computing power network graph.

[0124] In some embodiments, a local computing power network graph and a global computing power network graph are constructed based on the standard data obtained in step S1.

[0125] S22, Extract periodic data.

[0126] S23, Computational spatial attention.

[0127] S24, extract spatiotemporal features.

[0128] S25 completes the computing power demand forecast and outputs the "demand forecast result".

[0129] S3, the layout rationality prediction stage.

[0130] Simultaneously receive standardized data output from the data preparation phase to complete the prediction of layout rationality.

[0131] In some embodiments, when performing step S3, the following step is further included: S31, dividing the training and testing datasets.

[0132] The standard data obtained from S1 is divided into training and test sets.

[0133] S32, Training the Random Forest Model.

[0134] The random forest algorithm was used for model training and optimization to obtain the optimal model.

[0135] S33, Evaluate and optimize the random forest model.

[0136] S34 uses the optimized model to predict the rationality of computing node layout.

[0137] Input the predicted demand value obtained from S2 into this model to achieve intelligent prediction of the rationality of computing node layout.

[0138] S4, Layout optimization and planning stage.

[0139] By combining the "demand forecast results" from the intelligent demand forecasting stage with the forecast conclusions from the layout rationality forecasting stage, layout optimization and planning are carried out.

[0140] S41 triggered an alert for unreasonable computing power layout.

[0141] For computing nodes with unreasonably predicted computing power, the system automatically triggers warnings and optimization requests. Based on these warnings and requests, targeted computing power layout adjustment strategies are implemented to optimize resource allocation and improve overall performance.

[0142] S42 triggers a computing power layout optimization request.

[0143] S43 executes the operation of optimizing the allocation of computing resources.

[0144] This application also provides a computing power network demand assessment device for performing the above-described computing power network demand assessment method.

[0145] As shown in Figures 7a and 7b, in some embodiments, the computing power network demand assessment device includes: an intelligent demand prediction platform and a computing power layout inappropriate prediction system.

[0146] The intelligent demand prediction platform (also known as the intelligent computing demand precision prediction platform) may include an intelligent computing demand prediction module based on an attention-based multi-scale temporal-spatial graph convolutional network (AMSTGCN) (which can be referred to as the intelligent computing demand prediction module). In this module, to address the impact of potential cross-regional connections between computing nodes on computing demand prediction, a multi-dimensional spatial computing node network graph is constructed. This includes a local computing node network graph based on the actual geographical location of the computing nodes, and a global computing node network graph constructed by calculating the demand similarity between cross-regional computing nodes using Euclidean distance. To address the periodicity of intelligent computing demand, three types of time-series historical data—recent, daily, and weekly—are extracted as the basic input for intelligent computing demand prediction. To address the differences in the degree of influence between different computing nodes, the spatial dimension attention matrix of the local and global computing node network graphs is calculated. The spatiotemporal features of the two types of computing node network graphs are extracted and fused based on the spatiotemporal convolutional module (ST-Block), and these features are input into a fully connected layer to output the predicted intelligent computing demand value.

[0147] The computing power layout inappropriate prediction system includes a multi-dimensional computing power layout rationality judgment module based on random forest-based supply and demand efficiency. In the intelligent computing layout rationality judgment module, a multi-dimensional index system of supply and demand efficiency is constructed, trained and tested using a random forest algorithm, and dynamically and intelligently judged on the rationality of computing power layout based on the prediction results of the intelligent computing power demand prediction module. It also relies on the system device to provide early warning of nodes with inappropriate computing power layout and to intelligently adjust and optimize the layout.

[0148] As shown in Figure 7a, in some embodiments, the intelligent computing demand prediction platform is an intelligent computing demand prediction device that predicts the short- and long-term intelligent computing demands of each computing power node in the existing network. In one implementation, the intelligent computing demand prediction device includes a computing power node basic data preparation module, a network graph construction module, a periodic data extraction module, a spatial attention module, a spatiotemporal convolution module (including GCN and GRU), and a computing power demand prediction module.

[0149] The data preparation module is configured to input the raw data of the computing nodes extracted by the computing network management system into the data preparation module, perform data preprocessing and data format conversion, and obtain standardized data.

[0150] The network graph construction module is configured to input standardized data to build a computing power network graph. Specifically, the local computing power network graph is built based on the geographical adjacency of computing power nodes, while the global computing power network graph is built based on the demand similarity of computing power nodes.

[0151] The periodic data extraction module is configured to input standardized data into the periodic data extraction module and extract three types of periodic time series data: adjacent, daily, and weekly, based on the feature dimension of computing power nodes.

[0152] The spatial attention module is configured to input the computing network graph and periodic temporal data into the spatial attention module. For local and global network structures, attention matrices are calculated using Softmax based on three types of periodic temporal data.

[0153] The computational network graph, attention matrix, and periodic temporal data are input into the spatiotemporal convolution module. Based on the three types of periodic temporal data, spatial and temporal features are extracted using GCN and GRU respectively.

[0154] The spatiotemporal features are input into the fully connected layer to obtain intelligent computing demand prediction results based on three types of periodic time series data. The prediction results are then fused by weighted averaging to output the final intelligent computing demand prediction value.

[0155] As shown in Figure 7b, in some embodiments, there is a computing power layout unreasonable prediction system: This system is a computing power layout unreasonable discrimination device, which realizes the prediction of the multi-dimensional benefits and efficiency of computing power nodes. This part includes a data preparation module, a model training module, a model testing module, and a computing power layout unreasonable prediction module.

[0156] The data preparation module is configured to input the raw data of computing nodes extracted from the computing network management system and the subsystem into the data preparation module, perform data preprocessing and data format conversion, and obtain standardized data.

[0157] The model training module is configured to split the standardized data, with the initial proportion of the training set set to 80%, and input the training set into the model training module to train the random forest model.

[0158] The model testing module is configured to input 20% of the standardized data as a test set into the model testing module to evaluate the performance of the random forest model, adjust parameters to optimize the model, and generate the optimal random forest model.

[0159] The prediction module is configured to input the predicted value of intelligent computing demand from the intelligent computing demand precision prediction platform into the prediction module, and obtain the rationality judgment result of the computing power node to be judged based on the trained and tested random forest model.

[0160] This application also provides a computing power network demand assessment device. This device can be a blockchain-based intelligent computing demand prediction and intelligent early warning device for unreasonable computing power allocation. The device provides users with private data asset query, management, and accurate prediction and early warning functions, and leverages blockchain technology to ultimately realize the aforementioned methods and technologies.

[0161] As shown in Figure 8, in some embodiments, the computing power network demand assessment device includes: a distributed data asset management terminal, a computing power node data asset think tank management platform, an intelligent computing demand prediction platform, an intelligent early warning platform for unreasonable computing power node layout, an integrated optimization center for computing power layout, and a computing power resource visualization system.

[0162] The distributed data asset management terminal, as the core of user operations, fully supports users in managing their own data assets comprehensively, efficiently, and securely. In some embodiments, the distributed data asset management terminal includes a data receiving and storage module, a data asset management module, an identity authentication and authorization module, a key management and encryption module, a data asset token management module, and a data backup and recovery module.

[0163] The data receiving and storage module features a standardized interface design and efficiently achieves data reception, transmission, and distributed storage through network protocols, ensuring seamless integration of business information flow and function execution.

[0164] The data asset management module includes data asset catalog management, which supports operations such as catalog creation, updating, deletion, and querying; data asset library management, which provides functions such as data classification, tagging, searching, and statistics; and data lifecycle management, which manages the entire lifecycle of data, including generation, reception, storage, deletion, and modification.

[0165] The identity authentication and authorization module supports user identity information management, including user registration, updating, and deletion. The identity authentication mechanism uses multi-factor authentication to ensure the authenticity of user identity. The permission management is based on role or attribute-based permission control to ensure that users can only access data assets within their authorized scope.

[0166] The key management and encryption module supports the generation and local or distributed storage of symmetric and asymmetric keys. The data encryption and decryption module uses advanced encryption algorithms to ensure the security of data during transmission and storage. The key lifecycle management module includes the entire lifecycle management of keys, such as generation, distribution, update, revocation, and destruction.

[0167] The Data Asset Token Pass Management Module supports the generation and storage of data asset access tokens, token circulation and verification ensure the security and validity of tokens during transmission, and token tracing and auditing records the circulation trajectory of tokens and provides auditing functions.

[0168] The data backup and recovery module formulates reasonable data backup strategies and provides fast data recovery functions to reduce losses caused by data loss.

[0169] The computing power node data asset management platform is a device for managing the computing power node data asset repository. It enables the collection of full data from computing power nodes, the generation and revision of computing power node data catalogs, the creation and classification of computing power node data tags, and the storage of data in a distributed ledger. The platform mainly includes a data acquisition module, a data catalog management module, a data tag management module, and a data storage module.

[0170] The intelligent computing demand prediction platform is a device for predicting the demand of computing nodes, enabling the prediction of the short-term and long-term demand of computing nodes. It includes a data preparation module, a computing network graph construction module, a periodic data extraction module, a spatial attention module, a spatiotemporal convolution module, and a computing demand prediction module.

[0171] The data preparation module, based on the distributed data asset management terminal and the trusted chain interaction and rights confirmation process, supports users to select computing power nodes by ID and realize the retrieval and standardized processing of data from the selected computing power nodes.

[0172] The computing power network graph construction module, based on trusted chain interaction for rights confirmation, supports users in generating local and global computing power network graphs based on selected computing power nodes.

[0173] The periodic data extraction module, based on trusted chain interaction and rights confirmation, allows users to extract periodic data for selected computing power nodes.

[0174] The Time Attention module, based on trusted chain interaction for rights confirmation, supports users in automatically completing attention calculations using an embedded Python package.

[0175] The spatiotemporal convolution module, consisting of GCN and TCN, supports users in extracting spatiotemporal features from the computing power network graph.

[0176] The computing power demand prediction module, based on the interaction and rights confirmation with the trusted chain, allows users to select a prediction time window and supports users to automatically obtain computing power demand prediction results based on an embedded Python program package.

[0177] The intelligent early warning platform for unreasonable computing node layout is a platform for predicting and warning of unreasonable computing node layout. It can judge the rationality of computing node layout, including a random forest model training module, a random forest model testing module, and a judgment result output module.

[0178] The integrated optimization center for computing power layout is a center for early warning processing and optimization of unreasonable computing power layout. Based on the early warning and optimization request information output by the early warning platform, it realizes the optimization and adjustment of computing power layout. It includes an early warning receiving module for unreasonable computing power layout, a node location module for unreasonable computing power layout, a node optimization module for unreasonable computing power layout, and a computing power layout optimization result reporting module.

[0179] The computing power resource visualization system enables the visualization of basic data of computing power nodes, computing power network diagrams, and nodes with unreasonable computing power layouts through data communication and transmission between devices 1-5, as well as user identity authentication and on-chain data storage.

[0180] This application constructs multi-dimensional spatial relationships among computing power nodes. It characterizes local spatial connections through the geographical location relationships of these nodes and establishes global spatial connections using Euclidean distance similarity. This accurately captures the connections between geographically distant but strongly related computing power nodes, avoiding the omission of cross-regional computing power demand correlation information and thus improving the accuracy of computing power demand prediction. Including geographically distant but highly correlated computing power nodes in the local computing power demand prediction criteria solves the problem of missing long-distance spatial dependency metrics between computing power nodes. It more closely reflects the actual spatial relationships between computing power nodes in the computing power network, improving the accuracy of intelligent computing demand prediction.

[0181] This application fully considers the periodic fluctuations in computing power demand. In addition to incorporating time-series data adjacent to the prediction period, it also introduces non-adjacent time-series data from the same time period within a day and the same time period within a week as model inputs. The prediction results from these three types of data are then integrated to adapt to the temporal regularity of computing power demand, ensuring that the prediction results accurately reflect time-specific differences. By fully considering various periodic fluctuation factors in intelligent computing demand, the accuracy of predictions is effectively improved.

[0182] This application enables differentiated measurement of the influence of spatial nodes. By introducing an attention mechanism in the spatial dimension, it assigns differentiated weights to different computing power nodes corresponding to the nodes to be predicted, accurately identifying key nodes that significantly affect the prediction results, enhancing model focus, and further improving the accuracy of intelligent computing demand prediction. For the multi-dimensional computing power node network graph, the differences in the influence of different computing power nodes on the nodes to be predicted are measured separately. The spatial convolution module effectively captures the historical demand features of computing power nodes with high influence, and the attention mechanism fuses these features. This improves and refines the spatial influence measurement of intelligent computing demand prediction, enhancing its accuracy.

[0183] This application fills the technological gap in identifying irrationally distributed computing nodes. Based on computing node traffic prediction results, it integrates a three-dimensional framework of supply and demand matching, economic benefits (including operating revenue, construction costs, maintenance costs, energy consumption, etc.), and service efficiency (including network transmission rate, network latency, market response speed, etc.). It employs a random forest algorithm to generate an irrationally distributed node discrimination model, enabling the prediction of the rationality of computing node distribution. This provides effective decision support for the closure, migration, or construction of computing resources, avoiding waste of computing resources and insufficient service capacity. The random forest algorithm generates a random forest model to judge the rationality of computing node distribution, predicting whether the future computing node distribution is reasonable. Based on the prediction results, it enables the adjustment and optimization of the computing network resource layout, supporting the efficient formulation of computing network resource planning schemes.

[0184] In summary, this application can comprehensively address the shortcomings of existing technologies in terms of the accuracy of computing power demand prediction and the optimization of computing power node layout, providing efficient and accurate technical support for computing power operation enterprises.

[0185] In some solutions, multiple embodiments of this application can be combined, and the combined solution can be implemented. Optionally, some operations in the processes of each method embodiment may be combined, and / or the order of some operations may be changed. Furthermore, the execution order between the steps of each process is merely exemplary and does not constitute a limitation on the execution order between steps; other execution orders are also possible. It is not intended to indicate that the execution order is the only possible order in which these operations can be performed. Those skilled in the art will conceive of various ways to reorder the operations described herein. In addition, it should be noted that the process details involved in one embodiment of this document are similarly applicable to other embodiments, or different embodiments may be combined.

[0186] Furthermore, some steps in the method embodiments can be equivalently replaced with other possible steps. Alternatively, some steps in the method embodiments may be optional and can be deleted in certain use cases. Or, other possible steps may be added to the method embodiments. Moreover, the various method embodiments can be implemented individually or in combination.

[0187] According to embodiments of the present application, the present application also provides an electronic device, a readable storage medium, and a computer program product.

[0188] Figure 9 is a schematic block diagram of an example electronic device provided in an embodiment of this application. As shown in Figure 9, the electronic device 1000 includes a computing unit 1001, which can perform various appropriate actions and processes according to a computer program stored in ROM (Read-Only Memory) 1002 or a computer program loaded from storage unit 1008 into RAM (Random Access Memory) 1003. The RAM 1003 may also store various programs and data required for the operation of the electronic device 1000. The computing unit 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An I / O (Input / Output) interface 1005 is also connected to the bus 1004.

[0189] Multiple components in electronic device 1000 are connected to I / O interface 1005, including: input unit 1006, such as keyboard, mouse, etc.; output unit 1007, such as various types of monitors, speakers, etc.; storage unit 1008, such as disk, optical disk, etc.; and computing network demand assessment unit 1009, such as network card, modem, wireless computing network demand assessment transceiver, etc. The computing network demand assessment unit 1009 allows electronic device 1000 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0190] The computing unit 1001 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, CPUs (Central Processing Units), GPUs (Graphics Processing Units), various special-purpose AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, DSPs (Digital Signal Processors), and any suitable processor, controller, microcontroller, etc. The computing unit 1001 performs the various methods and processes described above, such as the computing power network demand assessment method. For example, in some embodiments, the computing power network demand assessment method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1008. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 1000 via ROM 1002 and / or computing power network demand assessment unit 1009. When the computer program is loaded into RAM 1003 and executed by computing unit 1001, one or more steps of the method described above can be performed. Alternatively, in other embodiments, computing unit 1001 can be configured to perform the aforementioned computing power network demand assessment method by any other suitable means (e.g., by means of firmware).

[0191] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application-Specific Standard Products), SOCs (System-on-Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0192] Program code used to implement the methods of the embodiments of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0193] In the context of embodiments of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only Memory) or flash memory, optical fibers, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0194] To provide interaction with the external environment, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the external environment (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball), through which the external environment can provide input to the computer. Other types of devices can also be used to provide interaction with the external environment; for example, feedback provided to the external environment can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the external environment can be received in any form (including sound input, voice input, or tactile input).

[0195] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., external environment computers with a graphical external environment interface or web browser, through which the external environment can interact with the implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected through digital data computing power network demand assessments (e.g., computing power network demand assessment networks) of any form or medium. Examples of computing power network demand assessment networks include: LANs (Local Area Networks), WANs (Wide Area Networks), the Internet, and blockchain networks.

[0196] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact through a network of computing power demand assessments. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, also known as cloud computing servers or cloud hosts, a hosting product within the cloud computing service system that addresses the shortcomings of traditional physical hosts and VPS (Virtual Private Server, or simply "VPS") services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0197] It's important to note that artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0198] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for assessing computing power network demand, characterized in that, The method includes: constructing a spatial association network corresponding to the computing power node based on a preset spatial relationship type of the computing power node; acquiring multiple types of time-series data corresponding to the computing power node; the multiple types of time-series data corresponding to different time period attributes; extracting multi-dimensional time association information corresponding to the computing power node based on the multiple types of time-series data; and obtaining a predicted value of computing power demand through an intelligent computing demand model based on the spatial association network and the multi-dimensional time association information.

2. The computing power network demand assessment method according to claim 1, characterized in that, The preset spatial relationship types include at least a first spatial relationship and a second spatial relationship; The step of constructing a spatial association network corresponding to computing power nodes based on a preset spatial relationship type includes: determining a local computing power node network graph based on the first spatial relationship; determining the first spatial relationship based on the geographical adjacency attribute of the computing power nodes; determining a global computing power node network graph based on the second spatial relationship; determining the second spatial relationship based on the demand similarity attribute of the computing power nodes; and constructing a spatial dimension attention matrix based on the local computing power node network graph and the global computing power node network graph to obtain the spatial association network.

3. The computing power network demand assessment method according to claim 2, characterized in that, The multiple types of time-series data include first time-series data and second time-series data; the first time-series data and the second time-series data have different time period attributes; the step of extracting multi-dimensional time-related information corresponding to the computing power node based on the multiple types of time-series data includes: using the first time-series data as first input data to obtain a first time-related feature; using the second time-series data as second input data to obtain a second time-related feature; and fusing the first time-related feature and the second time-related feature to jointly constitute the multi-dimensional time-related information.

4. The computing power network demand assessment method according to any one of claims 1-3, characterized in that, The method further includes: inputting the predicted computing power demand value into the trained random forest model to obtain the layout rationality judgment result of the computing power nodes; and generating a corresponding computing power layout unreasonable warning and optimization request in response to the layout rationality judgment result being unreasonable.

5. The computing power network demand assessment method according to claim 4, characterized in that, The method further includes: constructing a three-dimensional index framework of supply and demand, benefits and efficiency; collecting the raw data of the computing nodes based on the three-dimensional index framework, performing data preprocessing and data format conversion on the raw data in sequence to generate standardized basic data; dividing the standardized basic data into training set and test set, constructing an initial random forest model based on the training set, and using the test set to evaluate and adjust the parameters of the initial random forest model to obtain the trained random forest model.

6. A computing power network demand assessment device, characterized in that, The computing power network demand assessment device is used to perform the method according to any one of claims 1 to 5.

7. The computing power network demand assessment device according to claim 6, characterized in that, Also includes: A distributed data asset management terminal, used to manage proprietary data assets based on distributed blockchain technology; The computing power node data asset think tank management platform is used to collect full data of computing power nodes based on distributed blockchain technology, generate and revise the computing power node data catalog, formulate and classify computing power node data tags, and store the data in a distributed ledger.

8. An electronic device, characterized in that, The electronic device is used to perform the method according to any one of claims 1 to 5.

9. A non-transient computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 5.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 5.

Images