A method and system for cross-domain computing power scheduling based on blockchain and digital twins

Abstract

This invention discloses a cross-domain computing power scheduling method and system based on blockchain and digital twins. The cross-domain computing power scheduling method includes: responding to a user-initiated computing power demand task, obtaining available computing power nodes in computing power centers of various regions, and identifying and storing the computing power demand task and available computing power nodes through a blockchain network; using digital twin technology to perform dynamic twin mapping of the computing power centers in each region and generating feature parameters of available computing power nodes; using a hyperbolic causal graph reinforcement learning engine to obtain causal representations and hierarchical geometric representations of the computing power network based on the feature parameters, and generating a computing power scheduling strategy based on the causal representations and hierarchical geometric representations of the computing power network, thereby realizing cross-domain computing power scheduling. This invention solves the scheduling suboptimal problems caused by opaque cost metrics, dynamic network topology distortion, and multivariate confusion in cross-domain computing power scheduling, achieving efficient, economical, and reliable flow of computing power resources.

Description

Technical Field

[0001] This invention relates to the field of distributed computing and artificial intelligence, specifically to a cross-domain computing power scheduling method and system based on blockchain and digital twins. Background Technology

[0002] The rapid development of the digital economy has driven the continuous implementation of new businesses such as large-scale AI model training, autonomous driving simulation, and metaverse scene rendering, leading to a sustained high-speed growth in the demand for computing resources from these applications. Against the backdrop of nationwide integrated computing power scheduling and the comprehensive advancement of the "Eastern Data, Western Computing" project, building a unified and collaborative computing power scheduling network has become an inevitable direction for the development of computing infrastructure.

[0003] Due to factors such as natural conditions, energy distribution, and land costs, the construction and operation costs of computing centers vary significantly across different regions. Even when providing computing power services of the same specifications, there are substantial differences in the underlying physical costs, such as electricity prices and cooling energy consumption. Taking the comparison between the eastern and western regions as an example, data centers in the west have significant advantages in energy prices and natural heat dissipation conditions, which can effectively reduce computing power production costs. However, due to limitations in physical transmission distance, business access latency is generally higher than in the eastern region, creating an inherent contradiction between cost and latency.

[0004] Current mainstream cross-domain computing power scheduling technologies and systems still have significant technical shortcomings in practical applications: the lack of unified standards for computing power metrics and weak trust foundation for cross-domain scheduling. Most regional computing centers operate independently, lacking unified identification and measurement standards for computing power output. Due to factors such as the confidentiality of commercial data, computing power providers typically do not disclose core information such as hardware power consumption and actual operating costs. Scheduling platforms struggle to obtain accurate data for comparing computing power across the entire network, resulting in significant trust barriers in cross-domain computing power transactions and scheduling.

[0005] In addition, existing computing power management platforms mostly adopt traditional log reporting monitoring methods such as Zabbix and Prometheus, which make it difficult to capture key physical characteristics such as the operating wear and tear of computing power equipment and the dynamic changes in data center PUE in real time. Computing power pricing mostly adopts a fixed static model, which cannot match the actual operating status such as real-time load, ambient temperature, and bandwidth usage, and cannot truly reflect the dynamic cost-effectiveness of computing power in different regions. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a cross-domain computing power scheduling method and system based on blockchain and digital twins. It aims to solve the scheduling suboptimal problems caused by opaque cost measurement, dynamic network topology distortion, and multivariate confusion in cross-domain computing power scheduling, and realizes efficient, economical, and reliable flow of computing power resources.

[0007] This invention discloses a cross-domain computing power scheduling method based on blockchain and digital twins, comprising: S1. In response to the computing power demand task initiated by the user, obtain the available computing power nodes of the computing power centers in each region, and identify and store the computing power demand task and available computing power nodes through the blockchain network. S2. Utilize digital twin technology to perform dynamic twin mapping of the computing power center in each region and generate characteristic parameters of available computing power nodes. The characteristic parameters include computing power performance parameters, network latency parameters, and dynamic pricing costs. S3. Input the feature parameters of the available computing power nodes into the pre-built hyperbolic causal graph reinforcement learning engine to obtain the causal representation of the computing power network and the hierarchical geometric representation. The hyperbolic causal graph reinforcement learning engine includes a spatiotemporal causal attention inference module and a hyperbolic graph embedding module. S4. Based on the causal representation and hierarchical geometric representation of the computing power network, a computing power scheduling strategy is generated through a hyperbolic causal graph reinforcement learning engine, and cross-domain computing power scheduling is realized according to the computing power scheduling strategy.

[0008] Preferably, the computing power requirement tasks and available computing power nodes are identified and stored through a blockchain network, including: S11. Extract features from the computing power requirement task to obtain a first feature vector, and generate a first hash identifier based on the first feature vector. The first feature vector includes computing power requirement features, memory requirement features and latency tolerance features. S12. Extract features from available computing power nodes to obtain a second feature vector, and generate a second hash identifier based on the second feature vector. The second feature vector includes heterogeneous hardware specification features, current concurrent task count features, and region identifier features. S13. Use an asymmetric encryption algorithm to digitally sign the first hash identifier and the second hash identifier, and write the signature result into the blockchain distributed ledger for storage.

[0009] Preferably, the blockchain distributed ledger is deployed with a computing power pricing smart contract, which integrates zero-knowledge verification parameters. These parameters are used to verify the authenticity and compliance of the computing power price of available computing power nodes.

[0010] Preferably, digital twin technology is used to perform dynamic twin mapping of the computing power center in each region and generate characteristic parameters of available computing power nodes, including: S21. Deploy a digital twin agent in each computing center. The digital twin agent collects the operating status parameters of the physical server corresponding to the computing center at a preset frequency. The operating status parameters include environmental sensor data, network monitoring data and computing node operating data. S22. Based on the operating status parameters, construct a dynamic virtual mapping of available computing power nodes from physical state to virtual state in the digital twin engine; S23. Collect the computing power performance parameters and network latency parameters of available computing power nodes based on the dynamic virtual mapping, and calculate the dynamic pricing cost of available computing power nodes based on the running status parameters.

[0011] Preferably, the environmental sensing data includes the temperature of the rack's inlet and outlet air vents and the power consumption of the cooling system; Network monitoring data includes switch queue depth and packet loss rate; Compute node operation data includes compute unit load rate, temperature, and utilization rate; The dynamic pricing cost of available computing power nodes is calculated based on operational status parameters, including: The dynamic pricing cost is calculated using a multidimensional cost function, which is expressed by the following formula: ; in, Indicates available computing power nodes In time slice Dynamic pricing costs within the scope, Indicates the device's base power consumption. Indicates energy efficiency. This indicates tiered electricity pricing by region. This represents the computing power generated per unit of time. This indicates the dynamic depreciation rate of the hardware. Indicates available uplink bandwidth. This indicates the preset bandwidth threshold. All of these represent weighting coefficients.

[0012] Preferably, the feature parameters of available computing power nodes are input into a pre-constructed hyperbolic causal graph reinforcement learning engine to obtain a causal representation and hierarchical geometric representation of the computing power network, including: S31. Construct a state feature matrix based on the feature parameters of available computing power nodes, and input the state feature matrix into the hyperbolic graph embedding module to obtain the feature vector of available computing power nodes in hyperbolic space; S32. Based on the feature vectors, map the topology of the available computing power nodes to the Poincaré sphere model in hyperbolic space to obtain the hierarchical geometric representation. S33. Input the hierarchical geometric representation into the spatiotemporal causal attention inference module, and use the spatiotemporal causal attention inference module to construct a structural causal graph. The structural causal graph is used to describe the causal relationship of computing power flow. S34. Calculate the counterfactual spatiotemporal attention weights based on the structural causal graph, and generate a computational causal representation based on the counterfactual spatiotemporal attention weights and the structural causal graph.

[0013] Preferably, the cross-domain computing power scheduling method based on blockchain and digital twins further includes: The distance between target computing power nodes is calculated using a distance metric formula, and the topology of available computing power nodes is mapped to a Poincaré sphere model based on the node distance. The distance metric formula is expressed as follows: ; in, Indicates the distance between nodes. and These represent the feature vector coordinates of different target computing power nodes in hyperbolic space; The counterfactual spatiotemporal attention weights are calculated using the following formula: ; in, The query vector represents the task requiring computing power. This represents a predefined budget quantifier. The key vector representing the available computing power nodes. A value vector representing the available computing power nodes. and Represents the learnable projection matrix. Z represents the scaling factor, and Z represents the environmental confounding variable. This represents the causal expectation of the state characteristics of the computing node after applying a counterfactual intervention do(Z) to the environmental confusion variable Z.

[0014] Preferably, based on the causal representation and hierarchical geometric representation of the computing power network, a computing power scheduling strategy is generated through a hyperbolic causal graph reinforcement learning engine, including: S41. Integrate the causal representation of the computing power network and the hierarchical geometric representation to generate the state space of the Markov decision process. S42. Define the scheduling action space corresponding to the computing power requirement task. The scheduling action space includes the task splitting ratio, target computing power node allocation, and network routing path selection. S43. Construct a multidimensional dynamic reward function. The multidimensional dynamic reward function is used to update the policy gradient of the policy network in the hyperbolic causal graph reinforcement learning engine to obtain the scheduling policy network. S44. Generate computing power scheduling strategies for computing power demand tasks based on scheduling strategies.

[0015] Preferably, the multidimensional dynamic reward function is expressed using the following formula: ; in, This represents a multidimensional dynamic reward function. Representing the state space, Represents the scheduling action space, Indicates end-to-end prediction delay. Indicates the actual cost of pricing in the region. Indicates the user's budget. This indicates penalties for violating the service level agreement. , and This represents the preset preference coefficient.

[0016] This invention discloses a cross-domain computing power scheduling system based on blockchain and digital twins, used to execute the aforementioned cross-domain computing power scheduling method based on blockchain and digital twins. The cross-domain computing power scheduling system based on blockchain and digital twins includes: The blockchain trusted identification module is configured to: respond to the computing power demand task initiated by the user, obtain the available computing power nodes of computing power centers in various regions, and identify and store the computing power demand task and available computing power nodes through the blockchain network; The digital twin quantization module is configured to: use digital twin technology to perform dynamic twin mapping of the computing power center in each region, and generate characteristic parameters of available computing power nodes, including computing power performance parameters, network latency parameters, and dynamic pricing costs; The hyperbolic causal graph inference module is configured to input the feature parameters of available computing power nodes into a pre-built hyperbolic causal graph reinforcement learning engine to obtain the causal representation of the computing power network and the hierarchical geometric representation. The hyperbolic causal graph reinforcement learning engine includes a spatiotemporal causal attention inference module and a hyperbolic graph embedding module. The strategy scheduling recommendation module is configured to generate computing power scheduling strategies based on the causal representation and hierarchical geometric representation of the computing power network, and to realize cross-domain computing power scheduling according to the computing power scheduling strategies.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires cross-domain computing power demand tasks submitted by users and uses blockchain smart contracts to identify and privately store the computing power demand and all available computing power nodes on the network. It utilizes digital twin technology to construct a dynamic twin mapping of physical computing power centers, measuring the computing power performance and energy consumption costs of each region on the network in real time and dynamically pricing them. The dynamic twin mapping data is input into a hyperbolic causal graph reinforcement learning engine for training and inference to obtain a computing power topological causal representation stripped of confounding variables. Based on this computing power topological causal representation, a globally optimal computing power scheduling strategy is output to optimize scheduling cost-effectiveness. This application solves the problems of opaque cost measurement, dynamic network topology distortion, and suboptimal scheduling caused by multivariate confounding in cross-domain computing power scheduling, achieving efficient, economical, and reliable flow of computing power resources. Attached Figure Description

[0018] Figure 1A flowchart illustrating the cross-domain computing power scheduling method based on blockchain and digital twins provided by this invention; Figure 2 This is a schematic diagram of the cross-domain computing power scheduling system based on blockchain and digital twins provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The present invention will now be described in further detail with reference to the accompanying drawings.

[0021] like Figure 1 As shown, this embodiment of the invention provides a cross-domain computing power scheduling method based on blockchain and digital twins, including the following steps.

[0022] S1. In response to the computing power demand task initiated by the user, obtain the available computing power nodes of computing power centers in various regions, and identify and store the computing power demand task and available computing power nodes through the blockchain network.

[0023] In this embodiment of the invention, a user initiates a computing power request task, which includes key information such as specific computing power requirements, memory requirements, and latency tolerance. Upon receiving the computing power request task, the system first scans and queries computing power centers across the entire network to obtain information on available computing power nodes that are currently idle or available for allocation. This information covers the node's hardware configuration, geographical location, and current load. To ensure the authenticity, completeness, and immutability of the computing power request task and the available computing power node information, the system uses a blockchain network to identify and store the computing power requirement task and available computing power nodes.

[0024] For example, a hybrid blockchain structure based on a directed acyclic graph (DAG) combined with the main chain is adopted to accommodate high-throughput microtransactions. When an end user (such as an AI company needing to pre-train a language model with billions of parameters) initiates a computing power demand task, the task is encapsulated as a timestamped digital envelope. In this embodiment of the invention, computing power is no longer an abstract concept, but is identified in fine granular terms.

[0025] Specifically, step S1 includes the following sub-steps: S11. Extract features from the computing power requirement task to obtain the first feature vector, and generate the first hash identifier based on the first feature vector.

[0026] S12. Extract features from available computing power nodes to obtain a second feature vector, and generate a second hash identifier based on the second feature vector.

[0027] S13. Use an asymmetric encryption algorithm to digitally sign the first hash identifier and the second hash identifier, and write the signature result into the blockchain distributed ledger for storage.

[0028] It should be noted that, in this embodiment of the invention, feature extraction is performed on the computing power requirement task to obtain a first feature vector containing computational demand features, memory demand features, and latency tolerance features, and a unique first hash identifier is generated based on the first feature vector. The computational demand feature characterizes the total computing resource requirement of the task, typically measured by metrics such as floating-point operations (FLOPS) or task processing time. The memory demand feature reflects the amount of memory space required during task execution, including running memory (RAM) and temporary storage requirements. The latency tolerance feature defines the maximum acceptable latency threshold from task submission to completion, and is one of the key factors affecting scheduling priority.

[0029] Furthermore, features are extracted from available computing power nodes to obtain a second feature vector containing heterogeneous hardware specification features, current concurrent task count features, and region identifier features. The heterogeneous hardware specification features represent hardware parameters such as the node's CPU model, number of cores, clock speed, GPU configuration, memory capacity and type, storage medium, and read / write speed. The current concurrent task count feature indicates the number of tasks the node is currently processing, directly affecting its remaining computing power. The region identifier feature clarifies the geographical region where the node is located.

[0030] Furthermore, an asymmetric encryption algorithm is used to digitally sign the first and second hash identifiers. During the signing process, the hash value is encrypted using a private key to generate a unique digital signature. Finally, the signature result, along with the corresponding hash identifier, feature vector digest, and other information, is written into the blockchain distributed ledger for evidence storage.

[0031] It should be noted that the decentralized and immutable nature of blockchain ensures that the information of computing power demand tasks and available computing power nodes cannot be unilaterally tampered with, thus providing a reliable data foundation for subsequent computing power scheduling.

[0032] In some embodiments, a computing power pricing smart contract is deployed on the blockchain distributed ledger, which integrates zero-knowledge verification parameters. These parameters are used to verify the authenticity and compliance of the computing power prices of available computing power nodes. When an available computing power node reports its computing power price, the zero-knowledge verification parameters can verify the authenticity and compliance of that price without disclosing sensitive information such as the node's specific cost structure. For example, they can verify whether the price complies with the region's basic electricity price regulations and whether reasonable hardware depreciation has been considered, further ensuring the fairness and transparency of computing power transactions.

[0033] For example, the computing power pricing smart contract receives and records the computing power prices reported by computing power centers in various regions. When submitting their dynamic pricing costs, each computing power center does not need to disclose its underlying sensitive raw data (such as commercial electricity fee agreements with the power grid, hardware procurement base prices, and real-time PUE details of the data center). It only needs to submit zero-knowledge verification parameters for the proof process to the computing power pricing smart contract. Based on the zero-knowledge verification parameters, the computing power pricing smart contract uses a built-in zero-knowledge proof verification algorithm to automatically verify these parameters, verifying the authenticity and compliance of the computing power prices of available computing power nodes.

[0034] The verified prices are recorded on an immutable blockchain distributed ledger, providing a consistent and reliable computing power price benchmark for the entire scheduling system. This enables the subsequent hyperbolic causal graph reinforcement learning engine to make globally optimal scheduling decisions based on real and comparable price data.

[0035] S2. Utilize digital twin technology to perform dynamic twin mapping of the computing power center in each region and generate characteristic parameters of available computing power nodes.

[0036] In this embodiment of the invention, S2 includes the following steps: S21. Deploy a digital twin agent in each computing center. The digital twin agent collects the operating status parameters of the physical server corresponding to the computing center at a preset frequency. The operating status parameters include environmental sensor data, network monitoring data, and computing node operating data.

[0037] S22. Based on the operating status parameters, construct a dynamic virtual mapping of available computing power nodes from physical state to virtual state in the digital twin engine.

[0038] S23. Collect the computing power performance parameters and network latency parameters of available computing power nodes based on the dynamic virtual mapping, and calculate the dynamic pricing cost of available computing power nodes based on the running status parameters.

[0039] Specifically, to achieve accurate and dynamic twin mapping of computing centers in each region, a full-element digital mirror of the physical computing centers is first constructed. A digital twin agent is deployed in each computing center, and this agent collects the operational status parameters of the corresponding physical servers at a preset frequency. These operational status parameters include environmental sensor data, network monitoring data, and compute node operational data. Environmental sensor data includes rack air intake and exhaust temperatures and cooling system power consumption. Network monitoring data includes switch queue depth and packet loss rate. Compute node operational data includes compute unit load rate, temperature, and utilization rate.

[0040] For example, a digital twin agent of an edge computing microservice architecture is deployed at the bottom layer of each data center. The digital twin agent collects environmental sensor data, network switch SNMP data, and compute node monitoring data at a frequency of seconds. This data covers the data center spatial layout, server rack arrangement, network topology connections, and infrastructure such as air conditioning and UPS.

[0041] The collected raw data undergoes preprocessing via an edge computing gateway, including outlier removal, data normalization, and timestamp alignment, before being transmitted to the digital twin platform. Based on operational status parameters, a dynamic virtual mapping from the physical to the virtual state of available computing nodes is constructed within the digital twin engine. The digital twin platform utilizes a multiphysics-coupled simulation engine to simulate the operational state of the computing nodes in real time. For example, it predicts performance degradation trends based on CPU temperature, simulates data transmission bottlenecks in conjunction with network topology, and fuses and calibrates these simulation results with real-time collected data. Ultimately, it generates high-fidelity characteristic parameters of available computing nodes, providing accurate quantitative input for subsequent hyperbolic causal graph inference.

[0042] In this embodiment of the invention, a multidimensional cost function is used to calculate the dynamic pricing cost. The multidimensional cost function is expressed by the following formula: ; in, Indicates available computing power nodes In time slice The dynamic pricing cost within the region, due to the different electricity prices corresponding to different geographical locations (for example, the electricity cost is different in the east and west), will serve as the core basis for recommending cost-effectiveness. Indicates the device's base power consumption. This indicates the power utilization rate (in western regions, when nighttime temperatures are low, natural wind cooling or water cooling is used, causing a significant drop in PUE, and this value becomes smaller). This indicates the regional tiered electricity pricing, which varies due to different regional regulations (e.g., as low as 0.3 yuan / kWh in the west and as high as 0.8 yuan / kWh in the east). This represents the computing power generated per unit of time. This represents the dynamic depreciation rate of hardware (which takes into account the sunk cost of hardware depreciation caused by high-load operation). Indicates available uplink bandwidth. This indicates a preset bandwidth threshold. When the available bandwidth is below the threshold, the cost index increases. All of these represent weighting coefficients.

[0043] Through this twin computing engine, the system transforms cold physical servers into multi-dimensional, dynamically fluctuating computing power commodities, completely breaking the outdated model of traditional computing power leasing with a fixed monthly price.

[0044] S3. Input the feature parameters of the available computing power nodes into the pre-built hyperbolic causal graph reinforcement learning engine to obtain the causal representation and hierarchical geometric representation of the computing power network.

[0045] In this embodiment of the invention, the hyperbolic causal graph reinforcement learning engine includes a spatiotemporal causal attention inference module and a hyperbolic graph embedding module. The hyperbolic causal graph reinforcement learning engine is an integrated high-level artificial intelligence decision-making framework. It receives a state feature matrix provided by a digital twin system, containing dynamic pricing, performance, and latency information. Based on a precisely represented computing power network and clarified causal logic, it generates a computing power scheduling strategy that maximizes long-term returns (cost-effectiveness) through game theory and trial and error.

[0046] The hyperbolic graph embedding module is responsible for accurately perceiving and understanding the structure of complex networks, mapping computational networks with complex hierarchical relationships in the physical world (such as the tree structure of national backbone network - provincial hub - municipal edge nodes) into a mathematical space. This invention abandons the traditional Euclidean space and uses a Poincaré sphere model of hyperbolic space for embedding. Embedding a tree graph in a flat Euclidean space will lead to congestion of distant nodes and distance distortion, affecting route prediction. However, the boundary space capacity of hyperbolic space (Poincaré sphere) grows exponentially, which is consistent with the characteristic that the number of nodes in a hierarchical tree network grows exponentially with the level.

[0047] The spatiotemporal causal attention inference module receives hierarchical geometric representations from the hyperbolic graph embedding module and uses causal inference techniques to extract the true causal structure affecting computing power flow performance from mixed observation data. Traditional models suffer from "confounding." For example, when an increase in latency is detected at a node, it cannot distinguish whether this is due to "backbone fiber optic cable congestion" (a network problem that should be avoided) or "a surge in computing power tasks at this node" (a computing power problem that can be scheduled to other nodes in the same region). Decisions based on spurious correlations can lead to incorrect scheduling. This application introduces a structural causal model and counterfactual reasoning theory. By defining an interference quantifier do(·), independent scheduling actions are simulated, and counterfactual questions such as "what would happen to latency if network congestion had not occurred" are calculated.

[0048] Specifically, step S3 includes the following sub-steps: S31. Construct a state feature matrix based on the feature parameters of available computing power nodes, and input the state feature matrix into the hyperbolic graph embedding module to obtain the feature vector of available computing power nodes in hyperbolic space.

[0049] S32. Based on the feature vectors, map the topology of the available computing power nodes to the Poincaré sphere model in hyperbolic space to obtain the hierarchical geometric representation.

[0050] S33. Input the hierarchical geometric representation into the spatiotemporal causal attention inference module, and use the spatiotemporal causal attention inference module to construct a structural causal graph. The structural causal graph is used to describe the causal relationship of computing power flow.

[0051] S34. Calculate the counterfactual spatiotemporal attention weights based on the structural causal graph, and generate a computational causal representation based on the counterfactual spatiotemporal attention weights and the structural causal graph.

[0052] Specifically, a state feature matrix is ​​constructed based on computing power performance parameters, network latency parameters, and dynamic pricing costs. During the construction process, various feature parameters of available computing power nodes need to be standardized to eliminate the impact of dimensional differences on subsequent embedding computation. The constructed state feature matrix has a dimension of N×M, where N is the number of available computing power nodes and M is the total dimension of the feature parameters. The hyperbolic graph embedding module employs a gradient descent-based optimization algorithm. By minimizing the distance loss function between nodes, each node in the state feature matrix is ​​mapped to a feature vector in the hyperbolic space. This distance loss function comprehensively considers the physical connectivity of nodes (such as adjacency relationships in network topology) and feature similarity (such as similarity in hardware configuration), ensuring that nodes that are close in distance in the hyperbolic space are not only physically closely connected but also have more similar feature attributes.

[0053] Furthermore, when mapping the topology of available computing power nodes to the Poincaré sphere model, the approximate radial position of each node within the Poincaré sphere is first determined based on its hierarchical relationship (e.g., national backbone network nodes, provincial hub nodes, municipal edge nodes). Typically, higher-level nodes (e.g., national backbone network nodes) are closer to the center of the Poincaré sphere, while lower-level nodes (e.g., municipal edge nodes) are distributed along the sphere's edges, taking advantage of the large capacity of hyperbolic space boundaries. Then, the node distances between target computing power nodes are calculated using a distance metric formula, and the topology of available computing power nodes is mapped to the Poincaré sphere model based on these distances.

[0054] The distance metric formula is expressed as follows: ; in, Indicates the distance between nodes. and These represent the feature vector coordinates of different target computing power nodes in hyperbolic space. Describes the Euclidean norm. This represents the inverse hyperbolic cosine function.

[0055] In the aforementioned implementation process, the national computing power network is a typical tree-like hierarchical structure. Traditional algorithms embedding such data into a flat Euclidean space can cause congestion and distorted representations. The boundary space capacity of the Poincaré sphere grows exponentially, perfectly matching the exponential growth of the tree-like nodes. This allows for accurate reconstruction of the distance between remote nodes (such as western computing power) and central nodes (such as core routing), improving the model's accuracy in calculating long-distance physical features across domains.

[0056] Furthermore, the hierarchical geometric representation is input into the spatiotemporal causal attention inference module. The module uses the position and distance information of nodes in the hierarchical geometric representation to initially screen candidate node pairs that may have causal relationships. For example, nodes that are close to each other and are in the same or adjacent levels may have a closer causal relationship in terms of computing power flow.

[0057] Next, the spatiotemporal causal attention inference module introduces the time dimension, analyzing the computing node operation data (such as load changes, latency fluctuations, and data transmission volume) collected at different time slices to construct an observation dataset containing time-series information. Based on this observation dataset, causal discovery algorithms (such as PC algorithms or score-based search algorithms) are used to learn the potential causal structure from the data, generating a preliminary structural causal graph. This structural causal graph uses directed edges to represent causal relationships in the computing power flow process; for example, "increased load at node A" may be the cause of "increased latency at node B," and the edge weights represent the strength of the causal influence.

[0058] Finally, the spatiotemporal causal attention inference module defines counterfactual scenarios for each causal path in the structural causal graph. For example, for the causal relationship "node C's load increases due to task migration from node D," the counterfactual scenario is set as "what would node C's load be if node D did not migrate its task to node C?" The module simulates the operational parameters of node C under this counterfactual scenario using a digital twin engine to obtain counterfactual data. Then, the counterfactual data is compared with actual observed data, and the differences between the two in the spatiotemporal dimension are calculated, serving as a quantitative indicator of the counterfactual impact. Finally, combining the spatiotemporal location information of nodes in the hierarchical geometric representation, an attention mechanism is designed to integrate the quantitative indicator of counterfactual impact into the calculation of attention weights.

[0059] Specifically, for node pairs with strong causal relationships and significant counterfactual effects in the structural causal graph, higher spatiotemporal attention weights are assigned, allowing them to receive more attention in subsequent scheduling decisions. Finally, the counterfactual spatiotemporal attention weights are fused with the structural causal graph, and the causal relationship information of each node is aggregated through weighted aggregation to generate the final computational power causal representation. This representation clearly reveals the deep causal connections between different computational power nodes in the spatiotemporal dimension, providing strong support for the generation of subsequent optimal scheduling strategies.

[0060] In this embodiment of the invention, the counterfactual spatiotemporal attention weight is calculated using the following formula: ; in, The query vector represents the task requiring computing power. This represents a predefined budget quantifier. The key vector representing the available computing power nodes. A value vector representing the available computing power nodes. and Represents the learnable projection matrix. Z represents the scaling factor, and Z represents the environmental confounding variable. This represents the causal expectation of the state characteristics of the computing node after applying a counterfactual intervention do(Z) to the environmental confusion variable Z.

[0061] Specifically, by performing a linear transformation on the feature vectors of computing nodes in hyperbolic space, key vectors and value vectors of the computing nodes can be generated. The projection matrix is ​​a pre-defined learnable parameter matrix for the model. It is randomly initialized in the initial stage of the hyperbolic causal graph reinforcement learning engine and continuously optimized and learned during subsequent model training by using a multi-dimensional dynamic reward function for gradient updates through backpropagation. The scaling factor is a pre-defined numerical constant hyperparameter. It is usually taken as the dimension of the key vector to prevent the gradient of the Softmax function from vanishing due to an excessively large dot product result. The environmental confounding variable is a variable that interferes with the causal relationship of the real computing power scheduling, identified by the spatiotemporal causal attention inference module when constructing the structural causal graph based on the node temporal observation data through causal discovery algorithms (such as the PC algorithm). (For example, separating network backbone congestion from excessive node computing power load, identifying congestion as an environmental confounding variable.) The causal expectation value is obtained by calling the digital twin engine, applying interference quantifiers (i.e. simulating counterfactual scenarios) to the selected environmental confounding variables, simulating the operating state parameters of the computing nodes under the counterfactual scenario in the digital twin engine, and then calculating the mathematical expectation value of the feature vector.

[0062] It should be noted that, in the above implementation process, causal inference is introduced. Theoretically, in complex network scheduling, a spike in latency in a certain area could be due to a surge in computing power tasks or backbone fiber maintenance. The former can continue to dispatch tasks to other idle nodes in the same area, while the latter requires cross-domain avoidance. By eliminating unobserved confusion factors through the backdoor adjustment mechanism in the formula, blind scheduling caused by spurious correlations is avoided, significantly improving the robustness of cross-domain scheduling.

[0063] S4. Based on the causal representation and hierarchical geometric representation of the computing power network, a computing power scheduling strategy is generated through a hyperbolic causal graph reinforcement learning engine, and cross-domain computing power scheduling is realized according to the computing power scheduling strategy.

[0064] In this embodiment of the invention, step S4 includes the following sub-steps: S41. Integrate the causal representation of the computing power network and the hierarchical geometric representation to generate the state space of the Markov decision process.

[0065] S42. Define the scheduling action space corresponding to the computing power requirement task.

[0066] S43. Construct a multidimensional dynamic reward function. The multidimensional dynamic reward function is used to update the policy gradient of the policy network in the hyperbolic causal graph reinforcement learning engine to obtain the scheduling policy network.

[0067] S44. Generate computing power scheduling strategies for computing power demand tasks based on scheduling strategies.

[0068] Specifically, when fusing the causal representation and hierarchical geometric representation of the computing power network, the causal strength information in the computing power causal representation is first concatenated with the node spatial location and distance information in the hierarchical geometric representation. For example, the causal influence weight vector of each node is superimposed with the feature vector in the hyperbolic space to form a joint feature vector that integrates structural causal information and spatial topological information. Subsequently, these joint feature vectors are processed using a self-attention mechanism, enabling the model to adaptively focus on node features and causal relationships that are more important to the current scheduling task, ultimately constructing the state space of the Markov decision process. This state space not only contains the dynamic performance parameters of each available computing power node (such as real-time load, network latency, and dynamic pricing cost), but also implies the hierarchical structural relationships and deep causal connections between nodes, providing a comprehensive environmental state perception for the reinforcement learning agent.

[0069] The definition of the scheduling action space must cover all types of operations that may be taken during cross-domain computing power scheduling. Specifically, scheduling actions include, but are not limited to: migrating a task with specific computing power requirements to an available computing power node, decomposing a task into multiple sub-tasks and allocating them to multiple computing power nodes in different regions, adjusting the allocation ratio of computing power resources (such as the number of CPU cores and memory usage) of a task among different nodes, and selecting the optimal data transmission path based on real-time network conditions. Each scheduling action must clearly define specific parameters such as task identifier, target node identifier, and resource allocation amount to ensure the executability and accuracy of the action. For example, a specific scheduling action can be represented as "migrating 30% of the computing power requirement of task T1 from eastern node A to western node B, allocating 4 CPU cores and 16GB of memory".

[0070] The construction of a multidimensional dynamic reward function aims to comprehensively evaluate the merits of scheduling policies and guide reinforcement learning agents to optimize policies in the direction of maximizing long-term rewards. The multidimensional dynamic reward function is expressed by the following formula: ; in, This represents a multidimensional dynamic reward function. Representing the state space, Represents the scheduling action space, Indicates end-to-end prediction delay. Indicates the actual cost of pricing in the region. Indicates the user's budget. This indicates penalties for violating the service level agreement. , and This represents the preset preference coefficient.

[0071] In this embodiment of the invention, the end-to-end predicted latency is obtained as follows: network latency parameters collected by the digital twin agent are combined with the physical topology distance between nodes in the current state space and the network routing path selection specified in the action space to calculate the total estimated transmission and execution latency of the task. The actual execution pricing cost of the region comes from the dynamic pricing cost of available computing power nodes calculated using a multi-dimensional cost function, and multiplied by the task splitting ratio and resource allocation in the action space to finally obtain the total pricing amount. The user budget is a value input by the user. The penalty for violating the service level agreement is a preset numerical constant or a value generated based on rules. When the target node scheduled in the action space crashes, or the predicted latency exceeds the latency tolerance feature threshold, the system triggers and assigns a preset maximum negative penalty constant to the parameter to guide the reinforcement learning agent to avoid such scheduling. The preset preference coefficient is a preset weight value. It is preset by the system administrator or adaptively preset according to the type of computing power demand task initiated by the user, and is used to adjust the proportion of latency, cost and reliability in the final reward.

[0072] In the above implementation process, the constructed multidimensional dynamic reward function directly corresponds to the invention's objective of "optimizing the cost-effectiveness of computing power for recommendation scheduling." During training, the reinforcement learning agent continuously balances latency (limited by physical distance), budget, and fault penalties. Combined with the accurate forward prediction capability brought by hyperbolic causal embedding, after massive iterations, the policy network can automatically match the most suitable computing power center in the national map for different types of tasks (such as latency-sensitive autonomous driving inference tasks or cost-sensitive large model offline rendering tasks).

[0073] After obtaining the scheduling policy network, for tasks requiring computing power, their requirement parameters (such as computational load, latency requirements, and data volume) are first converted into a query vector Q and input into the scheduling policy network. The policy network utilizes learned state-space representations and action-space knowledge, combined with the real-time state characteristics of currently available computing power nodes, to output the probability distribution of each possible scheduling action through forward computation. The action with the highest probability is selected as the current optimal computing power scheduling policy, and it is transformed into a specific scheduling instruction, which is then issued to the corresponding computing power nodes for execution. Simultaneously, during the execution of the scheduling policy, actual running data is continuously collected as feedback to update the multi-dimensional dynamic reward function and policy network parameters, achieving closed-loop optimization through reinforcement learning and continuously improving the efficiency and cost-effectiveness of cross-domain computing power scheduling.

[0074] like Figure 2 As shown, the present invention also provides a cross-domain computing power scheduling system based on blockchain and digital twins, used to execute a cross-domain computing power scheduling method based on blockchain and digital twins. The cross-domain computing power scheduling system based on blockchain and digital twins includes: a blockchain trusted identification module 201, a digital twin quantization module 202, a hyperbolic causal graph inference module 203, and a strategy scheduling recommendation module 204.

[0075] Specifically, the blockchain trusted identification module 201 is configured to: respond to the computing power demand task initiated by the user, obtain the available computing power nodes of the computing power centers in each region, and identify and store the computing power demand task and available computing power nodes through the blockchain network; The digital twin quantization module 202 is configured to: use digital twin technology to perform dynamic twin mapping of the computing power center in each region, and generate characteristic parameters of available computing power nodes, including computing power performance parameters, network latency parameters and dynamic pricing costs; The hyperbolic causal graph inference module 203 is configured to input the feature parameters of available computing power nodes into a pre-built hyperbolic causal graph reinforcement learning engine to obtain the causal representation of the computing power network and the hierarchical geometric representation. The hyperbolic causal graph reinforcement learning engine includes a spatiotemporal causal attention inference module and a hyperbolic graph embedding module. The strategy scheduling recommendation module 204 is configured to generate a computing power scheduling strategy based on the causal representation and hierarchical geometric representation of the computing power network, and to realize cross-domain computing power scheduling according to the computing power scheduling strategy.

[0076] As can be seen from the above technical solution, this application provides a cross-domain computing power scheduling method and system based on blockchain and digital twins. The cross-domain computing power scheduling method includes: responding to a user-initiated computing power demand task, obtaining available computing power nodes in computing power centers of various regions, and identifying and storing the computing power demand task and available computing power nodes through a blockchain network; using digital twin technology to perform dynamic twin mapping of computing power centers in each region, and generating feature parameters of available computing power nodes; using a hyperbolic causal graph reinforcement learning engine to obtain causal representation and hierarchical geometric representation of the computing power network based on the feature parameters, and generating a computing power scheduling strategy based on the causal representation and hierarchical geometric representation of the computing power network, thereby realizing cross-domain computing power scheduling. This invention solves the scheduling suboptimal problems caused by opaque cost measurement, dynamic network topology distortion, and multivariate confusion in cross-domain computing power scheduling, and realizes efficient, economical, and reliable flow of computing power resources.

[0077] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cross-domain computing power scheduling method based on blockchain and digital twins, characterized in that, include: S1. In response to the computing power demand task initiated by the user, obtain the available computing power nodes of computing power centers in each region, and identify and store the computing power demand task and the available computing power nodes through the blockchain network. S2. Utilize digital twin technology to perform dynamic twin mapping of the computing power center in each region, and generate characteristic parameters of the available computing power nodes. The characteristic parameters include computing power performance parameters, network latency parameters, and dynamic pricing costs. S3. Input the feature parameters of the available computing power nodes into a pre-constructed hyperbolic causal graph reinforcement learning engine to obtain a causal representation and hierarchical geometric representation of the computing power network. The hyperbolic causal graph reinforcement learning engine includes a spatiotemporal causal attention inference module and a hyperbolic graph embedding module. Step S3 includes: S31. Construct a state feature matrix based on the feature parameters of the available computing power nodes, and input the state feature matrix into the hyperbolic graph embedding module to obtain the feature vector of the available computing power nodes in hyperbolic space; S32. Map the topology of the available computing power nodes to a Poincaré sphere model in hyperbolic space according to the feature vector to obtain the hierarchical geometric representation; S33. Input the hierarchical geometric representation into the spatiotemporal causal attention inference module, and use the spatiotemporal causal attention inference module to construct a structural causal graph, which is used to describe the causal relationship of computing power flow; S34. Calculate counterfactual spatiotemporal attention weights based on the structural causal graph, and generate the causal representation of the computing power network according to the counterfactual spatiotemporal attention weights and the structural causal graph. S4. Based on the causal representation of the computing power network and the hierarchical geometric representation, a computing power scheduling strategy is generated through the hyperbolic causal graph reinforcement learning engine, and cross-domain computing power scheduling is implemented according to the computing power scheduling strategy, including: S41. fusing the causal representation of the computing power network and the hierarchical geometric representation to generate the state space of a Markov decision process; S42. defining a scheduling action space corresponding to the computing power demand task, the scheduling action space including task splitting ratio, target computing power node allocation, and network routing path selection; S43. constructing a multidimensional dynamic reward function, the multidimensional dynamic reward function being used to update the policy gradient of the policy network in the hyperbolic causal graph reinforcement learning engine to obtain a scheduling policy network; S44. generating a computing power scheduling strategy corresponding to the computing power demand task based on the scheduling policy network.

2. The cross-domain computing power scheduling method according to claim 1, characterized in that, The step of identifying and storing the computing power requirement task and the available computing power nodes through a blockchain network includes: S11. Extract features from the computing power requirement task to obtain a first feature vector, and generate a first hash identifier based on the first feature vector. The first feature vector includes computing power requirement features, memory requirement features, and latency tolerance features. S12. Extract features from the available computing power nodes to obtain a second feature vector, and generate a second hash identifier based on the second feature vector. The second feature vector includes heterogeneous hardware specification features, current concurrent task count features, and region identifier features. S13. Use an asymmetric encryption algorithm to digitally sign the first hash identifier and the second hash identifier, and write the signature result into the blockchain distributed ledger for storage.

3. The cross-domain computing power scheduling method according to claim 2, characterized in that, The blockchain distributed ledger is deployed with a computing power pricing smart contract, which integrates zero-knowledge verification parameters. These parameters are used to verify the authenticity and compliance of the computing power price of the available computing power nodes.

4. The cross-domain computing power scheduling method according to claim 1, characterized in that, The process of using digital twin technology to dynamically map the computing power centers of each region and generate characteristic parameters of the available computing power nodes includes: S21. Deploy a digital twin agent in each of the computing power centers. The digital twin agent collects the operating status parameters of the physical server corresponding to the computing power center at a preset frequency. The operating status parameters include environmental sensor data, network monitoring data, and computing node operating data. S22. Based on the operating status parameters, construct a dynamic virtual mapping of the available computing power nodes from physical state to virtual state in the digital twin engine; S23. Collect the computing power performance parameters and network latency parameters of the available computing power nodes according to the dynamic virtual mapping, and calculate the dynamic pricing cost of the available computing power nodes according to the running status parameters.

5. The cross-domain computing power scheduling method according to claim 4, characterized in that, The environmental sensing data includes the temperature of the cabinet's air inlet and outlet and the power consumption of the cooling system. The network monitoring data includes switch queue depth and packet loss rate; The computing node operation data includes computing unit load rate, temperature, and utilization rate; The calculation of the dynamic pricing cost of the available computing power node based on the operating status parameters includes: The dynamic pricing cost is calculated using a multidimensional cost function, which is expressed by the following formula: ； in, Indicates available computing power nodes In time slice Dynamic pricing costs within the scope, Indicates the device's base power consumption. Indicates energy efficiency. This indicates tiered electricity pricing by region. This represents the computing power generated per unit of time. This indicates the dynamic depreciation rate of the hardware. Indicates available uplink bandwidth. This indicates the preset bandwidth threshold. All of these represent weighting coefficients.

6. The cross-domain computing power scheduling method according to claim 1, characterized in that, Also includes: The distance between target computing power nodes is calculated using a distance metric formula, and the topology of the available computing power nodes is mapped to the Poincaré sphere model based on the node distance. The distance metric formula is expressed as follows: ； in, Indicates the distance between nodes. and These represent the feature vector coordinates of different target computing power nodes in hyperbolic space; The counterfactual spatiotemporal attention weights are calculated using the following formula: ； in, This represents the query vector for the computing power requirement task. This represents a predefined budget quantifier. The key vector representing the available computing power node. This represents the value vector of the available computing power nodes. and Represents the learnable projection matrix. Z represents the scaling factor, and Z represents the environmental confounding variable. This represents the causal expectation of the state characteristics of the computing node after applying a counterfactual intervention do(Z) to the environmental confusion variable Z.

7. The cross-domain computing power scheduling method according to claim 1, characterized in that, The multidimensional dynamic reward function is expressed by the following formula: ； in, This represents a multidimensional dynamic reward function. Representing the state space, Represents the scheduling action space, Indicates end-to-end prediction delay. Indicates the actual cost of pricing in the region. Indicates the user's budget. This indicates penalties for violating the service level agreement. , and This represents the preset preference coefficient.

8. A cross-domain computing power scheduling system based on blockchain and digital twins, used to execute the cross-domain computing power scheduling method based on blockchain and digital twins as described in any one of claims 1-7, characterized in that, include: The blockchain trusted identification module is configured to: respond to a user-initiated computing power demand task, obtain available computing power nodes in computing power centers in various regions, and identify and store the computing power demand task and the available computing power nodes through the blockchain network; The digital twin quantization module is configured to: use digital twin technology to perform dynamic twin mapping of the computing power center in each region, and generate characteristic parameters of the available computing power nodes, including computing power performance parameters, network latency parameters, and dynamic pricing costs; The hyperbolic causal graph inference module is configured to: input the feature parameters of the available computing power nodes into a pre-built hyperbolic causal graph reinforcement learning engine to obtain the computing power network causal representation and hierarchical geometric representation. The hyperbolic causal graph reinforcement learning engine includes a spatiotemporal causal attention inference module and a hyperbolic graph embedding module. The strategy scheduling recommendation module is configured to: generate a computing power scheduling strategy through the hyperbolic causal graph reinforcement learning engine based on the computing power network causal representation and the hierarchical geometric representation, and implement cross-domain computing power scheduling according to the computing power scheduling strategy.

Images