A Business Resource Optimization Management Method and System Based on Computing Power Demand Forecasting

Abstract

This application provides a business resource optimization management method and system based on computing power demand prediction. The method includes: real-time collection of multi-source heterogeneous data to construct and continuously update a topological graph representing the dynamic relationships between business entities, logical service entities, and physical resource entities; inputting the temporal state of the topological graph into a prediction model that integrates graph structure and continuous-time modeling, which performs multi-dimensional evolution calculations of node states; and, under a hierarchical decision-making framework, generating dynamic resource value signals by a central scheduler to guide multiple business unit agents in distributed collaborative decision-making, iteratively generating the optimal resource allocation scheme; and performing potential risk assessment and reliable automated execution before executing the optimal resource allocation scheme. Thus, this application achieves closed-loop management and adaptive configuration of computing power resources in complex and dynamic business environments, from global perception, accurate prediction, intelligent optimization to risk control.

Description

Technical Field

[0001] This application relates to the field of resource optimization technology, and in particular to a commercial resource optimization management method and system based on computing power demand prediction. Background Technology

[0002] In the operation of modern digital enterprises, there is a highly complex and dynamically evolving nonlinear dependency between business activities such as marketing, e-commerce promotions, and peak financial transactions and the underlying information technology infrastructure. For example, a successful marketing campaign may trigger a surge in traffic on the microservice call chain in a very short period of time, ultimately impacting the underlying physical or cloud computing resources. Current resource management solutions commonly used in the industry typically first use models such as Long Short-Term Memory (LSTM) networks to predict the demand for single business or local resources, and then combine various heuristic search algorithms to generate scheduling strategies.

[0003] However, in the context of increasingly complex microservice architectures and fierce competition for resources across multiple business lines, these traditional methods have gradually revealed several fundamental bottlenecks:

[0004] First, most existing methods treat business units and IT components as independent time-series signals. For example, a surge in load on an upstream API not only affects itself but also cascades down the call chain to downstream databases and cache layers. Traditional point-based forecasting cannot predict the path and intensity of such "butterfly effect" traffic transmission, resulting in locally optimal rather than globally optimal computing power demand forecasts. The scheduling scheme lacks foresight, causing the forecast results to deviate from the actual global demand, and the scheduling strategy also loses its foresight.

[0005] Secondly, traditional optimization algorithms often aim at a single global metric, such as maximizing overall resource utilization or minimizing total cost. In actual operations, the urgency of different business units' demand for computing resources is closely related to the commercial value they create, such as GMV contribution and user retention rate. Traditional scheduling mechanisms cannot effectively handle resource conflicts under such multi-objective and multi-business priorities, and cannot automatically prioritize high-value businesses when resources are scarce, which can easily lead to losses in commercial interests.

[0006] Third, traditional RNN / LSTM models rely heavily on equally spaced training data, making them ill-suited to the sudden, non-uniform, and sparse nature of events in real-world business environments, such as instantaneous hotspots and system failures. Furthermore, most decision-making systems directly output scheduling instructions based solely on prediction results, which carries decision-making risks.

[0007] In summary, existing technologies have significant shortcomings in dealing with complex topology transmission, multi-business value conflicts, and decision-making risk control. Summary of the Invention

[0008] To address the aforementioned problems in the prior art, this application provides a business resource optimization management method and system based on computing power demand prediction, which realizes closed-loop management and adaptive configuration of computing power resources from global perception, accurate prediction, intelligent optimization to risk control in complex and dynamic business environments.

[0009] To achieve the above objectives, the technical solution adopted in this application is as follows:

[0010] Firstly, this application provides a commercial resource optimization management method based on computing power demand forecasting, including:

[0011] Step S1: Collect and integrate multi-source heterogeneous data from business activities, business links and infrastructure in real time, and construct and continuously update a topological map representing the dynamic relationship between business entities, logical service entities and physical resource entities based on the multi-source heterogeneous data.

[0012] Step S2: Input the temporal state of the topology graph into a prediction model that integrates graph structure and continuous time modeling. The prediction model performs multi-dimensional evolution calculation of node states and outputs the quantitative value of computing power demand intensity and the business value density assessment value of each node in the future time period.

[0013] Step S3: Based on the quantified value of computing power demand intensity and the assessed value of business value density, under a hierarchical decision-making framework, the central scheduler generates a dynamic resource value signal to guide multiple business unit agents to use multi-agent reinforcement learning algorithms to make distributed collaborative decisions and iteratively generate the optimal resource allocation scheme.

[0014] Step S4: Before the optimal resource allocation scheme is executed, counterfactual reasoning is performed based on the topology map to quantitatively assess its potential risks. When the potential risks are lower than a preset threshold, the optimal resource allocation scheme is converted into an automatically executable and tamper-proof scheduling instruction and issued for execution.

[0015] The beneficial effects of this application are as follows: By constructing and continuously updating a dynamic topology graph that integrates business, logical, and physical entities, this invention achieves a unified representation and real-time perception of the full-link relationships in complex systems; furthermore, by using a prediction model based on a fused graph structure and continuous-time modeling to perform evolutionary calculations on the temporal state of this graph, it achieves a comprehensive prediction of future computing power requirements and business value; on this basis, by guiding multi-agents to perform distributed collaborative learning and optimization through a hierarchical decision-making framework, a dynamic resource allocation scheme matching business value is generated; finally, before the scheme is executed, counterfactual reasoning is performed based on the same topology graph to quantify risks, and the verified scheme is converted into reliable instructions that can be automatically executed, thereby achieving a closed-loop management and adaptive configuration of commercial computing power resources from global perception, accurate prediction, intelligent optimization to risk control.

[0016] Optionally, step S2 includes:

[0017] Step S21: Input the temporal state of the topological graph into a prediction model that fuses graph structure and continuous time modeling;

[0018] Step S22: The prediction model uses a multi-head graph attention network to process the topology graph, aggregate the neighbor information of each node, and obtain the node state representing the spatial transmission relationship between nodes.

[0019] Step S23: The prediction model takes the node state, which represents the spatial transmission relationship between nodes, as the initial state of the neural ordinary differential equation. By solving the neural ordinary differential equation, the evolution of the node state in the continuous time domain is modeled and predicted, and the quantitative value of computing power demand intensity and the business value density assessment value of each node in the future period are output.

[0020] As described above, by combining graph attention networks with neural differential equations, this scheme simultaneously captures the spatial propagation characteristics of business load in the service topology and the continuous temporal evolution patterns under non-uniform business events, thus achieving more accurate prediction of computing power requirements and business value.

[0021] Optionally, step S23 includes:

[0022] Step S231: The neural ordinary differential equation defines the evolution of the node hidden state as follows:

[0023] ,

[0024] In the formula, h(t) represents the hidden state vector of the node at continuous time t, and f is a vector vector defined by the parameter θ. ODE A defined neural network is used to model the dynamics of state changes;

[0025] Step S232: Using the node state, which characterizes the spatial transmission relationship between nodes, as the initial state, the neural ordinary differential equation is solved by integration using an ordinary differential equation numerical solver to obtain the predicted node state value at future time.

[0026] Step S233: Based on the node state prediction value, the prediction model outputs the quantitative value of computing power demand intensity and the evaluation value of business value density for each node in the future time period.

[0027] Optionally, step S3 includes:

[0028] Step S31: Based on the quantitative value of computing power demand intensity and the evaluation value of business value density, under a hierarchical decision-making framework, the central scheduler dynamically generates resource value signals according to global resource supply, total demand forecast, operating costs and service level agreement constraints.

[0029] Step S32: Each business unit, acting as an intelligent agent, obtains a virtual budget based on its corresponding business value density assessment value. Guided by the resource value signal, it uses a multi-agent reinforcement learning algorithm to make distributed collaborative decisions and determine its respective resource request amount.

[0030] Step S33: Generate the optimal resource allocation scheme by iteratively optimizing the distributed collaborative decision-making.

[0031] Optionally, step S32 includes:

[0032] Step S321: Each business unit, acting as an intelligent agent, obtains a virtual budget based on its corresponding business value density assessment value.

[0033] Step S322: Each business unit agent i, under the constraint of the virtual budget, maximizes its utility function U. i Determine its resource request size r i The utility function U i Defined as:

[0034] ,

[0035] In the formula, Value i (r i ) Obtain the resource request quantity r for business intelligence agent i. i The resulting expected business value function is positively correlated with the business value density assessment value; P t (r i () refers to the unit price of resources published by the central scheduler; Penalty i (QoS) is a penalty for service agent i failing to meet the quality of service under the current resource allocation;

[0036] Step S323: Each business unit agent uses a multi-agent deep deterministic policy gradient algorithm to learn and optimize policies based on the global resource supply and demand status, resource unit price, and the behavioral history of other agents, through interaction with the environment. Each business unit agent has an independent policy network and shares a centralized commentator network.

[0037] As described above, by constructing a multi-agent bidding learning framework based on utility functions and virtual budget constraints, this scheme transforms resource allocation into a market-based value game process, automatically achieving efficient allocation of resources towards high-value businesses.

[0038] Optionally, it also includes:

[0039] Step S5: Monitor the historical accumulation of the multi-source heterogeneous data. When the historical accumulation reaches a preset data maturity threshold, automatically trigger the update of the prediction model and / or the multi-agent reinforcement learning algorithm.

[0040] Optionally, step S1 includes:

[0041] Step S11: Collect and integrate multi-source heterogeneous data from business activities, business links, and infrastructure in real time;

[0042] Step S12: Based on the multi-source heterogeneous data, construct a heterogeneous information network containing business activity nodes, logical service nodes, data instance nodes and physical resource nodes as the topology graph, wherein the nodes are connected through business event triggering edges, service call edges, deployment attribution edges and causal influence edges.

[0043] Step S13: Introduce dynamic attributes to the node features and edge weights in the heterogeneous information network, and update them in real time through a sliding time window mechanism to form a dynamic evolution graph sequence that serves as the temporal state of the topology graph.

[0044] Optionally, step S12 includes:

[0045] Step S121: Learn the causal relationships between nodes from historical time-series data using a causal discovery algorithm, and generate the causal influence edge;

[0046] Step S122: Based on the multi-source heterogeneous data, determine the connection relationships and weights of the business event triggering edge, service call edge, and deployment affiliation edge.

[0047] As described above, by using causal discovery algorithms and sliding time window update mechanisms, a dynamic graph reflecting the causal and real-time state of the real system is constructed and continuously maintained, providing a reliable data foundation for accurate prediction and credible inference.

[0048] Optionally, step S4 includes:

[0049] Step S41: Based on the topological map, define a structural causal model, and in the digital twin environment, intervene in the optimal resource allocation scheme to perform counterfactual reasoning and quantify its potential risks.

[0050] Step S42: When the potential risk is lower than a preset threshold, the optimal resource allocation scheme is encoded into a smart contract and recorded in a private consortium blockchain for evidence storage.

[0051] Step S43: Execute the resource scheduling operation defined in the smart contract by using the automated orchestration engine that listens to the events of the private consortium blockchain.

[0052] As described above, by assessing decision-making risks through counterfactual reasoning and combining blockchain evidence storage with automated execution, the scheduling strategy achieves pre-emptive risk quantification, reliable process auditing, and a reliable execution closed loop.

[0053] Secondly, this application provides a business resource optimization management system based on computing power demand prediction, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the business resource optimization management method based on computing power demand prediction of the first aspect.

[0054] The technical effects of the business resource optimization management system based on computing power demand prediction provided in the second aspect are described in the relevant description of the business resource optimization management method based on computing power demand prediction provided in the first aspect. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the main process of a business resource optimization management method based on computing power demand forecasting according to an embodiment of this application;

[0056] Figure 2 This is a schematic diagram of the structure of a business resource optimization management system based on computing power demand prediction, according to an embodiment of this application.

[0057] Explanation of reference numerals in the attached figures:

[0058] 1. A business resource optimization management system based on computing power demand forecasting;

[0059] 2. Processor;

[0060] 3. Memory. Detailed Implementation

[0061] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0062] This embodiment is applicable to business, financial, and management applications that require commercial activities. It is particularly suitable for scenarios that require dynamic computing resource management that handles large-scale, high-concurrency, highly dependent, and multi-objective tasks. It can be integrated into e-commerce platforms, financial service environments, or enterprise administrative management platforms to support electronic payment services, financial transaction processing, and management of specific business areas. Existing technologies have significant shortcomings in addressing the complex topology transmission, multi-business value conflicts, and decision-making risk control required by current commercial activities.

[0063] Therefore, this application provides a business resource optimization management method based on computing power demand prediction. It collects and integrates multi-source heterogeneous data from business activities, business links, and infrastructure in real time. Based on this multi-source heterogeneous data, it constructs and continuously updates a topology map representing the dynamic relationships between business entities, logical service entities, and physical resource entities. The temporal state of the topology map is input into a prediction model that integrates graph structure and continuous-time modeling. The prediction model performs multi-dimensional evolution calculations of node states, outputting a quantitative value of computing power demand intensity and a business value density assessment value for each node in future time periods. Based on the quantitative value of computing power demand intensity and the business value density assessment value, under a hierarchical decision-making framework, a central scheduler generates dynamic resource value signals to guide multiple business unit agents to perform distributed collaborative decision-making using multi-agent reinforcement learning algorithms, iteratively generating the optimal resource allocation scheme. Before executing the optimal resource allocation scheme, counterfactual reasoning is performed based on the topology map to quantitatively assess its potential risks. When the potential risks are below a preset threshold, the optimal resource allocation scheme is converted into an automatically executable and tamper-proof scheduling instruction and issued for execution. Therefore, this application achieves closed-loop management and adaptive configuration of computing resources in complex and dynamic business environments, from global perception, accurate prediction, intelligent optimization to risk control.

[0064] The present application will now be described with reference to specific embodiments.

[0065] Reference Figure 1 As can be seen, embodiments of this application provide a business resource optimization management method based on computing power demand prediction, including:

[0066] Step S1: Collect and integrate multi-source heterogeneous data from business activities, business links and infrastructure in real time, and build and continuously update a topology map representing the dynamic relationship between business entities, logical service entities and physical resource entities based on the multi-source heterogeneous data.

[0067] This involves real-time acquisition of event logs for business activities from the enterprise's internal Apache Kafka message queue, such as promotion IDs, payment success / failures, and large transaction notifications. It also includes Span data from business processes like OpenTelemetry and Zipkin, as well as business value data such as GMV contribution rate and profit margin from financial systems like SAP. Finally, it collects infrastructure metrics such as CPU / memory / network I / O for Pods and Nodes from monitoring systems like Prometheus and Grafana, resulting in a multi-source heterogeneous data set integrating business activities, business processes, and infrastructure data.

[0068] In this process, based on the aforementioned collection of multi-source heterogeneous data, the above data types are respectively set as corresponding nodes and edges are established, thereby forming a topological graph that represents the dynamic relationship between commercial entities, logical service entities and physical resource entities.

[0069] Step S2: Input the temporal state of the topology graph into a prediction model that integrates graph structure and continuous time modeling. The prediction model performs multi-dimensional evolution calculations of node states and outputs the quantitative values ​​of computing power demand intensity and business value density assessment values ​​of each node in the future time period.

[0070] In this embodiment, spatial dimension modeling is achieved through graph structure, and temporal dimension modeling is achieved through continuous time modeling. Thus, the prediction model performs spatiotemporal transmission prediction to obtain the quantitative value of computing power demand intensity and the business value density assessment value of each node in the future time period required for subsequent steps.

[0071] Step S3: Based on the quantitative value of computing power demand intensity and the evaluation value of business value density, under a hierarchical decision-making framework, the central scheduler generates a dynamic resource value signal to guide multiple business unit agents to use multi-agent reinforcement learning algorithms to make distributed collaborative decisions and iteratively generate the optimal resource allocation scheme.

[0072] In this system, the central scheduler acts as the upper-level decision-maker, generating signals to guide resource allocation. Each business unit is abstracted as an independent agent to perform distributed learning and decision-making. Through adaptive learning and iterative optimization, guided by value pricing, it autonomously decides how much computing power to purchase or request in order to maximize its contribution to the total value of the enterprise, thereby iteratively generating the optimal resource allocation scheme.

[0073] Step S4: Before executing the optimal resource allocation scheme, counterfactual reasoning is performed based on the topology map to quantitatively assess its potential risks. When the potential risks are lower than a preset threshold, the optimal resource allocation scheme is converted into an automatically executable and tamper-proof scheduling instruction and issued for execution.

[0074] Specifically, before the implementation of the solution, counterfactual reasoning is performed using the constructed topology map to quantitatively assess scheduling risks, and the reliability and trustworthiness of the decision are ensured by scheduling instructions that can be automatically executed and cannot be tampered with. Ultimately, the solution achieves topology awareness and accurate prediction of computing power demand in complex and dynamic business environments, efficient allocation of resources based on value, and risk control and automated closed-loop execution of the decision-making process.

[0075] In one embodiment, step S1 includes:

[0076] Step S11: Collect and integrate multi-source heterogeneous data from business activities, business links and infrastructure in real time.

[0077] Step S12: Based on multi-source heterogeneous data, construct a heterogeneous information network as a topology graph, which includes business activity nodes, logical service nodes, data instance nodes and physical resource nodes. The nodes are connected through business event triggering edges, service call edges, deployment attribution edges and causal influence edges.

[0078] Step S12 includes:

[0079] Step S121: Learn the causal relationships between nodes from historical time series data using the causal discovery algorithm, and generate causal influence edges.

[0080] Step S122: Based on multi-source heterogeneous data, determine the connection relationships and weights of the business event triggering edge, service call edge, and deployment affiliation edge.

[0081] Specifically, in this embodiment, the Heterogeneous Information Network (HIN) constructed based on multi-source heterogeneous data includes business activity nodes, logical service nodes, data instance nodes, and physical resource nodes. Edges represent causal dependencies, invocation relationships, and deployment affiliations. Business activity nodes include, for example, holiday promotions; logical service nodes include, for example, order creation services; data instance nodes include, for example, user databases; and physical resource nodes include, for example, physical machine / virtual machine / container nodes. The edge E that triggers business events... trigger This indicates a connection between a business activity and the first API service it triggers; service call edge E call This represents the dependencies between API services, with their weights determined by historical call frequency and real-time QPS; the deployment belongs to the edge E. hostThis indicates that logical services are deployed on physical resources. Causal influence edges represent potential causal relationships from business events to resource metrics. Causal relationships between nodes are learned from historical time-series data based on causal discovery algorithms such as Granger Causality or PCMCI+.

[0082] Therefore, in this embodiment, a dynamic graph reflecting the causal relationship and real-time state of the real system is constructed and continuously maintained through the causal discovery algorithm and the sliding time window update mechanism, providing a reliable data foundation for accurate prediction and credible reasoning.

[0083] Step S13: Introduce dynamic attributes to the node features and edge weights in the heterogeneous information network, and update them in real time through a sliding time window mechanism to form a dynamic evolution graph sequence as the temporal state of the topological graph.

[0084] Node characteristics include factors such as the scale of business activities, average API response time, and CPU utilization, while edge weights include factors such as call frequency. Therefore, this embodiment introduces dynamic attributes to node characteristics and edge weights in heterogeneous information networks, and uses a sliding time window mechanism to update the graph snapshot at each time step in real time, forming a dynamic evolution graph sequence G. t =(V t E t In the formula, V t Let E be the node feature matrix. t This is the edge weight matrix, which captures the full-topology dynamic causal relationship between business activities and IT resources.

[0085] In one embodiment, step S2 includes:

[0086] Step S21: Input the temporal state of the topological graph into a prediction model that integrates graph structure and continuous time modeling.

[0087] Step S22: The prediction model uses a multi-head graph attention network to process the topology graph, aggregate the neighbor information of each node, and obtain the node state representing the spatial transmission relationship between nodes.

[0088] Among them, the Graph Attention Network (GAT) is used to calculate the influence weights between different neighbor nodes in the graph, so as to capture and quantify the complex spatial transmission characteristics and cascading effects of business traffic or load in the microservice call chain.

[0089] Specifically, GAT is used for each node v i Aggregate its neighbor nodes v j Features h j GAT calculates the attention coefficient α through a self-attention mechanism. ijThis reflects the influence of neighboring nodes on the central node, and the formula is:

[0090] ;

[0091] In the formula, N i Let be the set of neighboring nodes of node i, W be the shared learnable weight matrix, and σ be the activation function. The goal is to capture the spatial propagation characteristics and cascading effects of traffic across nodes in the topology graph, for example, from gateway services to authentication services and then to product services.

[0092] Step S23: The prediction model takes the node state, which represents the spatial transmission relationship between nodes, as the initial state of the neural ordinary differential equation. By solving the neural ordinary differential equation, the evolution of the node state in the continuous time domain is modeled and predicted, and the quantitative value of computing power demand intensity and the business value density assessment value of each node in the future period are output.

[0093] In this approach, Neural Ordinary Differential Equations (Neural ODEs) are introduced to replace traditional recurrent neural networks. Neural ODEs model the state evolution of each node using a continuous time manifold, which can perfectly adapt to the uneven triggering time of business events and solve the prediction problem of traditional time series models for sudden and sparse events.

[0094] Finally, based on the above spatiotemporal modeling, the prediction model outputs the resource demand potential of each node within the future time window, including the quantitative value of the computing power demand intensity and the business value density assessment value of each node within the future time period.

[0095] Specifically, step S23 includes:

[0096] Step S231: The neural ordinary differential equation defines the evolution of the hidden state of a node as follows:

[0097] ,

[0098] In the formula, h(t) represents the hidden state vector of the node at continuous time t, and f is a vector vector defined by the parameter θ. ODE A defined neural network is used to model the dynamics of state changes.

[0099] Step S232: Using the node state, which represents the spatial transmission relationship between nodes, as the initial state, the neural ordinary differential equation is solved by integration using an ordinary differential equation numerical solver to obtain the predicted node state value at future time.

[0100] Step S233: Based on the node status prediction values, the prediction model outputs the quantitative values ​​of computing power demand intensity and the business value density assessment values ​​for each node in the future time period.

[0101] The predictive model outputs the resource demand potential of each node within a future time window, quantifying its abstract demand for different types of computing resources such as CPU cores, memory capacity, storage IOPS, and network bandwidth. It also outputs the business value density of each node, which is positively correlated with business KPIs such as the GMV contribution rate and user retention rate of the corresponding business of that node, and is dynamically adjusted according to real-time operational data.

[0102] Therefore, by combining graph attention networks with neural differential equations, this embodiment captures both the spatial propagation characteristics of business load in the service topology and the continuous-time evolution of non-uniform business events, thus achieving more accurate prediction of computing power requirements and business value.

[0103] In one embodiment, step S3 includes:

[0104] Step S31: Based on the quantitative value of computing power demand intensity and the assessment value of business value density, under a hierarchical decision-making framework, the central scheduler dynamically generates resource value signals according to global resource supply, total demand forecast, operating costs and service level agreement constraints.

[0105] The central scheduler dynamically sets value-based pricing, or shadow price, for different types of computing resources based on the real-time supply of global computing resources, the predicted total demand potential, and the overall operating costs, and uses this pricing as a signal to guide resource allocation.

[0106] Step S32: Each business unit, acting as an intelligent agent, obtains a virtual budget based on its corresponding business value density assessment value. Guided by the resource value signal, it uses a multi-agent reinforcement learning algorithm to make distributed collaborative decisions and determine its respective resource request amount.

[0107] Virtual budgets include metrics such as GMV contribution rate and user activity.

[0108] In this embodiment, step S32 includes:

[0109] Step S321: Each business unit, acting as an intelligent agent, obtains a virtual budget based on its corresponding business value density assessment value.

[0110] Step S322: Each business unit agent i, under the constraint of the virtual budget, maximizes its utility function U.i Determine its resource request size r i Utility function U i Defined as:

[0111] ,

[0112] In the formula, Value i (r i ) Obtain the resource request quantity r for business intelligence agent i. i The resulting expected business value function is positively correlated with the business value density assessment value; P t (r i () refers to the unit price of resources published by the central scheduler; Penalty i (QoS) is a penalty for service agent i failing to meet the quality of service under the current resource allocation.

[0113] Step S323: Each business unit agent uses the multi-agent deep deterministic policy gradient algorithm to learn and optimize policies based on the global resource supply and demand status, resource unit price, and the behavioral history of other agents, through interaction with the environment. Each business unit agent has an independent policy network and shares a centralized commentator network.

[0114] Step S33: Generate the optimal resource allocation scheme by iteratively optimizing the distributed collaborative decision-making.

[0115] The hierarchical decision-making model is iteratively optimized using the Multi-Agent Deep Deterministic Policy Gradient Algorithm (MADDPG). MADDPG works by having each business agent i learn a policy π. i A centralized comment network Q i Under shared global observation, the system learns how to autonomously determine the optimal amount of resources to request given a value price and competitor requests. This algorithm iterates until the resource request strategies of all agents no longer change significantly, eventually converging to an optimal allocation strategy. The system automatically generates the optimal computing power resource allocation scheme for each business unit, ensuring that high-value businesses receive scarce resources first.

[0116] Therefore, through iterative optimization using the Multi-Agent Deep Deterministic Policy Gradient Algorithm (MADDPG), the system gradually converges the resource request strategies of each agent, ultimately forming an efficient resource allocation strategy based on business value. This ensures that high-value businesses automatically adapt and prioritize access to scarce resources. Compared to the heuristic optimization framework of global exploration using the Improved Sparrow Search Algorithm (SSA) and local tuning using the Adaptive Genetic Algorithm (AGA) in existing technologies, this step completely shifts to an optimization paradigm based on hierarchical decision-making and multi-agent deep learning. It focuses on the implementation of value pricing and intelligent bidding allocation, avoiding the complexity of directly addressing the Nash equilibrium in game theory, while still achieving the innovative goal of prioritizing resource allocation for high-value businesses.

[0117] In one embodiment, step S4 includes:

[0118] Step S41: Define a structural causal model based on the topological graph, and in the digital twin environment, intervene in the optimal resource allocation scheme to perform counterfactual reasoning and quantify its potential risks.

[0119] Specifically, the system utilizes a causal inference module to perform counterfactual reasoning within a digital twin sandbox environment. For example, it simulates scenarios such as "What is the probability of actual business loss if resource expansion is implemented according to the current plan?", "What is the probability that the response time of service X will deteriorate to that of service Y if the scheduler does not execute the expansion instruction but maintains the status quo?", or "If resources for service A are reduced by 10%, how much will its GMV decrease?" Through this What-if analysis, the system can quantify and assess the potential risks of scheduling decisions.

[0120] Step S42: When the potential risk is lower than the preset threshold, the optimal resource allocation scheme is encoded into a smart contract and recorded on the private consortium blockchain for evidence storage.

[0121] Among them, preset thresholds include a 1% decrease in GMV, a 5% probability of SLA default, and so on.

[0122] The smart contract contains information such as scheduling instructions, resource ID, timestamp, and responsible party.

[0123] Only scheduling schemes that pass counterfactual risk assessments below a preset threshold are adopted. The final scheme is encapsulated as a smart contract, which is recorded and appended to the enterprise's private consortium blockchain for evidence storage. This ensures the immutability, transparency, and traceability of the scheduling history, supporting subsequent audit trails, financial settlements, compliance checks, and fault diagnosis.

[0124] Step S43: Execute the resource scheduling operations defined in the smart contract through the automated orchestration engine that listens for events on the private consortium blockchain.

[0125] This involves using pre-built automated orchestration engines such as Kubernetes Operator to automatically execute physical layer computing resource allocation instructions, such as container scaling, virtual machine creation / destruction, and network bandwidth adjustment. Compared to existing technologies, this step introduces more advanced causal reasoning and digital twin sandboxes for risk pre-assessment, and combines blockchain technology to achieve transparency and immutability in decision-making, significantly improving the robustness and credibility of the decisions.

[0126] In one embodiment, the above embodiments further include:

[0127] Step S5: Monitor the historical accumulation of multi-source heterogeneous data. When the historical accumulation reaches the preset data maturity threshold, automatically trigger the update of the prediction model and / or multi-agent reinforcement learning algorithm.

[0128] This embodiment automatically acquires and stores massive amounts of historical data on all business and computing power metrics. When the accumulated historical data reaches a preset data maturity threshold, it triggers updates such as retraining or model replacement of the prediction and optimization models based on more complex deep learning algorithms or reinforcement learning algorithms, such as Transformer and Seq2Seq models with Attention mechanisms. Through this mechanism, the algorithm upgrades and performance iterations of the neural graph prediction module and the intelligent optimization allocation module are achieved, ensuring that the system always maintains its optimal state and further optimizing the dynamic scheduling of business resources.

[0129] Therefore, the business resource optimization management method based on computing power demand prediction provided by the embodiments of this application has the following advantages:

[0130] 1. It has achieved comprehensive topological causal awareness and accurate computing power demand prediction capabilities.

[0131] By constructing a multi-source heterogeneous dynamic causal graph and utilizing a spatiotemporal graph attention network to accurately capture the spatial transmission characteristics and cascading effects of traffic and load in complex microservice links, this invention can proactively predict the depth and breadth of the impact of traffic peaks, significantly improving prediction accuracy.

[0132] 2. It possesses excellent robustness in dealing with non-uniform time events.

[0133] By introducing neural ordinary differential equations for time-dimensional modeling, continuous-time modeling of irregular, non-uniformly triggered sudden business events is achieved. This fundamentally solves the problems of high data sampling requirements and poor prediction performance faced by traditional RNN / LSTM models when dealing with business impulses and sparse events, greatly improving the system's adaptability and robustness in real-world business scenarios.

[0134] 3. A market-based tiered value pricing and intelligent bidding allocation mechanism has been established.

[0135] By innovatively introducing hierarchical value pricing and multi-agent deep learning, the resource optimization allocation problem is transformed into an intelligent decision-making process based on value contribution. The resource scheduler dynamically sets value pricing, and the agents of each business unit conduct adaptive bidding. The system can automatically balance resource conflicts between different business units, achieving efficient priority allocation based on business value, avoiding the bias and inefficiency of manual intervention, and greatly improving resource utilization.

[0136] 4. Provides forward-looking decision-making risk control and reliable execution assurance.

[0137] By introducing counterfactual reasoning mechanisms and digital twin sandboxes, a question-and-answer-style risk pre-assessment of scheduling decisions is conducted before execution, significantly reducing decision-making risks. Combined with blockchain evidence storage technology, the transparency, traceability, and immutability of the scheduling scheme are ensured, enhancing system security and compliance. This entire mechanism is more forward-looking and reliable than traditional statistical threshold detection and rule filtering methods.

[0138] Please refer to Figure 2 The embodiments of this application also provide a business resource optimization management system 1 based on computing power demand prediction, including a memory 3, a processor 2, and a computer program stored on the memory 3 and executable on the processor 2. When the processor 2 executes the computer program, it implements the steps of the above embodiments.

[0139] In addition, embodiments of this application also provide a business resource optimization management device based on model-based computing power demand prediction, comprising:

[0140] The graph construction module is used to collect and integrate multi-source heterogeneous data from business activities, business links and infrastructure in real time, and build and continuously update a topological graph representing the dynamic relationship between business entities, logical service entities and physical resource entities based on the multi-source heterogeneous data.

[0141] The neural graph prediction module is used to input the temporal state of the topology graph into a prediction model that integrates graph structure and continuous time modeling. The prediction model performs multidimensional evolution calculations of node states and outputs quantitative values ​​of computing power demand intensity and business value density assessment values ​​for each node in the future time period.

[0142] The intelligent optimization allocation module is used to generate dynamic resource value signals by the central scheduler based on the quantitative value of computing power demand intensity and the evaluation value of business value density within a hierarchical decision-making framework. This signals guide multiple business unit agents to conduct distributed collaborative decision-making using multi-agent reinforcement learning algorithms, and iteratively generate the optimal resource allocation scheme.

[0143] The risk control and feasible execution module is used to perform counterfactual reasoning based on the topology map to quantitatively assess the potential risks before the optimal resource allocation scheme is executed. When the potential risks are lower than a preset threshold, the optimal resource allocation scheme is converted into an automatically executable and tamper-proof scheduling instruction and issued for execution.

[0144] The further technical description and corresponding technical effects of the above-mentioned business resource optimization management system 1 based on computing power demand prediction and the business resource optimization management device based on model computing power demand prediction are all based on the description of various embodiments in the business resource optimization management method based on computing power demand prediction. For example, one embodiment of the business resource optimization management device based on model computing power demand prediction also includes an update module corresponding to step S5, etc.

[0145] Since the systems / devices described in the above embodiments of this application are systems / devices used to implement the methods of the above embodiments of this application, those skilled in the art can understand the specific structure and modifications of the system / devices based on the methods described in the above embodiments of this application, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of this application fall within the scope of protection of this application.

[0146] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0147] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0148] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The words "a" or "an" preceding a component do not exclude the presence of a plurality of such components. This application can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0149] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0150] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0151] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations fall within the scope of the claims of this application and their equivalents, then this application should also include these modifications and variations.

Claims

1. A business resource optimization management method based on computing power demand forecasting, characterized in that, include: Step S1: Collect and integrate multi-source heterogeneous data from business activities, business links, and infrastructure in real time. Based on the multi-source heterogeneous data, construct a heterogeneous information network as a topology graph, which includes business activity nodes, logical service nodes, data instance nodes, and physical resource nodes. Nodes are connected through business event triggering edges, service call edges, deployment attribution edges, and causal influence edges. Introduce dynamic attributes to the node characteristics and edge weights in the heterogeneous information network, and update them in real time through a sliding time window mechanism to form a dynamic evolution graph sequence that represents the temporal state of the topology graph. Step S2: The temporal state of the topology graph is input into a prediction model that integrates graph structure and continuous time modeling. The prediction model uses a multi-head graph attention network to process the topology graph, aggregating the neighbor information of each node to obtain the node state representing the spatial transmission relationship between nodes. The prediction model uses the node state representing the spatial transmission relationship between nodes as the initial state of the neural ordinary differential equation. By solving the neural ordinary differential equation, the evolution of the node state in the continuous time domain is modeled and predicted, and the quantitative value of computing power demand intensity and the business value density assessment value corresponding to each node in the future time period are output. The business value density assessment value is positively correlated with GMV contribution rate and user retention rate. Step S3: Based on the quantified value of computing power demand intensity and the assessed value of business value density, under a hierarchical decision-making framework, the central scheduler dynamically generates a resource value signal according to global resource supply, total demand forecast, operating costs, and service level agreement constraints. Each business unit, as an agent, obtains a virtual budget based on its corresponding assessed value of business value density. Guided by the resource value signal, it uses a multi-agent reinforcement learning algorithm to perform distributed collaborative decision-making to determine its respective resource request amount. By iteratively optimizing the distributed collaborative decision-making, an optimal resource allocation scheme is generated. Step S4: Before the optimal resource allocation scheme is executed, counterfactual reasoning is performed based on the topology map to quantitatively assess its potential risks. When the potential risks are lower than a preset threshold, the optimal resource allocation scheme is converted into an automatically executable and tamper-proof scheduling instruction and issued for execution.

2. The commercial resource optimization management method based on computing power demand forecasting according to claim 1, characterized in that, In step S2, the prediction model uses the node states, which characterize the spatial transmission relationship between nodes, as the initial states of the neural ordinary differential equation. By solving the neural ordinary differential equation, the evolution of the node states in the continuous time domain is modeled and predicted, and the quantitative values ​​of computing power demand intensity and business value density assessment values ​​corresponding to each node in the future time period are output, including: Step S231: The neural ordinary differential equation defines the evolution of the node hidden state as follows: ， In the formula, h(t) represents the hidden state vector of a node over continuous time t, and f is a vector vector defined by parameters θ. ODE A defined neural network is used to model the dynamics of state changes; Step S232: Using the node state, which characterizes the spatial transmission relationship between nodes, as the initial state, the neural ordinary differential equation is solved by integration using an ordinary differential equation numerical solver to obtain the predicted node state value at future time. Step S233: Based on the node state prediction value, the prediction model outputs the quantitative value of computing power demand intensity and the evaluation value of business value density for each node in the future time period.

3. The commercial resource optimization management method based on computing power demand forecasting according to claim 1, characterized in that, In step S3, each business unit, acting as an agent, obtains a virtual budget based on its corresponding business value density assessment value. Guided by the resource value signal, it uses a multi-agent reinforcement learning algorithm to perform distributed collaborative decision-making and determine its respective resource request amount, including: Step S321: Each business unit, acting as an intelligent agent, obtains a virtual budget based on its corresponding business value density assessment value. Step S322: Each business unit agent i, under the constraint of the virtual budget, maximizes its utility function U. i Determine its resource request size r i The utility function U i Defined as: ， In the formula, Value i (r i ) Obtain the resource request quantity r for business intelligence agent i. i The resulting expected business value function is positively correlated with the business value density assessment value; P t (r i () refers to the unit price of resources published by the central scheduler; Penalty i (QoS) is a penalty for service agent i failing to meet the quality of service under the current resource allocation; Step S323: Each business unit agent uses a multi-agent deep deterministic policy gradient algorithm to learn and optimize policies based on the global resource supply and demand status, resource unit price, and the behavioral history of other agents, through interaction with the environment. Each business unit agent has an independent policy network and shares a centralized commentator network.

4. The business resource optimization management method based on computing power demand forecasting according to claim 1, characterized in that, Also includes: Step S5: Monitor the historical accumulation of the multi-source heterogeneous data. When the historical accumulation reaches a preset data maturity threshold, automatically trigger the update of the prediction model and / or the multi-agent reinforcement learning algorithm.

5. The commercial resource optimization management method based on computing power demand forecasting according to claim 1, characterized in that, In step S1, based on the multi-source heterogeneous data, a heterogeneous information network including business activity nodes, logical service nodes, data instance nodes, and physical resource nodes is constructed as the topology map, including: Step S121: Learn the causal relationships between nodes from historical time-series data using a causal discovery algorithm, and generate the causal influence edge; Step S122: Based on the multi-source heterogeneous data, determine the connection relationships and weights of the business event triggering edge, service call edge, and deployment affiliation edge.

6. A commercial resource optimization management method based on computing power demand forecasting according to any one of claims 1 to 5, characterized in that, Step S4 includes: Step S41: Based on the topological map, define a structural causal model, and in the digital twin environment, intervene in the optimal resource allocation scheme to perform counterfactual reasoning and quantify its potential risks. Step S42: When the potential risk is lower than a preset threshold, the optimal resource allocation scheme is encoded into a smart contract and recorded in a private consortium blockchain for evidence storage. Step S43: Execute the resource scheduling operation defined in the smart contract by using the automated orchestration engine that listens to the events of the private consortium blockchain.

7. A business resource optimization management system based on computing power demand prediction, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a business resource optimization management method based on computing power demand prediction as described in any one of claims 1 to 6.

Images