Capacity management method and device

Abstract

The invention provides a capacity management method and device. The method comprises the following steps: determining a first feature tensor according to a preset time window, business event features and environment features; wherein the first feature tensor comprises a plurality of data dimensions; inputting the first feature tensor into a hybrid neural network model, and predicting the resource demand quantity at a specific time point in the future to obtain prediction information corresponding to the resource demand quantity at the specific time point in the future; performing abnormal data identification on the real-time data flow according to an unsupervised learning algorithm to obtain abnormal information; and determining an optimal capacity management strategy according to the prediction information and the abnormal information. Through the embodiment of the invention, the problem of low capacity management efficiency caused by the fact that the traditional CMP passively responds to resource alarm, depends on manual decision and execution and cannot predict future demands or global optimization in the related technology is solved.

Description

Technical Field

[0001] This invention relates to the fields of cloud computing operation and maintenance management and artificial intelligence operation and maintenance, and more specifically, to a capacity management method and apparatus. Background Technology

[0002] Effective capacity management is a continuous, closed-loop optimization process. Ideally, it should encompass all stages—sensing, analysis, decision-making, execution, and verification—and be able to anticipate future needs. Traditional cloud management platforms (CMPs) typically only implement the "sensing," "analysis," and "verification" stages. The "decision-making" and "execution" stages rely entirely on external human intervention.

[0003] In the "analysis" phase, the relevant technologies are limited to comparing current or historical data with static thresholds, unable to predict future state changes, and often only triggering alarms after resources are exhausted and performance is compromised. Furthermore, because threshold comparisons are typically based on a single resource metric (e.g., high CPU usage on a particular virtual machine), they cannot comprehensively analyze multi-dimensional data such as performance, cost, cross-cloud resource lists, and business requirements. Decisions made based on such a single alarm are likely to be locally optimal but globally suboptimal.

[0004] Furthermore, in related technologies, because the system itself lacks "decision-making" and "execution" capabilities, the workflow ends with sending the analysis results (alarms) to the administrator, lacking subsequent root cause analysis and solution development. A series of actions, such as logging into the cloud platform and manual operations, all rely on manual completion. This leads to time delays and operational risks due to personnel skills and fatigue, failing to meet the rapid scaling requirements of cloud-native applications. Summary of the Invention

[0005] This invention provides a capacity management method and apparatus to at least solve the problem in the related art where traditional CMP passively responds to resource alarms, relies on manual decision-making and execution, cannot predict future demand or global optimization, and thus results in low capacity management efficiency.

[0006] According to an embodiment of the present invention, a capacity management method is provided, comprising: determining a first feature tensor based on a preset time window, business event characteristics, and environmental characteristics; wherein the first feature tensor includes multiple data dimensions; inputting the first feature tensor into a hybrid neural network model to predict resource demand at a specific future time point, thereby obtaining prediction information corresponding to the resource demand at the specific future time point; identifying abnormal data in a real-time data stream using an unsupervised learning algorithm to obtain abnormal information; and determining an optimal capacity management strategy based on the prediction information and the abnormal information.

[0007] In one embodiment, a multidimensional feature tensor is determined based on a preset time window, business event characteristics, and environmental characteristics. Before determining the first feature tensor, the method further includes: collecting various types of data from multiple heterogeneous cloud platforms; wherein the various types of data include at least one of the following: external environment data, performance data, configuration data, and cost data; standardizing the various types of data and mapping the standardized data to a predefined metadata model to obtain metadata; and storing the metadata in a data lake.

[0008] In one embodiment, determining a multidimensional feature tensor based on a preset time window, business event characteristics, and environmental characteristics, and determining a first feature tensor, includes: obtaining the business event characteristics and the environmental characteristics from various metadata in the data lake; and determining the first feature tensor based on the time window, the business event characteristics, and the environmental characteristics.

[0009] In one embodiment, the hybrid neural network model includes a convolutional neural network (CNN) model and a long short-term memory (LSTM) network model. The first feature tensor is input into the hybrid neural network model to predict resource demand at a specific future time point, obtaining prediction information corresponding to the resource demand at that specific future time point. This includes: inputting the first feature tensor into the CNN model to obtain a second feature tensor; wherein the second feature tensor has a higher level of abstraction than the first feature tensor; and inputting the second feature tensor into the LSTM model to obtain prediction information corresponding to the resource demand at that specific future time point.

[0010] In one embodiment, inputting the second feature tensor into the LSTM model to obtain prediction information of resource demand corresponding to the specific future time point includes: inputting the second feature tensor into the LSTM model to perform the first layer prediction of the LSTM model to obtain application layer resource demand prediction information; inputting the application layer resource demand prediction information into the LSTM model to perform the second layer prediction of the LSTM model to obtain the prediction information of resource demand at the specific future time point.

[0011] In one embodiment, the capacity management method further includes: optimizing the hybrid neural network model with the goal of minimizing the loss function of the business objective.

[0012] In one embodiment, determining the optimal capacity management strategy based on the prediction information and the anomaly information includes: determining multiple capacity management strategies according to predefined business strategy rules; simulating the multiple capacity management strategies based on the prediction information and the anomaly information to obtain multiple simulation results; analyzing the multiple simulation results using a multi-objective optimization algorithm, and determining the optimal capacity management strategy from the multiple capacity management strategies based on the analysis results.

[0013] In one embodiment, after determining the optimal capacity management strategy based on the prediction information and the anomaly information, the method further includes: obtaining multiple atomic operation steps and constructing a directed acyclic graph based on the optimal capacity management strategy; and executing the optimal capacity management strategy based on the atomic operation steps and the directed acyclic graph.

[0014] According to another embodiment of the present invention, a capacity management device is also provided, comprising: a determination module, configured to determine a first feature tensor based on a preset time window, business event characteristics, and environmental characteristics; wherein the first feature tensor includes multiple data dimensions; a prediction module, configured to input the first feature tensor into a hybrid neural network model to predict the resource demand at a specific future time point, thereby obtaining prediction information corresponding to the resource demand at the specific future time point; an identification module, configured to identify abnormal data in a real-time data stream using an unsupervised learning algorithm, thereby obtaining abnormal information; and a management module, configured to determine an optimal capacity management strategy based on the prediction information and the abnormal information.

[0015] According to yet another embodiment of the present invention, a storage medium is also provided, wherein a computer program is stored therein, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0016] According to yet another embodiment of the present invention, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0017] According to yet another embodiment of the present invention, a computer program product is also provided, comprising a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0018] Through the above embodiments of the present invention, by constructing a multi-dimensional first feature tensor and utilizing a hybrid neural network model to predict future resource demands, proactive prediction, automatic execution, and closed-loop optimization of resource management are achieved. Therefore, this addresses the problem in related technologies where traditional CMP passively responds to resource alarms, relies on manual decision-making and execution, and cannot predict future demands or global optimization, resulting in low capacity management efficiency. This ultimately improves the efficiency and accuracy of resource management. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0020] Figure 1 This is a hardware structure block diagram of a computer terminal for the operation capacity management method according to an embodiment of the present invention;

[0021] Figure 2 This is a flowchart of a capacity management method according to an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the complete process of the capacity management method according to an embodiment of the present invention;

[0023] Figure 4 This is a structural block diagram of a capacity management device according to an embodiment of the present invention;

[0024] Figure 5 This is a structural block diagram of a capacity management device according to another embodiment of the present invention. Detailed Implementation

[0025] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0027] The following is an explanation of the technical terms used in the embodiments of the present invention:

[0028] Multi-cloud architecture: is an architecture that includes multiple public clouds from different cloud service providers and can be integrated with private cloud environments.

[0029] Capacity management is a core area of ​​IT operations, designed to ensure that IT resources can meet current and future business needs in a cost-effective and timely manner. Its scope includes current capacity monitoring as well as future planning and optimization.

[0030] End-to-end management paradigm: End-to-end is a complete closed-loop process in capacity management that includes data collection, monitoring and analysis, predictive planning, decision making, automated execution, and effect verification.

[0031] Unified Data Model: A standardized data schema used to integrate heterogeneous monitoring data from different cloud platforms (such as AWS EC2 and Azure VMs). It "translates" and "maps" data from different sources and in different formats into a unified language.

[0032] Declarative adapters: A type of software component that defines its behavior by declaring the required configuration (such as a YAML or JSON file). For example, defining how to parse API responses from a cloud platform through a configuration file.

[0033] CNN-LSTM Hybrid Model: A deep learning model that combines a convolutional neural network (CNN) and a long short-term memory network (LSTM).

[0034] Among them, CNN excels at extracting local features and spatial patterns from data (e.g., identifying short-term fluctuation patterns from time series).

[0035] LSTM excels at capturing long-term dependencies and periodic patterns in time series data (e.g., daily peaks, weekly peaks).

[0036] Multi-objective optimization: A mathematical optimization method that can simultaneously optimize multiple objectives, typically conflicting objectives. In embodiments of this invention, objectives may include lowest cost, best performance (i.e., lowest latency), highest resource utilization, and highest reliability.

[0037] Sandbox simulation environment: A virtual replica isolated from the production environment, used to safely test and verify the potential impact of change scenarios (such as scaling up or migrating) without interfering with actual business operations.

[0038] Automated orchestration: Orchestration refers to the process of coordinating multiple automated tasks to form a complex workflow. For example, a scaling operation might require sequentially invoking: creating virtual machines on cloud platform A, configuring load balancing, and updating DNS records.

[0039] What-if analysis: A hypothesis analysis technique that allows users to change input parameters (e.g., "user traffic increases by 50%" or "all instances are switched to Spot instances") and simulate the potential impact of these changes on the system (e.g., cost, performance).

[0040] Closed-loop feedback refers to the system using the execution results (e.g., actual performance data after scaling up) as input again to evaluate the correctness of the decision and to optimize future decision models.

[0041] Service Level Agreement (SLA): A formal contract between a service provider and a customer that specifies quantifiable service metrics (such as monthly uptime of at least 99.9%).

[0042] Total Cost of Ownership: The total cost of a technology or solution, including not only direct procurement costs (cloud instance fees) but also indirect costs such as management expenses, maintenance manpower costs, and business losses due to downtime.

[0043] The capacity management method and system provided in this invention include the following technologies:

[0044] a) Cloud Computing and Multi-Cloud Architecture: Infrastructure as a Service has become the standard paradigm for modern IT. To avoid vendor lock-in, meet compliance requirements, and pursue cost-effectiveness, enterprises generally use resources from multiple cloud service providers (such as AWS, Azure, GCP, Alibaba Cloud, etc.), forming a complex heterogeneous multi-cloud environment. This is the core scenario that this invention aims to address.

[0045] b) Cloud Application Programming Interfaces and Software Development Kits: All major cloud service providers offer comprehensive APIs and SDKs, allowing programmatic querying, configuration, management, and monitoring of their cloud resources. This is the technological foundation for achieving unified data collection and automated control across cloud resources.

[0046] c) Unified Monitoring and Cloud Management Platform: Well-known monitoring tools (such as Prometheus and Zabbix) and commercial CMPs (such as VMware vRealize and Flexera) can already collect basic resource inventory, configuration, and performance metric data (such as CPU utilization and memory usage) from multiple cloud platforms via proxies or APIs, and display them on a unified dashboard. This solves the "visibility" problem and is the prototype of the "data acquisition layer" in this invention.

[0047] d) Infrastructure as Code: Tools such as Terraform, Ansible, and AWS CloudFormation have become the standard for cloud resource management. They define and deploy infrastructure through declarative code templates, achieving versioning and a certain degree of automation in resource management. This is one of the core technologies upon which the "automation execution layer" in this invention relies.

[0048] e) Time Series Forecasting and Machine Learning: In the field of operations and maintenance, using statistical models (such as ARIMA) or machine learning models (such as LSTM) to analyze historical monitoring indicators (time series data) to predict future trends and detect anomalies is a well-known technical direction in the AIOps field. This provides the theoretical and technical foundation for the intelligent forecasting achieved in this invention.

[0049] The method embodiment provided in Embodiment 1 of this application can be executed on a computer terminal or similar computing device. Taking running on a computer terminal as an example, Figure 1 This is a hardware structure block diagram of a computer terminal for implementing the capacity management method according to an embodiment of the present invention. Figure 1 As shown, computer terminal 10 may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. Optionally, the computer terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the computer terminal described above. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0050] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the capacity management method in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0051] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0052] This embodiment provides a method that runs on the aforementioned computer terminal or network architecture. Figure 2 This is a flowchart of a capacity management method according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps:

[0053] Step S202: Determine a first feature tensor based on a preset time window, business event characteristics, and environmental characteristics; wherein the first feature tensor includes multiple data dimensions.

[0054] Step S204: Input the first feature tensor into the hybrid neural network model to predict the resource demand at a specific future time point, and obtain the prediction information corresponding to the resource demand at the specific future time point.

[0055] Step S206: Identify abnormal data in the real-time data stream using an unsupervised learning algorithm to obtain abnormal information;

[0056] Step S208: Determine the optimal capacity management strategy based on the predicted information and the anomaly information.

[0057] Through the above steps, by constructing a multi-dimensional first feature tensor and utilizing a hybrid neural network model to predict future resource demands, proactive prediction, automatic execution, and closed-loop optimization of resource management are achieved. Therefore, this addresses the problem of traditional CMP (Capacity Management Program) passively responding to resource alarms and relying on manual decision-making and execution, which fails to predict future demands or global optimization, resulting in low capacity management efficiency. This approach ultimately improves the efficiency and accuracy of resource management.

[0058] In a multi-cloud architecture, related technologies cannot connect the data silos of multiple clouds and lack a unified understanding and modeling of the data. The embodiments of the present invention can build a unified data collection and standardization layer, break down cloud platform barriers, and realize unified modeling and real-time aggregation of multi-cloud resource data.

[0059] Specifically, prior to step S202, the capacity management method further includes: collecting various types of data from multiple heterogeneous cloud platforms; wherein the various types of data include at least one of the following: external environment data, performance data, configuration data, and cost data; standardizing the various types of data and mapping the standardized data to a predefined metadata model to obtain metadata; and storing the metadata in a data lake to provide data support for upper-layer analysis.

[0060] In one exemplary embodiment, collecting various types of data from multiple heterogeneous cloud platforms includes: collecting various types of data from multiple heterogeneous cloud platforms through declarative adapters; wherein the declarative adapters are a set of pluggable components, each adapter defining how to interface with the APIs of a specific cloud platform (such as AWS, Azure, Alibaba Cloud) or system (such as Kubernetes) through a configuration file, and collecting performance data, configuration data, and cost data.

[0061] In this embodiment, performance data includes, but is not limited to, one of the following: CPU, memory, disk I / O, and network;

[0062] Configuration data includes, but is not limited to, one of the following: instance type, region;

[0063] Cost data includes, but is not limited to, one of the following: bills, rates.

[0064] In one embodiment, the standardization process for the various data includes: cleaning, transforming, and standardizing the collected data, and then mapping them uniformly to a predefined metadata model to eliminate differences between cloud platforms and form a consistent data view.

[0065] In one embodiment, determining a multidimensional feature tensor based on a preset time window, business event characteristics, and environmental characteristics, and determining a first feature tensor, includes: obtaining the business event characteristics and the environmental characteristics from various metadata in the data lake; and determining the first feature tensor based on the time window, the business event characteristics, and the environmental characteristics.

[0066] For massive, multi-dimensional operational data, existing technologies lack intelligent methods for processing, making it impossible to predict the future and extract deeper value from the data. This invention can introduce an AI intelligent prediction engine, which can predict future resource needs based on historical data and external characteristics (such as business activities), transforming passive response into proactive planning.

[0067] In this embodiment, the AI ​​intelligent prediction engine is a CNN-LSTM hybrid neural network model. The CNN component is used to extract local and short-term patterns in the data, the LSTM component is used to capture long-term dependencies and periodic patterns, and it can also integrate external features (such as marketing calendars and weather forecasts) to output accurate resource demand predictions for specific future time points.

[0068] Specifically, the hybrid neural network model includes a convolutional neural network (CNN) model and a long short-term memory (LSTM) network model. The first feature tensor is input into the hybrid neural network model to predict resource demand at a specific future time point, obtaining prediction information corresponding to the resource demand at that specific future time point. This includes: inputting the first feature tensor into the CNN model to obtain a second feature tensor; wherein the second feature tensor has a higher level of abstraction than the first feature tensor; and inputting the second feature tensor into the LSTM model to obtain prediction information corresponding to the resource demand at that specific future time point.

[0069] In this embodiment, the first feature tensor is a three-dimensional feature tensor, including a time dimension, a spatial dimension, and an instance dimension.

[0070] The time dimension represents the time-series characteristics of the data, allowing the model to understand the trends and periodic patterns of data changes over time. In this dimension, the data at each point in time records the state of resource requirements and other relevant indicators at that specific moment.

[0071] By analyzing past time points, the model can identify seasonal variations in resource demand, differences between weekdays and weekends, and even fluctuations at different times of the day. This dimension of data forms the basis for time-series forecasting, helping the model learn and predict future trends in resource demand.

[0072] Spatial Dimension: This reflects the geographical distribution of data, encompassing information from different cloud service providers, regions, and availability zones. This dimension of data is not limited to a single cloud platform or geographical location, but spans multiple cloud environments and geographical areas.

[0073] Since resource prices, network latency, and regulatory requirements can vary across different cloud service providers and regions, spatial data helps models understand the spatial distribution patterns of resource demand and how geographical location influences it. This is crucial for cost optimization and meeting performance and compliance requirements, especially in multi-cloud heterogeneous environments.

[0074] Instance Dimension: This refers to the attributes and status of a specific resource instance, such as a virtual machine, container, or database instance. Performance metrics (such as CPU utilization and memory utilization), configuration details (such as instance type and storage size), and cost information (such as billing method and unit price) for each resource instance are all reflected in this dimension.

[0075] Instance-level data enables models to differentiate and understand the differences between instances, including their performance characteristics, cost-effectiveness, and behavior under specific workloads. This is crucial for granular resource demand forecasting, as the same type of resource can exhibit drastically different demand patterns across different instances.

[0076] In an exemplary embodiment, the dimensions of the first feature tensor are as follows:

[0077] Dimension 1 (Time Step): Preset time window (e.g., the past 14 days) to provide time-series dependencies.

[0078] Dimensions two and three (feature channels): Each feature channel represents a different data source or perspective, collectively forming a panoramic view.

[0079] For example: Channel A (infrastructure performance metrics): Multi-dimensional metrics such as CPU, memory, disk I / O, and network throughput across multiple cloud platforms have been standardized through a unified data layer.

[0080] Channel B (Business and Application Metrics): Business data directly related to resource consumption, such as queries per second (QPS), daily active users (DAU), and order volume. This establishes a causal relationship between resources and business, transforming forecasting from "resource-driven" to "business-driven."

[0081] Channel C (External Events and Calendar Features): Known future events encoded in vector form, such as "marketing campaign dates," "holidays," "paydays," and "cloud platform maintenance windows." This gives the model the ability to predict known, recurring events.

[0082] Channel D (Cost and Configuration Metadata): Information such as resource specifications, unit price, and availability zone location, providing context for subsequent cost optimization decisions.

[0083] Channel E (Resource Instance Characteristics): Each instance-level channel covers detailed information about a specific resource instance, including but not limited to instance type, hardware configuration (such as the number of CPU cores and memory size), operating system version, and storage type (SSD or HDD). This channel helps the model distinguish the characteristics and performance differences of each instance, thus taking into account the heterogeneity and diversity of resources in predictions.

[0084] Channel F (Instance Historical Behavior): Collects and analyzes historical usage data for each resource instance, including peak resource consumption, average load, and number of anomalies. By understanding the past performance of each instance, the model can predict its possible behavior patterns under specific business loads or external events, enabling more accurate personalized resource demand forecasting.

[0085] This invention combines machine learning and time series analysis techniques to intelligently predict resource demands and potential bottlenecks at specific future points in time. This proactive predictive capability can trigger corresponding optimization processes before performance issues occur. In contrast, traditional CMP typically relies solely on comparing current or historical data with static thresholds, failing to predict future changes. This invention overcomes this limitation, avoiding alerts only after resource depletion or performance degradation.

[0086] In this embodiment, the CNN is not used to process images, but plays the role of an "intelligent feature detector", that is, the CNN layer can extract spatial features across clouds and across metrics.

[0087] In images, CNNs are used to detect spatial local patterns such as edges and textures; in this embodiment of the invention, in the constructed three-dimensional tensor, CNNs perform convolution operations on the "feature channel" dimension to automatically discover and fuse complex, nonlinear collaborative change relationships between different indicators.

[0088] For example, a convolutional kernel might automatically learn cross-domain association rules such as "when the business QPS (channel B) and network inflow traffic (channel A) rise together under a specific pattern, and it is a weekday (channel C), it indicates that the CPU (channel A) will reach its peak in 2 hours." This is something that manual rules or a single LSTM model can hardly capture efficiently.

[0089] LSTM layers can learn long-cycle dependencies and dynamic modeling of cloud resources.

[0090] Specifically, the high-level feature sequences refined by CNN are fed into the LSTM layer, which learns and memorizes the compound cyclical patterns of business load at the daily, weekly, monthly, and even quarterly levels.

[0091] For example, learn about the peak morning patterns on Mondays, the surge in computing load during the end-of-month settlement cycle, and the cyclical impact of price fluctuations of specific cloud services (such as AWS Spot instances) on resource demand.

[0092] Through the collaborative learning of LSTM and CNN—CNN is responsible for "seeing" the complex status quo across clouds and indicators at every moment, while LSTM is responsible for "understanding" the evolution of these status quo over time—the combination of the two can achieve "dual understanding of time and space".

[0093] In one embodiment, inputting the second feature tensor into the LSTM model to obtain prediction information of resource demand corresponding to the specific future time point includes: inputting the second feature tensor into the LSTM model to perform the first layer prediction of the LSTM model to obtain application layer resource demand prediction information; inputting the application layer resource demand prediction information into the LSTM model to perform the second layer prediction of the LSTM model to obtain the prediction information of resource demand at the specific future time point.

[0094] The capacity management method provided in this embodiment of the invention is a hierarchical-joint training strategy oriented towards management objectives.

[0095] Hierarchical prediction means that the model does not only predict a single general "load," but can also output multiple related targets simultaneously, forming a prediction pyramid:

[0096] Top level: Business metric forecast (such as the next day's QPS peak).

[0097] Middle layer: Application layer resource requirements derived from this (such as the number of container replicas required).

[0098] The bottom layer: the final IaaS layer resource requirements (such as vCPU cores and memory GB). This layered structure makes the prediction process interpretable and traceable.

[0099] Joint training, on the other hand, involves training the model end-to-end with a loss function that minimizes the final business objective (such as the cost of excess resources plus the SLA penalty cost of insufficient resources). This means that it is not for the sake of prediction, but directly serves the optimal business decision (a balance between cost and stability).

[0100] In one embodiment, the capacity management method further includes: optimizing the hybrid neural network model with the goal of minimizing the loss function of the business objective.

[0101] Specifically, the hybrid neural network model is optimized with the goal of minimizing the loss function of the business objective as the optimization objective, including: end-to-end training of the hybrid neural network model with the goal of minimizing the loss function of the business objective as the optimization objective.

[0102] End-to-end training refers to the process where all components and layers of the model are treated as a whole and optimized from the moment it receives raw input data until it produces the final business decision or prediction result. This training method avoids the information loss and performance degradation that may occur with layer-by-layer training, ensuring that the model's input-to-output transformation process can maximize its service to the final business objective.

[0103] In traditional machine learning tasks, loss functions are typically directly related to prediction error, such as mean squared error (MSE) or cross-entropy loss. However, in business-oriented scenarios, the design of loss functions needs to consider broader business impacts, such as cost, performance, and customer satisfaction.

[0104] For example, when the ultimate business objective is to minimize the cost of excess resources and the SLA penalty cost of insufficient resources, the loss function needs to comprehensively consider the impact of prediction accuracy and resource allocation decisions on costs and SLAs.

[0105] In this embodiment, during end-to-end training, each layer and component of the model is jointly optimized based on the loss function of the final business objective. For the hybrid neural network model (CNN-LSTM model), the feature extraction capability of the CNN layer, the time series prediction capability of the LSTM layer, and the resource optimization strategy of the decision layer are all adjusted to minimize the combined cost of resource excess and resource deficiency.

[0106] This optimization process is achieved through the backpropagation algorithm, which starts from the output layer and calculates the contribution of each parameter to the final loss function layer by layer, thereby adjusting the parameters to minimize this loss.

[0107] In this embodiment of the invention, the CNN-LSTM model differs from existing academic models in the following ways:

[0108] 1. After each automated decision (such as scaling up / down), the system records the "predicted value," "actual observed value," and "decision action" and its effect. This data constitutes new high-quality training samples.

[0109] 2. CNN-LSTM models support online learning or periodic incremental training. They continuously absorb feedback data to correct biases, learn and adjust strategies, and personalize settings.

[0110] Correcting deviations: Adapting to gradual changes in business models (such as user growth trends).

[0111] Learn and adjust strategies: Understand the feedback effect of automation strategies on resource usage patterns, thereby making predictions that are more in line with system dynamics.

[0112] Personalization: Evolving unique predictive models for different applications or departments.

[0113] In one embodiment, determining the optimal capacity management strategy based on the prediction information and the anomaly information includes: determining multiple capacity management strategies according to predefined business strategy rules; simulating the multiple capacity management strategies based on the prediction information and the anomaly information to obtain multiple simulation results; analyzing the multiple simulation results using a multi-objective optimization algorithm, and determining the optimal capacity management strategy from the multiple capacity management strategies based on the analysis results.

[0114] In an exemplary embodiment, determining an optimal capacity management strategy based on the prediction information and the anomaly information includes:

[0115] Store predefined business strategy rules, such as: "cost priority", "performance priority", "SLA 99.99% must be met", "production environment resources must not be oversold", etc.

[0116] Provides a simulation environment isolated from the production environment; this simulation environment can simulate the execution effects of different operation schemes (such as: expansion, migration, purchasing reserved instances, and switching to Spot instances) based on the prediction results, and calculate the estimated cost, performance and risk index of each scheme.

[0117] The multi-objective decision-maker receives the results from the simulator, uses multi-objective optimization algorithms (such as genetic algorithms) to perform trade-off analysis, and finally selects the optimal capacity management strategy that best suits the business strategy, and generates an executable plan.

[0118] Multi-objective optimization algorithms may involve the following steps:

[0119] Step 1: Quantification of decision-making objectives and strategy modeling, transforming vague business intentions into calculable mathematical objectives.

[0120] 1. Definition of Multi-Objective Function: The system quantifies business requirements into objective functions that can be optimized in parallel, typically including:

[0121] Cost function (Minimize Cost): Minimizes the estimated total cost over a future period (such as next month), including on-demand instances, reserved instances, Spot instances, and data transfer costs.

[0122] Performance function (Maximize Performance / Minimize Violation): Minimize the risk of performance SLA violation, such as ensuring that the probability of application response time <200ms is greater than 99.9%.

[0123] Stability function (Maximize Reliability): Maximizes resource redundancy, such as ensuring that critical services are distributed across availability zones.

[0124] 2. Policy and rule library: Define business constraints and policies in code form, such as: "Production database instances must not use Spot instances", "EU user data must remain in the EU region", and "Automatically merge idle resources during off-peak hours in the early morning".

[0125] Step Two: Simulation-Based Quantitative Analysis and "What-if" Deduction This is the "sand table" layer of decision-making, where the consequences of various choices are evaluated in a virtual environment.

[0126] 1. Solution Generation: Based on AI-predicted future load, the system automatically generates multiple candidate solutions. For example, to cope with a predicted traffic peak, solutions might include:

[0127] Option A: Horizontally expand 10 standard instances in the existing region of Cloud Merchant X.

[0128] Option B: Activate a new cluster of Spot instances in a region with lower costs from cloud provider Y, and redirect traffic through global load balancing.

[0129] Option C: Temporarily upgrade the existing instance to a higher specification.

[0130] 2. Sandbox Simulation: The system accurately simulates the operation of each scenario within a future time window (e.g., 24 hours) in an isolated simulation environment. The simulator calculates the Key Performance Indicators (KRIs) for each scenario:

[0131] Economic KRI: Estimated costs accurate to the dollar.

[0132] Performance KRI: Estimated peak CPU utilization, network latency, and SLA compliance rate.

[0133] Risk KRI: Spot instance outage probability, single point of failure risk score.

[0134] 3. Generate a decision matrix: Summarize the simulation results of all schemes into a clear decision matrix for objective comparison by optimization algorithms or administrators.

[0135] Step 3: Multi-objective optimization and automatic decision-making. This is the "decision-making" layer, which automatically selects or recommends the optimal solution based on the quantitative results.

[0136] 1. Optimization Solution: The system inputs a multi-objective optimization problem (such as "keeping costs as low as possible while keeping the SLA violation rate less than 0.1%) into a solver (such as a solver based on genetic algorithms or reinforcement learning). This solver searches a large solution space to find a set of "Pareto optimal" solutions (i.e., a set of solutions that cannot further improve either objective without compromising the other objective).

[0137] 2. Automated Decision-Making: Based on pre-defined high-level strategies (such as "the core objective for this quarter is cost reduction"), the system automatically selects the final execution plan from the Pareto optimal set. For complex or high-risk decisions, the system generates an approval work order with detailed simulation comparison data and pushes it to the administrator for final decision-making, realizing human-machine collaborative decision-making.

[0138] In related technologies, there is a lack of automated and closed-loop methods to translate analytical insights into effective operational actions. This invention, through the aforementioned multi-objective optimization decision engine, prioritizes the scheduling of resources with the best cost-effectiveness while satisfying performance and reliability requirements, and provides optimization suggestions. An automated orchestration and execution engine is designed to transform optimization schemes into automatically executable instruction sequences, thereby achieving closed-loop management.

[0139] Specifically, after step S208, the capacity management method further includes: obtaining multiple atomic operation steps according to the optimal capacity management strategy, and constructing a directed acyclic graph; and executing the optimal capacity management strategy according to the atomic operation steps and the directed acyclic graph.

[0140] In one exemplary embodiment, the secure and reliable translation of a decision plan into actual cloud operations includes:

[0141] For high-risk operations or operations that require manual confirmation according to strategy, a work order is generated and routed to the appropriate approver.

[0142] The plan generated by the multi-objective decision maker is translated into a series of executable atomic operation steps (e.g., 1. Call Terraform to create resources; 2. Call Ansible to configure the application; 3. Update the load balancer configuration) and constructed into a directed acyclic graph (DAG) to ensure execution order and dependencies.

[0143] The executor securely executes the DAG workflow provided by the orchestrator, completing the final resource operations by calling the APIs of various cloud platforms. Compliance checks are performed before execution, and the status is written back after execution.

[0144] In one embodiment, the capacity management method further includes: providing a human-machine interface via a console to display the global status, send alarms, and generate reports.

[0145] The console includes at least one of the following: global dashboard, alert and alarm center, reporting center, and What-if analysis simulator.

[0146] In this embodiment of the invention, the data acquisition layer provides fuel for the AI ​​engine; the AI ​​engine provides input for the decision engine; the decision engine directs the execution layer's actions; and the results of the execution layer are fed back to the system through the acquisition layer, forming a continuously optimized intelligent closed loop. A visual console is used throughout, providing an interface for interaction between humans and the system.

[0147] Related technologies either focus only on the real-time load of a single instance or only on resource quotas without providing insight into the actual usage of the resource pool. However, in the multi-cloud environment of this invention embodiment, a sustainable monitoring and feedback mechanism addresses the prominent issues of resource idleness and waste. It not only focuses on real-time load but also analyzes resource idleness patterns from a macro-time span and provides recycling suggestions.

[0148] Figure 3 This is a schematic diagram of the complete flow of the capacity management method according to an embodiment of the present invention, as shown below. Figure 3 As shown, the method includes the following steps:

[0149] Step S301: Unify the data acquisition and adaptation layer, and input the obtained standardized data into the core AI intelligent engine; wherein, the core AI intelligent engine is equipped with a prediction module and an anomaly detection module, and the prediction module is further equipped with a CNN-LSTM model.

[0150] In this embodiment, the standardized data is a multidimensional feature vector, which is functionally equivalent to the first feature tensor in the above embodiment.

[0151] In step S302, the core AI intelligent engine predicts the resource demand at a specific future time point based on the received standardized data, and identifies abnormal data in the real-time data stream; the obtained prediction information and abnormal information are then input into the capacity decision and simulation module.

[0152] The capacity decision and simulation module includes a strategy center, a sandbox simulator, and a multi-objective decision maker.

[0153] In step S303, the capacity decision and simulation module determines the optimal capacity management strategy based on the predicted information and the anomaly information, and inputs the optimal capacity management strategy into the core AI intelligent engine.

[0154] It also includes a work order management module, an orchestrator, and an executor.

[0155] Work order management module: Used for operations that require manual confirmation due to high risk or policy requirements, to generate work orders and route them to the appropriate approvers.

[0156] Orchestrator: Used to translate the plan generated by the decision-maker into a series of executable atomic operation steps and build a directed acyclic graph to ensure execution order and dependencies.

[0157] Executor: Used to securely execute the DAG workflow provided by the orchestrator, completing the final resource operations by calling the APIs of various cloud platforms. Compliance checks are performed before execution, and the status is written back after execution.

[0158] In step S304, the core AI intelligent engine executes the optimal capacity management strategy and sends the execution result to the console. The console provides a human-computer interaction interface to display the global status, send alarms, and generate reports.

[0159] In step S305, the console sends feedback to the data lake storage based on the execution result.

[0160] The complete process of the capacity management method described in this embodiment of the invention is based on a complete data closed loop of "perception-decision-execution-feedback":

[0161] Step 1: Establishing a closed-loop data feedback mechanism:

[0162] 1. Input features (X_t): All data used to trigger this decision (historical performance, business metrics, events, etc.).

[0163] 2. Model Prediction (Y_pred): The AI ​​engine's prediction at that time (e.g., it is expected to require 100 vCPUs).

[0164] 3. Actual Observation (Y_true): The actual resource usage observed after the fact (e.g., 92 or 110 vCPUs were actually used).

[0165] 4. Decision Action (A): The specific operation taken by the system (such as scaling up a c5.2xlarge instance in AWS us-east-1).

[0166] 5. Decision Effect (R): A quantified reward signal, for example: Effect R = Cost Savings - α Performance deviation penalty. This R value is the core compass driving the system to evolve in a "better" direction.

[0167] Step Two: Triggering Model Updates: The system will not blindly and frequently retrain the model, but will trigger updates through intelligent judgment.

[0168] Periodic triggering: Perform gentle incremental learning at fixed intervals (such as weekly).

[0169] Event-driven triggering: Triggered immediately when the following conditions are detected: Prediction bias continues to exceed the standard: prediction error exceeds the threshold multiple times in a row; statistical tests (such as KS test) reveal a significant difference between the recent data distribution and the historical training data distribution, indicating that the business model may have undergone fundamental changes; after major events: such as large-scale promotions or application architecture reconstruction and launch.

[0170] Step 3: Incremental Learning and Model Evolution: This is the core algorithmic step in adaptive learning. The system uses incremental learning instead of full retraining.

[0171] For models such as CNN and LSTM, online learning algorithms are used or recent data (e.g., the past 3 months) is used periodically for fine-tuning with some historical data. This absorbs new knowledge while avoiding the complete forgetting of long-term patterns.

[0172] Evolutionary Result: Models become increasingly personalized. For example, a model learned for an "e-commerce front-end application" will become more adept at handling "promotional traffic spikes"; a model learned for an "internal reporting system" will more accurately grasp "month-end and quarter-end calculation peaks." The general-purpose predictor thus evolves into a dedicated business expert.

[0173] Step 4: Adaptive Adjustment of the Policy Base: Adaptive learning occurs not only in the AI ​​model but also at the rule and policy level.

[0174] Strategy effectiveness evaluation: The system records the execution history and results of each automated strategy (such as "expand capacity if CPU > 80%)" and calculates its success rate, cost-effectiveness and other indicators.

[0175] Strategy optimization and generation, including:

[0176] Parameter tuning: Automatically adjust policy thresholds (e.g., optimize the expansion threshold from 80% to 75%).

[0177] Strategy withdrawal: Reduce the weight or disable strategies with poor long-term performance or extremely low trigger frequency.

[0178] New strategy discovery: Based on historical success stories, we explore and recommend new, more efficient strategy rules through reinforcement learning.

[0179] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0180] This embodiment also provides a capacity management device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0181] Figure 4 This is a structural block diagram of a capacity management device according to an embodiment of the present invention, such as... Figure 4 As shown, the capacity management device 400 includes: a determination module 41, a prediction module 42, an identification module 43, and a management module 44.

[0182] The determination module 41 is used to determine a first feature tensor based on a preset time window, business event characteristics, and environmental characteristics; wherein the first feature tensor includes multiple data dimensions.

[0183] Prediction module 42 is used to input the first feature tensor into a hybrid neural network model to predict the resource demand at a specific future time point, and obtain prediction information corresponding to the resource demand at the specific future time point.

[0184] The identification module 43 is used to identify abnormal data in the real-time data stream based on an unsupervised learning algorithm and obtain abnormal information.

[0185] The management module 44 is used to determine the optimal capacity management strategy based on the predicted information and the anomaly information.

[0186] Figure 5 This is a structural block diagram of a capacity management device according to another embodiment of the present invention, such as... Figure 5 As shown, this capacity management device, in addition to Figure 4In addition to the included modules, it also includes: a collection module 45, used to collect various types of data from multiple heterogeneous cloud platforms; wherein the various types of data include at least one of the following: external environment data, performance data, configuration data, and cost data; a standardization module 46, used to standardize the various types of data and map the standardized data to a predefined metadata model to obtain metadata; and a storage module 47, used to store the metadata in a data lake.

[0187] In one embodiment, the acquisition module is further configured to obtain the business event features and the environmental features from various metadata in the data lake; and determine the first feature tensor based on the time window, the business event features, and the environmental features.

[0188] In one embodiment, the hybrid neural network model includes a convolutional neural network (CNN) model and a long short-term memory (LSTM) network model. The prediction module is further configured to input the first feature tensor into the CNN model to obtain a second feature tensor; wherein the second feature tensor has a higher level of abstraction than the first feature tensor; and input the second feature tensor into the LSTM model to obtain prediction information of resource demand corresponding to the future specific time point.

[0189] In one embodiment, the prediction module is further configured to input the second feature tensor into the LSTM model to perform the first layer prediction of the LSTM model to obtain resource demand prediction information for the application layer; and input the resource demand prediction information for the application layer into the LSTM model to perform the second layer prediction of the LSTM model to obtain the prediction information of the resource demand at a specific future time point.

[0190] In one embodiment, the container management device further includes an optimization module, configured to optimize the hybrid neural network model with the goal of minimizing the loss function of the business objective.

[0191] In one embodiment, the management module is further configured to: determine multiple capacity management strategies according to predefined business strategy rules; simulate the multiple capacity management strategies according to the prediction information and the anomaly information to obtain multiple simulation results; analyze the multiple simulation results according to a multi-objective optimization algorithm; and determine the optimal capacity management strategy from the multiple capacity management strategies based on the analysis results.

[0192] In one embodiment, the capacity management device further includes: an execution module, configured to obtain multiple atomic operation steps and construct a directed acyclic graph according to the optimal capacity management strategy; and execute the optimal capacity management strategy according to the atomic operation steps and the directed acyclic graph.

[0193] It should be noted that the above modules can be implemented by software or hardware. For the latter, they can be implemented in the following ways, but are not limited to: all the above modules are located in the same processor; or, the above modules are located in different processors in any combination.

[0194] Embodiments of the present invention also provide a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0195] Optionally, in this embodiment, the storage medium may be configured to store a computer program for performing the following steps:

[0196] S1, determine a first feature tensor based on a preset time window, business event characteristics, and environmental characteristics; wherein, the first feature tensor includes multiple data dimensions;

[0197] S2, input the first feature tensor into the hybrid neural network model to predict the resource demand at a specific future time point, and obtain the prediction information corresponding to the resource demand at the specific future time point;

[0198] S3, use an unsupervised learning algorithm to identify abnormal data in the real-time data stream and obtain abnormal information;

[0199] S4. Determine the optimal capacity management strategy based on the predicted information and the anomaly information.

[0200] Optionally, the storage medium is also configured to store a computer program for performing the following steps:

[0201] S1, collect various types of data from multiple heterogeneous cloud platforms; wherein, the various types of data include at least one of the following: external environment data, performance data, configuration data, and cost data;

[0202] S2, standardize the various types of data, and map the standardized data to a predefined metadata model to obtain metadata;

[0203] S3, store the metadata in the data lake.

[0204] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0205] Embodiments of the present invention also provide an electronic device including a memory and a processor, the memory storing a computer program and the processor being configured to run the computer program to perform the steps in any of the above method embodiments.

[0206] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0207] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0208] S1, determine a first feature tensor based on a preset time window, business event characteristics, and environmental characteristics; wherein, the first feature tensor includes multiple data dimensions;

[0209] S2, input the first feature tensor into the hybrid neural network model to predict the resource demand at a specific future time point, and obtain the prediction information corresponding to the resource demand at the specific future time point;

[0210] S3, use an unsupervised learning algorithm to identify abnormal data in the real-time data stream and obtain abnormal information;

[0211] S4. Determine the optimal capacity management strategy based on the predicted information and the anomaly information.

[0212] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0213] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0214] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A capacity management method, characterized in that, include: Based on a preset time window, business event characteristics, and environmental characteristics, a first feature tensor is determined; wherein, the first feature tensor includes multiple data dimensions; The first feature tensor is input into a hybrid neural network model to predict the resource demand at a specific future time point, thereby obtaining the predicted information of the resource demand at the specific future time point. Anomaly information is obtained by identifying abnormal data in real-time data streams using unsupervised learning algorithms. Based on the predicted information and the anomaly information, the optimal capacity management strategy is determined.

2. The method according to claim 1, characterized in that, Based on a preset time window, business event characteristics, and environmental characteristics, a multidimensional feature tensor is determined. Before determining the first feature tensor, the method further includes: Collect various types of data from multiple heterogeneous cloud platforms; wherein, the various types of data include at least one of the following: external environment data, performance data, configuration data, and cost data; The various types of data are standardized, and the standardized data is mapped to a predefined metadata model to obtain metadata; The metadata is stored in a data lake.

3. The method according to claim 2, characterized in that, Based on the preset time window, business event characteristics, and environmental characteristics, a multidimensional feature tensor is determined, including the first feature tensor, which includes: The business event characteristics and the environmental characteristics are obtained from various metadata in the data lake; The first feature tensor is determined based on the time window, the business event characteristics, and the environmental characteristics.

4. The method according to claim 1, characterized in that, in, The hybrid neural network model includes a convolutional neural network (CNN) model and a long short-term memory (LSTM) network model. The first feature tensor is input into the hybrid neural network model to predict resource demand at a specific future time point, obtaining predicted information corresponding to the resource demand at that specific future time point, including: The first feature tensor is input into the CNN model to obtain the second feature tensor; wherein the second feature tensor has a higher level of abstraction than the first feature tensor. The second feature tensor is input into the LSTM model to obtain the predicted information of resource demand corresponding to the specific future time point.

5. The method according to claim 4, characterized in that, The second feature tensor is input into the LSTM model to obtain prediction information of resource demand corresponding to the specific future time point, including: The second feature tensor is input into the LSTM model to perform the first layer prediction of the LSTM model, thereby obtaining the resource demand prediction information of the application layer. The resource demand prediction information of the application layer is input into the LSTM model to perform the second layer prediction of the LSTM, which is the prediction information of the resource demand at a specific future time point.

6. The method according to claim 1, characterized in that, The method further includes: The hybrid neural network model is optimized with the goal of minimizing the loss function of the business objective.

7. The method according to claim 1, characterized in that, Based on the predicted information and the anomaly information, the optimal capacity management strategy is determined, including: Based on predefined business strategy rules, multiple capacity management strategies are determined; Based on the predicted information and the anomaly information, the multiple capacity management strategies are simulated to obtain multiple simulation results; The multiple simulation results are analyzed using a multi-objective optimization algorithm, and the optimal capacity management strategy is determined from the multiple capacity management strategies based on the analysis results.

8. The method according to claim 1, characterized in that, After determining the optimal capacity management strategy based on the predicted information and the anomaly information, the method further includes: Based on the optimal capacity management strategy, obtain multiple atomic operation steps and construct a directed acyclic graph; The optimal capacity management strategy is executed based on the atomic operation steps and the directed acyclic graph.

9. A capacity management device, characterized in that, include: The determination module is used to determine a first feature tensor based on a preset time window, business event characteristics, and environmental characteristics; wherein the first feature tensor includes multiple data dimensions. The prediction module is used to input the first feature tensor into a hybrid neural network model to predict the resource demand at a specific future time point, and obtain prediction information corresponding to the resource demand at the specific future time point. The identification module is used to identify abnormal data in the real-time data stream based on an unsupervised learning algorithm and obtain abnormal information. The management module is used to determine the optimal capacity management strategy based on the predicted information and the anomaly information.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method described in any one of claims 1 to 8 when it is run.

11. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method as described in any one of claims 1 to 8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1 to 8.

Images