Intelligent education cloud platform elastic resource scheduling system based on big data

Abstract

The invention relates to the technical field of cloud computing and resource scheduling, and discloses a smart education cloud platform elastic resource scheduling system based on big data, which mainly comprises a data sensing module, a data lake bin module, an intelligent prediction module, a strategy generation module, a transition control module, an elastic execution module, a dynamic evaluation module and the like. The multi-source data is integrated through the data perception and lake and warehouse module, the information island is broken, the historical load, the user behavior and the service period are fused by using the intelligent prediction module to carry out accurate demand prediction, and the conversion from passive response to active pre-configuration is realized; the system automatically generates and safely executes an optimization scheduling strategy according to a prediction result, and then forms closed-loop optimization through dynamic evaluation, so that the resource utilization efficiency and the service stability are remarkably improved; the system can effectively cope with periodic and sudden loads in a cloud platform education scene, and provides intelligent, elastic and continuously optimized resource guarantee for high-concurrency and high-availability internet data services.

Description

Technical Field

[0001] This invention relates to the field of cloud computing and resource scheduling technology, and in particular to an elastic resource scheduling system for a smart education cloud platform based on big data. Background Technology

[0002] With the widespread application of smart education cloud platforms, their workloads exhibit significant periodicity and burstiness, such as concentrated access during the start of the school year and the concurrent pressure of large-scale online examinations. To ensure service stability and user experience, the platform must possess the ability to accurately predict and elastically schedule computing, storage, and network resources. Currently, utilizing big data and machine learning technologies to analyze historical and real-time data to predict resource demand has become an important research direction in this field. Among existing technologies, a common approach is resource prediction based on time series models (such as ARIMA and LSTM). This method primarily relies on time series data of historical monitoring indicators (such as CPU utilization and request QPS) to train models and predict short-term resource demand trends. Another approach combines simple business calendar information, such as manually or semi-automatically increasing resource reservations before and after known exam dates. These methods, to some extent, achieve a shift from completely passive response to preliminary prediction.

[0003] The aforementioned and existing related technologies often suffer from the following drawbacks: Although the platform aggregates a large amount of data, the data itself becomes isolated due to system fragmentation. The scheduling system lacks the ability to integrate, analyze, and intelligently predict data from multiple dimensions such as historical load, user behavior, and business cycles, thus failing to achieve the leap from "passive response" to "proactive pre-provisioning." Summary of the Invention

[0004] The technical problem to be solved by this invention is that the existing technology has the disadvantage that although the platform has a large amount of data, it forms data islands due to system fragmentation, and the scheduling system lacks multi-dimensional data fusion analysis and intelligent prediction capabilities, making it difficult to realize the transformation from "passive response" to "proactive pre-allocation". To this end, we propose an elastic resource scheduling system for a smart education cloud platform based on big data.

[0005] To achieve the above objectives, this application adopts the following technical solution: a smart education cloud platform elastic resource scheduling system based on big data, comprising:

[0006] The data sensing module is used to collect real-time and historical operational data from the teaching business platform, infrastructure and user terminals, and transmit the data to the data lake warehouse module.

[0007] The data lake warehouse module is used to receive and store raw data from the data sensing module, clean, integrate and structure the data to form a unified data asset, and provide standardized data to the intelligent prediction module.

[0008] The intelligent forecasting module is used to perform multi-dimensional time series analysis based on standardized data provided by the data lake warehouse module, and generate quantitative forecast results of resource demand within the forecast time window through machine learning algorithms.

[0009] The strategy generation module is used to automatically generate the corresponding optimal resource scheduling strategy instruction based on the quantitative prediction results output by the intelligent prediction module and in combination with preset optimization objectives and constraints.

[0010] The transition control module is used to perform risk assessment and step decomposition of resource scheduling instructions issued by the strategy generation module, and to plan and control the specific execution process of resource changes.

[0011] The elastic execution module is used to parse and convert the resource scheduling policy instructions issued by the policy generation module into specific control commands that the resource management module can recognize, thereby driving the underlying resources to complete automatic scaling and configuration changes.

[0012] The dynamic evaluation module is used to monitor the actual system performance indicators after the elastic execution module schedules, calculate the deviation between them and the expected goals, and feed the evaluation results back to the data lake warehouse module and the strategy generation module to form an optimization closed loop.

[0013] Preferably, the system also includes:

[0014] The resource management module is used to virtualize and integrate all heterogeneous physical resources at the bottom layer into a globally unified logical resource pool through a unified abstract model and standard interface, providing a resource operation basis for all upper-layer modules.

[0015] The anomaly handling module is used to monitor the system's operating indicators and abnormal signals from the data sensing module in real time. When a sudden failure or predicted external load is detected, it immediately triggers and executes a predefined emergency dispatch plan.

[0016] The global collaboration module provides system administrators with a unified visual interactive interface that includes a resource overview, prediction curves, scheduling status, and evaluation reports, and receives manual intervention instructions and configuration input from the strategy generation module.

[0017] Preferably, the resource management module includes:

[0018] The resource abstraction module is used to virtualize computing devices, storage devices and network devices in heterogeneous physical resources into logical resource objects with consistent attributes through a preset unified abstraction model and standard interface, thereby forming a logical resource pool that is uniformly identified and managed by the upper layer system.

[0019] The profile management module is used to create and maintain a dynamic resource profile for each logical resource object generated by the resource abstraction module. The profile is dynamically updated by collecting static configuration information, real-time performance indicators and historical load records of the resource, thereby digitally encapsulating the inherent attributes, real-time status and historical performance of the resource.

[0020] The supply execution module receives resource supply intent descriptions from the strategy generation module, matches, filters, and combines them in all resource profiles maintained by the profile management module according to the constraints and optimization objectives in the intent, and finally issues specific resource allocation, configuration change, or recycling instructions to the underlying infrastructure to complete the actual supply of physical resources.

[0021] Preferably, the intelligent prediction module specifically includes:

[0022] The trend forecasting module is used to receive and analyze historical load data spanning one semester or one academic year, and calculate the overall baseline of resource demand changes through a long-term time series model, providing a macro-level and trend-based framework input for subsequent forecasts.

[0023] The cycle prediction module is used to receive the trend baseline output by the trend prediction module and focus on recent historical data with weekly and daily granularity. By identifying fixed patterns in course arrangements and teaching activities, it can accurately predict the periodic load peaks and troughs that will recur in the next few days.

[0024] The event correction module is used to monitor announcements and event streams from the academic affairs system in real time, proactively identify unplanned teaching activities, and generate correction signals in real time based on these activities. It dynamically overlays and adjusts the composite prediction results jointly generated by the trend prediction module and the cycle prediction module to output a final accurate resource demand prediction sequence that includes sudden loads.

[0025] Preferably, the intelligent forecasting module performs the following calculation and processing steps for forecasting resource demand within the forecast time window:

[0026] S1. The trend prediction module applies a time series decomposition algorithm to the input historical resource utilization time series to separate the trend component series, and uses an autoregressive integral moving average model to fit and predict the trend component series, and outputs the long-term trend baseline value series corresponding to the prediction time window.

[0027] S2. The periodic prediction module applies a seasonal decomposition algorithm to the same historical resource utilization time series, extracts the seasonal component series with a basic period of seven days or twenty-four hours, and uses a seasonal autoregressive integral moving average model to fit and predict the seasonal component series, outputting the periodic fluctuation value series corresponding to the prediction time window.

[0028] S3. The event correction module parses the external structured event data stream. When it identifies an unplanned event of a predefined type, it immediately queries the preset resource consumption coefficient table according to the event type code, and performs multiplication calculation according to the event scale parameter to obtain the incremental value of resource demand caused by the event. Then, it maps the event occurrence time to the prediction time window to generate an event incremental value sequence.

[0029] S4. The prediction engine of the intelligent prediction module aligns the long-term trend baseline value sequence, the periodic fluctuation value sequence, and the event increment value sequence in the time dimension, and performs point-by-point addition to obtain the final resource demand prediction value sequence. At the same time, based on the variance of the historical prediction error of each sequence, the upper and lower bounds of the prediction interval for each point in the final resource demand prediction value sequence are calculated using the error propagation law.

[0030] Preferably, the intelligent forecasting module calculates resource demand forecasting using the following formula:

[0031] ;

[0032] In the formula: R represents the resource demand forecast; i represents the hour number of the day; T represents the long-term trend baseline value; C i W represents the periodic amplitude in the i-th hour; i γ represents the weight of the i-th hour; γ represents the time period decay coefficient; E represents the incremental event base resources; K represents the event scale coefficient; σ 2 Z represents the variance of the historical comprehensive prediction error; Z represents the confidence coefficient; and N represents the total number of historical data samples.

[0033] Preferably, the dynamic evaluation module includes:

[0034] The real-time indicator acquisition module is used to collect raw performance indicator data, including resource utilization, service response time and transaction success rate, from the resource management module, elastic execution module and business application interface, and send the data to the deviation quantification calculation module in time series form.

[0035] The deviation quantification calculation module is used to receive the performance index time series sent by the real-time index acquisition module, and compare it point by point with the expected target value output by the strategy generation module or the prediction benchmark value generated by the intelligent prediction module, and output the specific quantified deviation value of each index at each moment through a preset difference calculation function.

[0036] The performance compliance determination module is used to receive the quantified deviation value output by the deviation quantification calculation module, and automatically determine whether the overall and local service levels of the current system meet the agreed requirements based on the preset performance compliance thresholds at each level, while identifying specific resource entities or service components that do not meet the agreement.

[0037] The closed-loop feedback generation module is used to receive the judgment result output by the performance compliance judgment module, convert it into specific optimization parameter adjustment instructions or strategy correction suggestions according to the predefined feedback logic, and send the instructions and suggestions to the intelligent prediction module and the strategy generation module respectively to drive the iterative update of the prediction model and the scheduling strategy.

[0038] Preferably, the deviation quantification calculation module calculates the comprehensive quantification deviation value of the output index based on the difference calculation function as follows:

[0039] ;

[0040] In the formula: M represents the quantization deviation value; x represents the performance index number; n represents the total number of performance indices involved in the calculation; ω x A represents the weight of the x-th category of indicators; x D represents the real-time actual value of the x-th category indicator; x ε represents the preset target value of the x-th type of indicator; α represents the bias sensitivity coefficient; d represents the time decay factor; t represents the current evaluation time step; t x k represents the time step for collecting data of the x-th category of indicators; x Indicates the coefficient of the indicator type.

[0041] Preferably, the feedback logic of the closed-loop feedback generation module is dynamically adjusted according to the deviation level. When the quantified deviation value is lower than the first-level threshold, only the model parameter fine-tuning instruction is output to the intelligent prediction module. When the deviation value is between the first-level and second-level thresholds, parameter adjustment and strategy optimization suggestions are output to both the intelligent prediction module and the strategy generation module. When the deviation value is higher than the second-level threshold, in addition to outputting optimization instructions, an alarm notification from the global collaboration module is triggered simultaneously to remind the administrator to intervene and verify. The execution effect of the feedback instructions is verified twice by the dynamic evaluation module to form a complete feedback closed-loop record.

[0042] Preferably, the strategy generation module includes a strategy knowledge base and a dynamic optimizer; the strategy knowledge base stores and maintains preset scheduling strategy templates for different educational scenarios, including live teaching, online examinations, and scientific research computing; the dynamic optimizer retrieves matching strategy templates from the strategy knowledge base and performs online parameter calculation and multi-objective optimization based on the specific prediction sequence output by the intelligent prediction module and the real-time resource status and cost information fed back by the resource management module, and generates a specific resource scheduling instruction sequence to be executed within a specified time window.

[0043] The technical effects and advantages of this invention are as follows:

[0044] In this invention, the collaboration between the data perception module and the data lake warehouse module effectively breaks down data barriers between teaching operations, infrastructure, and user terminals, integrating previously scattered, multi-source, heterogeneous data into high-quality, unified data assets, laying a solid foundation for subsequent analysis. Based on this, the intelligent prediction module integrates multi-dimensional data such as historical load, user behavior, and business cycles, and uses machine learning algorithms to perform accurate time-series predictions, achieving a fundamental shift in resource demand from "passive response" to "proactive pre-allocation." The strategy generation and elastic execution module can automatically generate and execute optimized scheduling strategies based on prediction results, significantly improving resource utilization efficiency and system response agility. Combined with the closed-loop feedback mechanism formed by the dynamic evaluation module, the system can continuously optimize prediction accuracy and scheduling effectiveness. Overall, this system not only significantly enhances the stability and service quality of the smart education cloud platform when facing periodic and sudden loads, but also provides efficient, intelligent, and adaptively adjustable resource scheduling guarantees for supporting large-scale, high-concurrency Internet data services. Attached Figure Description

[0045] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts:

[0046] Figure 1 This is an overall architecture diagram of a smart education cloud platform elastic resource scheduling system based on big data, according to the present invention.

[0047] Figure 2 This is a schematic diagram illustrating the overall working principle of an elastic resource scheduling system for a smart education cloud platform based on big data, according to the present invention.

[0048] Figure 3 This is an architecture diagram of the intelligent prediction module of a smart education cloud platform elastic resource scheduling system based on big data, according to the present invention.

[0049] Figure 4This is a resource management module architecture diagram of an elastic resource scheduling system for a smart education cloud platform based on big data, according to the present invention. Detailed Implementation

[0050] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0051] Reference Figures 1-2 As shown, the present invention provides a technical solution: an elastic resource scheduling system for a smart education cloud platform based on big data. The system mainly includes a data perception module, a data lake warehouse module, an intelligent prediction module, a strategy generation module, a transition control module, an elastic execution module, and a dynamic evaluation module.

[0052] Data Sensing Module: As the system's information entry point, the data sensing module is responsible for continuously and in real-time collecting operational data from various levels of the smart education cloud platform. These data sources include:

[0053] 1. Teaching Business Platform: Collects business indicator data such as concurrent access volume, number of online users, course viewing time, exam submission volume, and file download and upload volume for teaching applications.

[0054] 2. Infrastructure: Collect performance metrics of underlying computing (CPU utilization, memory utilization), storage (I / O throughput, storage capacity), network (bandwidth utilization, latency), and other virtualized or physical resources.

[0055] 3. User terminals: Collect information such as the geographical distribution of user access, terminal type, and network quality to provide a more comprehensive context for prediction and scheduling.

[0056] The data awareness module will use various APIs, agents, log parsing and other technologies to transmit the collected real-time and historical operational data to the data lake warehouse module at high frequency (e.g., second-level or minute-level) and in a historical batch processing manner.

[0057] Data Lake Warehouse Module: As the system's data hub, the data lake warehouse module is responsible for receiving and storing various types of raw data from the data sensing module. Its core functions include:

[0058] 1. Raw data storage: Supports massive, multi-source, and heterogeneous raw data storage, such as using HDFS, object storage and other technologies.

[0059] 2. Data cleaning and integration: Perform operations such as deduplication, format conversion, missing value imputation, and outlier handling on the original data to ensure data quality.

[0060] 3. Structured governance: Model and label the cleaned data to form a unified data asset, such as building a dimensional model or star schema, to facilitate subsequent querying and analysis.

[0061] 4. Standardized data provision: Provide a unified API interface or data service to external parties, output processed and standardized data to the intelligent prediction module, and ensure that the historical data input into the prediction model is of high quality and consistency.

[0062] Intelligent Prediction Module: This module is the core intelligent component of the system's elastic scheduling. Based on standardized historical data (such as historical resource utilization and concurrent user counts) provided by the data lake warehouse module, the intelligent prediction module uses machine learning algorithms (such as deep learning and time series analysis models) to perform multi-dimensional time-series analysis. For example, it considers multiple dimensions such as course type, semester cycle, exam schedule, and user behavior patterns to generate quantitative predictions of the system's resource requirements within a future prediction window (e.g., the next 30 minutes, 1 hour, or 1 day). The prediction results include specific indicators such as the number of CPU cores, memory GB, disk IOPS, and network bandwidth Mbps.

[0063] The strategy generation module receives the quantitative forecasts of future resource demands from the intelligent prediction module and combines them with preset optimization objectives (e.g., maximizing resource utilization, minimizing costs, and ensuring SLA service level agreements) and constraints (e.g., resource pool capacity limits, system disaster recovery requirements, and business priorities) to automatically generate corresponding optimal resource scheduling policy instructions. These instructions may include: adding / reducing a specified number of virtual machine / container instances, adjusting instance configurations (CPU / memory), modifying network policies, and expanding storage capacity. The strategy generation module uses heuristic algorithms, linear programming, or reinforcement learning to balance multiple objectives and find the optimal solution.

[0064] The strategy generation module includes a strategy knowledge base and a dynamic optimizer.

[0065] Strategy Knowledge Base: This module stores and maintains preset scheduling strategy templates for different educational scenarios within the smart education cloud platform. These templates are best practices summarized based on expert experience, historical scheduling data, and typical scenario requirements. Different educational scenarios have different resource demand patterns and SLA requirements.

[0066] Live teaching requires high concurrency, low latency network bandwidth and computing resources, and has high requirements for instantaneous expansion capabilities. The strategy template may focus on fast cold start and elastic scaling.

[0067] Online exams: have extremely high requirements for system stability and data security, with a concentrated number of concurrent users and a fixed duration. Policy templates may focus on resource reservation, fault isolation, and preventative scaling.

[0068] Scientific computing: This may involve long-running, resource-intensive tasks (such as GPU computing), which have high requirements for the management and cost optimization of heterogeneous computing resources. The strategy template may focus on batch scheduling, task prioritization, and cost control.

[0069] Each template in the strategy knowledge base not only contains general rules for resource configuration, but may also include scheduling logic, optimization target weights, elastic scaling thresholds, etc., for specific businesses.

[0070] Dynamic Optimizer: This module is the core decision engine of the strategy generation module. It receives specific prediction sequences (including the demand and prediction range of various resources) from the intelligent prediction module, as well as real-time resource status and cost information (such as current available resource capacity, real-time cost of various types of resources, and resource health status) from the resource management module.

[0071] Based on this input data, the dynamic optimizer retrieves the policy template that best matches the current prediction scenario from the policy knowledge base. For example, if it predicts a large-scale online exam tomorrow morning, it will select the "online exam" policy template.

[0072] Subsequently, the dynamic optimizer combines real-time prediction data and resource status information to perform online parameter calculations and multi-objective optimization solutions for the selected template. These multi-objective optimizations may include:

[0073] 1. Maximize performance: Ensure that service response time, throughput and other indicators meet expectations.

[0074] 2. Minimize costs: Under the premise of meeting performance requirements, choose the lowest cost resources possible.

[0075] 3. Maximize resource utilization: Reduce resource idleness.

[0076] 4. Risk Minimization: Considering the uncertainty of prediction, a certain amount of redundant resources is reserved. The dynamic optimizer uses techniques such as genetic algorithms, particle swarm optimization, and reinforcement learning to weigh the trade-offs among these conflicting objectives, ultimately generating a specific sequence of resource scheduling instructions to be executed within a specified time window. This instruction sequence is highly granular, for example, "expand 10 C4.large instances in region A at time T1, adjust the load balancing weight to the new instances at time T2, and release 5 C4.medium instances at time T3," and is finally passed to the transition control module for risk assessment and execution.

[0077] Transition control module: To ensure the safety and smoothness of resource scheduling, the transition control module intervenes before the resource scheduling instructions issued by the policy generation module are executed.

[0078] 1. Risk Assessment: Conduct a potential risk assessment for the scheduling instructions to be executed, such as determining whether the operation may cause service interruption, performance fluctuations, or resource exhaustion.

[0079] 2. Step decomposition: Decompose a complex scheduling instruction into a series of controllable, atomic execution steps, and define the execution order, dependencies and rollback mechanism of each step.

[0080] 3. Planning and Control: Plan the specific execution path for resource changes, such as expanding and then shrinking capacity first, or making changes in batches, and control the execution progress in real time to ensure a smooth scheduling process and avoid impacting existing business.

[0081] Elastic Execution Module: The elastic execution module is responsible for parsing and converting the resource scheduling policy instructions issued by the transition control module or policy generation module into specific control commands that the resource management module can recognize. These commands are typically operation requests compatible with the underlying infrastructure APIs, such as calling cloud platform APIs to create virtual machines, container orchestration tool APIs to scale up, storage management APIs to adjust storage volume sizes, and network controller APIs to configure load balancing, thereby driving the underlying resources to complete automated scaling and configuration changes, ultimately turning predicted resource demand into actual resource supply.

[0082] Dynamic Evaluation Module: The dynamic evaluation module is crucial for achieving closed-loop optimization of the system. It continuously monitors the actual system performance metrics after scheduling by the elastic execution module, including resource utilization, application response time, error rate, and throughput. By comparing these actual metrics with the prediction targets of the intelligent prediction module or the expected results of the strategy generation module, it calculates the deviation. For example, if the actual CPU utilization consistently exceeds the predicted value, it indicates a potential bias in the prediction model. The evaluation results, including deviation status and performance satisfaction, are fed back to the data lake warehouse module for updating historical data, to the strategy generation module for adjusting subsequent scheduling strategy parameters and optimization targets, and even to the intelligent prediction module for model parameter tuning or model retraining, thus forming a continuous learning and self-optimization closed loop.

[0083] Combination Figure 1 and Figure 2 As shown, the system of the present invention further includes a resource management module, an anomaly handling module, and a global collaboration module.

[0084] Resource Management Module: This module virtualizes and integrates all underlying heterogeneous physical resources (such as servers, storage devices, network devices from different vendors, or hybrid cloud resources in different data centers) through a unified abstract model and standard interfaces, forming a globally unified logical resource pool. The resource management module hides the complexity and heterogeneity of the underlying resources, providing a unified and standardized resource operation foundation for all upper-layer modules (especially the policy generation module and elastic execution module). It can provide abstract operations such as "creating a general-purpose computing instance with N CPU cores and MGB of memory" without needing to know whether the underlying system is a VMware virtual machine, a Kubernetes container, or a physical server.

[0085] Anomaly Handling Module: This module is responsible for ensuring system robustness. It monitors system operational metrics in real time (e.g., collected via the data awareness module or directly from the infrastructure) and anomaly signals transmitted by the data awareness module (e.g., numerous errors in logs, network connection loss, etc.). When a sudden failure (e.g., server crash, network link interruption) or predicted external load (e.g., a large influx of users due to a popular external event that was not predicted by the intelligent prediction module) is detected, the anomaly handling module immediately triggers and executes a predefined emergency response plan. The emergency response plan may include: rapidly migrating services from the failed instance to a healthy instance, activating backup resources, and triggering rate limiting and degradation to minimize system downtime and ensure the availability of core services.

[0086] Global Collaboration Module: The Global Collaboration Module serves as the interface between the system and the administrator, providing the system administrator with a unified visual interface that includes a resource overview, prediction curves, scheduling status, and evaluation reports.

[0087] 1. Resource Panorama: Displays the topology, real-time status, health status, and utilization rate of all current logical and physical resources.

[0088] 2. Prediction Curve: Visualizes the future resource demand prediction curve output by the intelligent prediction module, including the prediction range.

[0089] 3. Scheduling Status: Displays currently executing scheduling tasks, completed scheduling history, scheduling results, etc.

[0090] 4. Evaluation Report: Displays the system performance indicators, deviation analysis, and optimization suggestions output by the dynamic evaluation module.

[0091] In addition, the global collaboration module also receives manual intervention commands from administrators (such as manually triggering expansion, forcing rollback policies, adjusting scheduling priorities, etc.) and configuration inputs from the policy generation module (such as adjusting optimization goals, modifying constraints, and uploading new policy templates). These manual interventions and configurations are passed to the corresponding modules through the global collaboration module to adapt to special situations or specific needs of administrators.

[0092] Based on the above embodiments, the resource management module of the present invention specifically includes a resource abstraction module, a profile management module, and a supply execution module, the internal structure of which is as follows: Figures 1-2 and Figure 4 As shown.

[0093] Resource Abstraction Module: This module virtualizes underlying heterogeneous physical resources through a pre-defined unified abstraction model and standard interfaces. It can specifically virtualize computing devices, storage devices, and network devices from different vendors and technology stacks (e.g., x86 servers, ARM servers, GPU cards, NVMe storage, SDN network devices) into logical resource objects with consistent attributes. For example, different CPU models are abstracted into "number of computing cores" and "clock speed," different storage media are abstracted into "storage capacity" and "IOPS guarantee level," and different network switches are abstracted into "bandwidth" and "latency." Through this abstraction, a logical resource pool is formed that is uniformly identified and managed by the upper-layer system, greatly simplifying the upper-layer modules' understanding and operation of resources.

[0094] Profile Management Module: This module creates and maintains a dynamic resource profile for each logical resource object generated by the resource abstraction module. This profile is a digital representation of the resource, dynamically encapsulating the resource's inherent attributes, real-time status, and historical performance, specifically including:

[0095] Static configuration information: such as the virtual machine's operating system type, CPU model, memory size, hard disk capacity, network interface configuration, etc.

[0096] Real-time performance metrics: Real-time data such as CPU utilization, memory usage, disk IOPS, network traffic, and process count are continuously collected by the underlying monitoring agent.

[0097] Historical load records: These records store the resource's load status, performance baseline, and failure history over different time periods, used to assess the resource's reliability and load capacity. This information is continuously updated through the data awareness module or direct monitoring interfaces, ensuring that the resource profile always reflects the latest resource status and providing an accurate basis for intelligent decision-making by the supply execution module.

[0098] The Supply Execution Module receives resource supply intent descriptions from the Policy Generation Module. These intents typically include constraints and optimization objectives such as the required resource type, quantity, performance requirements (e.g., minimum IOPS, maximum latency), and cost preferences. Based on these intents, the Supply Execution Module intelligently matches, filters, and combines resources from all resource profiles maintained by the Profile Management Module. For example, it intelligently selects the most suitable physical resources based on factors such as the actual load, health status, and cost-effectiveness of the current resource pool. Ultimately, it issues specific resource allocation (e.g., creating virtual machines or containers with specified configurations), configuration change (e.g., adjusting the CPU / memory of virtual machines, mounting new storage volumes), or reclamation instructions (e.g., releasing unused instances) to the underlying infrastructure to complete the actual supply or reclamation of physical resources, thereby achieving refined resource management and efficient utilization.

[0099] Reference Figures 1-3 As shown in this implementation plan, the intelligent prediction module specifically includes a trend prediction module, a cycle prediction module, an event correction module, and a prediction engine.

[0100] Trend Forecasting Module: This module is specifically designed to analyze and forecast long-term trends in resource demand. It receives and analyzes historical workload data (provided by the data lake warehouse module) spanning typically a semester, academic year, or even longer, such as the average online user growth trend over the past few years and seasonal variations in total course visits. By employing long-term time series models (such as ARIMA, Prophet, and exponential smoothing), the Trend Forecasting module can calculate an overall baseline for resource demand changes. This baseline reflects macroscopic, slowly changing trends, providing crucial macroscopic framework input for subsequent forecasting.

[0101] Cyclical Prediction Module: This module focuses on the cyclical fluctuations in resource demand. It receives the trend baseline from the trend prediction module as a reference and primarily analyzes recent historical data at weekly and daily granularities. By identifying patterns such as fixed course schedules, teaching activities (e.g., weekly live classes at fixed times, daily morning and evening user activity peaks), and rest day effects within the smart education platform, the cyclical prediction module uses seasonal time series models (such as SARIMA) to accurately predict recurring cyclical load peaks and troughs over the next few days. For example, it predicts that concurrent user numbers will peak on Monday mornings, while load will be relatively lower on Saturday and Sunday afternoons.

[0102] Event Correction Module: This module handles non-periodic, sudden changes in resource demand. It monitors event streams from the academic affairs system in real time, including announcements, news, schedules, and system notifications, proactively identifying unplanned teaching activities (such as temporary make-up classes, urgent online meetings, special lectures, sudden school-wide system maintenance, and additional mock tests for large-scale exams). Based on the identified event type, scale, and warning time, the event correction module instantly generates correction signals, such as a positive or negative increase in resource demand. This increase is superimposed on the composite prediction result generated jointly by the trend prediction module and the periodic prediction module, dynamically overlaying and adjusting the prediction results to output a final accurate resource demand prediction sequence that includes sudden loads.

[0103] Prediction Engine: The prediction engine is the core coordinator of the intelligent prediction module. It integrates and processes the outputs of the trend prediction module, the periodic prediction module, and the event correction module. Specifically, it aligns the long-term trend baseline value sequence, the periodic fluctuation value sequence, and the event increment value sequence in the time dimension (using timestamps), and then performs point-by-point addition to generate the final resource demand prediction value sequence. In addition, based on the historical prediction error statistics of each submodule (e.g., mean squared error, variance), the prediction engine uses the error propagation law to calculate the upper and lower bounds of the prediction interval for each point in the final resource demand prediction value sequence, providing confidence information of the prediction results and providing risk reference for subsequent strategy generation.

[0104] The intelligent forecasting module performs the following calculation and processing steps for predicting resource demand within the forecast time window:

[0105] S1. Extraction and Prediction of Trend Components: The trend prediction module applies a time series decomposition algorithm, such as the classic additive or multiplicative model, to the input historical resource utilization time series (e.g., average daily CPU utilization over the past three years, or average weekly concurrent connections), to decompose the original time series into trend components, seasonal components, and residual components. Here, particular attention is paid to the trend component series. Next, the trend prediction module uses an Autoregressive Integral Moving Average (ARIMA) model (with specific parameters adaptively set based on historical data or empirically) to fit and predict this trend component series. The fitting process determines the model parameters by minimizing the prediction error. After prediction, the module outputs the long-term trend baseline value series corresponding to the prediction time window (e.g., the next month). This series represents the macro-level development direction of resource demand over time.

[0106] S2. Extraction and Prediction of Periodic Components: The periodic prediction module applies a seasonal decomposition algorithm (e.g., STL decomposition or X-13ARIMA-SEATS) to the same historical resource utilization time series to extract seasonal component sequences with a basic period of seven days (weekly cycle) or twenty-four hours (daily cycle) from the original sequence. If the data exhibits both weekly and daily periodicity, a multiple seasonal time series decomposition method can be used. Next, the periodic prediction module uses a seasonal autoregressive integral moving average (SARIMA) model (specific parameters P, D, Q, S are set according to the periodicity characteristics of historical data and the model's fit) to fit and predict this seasonal component sequence, outputting a periodic fluctuation value sequence corresponding to the prediction time window. This sequence captures the regular changes in resource demand caused by fixed periodic events (such as weekday / weekend, daytime / nighttime teaching activities).

[0107] S3. Event Incremental Correction: The event correction module continuously parses external structured event data streams, which may come from the academic affairs management system, the school's official website, news announcements, etc. This data includes information such as event type (e.g., final exam, new student enrollment, large online lecture), event scale (e.g., number of participants, duration), and event occurrence time. When a predefined type of unplanned event is identified (e.g., a large online open course that doesn't usually occur), the event correction module immediately queries a pre-defined resource consumption coefficient table based on the event type code (e.g., code A for "final exam," code B for "large lecture"). The resource consumption coefficient table stores the unit impact of different types of events on various resources (CPU, memory, network bandwidth, etc.). For example, for an "online exam" event, 1000 examinees may require an additional 50 CPU cores and 200GB of memory. The event correction module performs a multiplication calculation based on the event scale parameter (e.g., if the number of examinees is 10,000, the resource increment is 50 * (10000 / 1000) = 500 CPU cores) to obtain the incremental resource demand caused by the event. Then, the incremental value is mapped to the prediction time window according to the event occurrence time to generate an event incremental value sequence.

[0108] S4. Generation of Final Prediction Results and Calculation of Confidence Intervals: The prediction engine of the intelligent prediction module precisely aligns the long-term trend baseline value sequence generated in step S1, the periodic fluctuation value sequence generated in step S2, and the event increment value sequence generated in step S3 along the time dimension (e.g., accurate to each hour or minute). Then, it performs point-by-point addition, that is, adds the values ​​of the three sequences at each time point to obtain the final resource demand prediction value sequence. This sequence is the final prediction result that integrates long-term trends, periodic patterns, and the impact of sudden events.

[0109] Meanwhile, to provide reliable predictions, the prediction engine will calculate the upper and lower bounds of the prediction interval for each point in the final resource demand prediction sequence based on the variance (or standard deviation) of the historical prediction errors of each sub-module (trend prediction, cycle prediction, event correction), combined with the overall historical prediction error statistics of the system, and using the error propagation law (for example, if each error is independent, then the total error variance is equal to the sum of the variances of each error).

[0110] Furthermore, the intelligent forecasting module calculates resource demand forecasts using the following formula:

[0111] ;

[0112] In the formula:

[0113] R represents the predicted resource demand for the i-th hour of the day. Here, i is a time index from 1 to 24. For example, when predicting the number of CPU cores, R is the predicted number of CPU cores.

[0114] T represents the long-term trend baseline value, which is usually predicted by the trend forecasting module using models such as ARIMA to fit historical long-term data. It reflects the macro-growth or decline trend of resource demand.

[0115] C i This represents the periodic amplitude of the i-th hour, which is identified and predicted from historical periodic data by the periodic prediction module using models such as SARIMA. It represents the periodic fluctuation of this time point compared to the average value within a day or week.

[0116] W i This represents the weight of the i-th hour. It is a coefficient used to adjust the periodic amplitude and can be preset according to the importance of different time periods based on the actual business scenario. For example, higher weight can be assigned to peak teaching periods.

[0117] γ represents the time-lapse coefficient, used to soften periodic effects, especially when periodic fluctuations may weaken or become less significant over time. This coefficient is typically between 0 and 1.

[0118] E represents the basic resource increment of an event, which is obtained by the event correction module by querying the resource consumption coefficient table according to the event type code. It is the basic impact of a single event on resource demand.

[0119] K represents the event scale coefficient, which is used to quantify the impact of the event scale (such as the number of participants, duration, etc.) on the incremental resource demand. It is usually obtained by multiplying the event scale (such as the ratio of the actual number of participants to the basic parameters) with the basic increment.

[0120] Z represents the confidence coefficient, which is usually a constant related to the confidence level. For example, in a normal distribution, the Z value for a 95% confidence interval is approximately 1.96.

[0121] σ represents the standard deviation of the historical aggregate forecast error. It is a comprehensive measure of the forecast errors of all sub-modules and is used to assess the uncertainty of the forecast. It is likely calculated by combining the error propagation law with the variance of the errors of each module. 2 This represents the variance of the historical comprehensive forecast error.

[0122] N represents the total number of historical data samples. In the square root term, it is used to adjust for prediction uncertainty. Statistically, the more samples there are, the smaller the variance of the prediction, meaning the prediction is more stable.

[0123] By integrating long-term trend baselines, adjusting for cyclical fluctuations, and incorporating incremental data from unforeseen events, and by introducing confidence interval calculations, this method achieves accurate quantitative predictions of future resource demands. First, macroeconomic trends and cyclical patterns are extracted from historical data. Then, real-time event information is used for dynamic correction. Finally, the prediction uncertainty is assessed using the error propagation law, resulting in a complete prediction sequence that combines point prediction and interval estimation. This enhances the relevance and reliability of resource prediction on the education cloud platform, effectively responding to load fluctuations caused by semester cycles, course schedules, and unforeseen teaching events, providing accurate and reliable data for elastic scheduling.

[0124] Reference Figures 1-2 The dynamic evaluation module of this invention specifically includes a real-time indicator acquisition module, a deviation quantification calculation module, an efficiency compliance judgment module, and a closed-loop feedback generation module.

[0125] Real-time performance indicator acquisition module: This module serves as the input layer for the dynamic evaluation module, continuously collecting raw performance indicator data from various key points in system operation to ensure the real-time nature and accuracy of the evaluation. These data sources may include:

[0126] Resource management module: Collects real-time utilization rates (such as CPU utilization, memory usage, storage I / O), and remaining available capacity of various underlying logical resources.

[0127] Elastic execution module: Collects the success rate, time taken, and actual effective time of each scaling operation.

[0128] Business application interface: Collects key performance indicators such as business service response time, transaction success rate, concurrent users, and error logs obtained through APM (Application Performance Management) tools. This data is sent to the deviation quantification calculation module in time series format (with timestamps) to provide a basis for subsequent deviation analysis.

[0129] Deviation Quantification Calculation Module: This module receives the performance indicator time series sent by the real-time indicator acquisition module. It compares these actual performance indicators with the expected target values ​​output by the strategy generation module (e.g., CPU utilization not exceeding 80%, response time less than 100ms) or the baseline predictions generated by the intelligent prediction module (e.g., predicted concurrent connections). Using a preset difference calculation function (specific formulas detailed in Example 7, such as weighted deviation sum, percentage deviation, relative error, etc.), it outputs the specific quantified deviation value for each indicator at each moment. Positive values ​​may indicate that the actual value is higher than expected, negative values ​​indicate that the actual value is lower than expected, and the absolute value indicates the degree of deviation.

[0130] Performance Compliance Determination Module: This module receives the quantified deviation value output by the deviation quantification calculation module. Based on preset performance compliance thresholds at various levels (e.g., Level 1 threshold for slight deviation, Level 2 threshold for severe deviation), it automatically determines whether the overall and partial service levels of the current system meet the agreed requirements. For example, if the response time deviation of core business exceeds the preset "service quality degradation" threshold, or the utilization rate of critical resources is consistently higher than the "overload" threshold, it is determined to be non-compliant. Simultaneously, this module can identify specific non-compliant resource entities (such as a specific virtual machine or storage volume) or service components (such as a microservice or database instance), providing precise location for subsequent feedback and optimization.

[0131] Closed-loop feedback generation module: This module receives the judgment results output by the performance compliance judgment module. Based on predefined feedback logic (e.g., according to deviation level and type), it converts them into specific optimization parameter adjustment instructions or strategy correction suggestions.

[0132] Parameter adjustment instructions: For the intelligent prediction module, instructions such as "adjust the prediction model parameters to reduce overestimation bias" or "retrain the model" may be issued.

[0133] Strategy Revision Suggestions: The strategy generation module may issue suggestions such as "adjust threshold strategy," "increase buffer capacity," or "update resource optimization target weights." These instructions and suggestions are sent to the intelligent prediction module and the strategy generation module, respectively, to drive iterative updates of the prediction model and scheduling strategy. For example, if the evaluation results show that the system frequently triggers capacity expansion but resource utilization remains low, the strategy generation module may be advised to adjust its optimization target, prioritizing resource cost over solely pursuing performance.

[0134] The feedback logic of the closed-loop feedback generation module of this invention is dynamically adjusted according to the deviation level, as follows:

[0135] 1. Handling Minor Deviations (Below the First-Level Threshold): When the quantitative deviation value M detected by the performance compliance assessment module is lower than the preset "first-level threshold" (e.g., the overall service quality fluctuates only slightly and does not significantly affect the user experience), the closed-loop feedback generation module determines it as a minor deviation. In this case, the system only outputs model parameter fine-tuning instructions to the intelligent prediction module. For example, it may instruct the intelligent prediction module to fine-tune the parameters of its ARIMA or SARIMA model, or update the weights in the machine learning model to better adapt to recent data trends, thereby improving prediction accuracy. This feedback is gentle and internally optimized, aiming to continuously improve the adaptive capability of the prediction model and avoid over-adjustment that could cause system oscillations.

[0136] 2. Moderate Deviation Handling (between Level 1 and Level 2 Thresholds): When the quantified deviation value M falls between the "Level 1 Threshold" and the "Level 2 Threshold" (for example, performance degradation in some non-critical services, or slightly higher-than-expected utilization of critical resources), it is determined to be a moderate deviation. At this time, the closed-loop feedback generation module simultaneously outputs parameter adjustment instructions to the intelligent prediction module and strategy optimization suggestions to the strategy generation module.

[0137] Intelligent prediction module: In addition to fine-tuning model parameters, it may also suggest local model retraining or adjust the data collection frequency of the data perception module to obtain more timely data.

[0138] The policy generation module may suggest adjustments to the optimization objective weights in the scheduling policy (e.g., rebalancing cost and performance), modifications to the trigger threshold for elastic scaling, increasing the step size for resource expansion, or exploring new scheduling algorithms. These suggestions aim to better address identified moderate deviations by adjusting the scheduling policy.

[0139] 3. Severe Deviation Handling (Above the Secondary Threshold): When the quantified deviation value M exceeds the preset "secondary threshold" (e.g., a large-scale interruption of core services, severe overload of critical resources, or a sharp increase in business transaction failure rate), it is judged as a severe deviation. At this time, in addition to outputting optimization instructions and suggestions to the intelligent prediction module and the strategy generation module, the closed-loop feedback generation module also simultaneously triggers an alarm notification from the global collaboration module.

[0140] Alarm Notification: The global collaboration module will send emergency alarm messages to the system administrator through various means such as sound, light, electricity, email, SMS, and instant messaging tools, reminding the administrator to intervene and investigate immediately. The alarm message includes details of the deviation, the scope of impact, possible root causes, and the automatic intervention measures already taken by the system.

[0141] Manual intervention: Administrators can log in to the interface of the global collaboration module to view detailed operation logs, performance curves and scheduling status, troubleshoot problems, and issue manual intervention commands through the global collaboration module as needed (such as forced rollback, manual expansion, adjustment of emergency plans, etc.).

[0142] The effectiveness of feedback commands, whether automatically adjusted or manually intervened, undergoes secondary verification through a dynamic evaluation module. This means that after a command is executed, the dynamic evaluation module monitors system performance metrics again, calculates new deviation values, and continues to assess performance compliance. This process forms a complete feedback loop record, facilitating system auditing, experience learning, and continuous improvement, ensuring timely and effective handling of any deviations.

[0143] Furthermore, the deviation quantization calculation module of the present invention calculates the quantization deviation value of the output index based on the difference calculation function as follows:

[0144] ;

[0145] In the formula:

[0146] M represents the final quantification deviation value, which is a comprehensive indicator that reflects the overall degree of deviation between the system performance and the expected target.

[0147] x represents the index of the performance metric. For example, when there are multiple metrics such as CPU utilization, memory utilization, and network latency, x can be 1, 2, 3, etc.

[0148] n represents the total number of performance metrics involved in the calculation, i.e., how many types of performance metrics were considered in this evaluation.

[0149] ω x This represents the weight of the x-th category of metrics. This weight reflects the importance of different metrics in the overall performance evaluation. For example, for core business operations, the weight of service response time may be much higher than that of CPU utilization for secondary services. ∑ω x =1 is usually true.

[0150] A x This represents the real-time actual value (ActualValue) of the x-th category indicator, which is obtained by the indicator real-time acquisition module.

[0151] D x This represents the preset target value for the x-th type of metric. This target value can be the optimization target set by the strategy generation module or the performance baseline specified in the SLA.

[0152] ϵ represents a very small positive constant, whose function is to prevent the target value D from being affected. xThe error of dividing by zero occurs when the result is zero, thus ensuring the stability of the calculation.

[0153] α represents the bias sensitivity coefficient, which is a configurable exponent. When α=1, it is a linear sensitivity; when α>1, its aggravating effect is greater for larger biases, increasing the penalty for severe deviations; when 0<α<1, the sensitivity to bias is reduced. This allows the system to give different levels of attention to different degrees of bias.

[0154] d represents the time decay factor, which is a positive number. As the time step t−t increases... x As the number of data points increases, the impact of older metrics on current deviations decays exponentially. This makes the evaluation focus more on recent performance deviations, reflecting timeliness.

[0155] t represents the current time step of the evaluation.

[0156] t x This represents the time step for collecting data of the xth category of indicators, i.e., the moment when the data of this indicator is collected.

[0157] k x This represents the indicator type coefficient, a normalization coefficient used to adjust for dimensional differences between different types of indicators (e.g., percentages, absolute values, ratios). It ensures that indicators with different dimensions are weighted and summed fairly, or it can be used to apply additional adjustments to certain indicators to better integrate them into the overall deviation calculation. For example, for some indicators, a slight deviation may be serious, while for others, a large deviation is necessary; this coefficient can be adjusted.

[0158] By weighted fusion of real-time deviations from multiple performance metrics and incorporating time decay and sensitivity adjustment mechanisms, a comprehensive, time-aware system status assessment index is constructed. This formula not only considers the differences in importance among various metrics but also amplifies the impact of recent data through a time decay factor and amplifies the contribution of severe deviations using a sensitivity coefficient, thereby achieving a sensitive and comprehensive measurement of system performance deviations. This provides the system with an accurate and adaptive health assessment method, enabling timely identification of service level violations and resource anomalies. It also provides quantitative decision-making basis for the closed-loop feedback generation module, driving continuous optimization of predictive models and scheduling strategies, and enhancing the overall adaptability and robustness of the system.

[0159] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1.A big data-based intelligent education cloud platform elastic resource scheduling system, characterized in that, The system comprises: a data perception module for collecting real-time and historical operation data from a teaching business platform, infrastructure and user terminals, and transmitting the data to a data lake module; a data lake module for receiving and storing raw data from the data perception module, cleaning, integrating and structuring the data to form unified data assets, and providing standardized data to an intelligent prediction module; an intelligent prediction module for performing multi-dimensional time series analysis based on the standardized data provided by the data lake module, and generating quantitative prediction results of resource demand within a prediction time window; a strategy generation module for automatically generating optimal resource scheduling strategy instructions based on the quantitative prediction results output by the intelligent prediction module, in combination with preset optimization objectives and constraint conditions; a transition control module for risk assessment and step decomposition of the resource scheduling instructions issued by the strategy generation module, planning and controlling the specific execution process of resource changes; an elastic execution module for parsing and converting the resource scheduling strategy instructions issued by the strategy generation module into specific control commands recognizable by a resource management module, to drive the underlying resources to complete automatic scaling and configuration changes; a dynamic evaluation module for monitoring the actual operation indicators of the system after scheduling by the elastic execution module, calculating the deviation from the expected target, and feeding back the evaluation results to the data lake module and the strategy generation module to form an optimization closed loop. 2.The elastic resource scheduling system of a big data-based smart education cloud platform according to claim 1, characterized in that: The system further comprises: a resource management module for virtualizing and integrating all heterogeneous physical resources into a globally unified logical resource pool through a unified abstract model and standard interface, and providing resource operation basis for all upper modules; an exception handling module for real-time monitoring of system operation indicators and abnormal signals of the data perception module, and triggering and executing predefined emergency scheduling plans when detecting sudden failures or unexpected loads; a global coordination module for providing a unified visual interactive interface including resource panorama, prediction curve, scheduling status and evaluation report for system administrators, and receiving manual intervention instructions and configuration inputs of the strategy generation module. 3.The elastic resource scheduling system of a big data-based smart education cloud platform according to claim 2, characterized in that: The resource management module comprises: a resource abstraction module for virtualizing computing devices, storage devices and network devices in heterogeneous physical resources into logical resource objects with consistent attributes through a preset unified abstract model and standard interface, thereby forming a logical resource pool that is uniformly recognized and managed by the upper system; a portrait management module for creating and maintaining a dynamic resource portrait for each logical resource object generated by the resource abstraction module, which is dynamically updated by collecting static configuration information, real-time performance indicators and historical load records of the resource, thereby digitally encapsulating the inherent attributes, real-time state and historical performance of the resource; a supply execution module for receiving resource supply intent descriptions from the strategy generation module, matching, filtering and combining calculations among all resource portraits maintained by the portrait management module according to the constraint conditions and optimization objectives in the intent, and finally issuing specific resource allocation, configuration change or recovery instructions to the underlying infrastructure to complete the actual supply of physical resources. 4.The elastic resource scheduling system of a big data-based smart education cloud platform according to claim 1, characterized in that: The intelligent prediction module specifically comprises: A trend prediction module, configured to receive and analyze historical load data with a span of a semester or an academic year, calculate an overall change baseline of resource demand by a long-period time series model, and provide macro and trend framework input for subsequent prediction; A cycle prediction module, configured to receive the trend baseline output by the trend prediction module, focus on recent historical data with a granularity of weeks and days, accurately predict periodic load peaks and valleys that will repeatedly occur in the future days by identifying fixed patterns of course arrangement and teaching activities; An event correction module, configured to monitor announcements and event streams from a teaching affairs system in real time, actively identify unplanned teaching activities, and generate correction signals in real time based on the same, dynamically superimpose and adjust a composite prediction result generated by the trend prediction module and the cycle prediction module to output a final precise resource demand prediction sequence containing sudden load. 5.The elastic resource scheduling system of a big data-based smart education cloud platform according to claim 4, characterized in that: The intelligent prediction module performs the following calculation and processing steps for predicting the amount of resource demand in a prediction time window: S1, the trend prediction module applies a time series decomposition algorithm to the input historical resource utilization rate time series, separates out a trend component sequence, and uses an autoregressive integrated moving average model to fit and predict the trend component sequence, outputting a long-term trend baseline value sequence corresponding to the prediction time window; S2, the cycle prediction module applies a seasonal decomposition algorithm to the same historical resource utilization rate time series, extracts a seasonal component sequence with a basic period of seven days or twenty-four hours, and uses a seasonal autoregressive integrated moving average model to fit and predict the seasonal component sequence, outputting a periodic fluctuation value sequence corresponding to the prediction time window; S3, the event correction module analyzes external structured event data streams, and when a predefined type of unplanned event is identified, queries a preset resource consumption coefficient table according to the event type code, and performs multiplication calculation according to the event scale parameter to obtain the resource demand increment value caused by the event, and then maps the event occurrence time to the prediction time window to generate an event increment value sequence; S4, the prediction engine of the intelligent prediction module aligns the long-term trend baseline value sequence, the periodic fluctuation value sequence and the event increment value sequence in the time dimension, performs point-by-point addition operation, and obtains the final resource demand prediction value sequence; at the same time, based on the variance of the historical prediction error of each sequence, the error propagation law is used to calculate the upper and lower bounds of the prediction interval of each point in the final resource demand prediction value sequence. 6.The elastic resource scheduling system of a big data-based smart education cloud platform according to claim 5, characterized in that: The calculation formula of the intelligent prediction module for predicting the amount of resource demand is as follows: ； where: R represents the resource demand forecast value; i represents the hour number of the day; T represents the long-term trend baseline value; C i represents the i-th hour periodic amplitude; W i represents the ith hour weight; γ represents the time period decay coefficient; E represents the event base resource increment; K represents the event size coefficient; σ 2 represents the historical aggregate prediction error variance; Z represents the confidence coefficient; N represents the total number of historical data samples. 7.The elastic resource scheduling system of a big data-based smart education cloud platform according to claim 1, characterized in that: The dynamic evaluation module comprises: An index real-time collection module, configured to collect original performance index data including resource utilization rate, service response time and transaction success rate from a resource management module, an elastic execution module and a business application interface, and send the data in the form of a time series to a deviation quantification calculation module; The bias quantification calculation module is configured to receive the performance index time sequence sent by the index real-time collection module, and compare the performance index time sequence with the expected target value output by the strategy generation module or the prediction benchmark value generated by the intelligent prediction module point by point, and output the specific quantification bias value of each index at each time through a preset difference calculation function. The performance compliance determination module is configured to receive the quantification bias value output by the bias quantification calculation module, and automatically determine whether the overall and local service level of the current system meets the agreed requirements according to the preset performance compliance threshold at each level, and identify the specific resource entity or service component that does not meet the agreement. The closed-loop feedback generation module is configured to receive the determination result output by the performance compliance determination module, convert the determination result into specific optimization parameter adjustment instructions or strategy modification suggestions according to a predefined feedback logic, and send the instructions and suggestions to the intelligent prediction module and the strategy generation module, respectively, to drive the iterative update of the prediction model and the scheduling strategy. 8.The elastic resource scheduling system of a big data-based smart education cloud platform according to claim 7, characterized in that: The bias quantification calculation module outputs the comprehensive quantification bias value of the index based on the difference calculation function in the following manner: ； In the formula, M represents a quantization deviation value; x represents a performance index serial number; n represents a total number of performance indexes participating in calculation; ω x represents a weight of the xth index; A x represents a real-time actual value of the xth index; D x represents a preset target value of the xth index; ε represents a positive constant; α represents a deviation sensitivity coefficient; d represents a time decay factor; t represents a current evaluation time step; t x represents a collection time step of the xth index data; k x represents an index type coefficient. 9.The elastic resource scheduling system of a big data-based smart education cloud platform according to claim 7, characterized in that: The feedback logic of the closed-loop feedback generation module dynamically adjusts according to the bias level. When the quantification bias value is below the first-level threshold, only the model parameter fine-tuning instructions are output to the intelligent prediction module. When the bias value is between the first-level and second-level thresholds, the parameter adjustment and strategy optimization suggestions are output to the intelligent prediction module and the strategy generation module simultaneously. When the bias value is higher than the second-level threshold, in addition to outputting the optimization instructions, the alarm notification of the global coordination module is triggered synchronously to remind the administrator to intervene in the verification. The execution effect of the feedback instructions is checked again by the dynamic evaluation module to form a complete feedback closed-loop record. 10.The elastic resource scheduling system of a big data-based smart education cloud platform according to claim 1, characterized in that: The strategy generation module includes a strategy knowledge base and a dynamic optimizer. The strategy knowledge base stores and maintains the preset scheduling strategy templates for different education scenarios such as live teaching, online examination and scientific calculation. The dynamic optimizer retrieves the matching strategy template from the strategy knowledge base according to the specific prediction sequence output by the intelligent prediction module and the real-time resource state and cost information fed back by the resource management module, and performs online parameter calculation and multi-objective optimization solution to generate a specific resource scheduling instruction sequence for execution within a specified time window.

Images